Optimality-Theoretic Syntax (Language, Speech, and Communication)

Acknowledgments

The idea for a volume that brings together a growing body of syntactic research in Optimality Theory (OT) grew independently from two sources. In 1997, Johns Hopkins University and the University of Maryland at College Park sponsored an OT workshop. The success of the workshop prompted GeÂraldine Legendre to initiate a publication project, to include the papers and/or posters presented at the workshop. At the same time, Jane Grimshaw and Sten Vikner had begun discussions with potential publishers and with potential contributors, about the possibilities for a volume that would re¯ect current research topics and results in OT syntax. Not surprisingly, there was considerable overlap between the contributions envisaged for the two publications. Once we realized that we were all working on the same project, we decided to join forces and present the best single volume we could construct. This is it. The workshop, organized by GeÂraldine Legendre (Johns Hopkins University) and Linda Lombardi (University of Maryland), with the assistance of Paul Smolensky (Johns Hopkins University) and Laura Benua (University of Maryland), was titled the ``Hopkins Optimality Theory Workshop/University of Maryland Mayfest'' (HOT for short) and brought together researchers in phonology, syntax, sentence processing, and acquisition. It also drew an unexpectedly large audience, more typical of a linguistics conference than a workshop. This volume would never have seen the light of day if it were not for the dedication and support of many individuals and units in ®rst making the workshop happen. We wish to thank, at Johns Hopkins University, the O½ce of the Provost, the School of Arts and Sciences, the Center for Language and Speech Processing, and the Department of Cognitive Science, without whose generous ®nancial backing HOT would never have been possible; at the University of Maryland, College Park, the College of Arts and Humanities and the Department of Linguistics, which o¨ered additional ®nancial support; Linda Lombardi for help with organizing HOT and for handling with expertise and aplomb the consequences of having an audience three times as large as anticipated; Paul Smolensky and Laura Benua for their advice and support; and Tzu-Tzu Wang and her friends

viii

Acknowledgments

for handling many local arrangements. Last but not least, we wish to thank the speakers and the enthusiastic audience at HOT. The task of bringing this volume about turned out to be a much bigger project than any of us had anticipated. We wish to thank the contributors for their dedication and patience, the reviewers for their constructive comments and criticisms, Ben Slade for last-minute proofreading, and Barbara Fisher, Jill Clem, Kathleen Coakley, and Erica Pham for much welcomed help with the index, etc. Finally, we thank our MIT Press editor, Amy Brand, for her help in seeing this project through, our copyeditor, Elizabeth Judd, and our production editor, Sandra Minkkinen.

Contributors

Peter Ackema

Utrecht Institute of Linguistics OTS, Utrecht

Judith Aissen

Department of Linguistics, University of California, Santa Cruz

Eric BakovicÂ

Department of Linguistics, University of California, San Diego

Joan Bresnan

Department of Linguistics, Stanford University

Hye-Won Choi JoaÄo Costa

Department of Modern Languages and Literatures, SUNY Bu¨alo

Departamento de Estudos LinguõÂsticos, Universidade Nova de Lisboa

Jane Grimshaw Edward Keer

Department of Linguistics, Rutgers University

Hawthorne, New Jersey

GeÂraldine Legendre Gereon MuÈller Ad Neeleman

Department of Cognitive Science, Johns Hopkins University

Institut fuÈr Deutsche Sprache, Mannheim Department of Phonetics and Linguistics, University College London

Vieri Samek-Lodovici

Department of Italian, University College London

Peter Sells Department of Linguistics, Stanford University Margaret Speas Sten Vikner

Department of Linguistics, University of Massachusetts, Amherst

Institut fuÈr Linguistik/Germanistik, UniversitaÈt Stuttgart

Colin Wilson Ellen Woolford

Department of Linguistics, University of California, Los Angeles Department of Linguistics, University of Massachusetts, Amherst

Overview

This volume o¨ers a window on current research that extends Optimality Theory (OT), developed in Prince and Smolensky 1993, to the domain of syntactic representation. The range of topics addressed in the individual chapters is remarkably broad, with researchers developing proposals ranging from morphosyntax (case, voice, cliticization) through the syntax of reference (control, anaphora, and pronominalization) and the grammar of clauses (complementizers and their absence) to grammatical and discourse e¨ects in word order. The coherence and repeated themes of the chapters, despite their empirical breadth, re¯ects a broad consensus as to the important results of OT and its syntactic applications, and as to the central research questions of current work. A chapter summary may be helpful. The original contributions are preceded by an introduction to OT in syntax. Written by GeÂraldine Legendre, chapter 1 sets the stage for readers not familiar with the framework. In chapter 2, Peter Ackema and Ad Neeleman address the question of how apparently parallel morphological and syntactic expressions are assessed in grammatical competition. They argue that wherever possible, syntactic expression of lexical items is preferred, and that morphological expression is possible only where other factors intervene, such as the presence of a derivational a½x in an input. They apply this proposal in detail to compounds, idioms, and verb-particle combinations in Dutch, and argue that the model predicts the mixture of syntactic and morphological properties observed in these constructions. The privileged status of syntactic expression follows from a constraint against morphological operations, and the position of this constraint in the hierarchy underlies typological variation in the realization of lexical items. The research reported by Judith Aissen in chapter 3 concerns voice, and its relationship to person in particular. Why does a language have no passive with a ®rst- or second-person agent, or no active with ®rst- or second-person object? Aissen argues for interpreting language-particular hierarchies that ®gure in analyses of voice as consequences of universal constraints (derived from harmonic alignment of universal

xii

Overview

scales) governing combinations of person and grammatical relation, ranked in each individual grammar. This proposal makes it possible to unify otherwise unrelated restrictions on passives and actives. Aissen further develops the view that syntactic markedness (as determined by the ranked constraints) is accompanied by morphological marking, so that absence of morphology is linked to the unmarked association of person and semantic role. The proposed explanation for this relationship invokes local constraint conjunction of a constraint banning null morphology with markedness constraints that ban marked combinations of grammatical role and person. A candidate that has null morphology and contains a marked combination violates the conjoined constraint, while the constraint is satis®ed by a candidate that has overt morphology signaling the marked combination of person and role. Eric BakovicÂ's and Ed Keer's contributionÐin chapter 4Ðdiscusses the ``neutralization'' approach to optionality and ine¨ability in OT. Both optionality and ine¨ability at ®rst seem problematic for OT. The role of Eval is to ®nd out which of the possible realizations of a given input is optimal. Optionality problems arise when more than one of the realizations of a given input seem grammatical, and ine¨ability problems arise when none of the realizations of a given input seem grammatical. Bakovic and Keer argue that ine¨ability arises if a markedness constraint is higher ranked than a con¯icting faithfulness constraint, selecting as the output a form that is quite surprising given the input. This is the source, in their proposal, of the apparent lack of an output for some inputs. Such rankings also make it possible for the optimal outputs of two di¨erent competitions to be identical, a case of neutralization. When the inputs and also the optimal outputs of two di¨erent competitions di¨er only in features that have no apparent consequences for the interpretation (e.g., presence or absence of that in embedded ®nite declarative clauses in English), the resulting situation mimics free choice between alternativesÐthat is, optionality. Joan Bresnan's chapterÐchapter 5Ðtakes as its starting point universal markedness relations among pronouns. She distinguishes between zero pronouns (inaudible), bound pronouns (pronominal in¯ections on a host), clitic pronouns (cliticized to a host), weak pronouns (free morphemes unable to receive primary stress), and free pronouns (free morphemes capable of receiving primary stress). The earlier in this list a class occurs, the more marked it is: languages may lack bond and/or zero pronouns but never free pronouns; pidgin languages almost exclusively have free pronouns; and free pronouns are the ones used when a gap occurs in one of the other paradigms. Based on two other markedness relations, Bresnan suggests two Harmony constraints to the e¨ect that only reduced (nonfree) pronouns are specialized for topic anaphoricity and that only overt (nonzero) pronouns are inherently speci®ed for person, number, or gender. Also assumed are a feature faithfulness constraint and two constraints inspired by Haiman's work on iconicity and allotaxy (see Bresnan's chapter for references). With these constraints and with the assumption that accidental lexical

Overview

xiii

gaps result from higher-ranking constraints, a highly restrictive account is presented of the various asymmetries in pronoun distribution found across languages, including the markedness relations just mentioned. In chapter 6, Hye-Won Choi focuses on variation within a language and on how the nature of constraint interaction changes across discourse contexts or inputs. Along with Costa, Legendre, and Samek-Lodovici, she proposes that arguments are speci®ed for discourse status in the input. In her analysis, each argument may be speci®ed for two independent features [GNew], [GProminent]. She derives the mixed A/A0 -binding behavior of scrambling in German, and the fact that focus e¨ects do not emerge across the board, from a competition between syntactic constraints and information structure constraints. She relies on local conjunction of binding constraints to explain the pattern of A binding and A0 binding e¨ects. Only when both SyntacticRank (favoring a subject over an object binder) and LinearPrecedence (favoring a binder-bindee order) are violated in the same local domain (i.e., by the binder) is binding ungrammatical. In other words, the conjoined constraint SR&LP outranks the individual constraints SR and LP. Scrambling always wins when the binder is the object and the bindee is the subject because the nonscrambled order fatally violates the conjoined constraint. The constraint interaction is more complex and interesting when the GFs of the binder and bindee are reversed (binder ˆ subject; bindee ˆ object). In that case, the conjoined constraint is satis®ed and a lower-ranked constraint determines the optimal output: scrambling of the bindee when only the bindee is [‡Prominent], no scrambling when both the bindee and the binder are [‡Prominent], and the constraint Subject is decisive, favoring subjectobject order. If the binder is [‡Prominent] and the bindee [ Prominent], no scrambling is again optimal (all constraints are satis®ed). Finally, if neither binder nor bindee is prominent, no scrambling again prevails because discourse constraints are vacuously satis®ed, and Subject favors the binder-bindee order. JoaÄo Costa's chapterÐchapter 7Ðfocuses on a fundamental and pervasive pattern of constraint interaction in OT, resulting in the emergence of the unmarked structure (see the chapter for references). He argues that unmarked word order (i.e., the word order found in sentence-focus contexts, such as answers to the question ``What happened?'') is not to be understood as underived word order (i.e., word order instantiating theta relations) but rather as a consequence of optimatization for inputs in which all elements are speci®ed for focus. He assume constraints like AlignFocus, which regulates the position of focused elements, and SubjCase, which forces the subject to move to SpecIP. Given an input where all elements are speci®ed for focus, and a particular constraint ranking, say that of Portuguese, a high-ranked constraint like AlignFocus is satis®ed by all alternative candidates (hence inactive) and a lowerranked constraint like SubjCase (whose e¨ect is rendered invisible when AlignFocus is active) becomes decisive. The result is SVO order. Suboptimal candidates are elimi-

xiv

Overview

nated in sentence-focus contexts not because they violate high-ranked constraints but because they violate low-ranked constraints; the optimal candidates satisfy these and show the unmarked word order for a given language. Jane Grimshaw's proposal in chapter 8 revolves around the lexical representation of Romance clitic systems. She argues that the relative order of clitics in combination can be predicted from alignment constraints, given a particular set of assumptions about the morphosyntactic analysis of the clitics. Markedness and faithfulness constraints in turn derive the clitics' lexical representations. Hence the inventory of clitics is not lexically stipulated but is an epiphenomenon of ranked grammatical constraintsÐthat is, grammars. Further characteristics of the clitic systems, such as the absolute impossibility of some combinations and the ``spurious se'' phenomenon, are derived from the same constraint interactions. GeÂraldine Legendre's contribution in chapter 9 addresses the positioning of ®nite verbal categories and, more generally, the status of in¯ection in the grammar. Building on her previous work on cliticization as PF-spell-out of morphosyntactic features, she argues that the positioning of ®niteness takes place at PF rather than in the syntax (as is assumed in the functional category approach to morphosyntax). She examines V2 e¨ects in a VSO language (Breton) and an SVO language (Macedonian), and presents evidence that PF alignment is crucially involved. PF alignment may, however, be masked by overriding constraints. When e¨ects of overriding constraints are controlled for, the unmarked pattern emerges, here V2. The analysis exploits the fact that candidates are construed as global structures subsuming several levels of representation in standard, representationally oriented OT. In this view, word order is a complex interface phenomenon resulting from the parallel evaluation of markedness constraints on syntactic structures (including economy of movement and structure) and constraints on syntactic alignment, PF alignment, and information structure alignment with input-output faithfulness constraints. Chapter 10, by Gereon MuÈller, o¨ers an explanation for why syntactic movement is fundamentally, though not consistently, order preserving. MuÈller proposes a constraint ParMove whose e¨ect is to preserve asymmetric c-command relations among arguments across levels of representation (D-structure, S-structure, and LF). Three major patterns of constraint interaction are discussed. When ParMove con¯icts with a constraint requiring fronting, ParMove will be violated if it is the lower ranked of the two, with the result that its e¨ects are invisible and order is not preserved. When ParMove is the higher ranked, or when there is no con¯ict between the two constraints, ParMove is satis®ed and the optimal output respects order preservation. MuÈller grounds his analysis in a wide range of movement phenomena, including whphrases, pronouns, and QR in a number of languages. Vieri Samek-Lodovici's contribution in chapter 11 addresses the size of an OT typology in the empirical domain of presentational focus constructions. Four con-

Overview

xv

straints (Subject, Stay, AlignRight, and AlignLeft), which regulate the position of focused arguments, have 4! distinct rankings. For 3 distinct inputs (for example, with no argument focused, the subject focused, or the object focused) there appear to be 13,824 possible optima in principle, 24 for each input. Samek-Lodovici argues that two basic formal properties of OTÐharmonic bounding of candidates and constraint ranking consistency across inputsÐconspire to dramatically limit the size of cross-input typologies or possible languages to a mere 8. A candidate is said to be harmonically bounded by a potential winning candidate (i.e., a candidate optimal under at least one constraint ranking) if it incurs the same violations as a potential winner plus additional ones. A harmonically bounded candidate can never be optimal under any constraint ranking. Hence, all harmonically bounded candidates can be eliminated from consideration in a typology. Moreover, constraint ranking is ®xed for a given languageÐthat is, it is constant across inputs (or contexts). Hence, contradictory rankings are also eliminated from a typology. Samek-Lodovici demonstrates that this restrictive typology exhausts the interaction of the four relevant constraints and predicts the attested patterns of leftward focus, rightward focus, and partial focus in languages with basic SVO and VSO word order, such as Kera (Chadic), Italian, English, Aghem (Bantu), Scottish Gaelic, Spanish, and Podoko (Chadic). The core of Peter Sells's work in chapter 12 concerns the typology of voice systems. He posits inputs in which each argument has one of three degrees of prominence. Ranked markedness constraints then select a single least marked (and necessarily unfaithful) candidate for all distinct inputs. The winning candidate is morphologically simplest, to match its unmarked status with respect to voice. The derivation of inventories that include more than one voice crucially involves inventory-level constraints. These constraints force a system to encode one or more distinctions, and hence determine the number of contrasts a voice system admits. The contrasting forms are the least marked items that express contrasting prominence relations. In this proposal, then, the candidates are inventories, not individual verbal forms. The typology derived from this system predicts implicational hierarchies among voice forms. Some voices can exist in a given system only if another voice exists also. In chapter 13, Margaret Speas illustrates how constraint violability allows greatly increased generality not only for constraints but also for entire analyses. Null subjects of in®nitivals (found in all languages), null subjects of ®nite clauses (found only in some languages), and null objects (found only in a subset of the languages that have null subjects of ®nite clauses) are instances of one element, Pro (see citations in her chapter). Crosslinguistic variation in the licensing and interpretational properties of Pro are derivable from alternative rankings of a few violable constraints. Inputs may contain either an overt or a null pronoun. Max(Pro) is violated if a null pronoun in the input is not realized as a null pronoun in the output. Case is violated only if overt

xvi

Overview

pronouns are not assigned case. Control is violated if Pro has no antecedent in the ®nite clause, whereas BindingTheoryB is violated if Pro is bound in its clause: which among these two constraints is highest ranked determines the interpretational properties of Pro. The ranking of these two constraints with respect to Max(Pro) and Case determines whether Pro is possible at all. Speas discusses di¨erences between her own analysis and those of Bresnan and of Samek-Lodovici, arguing that a closer examination of ``point of view'' might make a uni®cation of the three analyses possible. In chapter 14, Sten Vikner analyzes the various positions of the ®nite verb in Danish, English, French, and Icelandic. The chapter tests the claim that the main or only source of language variation lies in di¨erences in constraint ranking. Earlier research (see Vikner's chapter for references) shows how constraints on clauses can explain why the complementizer is obligatory in embedded topicalizations; Vikner's extension explains why the complementizer is optional in embedded clauses only in those languages where the ®nite verb does not undergo V-to-I o movement. Vikner further shows how the relative ranking between two constraints, Economy of VPs (a variant of *Structure, which is violated by all and only VPs) and V-in-V o (violated by verbs not inserted under V o ), may determine whether nonthematic verbs are inserted under V o or above VP. The same ranking also determines whether a ®nite main verb may leave VP or has to be supported by do-insertion in negated sentences and questions. Chapter 15, by Colin Wilson, is centered around the notions of referential economy and relativized minimality. He shows how capturing such generalizations in terms of constraints that are able to compare between candidates (an example might be MostLocalAntecedent: An anaphor must be bound by the most local potential antecedent'') is undesirable not only empirically but also conceptually: the ability to compare between candidates is one that ideally belongs to grammars and not to individual constraints. Wilson advocates partitioning the optimization process into two parts: an interpretive optimization followed by an expressive optimization. In the former, the competition is between di¨erent semantic interpretations of one and the same syntactic expression (``Which is the optimal interpretation of a given expression?''), whereas in the latter, di¨erent syntactic expressions that share a single semantic interpretation compete against each other (``Which is the optimal expression of a given interpretation?''). The ordering of the two optimizations means that only well-formed outputs of an interpretive optimization may be input to an expressive optimization. This model is compared to a single optimization model, where candidates di¨er from each other with respect to both syntactic expressions and semantic interpretation (``Which is the optimal expression and the optimal interpretation of a given input?''). The analysis is based mainly on evidence from Chamorro (see citations in Wilson's chapter), where Wilson argues that the complex interplay between

Overview

xvii

surface ®lters on transitive clauses and anaphora cannot be successfully analyzed with a single optimization. In chapter 16, Ellen Woolford lays out a general view of case theory, focusing on languages in which the case of an argument can depend on the case of another argument, or on the transitivity of the verb. She suggests that case principles of a familiar sort are universal and inviolable: every NP has a case; nominative case is licensed by tense; accusative case is licensed by a verb. Moreover, a verb can assign inherent case when its lexical entry is appropriately speci®ed. Where these principles allow for a choice among case markings for an NP, the choice is made by ranked, violable faithfulness and markedness constraints. The faithfulness constraints favor assignment of an inherent case when the lexical entry of a verb provides one, and markedness constraints favor structural cases, hence a verb lexically speci®ed to assign inherent case may or may not occur with lexical case in the output, depending on the relative ranking of the markedness constraint against inherent case and the faithfulness constraint that preserves inherent case from the input. The markedness constraints for structural case are organized into a universal relative ranking, and hence always favor the least marked structural case. Nominative is preferred over accusative, which is more marked, hence nominative appears on subjects of unaccusatives, and on objects when the subject is inherently case marked. Some important theoretical themes run through the works presented here. Among constraints, the interplay between faithfulness and markedness of course ®gures prominently in most of the analyses; it is the foundation for the proposals of Bakovic and Keer, Wilson, and Woolford, for example. New constraints, or types of constraints, are introduced in some works: Vikner, for instance, posits a new kind of markedness constraint, and Woolford relativizes faithfulness to a domain. In addition, alignment plays a prominent role in a number of proposals, especially regarding word order, such as the proposals by Choi, Costa, Grimshaw, Legendre, and SamekLodovici. Problems of (apparent) free variation and its limits have naturally attracted the attention of several researchers, particularly Bakovic and Keer, MuÈller, and Choi. The proper understanding of markedness is central in several chapters, especially those of Aissen, Bresnan, and Sells, who are also concerned with the relationship between abstract syntactic markedness and morphological/syntactic overtness. The potential role of constraint conjunction in banning particularly marked forms (those that locally violate more than one constraint) is examined by Aissen, Choi, and Grimshaw. The emergence of the unmarked, resulting from satisfaction of lowerranking constraints when higher-ranking con¯icting constraints are irrelevant, lies behind the analyses developed by Ackema and Neeleman, Costa, Legendre, and MuÈller: when nothing interferes, an unmarked structure for syntactic elements emerges. Properties of factorial typologies are the prime focus for Samek-Lodovici,

xviii

Overview

but almost every chapter has typological concerns as an underlying theme (for example, complementation in languages with V-to-I raising, as discussed in Vikner's chapter). The status of inventories as derived entities is an important result in the chapters by Bresnan, Grimshaw, Sells, Speas, and Woolford. Ackema and Neeleman, Legendre, MuÈller, Sells, and Wilson all raise the question of the precise nature of syntactic competition: What levels of representation are involved in competition; does morphology compete with syntax; are inventories in competition with inventories? Is more than one kind of optimization involved? These chapters illustrate the empirical basis and fundamental content of current developments in OT research in syntax. They also lay out some of the results obtained so far and the challenges to be faced in the future. We hope that this collection proves valuable in shaping and informing the future research direction of syntactic theory. GeÂraldine Legendre, Baltimore, Maryland Jane Grimshaw, New Brunswick, New Jersey Sten Vikner, Stuttgart, Germany Reference Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html.

Chapter 1 An Introduction to Optimality Theory in Syntax

GeÂraldine Legendre

In developing generative theories of syntax the challenge has long been to reconcile two opposite kinds of empirical evidence. On the one hand, there is substantial evidence that languages of the world are deeply similar; on the other, there is also substantial evidence that they di¨er on the surface. For example, there are languages that allow sentences without a pronounced subject and those that do not, languages with prepositions and with postverbal objects and languages with postpositions and preverbal objects, languages that move question words to the front of the sentence and languages that leave question words inside the sentence. Historically, principles-and-parameters theory (P&P), dating back to Chomsky 1981, has taken one of the possible (but potentially weak) positions on the issue and claimed that principles of universal grammar (UG) are both universal (by de®nition of UG) and inviolable. A system of inviolable principles can be made compatible with the irrefutable evidence of crosslinguistic variation by appealing to parameters. For example, in earlier versions of P&P, whether a given principle is active or not at individual levels of syntactic representation in individual languages is determined by parameters. In this view, parameters circumvent inviolability by turning universal principles o¨. An important property of such a system is that once a principle is turned o¨ in a language, it is predicted not to show any e¨ects in that language. That is, parameter values are set once and for all for a given language. To take only one example, in the Minimalist Program (Chomsky 1995), the strength of functional features is speci®ed in the lexicon as either strong or weak on a language-particular basis. This imposes restrictions on the computational system that work largely independently of one another. Empirically, however, the presence of a particular syntactic pattern within a given language is frequently only partial. The expletive es subject in German impersonal passives illustrates this point. Sentences (1a) and (1b), but not (1c), are grammatical: (1) a.

Es wurde schoÈn getanzt. it was beautifully danced `The dancing was beautiful.'

GeÂraldine Legendre

2

b. SchoÈn wurde getanzt. c. *SchoÈn wurde es getanzt. In the standard V2 analysis of German whereby the ®nite verb is located in C (den Besten 1983; Vikner 1995; and many others), the abverbial phrase schoÈn is in specCP in (1b) and (1c) while SpecIP is empty in (1b) but ®lled with es in (1c). IfÐas everyone assumesÐthe presence of expletive subjects is to satisfy the Extended Projection Principle (EPP), a principle requiring that every clause has a subject (Chomsky 1982), then the EPP is satis®ed in (1a), but not obviously in (1b): schoÈn is an adverb in SpecCP (V2 constraint) and the expletive is missing.1,2 Note that the EPP is also satis®ed in (1c) and so is V2, yet the sentence is ungrammatical. Hence, any system of inviolable principles will require some ®ne-tuning in light of these facts: either the EPP or the principle linking expletives to the EPP would have to be complicated on a language-speci®c basis. Alternatively, violable constraints are involved and the surface pattern is the result of constraint interaction. As mentioned in Vikner 1995:186, Cardinaletti 1990:17, for example, appeals to ``Avoid Pronoun,'' a violable principle ``outranked'' by a V2 constraint. The intuition pursued in the line of work represented in this volume is that it is incorrect to assume that UG principles are inviolable while supplemented by languageparticular parameters. The correct assumption, we claim, is that principles or constraints are pure or uncomplicated in their form but violable. So, we can assume that the EPP is indeed satis®ed in (1a), but violated in (1b), under compulsion from some higher-ranked constraint, yet to be determined. This is the strong position taken in Optimality Theory (OT; Prince and Smolensky 1993). Principles of UG are universal but violable if necessary to allow satisfaction of a higher-ranked (more important) principle. The concepts of violability and constraint interaction are formalized in such a way that crosslinguistic variation derives from alternative rankings of the same principles. This introductory chapter reviews the main claims of Prince and Smolensky's original OT as they pertain to syntax and illustrates them, based on some appropriate examples (including an analysis of the above German pattern). This chapter also addresses a number of theoretical issues that syntax poses for OT, including the nature of the input and the question of ``ine¨able'' structures (inputs for which a given language may not have a grammatical output). Its purpose, however, is merely to present a somewhat simpli®ed overview and make the rest of the volume accessible to a reader knowledgeable in syntax but not previously exposed to OT.3 It does not articulate or defend many remaining subtleties of OT syntax. In fact, the rest of the volume builds on the claims reviewed in the present chapter and o¨ers a rich set of OT analyses that, to some extent, stretch the boundaries of the theory presented here. This is to be expected and welcome, since this approach to

Introduction to Optimality Theory in Syntax

3

syntax is now just coming of age by confronting the daunting complexity of crosslinguistic patterns. Aside from section 1.1, which sets the stage, each section of the chapter focuses on one major aspect of the theory. Section 1.2 presents the general architecture of the theory. Section 1.3 discusses the notion of competition central to OT constraint interaction. Section 1.4 focuses on the general and universal character of OT constraints, while section 1.5 discusses patterns of violations. Section 1.6 deals with typological issues. The concept of economy, central to OT, is discussed in section 1.7. Optionality is the topic of section 1.8. In section 1.9, we take up the question of ``ine¨ability,'' which, in turn, leads to a discussion of the input in section 1.10. Section 1.10 also includes some ®nal comments on the uni®ed OT approach to both syntax and phonology. 1.1

What Is Optimality Theory?

OT is a formal theory of constraint interaction in UG. It is not a substantive theory of any phenomenon, syntactic or otherwise. It is not committed to any speci®c type of structural (or markedness) constraints. This is why OT analyses inspired by di¨erent types of substantive theories of syntax can be found in the literature, including Government-Binding Theory (e.g., Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998; Grimshaw 1997; as well as many chapters in the present volume), Lexical-Functional Grammar (e.g., Bresnan, chap. 5, and Sells, chap. 12, this volume), and the Minimalist Program (e.g., MuÈller, chap. 10, and Speas, chap. 13, this volume). Rather, OT's main hypotheses are the following: (2) a. UG is an optimizing system of universal well-formedness constraints on linguistic forms. b. Well-formedness constraints are simple and general. They routinely come into con¯ict and are (often) violated by the surfacing form. c. Con¯icts are resolved through hierarchical rankings of constraints. The e¨ect of a given constraint is relative to its ranking, which is determined on a language-particular basis. d. Evaluation of candidates by the set of constraints is based on strict domination. For any two constraints C1 and C2 , either C1 outranks C2 or C2 outranks C1 . e. Alternative structural realizations of an input compete for the status of being the optimal output of a particular input. The most harmonic outputÐthe one that best satis®es, or minimally violates, the full set of ranked constraints in a given languageÐis the optimal one. Only the optimal structure is grammatical. f. Every competition yields an optimal output.

GeÂraldine Legendre

4

Each point is addressed in turn in the following sections, starting with some remarks on the general architecture of OT. 1.2

The Architecture of OT

OT relies on an input-output mapping architecture. Provisionally, the input to optimization in syntax can be assumed to consist of predicate-argument structure, functional features, and lexical items (See sections 1.3, 1.8, and 1.10 for further discussion). For a given input, the grammar generates and evaluates a potentially in®nite set of output candidatesÐthe candidate setÐwhich consists of alternative structural realizations of that input. The component of the grammar responsible for generating the candidate set corresponding to a particular input is called Gen (for Generator). In syntax, Gen has so far been assumed to generate only candidate structures that respect basic X 0 theory principles (thus in e¨ect taking X 0 theory to be a system of inviolable constraints). The set of universal well-formedness constraints is called Con (for Constraints). The component responsible for evaluating the candidate ouputs is called Eval (for Evaluator). Evaluation of candidate outputs relies on a set of hierarchically ranked constraints of Con: C1 g C2 g


Cn . Note that the constraint ranking constitutes the language-particular component of the grammarÐthat is, it is the only component that admits variationÐwhile the set of constraints itself is claimed to be universal. OT relies on a unique type of constraints to regulate the input-output mapping. These input-output faithfulness constraints limit how far candidate outputs may di¨er from the input. They require the output to express all and only the properties of the input. Faithfulness constraints are crucial to the OT conception and have played a pivotal role since the theory's inception (contra Chomsky 1995:380, note 4). Without them all input structures would map to the same, least marked, output. 1.3

Constraint Con¯ict and Ranking

OT makes the claim that languages cannot di¨er in their well-formedness criteria but only in which criteria have priority in cases of con¯ict.4 Two constraints con¯ict when satisfying one entails violating the other. A simple case of constraint con¯ict can be postulated for the existence of expletive elements like it in English, which has no counterpart in Italian, as discussed in Grimshaw and Samek-Lodovici 1998:203±208. (3) a. It rained. b. Piove.

Introduction to Optimality Theory in Syntax

5

The relevant constraints are two that have been proposed on the basis of do-support in English (Grimshaw 1997), and are given in (4). They capture the core ideas of the EPP and the Principle of Full Interpretation (Chomsky 1991), respectively. Grimshaw and Samek-Lodovici (1998:194) explicitly note that their Subject is violated by clauses without a subject in the canonical position. (4) a. Subject: The highest A-speci®er in an extended projection must be ®lled. (Grimshaw and Samek-Lodovici 1998) b. FullInt: Lexical items must contribute to the interpretation of a structure.5 These two constraints con¯ict in the case of weather verbs because the latter do not select for a thematic argument. One option is for a weather verb to surface without a subject, in which case it satis®es FullInt (the corresponding structure [rains] contains no lexical item that does not contribute to its interpretation). This, however, entails a necessary violation of Subject (the structure is an IP whose speci®er is not ®lled). Alternatively, a weather verb may satisfy Subject by surfacing with an expletive subject. This, however, entails a necessary violation of FullInt, because the expletive it does not contribute to the interpretation of the structure. Thus, there is no possible output that satis®es both constraints. The con¯ict is resolved by hierarchically ranking the constraints. Note that ``strict domination'' is assumed to hold over the rankings (rather than weighting), which means that any higher-ranked constraint takes absolute priority over any lowerranked constraint (i.e., a single violation of a higher-ranked constraint is always worse than any number of violations of any number of lower-ranked constraints). Note that strict domination is a particular theoretical assumption, which further speci®es constraint violability.6 Competitions can be made formally explicit in tableaux like T1.1. The optimal candidate is identi®ed by the pointing ®nger f. Constraint ranking is indicated by the left-to-right order, each constraint dominating the ones on its right. Violations of constraints are recorded as * in individual cells; *! are fatal violations for suboptimal candidates, while Y are violations incurred by optimal candidates. As mentioned, the input includes lexical items such as verbs, their argument structure, and tense speci®cations. In the case of weather verbs, however, the argument structure contains no argument position.7 The two optionsÐsurfacing with a subject and surfacing withoutÐconstitute the candidate set of structures to be evaluated by the constraints in Con {Subject, FullInt}.8 If FullInt outranks Subject, as it does in tableau T1.1, it is less important to satisfy the lower-ranked constraint, Subject, than the higher-ranked one, FullInt. The result is that candidate (b) is better or more harmonic than (a) with respect to the constraint ranking in tableau T1.1. Hence (b) emerges from the comparative evaluation as the optimal candidate and is thus the only one to be grammatical.

GeÂraldine Legendre

6

Tableau T1.1 Italian (Input: piovereV [present]) FullInt a. EXPL piove

Subj

*!

f b. Piove

Y

Tableau T1.2 English (Input: rainV [present]) Subj f a. It rained b. Rained

FullInt Y

*!

In tableau T1.1, the grammatical structure, piove, ends up violating some universal constraint: Subject. Violations are relative, however, and the one incurred by optimal candidate (b) is only of the lower-ranked constraint. Hence, it is a minimal violation. If, however, Subject outranks FullInt, the structure containing an expletive subject will be optimal. As shown in tableau T1.2, the minimal violation incurred by the optimal candidate is of lower-ranked FullInt. Its competitor (b) fares worse because it violates the higher-ranked constraint, Subject. Thus, the di¨erent English and Italian patterns result from the two logically possible rankings of the two constraints.9 Note an important aspect of this analysis of expletives. Whether a given language has expletive subjects or not depends on the relative ranking of Subject and FullInt. In other words, it is not the case that some languages have an expletive subject in their lexicon while others do not. It is the syntax that determines whether some element should be drafted to serve as an expletive.10 We can further exploit this analysis of expletive subjects, based on constraint interaction and minimal constraint violations, to explain the fact that German has expletive subjects though they need not be systematically deployed in impersonal passives lacking a thematic subject. The discussion of German also serves to illustrate the role of the input in OT syntax as well as various possible patterns of constraint interaction. The appropriateness of expletive es in impersonal passives is best seen in the context of question-and-answer pairs. If the question is a general one about the event (5a), only an answer with es (5b) is natural.

Introduction to Optimality Theory in Syntax

7

(5) a.

Was geschah? `What happened?' b. Es wurde schoÈn getanzt. it was beautifully danced c. aSchoÈn wurde getanzt.11 beautifully was danced wurde es getanzt. d. *SchoÈn beautifully was it danced

If, however, the question is about a property (or location) of the dancing itself, the pattern is a bit more complicated. If the answer simply conveys new information, then (6b) with es is the natural answer. The only di¨erence with (5b) lies with additional stress falling on the adverb schoÈn in (6b), represented in caps. If the adverb conveys information that is not only new but noteworthy (i.e., unexpected), the natural answer is (6c) with additional stress on the fronted adverb. Native speakers report that (6b) and (6c) are not instances of contrastive stress, which, in German, requires heavier stressing than new information does.12 (6) a.

Wie wurde getanzt? how was danced `How was the dancing?' b. Es wurde schoÈn getantzt. it was beautifully danced c. SchoÈn wurde getanzt. beautifully was danced d. *SchoÈn wurde es getanzt. beautifully was it danced

In any discourse situation, the answer with es plus fronted adverb (5d±6d) is ungrammatical.13 The intuitive explanation for the distribution of expletive es in (5) and (6) is that es surfaces only where absolutely necessary. The source of the contrast can be attributed to the input by assuming that it encodes information structure features like [new], [noteworthy], and so on (see Choi 1996; Samek-Lodovici 1996, 1998; Costa 1998; Legendre 1999; and several chapters in the present volume for related proposals). As discussed in these references, elements focalized by virtue of encoding these input features are subject to alignment constraints (McCarthy and Prince 1993a, 1993b) operating in a particular syntactic domain. Because German treats ``new'' information di¨erently from ``new and noteworthy'' information, it is necessary to assume two constraintsÐAlignNew and AlignNoteworthy, which align the focalized element with the left edge of VP and the clause, respectively.14 Alignment constraints interact with the two constraints discussed earlier, Subject and FullInt. Recall that Subject, as its name indicates,

GeÂraldine Legendre

8

requires every clause to have the highest A-speci®er (SpecIP) ®lled with a subject. Subject is violated whenever the relevant position is ®lled with the adverb schoÈn. To see how the German competitions play out, consider initially two of the relevant inputs in German. When schoÈn conveys ``new and noteworthy'' information as identi®ed in the input, SCHoÈN wurde getanzt is optimal. That is, it is less costly to violate Subject than the alignment constraints. But when schoÈn is not focalized (i.e., no information structure feature is present in the input), the alignment constraints are vacuously satis®ed; Subject requires that SpecIP be ®lled with an expletive subject at the cost of violating FullInt. The result is es wurde schoÈn getanzt. In neither competition can a candidate with both fronting and expletive es, *schoÈn wurde es getanzt, emerge as a winner. The reason is economy: Either fronting or expletive es is enough to produce an optimal candidate. Doing both is overkill. Its formal account relies on a constraint family favoring economy of building blocks, *Structure (Prince and Smolensky 1993). A particular member of the *Structure family, MinimalProjection (MinProj, Grimshaw 1993) penalizes the building up of maximal projections in syntax. (See further discussion and tableaux below.) We now turn to the individual competitions. Because es is the focus of discussion, I have simpli®ed the tableaux to the maximum. For example, I am ignoring the fact that prosodic constraints are relevant to a full analysis of focalization e¨ects in German and that candidates with a particular stress pattern compete with candidates with a di¨erent stress pattern. I am also ignoring violations of *t/Stay (Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998; Grimshaw 1997) since they do not a¨ect the outcome. Finally, CP structures compete with IP structures, with the former systematically violating MinProj. I am discounting maximal projections (IP, Tableau T1.3 German (Input: tanzenV (x); x ˆ 0; [past]; schoÈnAdv [new]) AlNewVP

MinProj

Subj

f a. [ IP es wurde [VP schoÈn getanzt ]] b. [ IP

Y

wurde [VP schoÈn getanzt ]]

*!

c. [ IP schoÈnj wurde [VP tj getanzt]]

*!

d. [CP schoÈnj wurdei [ IP es ti [VP tj getanzt]]]

*!

*

e. [CP schoÈnj wurdei [ IP getanzt]]]

*!

*

ti [VP tj

Full-Int

* * *

Introduction to Optimality Theory in Syntax

9

VP) that do not distinguish candidates from one another in tableaux T1.3±T1.5 (pp. 8±11); hence only MinProj violations incurred by CPs are recorded. Tableau T1.3 represents the competition for an input in which schoÈn carries the input feature [new].15 An IP structure is all that is needed to best satisfy the constraint ranking in tableau T1.3. As we saw earlier in English, expletive subjects result from the basic ranking: Subject g FullInt, which also eliminates candidate (b). Candidate (c) is eliminated because it violates AlignNew: schoÈn is aligned with the left edge of IP, not VP. Candidates (d) and (e) violate AlignNew and Economy (MinProj ), hence they also are eliminated. Only candidate (a) incurs a minimal violation (FullInt), hence it is optimal and grammatical. The fact that German has expletive subjects does not, however, imply that the expletive subject structure is always optimal. This is shown in tableau T1.4, where schoÈn carries two input features [new] and [noteworthy]. Assuming that each feature is sensitive to its own alignment requirement, candidate (c) is optimal if AlignNoteworthy outranks AlignNew. In fact, candidate (c) is optimal in tableau T1.4 despite the absence of an expletive subject because all its competitors fare worse. IP structures in which schoÈn is not fronted (candidates (a) and (b)) fatally violate AlignNoteworthy, while other competitors succumb to MinProj (candidates (d) and (e)). Note that the optimal candidate (c) in tableau T1.4 violates Subject, resulting in the focalized adverb appearing in SpecIP, with no violation of MinProj. The nature of the competition changes again for an input in which schoÈn does not carry any information structure features. AlignNew and AlignNoteworthy are vacuously satis®ed, which I have represented by their absence in tableau T1.5. The e¨ect of MinProj and Subject becomes visible as they eliminate suboptimal candidates. Once more, nothing is gained from the addition of structure. SpecIP can be ®lled, either with expletive es (a) or adverbial schoÈn (c). The di¨erence is a violation of Subject, fatally incurred by schoÈn. Candidate (c)Ðoptimal in tableau T1.4Ðnow loses, given the ranking Subject g FullInt established earlier. Candidate (a), in fact, beats all its competitors, all of which violate a constraint that outranks FullInt.16,17 The outcome of the competition in tableau T1.5 is an instance of the ``Emergence of the Unmarked'' (McCarthy and Prince 1994). The unmarked pattern emerges from the e¨ect of low-ranked constraints like Subject when dominating constraints are controlled for (for example, when the latter are vacuously satis®ed).18 The comparative discussion of tableaux T1.3±T1.5 highlights the fact that the outcome of each competition is indirectly determined by the input. If one adds or removes a feature of the input, the nature of the competition changes because the input determines which of the constraints are applicable. This was illustrated above with the features [new] and [noteworthy] activating alignment constraints in two competitions.

*!

wurde [VP schoÈn getanzt ]] * *

d. [CP schoÈnj wurdei [ IP es ti [VP tj getanzt]]]

e. [CP schoÈnj wurdei [ IP getanzt]]] ti [VP tj

Y

AlNewVP

f c. [ IP schoÈnj wurde [VP tj getanzt]]

b. [ IP

*!

a. [ IP es wurde [VP schoÈn getanzt ]]

AlNotewip

Tableau T1.4 German (Input: tanzenV (x); x ˆ 0; [past]; schoÈnAdv [new], [noteworthy]

*!

*!

MinProj

*

Y

*

Subj

*

*

FullInt

10 GeÂraldine Legendre

Introduction to Optimality Theory in Syntax

11

Tableau T1.5 German (Input: tanzenV (x); x ˆ 0; [past]; schoÈnAdv ) MinProj

Subj

f a. [ IP es wurde [VP schoÈn getanzt ]] b. [ IP

Y

wurde [VP schoÈn getanzt ]]

*!

c. [ IP schoÈnj wurde [VP tj getanzt]]

*!

d. [CP schoÈnj wurdei [ IP es ti [VP tj getanzt]]]

*!

e. [CP schoÈnj wurdei [ IP

*!

ti [VP tj getanzt]]]

FullInt

* *

Further comparison between the competitions in tableaux T1.3±T1.5 highlights the most fundamental property of constraints in OT. Within a single language, the same constraint can be both violated by a grammatical structure in one context and fatal to an ungrammatical structure in another context. For example, Subject is both violated by a grammatical structure (optimal candidate (c) in tableau T1.4) and fatal to ungrammatical ones (candidates (b) in tableau T1.3 and (b, c) in tableau T1.5). AlignNew is violated by the optimal candidate (c) in tableau T1.4 and fatally so by suboptimal candidates (c, d, e) in tableau T1.3.19 We may now return to the general question of constraint violability versus parameters and why the former is a desirable feature of any theory of UG. In P&P terms, having an overt expletive pronoun in one's lexicon is subject to crosslinguistic parametrization. German and French have one, Italian does not. Yet German does not systematically deploy expletive es to satisfy EPP/Subject, contrary to what is expected under a system of parameters ®xed once and for all in a given language. From an OT point of view, the occurrence of es is tied to a particular constraint interaction that is determined by a particular input. Hence, its occurrence is context sensitive. Inviolable principles and parameters seem to necessarily require ad hoc solutions, some of which easily come to mind: (1) posit a little pro in SpecIP in (6c) and ensure the inviolability of SubjectÐbut German does not allow null subjects, (2) stipulate that es may only occur in SpecCP position in German in order to eliminate (6d), and so on. A mixed pattern, however, is what is to be expected under a system of universal constraints that (1) are activated by features of the input, (2) are ranked for priority in a given language, and (3) can be violated by grammatical structures. Simply stated, the nature of the competition and its outcome will change with every input, as exempli®ed in German.

GeÂraldine Legendre

12

1.4

Constraints Are General and Universal

Why do well-formedness constraints often con¯ict, leading to surface violations? It is because they are stated in very general terms. See examples (4a) and (4b). OT constraints eschew logical complexity because complexity is derivative in an OT system: It is the product of the interaction of constraints, not the constraints themselves. In fact, any empirical generalization formulated as a disjunction is the clearest clue to the existence of violable constraints, as pointed out in Speas 1997: 184±185. Unlike in P&P, the content of a OT constraint is not responsible for ensuring its universal application. That aspect of the job derives from the constraint ranking. All constraints are universal in the sense that they are present in every languageparticular ranking, but their relative priority in a given language will be determined by the ranking itself. Thus, the two constraints Subject and FullInt are equally present in the grammars of English, Italian, and German, but their e¨ect is di¨erent because their relative rankings vary with respect to one another as well as with respect to other constraints they interact with. OT constraints belong to constraint families that instantiate two types of possible interactions between constraints. One type consists of families whose member constraints are not universally ranked with respect to one another; they simply belong to the same family by virtue of their content. The other type consists of subhierarchies within a single constraint family whose relative ranking is universally ®xed. Freely rankable constraints include faithfulness constraints (Parse and Fill, Prince and Smolensky 1993) applied to syntax in Legendre et al. 1995 and Legendre, Smolensky, and Wilson 1998 and in the chapters by Bakovic and Keer (chap. 4), Bresnan (chap. 5), Grimshaw (chap. 8), Wilson (chap. 15), and Woolford (chap. 16) in this volume; economy constraints (Stay/*t, Grimshaw 1997, Legendre et al. 1995 and Legendre, Smolensky, and Wilson 1998; NoLexicalMovement, Grimshaw 1997); structural constraints (Subject, ObligatoryHeads, Grimshaw 1997); and so on. They are, by far, the most common constraint type found in this volume and elsewhere. The second type of constraints, subhierarchies, include the MinLink family of constraints against long movements proposed in Legendre et al. 1995 and Legendre, Smolensky, and Wilson 1998, the OpSpec family of Bakovic 1998, and various prominence hierarchies (Burzio 1998; Aissen, chap. 3, and Sells, chap. 12, (this volume). To illustrate with an example, OpSpec, a general constraint requiring wh-operators to be in speci®er position (Grimshaw 1997), is individualized in Bakovic 1998 for each argumental type of wh-phrase: core argument, manner, reason, and so on. These individualized OpSpec constraints are universally ranked with respect to one another, yielding a markedness subhierarchy: ArgOpSpec g LocOpSpec g ManOpSpec g

Introduction to Optimality Theory in Syntax

13

ReasOpSpec. That is, this ranking is ®xed and present in all languages. A con¯icting constraint like Stay (which penalizes wh-fronting and head movement) can be ranked anywhere in the ®xed OpSpec subhierarchy. All wh-operators of the type whose OpSpec constraint is ranked below Stay will be fronted, because of a scope requirement (OpScope), but to an adjoined, rather than to a speci®er position, so as to minimally violate StayÐonce instead of twice if head movement is involved. (Bakovic assumes that fronted wh-phrases without inversion are adjoined to IP.) There are ®ve possible ways of ranking Stay in the markedness subhierarchy, hence his analysis predicts ®ve di¨erent grammars or dialects (abstracting away from additional matrix/subordinate distinctions that Bakovic includes in his analysis). Indeed, Spanish dialects vary as to what type of wh-phrase triggers inversion (i.e., head movement to C) in wh-questions. This type of constraint family interaction, in conjunction with the proposed subhierarchy, derives the familiar type of ``referentiality'' e¨ects in wh-extraction that are prominent crosslinguistically. The question of what de®nes the constraints of a subhierarchy in terms of type or content is at present an open question. Constraints are, however, restricted to markedness constraints. Most seem to pertain to cognitively salient categories, including referentiality distinctions (Bakovic 1998), person/animacy distinctions (Aissen, chap. 3, this volume), and possibly processing-related constraints such as short movement (Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998). Their use in OT syntax, however, has been limited. Much more work is needed before we can extract any generalizations. One important consequence of constraint universality pertaining to the more common freely rankable constraint type is that any new constraint Cn invented on the basis of some phenomenon in, say, German will need to be present in all languages. That is, positing a new constraint in a hierarchy a¨ects the grammar of the particular language but also a¨ects the analyses of all languages. The basic idea is that a ranking C1 g C2 can only be exploited in the analysis of language L1 , all other things being equal, if the reverse ranking C2 g C1 can be demonstrated in some other language. This property of the system imposes strong limitations on the constraints themselves as well as on possible analyses. In practice, this means that any analysis of language L1 remains tentative until a comparative componentÐmore precisely, a factorial typologyÐis added. See section 1.6 for further discussion. 1.5

Constraint Violations

As already mentioned, con¯icts among general constraints are resolved by strict domination rankings (Prince and Smolensky 1993). Violation of higher-ranked constraints cannot be compensated for by satisfaction of lower-ranked constraints. Thus, there are no trade-o¨s in OT.20

14

GeÂraldine Legendre

A constraint cannot hold in one language and simply disappear in another. It can only be subordinated to other, con¯icting, universal constraints. The persistence of low-ranked constraints is supported by the following observations. In a given language (7) a. A constraint may be violated in one context but remain unviolated in another. b. Violation of low-ranked constraints may be fatal. c. Violation of high-ranked constraints is not necessarily fatal. The ®rst pattern of constraint violation is illustrated in German: Subject is violated by the optimal candidate in tableau T1.4 but fatal to two suboptimal candidates in tableau T1.5. Patterns (7b) and (7c) can be illustrated in English, on the basis of Grimshaw's analysis of subject-auxiliary inversion (Grimshaw 1997:377±379). Matrix interrogatives require inversion (i.e., head movement to C) while declaratives forbid it. Consider the pattern in (8). (8) a. Which book will they read? b. *Which book they will read? c. John will read a book. d. *Will John read a book. Grimshaw proposes that movement to C in (8a) is induced by an Obligatory Heads (ObHd ) constraint to provide a head for a projection independently needed to house the wh-operator and satisfy OpSpec (discussed earlier). Sentence (8a) only violates Stay twice (wh-movement and head movement). Given the ranking OpSpec g ObHd g Stay (tableau T1.6), (8b) with an empty C head is eliminated by a fatal violation of ObHd. Grimshaw's account derives the basic generalization that matrix wh-questions are CP structures, while matrix declaratives are IPs. That is, IP candidates systematically compete with CP candidates in both competitions. An IP competitor to the whquestion in (8a) is one with in situ wh-that incurs a fatal violation of OpSpec. In matrix declaratives OpSpec is vacuously satis®ed since no operator is present. An extra CP projection is not needed. Candidate (a) in tableau T1.6 satis®es all constraints. It is optimal. The target of discussion, however, is candidate (b), which is eliminated by the lowest-ranked constraint Stay, thereby illustrating the fact that even low-ranked constraints play a crucial role in competitions (pattern (7b) above). The third and last pattern of constraint violations (pattern 7c) is illustrated in multiple interrogatives (Grimshaw 1997:379±380). As is well known, only one whphrase moves; the other remains in situ. (9) a. What will they put where? b. *What will where they put?

Introduction to Optimality Theory in Syntax

15

Tableau T1.6 English (Input: readV (x, y), [future] x ˆ JohnN , y ˆ a bookDP ) OpSpec

ObHd

Stay

f a. [ IP John will [VP read a book ]] b. [CP willi [ IP John ti [VP read a book ]]]

*!

Tableau T1.7 English (Input: putV (x, y, z), [future] x ˆ theyD , y ˆ what [wh], z ˆ where [wh]) OpSpec f a. [CP whatj willi [ IP they ti [VP put tj where]]] b. [CP whatj willi [ IP wherek [ IP they ti [VP put tj tk ]]]

ObHd

Stay

Y

YY

*

***!

Consider the pattern of violations in tableau T1.7. Both candidates violate the highest-ranked constraint (OpSpec), yet one is optimal. This illustrates the claim that violation of high-ranked constraints is not necessarily fatal. Note in passing that exchanging the two wh-phrases and fronting where instead of what would still violate OpSpec, and so would leaving both wh-phrases in situ (thereby incurring two violations of OpSpec). There is simply no candidate (for this input) that fails to violate the highest-ranked constraint. The decision then falls to a low-ranked one, here again Stay. In fact, both candidates violate Stay, but (9a) does so to a lesser degree. 1.6

Typology by Reranking

OT relies on factorial typology to explain crosslinguistic variation. To propose a constraint ranking for one language in OT is to claim that all possible rerankings of those constraints yield all and only the possible human languages. Two examples are listed in (10). (10) a. Factorial typology of basic case marking/grammatical voice systems (Legendre, Raymond, and Smolensky 1993) b. Factorial typology of presentational focus (Samek-Lodovici, chap. 11, this volume) It is important not to confuse the number of possible languages with the number of possible rerankings. An example taken from Legendre, Raymond, and Smolensky

GeÂraldine Legendre

16

1993 makes this point clearly. Legendre et al. propose a set of eight constraints governing the mapping between thematic roles and their morphosyntactic realizations. This e¨ectively predicts about 40,000 rankings. Yet, as we show, the typology contains thirteen possible language types (only!). Techniques for computing typologies are discussed in Prince and Smolensky 1993: sec. 9.2. In chapter 11 of the present volume, Samek-Lodovici demonstrates how cross-input typologies are determined. Computational tools for determining typologies are also available (Raymond and Hogan 1994; Hayes 1998). 1.7

Economy

Economy plays an important role in the theory. It ranges from minimal constraint violations to structural economy resulting from economy constraints, and more generally, from constraint violability. Minimal violations govern each and every competition in the sense that constraints are violable but are always minimally violated by optimal candidates. When two candidates fare identically except for their number of violations of a single constraint, the candidate with the fewer violations is optimal with respect to that constraint. An example can be found in the multiple interrogatives example in tableau T1.7; the relevant constraint is Stay. When gradient constraints are involvedÐas is the case for alignment constraints like Edgemost, which have been argued to govern the positioning of clitics at the edge of some domain (Anderson, forthcoming; Legendre, forthcoming a, b)Ðit is always the candidate in which the relevant constituent is closer to the domain edge that wins, all other things being equal. Alignment constraints are further discussed in the contributions by Grimshaw (chap. 8) and Legendre (chap. 9) in this volume. Structural economy constraints generally fall under the *Structure family (Prince and Smolensky 1993). These are constraints that penalize the presence of elements in the output and are satis®ed by their absence.21 A well-known one embodies the idea that movement is marked. The main economy-of-movement constraint is called Stay in Grimshaw 1997 and *t (star-trace) in Legendre et al. 1995 and Legendre, Smolensky, and Wilson 1998. They penalize chains or links, favoring ``in situ'' structures. Other economy-of-movement constraints include the MinLink family (Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998), which penalizes longer links compared with shorter ones, for example in the domain of wh-movement. Other structural economy constraints invoked in the OT literature pertain to phrase structure itself, including Minimal Projection (see the analysis of German es above and Grimshaw 1993),22 Economy of Expression (Bresnan 1998), and *Adjunction (Legendre 1999). Structural economy extends beyond syntax: NoMorphology (Ackema and Neeleman, chap. 2, this volume; see also Burzio 1998).

Introduction to Optimality Theory in Syntax

17

More important still is the fact that economy of structure is a direct consequence of violable constraints. In P&P, abstract functional projections are routinely added in order for an element to move to its surface position or reach a particular position by LF, so as to avoid violating inviolable principles of the theory. Such building up of functional structure is often unnecessary in OT precisely because constraints can be violated. Moreover, from a general OT perspective on variation, there is no principled reason why structures should universally involve multiple functional projections ®lled with phonologically empty material such as traces of phrases and/or heads. Functional projections do not come for free (as they essentially do in P&P), since they can result in multiple violations of economy-of-movement and economy-of-structure constraints. Hence, by the economy principles built into OT, they are expected only where they are required by higher-ranked constraints. Consider the position of subject wh-phrases, as in (a) Who saw it?, and the fact that do-support is ungrammatical (under a neutral interpretation): (b) *Who did see it? Grimshaw (1997) demonstrates that her general analysis predicts (a) to have only as much structure as is needed to satisfy OpSpec (any speci®er position will do). The verbal projection is an IP or a VP, depending on assumptions pertaining to the licensing of ®nite in¯ection. In her analysis, vacuous movement (to SpecCP) is suboptimal because it entails a violation of Stay, plus a violation of ObHd or FullInt, if the C head is left empty or ®lled with expletive do, respectively.23 In contrast, nonsubject wh-phrases must move to the speci®er of a functional projection above IP. Otherwise, they fatally violate OpSpec. Thus, while OT predicts minimal syntactic representations, it does not fail to predict that complex representations under appropriate circumstances are favored over simpler ones. 1.8

Optionality

Economy and optimality have the important consequence of essentially excluding optionality, except under particular types of constraint interaction. (See MuÈller 1999 for an overview paper.24) Two candidates may incur exactly the same set of violations and thereby both be optimal. In practice, this is extremely unlikely, since candidates incurring exactly the same violations would tend to be identical. See further discussion in Bakovic and Keer (chap. 4, this volume). A second possibility is a constraint tie: When two constraints C1 and C2 are tied in ranking, a violation of C1 and a violation of C2 cancel each other. If C1 and C2 are high ranked, lower-ranked constraints become decisive. Such an example is found in chapter 9, where I discuss masked second-position e¨ects. A third possibility is that the relative ranking of two constraints may be indeterminate. This is known in the literature as partial ordering of constraints or ¯oating

18

GeÂraldine Legendre

constraints (e.g., Nagy and Reynolds 1997; Antilla 1997; Legendre et al. 1999). The di¨erence between partial orderings and constraint ties is the following. Constraint ties pertain to a single ranking in which violations of two constraints cancel each other. A partial constraint ordering yields a set of rankings. This set of rankings yields potentially di¨erent optimal outputs (hence variation). Both constitute departures from strict domination that might, at ®rst glance, seem fairly innocuous. They are, however, problematic from both a learnability (Tesar and Smolensky 1998; sec. 4.1) and a typological point of view. They substantially complicate learnability, and they can be shown to predict unattested typologies. Therefore, they should perhaps be restricted to contexts of change, with synchronic cases excluded by the theory. By far, the most common type of constraint interaction attested is one in which every competition yields a single optimal output. In fact, there is substantial evidence from OT research (to cite only a few examples, Grimshaw and Samek-Lodovici 1998; Legendre et al. 1995; Bakovic and Keer, chap. 4, this volume) that optionality is most often apparent (except in the unlikely case of identical violations discussed earlier), hiding important di¨erences in information status, register, dialect, and so on that can be formalized in OT in a natural way. Two formal options exist for the treatment of apparent optionality. First, it may result from di¨erent optimizations based on di¨erent inputs. German may serve as an example of the e¨ect of the presence or absence of the feature [noteworthy] in an input to optimization. The pairÐ(a) SchoÈn wurde getanzt `the dancing was beautiful' versus (b) Es wurde schoÈn getanztÐis only super®cially synonymous. Sentence (a) is felicitous only when the adverb schoÈn is speci®ed for noteworthiness in the input. In OT terms, the (high-ranked) constraint AlignNoteworthy favors (c) over (a) when schoÈn is marked [‡noteworthy] in the inputÐit eliminates (a) altogether from the competition (tableau T1.4). When schoÈn is not marked [‡noteworthy], however, AlignNoteworthy is vacuously satis®ed and lower-ranked constraints become active, favoring (a) over (c) (tableaux T1.3 and T1.5). Another source of apparent optionality is di¨erent rankings or grammars representative of di¨erent idiolects, registers, and so on. French exempli®es this possibility. It exhibits a rich typology of wh-questions that re¯ect register di¨erences (Legendre 1998). In colloquial French, contrary to more formal varieties, wh-questions do not undergo fronting. (11) a. Pierre est parti ouÁ? (in situ wh-) Peter has gone where `Where did Peter go?' b. OuÁ Pierre est-il parti? (Complex Inversion) where Peter has he gone All other things being equal, (11a) is optimal under a ranking Stay g OpSpec while (11b) is optimal under the reverse ranking OpSpec g Stay.25 Register di¨erences

Introduction to Optimality Theory in Syntax

19

are, formally speaking, a subcase of crosslinguistic variation. They correspond to di¨erent grammars. In Legendre 1998, I proposed that register variation di¨ers from crosslinguistic variation only with respect to how many constraints may be reranked. Register variation involves minimal constraint reranking, typically involving one constraint only. Crosslinguistic variation is obviously not so constrained. Note that both types of apparent optionality simply exploit existing mechanisms of the theory: inputs on the one hand, and rankings on the other. 1.9

Ine¨ability

In OT, each competition yields an optimal candidate, therefore a grammatical output. This raises a question about ``ine¨able'' structuresÐthat is, instances where some inputs yield no acceptable output. To take a simple example, some languages simply do not allow multiple questions like Who ate what? Ignoring d-linked ones, multiple questions are impossible in Standard Italian, Irish, and other languages (Legendre, Smolensky, and Wilson 1998). The existence of ine¨able structures has, in fact, been characterized as a fatal problem for an OT approach to core aspects of syntaxÐthat is, structure building and movement (Pesetsky 1997:147±150). Pesetsky argues that the existence of ineffable structures suggests instead that the domain of application of OT should be restricted to the pronunciation of the consequences of movement (i.e., the phonological interpretation of the structure). This conclusion, however, is unwarranted because it relies on arbitrarily restrictive assumptions about the nature of the candidate set. In OT, what competes is determined by the input. The main question is: To what extent must the competitors share the same LF? One answer is to incorporate an assumption inherited from P&P and posit that there is a direct mapping between the input and the interpretation of the structure. This entails that all competitors, which by de®nition, share an input, must also share an interpretation. This is the position taken, for example, in Grimshaw 1997 in the context of a study in which all inputs yield e¨able structures. This is also the assumption made in Pesetsky 1997 (though on pp. 150±151, he brie¯y considers the alternative described below). Legendre et al. (1995) and Legendre, Smolensky, and Wilson (1998) confront the language-particular ine¨ability question and argue that competitors need not have the same LF. We reason that the output of a competition has to be semantically interpreted. Thus, it is redundant to include a semantic mapping in the input and then check it after the optimal output has been determined. We o¨er the following account of ine¨ability in Italian wh-questions: Multiple wh-outputs in Italian are suboptimal because they lose to a competitor in which one of the input wh-features is not parsed. In other words, the optimal candidate violates the input-output faithfulness constraint Parse(wh), resulting in a declarative structure with an inde®nite

GeÂraldine Legendre

20

reading of the argument marked [wh] in the input.26 Under this analysis, ine¨ability is reduced to an input-output faithfulness violation. (See also Bakovic and Keer, chap. 4, and Wilson, chap. 15, this volume.) While allowing di¨erent LFs to compete enlarges the candidate set, input-output faithfulness constraints will routinely eliminate candidates with an LF distinct from that intended in the input. 1.10

The Input

I have saved the question of the input for last because it is a question that cannot be addressed independently. It depends, for example, on the answer to the question of what the candidate set is. The input and the candidate set are intimately connected under the architecture of OT. For a given input, the grammar generates and evaluates an in®nite set of output candidates that represent alternative structural realizations of that input. If the candidate set includes candidates with di¨erent LFs, as argued by some, then the input to, say, a wh-question must include target [wh] and operator scope speci®cations. The main role of the input is to determine what competesÐwhat wins is determined by the constraints. OT syntacticians generally agree that the input must specify predicate-argument structure, lexical items, information and illocutionary features, level of argument prominence, as well as familiar functional features (tense, aspect, and so on). For some, operator scope must also be included, as discussed above. Input speci®cations, however, are only target ones; they will be realized only if all faithfulness constraints can be satis®ed. Deviations from target speci®cations do occur, yielding structures unfaithful to the input under compulsion of a high-ranked constraint. Properties of the input and the candidate set are to a large extent determined by the underlying substantive theory of syntax. Thus the input described in the previous paragraph and the corresponding candidate set borrow extensively from the type of representation assumed in P&P. As noted earlier, the question of the substantive nature of syntactic constraints and representations is largely independent of the claims made by OT. This is, in fact, why alternative representations are often employed, as for example, in LFG-based OT work, in Legendre, Raymond, and Smolensky 1993 and several chapters in the present volume.27 The question of the input and the candidate set helps bring out similarities and di¨erences between syntax and phonology. Phonology o¨ers two models of the grammar, only one of which is relevant to current syntactic theorizing. On the one hand, the standard question of the surface form of a given morpheme in phonology leads to an input-output mapping view of the grammar (i.e., a device for mapping a particular underlying form deriving from a lexicon into its correct structural description). On the other, the analysis of basic CV syllable structure in Prince and Smolensky (1993: esp. chap. 9) leads to an inventory view of the grammar, based on the question:

Introduction to Optimality Theory in Syntax

21

What is the inventory of all possible output structures (e.g., syllable shapes), as the input is allowed to range over all possible inputs (e.g., strings of C's and V's)? The latter inventory view (familiar from preexisting constraint-based approaches to syntax) implies that the ``right'' sort of question to ask from an OT perspective is the following: ``What is the inventory of all possible questions in a given language, deduced by considering all possible inputs?''28 It is not: ``What is the input-output mapping?'', given a particular input. In other words, it is important that no output in Italian contain multiple [wh]sÐbeyond this, it is less important what the input-output mapping happens to produce, given a multiple [wh] input. The inventory view of syntax serves to highlight the closeness that exists in OT work in syntax and phonology. Departing from the view held by many syntacticians that syntax and phonology are di¨erent in architecture and formal constraint interaction, existing work in OT syntax has made the strong claim that the two modules operate on the basis of the same formal and markedness principles. For example, economy of structure, alignment with the edge of a domain, input-output faithfulness, and universal markedness subhierarchies have been shown to pervade both syntax and phonology, to name only a few. The remaining ®fteen chapters in this volume o¨er additional evidence that syntax operates on OT principles. I believe that they contribute considerable and solid evidence for a uni®ed grammatical theory. Notes I would like to thank Luigi Burzio, Jane Grimshaw, Paul Hagstrom, Paul Smolensky, and Sten Vikner for valuable comments on earlier drafts of this chapter, as well as Tanja Schmid and Ralf Vogel for sharing their native intuitions and for helping me grasp some of the subtleties of focalization in German. 1. Evidence for movement of adverbs bypassing SpecIP comes from passive examples like (i), in which SpecIP is ®lled with an overt lexical subject: (i) hier wurde sein Auto gestohlen `his car was stolen here'. 2. Arguably, the EPP is satis®ed only by elements that share some properties with canonical subjects, such as nominative Case. This excludes adverbs. See discussion of the constraint Subject (Grimshaw and Samek-Lodovici 1998) below. Construing the EPP as a requirement that the speci®er of I/T be ®lled instead (Chomsky 1995:232) does not a¨ect the point made here. In the standard V-in-C analysis of V2 assumed for the purpose of this illustration, SpecIP is empty in SchoÈn wurde getanzt. Invoking a phonologically null category in SpecIP only serves to maintain the inviolability of the EPP. 3. For a larger picture of the role of OT in the language faculty, see Tesar, Grimshaw, and Prince 1999. 4. Hence constraint universality is logically independent of constraint violability in OT. 5. Grimshaw (1997:374) and Grimshaw and Samek-Lodovici (1998:194) de®ne Full-Int differently: ``Lexical conceptual structure must be parsed.'' Here I restate it slightly because their wording raises unnecessary questions about input and input-output faithfulness that are irrelevant to the point of the text.

22

GeÂraldine Legendre

6. Historically, OT developed out of a theory called Harmonic Grammar (Legendre, Miyata, and Smolensky 1990, 1991), which assumed numerical weighting of constraints rather than strict domination. See Smolensky and Legendre (forthcoming) for further discussion. 7. The competitions in tableaux T1.1 and T1.2 are provided for illustrative purposes only. I am, in particular, abstracting away from the issue of whether it is an expletive or quasiargument, as argued in Vikner 1995:224±228. See note 16 for further discussion. 8. The set includes any other constraints that may turn out to be relevant. 9. Grimshaw and Samek-Lodovici's analysis of null expletives in Italian is o¨ered here as a simple enough illustration of constraint interaction in OT. Obviously, my presentation does not address the well-known generalization that null expletives are found only in null-subject languages. Grimshaw and Samek-Lodovici (1998) propose that null subjects in Italian result from dropping topic-referring subjects. As Samek-Lodovici (1996:46±48) shows, referential null subjects and overt expletives are possible only under rankings that are logically inconsistent with each other. Overt expletives only occur under the ranking Subject g FullInt, while null subjects require the ranking FullInt g Subject. Thus, Samek-Lodovici derives the universal correlation from a basic property of rankings, namely, that a given constraint ranking is ®xed in a given language. 10. For a discussion of the form of expletivesÐwhich pronoun in a language gets drafted for expletive purposesÐsee Grimshaw and Samek-Lodovici 1998: esp. 198, footnote 1. They argue that that too is derived from the syntax, under their formulation of FullInt. It is the pronoun that incurs minimal violations of FullIntÐthat is, the one that has the smaller number of phi features. This eliminates the pronoun he, for example. See Grimshaw (chap. 8, this volume) for a demonstration that language-particular inventories of clitic forms are also derived from constraint rankings. 11. Ralf Vogel (personal communication) points out that preposing of adverbs or PPs in anwers to Was geschah? `What happened?' is possible but is associated with a presuppositional reading (Diesing 1992) rather than with introducing a new entity in the discourse. Hence (6c) is marked aÐthat is, unnatural on a nonpresuppositional reading. 12. I am grateful to Ralf Vogel for clarifying the prosodic properties of these examples, only the most relevant of which are mentioned in the text. Among other things, the full picture involves destressing of the past participle when the adverb schoÈn conveys new information and some idiosyncratic behavior on the part of the adverb hier `here' in similar contexts. 13. Second-position e¨ects are controlled for in all examples, grammatical and ungrammatical alike. 14. The need to recognize di¨erent domains of alignment independently of the German pattern discussed here is amply demonstrated in the OT literature on ``focus'' phenomena. See the references mentioned in the text. The necessity of having constraints refer to features like [new] as opposed to [focus] is demonstrated in Choi 1996 and Legendre 1999. While the present discussion assumes that the domain of AlignNoteworthy is the clause, it is possible that it should be characterized as the Intonational Phrase instead. Further investigation is needed to be more conclusive. 15. My speci®cation of the input only includes the verb and its tense speci®cation, the adverb schoÈn, and the information structure status of the adverb. Passive is completely left out of the picture, since it derives from the mapping from theta roles to grammatical functions. For alter-

Introduction to Optimality Theory in Syntax

23

native OT analyses of grammatical voice, see Legendre, Raymond, and Smolensky 1993 as well as the chapters by Aissen (chap. 3), Sells (chap. 12), and Woolford (chap. 16) in the present volume. 16. It is interesting to note that expletive es in weather verb constructions has a di¨erent distribution: hier regnet es `it's raining HERE'. Under focalization of hier, es must appear, contrary to the impersonal passive pattern in (6d). If we adopt the idea that weather verbs take a quasi-argument (Vikner 1995), a candidate like *hier regnet would fatally fail to parse the quasi-argument speci®ed in its input. This, in turn, is likely to lead to a re®nement of the analysis of weather verb constructions in English and Italian discussed earlier. 17. The present analysis entails an asymmetric approach to V2 phenomena, to some extent similar to that proposed in Travis 1991:355±356 and Zwart 1997:191±196. They propose that ®nite verbs are in I in subject-initial root clauses but in C in non-subject-initial root clauses. In the present analysis, ®nite verbs are not even in C in non-subject-initial root clauses. V2 clauses are CPs only if a CP projection is needed to house an operator-like element. 18. The relative ranking of MinProj in German cannot be fully determined on the basis of the present competitions. The only evidence we have is that MinRroj outranks Subject in tableau T1.4 (hence justifying ranking it just above Subject). It could equally well outrank one or both of the alignment constraints with no e¨ect on the outcome. 19. The discussion of tableaux T1.3±T1.5 also highlights a practical aspect of any OT analysis, namely, that a given competition typically yields only pieces of the total constraint ranking. 20. Local conjunctions of constraints (Smolensky 1993, 1995, 1997) might appear to constitute a counterexample to this generalization. Local conjunctions express the intuition that two constraint violations are worse when they occur in the same location. While most constraints pertain to a single dimension of structure, a local conjunction of two constraints pertains to two dimensions simultaneously. To have any e¨ect, local conjunctions must outrank the constraints they are conjunctions of. See Legendre, forthcoming a, b, for analyses of clitic patterns and Legendre et al. 1995 as well as Legendre, Smolensky, and Wilson 1998 for analyses of whextraction that exploit local conjunctions. See also the contributions by Aissen (chap. 3), Choi (chap. 6), and Grimshaw (chap. 8) in the present volume. While the existence of local conjunctions is amply motivated, their status is still being debated. Smolensky (1997) argues that OT relies on two mechanisms working together: (1) ranking (universal) and (2) local conjunction (language-particular). 21. The overall constraint typology includes constraints on elements (*Structure), constraints on relations (e.g., theta-case mapping), and faithfulness constraints. 22. MinimalProjection is eliminated in the published version, Grimshaw 1997. 23. The idea that subject wh-phrases do not involve wh-movement is not novel to Grimshaw (1997)Ðsee her extended list of references (p. 389)Ðbut the OT explanation for it is. 24. The Minimalist Program (MP) essentially excludes optionality as well, thanks to its incorporating concepts of economy and optimality (Chomsky 1995). In MP, however, optimality is restricted to the determination of the next step in a syntactic derivation and economy is the only optimality metric. 25. Additional considerations pertaining to information status are ignored here, in the interest of exposition. See Legendre 1998, 1999 for details.

24

GeÂraldine Legendre

26. Ackema and Neeleman (1998a, 1998b) propose instead that the optimal output in a case of ine¨ability is the Null Parse of Prince and Smolensky (1993)Ðthat is, an empty structure. Legendre, Smolensky, and Wilson (1998) argue, however, that, based on the fact that an empty structure violates many Input-Output faithfulness constraints, a structure in which only one feature [wh] is unparsed is to be preferred. An empty structure does not minimally violate Input-Output faithfulness. 27. See Speas (chap. 13, this volume) for an alternative conception of the input. See also Wilson (chap. 15, this volume) for an interesting take on inputs and candidate sets in the context of bidirectional optimization. 28. The assumption that the set of possible inputs to grammar is universal and not subject to any language-particular restrictions is known as the Richness of the Base Principle (Prince and Smolensky 1993: chap. 9). References Ackema, Peter, and Ad Neeleman. 1998a. Optimal Questions. Natural Language and Linguistic Theory 16(3): 443±490. Ackema, Peter, and Ad Neeleman. 1998b. WHOT? In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 15±33. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Anderson, Stephen R. Forthcoming. Towards an Optimal Account of Second Position Phenomena. In J. Dekkers, F. van der Leeuw, and J. van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Antilla, Arto. 1997. Variation in Finnish Phonology and Morphology. Doctoral dissertation, Stanford University, Stanford, Calif. BakovicÂ, Eric. 1998. Optimality and Inversion in Spanish. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 35±58. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. den Besten, Hans. 1983. On the Interaction of Root Transformations and Lexical Deletive Rules. In W. Abraham (ed.), On the Formal Syntax of Westgermania, 47±131. Amsterdam: Benjamins. Bresnan, Joan. 1998. Morphology Competes with Syntax: Explaining Typological Variation in Weak Crossover E¨ects. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 59±92. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Burzio, Luigi. 1998. Anaphora and Soft Constraints. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 93±113. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Cardinaletti, Anna. 1990. Impersonal Constructions and Sentential Arguments in German. Padua: Unipress. Choi, Hye-Won. 1996. Optimizing Structure in Context: Scrambling and Information Structure. Doctoral dissertation, Stanford University, Stanford, Calif.

Introduction to Optimality Theory in Syntax

25

Chomsky, Noam. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, Noam. 1982. Some Concepts and Consequences of the Theory of Government and Binding. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1991. Some Notes on the Economy of Derivation. In Robert Freidin (ed.), Principles and Parameters in Comparative Grammar, 417±454. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Costa, JoaÄo. 1998. Word Order Variation: A Constraint-Based Approach. The Hague: Holland Academic Graphics. Diesing, Molly. 1992. Inde®nites. Cambridge, Mass.: MIT Press. Grimshaw, Jane. 1993. Minimal Projection, Heads, and Optimality. Ms., Rutgers University, New Brunswick, N.J. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28(3): 373±422. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 193±219. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Hayes, Bruce. 1998. Optimality Theory Software. Available at http://www.humnet.ucla.edu/ humnet/linguistics/people/hayes/otsoft/otsoft.html. Legendre, GeÂraldine. 1998. PF Cliticization and Complex Inversion in French. Handout of a talk delivered at the Second OT Workshop, University of Stuttgart, November. Legendre, GeÂraldine. 1999. Why French Stylistic Inversion Is Optimal. Ms., Johns Hopkins University, Baltimore. Legendre, GeÂraldine. Forthcoming a. Morphological and Prosodic Alignment of Bulgarian Clitics. In J. Dekkers, F. van der Leeuw, and J. van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Legendre, GeÂraldine. Forthcoming b. Optimal Romanian Clitics: A Crosslinguistic Perspective. In V. Motapanyane (ed.), Comparative Studies in Romanian Syntax. The Hague: Holland Academic Graphics. Legendre, GeÂraldine, Yoshiro Miyata, and Paul Smolensky. 1990. Can Connectionism Contribute to Syntax? Harmonic Grammar, with an Application. In Michael Ziolkowski, Manuela Noske, and Karen Deaton (eds.), Proceedings of the Twenty-Sixth Annual Meeting of the Chicago Linguistic Society, 237±253. Chicago Linguistic Society, University of Chicago. Legendre, GeÂraldine, Yoshiro Miyata, and Paul Smolensky. 1991. Integrating Semantic and Syntactic Accounts of Unaccusativity: A Connectionist Approach. In Laurel A. Sutton, Christopher Johnson, and Ruth Shields (eds.), Proceedings of the Seventeenth Annual Meeting of the Berkeley Linguistics Society, 156±167. Berkeley Linguistics Society, University of California, Berkeley. Legendre, GeÂraldine, William Raymond, and Paul Smolensky. 1993. An Optimality-Theoretic Typology of Case and Grammatical Voice Systems. In Joshua S. Guenther, Barbara A. Kauer, and Cheryl C. Zoll (eds.), Proceedings of the Nineteenth Annual Meeting of the Berkeley Linguistics Society, 464±478. Berkeley Linguistics Society, University of California, Berkeley.

26

GeÂraldine Legendre

Legendre, GeÂraldine, Paul Smolensky, and Colin Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Legendre, GeÂraldine, Anne Vainikka, Paul Hagstrom, and Marina Todorova. 1999. Partial Constraint Ordering in Child French Syntax. Ms., Johns Hopkins University, Baltimore. Legendre, GeÂraldine, Colin Wilson, Paul Smolensky, Kristin Homer, and William Raymond. 1995. Optimality and Wh-Extraction. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 607± 636. GLSA, University of Massachusetts, Amherst. McCarthy, John, and Alan Prince. 1993a. Generalized Alignment. Yearbook of Morphology, 79±153. McCarthy, John, and Alan Prince. 1993b. Prosodic Morphology I: Constraint Interaction and Satisfaction. Technical Report TR-3, Center for Cognitive Science, Rutgers University, New Brunswick, N.J. [Forthcoming, MIT Press, Cambridge, Mass.] Available at http:// www.ruccs.rutgers.edu/publicationsreports.html. McCarthy, John, and Alan Prince. 1994. The Emergence of the Unmarked: Optimality in Prosodic Morphology. In MerceÁ GonzaÁlez (ed.), NELS 24, vol. 2, 333±379. GLSA, University of Massachusetts, Amherst. MuÈller, Gereon. 1999. Optionality in Optimality-Theoretic Syntax. Glot International 4(5): 3± 8. Nagy, Naomi, and Bill Reynolds. 1997. Optimality Theory and Variable Word-Final Deletion in Fñtar. Language Variation and Change 9: 37±55. Pesetsky, David. 1997. Optimality Theory and Syntax: Movement and Pronunciation. In D. Archangeli and D. T. Langendoen (eds.), Optimality Theory: An Overview, 134±170. Oxford: Blackwell. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Raymond, William, and Apollo Hogan. 1994. A Users Guide to the Optimality Interpreter: A Software Tool for Optimality Theoretic Analysis. Technical Report CU-CS-734-94, Department of Computer Science, University of Colorado, Boulder. Samek-Lodovici, Vieri. 1996. Constraints on Subjects: An Optimality Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Samek-Lodovici, Vieri. 1998. Opposite Constraints: Left and Right Focus-Alignment in Kanakuru. Lingua 104(1±2): 111±130. Smolensky, Paul. 1993. Harmony, Markedness, and Phonological Activity. Paper presented at the Rutgers Optimality Workshop 1, New Brunswick, N.J. Smolensky, Paul. 1995. On the Internal Structure of the Constraint Component Con of UG. Paper presented at UCLA, Los Angeles.

Introduction to Optimality Theory in Syntax

27

Smolensky, Paul. 1997. Constraint Interaction in Generative Grammar II: Local Conjunction, or Random Rules in UG. Paper presented at the Hopkins Optimality Theory Workshop/ University of Maryland Mayfest 1997, Baltimore. Smolensky, Paul, and GeÂraldine Legendre. Forthcoming. Architecture of the Mind/Brain: Neural Computation, Optimality, and Universal Grammar in Cognitive Science. Ms., Johns Hopkins University, Baltimore. Speas, Margaret. 1997. Optimality Theory and Syntax: Null Pronouns and Control. In D. Archangeli and D. T. Langendoen (eds.), Optimality Theory: An Overview, 171±199. Oxford: Blackwell. Tesar, Bruce, Jane Grimshaw, and Alan Prince. 1999. Linguistic and Cognitive Explanation in OT. In E. Lepore and Z. Pylyshyn (eds.), What Is Cognitive Science?, 295±326. Oxford: Blackwell. Tesar, Bruce, and Paul Smolensky. 1998. Learnability in Optimality Theory. Linguistic Inquiry 29(2): 229±268. Travis, Lisa. 1991. Parameters of Phrase Structure and Verb Second Phenomena. In Robert Freidin (ed.), Principles and Parameters in Comparative Grammar, 339±364. Cambridge, Mass.: MIT Press. Vikner, Sten. 1995. Verb Movement and Expletive Subjects in the Germanic Languages. Oxford: Oxford University Press. Zwart, Jan-Wouter. 1997. The Morphosyntax of Verb Movement. Dordrecht: Kluwer.

Chapter 2 Competition between Syntax and Morphology

Peter Ackema and Ad Neeleman

Optimality Theory (OT) usually deals with competition within grammatical components. In work on syntax, for instance, various syntactic structures are compared (globally) with respect to a number of syntactic constraints. In this chapter, we will address the question whether there is also competition between components, in particular between syntax and morphology. Does a syntactic realization of some input compete with a morphological realization if both are in principle possible? One complication in addressing this question concerns blocking e¨ects related to the elsewhere condition (see Anderson 1969; Kiparsky 1973). This condition states that insertion of a general form A is blocked where a more speci®c form B can be inserted. As discussed by Di Sciullo and Williams (1987), Andrews (1990), Poser (1992), Sells (1997), and others, there are cases in which the speci®c form is morphological, whereas the general form is syntactic, so that morphology blocks syntax. A well-known example is comparative formation in English. The morphological comparative -er has a more limited distribution, and consequently, where it can attach, it blocks syntactic comparative formation with more (bigger excludes *more big). However, the e¨ects of the elsewhere condition are not limited to morphology blocking syntax. Many familiar examples concern morphology-internal blocking e¨ects, in particular in in¯ectional paradigms: irregular forms like went block regular forms like *goed. In addition, there may be cases in which the speci®c form is syntactic and the general form morphological. The English simple past, for instance, is morphological. Yet, in the perfect, it is blocked by a syntactic periphrastic construction, which is more speci®c because it roughly expresses past with present relevance. As pointed out by Williams (1997), there are also cases in which the elsewhere condition operates within syntax. For one thing, the minimal link condition (see Chomsky 1995) can be seen as an instance of blocking. A lower landing site can attract a subset of the elements that a higher landing site can attract, so in this sense the lower landing site is more speci®c than the higher one. Consequently, movement to a higher landing site is blocked by movement to a lower landing site.

30

Peter Ackema and Ad Neeleman

In conclusion, the elsewhere condition is a general grammatical principle that states that speci®c forms block more general forms, but that does not have anything speci®c to say about the component in which the competing forms are generated. (In OT, e¨ects of the elsewhere condition can be captured by interaction between parse conditions and markedness conditions, as argued by McCarthy and Prince 1994 and Bresnan 2000.1) This leaves open the question whether syntax and morphology as such compete, apart from cases involving the elsewhere condition. At ®rst sight, it seems unlikely that there is such competition, given that related words and phrases freely co-occur. For instance, driver of trucks and truck driver realize the same head-complement relation, but neither blocks the other. Nevertheless, in other examples exactly the relevant type of competition seems to occur. One could hypothesize, for instance, that the verbal compound *to truck-drive is blocked by its syntactic counterpart to drive trucks. (Note that neither form is more speci®c than the other, so that the elsewhere principle does not apply.) The hypothesis that we are dealing with competition here seems to be con®rmed, at least at ®rst glance, by the observation that N-V compounds are possible when there is no syntactic V-NP combination with the same semantics. An example is to Chomsky-adjoin, which exists in the absence of *to adjoin Chomsky (on the intended meaning). If there is competition between components, we must ask ourselves how it compares to competition within a component (the case most often discussed in OT). Of course, the usual questions concerning the nature of the constraints governing competition must be addressed. We will argue that two constraints come into play: a constraint that favors syntactic realization and a constraint that requires satisfaction of selectional properties. In addition, it appears that a di¨erent notion of candidate set is required for competition between components, one that relies on local rather than global comparison. We will see that the two constraints mentioned above are evaluated locally: when an item is taken from the lexicon, the constraints determine in which component it is merged with other material without taking into account the e¨ects this has on the total derivation. In the next section we will outline the proposal in more detail. 2.1

The Proposal

In the generation of structures, three interacting systems play a role: syntax, morphology, and the lexicon.2 The central question of this chapter is how they interact. We will argue for a model that can be characterized by the following four statements: (1) a. Syntax and morphology are independent generative systems. b. The lexicon is a list of syntactic and morphological irregularities.

Competition between Syntax and Morphology

31

c. Syntactic generation of structures is unmarked with respect to morphological generation. d. Lexical items can be underspeci®ed in various ways; one type of underspeci®cation concerns their locus of realization (that is, syntax or morphology). Although the statements in (1a) and (1b) may require some empirical and conceptual motivation, they are familiar from the literature. Di Sciullo and Williams (1987), for example, have presented arguments in favor of both of them. The statements in (1c) and (1d), however, are more controversial, and perhaps it is not immediately clear what they amount to. For that reason, we will give a brief description of the various components and of their interaction. Basic syntactic structures are built up by an operation of merger, which takes two expressions from the lexicon and puts them together in a binary-branching structure (see Chomsky 1995). It is important to realize that merger need not involve simplex expressions. In a sentence like the cat ate the rat, VP is merged with a complex subjectÐthat is, a subject itself created by merger. It must therefore be possible to generate a complex expression by merger, and then store it temporarily while another complex expression is being built. The ®rst expression can then be retrieved again, so that merger of the two becomes possible. Like syntax, morphology is a generative system. Trivially, the generation of complex morphological expressions can be described in terms of merger. Two simplex expressions can be taken from the lexicon and merged into a complex word. This is not to say that syntax and morphology are identical. For example, the position of arguments in syntax is determined by case theory, whereas case features do not seem to play a role in morphology. Instead, word-internal order can be described in terms of the right-hand head rule (see Williams 1981). Such di¨erences between syntax and morphology suggest that the two are separate structure-building components.3 The role of morphology in the derivation can now be understood as follows. Suppose that some complex expression has been generated in syntax and is temporarily stored, so that another complex expression, with which it is to be merged, can be constructed. The latter can of course be built in syntax, but it is also possible to build it in morphology. The output of morphology can then be introduced into the syntax and undergo further operations of merger. If the relation between syntax and morphology is as suggested, the two can be seen as competitors. For each lexical item, it must be decided whether it is merged with its dependents in syntax or in morphology. We assume that there is an important di¨erence between those two alternatives. Continuation in syntax implies that only one rule system will be operative. Continuation in morphology implies that in addition to the syntactic rule system, a second rule system must be activated. From

32

Peter Ackema and Ad Neeleman

a computational point of view, this procedure would seem costly, and it will therefore be avoided if possible. Syntactic generation of structure can hence be seen as unmarkedÐmorphological generation must be forced. This is formulated in the following constraint, which is a variation on a constraint proposed by Grimshaw (1997:382):4 (2) No Morphology Do not build structures in the morphological component. The choice between a syntactic and a morphological continuation of the derivation is made locally. This means that the grammar does not compare complete derivations in which shifts to morphology are made with derivations in which that is not the case. Rather, the choice to activate morphology at a particular point in the derivation is made on the basis of information available at that point. This issue is discussed further in section 2.7. The question that now arises is what can trigger a shift to morphology. We will argue that such shifts are forced by the material picked from the lexicon. One option is that this material consists of an a½x. Since a½xes select a morphological host, taking an a½x from the lexicon activates morphology.5 Thus, the principle in (2) can be overruled by the projection principle in (3). (3) Projection Principle Respect selectional requirements. Let us ®nally turn to the lexicon. Di Sciullo and Williams (1987) argue convincingly that the lexicon lacks generative capacity (see also Jackendo¨ 1997). It is a list of syntactic and morphological irregularities, containing heads (plus their selectional properties) and idiomatic expressions. These may be either words or phrases. One issue concerning the lexicon deserves further attention. Does a particular complex lexical item (that is, an idiomatic expression) uniformly receive a particular (syntactic or morphological) realization, or may its realization di¨er in di¨erent circumstances? Most theories implicitly assume uniform realization, but given the model just sketched, there is no reason to do so. Lexical items can be underspeci®ed in various ways; there is no reason a complex lexical expression consisting of a head and a dependent could not be underspeci®ed as to the component in which it is to be realized (see Groos 1989 and Ackerman and LeSourd 1997 for similar ideas). Suppose that a complex lexical item A5 B is indeed underspeci®ed with respect to its locus of realization. Its parts, A and B, may then be merged either in syntax (giving, for example, [ AP A BP]) or in morphology (giving, for instance, [ A B A]). Of course, the unmarked status of syntactic merger has the e¨ect that A and B will usually combine in syntax. Under speci®c circumstances, however, it is still possible for A and B to be merged in morphology and thus combine into a word. We will

Competition between Syntax and Morphology

33

argue that, once the assumption of uniform realization is abandoned, certain linguistic paradoxes disappear. We will now ®rst present a piece of evidence for the claim that syntax and morphology are separate components, based on synthetic compounds. 2.2

Synthetic Compounds

A strong argument for nonsyntactic word formation, and thus for the separation of syntax and morphology, can be made on the basis of synthetic compounds in English and other Germanic languages. For the purposes of this chapter, the question to be answered is how a compound like that in (4a) is related to the syntactic structure in (4b).6 (4) a. truck driver b. to drive a truck There are two basic possibilities. According to the ®rst, the compound is generated in an independent morphological component and the shared semantics of the two constructions is explained by the system of interpretation. The second possibility is that the compound is derived in syntax and that the shared semantics follows from the presence of a VP like (4b) in the compound's underlying structure. An analysis along these lines was proposed by Baker (1988) for noun incorporation in polysynthetic languages.7 Clearly, these two analyses assign di¨erent structures to truck driver. The morphological analysis assumes a structure as in (5), which can be imported into syntax and then project an NP (see for instance Lieber 1983). (5) [ N [V truck drive] er] The syntactic analysis also assumes a structure as in (5), but this structure is not derived in morphology. Rather, it is derived by the syntactic operation of head-tohead movement. The su½x -er, the noun truck, and the verb drive all start out as the head of an independent maximal projection, as in (6a). They are joined by movement of truck to drive and of the resulting complex verb to -er, as in (6b). (Analyses of this type have been proposed regularly; an early example is Roeper and Siegel 1978.) (6) a. [ NP er [VP drive [ NP truck]]] b. [ NP [ N [V truck drive] er] [VP tV [ NP tN ]]] It will be clear that the morphological and syntactic analyses make rather di¨erent predictions. In particular, the syntactic analysis predicts that material that can be inserted in the maximal projections of nouns such as truck and verbs such as drive, can also be inserted in the structure underlying synthetic compounds. This material

34

Peter Ackema and Ad Neeleman

would then be stranded by the various head movements. The morphological analysis, on the other hand, predicts that the relevant syntactic material cannot be inserted: the structure that should host it is absent if synthetic compounds are not formed in syntax. The data clearly support the morphological analysis. The insertion possibilities in the VP in (7a) do not reappear if a synthetic compound is formed, as in (7b) (see also Neeleman 1994 and Bresnan and Mchombo 1995; see Baker 1988, 1996 for a syntactic analysis, and Rosen 1989 and Ackema 1999a for a morphological analysis, of stranding in incorporating languages). (7) a. to [VP reluctantly drive [ NP a rusty truck] [ PP to Arizona]] b. *[ NP a [ N [V truck drive] er] [VP reluctantly tV [ NP a rusty tN ] [ PP to Arizona]]] Not all the material that must be omitted in synthetic compounds falls into the same class. Some elements, such as the adverbial reluctantly and the PP to Arizona, are optionally inserted in syntax. Therefore, the fact that they must be omitted in compounds only raises the problem why such omission is necessary, rather than optional. Omission of the determiner a leads to the further problem of why it is allowed at all. Singular count nouns in English must always be accompanied by some determiner. Unless other assumptions are made (see Drijkoningen 1994), this implies that the representation in (6b) should be ruled out. The obligatory omission of syntactic material is not the only fact that supports a morphological analysis. Although the su½x -er is combined with a complex verbal category in truck driver, it can also be combined with just a verb, as in driver. Later we will brie¯y discuss how this type of optionality can be analyzed. For now, let us just point out that the two rival analyses make di¨erent predictions with respect to the way the internal argument of driver is realized. According to the morphological analysis, driver is a complex noun, which, like any other noun, projects an NP in syntax. Its arguments should consequently be licensed through of-insertion, just like other nominal arguments. This is of course correct. (8) [ NP [ N drive er] of a truck] The syntactic analysis of driver would start out from the same structure as truck driver (see (6b)), but in this case only the verb moves to -er; the noun remains in situ. Since the noun is in the complement position of a regular VP in this structure, it is incorrectly predicted to appear in the accusative.8 (9) *[ NP [ N [V drive] er] [VP tV [ NP a truck]]] Facts like these can of course be reconciled with a syntactic analysis if further assumptions are made, but all in all they support the claim that there is no underlying verbal projection in nominals derived by -er su½xation. It seems, then, that there is

Competition between Syntax and Morphology

35

empirical motivation for at least one aspect of the model discussed above: morphology and syntax are separate components of grammar (see (1a)). There is one problem for the syntactic analysis of synthetic compounds that we have not mentioned yet. In many cases, the derivation of syntactic compounds seems to contain an ungrammatical intermediate step. The verbal compound *truck drive does not exist, and in general verbal compounding is rather unproductive in English when compared to the formation of synthetic compounds. So, on the one hand noun incorporation must be ruled out in English (see (10)), while on the other hand it is an essential part of the derivation in (6). (10) *to [VP [truck drive]V [ NP tN ]] This problem also holds in the morphological analysis of synthetic compounds, because the structure in (5) contains the verbal compound in (11). (11) *[V truck drive] The ungrammaticality of (11) might seem to indicate that truck driver is in fact an N-N compound, rather than a derivation based on a verbal compound. Dutch, however, provides a strong argument for the existence of verbal compounds inside larger words. As Di Sciullo and Williams (1987) argue, there is no inheritance from the left in compounds. This explains why *een bak-oven van brood `a bake-oven of bread' is ungrammatical: the compound bak-oven does not inherit the internal Y-role of its left-hand part, the verb bak `bake'. This means that the complex compound brood-bak-oven `bread-bake-oven' must be left-branching, since under a rightbranching analysis containing the compound bak-oven `bake-oven', the occurrence of the internal argument of bak `bake' is unexpected. That is, brood-bak-oven `breadbake-oven' must contain the verbal compound brood-bak `bread-bake' (see Ackema 1999b for additional discussion).9 If it is true that N-V compounds like truck-drive can be embedded under an a½x but do not occur in isolation, this can be taken as evidence for another aspect of the model sketched in the introduction, namely, that syntax and morphology are competitors and that generation of structures in syntax is the default (see (1c)); to drive a truck seems to block to truck drive.10 As pointed out in the introduction, however, blocking does not always occur. With respect to the realization of the noun truck, driver of trucks can be seen as a syntactic alternative to truck driver, but the former does not block the latter. Below, we will ®rst explain why syntax does not block morphology in this case. After that a straightforward account for the ungrammaticality of the compound in (11) can be given. Recall that the morphological component is only activated when it must be, and that this decision is taken locally. One way in which activation of morphology can be triggered is by taking an a½x from the lexicon. So, if the su½x -er is at some point

36

Peter Ackema and Ad Neeleman

included in the derivation, morphology is activated to satisfy the selectional properties of this a½x. The derivation can now be continued by importing a verb, say drive, from the lexicon and merging it immediately with the su½x. Since the resulting noun, driver, is well formed, it can be transferred to syntax and morphology can (and must) be deactivated. Any subsequent merger operation will take place in syntax, hence driver of a truck is derived.11 After -er is taken from the lexicon, the derivation can also be continued by importing both a verb and a noun, say drive and truck. If these two elements are merged ®rst, a possibility that must be allowed on independent grounds, morphology will remain active because the su½x -er is not yet attached to a suitable host. The result of the ®rst merger operation can hence be combined with this su½x into the synthetic compound truck driver. As will be clear, more morphological operations take place in this continuation of the derivation than in the one giving rise to driver of a truck. However, since the decision to activate or deactivate morphology is taken locally, at the point when the a½x is taken from the lexicon, this fact does not lead to blocking e¨ects. Such blocking would require global evaluation. Consider now why the verbal compound in (11) is ungrammatical. Since no a½x is taken from the lexicon, there is nothing that forces morphology to become active (see section 2.6 for discussion of in¯ectional a½xes). The verb drive clearly does not have a½xal properties, and neither does the noun truck. The activation of morphology can also not be triggered by the impossibility of combining drive and truck in syntax. These two elements can be combined as the lexical heads in a regular transitive VP (see (4b)). Hence, morphology remains dormant, and the formation of a compound is blocked.12 In section 2.5, we will come back to compound formation. First, however, we will present some further evidence for the claim that syntactic generation of structure is the default (see (1c) and (2)). 2.3

Particle Verbs

Dutch particle verbs give rise to a paradox re¯ected in the term used for them in the traditional literature, namely, separable compounds. The point is that particle verbs on the one hand must be classi®ed as morphological constructs, the main reason being that they appear productively as hosts for derivational a½xes. Some examples are given in (12). (12) a. opmerkelijk up-notice-able `remarkable'

Competition between Syntax and Morphology

37

b. toelaatbaar to-let-able `admissible' c. schoonmaker clean-make-er `cleaner' On the other hand, the verb and the particle, in contrast to verbal compounds, can be separated by the syntactic rule of verb second, suggesting that they are generated in syntax. (13) a. De onderzoeker merkte dit feit niet [op t] the researcher noticed this fact not up `The researcher did not notice this fact.' b. Een agent laat zulk gedrag zelden [toe t] a policeman lets such behavior seldom to `A policeman seldom allows such behavior.' c. De badkamer maakt niemand [schoon t] the bathroom makes nobody clean `Nobody cleans the bathroom.' The problem, of course, is that an expression cannot at the same time be generated in syntax and in morphology, at least not according to the generally accepted model of grammar. We will argue that this paradox arises as a consequence of the assumption that lexical items must be realized uniformly, either in syntax or in morphology. Once this assumption is abandoned, the problem disappears. Before discussing why this is so, we will ®rst consider analyses assuming uniform syntactic or uniform morphological realization of particle verbs. Consider ®rst the di½culties that arise when particle verbs are taken to be uniformly generated in syntax. Such analyses assume that the verb and the particle are inserted as separate heads. The verb heads a VP and the particle is usually taken to head a small clause complement (see, for example, Den Dikken 1995). (14) dat Jan [[SC boeken uit] geeft] that John books out gives `that John publishes books' Although the verb and the particle originate in di¨erent positions, they can be combined into a complex verbal head by a process of head-to-head adjunction. In general, head-to-head movement is the way to derive complex words in theories of this type (see Baker 1988).

38

Peter Ackema and Ad Neeleman

(15) dat Jan [[SC boeken tPrt ] [uit geeft]] that John books out gives Since the complex head is created in syntax, it seems reasonable to assume that its internal structure is accessible to syntactic rules like verb second, and hence that excorporation of the verb is allowed. (16) Jan geeft [[SC boeken tPrt ] [V uit tV ŠŠ John gives out books At the same time, it is expected that further processes of word formation can take place after the particle incorporates into the verb. The verb-particle complex could, for instance, move to the a½x -er, giving rise to the noun uitgever `publisher'. (17) [ NP [ N [V uit geef ] er] [VP [SC . . . tPrt ] tV ]]] out give er This analysis gives rise to various problems. For a start, many particle verbs are unergative, not transitive. Yet, these verbs show exactly the same behavior with respect to verb second and further word formation as other particle verbs. An example is samen-werken `together-work' (cooperate), in which samen `together' behaves like a particle in all respects. (18) a. dat Jan en Piet samenwerken that John and Pete together-work `that John and Pete cooperate' b. Jan en Piet werken samen John and Pete work together c. samenwerking together-work-ing `cooperation' If particles head a small clause complement, this small clause should have a subject, and hence particle verbs should always be transitive or unaccusative. Since no independent source position for particles like samen can be motivated, the most straightforward analysis would be to say that they are base-generated as adjuncts to the verb. But if that is the case, and if they behave just like other particles, there is no reason to assume an underlying small clause in examples like (14). (See Neeleman and Weerman 1993, 1999 for further argumentation to this e¨ect.) More relevant to this chapter, however, are di½culties having to do with the solution proposed for the paradox created by particle verbs. As we argued in the previous section, the theory that word formation is a syntactic process of head-to-head movement faces the problem that evidence for the required underlying structures is absent. Since the analysis of uitgever depends on this theory, it inherits this problem. In par-

Competition between Syntax and Morphology

39

ticular, it incorrectly predicts that material that can (or must) be present in VPs containing a particle, can (or must) also be present in related derivational structures. However, derivations as in (19b) are systematically ruled out, even though structures like (19a) exist. (19) a.

dat Jan [VP al jaren [VP [SC prachtige boeken tPrt ] [V uit geeft]]] that John for years beautiful books out gives b. *Jan is een [ NP [ N [V uit geef ] er] [VP al jaren [VP [SC prachtige John is a out give er for years beautiful boeken tPrt ] tV ]]] books

It seems, then, that if a particle and a verb jointly appear within a larger word, they form a unit generated in morphology and not in syntax. Consequently, the other option to maintain the uniform realization hypothesis is to assume that particle verbs are always realized in morphology, as verbal compounds. The fact that they can be separated by verb second then indicates that the principle of lexical integrity must be abandoned (see Neeleman and Weerman 1993 and Ackema 1999a). However, there are some empirical problems with such an analysis as well. There is strong evidence that Dutch has a syntactic process of complex predicate formation by which resultative expressions are base-adjoined to the verb (see Neeleman and Weerman 1993). Interestingly, this process is blocked when a particle is present: whereas (20a) is grammatical, (20b) is not (irrespective of the order of particle and resultative). (20) a.

dat Jan that John b. *dat Jan that John

en and en and

Piet Pete Piet Pete

zich themselves zich themselves

[kapot to-pieces [kapot to-pieces

werken] work [samen werken]] together-work

Neeleman and Weerman argue that this is due to a syntactic constraint that says that a complex head cannot have a head that is complex as well. This complexity constraint also correctly rules out verbs taking two particles (irrespective of the order of the two particles). (21) a.

dat Jan en Piet [samen werken] that John and Pete together-work b. dat Jan en Piet het voorstel [uit werken] that John and Pete the proposal out-work `that John and Pete develop the proposal' c. *dat Jan en Piet het voorstel [uit [samen werken]] that John and Pete the proposal out-together-work

40

Peter Ackema and Ad Neeleman

Now, in addition to particle verbs, Dutch has complex verbal heads that are uncontroversially derived by morphological processes. These fall into three groups: verbs derived by compounding, verbs derived by pre®xation, and verbs derived by su½xation. As it turns out, none of these verbs violates the complexity constraint when inserted in a larger verbal complex. (22) a. dat Jan [stijl danst] that John style dances `that John is a ballroom dancer' a 0 . dat Jan zich [suf [stijl danst]] that John himself drowsy style dances b. dat Jan de foto's [ver groot] that John the pictures en larges b 0 . dat Jan de foto's [uit [ver groot]] that John the pictures out en larges `that John completely enlarges the picture' c. dat Jan het gedicht [analyse eert] that John the poem analysis izes `that John analyses the poem' c 0 . dat Jan het gedicht [stuk [analyse eert]] that John the poem to-pieces analysis izes One account of these data would be to say that the complexity constraint holds without exception, but that the internal structure of words is not visible in syntax, and that consequently compounds and verbs derived by a½xation may head complex predicates. This, however, would lead to the conclusion that particle verbs are not words, at least not when they occur in isolation (that is, without an a½x), because their structure must be visible to account for the data in (20) and (21). Obviously, this is at odds with the claim that particle verbs are uniformly generated in morphology. If this claim is maintained, then, the complexity constraint must be amended. This is not su½cient, however. Another problem for theories assuming uniform morphological realization has to do with the X-bar theoretical status of the element adjoined to the verb. If particles are attached to the verb in morphology, the prediction is that they cannot project. This prediction is hard to test, since in many cases particle verbs are idiomatic, with the consequence that the particle cannot be modi®ed. If the semantics of the particle is su½ciently transparent, however, modi®cation is possible. This is observed by LuÈdeling (1997) for German; a Dutch example is the verb schoonmaken `to clean'. It can be the basis for further derivation (see (23a)), but at the same time the adjective can be modi®ed (see (23b)). Crucially, this type of modi®cation is impossible if derivation takes place (see (23c)).

Competition between Syntax and Morphology

41

(23) a.

[[schoon maak] er] clean make er `cleaner' b. dat Jan de badkamer waarschijnlijk [[hardstikke schoon] maakt] that John the bathroom probably incredibly clean makes c. *[[[hardstikke schoon] maak] er] incredibly clean make er

This combination of data can only be accounted for if in addition to a morphological variant of schoonmaken, a syntactic variant exists.13 A similar conclusion can be drawn on the basis of data discussed in Hoeksema 1991. In general, particles cannot be moved to SpecCP, presumably because their idiomatic status makes it impossible to focus them. Hoeksema observes, however, that if a particle can be contrasted with another particle, it can be fronted. An example is given in (24). Under the assumption that movement to SpecCP can a¨ect maximal projections only, the moved particle in (24) must have projected. Again, this implies that in addition to the morphological variant of uitvoeren `to export', a syntactic variant exists. (24) Angola voert veel goederen in; uit voert het alleen ko½e tPrt Angola moves many goods in; out moves it only co¨ee `Angola imports many goods; it only exports co¨ee.' Of course, the assumption that particle verbs can be generated in either syntax or morphology implies that the hypothesis of uniform realization must be abandoned (see Groos 1989 and Ackerman and LeSourd 1997 for related discussion). Let us therefore consider how an analysis of particle verbs based on the model sketched in the introduction fares when confronted with the above data. According to this model, structure is only generated in morphology if that is required by the material taken from the lexicon. Suppose that particle verbs are underspeci®ed with respect to their locus of realization. This means that, unless morphological realization is forced, the particle and the verb will be joined in syntax, forming a complex predicate comparable to the ones formed by merging a verb and a resultative expression. Now, as explained in section 2.1, the choice to activate morphology is made locally. This implies that importing a derivational a½x from the lexicon will trigger a shift to morphology. Such a shift cannot be motivated by global considerations. The consideration that it avoids a fatal violation of syntactic constraints (that is, a crashing derivation) will in other words not be su½cient to activate morphology. These assumptions straightforwardly account for the data introduced above, as we will now explain. Suppose that after the nominalizing su½x -er has been imported, the derivation is continued by importing a complex lexical item consisting of a particle (say uit `out')

42

Peter Ackema and Ad Neeleman

and a verb (say geven `give'). These two elements can then be merged in morphology, given that this component must remain active as long as the a½xal properties of -er are not satis®ed. The result is a verbal compound uitgeef `publish', which in turn can be combined with -er, yielding the noun uitgever `publisher'. The same procedure explains the existence of the words in (12). Note that since the particle and the verb are merged in morphology in the cases at hand, it is predicted that neither the particle nor the verb can project to a full XP. This implies that modi®ers and other material that can otherwise accompany the particle or the verb cannot occur. The ill-formedness of examples like (19b) and (23c) thus follows. In the absence of a derivational a½x, nothing triggers activation of morphology, and consequently, if the particle uit and the verb geven are taken from the lexicon, they must be merged in syntax. Crucially, the morphological analysis is not available. The result is that the particle and the verb form a complex predicate of the type required independently for resultative constructions. Of course, structure generated in syntax is accessible to syntactic operations, and therefore the particle and the verb are separated by verb second, as in (13). A further consequence is that the particle may project. This explains why, in the right contexts, it can be topicalized (see (24)) or be accompanied by modi®ers and other syntactic material (see (23b)). Finally, the complexity constraint can be seen as a regular syntactic condition that applies to syntactic structures only. An analysis of Dutch particle verbs in terms of nonuniform realization receives further support from the behavior of particle verbs in Swedish. In this VO language, syntactic dependents of the verb typically follow it. Hence, we expect the particle to follow the verb if the particle verb appears in isolation (as argued above, particle verbs in isolation must be realized syntactically). This is corroborated by examples like (25) (from GunloÈg Josefsson, personal communication, and Holmes and Hinchli¨e 1994; see section 2.5 for further discussion). (25) stiga upp transportera bort trubba av hyra ut gaÊ vilse

`rise' `transport away' `blunt' `hire out' `get lost'

However, if the particle verb is selected as host by an a½x and hence occurs within a larger word, the order between particle and verb is reversed. (26) upstigning borttransportering avtrubbning uthyrare vilsegaÊngen

`ascent' (of an airplane) `sending away' `blunting' `letter' `lost'

Competition between Syntax and Morphology

43

This is to be expected if particle verbs are realized morphologically when selected by an a½x, since Swedish morphology is in general right-headed (see Josefsson 1998 and references cited there). The illustrated nonuniform realization of Swedish particle verbs therefore corroborates our analysis. The paradoxical behavior of particle verbs thus supports the following assumptions. First, it shows that syntax and morphology are separate modules; otherwise the absence of modi®ers in the presence of a derivational a½x could not be explained. Second, it shows that syntactic and morphological merger are alternative ways of generating structure. If there were only one way of generating particle verbs, they should show uniform behavior. The data indicate, however, that particle verbs show syntactic behavior (separation by movement, projection, sensitivity to the complexity constraint and to syntactic ordering conditions) in some contexts and morphological behavior (no projection, no sensitivity to the complexity constraint, sensitivity to morphological ordering conditions) in others. Third, it shows that syntactic realization is preferred over morphological realization and that the choice between the two is made locally. If syntactic realization did not block morphological realization in the absence of a derivational a½x, we could still not account for the fact that particle verbs are subject to the complexity constraint. Their morphological alternate should freely combine with resultatives, contrary to fact (see (20b)). This morphological alternate could also undergo verb second as a whole, which is again impossible. In the next section we will strengthen the argument, using clearly syntactic idioms that, in the presence of a derivational a½x, must nevertheless be realized morphologically. 2.4

Verbal Idioms

If the analysis of particle verbs presented in the previous section is correct, complex lexical items containing a verb and a particle must be underspeci®ed with respect to their locus of realization. This does not mean that underspeci®cation is a necessity, because there are clearly syntactic and clearly morphological complex lexical items. However, the possibility of underspeci®cation is crucial since underspeci®ed items allow us to study the competition between syntax and morphology. We will start this section by considering underspeci®cation (or the lack thereof ) in more detail. Then we will illustrate its consequences for verbal idioms. Research into the form of lexical entries is guided by the assumption that lexical storage should be kept to a minimum. As a research strategy, it is assumed that knowledge that can be represented by rules should not be duplicated by lexical stipulations. Attempts have been made, for example, to derive categorial selection from semantic selection where possible (see Grimshaw 1981; Pesetsky 1982).

44

Peter Ackema and Ad Neeleman

It follows from this research strategy that only idioms and simplex words will be listed.14 This implies that the word reattach, which has completely transparent semantics, will not be in the lexicon as such (although re- and attach must of course be listed). On the same grounds there is no need to store the phrase attach (something) again. A more interesting consequence of the strategy of lexicon minimization concerns the information stored in a lexical entry. It is well known that the structural properties of at least some complex lexical items (that is, idioms) can be predicted on the basis of the pertinent grammar. The position of the verb in pull someone's leg, for example, is a consequence of the VO character of English. Similarly, the linear realization of beefeater follows the right-hand head rule of English morphology. This suggests that, as long as an idiom adheres to syntactic or morphological wellformedness conditions, its internal structure will not be speci®ed in the lexicon. The minimum that each complex lexical item must contain is a speci®cation of the grammatical relation between its parts. This hypothesis goes back to work on verbal idioms by Bresnan (1982) and Coopmans and Everaert (1988). In the examples at hand, the crucial relation is one of internal theta-role assignment. In beefeater the internal theta role of eat is bound by beef. Similarly, the DP headed by leg binds the internal theta role of pull in pull someone's leg. So, abstracting away from various details, the representation of the relevant lexical items is as in (27), where the brackets indicate junctures unspeci®ed for type (syntactic or morphological) or ordering (head®rst or headlast). (27) a. hheat beefi i eri (r, theme i ); semantics: member-of-the-Royal-Guard (r) b. hpull legi i (agent, theme i ); semantics: tease (agent) As will be clear, the complex lexical items in (27) must contain an instruction concerning their interpretation. Similarly, it must be speci®ed whether they are to be realized in syntax or morphology, an issue to which we will turn shortly. Once this is arranged, however, their internal structures need not be speci®ed any further. These simply follow from the principles that govern the realization of arguments in morphology and syntax. If complex lexical items can indeed be listed without structural speci®cations, it must be indicated whether they are to be realized in syntax or morphology. The simplest way of doing this is by adding diacritics, say S for syntactic and M for morphological realization. Simple as this may be, it is not the most economical solution. If one of the two ways of realizing a complex lexical item is unmarked with respect to the other, only the marked option needs to be accompanied by a diacritic. Idioms that lack the relevant diacritic could in principle be realized either syntactically or

Competition between Syntax and Morphology

45

morphologically. Confronted with this choice, however, the system will opt for the default realization. Thus, the amount of information stored in the lexicon can be further reduced. This strategy ties in naturally with the model sketched in section 2.1. There, it was claimed that of the two modes of structure generation, syntax is the default. If this is so, only morphological idioms need to be marked as such. This can be implemented by specifying the brackets in lexical entries like (27) with an M in case of morphological idioms (see (28)). Junctures encoded by unspeci®ed brackets are underspeci®ed with respect to their locus of realization, but due to the unmarked status of syntax these will give rise to a syntactic idiom. (28) hhM eat beefi i eri (r, theme i ); semantics: member-of-the-Royal-Guard (r) To be sure, we are not claiming that syntactic idioms cannot be marked as such (through the diacritic S). The crucial point, however, is that underspeci®cation with respect to locus of realization is to be expected in at least certain idioms. A survey of words derived by su½xation in Dutch veri®es the existence of such underspeci®cation. Several syntactic idioms in Dutch may be realized morphologically in the presence of a derivational su½x. A ®rst example is iemands hart breken `to break someone's heart', which clearly is a syntactic idiom, but which nevertheless can be part of the synthetic compound hartenbreker `heart breaker'.15 (29) a. dat hij Marie's hart vaak heeft gebroken that he Mary's heart often has broken b. Hij is een echte hartenbreker he is a real heartbreaker Like particle verbs, these data give rise to a paradox in traditional theories of morphology. On the one hand, the idiom in (29) is clearly syntactic in nature. A speci®er can accompany the noun and the verb can undergo movement to C in main clauses. On the other hand, the idiom can appear within a word derived by attachment of a su½x. However, whereas particle verbs could be analyzed in various ways (at least at ®rst sight), the only plausible analysis of the data in (29) seems to be one in terms of nonuniform realization and lexical underspeci®cation. As far as we know, the existence of pairs like in (29) has not been discussed before. Before presenting the argument, we will therefore give a more complete overview of the data. The examples given below are extracted from Nieuwborg's (1969) Retrograde woordenboek. What we have searched for are synthetic compounds that end in the deverbal a½xes -er, -end, and -ing and that seem to be derived from syntactic idioms. We have used Van Dale's (1984) dictionary to check whether the relevant syntactic idioms exist. For each of the a½xes, some examples are given in (30)±(32).

46

Peter Ackema and Ad Neeleman

(30) a. dat Jan iedereen stroop om de mond smeert that John everyone syrup around the mouth smears `John ¯atters everyone.' a 0 . Zo'n stroopsmeerder wekt altijd afkeer such a syrup-smearer causes always disgust `Such a ¯atterer is disgusting.' b. dat die mededeling grote onrust heeft gezaaid that that announcement big unrest has sown `that that announcement has caused much anxiety' b 0 . Hij heeft met Piet een onrustzaaier in dienst genomen he has with Pete a unrest-sower in employment taken `In Pete, he has employed a troublemaker.' c. dat Karel over deze brigade het bevel voert that Karel over this brigade the order leads `that Karel commands this brigade' c 0 . Karel is de bevelvoerder over deze brigade Karel is the order-leader over this brigade `Karel is the commander of this brigade.' d. dat hij voor Picasso de weg bereidde that he for Picasso the road prepared `that he paved the way for Picasso' d 0 . Hij was wegbereider voor Picasso he was road-preparer for Picasso `He was someone who paved the way for Picasso.' (31) a. dat deze voorstelling iedereen de adem heeft benomen that this performance everyone the breath has taken `that this performance took everyone's breath away' a 0 . De voorstelling was adembenemend the performance was breathtaking b. dat Jan zich altijd over van alles het hoofd breekt that John himself really about of everything the head breaks `that John always worries about everything' b 0 . dat was werkelijk een hoofdbrekend karwei that was really a head-breaking job `That really was a tough job.' c. dat de uitslag van die wedstrijd veel opzien heeft gebaard that the result of that match lots-of up-looking has given-birth-to `that the result of that match was very surprising'

Competition between Syntax and Morphology

47

c 0 . De uitslag was opzienbarend the outcome was up-looking-give-birth-to-ing `The result was surprising.' d. dat haar kwaliteiten uiteindelijk de doorslag hebben gegeven that her qualities eventually the through-hit have given `that her qualities eventually tipped the balance' 0 d . Haar kwaliteiten waren doorslaggevend her qualities were through-hit-giving `Her qualities tipped the balance.' (32) a. dat die lui altijd ergens de hand mee lichten that such people always something the hand with lift `that such people always ignore some regulation' a 0 . Zo'n handlichting moet consequenties hebben such a hand-lifting must consequences have `Ignoring regulations in this way must have consequences.' b. Laten wij elkander de hand reiken let we each-other the hand reach `Reach out and touch.' b 0 . Zijn handreiking had weinig resultaat his hand-reaching had few results `His concession had few results.' c. dat de bestuursleden enige harde woorden wisselden that the members-of-the-board some harsh words exchanged c 0 . Ik vond die woordenwisseling nogal geÃnant I found that words-exchanging rather embarrassing `I found that argument rather embarrassing.' d. dat dit gerecht de tong streelt that this dish the tongue strokes `that this dish is delicious' d 0 . Dat was een ware tongstreling that was a real tongue-stroking `That was really delicious.' Although the paradox these data give rise to is comparable to the one caused by particle verbs, at least one of the alternative analyses proposed for particle verbs cannot be extended to this case. In view of the material indicating a full-blown syntactic structure in every ®rst example, it seems impossible to maintain that the idioms in (29) to (32) are uniformly realized in morphology. If one wants to hold on to uniform realization of lexical items, the only option seems to be that, if the idiom shows

48

Peter Ackema and Ad Neeleman

up word internally, this is the result of head-to-head movement. Thus, the word hartenbreker could be derived as follows: (33) [ NP [ A [V hartenN breekV ] erN ] [VP tV [ NP tN ]]] heart break er In sections 2.2 and 2.3 we have already considered to what extend head-to-head movement explains properties of particle verbs and (nonidiomatic) synthetic compounds. The problems this analysis faces are therefore familiar. For a start, the incorporation analysis fails to explain why idioms are stripped of all material than can appear in syntactic projections when they appear as part of a word. If a full-¯edged syntactic structure is present, there is no reason such material cannot be stranded by the two head movements in (33). Nevertheless the example in (34), based on (29a), is ungrammatical. (34) *hij is een [ NP [ N [V harten breek] er] [VP vaak tV [ NP Marie's en Sue's tN ]]] he is a heart breaker often Marie's and Sue's As in the case of particle verbs and nonidiomatic synthetic compounds, material obligatorily present in syntax (like determiners) must sometimes be omitted. Even if problems of this type could be solved, it would have to be stipulated that of the two operations needed to derive hartenbreker in (33), the incorporation of the noun into the verb is impossible unless the resulting complex verb moves on to a derivational su½x. At least, the verbal compounds *hartenbreken `heart-break', *adembenemen `breath-take', and *handlichten `hand-lift' do not exist. As it turns out, then, neither uniform morphological nor uniform syntactic realization of the idioms under discussion leads to a plausible analysis of the alternation found in (29) to (32). This alternation falls out naturally, however, from the model proposed here. Suppose that the idiom iemand's hart breken is stored in the lexicon without being speci®ed for syntactic or morphological realization. (35) hbreek harti i (agent, theme i ); semantics: hurt-in-love (a) Unless morphology is activated by the inclusion of an a½x in the derivation, this idiom must be realized in syntax, given that this is the default option. Consequently, the verb will project a VP that contains a DP headed by the noun hart and, like other VPs and DPs, these projections allow insertion of further material, as illustrated by (29a). Crucially, in the absence of a derivational a½x a morphological realization of the idiom is excluded, which accounts for the absence of the compound *hartenbreken `to heart-break'. If a derivational a½x, say -er, is taken from the lexicon, morphology must be activated. This implies that the idiom in (33) is now realized as a verbal compound,

Competition between Syntax and Morphology

49

thus providing a category with which the su½x can be merged. In this way, generation of synthetic compounds like hartenbreker is possible. Since the idiom is now structurally realized in morphology, it is predicted that the material that usually accompanies it in syntax must be omitted. This accounts for the ungrammaticality of the example in (34) (where this material is stranded by alleged head movement) and of the ones in (36) (where it appears inside the compound). (36) a. *hij is een [ NP [ N [VP [ NP Marie's hart] vaak breek] er] he is a Mary's heart often break er b. *de voorstelling was [ AP [ A [VP iedereen [ NP de adem] beneem] end] the performance was everyone the breath take ing c. *zo'n [ NP [ N [VP altijd ergens [ NP de hand] mee licht] ing] moet such always something the hand with lift ing must consequenties hebben consequences have So, once the assumption that complex lexical items are realized uniformly is abandoned, a simple account of the alternation in (29) to (32) can be given. This account runs parallel to the one given for particle verbs. It thus also supports the following aspects of the model defended here: (1) syntax and morphology are alternative modes of generating structure, (2) syntactic generation is unmarked with respect to generation in morphology, and (3) complex lexical items can be underspeci®ed for their locus of realization. There is a di¨erence between particle verbs and idioms with respect to their occurrence within words. Particle verbs form a much more productive basis for further word formation. This need not be stipulated, however, since particle verbs and idioms have di¨erent properties anyway. Particle verbs are complex predicates, which means that they already have a structure that closely resembles that of verbal compounds when they are realized in syntax. Morphological realization would consequently seem unproblematic. Syntactic idioms, however, often consist of a verb that combines with more than one ®xed expression (consider van een mug een olifant maken `to make an elephant out of a mosquito'Ðthat is, `to make a big a fuss about nothing'). Usually it is not possible to create a well-formed morphological expression that can host all this idiomatic material, which means that a morphological realization of the idiom is ruled out. It is correctly predicted, then, that only simple idioms, consisting of just a verb and one dependent, can be realized within a synthetic compound.16 2.5

Compounds

So far, we have only considered the selectional properties of a½xes as a reason for morphological realization. However, there is a second possibility: a complex lexical

50

Peter Ackema and Ad Neeleman

item can simply be listed as a morphological entity, a possibility mentioned in section 2.4 for idioms that can only be realized morphologically. Listing, however, only occurs when the listed item has a (partially) unpredictable interpretation. This implies that simple compounds, at least in languages of the type under discussion here, can exist only if they have an idiomatic meaning. Conversely, a morphological process that has transparent semantics must involve a½xation. (See section 2.6 for language variation on this point.) Starting with the latter point, it indeed seems to be the case that in the languages under discussion morphological operations leading to predictable changes in meaning are always instantiated by a½xation. For example, diminutive formation in Dutch is expressed by the su½x -tje, repetition of action is expressed by the pre®x her-, and causativization is expressed by the su½x -iseer.17 Crucially, none of these processes can be expressed by compounding in Dutch (or in the other languages under discussion). The same point can be made on the basis of in¯ectional morphology, which is semantically fully transparent and which never takes the form of compounding. It also seems that the instances of simple compounding attested in the languages under discussion cannot be described on the basis of a predictable resultant meaning. The general right-headedness of morphological constructs ensures that the ``is a'' relation holds between the compound and its right-hand member, but the relation between the right-hand and the left-hand part of a compound is unpredictable (see also Carstairs-McCarthy 1992). Consider the following examples of noun-verb compounding in English: (37) to to to to to

breast-feed play-act window-shop base-generate head-adjoin

to to to to to

hand-make air-condition c-command chomsky-adjoin head-govern

It is clear that no uniform characterization of the semantic relation between the members of these compounds can be given (see Preuss 1960). To chomsky-adjoin, for example, means ``to adjoin in a certain way proposed by Chomsky,'' but to headadjoin does not have a meaning parallel to this; it means `to adjoin to a head'. In turn, to head-govern does not have an interpretation of this type, but means `to govern in the capacity of a head'. There is one restriction on the semantics of N-V compounding, however: the noun cannot be interpreted as the internal argument of the verb (see Shimamura 1983). This can be understood if listing is costly. In that case, a form will not be listed if its meaning, using the same lexical material, can also be derived by regular processes independently available in the language. Since the internal argument relation has a fully compositional meaning, semantics cannot be used to motivate listing. There is,

Competition between Syntax and Morphology

51

in other words, no option of listing a head and its internal argument as a morphological item. The other means of activating morphology, namely the presence of an a½x, does not work in cases of simple compounding either. As a consequence, simple N-V compounds in which N is an argument of V are blocked. This implies that whenever the noun in an N-V combination is interpreted as the internal argument of the verb, the two must have been combined in syntax. Indeed, Dutch N-V combinations in which the noun is an argument, such as ko½e-drinken `to co¨ee-drink', may behave as complex verbal heads, but they are, without exception, separable through head movement of the verb (cf. Jan dronk gisteren ko½e `John drank yesterday co¨ee' versus *Jan ko½e-dronk gisteren `John co¨ee-drank yesterday'). N-V combinations of the type in (37), on the other hand, behave as indivisible units (cf. *Jan adjungeert zijn topics altijd Chomsky `John adjoins his topics always Chomsky' versus Jan Chomsky-adjungeert zijn topics altijd `John Chomsky-adjoins his topics always'). (For more data see Ackema 1999b, where this generalization is derived under di¨erent assumptions; incorporating languages will be discussed in section 2.6.) The function of listing as a trigger for morphological activation can be further demonstrated when we return to verb-particle constructions in Swedish. In section 2.3 we derived the generalization that the particle follows the verb in the absence of an a½x (verb and particle are merged in syntax), whereas the particle precedes the verb when an a½x is present (verb and particle are merged in morphology). In fact, however, a quali®cation is in order: there are also cases in which the particle precedes the verb when no further a½xation takes place. This shows that simple compounding of particle and verb is possible after all. Given what we have said so far, simple compounding of this type should only occur if the resulting word has a (partially) unpredictable meaning, which triggers listing. This prediction appears to be correct. With regard to Swedish verb-particle combinations, Holmes and Hinchli¨e (1994:321) remark that ``the separated form has a literal meaning. . . , whereas the corresponding integrated form has a ®gurative meaning.'' Examples are given in (38). (38) a. Jag bryter av kvisten I break o¨ the-branch a 0 . Jag avbryter samtalet I o¨-break the-conversation `I interrupt the conversation.' b. Han stroÈk under ordet he lined under the-word `He underlined the word.' b 0 . Han understroÈk ordets betydelse he underlined the-word's meaning `He emphasized the word's meaning.'

52

Peter Ackema and Ad Neeleman

c. Bordsbenet gick av the-table-leg broke o¨ 0 c . Statsministern avgick the-prime-minister o¨-broke `The prime minister resigned.' Note, ®nally, that there can be other reasons for listing a complex form than unpredictable semantics. More speci®cally, it appears that some compounded particle verbs in Swedish have the same meaning as their syntactic counterparts. However, as observed by Holmes and Hinchli¨e (1994:320), these ``are reserved almost solely for o½cial documents and more formal usage'' and are not used in the spoken language. Under the plausible assumption that specialized, formal expressions must be listed, the fact that these forms are compounds further corroborates the idea that listedness is a prerequisite for simple compound formation. 2.6

In¯ection

If synthetic compounding is possible because of the presence of an a½x, one may expect that in¯ection can license N-V compounding in English as well, contrary to fact. (39) *John truck-drives all day long Apparently, in¯ectional and derivational morphology di¨er in this respect. A closer look at the data reveals that in fact it is not this distinction that is crucial, but rather that between category-neutral and category-changing morphology.18 Thus, it turns out that in English synthetic compounding is never triggered by an a½x that does not change the category of the verbal base. (40) a. *Mary re-truck-drives to Mississippi tomorrow b. *This system over-tree-generates Indeed, in the languages under discussion so far the a½xes involved in synthetic compounding are always category changing. What this indicates is that category-changing a½xes trigger morphology in a way that does not, or not necessarily, carry over to non-category-changing ones. We suggest that the following rationale underlies this di¨erence. If one wants to attach a category-changing a½x to a base, one always has to leave the syntactic representation and perform an operation in morphology. However, when a non-category-changing a½x is to be attached, it is possible to do so without leaving the syntactic representation: one can simply apply the morphological rule system to a part of the syntactic tree (see Borer 1993).

Competition between Syntax and Morphology XP X

*

53

XP Y

ƒƒ! X

Y

Figure 2.1

Consider what would happen if one would try to attach a category-changing a½x without leaving the syntactic tree. The node created by attachment of such an a½x would have a di¨erent category than the projections that dominate it, as ®gure 2.1 shows. The result is a structure that violates the principle of endocentricity. However, endocentricity is presumably part of GEN (Grimshaw 1997) and can therefore not be violated. As a consequence, attachment of category-changing a½xes must take place outside the syntactic representation. In the case of non-category-changing a½xes, there is, at least in principle, a choice between independent application of morphology and application to the syntactic tree.19 The latter point, we propose, is subject to crosslinguistic variation. In languages like English, category-neutral a½xes are attached to their host via application of morphology to the syntactic tree. In some other languages, they are attached via independent application of morphology. This variation can be captured by reformulating the No Morphology constraint as in (41a) and adding the constraint in (41b).20 (41) a. No Morphology Operate on the syntactic tree. b. Morphological Exclusivity Do not apply morphological operations to the syntactic tree. If (41a) outranks (41b), a language like English results. Synthetic compounding cannot be licensed by in¯ection, since in¯ection must be applied to the syntactic representation, and the syntax does not generate N-V compounds. For reasons outlined earlier, category-changing a½xes are always attached to their host outside the syntactic representation, so they do license synthetic compounds. In languages in which (41b) outranks (41a), non-category-changing a½xes are predicted to license compounding of verbs and their internal arguments just like other a½xes. This gives rise to incorporating languages such as Mohawk (Baker 1996), Hindi (Mohanan 1995), and many others (Mithun 1984).21 Note that, indeed, whenever an in¯ectional a½x and the incorporated noun appear on the same side of the verb, the in¯ectional a½x is attached external to the noun. This follows if the in¯ectional a½x is the trigger for compounding. An example from Mohawk is given in (42) (from Baker 1988).

54

Peter Ackema and Ad Neeleman

(42) Yao-wir-a?a ye-nuhs-nuhwe?-s pre-baby-suf 3fS/3n-house-like-asp `The baby likes the house.' Thus the idea that syntax and morphology are competitors cannot only be used to capture data internal to Dutch and English, but also to capture crosslinguistic variation. 2.7

Concluding Remarks

Let us ®nally compare the kind of competition discussed above, that between two di¨erent rule systems in the grammar, with the more familiar type of competition, that between di¨erent structures generated by one rule system. The main di¨erence between the two types of competition is that evaluation in the former type is local in character, while the latter involves global evaluation. In a modular grammar, rules contained within a component may not (directly) interact with rules external to that component. This means that if there are constraints that govern the competition between morphology and syntax, these constraints cannot interact with either syntaxinternal or morphology-internal constraints. To put it di¨erently, a prerequisite for (global) evaluation within a component is that all else is equal: candidates in syntactic evaluation di¨er only in the application of syntactic operations, candidates in morphological evaluation di¨er only in the application of morphological operations, and so on. But competition between components concerns cases where all else is not equal: candidates di¨er precisely in whether either syntactic or morphological operations apply. Hence, the question whether the syntactic or the morphological rule system is to be activated must be decided locally. From this a further di¨erence follows, namely, one concerning the de®nition of candidate set. Within syntax and morphology, candidates are de®ned in terms of shared lexical material and shared semantics (see Grimshaw 1997). Such a de®nition is impossible in the case of competition between components, because at the moment of competition the overall semantics of the structure is unknown (the morphological output may give rise to a di¨erent semantics than the syntactic one). A similar remark can be made with respect to the notion of numeration. However, as we have seen, competition is still limited to cases in which candidates share two lexical items (the ones to be merged either in syntax or in morphology) and a particular semantic relation between these items. For instance, competition is between alternative realizations of a verbal head and its internal argument. Of course, this is the most that can be shared between candidates in local evaluation. What does not di¨er is the way language variation is handled: in OT this is done by constraint reranking. As shown in section 2.6, constraints that trigger activation of

Competition between Syntax and Morphology

55

a rule system and constraints that block this are indeed ranked di¨erently in di¨erent languages. Notes We would like to thank Joan Bresnan, Annabel Cormack, Jane Grimshaw, GunloÈg Josefsson, Henk Schultink, Kriszta Szendroi, Yoko Yumoto, and the audiences at the University of Durham and the Congress on Storage and Computation in Linguistics at Utrecht University. 1. Bresnan shows that this approach can in fact handle more cases of blocking than the elsewhere condition, since it also applies when the features of one of the competing forms are not a proper subset of the features of the other. 2. And of course semantics and phonology, but these are irrelevant to the present discussion. 3. This leaves open the question of to what extent morphological and syntactic rules have the same format. See Sproat 1985, Di Sciullo and Williams 1987, and Ackema 1999a, among others, for various proposals. 4. Sells (1997) argues that morphological realization is preferred over syntactic realization. The cases he discusses can all be subsumed under the elsewhere condition, and consequently they do not bear on the issue at hand (see the introductory section of the present chapter). 5. We will only discuss overt a½xes in this chapter. In principle, insertion of silent a½xes should yield identical e¨ects, but this might not be so if some of the selectional relations under discussion are partly phonological in nature. As discussed by Sproat (1985), Beard (1988), and others, there are morphosyntactic a½xes and morphophonological ones and the mapping between the two is not necessarily one to one. We do not go into this issue here, but simply restrict the discussion to clear cases in which the mapping is one to one. 6. For an overview of theories of English synthetic compounds, see Spencer 1991:324¨. 7. Note that Baker explicitly claims that synthetic compounds and simple N-V compounds in languages like English do not involve syntactic head-to-head movement. 8. Certain other a½xes allow inheritance of accusative arguments. An example in English is -ing in John's singing the Marseillaise. Such a½xes derive ``mixed categories,'' so called because the lower and higher parts of their projection seem to be of di¨erent categories. At ®rst sight, mixed categories present a stronger argument for a½xation in syntax (see Jackendo¨ 1977; Abney 1987; Hazout 1995; and others). However, two alternative analyses are available. One is that the stem and the a½x introduce two sets of categorial features that are projected simultaneously at ®rst. The a½x's features project higher than those of the stem, due to independent well-formedness conditions on structure, as argued by Reuland (1988). The second alternative is provided by Bresnan's (1997) head-sharing analysis, which makes use of the modularity of grammar in LFG: two distinct heads in categorial structure are mapped onto a single word in functional structure. Neither analysis needs to resort to syntactic a½xation. 9. A theoretical argument in favor of a left-branching ([[N V] -er]) rather than right-branching ([N [V -er]]) analysis of synthetic compounds can be given as well. Thematic roles are discharged in accordance with the thematic hierarchy: a theta role lower on the hierarchy is discharged before a theta role higher on the hierarchyÐthat is, it is discharged to a position lower in the tree. Since the agentive su½x -er binds the verb's external theta role, its internal theta role must be discharged before this su½x is attached. Truck driver should hence have a leftbranching structure.

56

Peter Ackema and Ad Neeleman

The grammaticality of driver of trucks is not incompatible with this idea. Although the thematic hierarchy holds within both syntax and morphology, it arguably does not apply across components. In driver of trucks one of the verb's theta roles is discharged in morphology (it is bound by -er), while the other is assigned in syntax. Hence, no violation of the thematic hierarchy obtains in either the syntactic or morphological component. 10. This is not true for all languages; see section 2.6 for discussion. 11. Van Hout and Roeper (1998) argue that forms like driver of a truck must involve syntactic a½xation of -er to a full-blown extended verbal projection, including at least VP and AspP. Their argument is based on a semantic contrast between driver of a truck and truck driver: only the former entails that there is a truck that has been driven at some point. According to Van Hout and Roeper, this necessitates the presence of AspP, since this is the locus of event interpretation. To us it seems that the contrast reduces to the presence of a determiner in driver of a truck. Nouns in compounds cannot be accompanied by determiners (see Hoeksema 1988), hence they must be interpreted as nonreferential. 12. Independent evidence for the assumption that the inclusion of an a½x in the derivation triggers morphology comes from the acquisition of synthetic compounds, as discussed by Clark, Hecht, and Mulford (1986). It turns out that children only acquire the morphological OV order after they have acquired su½xes like -er. 13. An alternative is that there is only one variant of particle verbs like schoonmaken `cleanmake', but that this form has special properties that make it transparent to both syntactic and morphological constraints. A proposal that comes close to this is made by Booij (1990), who introduces a V* category, in between V 0 and V 0 . Such a proposal can be interpreted in one of two ways. In one interpretation, V* behaves as a word if embedded under an a½x, and as a syntactic unit otherwise. Since this in e¨ect gives up on uniform realization, just like the analysis proposed in the main text, the introduction of a V* category is rendered super¯uous. In the other interpretation, V* is subject to both syntactic and morphological constraints in all its occurrences. This is not in accordance with the data discussed in the main text, however. These data show that if embedded in a word the particle verb behaves like any other compound. If it occurs without a su½x, it behaves like any other syntactic complex predicate. 14. This is orthogonal to the conclusion drawn by Baayen and Schreuder (1998) and others that frequently used rule-governed complex words are stored to minimize computational load. (The same may be true of frequently used phrases; see Jackendo¨ 1997). Although such forms may be stored with their (regular) semantics, this does not imply that they are stored with a complex structure. For all intents and purposes, such words function as simplexes. This is corroborated by the fact that frequently used compounds sometimes receive noncompound stress. Thus, inhabitants of Utrecht speak of the domPLEIN `Dom-square' (the name of the square next to the Dom tower in Utrecht), while strangers will ask for the DOMplein, with stress on the left-hand part, as in compounds. 15. The morpheme -en- that shows up in hartenbreker is a linking morpheme, not the plural. 16. In some cases it is possible, though, to suppress some of the material of the syntactic idiom when it is realized morphologically. In that case, we assume that the possibility of suppression must be stored as part of the morphological idiom. An example is the suppression of om de mond `around the mouth' in the compound stroopsmeerder `syrup smearer' (see (30a,a 0 )).

Competition between Syntax and Morphology

57

17. We are claiming that if a process is semantically transparent, it must be expressed by a½xation. This does not imply that every single form derived by a½xation must be semantically transparent. Like other expressions, derived words can be listed with a particular, nonpredictable, meaning. This is also true of the a½xes mentioned in the main text. 18. Of course, it is often assumed that in¯ectional morphology is never category changing. Whether this is correct or not is irrelevant to the argumentation in this section. Haspelmath (1996), for example, discusses cases of mixed categories in terms of category-changing in¯ection; see also note 8. 19. Note that the hypothesis that morphology can apply to the syntactic tree does not mean that syntax and morphology are a single rule system. 20. Recall that No Morphology was motivated by the fact that it is more costly to activate two rule systems than it is to activate just one. Given the reformulation in (41a), we can be more precise: what is costly is leaving a rule system in the middle of a derivation and storing the state the derivation is in at that point in short-term memory. 21. Somewhat surprisingly from a typological perspective, the Germanic language Frisian also appears to be of this type; see Dijk 1997. References Abney, Stephen. 1987. The English Noun Phrase in Its Sentential Aspect. Doctoral dissertation, MIT, Cambridge, Mass. Ackema, Peter. 1999a. Issues in Morphosyntax. Amsterdam: Benjamins. Ackema, Peter. 1999b. The Nonuniform Structure of Dutch N-V Compounds. In Geert Booij and Jaap van Marle (eds.), Yearbook of Morphology 1998, 127±158. Dordrecht: Kluwer. Ackerman, Farrell, and Philip LeSourd. 1997. Toward a Lexical Representation of Phrasal Predicates. In Alex Alsina, Joan Bresnan, and Peter Sells (eds.), Complex Predicates, 67±106. Stanford, Calif.: CSLI Publications. Anderson, Stephen. 1969. West Scandinavian Vowel Systems. Doctoral dissertation, MIT, Cambridge, Mass. Andrews, Avery. 1990. Uni®cation and Morphological Blocking. Natural Language and Linguistic Theory 8: 507±557. Baayen, Harald, and Rob Schreuder. 1998. The Balance of Storage and Computation in the Mental Lexicon: The Case of Morphological Processing in Language Comprehension. Paper presented at the Congress on Storage and Computation in Linguistics, Utrecht University. Baker, Mark. 1988. Incorporation. Chicago: University of Chicago Press. Baker, Mark. 1996. The Polysynthesis Parameter. Oxford: Oxford University Press. Beard, Robert. 1988. On the Separation of Derivation from A½xation: Toward a Lexeme/ Morpheme-Based Morphology. Quaderni di Semantica 9: 3±59. Booij, Geert. 1990. The Boundary between Morphology and Syntax: Separable Complex Verbs in Dutch. In G. Booij and J. van Marle (eds.), Yearbook of Morphology 1990, 45±63. Dordrecht: Foris. Borer, Hagit. 1993. Parallel Morphology. Ms., University of Massachusetts, Amherst.

58

Peter Ackema and Ad Neeleman

Bresnan, Joan. 1982. The Passive in Lexical Theory. In Joan Bresnan (ed.), The Mental Representation of Grammatical Relations, 3±86. Cambridge, Mass.: MIT Press. Bresnan, Joan. 1997. Mixed Categories as Head Sharing Constructions. In Miriam Butt and Tracy Holloway King (eds.), Proceedings of the LFG 97 Conference. Stanford, Calif.: CSLI Publications. Bresnan, Joan. 2000. Explaining Morphosyntactic Competition. In Mark Baltin and Chris Collins (eds.), Handbook of Contemporary Syntactic Theory. Oxford: Blackwell. Bresnan, Joan, and Sam Mchombo. 1995. The Lexical Integrity Principle: Evidence from Bantu. Natural Language and Linguistic Theory 13: 181±252. Carstairs-McCarthy, Andrew. 1992. Current Morphology. London: Routledge. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Clark, Eve, Barbara Hecht, and Randa Mulford. 1986. Coining Complex Compounds in English: A½xes and Word Order in Acquisition. Linguistics 24: 7±29. Coopmans, Peter, and Martin Everaert. 1988. The Simplex Structure of Complex Idioms: The Morphological Status of laten. In M. Everaert, A. Evers, R. Huybregts, and M. Trommelen (eds.), Morphology and Modularity, 75±104. Dordrecht: Foris. Dijk, Siebren. 1997. Noun Incorporation in Frisian. Doctoral dissertation, University of Groningen (Ljouwert: Fryske Akademy). Dikken, Marcel den. 1995. Particles. Oxford: Oxford University Press. Di Sciullo, Anna-Maria, and Edwin Williams. 1987. On the De®nition of Word. Cambridge, Mass.: MIT Press. Drijkoningen, Frank. 1994. A½xation and Logical Form. In Reineke Bok-Bennema and Crit Cremers (eds.), Linguistics in the Netherlands 1994, 25±36. Amsterdam: Benjamins. Grimshaw, Jane. 1981. Form, Function, and the Language Acquisition Device. In C. L. Baker and John McCarthy (eds.), The Logical Problem of Language Acquisition, 165±182. Cambridge, Mass.: MIT Press. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28: 373±422. Groos, Anneke. 1989. Particle-Verbs and Adjunction. In Hans Bennis and Ans van Kemenade (eds.), Linguistics in the Netherlands 1989, 51±60. Dordrecht: Foris. Haspelmath, Martin. 1996. Word-Class-Changing In¯ection and Morphological Theory. In Geert Booij and Jaap van Marle (eds.), Yearbook of Morphology 1995, 43±66. Dordrecht: Kluwer. Hazout, Ilan. 1995. Action Nominalizations and the Lexicalist Hypothesis. Natural Language and Linguistic Theory 13: 355±404. Hoeksema, Jack. 1988. Head-Types in Morphology. In Geert Booij and Jaap van Marle (eds.), Yearbook of Morphology 1988, 123±137. Dordrecht: Foris. Hoeksema, Jack. 1991. Theoretische aspecten van partikelvooropplaatsing. Tabu 21: 18± 26. Holmes, Philip, and Ian Hinchli¨e. 1994. Swedish: A Comprehensive Grammar. London: Routledge. Jackendo¨, Ray. 1977. X-Bar Syntax. Cambridge, Mass.: MIT Press.

Competition between Syntax and Morphology

59

Jackendo¨, Ray. 1997. The Architecture of the Language Faculty. Cambridge, Mass.: MIT Press. Josefsson, GunloÈg. 1998. Minimal Words in a Minimal Syntax. Amsterdam: Benjamins. Kiparsky, Paul. 1973. Elsewhere in Phonology. In Stephen Anderson and Paul Kiparsky (eds.), A Festschrift for Morris Halle, 93±106. New York: Holt, Rinehart and Winston. Lieber, Rochelle. 1983. Argument Linking and Compounds in English. Linguistic Inquiry 14: 251±285. LuÈdeling, Anke. 1997. Between Particle Verbs and Resultative Constructions. Paper presented at the DGfS Jahrestagung, DuÈsseldorf. McCarthy, John, and Alan Prince. 1994. The Emergence of the Unmarked: Optimality in Prosodic Morphology. In MerceÁ GonzaÁlez (ed.), NELS 24, 333±379. GLSA, University of Massachusetts, Amherst. Mithun, Marianne. 1984. The Evolution of Noun Incorporation. Language 60: 847±894. Mohanan, Tara. 1995. Wordhood and Lexicality: Noun Incorporation in Hindi. Natural Language and Linguistic Theory 13: 75±134. Neeleman, Ad. 1994. Complex Predicates. Doctoral dissertation, Utrecht University. Neeleman, Ad, and Fred Weerman. 1993. The Balance between Syntax and Morphology: Dutch Particles and Resultatives. Natural Language and Linguistic Theory 11: 433±475. Neeleman, Ad, and Fred Weerman. 1999. Flexible Syntax. Dordrecht: Kluwer. Nieuwborg, E. R. 1969. Retrograde Woordenboek van de Nederlandse Taal. Antwerp: Plantyn. Pesetsky, David. 1982. Paths and Categories. Doctoral dissertation, MIT, Cambridge, Mass. Poser, William. 1992. Blocking of Phrasal Constructions by Lexical Items. In Ivan Sag and Anna Szabolcsi (eds.), Lexical Matters, 111±130. Stanford, Calif.: CSLI Publications. Preuss, F. 1960. Backformation oder Noun-Incorporation. Lebende Sprachen 5: 110±112, 165± 167. Reuland, Eric. 1988. Relating Morphological and Syntactic Structure. In Martin Everaert, Arnold Evers, Riny Huybregts, and Mieke Trommelen (eds.), Morphology and Modularity, 303±337. Dordrecht: Foris. Roeper, Thomas, and Mu¨y Siegel. 1978. A Lexical Transformation for Verbal Compounds. Linguistic Inquiry 9: 199±260. Rosen, Sara Thomas. 1989. Two Types of Noun Incorporation: A Lexical Analysis. Language 65: 294±317. Sells, Peter. 1998. Optimality and Economy of Expression in Japanese and Korean. In N. Akatsuka, H. Hoji, S. Iwasaki, S.-O. Sohn, and S. Strauss (eds.), Japanese/Korean Linguistics 7. Stanford, Calif.: CSLI Publications. Shimamura, R. 1983. Backformation of English Compound Verbs. In J. F. Richardson, M. Marks, and A. Chuckerman (eds.), Papers from the Parasession on the Interplay of Phonology, Morphology, and Syntax, 271±282. Chicago Linguistic Society, University of Chicago. Spencer, Andrew. 1991. Morphological Theory. Oxford: Blackwell. Sproat, Richard. 1985. On Deriving the Lexicon. Doctoral dissertation, MIT, Cambridge, Mass.

60

Peter Ackema and Ad Neeleman

Van Dale. 1984. Groot Woordenboek der Nederlandse Taal. Utrecht: Van Dale Lexicogra®e. Van Hout, Angeliek, and Thomas Roeper. 1998. Events and Aspectual Structure in Derivational Morphology. In Heidi Harley (ed.), MIT Working Papers in Linguistics 32: Papers from the UPenn/MIT Roundtable on Argument Structure and Aspect, 175±200. MITWPL, Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Williams, Edwin. 1981. On the Notions ``Lexically Related'' and ``Head of a Word.'' Linguistic Inquiry 12: 245±274. Williams, Edwin. 1997. Blocking and Anaphora. Linguistic Inquiry 28: 577±628.

Chapter 3 Markedness and Subject Choice in Optimality Theory

Judith Aissen

Among the most robust generalizations in syntactic markedness is the association of semantic role with person/animacy rank, discussed ®rst in Michael Silverstein's 1976 paper ``Hierarchy of Features and Ergativity.'' A version of Silverstein's hierarchy is given in (1a) (the format is di¨erent from his original): ®rst and second personsÐhere called the ``local'' personsÐoutrank third, and within the third person there is a further ranking of various subcategories. (1) a. Local Person > Pronoun 3rd > Proper Noun 3rd > Human 3rd > Animate 3rd > Inanimate 3rd b. Agent > Patient The hierarchy in (1a) must be understood in connection with the semantic role hierarchy in (1b). Silverstein's claim is that the unmarked situation is for elements on the upper end of (1a) to be agents in transitive propositions and for elements on the lower end to be patients (Silverstein 1976:123). Evidence for this is that in many languages, expression of more marked con®gurations is morphologically more complex than that of less marked ones. Interestingly, the markedness expressed in (1) is realized in a number of di¨erent ways: through case marking, through the category of direction (direct vs. inverse), and through the category of voice (active vs. passive). The generalization expressed in (1) is an important result in universal grammar, and appropriately occupies a prominent place in the typological and functional literature. The present chapter explores how Silverstein's generalization might be expressed in a formal theory of grammar, and how it can play a role in individual grammars. The account is developed within Optimality Theory (OT; Prince and Smolensky 1993), since OT is ``in many respects a formal theory of markedness'' (Smolensky 1996a). Central to the account are two formal devices that have been proposed in connection with phonology: harmonic alignment of prominence scales, and local conjunction of constraints. I begin by sketching the ways in which the markedness expressed by (1) is realized. Throughout, attention is restricted to the part of the hierarchy in (1a) that concerns person. This narrows the scope of the present inquiry

62

Judith Aissen

Table 3.1 Person-based split-ergative case-marking system (Silverstein 1976) Local persons 3rd person Case

Unmarked

Marked

Subject Object Nominative/Absolutive

Object Subject (of transitive) Accusative/Ergative

to manageable size; whether the approach suggested here extends to the rest of the hierarchy remains to be seen. 3.1

The Realization of Markedness

Silverstein proposed that the markedness represented by (1) underlies split-ergative case marking in languages where the split is based on person and/or animacy (see also Dixon 1994). Table 3.1 shows case marking in a language like Dyirbal (Dixon 1972; Silverstein 1976) where, reading across, ®rst-/second-person pronouns are marked when they are objects, but not when they are subjects; conversely, third persons are marked when they are subjects of transitive clauses, but not when they are objects (or subjects of intransitive clauses).1,2 Since morphological marking coincides with syntactic markedness, systems of this sort involve a ®ne-grained matching of the two (Silverstein 1976:149¨.). Reading down table 3.1, note that the maximally unmarked clause is one in which the subject is a local person and the object is third person; the maximally marked clause is one in which the subject is third person and the object is a local person. Dyirbal is a dependent-marking language (Nichols 1986), one in which grammatical relations are marked on nominals themselves. Silverstein (1976:160) notes and DeLancey (1981) has developed further the idea that the same markedness that is expressed through the case system in Dyirbal underlies the distribution of direct and inverse verbs in other languages. The category of direction, within which direct and inverse verbs contrast, is generally limited to head-marking languages (Nichols 1986; Klaiman 1993). Direction is relevant only in transitive clauses, with the distribution of direct and inverse verbs determined by the relative position of subject and object along some dimension. In Nocte, for example, a Tibeto-Burman language discussed in DeLancey (1981), the direct verb is used when the subject is ®rst or second person and the object is third (the unmarked con®guration), while the inverse verb is used when the subject is third person and the object is ®rst or second (the marked con®guration). See table 3.2. Here, the crucial assumption is that the inverse verb is the marked member of the inverse-direct opposition, an assumption motivated in Nocte by the fact that the inverse category is marked by overt morphological material and

Markedness and Subject Choice in Optimality Theory

63

Table 3.2 Person-based direction system (Nocte, DeLancey 1981) Local persons 3rd person Verb form

Unmarked

Marked

Subject Object Direct

Object Subject Inverse

Table 3.3 Person-based voice system (Lummi, Jelinek and Demers 1983) Local persons 3rd person Verb form

Unmarked

Marked

Subject (Agent) Object (Patient) Active

Subject (Patient) Oblique (Agent) Passive

the direct category by j (DeLancey 1981). Again, morphological marking coincides with syntactic markedness. The same markedness relations determine the distribution of voice in yet other languages (DeLancey 1981; Jelinek 1990, 1993; GivoÂn 1994). Lummi (Coast Salish), for example, has an active-passive opposition, but the two are in partial complementary distribution, depending on the person of agent and patient (Jelinek and Demers 1983). Per table 3.3, only the active occurs when the agent is a local person and the patient third person; only the passive occurs when the agent is third person and the patient a local person.3,4 Here, the unmarked voice (active) is only compatible with the unmarked association of person with semantic role. To express the marked association, a marked construction is required, the passive (on relative markedness of active and passive, see Greenberg 1966:46). 3.2

Language-Particular Hierarchies

Analyses of facts such as these often involve the positing of language-particular hierarchies. It is indisputable that languages vary in the elements they rank. Nonetheless, language-particular hierarchies pose signi®cant problems for a general theory of markedness. For the purposes of discussion, (2) lists several hierarchies posited for languages with a direct-inverse opposition; (3) lists hierarchies in several languages with an active-passive opposition. These illustrate how languages can vary in the inventory of elements they rank.

64

Judith Aissen

(2) Some hierarchies in languages with direction systems a. 1 > 2 > 3 (Nocte, DeLancey 1981) b. local > proximate > obviative > inanimate (Fox, Dahlstrom 1995:99) (3) Some hierarchies in languages with voice systems a. local > 3 (Lummi, Jelinek and Demers 1983) 2>3 (Squamish, Jelinek and Demers 1983) j (Lushootseed, Jelinek, and Demers 1983) b. 2 > 3 animate pronoun > animate > inanimate (Chamorro, Chung 1998:34) c. proximate > obviative (Tzotzil, Aissen 1997) (The categories proximate and obviative that ®gure in (2b) and (3c) refer to subcategories within the third person; see the references cited.) Thus, Nocte, Fox, and Lummi rank local persons over third, while Tzotzil does not. On the other hand, Fox and Tzotzil rank proximate over obviative within the category of third person, while Nocte and the Coast Salish languages do not.5 Chamorro ranks second person (alone) over third, and so does Squamish, a property not shared with the other languages. While languages vary a good deal in which elements they rank, the rankings they do observe are largely invariant. When they are ranked, local persons always outrank third; where there is ranking between proximate and obviative, proximate outranks obviative; animate always outranks inanimate; and so on. To the extent that the rankings found in particular languages instantiate universal rankings, they should not be stipulated. But this seems unavoidable when one and the same format is used to indicate both the elements ranked (a language-particular matter) and their rankings (largely universal). There is also a descriptive problem with the hierarchies in (3), which are all designed speci®cally to express restrictions on the pairing of person with grammatical function. All these hierarchies are really constraints on pairings in active clauses (and all the sources cited in (3) restrict them accordingly). However, all the languages listed in (3) are subject to additional restrictions in passive clauses that are not expressed by the hierarchies in (3): namely, in none of these languages can local persons be agents in passive clauses (Jelinek and Demers 1983; Aissen 1987:63; Cooreman 1987:112). This restriction is as much a consequence of the markedness expressed by the hierarchies in (1) as are the restrictions on active clauses stated in (2) and (3): the maximally unmarked situation is for local persons to be agents, hence propositions in which they are should be expressed in the unmarked voiceÐactive, not passive. If person-based restrictions in active and passive clauses express the same markedness, they should be expressed in a formally uni®ed way. But in general, the hierarchies in (3) do not generalize to passive clauses because person-based constraints are di¨erent in active and passive clauses. Among other things, in all the languages

Markedness and Subject Choice in Optimality Theory

65

except Lummi, there are no restrictions on ®rst person in active clauses, but there are in passive clauses. Addition of ®rst person to these hierarchies to account for the passive yields incorrect predictions about the active. The language-particular hierarchy format for representing person-based constraints seems to stand in the way of formally unifying person-based constraints in all clauses. In the actual practice of grammatical description, person-based restrictions in passive clauses are stated separately, with no suggestion of any substantive connection to those in active clauses. This suggests the need for a di¨erent approach. Apparently it is necessary to pick out very speci®c associations of person and grammatical functionÐfor example, Squamish excludes second-person objects (relevant only in active clauses) and local-person agents (relevant only in passive clauses). The fact that these very speci®c language-particular constraints re¯ect universal principles of markedness must be expressed via the relation between universal grammar and language-particular grammars. Language-particular hierarchies pose a particularly pressing problem for OT: in short, they have no status. An OT-grammar is a ranking of universal constraints, but language-particular hierarchies like those of (2) and (3) are not universal constraints. 3.3

Eliminating Language-Particular Hierarchies

We can make headway on these various fronts by replacing language-particular hierarchies with language-particular rankings of simple, universal constraints, an approach that has been developed within OT for phonology. In what follows, I draw heavily on ideas concerning voice and typology presented in Legendre, Raymond, and Smolensky 1993, and on harmonic alignment as developed in Prince and Smolensky 1993. Once we have reconceived hierarchies as a set of ranked constraints, we can solve the conceptual and descriptive problems discussed earlier. The analysis rests on a set of proposed universal prominence scales that are part of universal grammar. Some candidates for this status are listed in (4) (these correspond to a subset of those needed to account for the facts discussed below). (4) a. Person scale: b. Role scale: c. Relational scale:

Local > 3rd [Local ˆ 1st, 2nd] Agent > Patient Subject > Nonsubject

For the most part, these rankings are not controversial and have been assumed (or proposed) by linguists working in a variety of frameworks.6 Since the analyses developed below depend in a fundamental way on these scales, it is positive that a broad consensus surrounds them. In the analyses that follow, however, these scales are not taken to express markedness per se, but a ranking. Thus, local persons are not inherently less marked than third: they are less marked qua (transitive) subjects and more marked qua objects. (This situation is what is characterized in the literature on

66

Judith Aissen

markedness as markedness reversalÐi.e., situations in which what is relatively marked in one context is relatively unmarked in another (Battistella 1990, 1996; Croft 1990). The analysis proposed below predicts markedness reversal in this domain.) Expressing this generalization, which I assume to be universal, requires that the person scale and the relational scale be brought into alignment (DeLancey 1981), for it is particular associations of person and grammatical function that are relatively marked or unmarked. A formal technique for aligning two prominence scales (or dimensions) is de®ned in Prince and Smolensky 1993 and discussed there in the context of syllable structure and sonority. (5) Alignment. Suppose given a binary dimension D1 with a scale X > Y on its elements {X, Y}, and another dimension D2 with a scale a > b


> z on its elements. The harmonic alignment of D1 and D2 is the pair of Harmony scales. Hx : X/a  X/b 


 X/z Hy : Y/z 


 Y/b  Y/a The constraint alignment is the pair of constraint hierarchies. Cx : *X/z g


g *X/b g *X/a Cy : *Y/a g *Y/b g


g *Y/z

(Prince and Smolensky 1993:136)

In the original example, D1 is a scale of structural positions within the syllable (Peak > Margin), and D2 is the sonority hierarchy (where, inter alia, vowels > nonvowel sononants > obstruents). Harmonic alignment of the two prominence scales associates the high-ranking elements on the two scales (vowels with peaks), as well as the low-ranking elements on the two (obstruents with margins). The result is two harmonic scales (Hx , Hy above), one of which concerns peaks and the other margins. (These scales contain the connective , which is read as ``more harmonic than.'') The scale on peaks says that it more harmonic (less marked) for vowels to be peaks than for nonvowels, and more harmonic for nonvowel sonorants to be peaks than for obstruents; the scale on margins says that it is more harmonic for obstruents to be margins than sonorants, and more harmonic for nonvowels to be margins than vowels. The basic idea is that a position that is structurally prominentÐfor example, the peakÐattracts elements prominent along some relevant substantial dimension, like sonority, while a position low in structural prominence, such as the margin, attracts elements low on relevant structural dimensionsÐfor instance, sonority. Harmonic alignment thus derives markedness reversal. Our concern here is the relative markedness (or Harmony) that characterizes the association of structural positions within the clause (e.g., subject, object) with the dimension of person. In most languages, subject position plays a role in the clause

Markedness and Subject Choice in Optimality Theory

67

Table 3.4 Alignment of person and grammatical relation (encapsulated) Scales

Harmonic alignment

Constraint alignment

Local >3 Su > Non-Su

Su/Local  Su/3 Non-Su/3  Non-Su/Local

*Su/3 g *Su/Local *Non-Su/Local g *Non-Su/3

analogous to that played by the peak in syllable structure: it is the most prominent structural position and thereby attracts elements relatively prominent on other substantial dimensions. In syntax, perhaps more than in phonology, there are a number of dimensions relevant to subject choice. Depending on the language, subject position may attract local persons over third, proximates over obviatives, agents over patients, and so on. It is the job of constraint ranking to adjudicate in particular languages between the various dimensions that play a role, crosslinguistically, in subject choice. The generation of constraint hierarchies through harmonic alignment of person and grammatical relation is illustrated in table 3.4 (for other applications of harmonic alignment in syntax, see Artstein 1998 and Asudeh 1999). The left-hand column shows the universal scales involved, person, and grammatical relation. The person scale ranks the local persons (®rst and second) with respect to third, but does not rank ®rst and second with each other.7 The relational scale is presented in skeletal form here for the sake of illustration, and simply ranks subject over nonsubject.8 The middle column shows the Harmony scales, which are derived through harmonic alignment (5). These express the fact that a local-person subject is more harmonic (less marked) than a third-person subject. Conversely, a third-person nonsubject is less marked than a local-person nonsubject. The relative markedness of particular associations is stated independently for subject and nonsubject, a point that the Australian literature in particular argues is necessary (Silverstein 1976; Dixon 1994). The right-hand column contains the constraints that actually ®gure in grammar construction. These are stated as Avoid constraints, and are derived by inverting the rankings in the middle box and pre®xing the Avoid operator ``*''. The ranking of *Su/3 over *Su/Local means that in the absence of any relevant higher-ranking constraints, a clause with a third-person subject will lose out in direct competition to a clause with a ®rst- or second-person subject. (I assume that faithfulness to person and to semantic role must be satis®ed; given alternative expressions of the same propositional content, one with a third-person subject and another with a ®rst- or secondperson subject, the latter will win.) The constraints in the right-hand column are proposed as universal, as are their rankings. That is, each set constitutes a constraint subhierarchy. The ranking of constraints in a subhierarchy is universally ®xed and expresses the universal markedness relations in this domain. Language-particular

68

Judith Aissen

Table 3.5 Alignment of person and grammatical relation Scales

Harmonic alignment

Constraint alignment

Local > 3 1>3 2>3

Su/Local  Su/3 Su/1  Su/3 Su/2  Su/3

*Su/3 g *Su/Local *Su/3 g *Su/1 *Su/3 g *Su/2

Su > Non-Su Su > Oj Su > Obl

Oj/3  Oj/Local Oj/3  Oj/1 Oj/3  Oj/2

*Oj/Local g *Oj/3 *Oj/1 g *Oj/3 *Oj/2 g *Oj/3

Obl/3  Obl/Local Obl/3  Obl/1 Obl/3  Obl/2

*Obl/Local g *Obl/3 *Obl/1 g *Obl/3 *Obl/2 g *Obl/3

variation can be described through the interpolation of other constraints among those in a subhierarchy, but not through di¨erences in ranking within the subhierarchy itself. In table 3.4, local person subsumes ®rst and second persons, and nonsubject subsumes object and passive agent. In some grammars, however, relational constraints on ®rst and second persons are di¨erent, as are person constraints on direct object and passive agent. Hence, the constraints in table 3.4 encapsulate the more expanded set shown in table 3.5, where the two scales of table 3.4 are each decomposed, yielding a total of four scales. I use the term Oblique (Obl) here to refer exclusively to the syntactic relation borne by the agent in a passive clause. Some of the constraints in the right-hand column are restricted to particular clause types: those that begin *Oj/ penalize only active clauses, since passive clauses contain no object. Those that begin *Obl/ penalize only passive clauses, since active clauses contain no passive agent. Hence, although these constraints are generated from the same initial scale (Su > Non-Su), decomposition of Non-Su into Oj and Obl yields constraints that distinguish active from passive clauses, and are ranked independently. At the same time, since these constraints are ultimately derived from alignment of the same two initial scales (the person scale, the relational scale), they express the same markedness relations. Conceptually, then, this account uni®es person-based constraints in active and passive clauses while making it possible to distinguish them. Among other dimensions relevant to subject choice are semantic role and thematic (discourse) prominence. These are treated here in much the same way as personÐ that is, through harmonic alignment. Relevant constraints are shown in table 3.6. Analogous constraints were proposed in di¨erent formalism by Legendre, Raymond, and Smolensky (1993). The ®rst set (top row) concerns the association of semantic role and grammatical function, and asserts that the unmarked situation is

Markedness and Subject Choice in Optimality Theory

69

Table 3.6 Alignment of grammatical relation with semantic role and thematic prominence Scales

Harmonic alignment

Constraint alignment

Agt > Pat

Su/Agt  Su/Pat

*Su/Pat g *Su/Agt

Su > Oj

Oj/Pat  Oj/Agt

*Oj/Agt g *Oj/Pat

X>x

Su/X  Su/x

*Su/x g *Su/X

Su > Oj Su > Obl

Oj/x  Oj/X Obl/x  Obl/X

*Oj/X g *Oj/x *Obl/X g *Obl/x

for agent to be subject and for patient to be object. I intend the terms agent and patient here to correspond to the Proto-Agent and Proto-Patient roles of Dowty 1991, roles conceived as cluster concepts de®ned in terms of P-Agent and P-Patient properties. The Proto-Agent is the argument with the greater number of P-Agent properties; the Proto-Patient is the argument with the greater number of P-Patient properties. Each role subsumes a large number of speci®c roles and there is no implication, for example, that the Proto-Agent acts with intentionality or is a causal agent. The constraint hierarchies in the top half of table 3.6 correspond to Dowty's Argument Selection Principle (ASP), which says that the argument with the greatest number of P-Agent properties is realized as the subject, and the argument with the greatest number of P-Patient properties is realized as the object. *Su/Pat is violated then if the argument with the greater number of P-Patient properties is realized as the subject; *Oj/Agt is violated if the argument with the greater number of P-Agent properties is realized as the object. However, these constraints di¨er in two important ways from Dowty's ASP. First, the ASP determines subject and object only for active, transitive clauses, while the constraints *Su/Pat and *Oj/Agt apply to both active and passive clauses. Of course, *Su/Pat is systematically violated in passive clauses, since patient subjects are not avoided in passive clauses: the passive functions precisely to realize the patient as subject. This underscores the second di¨erence: *Su/Pat and *Oj/Agt are conceived as violable constraints, while Dowty's ASP is not (see also Asudeh 1999). The constraint *Su/Pat (``avoid subject patient'') (italicized in table 3.6) is pivotal in the present analysis because it penalizes any passive, and depending on its ranking relative to other constraints on subject choice, can selectively admit passive clauses or can eliminate the possibility of passive altogether. Passive clauses will be possible, then, when some constraint higher ranked than *Su/Pat penalizes a patient object, or an agent subject. The second set of constraints (bottom row) deals with pressures on subject choice that are grounded in discourse prominenceÐthat is, are related to topicality, empathy, perspective, discourse coherence, and so on. The basic insight is that the

70

Judith Aissen

unmarked situation is for a more prominent argument to be subject and for a less prominent one to be nonsubject (Chafe 1976; GivoÂn 1976; Keenan 1976; Kuno and Kaburaki 1977; and many others). The harmonic alignments in the middle column express this where, following Legendre, Raymond, and Smolensky (1993), capital X represents a discoursally more prominent element and lowercase x a less prominent one. I assume that the default is for all nominals to be neutral in discourse prominence, but that asymmetries in prominence can result through the raising or depression of an argument's prominence. Again, some of the constraints in the right-hand column restrict only some clauses: *Oj/X penalizes active clauses (with prominent objects), while *Obl/X penalizes passives (with prominent agents). The constraints in tables 3.2 and 3.3 concern the harmonic alignment of grammatical function along three di¨erent dimensions: person, semantic role, and discourse prominence. The important point is that while the ranking of constraints within a subhierarchy is universally ®xed (e.g., *Su/3 always outranks *Su/Local), crosslinguistic variation arises because the interpolation of constraints from distinct subhierarchies can yield distinct overall constraint rankingsÐthat is, distinct grammars. In the construction of language-particular OT grammars, constraints that express preferences based on di¨erent dimensions can be ranked in di¨erent ways and thereby express di¨erent priorities for subject choice. In section 3.4, I show how voice systems with varying sensitivity to person can be described in terms of the constraints developed above. Section 3.5 considers in a somewhat preliminary way how the same constraints might play a role in determining the distribution of case and direction in languages like Dyirbal and Nocte. One feature of the analyses constructed below is worth mentioning at this point: in general, the low-ranked constraints in tables 3.1 to 3.3 do not ®gure crucially. This is because most of the subhierarchies that play a role here contain only two members and together exhaust the relevant domain. Thus all candidates that violate the highranked constraint will satisfy the low-ranked one, and vice versa. In such cases, the lower-ranked constraint can play no crucial role in the evaluation and can be dropped from consideration, a practice I follow here at various points. 3.4

Language-Particular Rankings

We consider now how the constraints of tables 3.2 and 3.3 can be interpolated to yield particular grammars. I assume that inputs consist of a predicate-argument structure, with (proto) semantic role, relative discourse prominence, and person of each argument speci®ed.9 Candidates are constructed by assigning a syntactic realization to each argument. Each candidate consists of a predicate with nominal arguments, where each of the latter is associated with a set of features that represents its morphosyntactic propertiesÐfor example, features indicating grammatical function,

Markedness and Subject Choice in Optimality Theory

71

person, and morphological case, when case is overtly marked. Following Legendre, Raymond, and Smolensky (1993) and Aissen (1997), I assume that active and passive clauses belong to the same candidate sets and thus compete with one another (see Ackema and Neeleman 1998 for an analysis in which active and passive clauses do not compete). In OT, the class of inputs and candidate sets is universal (see especially Prince and Smolensky 1993 and Smolensky 1996b); likewise, the set of constraints is essentially universal. Consequently, the source of language-particular variation lies not in differences in the class of inputs, candidate sets, or constraints, but in the ranking of the constraints. Further, there is no pre®ltering of inputs, and the fact that some languages have passive and some do not, for example, cannot be explained simply by assuming that the former have passive candidates (or inputs) while the later lack them. Since passive clauses occur in some languages, the grammar of every language must evaluate passive candidates and the grammar of a language lacking passive clauses must contain a high-ranking constraint (or constraints) that excludes such clauses. The fact that the lexicon (or morphology) of such a language lacks passive is a consequence of the language's constraint ranking. These assumptions are fundamental if OT is to provide an account of typological variation. In none of the languages under discussion here is the agent ever realized syntactically as object. That is, *Oj/Agt, which is one of the high-ranked constraints in table 3.6, is unviolated.10 I simplify by excluding candidates corresponding to this assignment, with the result that all candidates are either active (agent ! subject, patient ! object) or passive (patient ! subject, agent ! passive agent). 3.4.1 Inverse Languages Consider ®rst languages like Fox and Nocte in which the markedness relations represented by (1) are expressed through the direct-inverse opposition. In such languages, all sentences are, by hypothesis, active, with grammatical function entirely determined by semantic role (see, in particular, Dahlstrom 1995 for arguments that inverse clauses in Fox are active; see LeSourd 1976 for a di¨erent view). Following ideas of Legendre, Raymond, and Smolensky (1993), these languages can be described by ranking *Su/Pat (as well as *Oj/Agt) higher than any of the constraint subhierarchies that penalize associations of grammatical function with person or prominence. (These constraint subhierarchies are represented below as *GR/Pers and *GR/w, respectively, where GR ranges over {Su, Oj, Obl}, Pers ranges over {1, 2, 3}, and w ranges over the two degrees of prominence {X, x}.) Recall that *Su/Pat penalizes a structure in which the argument with more P-Patient properties is realized as subject. (6)

Fox/Nocte *Su/Pat g


g {*GR/Pers, *GR/w}

72

Judith Aissen

Tableau T3.1 Fox V (Agt/3, Pat/1)

*Su/Pat

f ACTIVE Agt/Su/3±Pat/Oj/1 PASSIVE Pat/Su/1±Agt/Obl/3

*GR/Pers **

*!

**

Tableau T3.1 shows the evaluation of a clause whose input speci®es a third-person agent (indicated as Agt/3) and a ®rst-person patient (Pat/1). There are two candidates: one active and one passive, depending on how grammatical function is assigned to each of the two arguments. In the active candidate, the third-person agent is subject (indicated as Agt/Su/3) and the ®rst-person patient is direct object (Pat/Oj/1); in the passive, the patient is subject (Pat/Su/1) and the agent is syntactically oblique (Agt/ Obl/3). (Semantic role is marked for the candidates so that it is easier to determine satisfaction of constraints. I assume that candidates are always faithful to the input in semantic role.) Discourse prominence plays no role in determining grammatical function in this evaluation, so it is omitted from the input. The high-ranking constraint, *Su/Pat, eliminates the passive entirely, leaving the active as the optimal candidate. This accounts for the nonexistence of passives, but not for the fact that the inverse form of the active verb must be used here. I return to this in section 3.5.3. 3.4.2 The Relevance of Thematic Prominence to Voice By ranking *Su/Pat lowerÐthat is, among the constraints that penalize particular active clausesÐwe can begin to admit passive under various conditions. Consider ®rst a language like English. There is general agreement in the functional syntax literature that the choice between active and passive in English is primarily determined by the relative prominence of agent and patient. Crudely put: the passive is used if the patient is more prominent than the agent; the active is used if either the agent is more prominent than the patient, or if neither has greater prominence. This description is meaningful only if the term prominence can be given content, and this is, of course, di½cult. Studies of the distribution of active and passive in English discourse suggest two factors. One has to do with attention: an individual that is the center of attention at a given moment is more prominent at that moment than one that is not (Tomlin 1985). Empathy and topicality (``what the sentence is about'') are often relevant to choosing the center of attention. The other factor is that an individual mentioned in recent, local discourse is more prominent than one not mentioned. This factor is rel-

Markedness and Subject Choice in Optimality Theory

73

evant to discourse coherence: discourse coherence is increased by selecting as subject of one clause a nominal whose referent has been mentioned in the immediately preceding discourse (Thompson 1987).11 These factors determine a ranking of nominals in a clause that I will refer to as thematic prominence. The preference for thematically prominent subjects and the use of passive in English to present a thematically prominent nonagent as subject is acquired early by English-speaking children (Slobin 1994). Both Tomlin (1985) and Thompson (1987) conceive of the choice between active and passive as involving a competition for subject position between the agent and the theme.12 According to Tomlin (1985:76±77), The analysis of naturally occurring texts . . . reveals that the interaction of subject, agent, and thematic information can be described in a precise and explicit manner. The subject relation, for the texts examined, always takes the NP whose referent is highest in the proposed hierarchy of thematicity. . . . If the relative thematicity of the NPs competing for the subject relation is equal or neutralized, the subject then takes the agent. Thus the function of subject in the synchronic grammar of English, at least for this kind of communicative event, can be described as: The function of subject in English Subject encodes thematic information over agent.

Thompson (1987:501) puts it this way: ``It appears that users of English are content to code the agent as subject unless broadly thematic or more local cross-clausal considerations require an alternative coding.'' These conceptions clearly lend themselves to formalization through a set of ranked violable constraints that expresses the preference for a thematically prominent subject over an agent subject. The constraint ranking in (7) expresses this and denies person per se a role in subject choice in English. (7)

English *Su/x g *Su/Pat g *GR/Pers

*Su/x penalizes clauses with nonthematic subjects and thus favors clauses with thematic subjects. Ranking *Su/x over *Su/Pat in English represents the fact that the passive is a better choice than the active when the patient is more thematic than the agent. The evaluation of a clause with third-person agent/®rst-person patient, where the patient is speci®ed in the input as thematically prominent (X), is shown in tableau T3.2. *Su/x is violated by the active, because the subject is not prominent, but it is not violated by the passive, where the subject is prominent. Tableau T3.3 shows that when no argument is prominent, *Su/x is not decisive (both candidates violate it) and the choice between active and passive passes to *Su/Pat, which selects the active (see Tomlin 1985). The fact that *Su/Pat dominates all the constraints that penalize particular associations of person with grammatical function eliminates a role for person per se in the determination of voice in English.13

74

Judith Aissen

Tableau T3.2 English (prominent patient) V (Agt/3/x, Pat/1/X) ACTIVE Agt/Su/3/x±Pat/Oj/1/X

*Su/x

*Su/Pat

*!

*GR/Pers **

f PASSIVE Pat/Su/1/X±Agt/Obl/3/x

*

**

*Su/Pat

*GR/Pers

Tableau T3.3 English (no prominent argument) V (Agt/3/x, Pat/1/x) f ACTIVE Agt/Su/3/x±Pat/Oj/1/x PASSIVE Pat/Su/1/x±Agt/Obl/3/x

*Su/x * *

** *!

**

3.4.3 The Relevance of Person to Voice We turn now to cases in which person does play a role. As indicated earlier, exclusion of local-person passive agents is quite common. Lushootseed (Coast Salish) is a language that has roughly the same distribution of active and passive as English, but that in addition bars passive clauses with ®rst- and second-person agents (Jelinek and Demers 1983). This yields the voice distribution by person of agent and patient shown in table 3.7. Active is possible for all combinations, but passive is not. Relative to English, Lushootseed can be described by ranking high just one of the constraints on association of grammatical function with person, *Obl/Local. *Obl/Local involves alignment with the Oblique relation (which refers here to the relation borne by the agent in a passive clause), so it penalizes passive clauses. (8)

Lushootseed *Obl/Local g *Su/x g *Su/Pat g *GR/Pers

Under the ranking in (8), passive clauses with local-person agents are categorically excluded, regardless of the thematic prominence of the patient. Tableau T3.4 shows how the active can win in a clause with a thematic patient. Here, the passive violates the highest-ranked constraint and is excluded despite the fact that the patient is higher in thematic prominence than the agent. *Su/Pat dominates all other con-

Markedness and Subject Choice in Optimality Theory

75

Table 3.7 Distribution of voice by person in Lushootseed Agt +

Pat )

1 2 3

1

2

3

Ð active/*passive active/passive

active/*passive Ð active/passive

active/*passive active/*passive active/passive

Tableau T3.4 Lushootseed V (Agt/1/x, Pat/3/X)

*Obl/ Local

f ACTIVE Agt/Su/1/x±Pat/Oj/3/X PASSIVE Pat/Su/3/X±Agt/Obl/1/x

*Su/x

*Su/Pat

* *!

*GR/Pers

** *

*

straints that penalize associations of grammatical relation and person, re¯ecting the fact that aside from local-person passive agents, person plays no role in determining voice in Lushootseed. The top-ranked constraint (*Obl/Local) is part of the same subhierarchy as some of the constraints encapsulated in the lowest-ranked constraint (*GR/Pers), but since *Obl/Local is the highest-ranked constraint in its own subhierarchy (see table 3.5), the ranking in (8) is consistent with the universal rankings of table 3.5. The three top-ranked constraints all come from distinct constraint subhierarchies (one concerning person, one concerning thematic prominence, one concerning semantic role) and may freely interpolate. In clauses with third-person agents, the highest-ranked constraint is irrelevant and both active and passive survive that constraint as candidates. Under the ranking in (8), *Su/Pat excludes any passive except where passive is favored over active by *Su/x, just as in English. Thus, when the patient is thematically prominent, the passive is selected (see tableau T3.2). In the absence of a prominent argument, both active and passive violate *Su/x, and the choice between active and passive passes to *Su/Pat, which, as in English, selects the active as optimal (see tableau T3.3). Lummi, also Coast Salish, is like Lushootseed, except that active is not possible for all combinations: it is disallowed when the subject is third person and the object ®rst or second (Jelinek and Demers 1983). (See table 3.8.) Except when both agent and patient are third person, active and passive are in complementary distribution in

76

Judith Aissen

Table 3.8 Distribution of voice by person in Lummi Agt +

Pat )

1 2 3

1

2

3

Ð active/*passive *active/passive

active/*passive Ð *active/passive

active/*passive active/*passive active/passive

Tableau T3.5 Lummi (®rst-person patient) V (Agt/3/X, Pat/1/x) ACTIVE Agt/Su/3/X±Pat/Oj/1/x f PASSIVE Pat/Su/1/x±Agt/Obl/3/X

*Obl/ Local

*Oj/ Local

*Su/x

*Su/ Pat

*!

*GR/ Pers *

*

*

**

Lummi. Lummi can be described by interpolating *Oj/Local between the two highranking constraints of Lushootseed (8). (9)

Lummi *Obl/Local g *Oj/Local g *Su/x g *Su/Pat g *GR/Pers

The two high-ranked constraints in (9) both penalize highly marked con®gurations and are the high-ranked constraints within their own subhierarchies (see table 3.5). Their ranking over all the remaining constraints in *GR/Pers is thus consistent with the universal rankings of tables 3.2 and 3.3. Tableau T3.5 shows that the ranking in (9) excludes actives with ®rst-person objects (likewise second-person objects), even when the subject is speci®ed in the input as thematically prominent. *Oj/Local rules out the active, so passive is the optimal form. To achieve this, it is necessary for *Oj/Local to dominate *Su/x and *Su/Pat, both of which penalize the passive in this case. In clauses where both arguments are third person, the top two constraints are irrelevant, and the next two become decisive. They are ranked as in Lushootseed and English, with the same consequences: both active and passive are possible with thirdperson agent and patient, but the passive is (I assume) reserved for cases in which the patient is more thematically prominent than the agent. The preference for passive in this case is assured by the ranking of *Su/x over *Su/Pat, which fatally penalizes the

Markedness and Subject Choice in Optimality Theory

77

Tableau T3.6 Lummi (local-person agent and patient) V (Agt/1/x, Pat/2/x)

*Obl/ Local

f ACTIVE Agt/Su/1/x±Pat/Oj/2/x PASSIVE Pat/Su/2/x±Agt/Obl/1/x

*Oj/ Local

*Su/x

*

*

*!

*Su/ Pat

*

*GR/ Pers *

*

*

Table 3.9 Distribution of voice by person in Squamish Agt + 1 2 3

Pat )

1

2

3

Ð active/*passive active/passive

active/*passive Ð *active/passive

active/*passive active/*passive active/passive

active (see tableau T3.2). In the absence of a prominent argument, *Su/Pat once again plays the key role, excluding the passive (see tableau T3.3). Given the adjacent ranking of *Oj/Local and *Obl/Local, one might consider combining them as *Non-Su/Local. But in fact *Obl/Local must outrank *Oj/Local to correctly determine voice in clauses where both agent and patient are local persons. In this case, only the active is possible, despite the fact that it violates *Oj/Local. Tableau T3.6 shows that this follows if the passive is excluded by a higher-ranked constraint, *Obl/Local. We have so far seen no reason to distinguish between ®rst and second person, but there are languages in which constraints on the association of person with grammatical function treat the two di¨erently. Squamish is such a language. As shown in table 3.9, active clauses with third-person subject are excluded if the object is second person. There is no analogous constraint if the object is ®rst person.14 Clauses with third-person agent/second-person patient are expressed then in the passive, while clauses with third-person agent/®rst-person patient allow either active or passive. Encapsulation of ®rst- and second-person constraints must be abandoned here, since the two cases are treated di¨erently. The situation can be described by ranking *Oj/2 above *Su/Pat and *Oj/1 below, as in (10). (10)

Squamish *Obl/Local g *Oj/2 g *Su/x g *Su/Pat g *Oj/1

78

Judith Aissen

Tableau T3.7 Squamish (second-person patient) V (Agt/3/x, Pat/2/x)

*Obl/ Local

ACTIVE Agt/Su/3/x±Pat/Oj/2/x

*Oj/2

*Su/x

*!

*

f PASSIVE Pat/Su/2/x±Agt/Obl/3/x

*Su/ Pat

*

*

*Su/x

*Su/ Pat

*Oj/1

Tableau T3.8 Squamish (®rst-person patient) V (Agt/3/x, Pat/1/x) f ACTIVE Agt/Su/3/x±Pat/Oj/1/x PASSIVE Pat/Su/1/x±Agt/Obl/3/x

*Obl/ Local

*Oj/2

* *

*Oj/1

* *!

The high ranking of *Oj/2 excludes active clauses with third-person subject, secondperson object. The passive then emerges as the only choice, even though it violates *Su/Pat (see tableau T3.7). In the case of a ®rst-person patient, however, there are no constraints in active clauses: as in Lushootseed, both active and passive are possible when agent is third person. In the Jelinek and Demers account, ®rst person is simply kept o¨ the hierarchy, but we cannot make the analogous move here, which would be to absent the constraint *Oj/1 from the grammar of Squamish. But we can prevent it from forcing passive by ranking it below *Su/Pat, per (10). The e¨ects of this ranking are shown in tableau T3.8. The ranking of *Oj/1 below *Su/Pat means that the passive candidate is excluded by a higher-ranked constraint than the active, and the active therefore surfaces. (Passive with a ®rst-person patient will be possible, however, if the patient is thematically prominent. Then *Su/x excludes the active, leaving the passive.) Di¨erences in the realization of clauses with ®rst- and second-person patient are derived in this account through di¨erent rankings with respect to *Su/Pat, which plays a pivotal role. Since *Su/Pat penalizes any surviving passive candidate, passive can emerge as optimal only when the competing active is penalized by a higher-ranked constraint. The higher-ranked *Oj/2 does exactly this in the case of

Markedness and Subject Choice in Optimality Theory

79

second-person patients. The ranking of *Oj/1 below *Su/Pat means that *Su/Pat plays the decisive role in clauses with ®rst-person patients and determines active voice. 3.4.4 Summary The preceding analyses illustrate how harmonic alignment, applied to a small number of universal scales, can yield a set of constraint subhierarchies each of which concerns the association of grammatical function with some substantive dimension (e.g., person, semantic role, thematic prominence). The harmonic alignments shown in tables 3.4 to 3.6 characterize universal preferences in these associations; the constraint subhierarchies enforce the observed preferences through universal ranking of constraints. The possibility of interpolating (reranking) constraints from distinct subhierarchies makes it possible to account for the di¨erent ways languages prioritize these preferences. The rankings illustrated by the languages discussed above are collected in (11). (11) Fox/Nocte: English: Lushootseed: Lummi: Squamish:

*Su/Pat g


g {*GR/Pers, *GR/w} *Su/x g *Su/Pat g *GR/Pers *Obl/Local g *Su/x g *Su/Pat g *GR/Pers *Obl/Local g *Oj/Local g *Su/x g *Su/Pat g *GR/Pers *Obl/Local g *Oj/2 g *Su/x g *Su/Pat g *Oj/1 g *GR/Pers

The grammars of these languages show increasing sensitivity of voice to person, re¯ected in the increasing domination of  Su/Pat by  GR/Pers constraints. This approach to the person hierarchy addresses the issues raised earlier. First, it draws on a set of universally ranked constraints (subhierarchies) whose ranking expresses the basic markedness relations that hold in this domain. Individual languages cannot change rankings that are universally ®xed, but by interpolating constraints from distinct subhierarchies in language-particular ways, grammars can determine whether person plays any role at all in voice determination, and if it does, which elements play a role. Di¨erences between languages lie then in the way the person constraints are ranked with respect to constraints that deal with the association of grammatical function with other substantive dimensions, like thematic prominence and semantic role. This account also resolves the descriptive problem noted earlier: the fact that in general it is not possible to devise a single hierarchy that works for both active and passive clauses. This stands as an obstacle to formally expressing the fact that constraints in both clause types re¯ect the same basic markedness relations. The constraints on the association of grammatical function/person that are derived here through harmonic alignment of prominence scales are very speci®c and simple: each penalizes a particular association of person with grammatical function. Because

80

Judith Aissen

of the particular grammatical functions that they reference (e.g., object, oblique (ˆ passive agent)), some constraints penalize only active clauses and some penalize only passives. The reanalysis of language-particular hierarchies as a set of simple, ranked constraints thereby provides the speci®city and ¯exibility needed to handle constraints that apply only to one clause type or the other. At the same time, because language-particular rankings of these constraints are constrained by those of the universal subhierarchies from which they are drawn, they re¯ect the universal markedness expressed by those subhierarchies. 3.5

Local Conjunction with *0/

In the absence of interference from other constraints, those that penalize marked associations of person and grammatical function will lead to grammars in which unmarked associations are optimized. Of the languages discussed above, Lummi comes the closest to realizing this situation: active clauses never contain third-person subjects with local-person objects, and passive clauses never contain third-person subjects with local-person agents. The other languages realize this situation to varying degrees. On the other extreme are languages like Fox, Nocte, and Dyirbal in which marked associations routinely surface because the unmarked con®guration is excluded by a higher-ranked constraint,  Su/Pat. But it is not the case that these languages are indi¨erent to the markedness of particular associations of person and grammatical function. Indeed, as seen earlier, one of the striking characteristics of these languages is that they do register the markedness of particular associations either through case marking or a direction system. It makes sense to try to describe such languages using the constraints derived above, since these constraints already express the relative markedness of various person/grammatical function associations. Consider the evaluation of a clause with third-person subject/®rst-person object in a language like Nocte in which all clauses are active (tableau T3.9 is a more speci®c version of tableau T3.1). The active involves a marked con®guration, as shown by violations of high-ranking  Su/3 and  Oj/1, but the con®guration is tolerated because these constraints are dominated by  Su/Pat. Nonetheless, the markedness of the con®guration (third-person subject, ®rst-person object) is expressed in Nocte, which requires the marked verbal categoryÐthe inverse. The question then is how to use constraints like  Su/3 and  Oj/1 in the description of these clauses, despite their apparent irrelevance to the evaluation in tableau T3.9. The association of marked con®guration with marked expression is a fundamental idea in markedness theory (Jakobson [1939] 1984; Greenberg 1966). Hence the question of how this property might be expressed formally is important. I pursue here a suggestion by Paul Smolensky (personal communication), that it might be expressed through local conjunction of a constraint hierarchy with a constraint that penalizes

Markedness and Subject Choice in Optimality Theory

81

Tableau T3.9 Fox V (Agt/3/x, Pat/1/x)

*Su/Pat

f ACTIVE Agt/Su/3/x±Pat/Oj/1/x PASSIVE Pat/Su/1/x±Agt/Obl/3/x

*Su/3

*Oj/1

*

*

*!

zero exponence. Local conjunction is a formal operation that conjoins existing constraints (Smolensky 1995). (12) The local conjunction of C1 and C2 in domain D, C1 & C2 , is violated when there is some domain of type D in which both C1 and C2 are violated. The intuition behind constraint conjunction is that ``two constraint violations are worse when they occur in the same location'' (Smolensky 1995). This idea is formalized by imposing a universally ®xed ranking of C1 & C2 with respect to each of its component constraints (Smolensky 1995). (13) Universally, C1 & C2 dominates C1 , C2 . In the present case, the idea is that the presence of a marked category is bad, but zero expression of a marked category is worse. The analysis involves two general constraints beyond those already discussed: *j (read: ``star zero'') and *struc (read: ``star structure''). The constraint *j can be understood as expressing the idea that morphological categories should be expressed. This is in tension with another principle, *struc, which penalizes linguistic structure and is understood broadly to exclude features, nodes, and so on (Prince and Smolensky 1993:25). Both *j and *struc are very general constraints and will be subscripted here to make speci®c refererence to particular categories. We want to express the idea that it is more harmonic for a marked association than an unmarked association to be morphologically marked. Subhierarchies that express the relative markedness of person/GR associations already exist (tables 3.5 and 3.6). So does *j, which penalizes absence of morphological expression. Conjunction of *j with the subhierarchies of tables 3.5 and 3.6 yields subhierarchies expressing exactly the idea described above, as long as the ranking of each resulting subhierarchy is predetermined by that of the input subhierarchy. I assume (14) then. (14) The local conjunction of C1 with subhierarchy [C2 g C3 g


g Cn ] yields the subhierarchy [C1 & C2 g C1 & C3 g


C1 & Cn ].

82

Judith Aissen

Table 3.10 Conjunction of *j with subhierarchies from table 3.5 Subhierarchies involving alignment of grammatical function and person

Conjunction of subhierarchies with *j

*Su/3 g *Su/2 *Su/3 g *Su/1

*j & *Su/3 g *j & *Su/2 *j & *Su/3 g *j & *Su/1

*Oj/2 g *Oj/3 *Oj/1 g *Oj/3

*j & *Oj/2 g *j & *Oj/3 *j & *Oj/1 g *j & *Oj/3

The subhierarchies that locally conjoin with *j are listed in the left column of table 3.10; the subhierarchies that result from local conjunction with *j are listed in the right column. Per (14), the rankings that characterize the subhierarchies on the right are all predetermined by those of the subhierarchies on the left. The subhierarchies on the right express the fact that it is more marked for a third-person subject to be associated with zero exponence than for a ®rst- or second-person subject; and inversely, more marked for a ®rst- or second-person object than a third-person object to be associated with zero exponence. Since the ranking of the locally conjoined constraints (right column) is predetermined by that of the subhierarchies from which they are derived, the derived constraints express directly the basic idea that marked con®gurations are more likely to be morphologically marked than unmarked ones. These rankings thus represent the intuition that Silverstein (1976) is at pains to develop. To see how these constraints might ®gure in actual grammar construction, I consider case marking in languages like Dyirbal (sections 3.5.1 and 3.5.2). I discuss brie¯y in section 3.5.3 how they might be used to describe the direction system of a language like Nocte. 3.5.1 Split Ergativity15 The literature on case marking in the Australian languages (Silverstein 1976; Blake 1977; Dixon 1994) establishes that it is common for third persons to be overtly case marked when they function as subjects of transitive clauses, but to be unmarked as objects or as subjects of intransitive clauses; on the other hand, it is common for ®rst and second persons to be overtly case marked when they function as objects, but not as subjects. The implicational universals in (15) appear to hold. (15) a. If 3rd-person objects are case marked, then so are local-person objects. b. If local person subjects are case marked, then so are 3rd-person subjects. These generalizations follow from the universal rankings established in table 3.10 and encapsulated in (16), if *j is understood to penalize the absence of morphological case (*j is subscripted with C to represent this).

Markedness and Subject Choice in Optimality Theory

83

Table 3.11 Case marking in Dyirbal Marked

Local subject±3rd object Local subject±local object 3rd subject±3rd object 3rd subject±local object

O

S S

Abs-Erg O O

O

ƒ!

Acc-Nom

!

!

ƒ!

Unmarked

S S

(16) a. *jc & *Su/3 g *jc & *Su/Local b. *jc & *Oj/Local g *jc & *Oj/3 Some domain must be speci®ed for local conjunction (see (12)). For case marking, the relevant domain is a nominal projection. Then, the ranking in (16a) asserts that it is more marked for a third-person subject than for a local-person subject not to be case marked. Therefore, if local-person subjects are case marked, third persons must also be (ˆ (15a)). Conversely, it is more marked for a local-person object than for a third-person object not to be case marked. Hence, if third persons are case marked, local-person objects must also be (ˆ (15b)). Dyirbal instantiates this system in a particularly clean way. Four case-marking ``patterns'' result from the independent case marking of subject and object. These are shown in table 3.11 in order of increasing markedness (compare the chart in Silverstein 1976:151). The maximally unmarked clause (top line) is one in which the subject is a local person, and the object third person. In this case, neither argument is case marked. The maximally marked clause (bottom line) is one in which the subject is third person and the object is a local person; both arguments are case marked in such clauses. The other two combinations (middle lines) are intermediate in markedness, with one argument case marked and the other not. Dyirbal seems then to provide clear evidence for the existence of the subhierarchies in (16a) and (16b). However, in the absence of any other constraints, the rankings in (16) will penalize zero marking on all subjects and on all objects, forcing overt case marking in all instances. This is not what is found, so there must be some other constraint involved. Following a suggestion of Paul Smolensky (personal communication), let us suppose that the relevant constraint is *struc. By penalizing structure (here, case speci®cation), *struc delimits the point at which zero exponence is preferred by the grammar over morphological expression. Ranking it between the two constraints in each of (16a) and (16b), for example, yields a system in which marked

84

Judith Aissen

associations are overtly marked and unmarked ones are notÐexactly the Dyirbal system. The overall ranking of constraints for Dyirbal is given then in (17), with *jc penalizing the absence of case marking and *strucc penalizing its presence. The ranking between the top constraints is not ®xed; likewise for the bottom two constraints. (17) {*jc & *Su/3, *jc & *Oj/Local} g *strucc g {*jc & * Su/Local, *jc & Oj/3} To see how this works, we consider several evaluations. As before, each candidate is a feature structure that speci®es (inter alia) values for person, grammatical relation, and morphological case. I assume that nominals are optionally speci®ed for case. Speci®cation for case means that case is morphologically expressed; absence of such speci®cation means that case is not morphologically expressedÐthat is, it is j. There are four candidates, arrived at by associating each nominal argument independently with the presence or absence of case. The domain of local conjunction is some node within the nominal projection, so the conjoined constraints are evaluated internal to each nominal constituent. Thematic prominence appears to play no role in Dyirbal case marking and is omitted from input and candidates. Consider ®rst the maximally unmarked clause type, one with a ®rst-person subject, third-person object (tableau T3.10). In this case, the winner should be the candidate in which neither argument is case marked. Neither of the top two constraints is applicable since no candidate has a third-person subject or a local-person object. Since no candidate violates *Su/3 or *Oj/Loc, no candidate can violate a local conjunction containing one of these constraints (per (12)). The next constraint is *strucc , which is violated once by the candidates that specify case for one argument, and twice by the one that speci®es it for both. Thus, the candidate with no case marking wins. The last two Tableau T3.10 Dyirbal (®rst-person subject/third-person object) V (Agt/1, Pat/3)

*jc & *Su/3

*jc & *Oj/Loc

*strucc

Agt/Su/1/case Pat/Oj/3

*!

Agt/Su/1 Pat/Oj/3/case

*!

f Agt/Su/1±Pat/Oj/3 Agt/Su/1/case Pat/Oj/3/case

*jc & *Su/Loc

* * *

*!*

*jc & Oj/3

*

Markedness and Subject Choice in Optimality Theory

85

constraints penalize zero marking in relatively unmarked con®gurations and are irrelevant here because they are ranked below *strucc . The ranking in (17) gives the right result here. Consider now the maximally marked clause type: one with a third-person subject and ®rst-person object (tableau T3.11). Both subject and object are marked and each requires morphological expression (i.e., overt case). Only one candidate in fact provides it. The other three each violate one or both of the top two constraints that penalize j marking in clauses with marked associations. The two intermediate cases involve clauses in which both arguments are third person, or both are local persons. These are ``mixed'' types in that one argument involves an unmarked association of person and grammatical function, the other a marked association. Accordingly, one argument is case marked and the other is not (see table 3.11). The ranking of constraints proposed in (17) treats these cases correctly.Veri®cation is left to the reader. The account just sketched presents a number of problems, which will not be solved here but should be noted. The ®rst is that the discussion above assumes a universal notion of ``subject,'' which is associated in Dyirbal transitive clauses with the agent. Dixon has argued that subject is not a universal notion and, in particular, that it is not useful in Dyirbal. In its place, he posits two types of syntactic relations. One is the relation ``syntactic pivot,'' which is associated with the nominal that functions as the ``pivot'' of a cross-clausal construction (Dixon 1979, 1994; Foley and Van Valin 1987). The other includes the syntactic relations A and S, which correspond to the agent argument of a transitive clause and the sole argument of an intransitive clause. The syntactic pivot in Dyirbal transitive clauses is consistently the patient, not the Tableau T3.11 Dyirbal (third-person subject/®rst-person object) V (Agt/3, Pat/1)

*jc & *Su/3

Agt/Su/3/case Pat/Oj/1 Agt/Su/3 Pat/Oj/1/case

*!

Agt/Su/3±Pat/Oj/1

*!

f Agt/Su/3/case Pat/Oj/1/case

*jc & *Oj/Loc

*strucc

*!

* *

* **

*jc & *Su/1

*jc & Oj/3

86

Judith Aissen

agent. Hence, if the pivot is to be equated with subject (in my sense), the analysis proposed here is not correct, since it identi®es the subject with the agent. Harmonic alignment of person with semantic role would, however, give the right results, and would be more in line with the analyses of Dixon and Silverstein. A possibly related problem is that this analysis makes incorrect claims about intransitive clauses in Dyirbal, where the subject is never marked. The constraint [*jc & *Su/3] is evaluated internal to each nominal and has no access to information outside the nominal. But the presence or absence of an object in the clause is crucial in determining case marking on third-person subjects, since they are case marked only in the presence of an objectÐthat is, only in transitive clauses. Since the subhierarchy in (16a) does not distinguish transitive from intransitive clauses, it leads to outputs in which overt case marking is forced for all third-person subjects. This is not insoluble, since it is possible to devise more complex constraints that draw the correct distinction between transitive and intransitive subjects. (For example, local conjunction of [*Su/3 & *jc ] with *Oj/Pers (where Pers ranges over the three persons) derives a constraint that penalizes a subject without case marking only in a clause containing an object; note that the outer conjunction must take the clause as its domain.) What is at issue, though, seems to be the functional motivation for case marking, which is generally seen as distinguishing subject from object. But the need to distinguish subject from object arises only in transitive clauses. The fact that the constraints in (16a) are only relevant then in transitive clauses seems reasonable, but the present analysis does not restrict them to transitive clauses. What appears to be needed here is a more systemic view of case marking, one that can formally express the functional motivation for case marking. A ®nal problem concerns the availability of both *j and *struc for local conjunction. The functional motivation of locally conjoining *j with subhierarchies like those of table 3.5 that express the relative markedness of particular associations of person and grammatical function is clear: the more marked a nominal is qua subject/ object, the more useful it is to overtly mark its grammatical function. From the functional perspective, it is pointless to locally conjoin the same subhierarchies with *struc, since the result would favor overt marking where it is least needed, and penalize it where it is most needed. However, formally, there is nothing in the present system that prevents local conjunction of the subhierarchies in table 3.5 with *struc, and this would yield a set of ranked constraints that could entirely neutralize the predictions derived above. 3.5.2 Split Ergativity and Constraint Reranking The Dyirbal pattern is common among the Australian languages, but within the limits sketched above (see (15)), there is a great deal of language-particular variation

Markedness and Subject Choice in Optimality Theory

87

concerning which elements are marked according to a nominative-accusative system and which according to an ergative-absolutive system. Some of this variation has to do with features that we have not discussed explicitly here (e.g., animate vs. inanimate, proper vs. common nouns, singular vs. plural); hopefully techniques like those used here will extend to such cases and it would be interesting to see whether they do. But some of the variation has to do with features that do ®gure in the constraints proposed here. We show brie¯y here that some of these cases can be easily described by constraint reranking. According to Blake (1977), Aranda (Pama-Nyungan group) has an ergativeabsolutive system for third persons and for ®rst-person singular (i.e., case marked when transitive subject), while second-person singular is marked according to a nominative-accusative system (not case marked as subject).16 The distribution of case in Aranda parallels the distribution of voice in Salish, with the ®rst and second person distinguished, and the ®rst person treated on a par with third person. Case marking of subjects in Aranda can be described by the ranking in (18), where [*jc & *Su/1] dominates *strucc , which in turn dominates [*jc & *Su/2] (compare (18) with the ranking in (17)).17 (18) *jc & *Su/3 g *jc & *Su/1 g *strucc g *jc & *Su/2 This ranking penalizes a ®rst-person subject without case marking, forcing overt (i.e., ergative) case;18 on the other hand, it penalizes a second-person subject with case marking, forcing zero exponence (i.e., nominative case). This case shows how the incursion of ergative case marking into the local persons is modeled by the reranking of *Su/Pers constraints over *struc. The extreme case will be one in which all subject constraints are so ranked (in (19) and (20), X ranges over any of the dimensions aligned with grammatical function). (19) *jc & *Su/X g *strucc g *jc & *Oj/X The result will be a language with ergative (not split-ergative) case marking: all transitive subjects are marked for case, but no objects. Examples are apparently found in Australia, as well as in other parts of the world. Blake (1977) classi®es Yalarnga and Rembarnga as instances. Incursion of the accusative system into the third person is also found in Australia, often with various subclasses of third person treated di¨erently (Silverstein 1976; Blake 1977). A simple case where apparently all third persons (as well as all local persons) are case marked according to an accusative system is Ngarluma (Blake 1977). In this case, the ranking would be that of (20). (20) *jc & *Oj/X g *strucc g *jc & *Su/X In languages with this ranking, all objects but no subjects will carry case marking.

88

Judith Aissen

Many case-marking systems of Europe approximate this system, though j marking of inanimate (neuter) objects is common. 3.5.3 Direction As mentioned earlier, Silverstein (1976) and DeLancey (1981) both elaborate the idea that (person-based) direction systems are based on the same markedness relations as (person-based) split-ergative case-marking systems but express this markedness through head marking rather than dependent marking. Here I consider brie¯y how the constraints developed so far for a language like Dyirbal might apply to a language like Nocte. One basic di¨erence is that markedness is expressed on the predicate, not on the nominal dependents themselves. This suggests ®rst of all that the domain for evaluation of the conjoined constraint set (*j & *GR/Pers) must be some projection of V rather than N and, second, that features of the verb's nominal arguments must be present in the V projectionÐfor example, via abstract agreement. Further, the category involved is not morphological case, but direction. Hence *j must penalize zero expression of direction, not case. The constraint *jd (``express direction'') will be satisifed when the category of direction is overtly marked and violated when it is not. In Nocte, the inverse verb carries a morphological mark and so satis®es *jd ; the direct verb carries no morphological mark and thus violates *jd . Correspondingly, *strucd penalizes expression of direction and is satis®ed in Nocte by the direct verb and violated by the inverse. There is another di¨erence, namely, the distribution of direct and inverse in Nocte requires simultaneous reference to properties of both subject and object. In Nocte, the direct form is used in the following subject-object combinations: 1-2, 1-3, 2-3, 3-3, and the inverse is used elsewhere (2-1, 3-1, 3-2). That is, the inverse is used in the most marked clause types, those with third-person subject and local-person object, as well as in clauses with second-person subject and ®rst-person object.19 There is a natural way to model this situation in the present framework, namely, through local conjunction of the subject and object subhierarchies established in section 3.4. Local conjunction automatically yields a ranking of the constraints referencing all combinations of subject and object by person.20 The ranking in table 3.12 (right-hand side) corresponds directly to the markedness registered in table 3.11 and thus correctly expresses the relative markedness of all combinations of subject and object by person: the most marked clause is one with a third-person subject/local-person object; the least marked is one with a local-person subject/third-person object; and other combinations are unranked with respect to one another. The ranking that characterizes the subhierarchy on the right-hand side of table 3.12 is completely determined by constraint conjunction (see note 20), and is (proposed as) universal. There should be no language in which a clause with a local-person subject/third-person object is

Markedness and Subject Choice in Optimality Theory

89

Table 3.12 Local conjunction of GR/person subhierarchies Subhierarchies

Conjunction of subhierarchies

*Su/3 g *Su/Local

*Su/3 & *Oj/Local g {*Su/3 & *Oj/3, *Su/Local & *Oj/Local} g *Su/Local & *Oj/3

*Oj/Local g *Oj/3

treated as more marked than one with a third-person subject/third-person object, and so on. Grammars of particular languages may di¨er, then, in how the constraints in table 3.12 are ranked with respect to constraints like *j or *struc, but they will not di¨er in the ranking of the subhierarchy. Active in the grammar of Nocte, therefore, are the basic constraints on the left side of table 3.12, the conjoined constraints on the right side, and a further set of constraints formed by local conjunction of the latter with *jd . This local conjunction yields a further subhierarchy whose ranking is strictly determined by that of the input subhierarchy (again, per (14)). (21) *jd & *Su/3 & *Oj/Loc g {*jd & *Su/3 & *Oj/3, *jd & *Su/Loc & *Oj/ Loc} g *jd & *Su/Loc & *Oj/3 The distribution of direct and inverse can be characterized by interpolating *strucd among the constraints in (21). This predicts that if direction is overtly marked for relatively unmarked con®gurations, it must be overtly marked for more marked ones. For a language like Nocte, where direct verbs have no morphological mark and inverse verbs do, the ranking in (22) describes the distribution of direct and inverse verbs.21 (22) Nocte {*jd & *Su/3 & Oj/Loc, *jd & *Su/2 & Oj/1} g *strucd g *jd & *GR/Pers This ranking penalizes the direct verb (j) whenever the subject is third person and the object is a local person, and when the subject is second person and the object ®rst; it penalizes the inverse verb in all other cases. Crucially, no reranking of the subhierarchy in (21) is involved in Nocte, simply interpolation of *strucd . 3.6

Conclusion

Let us step back and review how the constraints employed in the analysis presented here and their rankings were derived. Aside from *j and *struc, all the constraints and most of their rankings are derived through two operations, alignment and local conjunction, operating on two proposed universal scales (Local > 3; Su > Non-Su)

90

Judith Aissen

Table 3.13 Constraint subhierarchies summarized a. *Su/3 g *Su/Local *Oj/Local g *Oj/3 b. *j & *Su/3 g *j & *Su/Local *j & *Oj/Local g *j & *Oj/3 c. *Su/3 & *Oj/Local g {*Su/3 & *Oj/3, *Su/Local & *Oj/Local} g *Su/Local & *Oj/3 d. *j & [*Su/3 & *Oj/Local] g {*j & [*Su/3 & *Oj/3], *j & [*Su/Local & *Oj/Local]} g *j & [*Su/Local & *Oj/3]

and decompositions of those scales. Alignment generates the constraints listed in the (a) cell of table 3.13. These function by themselves in person-driven voice systems. The remaining constraints are generated by local conjunction. Those in (b) involve local conjunction of *j with the constraints in (a), those in (c) involve local conjunction of the constraints in (a) with one another, and those in (d) involve local conjunction of *j with the constraints in (c). The intuitions behind these applications of local conjunction are sound: conjunction with *j expresses the basic idea that morphological complexity is associated with marked con®gurations; local conjunction of the subject/person and object/person subhierarchies with one another expresses the relative markedness of clauses depending on person of both subject and object. The (b) constraints play a key role in languages with person-driven case and direction systems, where reference to a single nominal argument is su½cient. With the caveat discussed at the end of section 3.5.1, Dyirbal is such a language. Those of (d) are relevant in languages like Nocte where case marking or direction marking requires global reference to properties of both subject and object. The constraints in (c) are not by themselves crucial in any of the analyses discussed above, but would be relevant in a language where voice was determined jointly by subject and objectÐfor example, in a language where voice was driven by a hierarchy of the form 1 > 2 > 3. Of course, their existence is required if the constraints in (d) are to be derived through local conjunction. It is striking that starting with so little (a small number of prominence scales, harmonic alignment, and local conjunction), it is possible to construct grammars that not only succeed in characterizing the facts of individual languages but that do so within a universal theory of markedness. The analyses developed here depend crucially on harmonic alignment and local conjunction and thereby endorse a role for these operations in constraint generation. Further, the key role of harmonic alignment and local conjunction highlights certain structural parallels between phonology and syntaxÐparallels that have informed markedness theory but that have remained largely unexpressed in generative approaches to grammar.

Markedness and Subject Choice in Optimality Theory

91

Notes This is a slightly edited version of a paper that appeared in Natural Language and Linguistic Theory (Aissen 1999). Work on the paper was begun while the author was a fellow at the Center for Advanced Study in the Behavioral Sciences, 1996±1997. Financial support provided by National Science Foundation Grants aSBR-9022192 and aSBR-9818177, by the University of California O½ce of the President, and by the Academic Senate of the University of California at Santa Cruz is gratefully acknowledged. Thanks also to Alan Prince and Paul Smolensky, whose suggestions shaped the basic analysis presented here, and to Joan Bresnan, Donka Farkas, Jane Grimshaw, Eloise Jelinek, Kazutaka Kurisu, and several anonymous reviewers, who made detailed comments on earlier versions of the manuscript. This chapter is an expanded version of the paper ``Person and Voice,'' read at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest in Baltimore in May 1997. Questions raised by participants at the workshop and by members of Joan Bresnan's syntax seminar at Stanford, winter 1999, have also been very helpful. 1. I am ignoring here some aspects of the analysis of Dyirbal that may render this discussion problematic; see section 3.5.1. 2. This system is termed split-ergative because third persons are marked according to an ergative/absolutive pattern: transitive subjects are overtly marked, while intransitive subjects and transitive objects are unmarked. The marked case is called ergative, the unmarked case absolutive. On the other hand, the local persons are marked according to a nominative/ accusative pattern: subjects of all clauses are unmarked, while objects are marked. In this alignment, the marked case is called accusative, and the unmarked case nominative. 3. DeLancey (1981) suggests the relevance of Silverstein's hierarchy to voice, citing the marginality of local-person passive agents in English as evidence. The Coast Salish languages discussed here are a clearer case, if the Jelinek and Demers 1983 analysis in terms of voice is correct. Klaiman (1991) suggests that Lummi has a direction system (see below) rather than a voice system. The question of how Lummi should be analyzed is important, but as far as I can tell, neither Klaiman nor Jelinek and Demers present decisive evidence (for some discussion of the issue, see the Jelinek and Demers paper as well as Jelinek 1993). I assume the Jelinek and Demers analysis here. 4. Eloise Jelinek reminds me that a clari®cation is needed here. First- and second-person agents can be identi®ed in Lummi passive sentences by use of oblique (third-person) nominal adjuncts that are built on what Jelinek terms person-deitic roots (Jelinek 2000). These adjuncts allow reference to ®rst and second person, but they are syntactically third person, corresponding roughly to `the one who is me/you'. These are nominals, not pronouns, and are used for contrastive focus of ®rst and second person. 5. Aissen (1997) suggests that Chamorro also ranks third-person nominals in obviation status. This subsumes the ranking of animates indicated in the hierarchy in (3b). 6. On the relational scale, see for example Jakobson [1965] 1995; Keenan and Comrie 1977; Bell 1983; Croft 1990; Bresnan 1994, 1998, forthcoming. On the person scale, see Silverstein 1976; DeLancey 1981. On the role scale, see Jackendo¨ 1972; Silverstein 1976; Bresnan and Kanerva 1989; Grimshaw 1990. On the obviation scale, see Wolfart 1973. 7. DeLancey (1981) argues that ®rst and second persons are not universally ranked. This allows languages to di¨er in their ranking. Dixon (1994: 88¨.) argues that ®rst person usually

92

Judith Aissen

outranks second, but concedes the need to allow the opposite ranking in some cases. Several of the languages discussed here show the need for second person to outrank ®rst. 8. The following abbreviations are used in tables and tableaux: Su: subject, Non-Su: nonsubject; Oj: object; Obl: passive agent (syntactic function); Agt: agent (semantic role); Pat: patient; GR: grammatical relation; Pers: person. 9. In assuming that information about the discourse status of arguments may be speci®ed in inputs, I follow Legendre, Raymond, and Smolensky (1993), Grimshaw and Samek-Lodovici (1998), and Sells (chap. 12, this volume). Sells assumes a more articulated input, where three di¨erent degrees of prominence can be indicated in the input. 10. Languages in which agents are realized as object are not excluded by the present setup, but they will only emerge if there are higher-ranked constraints that force the appearance of object agents. A number of analyses positing agent objects have been proposed (see Rhodes 1990, 1994 on Ojibwe (Algonquian) and Kroeger 1993 on Tagalog; see also Manning 1996 for a more general discussion and Sells (chap. 12, this volume) for an OT analysis in which there is no asymmetry between agent and patient in access to the grammatical functions subject and object). 11. These two factors correspond fairly closely to the two topicality measures proposed in GivoÂn 1983, and developed further in work like Cooreman 1987 and the papers in GivoÂn 1994: referential distance (how far back is the most recent mention?) and topic persistence (how many times is the referent mentioned in the next x lines?). 12. Tomlin's conclusions were based on a study of texts taken from hockey play-by-play descriptions; Thompson's study included texts from a variety of genres: formal written English, less formal written English, informal spoken English. 13. Kuno (1976, 1987) and Kuno and Kaburaki (1977) argue that person does play a direct role in English voice, and in particular that passive is avoided when the agent is ®rst or second person. In their terms, there is a con¯ict between the speaker's natural empathy with ®rst or second person, and the fact that the focus of speaker empathy is associated with surface subject position. If the constraint against local-person passive agents (i.e., *Obl/Local) is active in English, it must be ranked below *Su/x, since examples with local-person passive agents are grammatical in the right discourse contexts (Kato 1979). This di¨erentiates English from the languages to be discussed below, where *Obl/Local is undominated. In the present account, person itself plays no direct role in determining voice, but it can play an indirect role via empathy, since empathy is one factor in choosing the center of attention. 14. An analogous situation is found in Chamorro, where active clauses with third-person subject/second-person object are generally excluded (Chung 1981; Cooreman 1987:96¨.; Chung 1998:34); similarly in K'iche' (Mayan), where active clauses with third-person subject/ second-person formal object are excluded (Mondloch 1981:35). In both Chamorro and K'iche', active clauses with third-person subject/®rst-person object are completely grammatical. 15. For a di¨erent approach to the split ergativity of Dyirbal within OT, see Woolford (chap: 16, this volume). There are three main di¨erences between her approach and mine. First, Woolford's approach makes no formal or systematic use of hierarchy alignment, though some of her principles re¯ect e¨ects derived here through harmonic alignment. Second, Woolford regards the case split of Dyirbal subjects as involving di¨erent morphological realization of the same abstract Case, but takes that of objects to involve di¨erent abstract Cases. Since her proposal deals only with abstract Case, it does not deal with the realization of subjects. Third,

Markedness and Subject Choice in Optimality Theory

93

Woolford's account is developed within the structural assumptions of the Minimalist Program, in which distinct abstract Cases correspond to distinct structural positions. Thus, Woolford assumes that morphologically marked objects occupy a di¨erent structural position than unmarked ones. 16. Apparently ®rst- and second-person plurals pattern like second-person singular. This does not follow from the constraint ranking proposed here. 17. Case marking of objects is like Dyirbal, so the object-referring constraints would be ranked with respect to *struc as in Dyirbal. 18. Again, it is necessary to restrict this constraint to transitive clauses, since case marking on ®rst persons in Aranda is limited to such clauses. 19. Most of the Nocte system can be described by reference solely to subject or object: all clauses with ®rst-person objects require inverse forms; all clauses with third-person objects require direct forms. But in clauses with third-person subjects or second-person objects, information about both subject and object is needed. 20. Several steps are involved in reaching the ranking on the right-hand side of table 3.12: Assume that the four constraints on the left-hand side are A g B and C g D. There are two ways to conjoin these two subhierarchies: (i) A conjoins with each of C and D; B conjoins with each of C and D. (ii) C conjoins with each of A and B; D conjoins with each of A and B. This yields: (i) A & C g A & D B & CgB & D (ii) C & A g C & B D & AgD & B Since conjunction is symmetric, this yields a partial ordering of the four constraints: (iii) A & C g {A & D, B & C} g B & D 21. The ranking in (22) does not, however, su½ce to describe a language like Fox, in which both direct and inverse verbs carry a morphological mark. References Ackema, Peter, and Ad Neeleman. 1998. Con¯ict Resolution in Passive Formation. Lingua 104: 13±29. Aissen, Judith. 1987. Tzotzil Clause Structure. Dordrecht: Reidel. Aissen, Judith. 1997. On the Syntax of Obviation. Language 73: 705±750. Aissen, Judith. 1999. Markedness and Subject Choice in Optimality Theory. Natural Language and Linguistic Theory 17: 673±711. Artstein, Ron. 1998. Hierarchies. Ms., Rutgers University, New Brunswick, N.J. Available at http://www.eden.rutgers.edu/@artstein/. Asudeh, Ash. 1999. Linking, Optimality, and Ambiguity in Marathi: An Optimality Theory Analysis. Ms., Stanford University, Stanford, Calif. Available at http://www.stanford.edu/ @asudeh/. Battistella, Edwin. 1990. Markedness: The Evaluative Superstructure of Language. Albany, N.Y.: SUNY Press.

94

Judith Aissen

Battistella, Edwin. 1996. The Logic of Markedness. Oxford: Oxford University Press. Bell, Sarah J. 1983. Advancements and Ascensions in Cebuano. In David M. Perlmutter (ed.), Studies in Relational Grammar, vol. 1, 143±218. Chicago: University of Chicago Press. Blake, Barry J. 1977. Case Marking in Australian Languages. Canberra: Australian Institute of Aboriginal Studies. Bresnan, Joan. 1994. Linear Order vs. Syntactic Rank: Evidence from Weak Crossover. In Katie Beals, Jeannette Denton, Bob Knippen, Lynette Melnar, Hisami Suzuki, and Erika Zeinfeld (eds.), Papers from the Thirtieth Regional Meeting of the Chicago Linguistic Society, 57±89. Chicago Linguistic Society, University of Chicago. Bresnan, Joan. 1998. Morphology Competes with Syntax: Explaining Typological Variation in Weak Crossover E¨ects. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 59±92. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Bresnan, Joan. Forthcoming. Lexical-Functional Syntax. Oxford: Blackwell. Ã a: A Case Study of Bresnan, Joan, and Jonni Kanerva. 1989. Locative Inversion in Chichew Factorization in Grammar. Linguistic Inquiry 20: 1±50. Chafe, Wallace. 1976. Givenness, Contrastiveness, De®niteness, Subjects, Topics, and Point of View. In Charles Li (ed.), Subject and Topic, 25±55. New York: Academic Press. Chung, Sandra. 1981. Transitivity and Surface Filters in Chamorro. In J. Hollyman and A. Pawley (eds.), Studies in Paci®c Languages and Cultures, in Honour of Bruce Biggs, 311±332. Auckland: Linguistic Society of New Zealand. Chung, Sandra. 1998. The Design of Agreement: Evidence from Chamorro. Chicago: University of Chicago Press. Cooreman, Ann. 1987. Transitivity and Discourse Continuity in Chamorro Narratives. Berlin: Mouton de Gruyter. Croft, William. 1990. Typology and Universals. Cambridge, England: Cambridge University Press. Dahlstrom, Amy. 1995. Morphology and Syntax of the Fox (Mesquakie) Language. Ms., University of Chicago. DeLancey, Scott. 1981. An Interpretation of Split Ergativity. Language 57: 626±657. Dixon, R. M. W. 1972. The Dyirbal Language of North Queensland. Cambridge, England: Cambridge University Press. Dixon, R. M. W. 1979. Ergativity. Language 55: 59±138. Dixon, R. M. W. 1994. Ergativity. Cambridge, England: Cambridge University Press. Dowty, David. 1991. Thematic Proto-Roles and Argument Selection. Language 67: 547±619. Foley, William, and Robert Van Valin. 1987. Functional Syntax and Universal Grammar. Cambridge, England: Cambridge University Press. GivoÂn, Talmy. 1976. Topic, Pronoun, and Grammatical Agreement. In Charles Li (ed.), Subject and Topic, 149±188. New York: Academic Press. GivoÂn, Talmy (ed.). 1983. Topic Continuity in Discourse. Amsterdam: Benjamins. GivoÂn, Talmy (ed.). 1994. Voice and Inversion. Amsterdam: Benjamins.

Markedness and Subject Choice in Optimality Theory

95

Greenberg, Joseph H. 1966. Language Universals, with Special Reference to Feature Hierarchies. The Hague: Mouton. Grimshaw, Jane. 1990. Argument Structure. Cambridge, Mass.: MIT Press. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 193±219. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Jackendo¨, Ray. 1972. Semantic Interpretation in Generative Grammar. Cambridge, Mass.: MIT Press. Jakobson, Roman. [1939] 1984. Zero Sign. In Linda Waugh and Morris Halle (eds.), Russian and Slavic Grammar: Studies, 1931±1981, 151±160. Berlin: Mouton. Jakobson, Roman. [1965] 1995. Quest for the Essence of Language. In Linda Waugh and Monique Monville-Burston (eds.), On Language: Roman Jakobson, 407±421. Cambridge, Mass.: Harvard University Press. Jelinek, Eloise. 1990. Grammatical Relations and Coindexing in Inverse Systems. In Katarzyna Dziwirek, Patrick Farrell, and Errapel MejõÂas-Bikandi (eds.), Grammatical Relations: A Cross-Theoretical Perspective, 227±246. Stanford, Calif: CSLI Publications. Jelinek, Eloise. 1993. Ergative ``Splits'' and Argument Type. In Jonathan Bobaljik and Colin Phillips (eds.), MIT Working Papers in Linguistics 18: Papers on Case and Agreement I, 15±42. MITWPL, Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Jelinek, Eloise. 2000. Predicate Raising in Lummi, Straits Salish. In Andrew Carnie and Eithne Guilfoyle (eds.), The Syntax of Verb-Initial Languages. Oxford: Oxford University Press. Jelinek, Eloise, and Richard Demers. 1983. The Agent Hierarchy and Voice in Some Coast Salish Languages. International Journal of American Linguistics 49: 167±185. Kato, Kazuo. 1979. Empathy and Passive Resistance. Linguistic Inquiry 10: 149±152. Keenan, Edward. 1976. Towards a Universal De®nition of ``Subject.'' In Charles Li (ed.), Subject and Topic, 303±333. New York: Academic Press. Keenan, Edward, and Bernard Comrie. 1977. Noun Phrase Accessibility and Universal Grammar. Linguistic Inquiry 8: 63±100. Klaiman, M. H. 1991. Grammatical Voice. Cambridge, England: Cambridge University Press. Klaiman, M. H. 1993. The Relationship of Inverse Voice and Head-Marking in Arizona Tewa and Other Tanoan Languages. Studies in Language 17: 343±370. Kroeger, Paul. 1993. Phrase Structure and Grammatical Relations in Tagalog. Stanford, Calif.: CSLI Publications. Kuno, Susumu. 1976. Subject, Theme, and the Speaker's EmpathyÐA Reexamination of Relativization Phenomena. In Charles Li (ed.), Subject and Topic, 417±444. New York: Academic Press. Kuno, Susumu. 1987. Functional Syntax. Chicago: University of Chicago Press. Kuno, Susumu, and Etsuko Kaburaki. 1977. Empathy and Syntax. Linguistic Inquiry 8: 627± 672. Legendre, GeÂraldine, William Raymond, and Paul Smolensky. 1993. An Optimality-Theoretic Typology of Case and Grammatical Voice Systems. Proceedings of the Nineteenth Annual

96

Judith Aissen

Meeting of the Berkeley Linguistics Society, 464±478. Berkeley Linguistics Society, University of California, Berkeley. LeSourd, Philip. 1976. Verb Agreement in Fox. In Judith Aissen and Jorge Hankamer (eds.), Harvard Studies in Syntax and Semantics 2, 445±528. Department of Linguistics, Harvard University, Cambridge, Mass. Manning, Christopher. 1996. Ergativity: Argument Structure and Grammatical Relations. Stanford, Calif.: CSLI Publications. Mondloch, James. 1981. Voice in Quiche-Maya. Doctoral dissertation, SUNY Albany, Albany, N.Y. Nichols, Johanna. 1986. Head-Marking and Dependent-Marking Grammar. Language 62: 56± 119. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Rhodes, Richard. 1990. Obviation, Inversion, and Topic Rank in Ojibwa. Proceedings of the Sixteenth Annual Meeting of the Berkeley Linguistics Society: Special Session on General Topics in American Indian Linguistics, 101±115. Berkeley Linguistics Society, University of California, Berkeley. Rhodes, Richard. 1994. Agency, Inversion, and Thematic Alignment in Ojibwe. Proceedings of the Twentieth Annual Meeting of the Berkeley Linguistics Society, 431±446. Berkeley Linguistics Society, University of California, Berkeley. Silverstein, Michael. 1976. Hierarchy of Features and Ergativity. In R. M. W. Dixon (ed.), Grammatical Categories in Australian Languages, 112±171. Canberra: Australian Institute of Aboriginal Studies. Slobin, Dan. 1994. Passives and Alternatives in Children's Narratives in English, Spanish, German, and Turkish. In Barbara Fox and Paul Hopper (eds.), Voice: Form and Function, 341±364. Amsterdam: Benjamins. Smolensky, Paul. 1995. On the Internal Structure of the Constraint Component Con of UG. Handout of talk at UCLA, Los Angeles, April 7. Smolensky, Paul. 1996a. On the Comprehension/Production Dilemma in Child Language. Linguistic Inquiry 27: 720±731. Smolensky, Paul. 1996b. The Initial State and ``Richness of the Base'' in Optimality Theory. Report JHU-CogSci-96-4. Department of Cognitive Science, Johns Hopkins University, Baltimore. Thompson, Sandra. 1987. The Passive in English: A Discourse Perspective. In Robert Channon and Linda Shockey (eds.), In Honor of Ilse Lehiste, 497±511. Dordrecht: Foris. Tomlin, Russell. 1985. Interaction of Subject, Theme, and Agent. In Jessica Wirth (ed.), Beyond the Sentence: Discourse and Sentential Form, 61±80. Ann Arbor, Mich.: Karoma. Wolfart, H. Christoph. 1973. Plains Cree: A Grammatical Study. Transactions of the American Philosophical Society n.s. 63, part 5.

Chapter 4 Optionality and Ine¨ability

Eric Bakovic and Edward Keer

English is well known for the optional presence of the complementizer that in declarative complements and relative clauses, as shown in (1a) and (1b). The presence of the complementizer here is truly optional, in that it triggers no discernible di¨erence in core meaning and/or discourse status. In addition to the optionality of the complementizer, object relative clauses in English are also generally optionally introduced by a wh-phrase, as shown in (1c); again, no discernible di¨erence in meaning and/or discourse status can be systematically gleaned. (1) Optionality in English a. I think [that the coat doesn't ®t him]. I think [the coat doesn't ®t him]. b. The coat [that he always wears t] doesn't ®t him. The coat [he always wears t] doesn't ®t him. c. The guy [who she dates t] doesn't respect her. The guy [she dates t] doesn't respect her. In this chapter we advance a theory of such (true) optionality1 within the framework of Optimality Theory (OT, Prince and Smolensky 1993; we assume the general OT approach to syntax of Grimshaw 1997b), using the cases of optionality exhibited in (1) as our core examples. In OT, candidate structural analyses of an underlying form are subject to evaluation by constraints against input-output disparity, or faithfulness constraints, as well as by constraints on structural outputs, or markedness constraints. The optimal realization of one input thus may or may not be the optimal realization of another; optimality is necessarily relativized. We propose that the locus of true optionality is in the underlying form or input. That is, the versions of the sentences in (1) with and without the complementizer or wh-phrase are derived from di¨erent inputs. However, because of the possibility of deletion or insertion of complementizers and wh-phrases, both versions of the sentence are possible output forms for each of the distinct inputs. We propose that there exist faithfulness constraints on distinctive formal properties of syntactic structures that disfavor deviation from the input in each case. If these faithfulness constraints

Eric Bakovic and Edward Keer

98

are ranked above markedness constraints against one or another of the possible outputs, the result is (apparent) optionality. As is generally assumed in OT, languageparticular ine¨ability results from the opposite ranking of faithfulness below markedness. This ranking forces avoidance of marked structure in the output through unfaithful input-output mappings (Prince and Smolensky 1993; see also Legendre et al. 1995; Legendre, Wilson, and Smolensky 1998). When the complementizer or wh-phrase is either obligatory or prohibited, it is because certain markedness constraints dominate the proposed faithfulness constraints. Under this schematic ranking, the markedness constraints choose the same output for both inputs. Optionality and ine¨ability thus arise as alternate e¨ects of the usual interaction between markedness and faithfulness constraintsÐan expected consequence of ranked and violable con¯icting constraints, the core assumption of OT. Moreover, since it is generally (and independently) assumed that OT constraints are universal and that their relative ranking is language particular, a necessary prediction of our theory is that what is optional in one language or context may be ungrammatical in another. This typological prediction, we argue, is con®rmed both within EnglishÐin contexts other than those shown in (1)Ðand in another language, Norwegian.2 There are at least two other possible approaches to optionality in OT, both claiming that the outputs in free variation arise from one and the same input. One approach is to ensure that no constraint distinguishes the outputs, so that if one emerges as optimal, the other(s) must also. In other words, the candidates fare equally or ``tie'' on all constraints, and under some rankings they tie for optimality. The second approach embellishes the basic theory with the notion of a constraint tie, with the same basic e¨ect: the outputs in free variation arise from the same input. Grimshaw (1997b) takes the former approach and Pesetsky (1998) the latter, in their respective analyses of the optionality of the English complementizer. Other work has shown these ``one-to-many'' input-output approaches to be empirically problematic; for instance, Legendre et al. (1995) argue based on extraction facts that candidates with and without the complementizer in English must arise from di¨erent inputs. Under the approach advocated here, these candidates do arise from di¨erent inputs, as necessary, with no problematic additions to the theory. The analysis laid out below is otherwise parallel to Grimshaw's in that it employs the same set of markedness constraints, thereby retaining the essential explanatory virtues of Grimshaw's overall system. 4.1

The Interaction of Markedness and Faithfulness

The optimality (1 grammaticality in OT) of a form in a given candidate output set is partially dependent on the input associated with that set: the optimal output of an input i1 may or may not be the optimal output of some other input i2 . Given only markedness constraints, this would of course be impossibleÐno matter what the

Optionality and Ine¨ability FgM inputs

outputs

99

MgF inputs

outputs

.................................. i1 ƒƒƒƒ! o1 i2 ƒƒƒƒ! o2 ... ƒƒƒƒ! ...

i1 i2 ...

fo1 o2 ...

Figure 4.1 The interaction of markedness and faithfulness

input, markedness constraints would battle it out among themselves and a unique, least marked form would invariably surface. Faithfulness constraints, penalizing disparity between input and output, have thus played a pivotal role in OT since its inception. Depending on the relative ranking of faithfulness and con¯icting markedness constraints, contrasting input speci®cations may or may not surface. If faithfulness dominates markedness (F g M), then F-dependent contrasts surface in the output. If, conversely, markedness dominates faithfulness (M g F), then F-dependent contrasts are neutralized in the output, in favor of the M-respecting end of the contrast spectrum. These two situations are depicted in ®gure 4.1. The speci®c proposal made here is that in addition to semantically discernible (lexical ) contrasts, there exist semantically inert ( functional ) contrasts governed by a set of faithfulness constraints. The notion of ``input'' that we assume is a modi®cation of that employed by Grimshaw (1997b:375±376): in addition to the lexical features, argument structure, tense and aspect of the Grimshavian input, we posit that there are functional features such as [Gcomp] and [Gwh].3 These functional features are described in more detail below. When functional contrasts are preserved in the output because F g M, the somewhat illusory e¨ect is optionality of forms in the context marked by M. Under the ranking M g F, what is predicted is the lack of optionality in the M-relevant context. This prediction is empirically supported within English, as demonstrated in section 4.2; since a faithfulness constraint can simultaneously dominate some markedness constraints and be dominated by others, this gives rise to optionality in some contexts but not in others. Comparative evidence between English and Norwegian indicates that this prediction is empirically supported crosslinguistically, where F g M in one grammar and M g F in another, as shown in section 4.3. 4.2

English

4.2.1 The Complementizer As noted above about (1a) and (1b), repeated below in (2), the complementizer that is often optional in English declarative complements and relative clauses. We follow

100

Eric Bakovic and Edward Keer

Doherty (1993) in assuming that the structural distinction between that-clauses and that-less clauses boils down to a distinction in verbal extended projection level, CP versus IP, and henceforth note it as such.4 (2) Complementizer optionality a. I think [CP that the coat doesn't ®t him]. I think [ IP the coat doesn't ®t him]. b. The coat [CP that he always wears t] doesn't ®t him. The coat [ IP he always wears t] doesn't ®t him. Sometimes the complementizer is obligatorily presentÐfor instance, when there is subject extraction from a relative (3) or when there is adjunction to a complement (4). (3) Complementizer obligatoriness in relatives with subject extraction a. The coat [CP that t doesn't ®t him] might ®t me. b. *The coat [ IP t doesn't ®t him] might ®t me. (4) Complementizer obligatoriness in complements with adjunction a. I think [CP that on him, no coat looks good t]. b. *I think [ IP on him, no coat looks good t]. In some other cases, the complementizer is obligatorily absentÐfor example, when there is subject extraction from a complement (5). This is known as the *that±trace e¨ect; see DeÂprez 1994, Keer 1996, and references therein. (5) Complementizer absence in complements with subject extraction a. *Which coat do you think [CP that t doesn't ®t]? b. Which coat do you think [ IP t doesn't ®t]? Suppose that the functional distinction between an embedded CP and an embedded IP is their speci®cation for a feature [comp]ÐCPs are speci®ed as [‡comp] and IPs are speci®ed as [ comp]. Suppose further that an embedded clause may be freely speci®ed in the input as [‡comp] or as [ comp]. To regulate the disparity between input and output in terms of the two values of this feature, we must have the following faithfulness constraint. (6) Faith[comp]: The output value of [comp] is the same as the input value. 4.2.2 Analytic Essentials If there are no relevant markedness constraints ranked higher than Faith[comp] that distinguish a particular pair of CP and IP forms, Faith[comp] ensures that the faithful output candidate for each type of input embedded clause is the optimal candidate in its candidate set, and hence a grammatical option. The input embedded clause speci®ed as [‡comp] will surface as a [comp]-faithful CP, and the one speci®ed as

Optionality and Ine¨ability

101

Tableau T4.1 Complementizer optionality: Faith[comp] g Mark Input: [‡comp]

Faith[comp]

f a. CP b. IP Input: [ comp] c. CP

Mark (*)

*!

(*)

Faith[comp]

Mark

*!

(*)

f d. IP

(*)

Tableau T4.2 Complementizer obligatoriness: MarkIP g Faith[comp] Input: [‡comp]

MarkIP

Faith[comp]

*!

*

MarkIP

Faith[comp]

f a. CP b. IP Input: [ comp] f c. CP d. IP

* *!

[ comp] will surface as a [comp]-faithful IP. The e¨ect will be the optionality of a complementizer, as in the examples in (2). This is shown in tableau T4.1. (Mark stands for any and all markedness constraints that militate against either the CP or IP form in these cases.) On the other hand, if the output structure of a particular form is such that a higher-ranked markedness constraint distinguishing the contrasting CP and IP forms is relevant, the result is neutralization of the contrast. This neutralization can be in favor of the CP form as in tableau T4.2, resulting in the obligatory complementizer e¨ect in (3) and (4), or in favor of the IP form as in tableau T4.3, resulting in the complementizer absence e¨ect in (5). (MarkXP stands for some markedness constraint that militates against the XP form in each of these cases.) In the following subsections we make some explicit claims as to the actual content of the schematic markedness constraints Mark, MarkIP, and MarkCP to account

Eric Bakovic and Edward Keer

102

Tableau T4.3 Complementizer absence: MarkCP g Faith[comp] Input: [‡comp] a. CP

MarkCP *!

f b. IP Input: [ comp] c. CP

Faith[comp]

* MarkCP

Faith[comp]

*!

*

f d. IP for the particular cases exempli®ed in (2)±(5). These constraints are the same ones employed by Grimshaw (1997b) in her account of the same set of data; the major di¨erence here is their necessary ranking with respect to the new constraint Faith[comp], required by our approach to optionality. We begin in the middle, with complementizer obligatoriness. 4.2.3 Complementizer Obligatoriness in Relatives with Subject Extraction Recall from (3), repeated below in (7), that subject extraction from a relative induces the obligatory presence of the complementizer. (7) Complementizer obligatoriness in relatives with subject extraction a. The coat [CP that t doesn't ®t him] might ®t me. b. *The coat [ IP t doesn't ®t him] might ®t me. In Grimshaw's account, this is because subject traces in relative clauses potentially run afoul of the constraint TGov, demanding that traces be governed (see also DeÂprez 1994). Relative clauses (in English) are adjoined structures and are thus not governed. The subject trace in (7b) is thereby also ungoverned, violating TGov. On the other hand, the subject trace in (7a) is governed by the complementizer that, satisfying the constraint. It should be quite clear how TGov naturally takes the place of MarkIP in tableau T4.2 to explain the obligatory complementizer e¨ect in (7), as shown in tableau T4.4. The rank of TGov above Faith[comp] explains why the particular con®guration of subject extraction from a relative clause requires a complementizer. If the input e¨ectively lacks one (that is, if it is speci®ed as [ comp]), the low-rank of Faith[comp] relative to TGov means that the optimal output is going to violate Faith[comp] and be a less marked CP, rather than the [comp]-faithful but more marked IP.

Optionality and Ine¨ability

103

Tableau T4.4 Complementizer obligatoriness in relatives: TGov g Faith[comp] Input: [‡comp] relative clause with subject extraction

TGov

Faith[comp]

b. The coat [IP t doesn't ®t him] might ®t me.

*!

*

Input: [ comp] relative clause with subject extraction

TGov

Faith[comp]

f a. The coat [CP that t doesn't ®t him] might ®t me.

f c. The coat [CP that t doesn't ®t him] might ®t me. d. The coat [IP t doesn't ®t him] might ®t me.

* *!

4.2.4 Complementizer Obligatoriness in Complements with Adjunction Now recall from (4), repeated below in (8), that adjunction to a complement also requires the presence of the complementizer. Again, following Grimshaw's account, adjunction to the highest node of an embedded clause violates PureEP (see also McCloskey 1992; Doherty 1993). (8) Complementizer obligatoriness in complements with adjunction a. I think [CP that on him, no coat looks good t]. b. *I think [ IP on him, no coat looks good t]. The presence of the higher complementizer that in (8a) means that the highest node of the complement (here, CP) is not adjoined to, satisfying PureEP. In (8b), with no complementizer, the highest node of the complement (here, IP) is adjoined to, violating the constraint. All that remains to be said is that PureEP, like TGov, dominates Faith[comp], as shown in tableau T4.5.5 The ranking of PureEP above Faith[comp] explains why adjunction to a complement clause requires a protective complementizer. If the input e¨ectively lacks one (that is, if it is speci®ed as [ comp]), the low rank of Faith[comp] relative to PureEP means that the optimal output is going to violate Faith[comp] and be a protected CP, rather than the [comp]-faithful IP. 4.2.5 Complementizer Absence Unlike relative clauses, complement clauses are complements (to verbs), and are hence (lexically) governed. Subject traces in complements thus satisfy TGov whether there is a complementizer or not. But recall from (5), repeated below in (9), that a

Eric Bakovic and Edward Keer

104

Tableau T4.5 Complementizer obligatoriness in complements: PureEP g Faith[comp] Input: [‡comp] complement clause with adjunction

PureEP

Faith[comp]

b. I think [IP on him, no coat looks good t].

*!

*

Input: [ comp] complement clause with adjunction

PureEP

Faith[comp]

f a. I think [CP that on him, no coat looks good t].

f c. I think [CP that on him, no coat looks good t]. d. I think [IP on him, no coat looks good t].

* *!

complementizer is ungrammatical here, unlike subject extraction from a relative clause. (9) Complementizer absence in complements with subject extraction a. *Which coat do you think [CP that t doesn't ®t]? b. Which coat do you think [ IP t doesn't ®t]? This is because the subject trace in (9a) violates TLexGov, demanding that traces not only be governed but lexically governed (DeÂprez 1994).6 If TLexGov also dominates Faith[comp], then given a choice between an IP with a lexically governed subject trace and a CP with a nonlexically governed trace, TLexGov prefers the former, at the expense of Faith[comp]. This is shown in tableau T4.6. The relative ranking of TGov and TLexGov is irrelevant here, since they make partially overlapping rather than con¯icting demands. TGov is satis®ed by both lexical and nonlexical government, so it fails to distinguish the forms in (9) and the work is left entirely up to the con¯ict between TLexGov and Faith[comp] in tableau T4.6. 4.2.6 Complementizer Optionality Any markedness constraint that prefers CPs to IPs or vice versa in the contexts in which the complementizer is optional must be dominated by Faith[comp] in order for the optionality to be possible. The assumption in Grimshaw 1997b is that no such constraints exist, and that optionality emerges as a consequence of the optimality of two indistinguishable structures. But there is at least one constraint in Grimshaw's system that does distinguish the forms in question. HdRt, demanding rightmostness of a head in its projections, militates against the CP candidate relative to the IP

Optionality and Ine¨ability

105

Tableau T4.6 Complementizer absence: TLexGov g Faith[comp] Input: [‡comp] complement clause with subject extraction

TLexGov

a. Which coat do you think [CP that t doesn't ®t]?

*!

f b. Which coat do you think [IP t doesn't ®t]?

Faith[comp]

*

Input: [ comp] complement clause with subject extraction

TLexGov

Faith[comp]

c. Which coat do you think [CP that t doesn't ®t]?

*!

*

f d. Which coat do you think [IP t doesn't ®t]? candidate. This and any other similar constraints must be outranked by Faith[comp] in order to prevent them from changing input [comp]-speci®cations, as shown in tableau T4.7. Given that constraints like HdRt exist, something like Faith[comp] must exist to account for the optionality of the complementizer in these forms. Note that it is possible that HdRt does not exist in English and other languages with the opposing constraint HdLft being dominant: as Grimshaw (1997a) has independently argued, the direct opposition of alignment constraints like HdLft and HdRt completely inactivates the lower ranked of the two (modulo the way that it can emerge to prevent optionality, as just shown above). HdLft and HdRt can thus be seen as di¨erent parametric settings of the same universal constraint schema, Hd{Lft/Rt}. Since HdLft must be dominant in English for independent reasons (see Grimshaw 1997b:406±409), HdRt need not be posited for English and thus will not be in the way to prevent complementizer optionality. The elimination of HdRt does not, however, change the di½cult-to-reconcile fact that any constraint distinguishing the candidates in free variation will subvert Grimshaw's (1997b) tied-candidate approach to optionality. In fact, this argument applies to other one-input/many-output approaches to optionality, in particular to Pesetsky's (1998) tied-constraint approach. Pesetsky's de®nition of a constraint tie crucially allows for a constraint to distinguish candidates that would otherwise tie.7 Advocates of this approach, like advocates of Grimshaw's, must deny the existence of such constraints applying to the cases where a tie is desired. One crucial di¨erence between Grimshaw's tied-candidate approach and that of Pesetsky is that Grimshaw's approach does not add anything to the mechanisms

Eric Bakovic and Edward Keer

106

Tableau T4.7 Complementizer optionality: Faith[comp] g HdRt Input: [‡comp] embedded clause

Faith[comp]

f a. The coat [CP that he always wears t] doesn't ®t him. I think [CP that the coat doesn't ®t him]. b. The coat [IP he always wears t] doesn't ®t him. I think [IP the coat doesn't ®t him]. Input: [ comp] embedded clause c. The coat [CP that he always wears t] doesn't ®t him. I think [CP that the coat doesn't ®t him].

HdRt

* *! Faith[comp]

HdRt

*!

*

f d. The coat [IP he always wears t] doesn't ®t him. I think [IP the coat doesn't ®t him]. of OT, whereas Pesetsky's introduces the notion of tied constraints. In a sense, then, Grimshaw's approach cannot be argued against in general. OT allows for the possibility that any two candidates could tie on all the possible constraints and thus that both could be optimal. In order for this tie to occur, no constraint may prefer one candidate to the other on any of the ways they di¨er. We are arguing here that there must be markedness constraints that prefer that-less clauses to clauses with that and that these constraints can only be inactivated through domination (by faithfulness). There are other compelling reasons to disprefer these one-input/many-output approaches. One inevitable prediction of Grimshaw's tied-candidate approach to complementizer optionality is that if two candidates c1 and c2 tie for optimality in some language L1 , then the optimality of c1 in another language L2 entails the optimality of c2 in L2 (and vice versa, of course). That is, tied candidates are inextricably bound to each other, and one cannot be optimal without the other.8 Pesetsky's tiedconstraint approach in particular predicts that any di¨erence between languages that is analyzed as the di¨erent available rankings of con¯icting constraints could be a point of free variation in some third language, through the crucial nonranking of the relevant constraints. There seems to us to be too much descriptive looseness and too little explanatory payo¨ with this approach, under any plausible de®nition of constraint tie.

Optionality and Ine¨ability

107

4.2.7 Summary Complementizers in English embedded clauses are only optionally present, except under certain conditions when they are either obligatorily present or obligatorily absent. The optionality itself is due to the purely functional nature of the values of the [comp] feature, its arbitrary speci®cation in the input, and the faithfulness constraint Faith[comp]. When imposed on by con¯icting markedness constraints such as TGov, PureEP, and TLexGov, Faith[comp] gives way and there is loss of optionality in just those contexts that the markedness constraints are sensitive to, and optionality otherwise. 4.3

English vs. Norwegian

Like English that, the Norwegian complementizer som is generally optional in declarative complements and in relative clauses. The conditions under which it is not optional, somewhat distinct from the conditions in English noted above, are discussed and analyzed in detail by Keer (1996). We concentrate here on another point of difference between English and Norwegian, namely, the optionality of wh-relatives in Ê farli English and their ungrammaticality in Norwegian (Taraldsen 1978, 1986; cf. A 1994).9 (10) Optional vs. impossible wh-relatives English Norwegian a. the man [CP who I know t] d. *mannen [CP hvem jeg kjenner t] b. the man [CP that I know t] e. mannen [CP som jeg kjenner t] c. the man [ IP I know t] f. mannen [ IP jeg kjenner t] Following the analyses of Grimshaw (1997b) and Keer (1996), we assume that a wh-relative ((10a) and (10d)) in these two languagesÐa headless CP with a wh-phrase in its speci®erÐincurs a violation of ObHd, requiring that a projection have a head. This violation is forced by higher-ranked HdLft, requiring that a head be leftmost in its projections; HdLft is satis®ed by complementizer-headed relatives ((10b) and (10e)), but violated by a relative introduced by both a wh-phrase and a complementizer (cf. the Doubly-Filled Comp Filter; Chomsky and Lasnik 1977). The violation of ObHd incurred by wh-relatives is obviously tolerated in English but not in Norwegian. We propose that this is a result of the di¨erent ranking of ObHd with respect to another faithfulness constraint, Faith[wh] (cf. FaithOp in Keer and Bakovic 1997). (11) Faith[wh]:

The output value of [wh] is the same as the input value.

The feature [wh] here stands for the presence versus absence of a wh-phrase: a [‡wh] input is faithfully realized by a clause with a wh-phrase, while a [ wh] input is

Eric Bakovic and Edward Keer

108

Tableau T4.8 wh-phrase optionality in English: Faith[wh] g ObHd Input: [‡wh] object relative clause

Faith[wh]

f a. the man [CP who I know t] b. the man [IP I know t] Input: [ wh] object relative clause c. the man [CP who I know t]

ObHd *

*! Faith[wh]

ObHd

*!

*

f d. the man [IP I know t] Tableau T4.9 wh-phrase absence in Norwegian: ObHd g Faith[wh] Input: [‡wh] object relative clause a. mannen [CP hvem jeg kjenner t]

ObHd *!

f b. mannen [IP jeg kjenner t] Input: [ wh] object relative clause c. mannen [CP hvem jeg kjenner t]

Faith[wh]

* ObHd

Faith[wh]

*!

*

f d. mannen [IP jeg kjenner t] faithfully realized by a clause without one.10 Faith[wh] must dominate ObHd in English, since wh-relatives are optional in that language, but the opposite ranking must hold in Norwegian, since in this language wh-relatives are ungrammatical. This is shown in tableaux T4.8 and T4.9. The winner for the [ wh] input in tableau T4.8 and for both inputs in tableau T4.9 is simply an IP with a trace in object position. We follow Doherty (1993) in assuming that the trace is bound by the head of the relative clause (man in English and mannen in Norwegian). Note that the traditional analysis of these relative clauses (Chomsky 1977) is that they are CPs with an empty operator in the speci®er of CP. Nothing we have said so far hinges on this choice. However, the empty operator approach requires one more ranking for both languages: Faith[comp] must dominate ObHd since the complementizer is optional in both Norwegian and English object relatives. This is shown in tableau T4.10 for Norwegian.11

Optionality and Ine¨ability

109

Tableau T4.10 Complementizer optionality with empty operators: Faith[comp] g ObHd Input: [‡comp] object relative clause

Faith[comp]

ObHd

*!

*

Faith[comp]

ObHd

f a. mannen [CP Op som [ jeg kjenner t]] b. mannen [CP Op [ jeg kjenner t]] Input: [ comp] object relative clause c. mannen [CP Op som [ jeg kjenner t]]

*!

f d. mannen [CP Op [ jeg kjenner t]]

*

Tableau T4.11 Complementizer optionality: Faith[comp] g HdRt Input: [‡comp] embedded clause

Faith[comp]

f a. the man [CP that I know t] mannen [CP som jeg kjenner t] b. the man [IP I know t] mannen [IP jeg kjenner t]. Input: [ comp] embedded clause c. the man [CP that I know t] mannen [CP som jeg kjenner t]

HdRt *

*! Faith[comp]

HdRt

*!

*

f d. the man [IP I know t] mannen [IP jeg kjenner t] Both languages exhibit complementizer optionality in non-wh object relative clauses ((10b) and (10c) for English, and (10e) and (10f ) for Norwegian). Since the trace in object relatives is guaranteed to be lexically governed by the embedded verb, both TGov and TLexGov are satis®ed by both relevant output candidates in both languages. Complementizer optionality results in both languages from the ranking of Faith[comp] above HdRt, as shown in tableau T4.11 (see also section 4.2.6).

Eric Bakovic and Edward Keer

110

4.4

Concluding Remarks

We have proposed here a general theory of optionality, adapting the tools of OT to the problem by admitting faithfulness constraints to distinctive formal properties of syntactic structures. When con¯icting markedness constraints are either irrelevant or subordinate to one of these faithfulness constraints, optionality of a formal property arises. When markedness is relevant and dominates faithfulness, the result is ine¨ability. We have attempted to show in this chapter that these predicted interactional possibilities are borne out both within a language and across languages, lending support to the approach. Notes This chapter consolidates, and in some parts improves on, work by both authors, independent and joint (see Bakovic 1997; Keer and Bakovic 1997); our names appear here in alphabetical order. Portions of the present version were presented independently by both authors at the Workshop on Optimality Theory Syntax at UniversitaÈt Stuttgart. We would like to thank the participants in that workshop as well as the audiences at RUM J, CLAM II, WCCFL XVI, and HOT-97; thanks also go to Viviane DeÂprez, Jane Grimshaw, Takeo Kurafuji, GeÂraldine Legendre, Susanne Preuss, Alan Prince, Vieri Samek-Lodovici, Paul Smolensky, Sten Vikner, Colin Wilson, and an anonymous reviewer. This work was supported by National Science Foundation Grant SBR-9511891 and by Rutgers University. 1. On discourse-based or apparent optionality within OT, see the chapters by Choi (chap. 6), Costa (chap. 7), and Samek-Lodovici (chap. 11) in this volume and the references cited therein. 2. Kurafuji (1997) provides additional evidence from di¨erent dialects of Japanese, arguing for the present approach to optionality as opposed to the approaches advocated by Grimshaw (1997b) and Pesetsky (1998). See section 4.2.6 below. 3. Our notion of the ``input'' thus corresponds more closely to the notion of the ``numeration'' in Chomsky 1995. 4. Grimshaw (1997b:381) argues further that clauses are only as big as they need to be and can be CP, IP, or VP. Our analysis is entirely compatible with this assumption; the further distinction between VP and IP is reduced here to IP for purely expository reasons. 5. The full form of Grimshaw's PureEP also rules out movement into the head of a subordinate clause (see Rizzi and Roberts 1989; McCloskey 1992), which is irrelevant to our immediate concerns here. Grimshaw (1998) argues for a di¨erent analysis of the e¨ects of PureEp that is entirely compatible with the essence of the analysis put forth here. 6. Subject traces in English relatives uniformly violate TLexGov because English relatives are adjuncts, hence the irrelevance of this constraint in the case of relatives. 7. Other de®nitions of constraint ties, with somewhat di¨erent consequences, have been proposed by Ackema and Neeleman (1998) and MuÈller (1997). 8. In the case of complementizers, this prediction seems discon®rmed by languages such as Spanish or French in which the complementizer is generally obligatory in declarative complements. But one could, of course, posit a structural di¨erence between the declarative com-

Optionality and Ine¨ability

111

plements of English and those of Spanish and French, as Vikner (chap. 14, this volume) does, such that the winning candidate with the complementizer in Spanish and French is somehow distinct from the corresponding (losing) candidate in English. 9. We restrict our attention to argument relatives here; adjunct relatives, for instance, can and must be introduced by a wh-phrase in Norwegian as well as in English. See Keer and Bakovic 1997 for discussion and analysis. 10. For recoverability reasons (see Keer and Bakovic 1997; cf. Pesetsky 1998), only argument relatives can be faithfully realized without a wh-phrase. 11. There is a further complication with the empty operator analysis: in English relatives with subject extraction (see section 4.2.3), there must be a trace in subject position for TGov to force the presence of a complementizer. To achieve this, empty operators must be forced to move. See Keer 1996 for discussion. References Ackema, Peter, and Ad Neeleman. 1998. WHOT? In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 15±33. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Ê farli, Tor. 1994. A Promotion Analysis of Restrictive Relative Clauses. Linguistic Review 11: A 81±100. BakovicÂ, Eric. 1997. Complementizers, Faithfulness, and Optionality. Available as ROA-212 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Chomsky, Noam. 1977. On Wh-Movement. In P. Culicover, T. Wasow, and A. Akmajian (eds.), Formal Syntax, 77±142. New York: Academic Press. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Chomsky, Noam, and Howard Lasnik. 1977. Filters and Control. Linguistic Inquiry 8: 425± 504. DeÂprez, Viviane. 1994. A Minimal Account of the that-t E¨ect. In G. Cinque, J. Koster, J.-Y. Pollock, and R. Zanuttini (eds.), Paths toward Universal Grammar: Studies in Honor of Richard S. Kayne, 121±135. Washington, D.C.: Georgetown University Press. Doherty, Cathal. 1993. Clauses without that: The Case for Bare Sentential Complementation in English. Doctoral dissertation, University of California, Santa Cruz. Grimshaw, Jane. 1997a. The Best Clitic and the Best Place to Put It. Paper presented at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest 1997, Baltimore. Grimshaw, Jane. 1997b. Projection, Heads, and Optimality. Linguistic Inquiry 28(3): 373±422. Grimshaw, Jane. 1998. Constraints on Constraints in OT Syntax. Ms., Rutgers University, New Brunswick, N.J. Keer, Edward. 1996. Anti-*that-Trace E¨ects in Norwegian. Ms., Department of Linguistics, Rutgers University, New Brunswick, N.J. Available as ROA-187 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Keer, Edward, and Eric BakovicÂ. 1997. Have Faith in Syntax. In E. Curtis, J. Lyle, and G. Webster (eds.), Proceedings of the Sixteenth West Coast Conference on Formal Linguistics,

112

Eric Bakovic and Edward Keer

255±269. Stanford, Calif.: CSLI Publications. Available as ROA-200 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Kurafuji, Takeo. 1997. Three OT Approaches to the Optionality of Complementizers. Ms., Rutgers University, New Brunswick, N.J. Legendre, GeÂraldine, Colin Wilson, and Paul Smolensky. 1998. When Is Less More? Faithfulness in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249± 289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Legendre, GeÂraldine, Colin Wilson, Paul Smolensky, Kristin Homer, and William Raymond. 1995. Optimality and Wh-Extraction. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 607± 636. GLSA, University of Massachusetts, Amherst. McCloskey, James. 1992. Adjunction, Selection, and Embedded Verb Second. Linguistics Research Report LRC-92-07. University of California, Santa Cruz. MuÈller, Gereon. 1997. Partial Wh-Movement and Optimality Theory. Linguistic Review 14: 249±306. Pesetsky, David. 1998. Some Optimality Principles of Sentence Pronunciation. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition, in Syntax, 337±383. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Rizzi, Luigi, and Ian Roberts. 1989. Complex Inversion in French. Probus 1: 1±30. Taraldsen, Knut Tarald. 1978. The Scope of wh-Movement in Norwegian. Linguistic Inquiry 9: 623±640. Taraldsen, Knut Tarald. 1986. Som and the Binding Theory. In L. Hellan and K. K. Christensen (eds.), Topics in Scandinavian Syntax, 149±184. Dordrecht: Reidel.

Chapter 5 The Emergence of the Unmarked Pronoun

Joan Bresnan

In pre±Optimality Theory (OT) generative syntax, unlike phonology, markedness theory has had only a marginal role in theory development.1 Yet appeals to markedness are often implicit in generative syntactic argumentation. For example, Baker (1993) points out that the free NP is a more basic structure than the incorporated noun: all noun-incorporating languages have free NPs and many languages with free NPs have no noun incorporation. This is a classic markedness asymmetry, indicating that among inventories of nominal expressions, morphologically bound nominal stems are marked and free NPs are unmarked as expressions of argument roles. Baker takes it as evidence that incorporated nouns are transformationally derived from free NPs by syntactic head movement. Indeed, the entire motivation for syntactic movement can be seen to be based on the idea that there are ``canonical,'' unmarked phrase structure positions in which arguments receive their logical interpretations (or theta roles), and that the more marked structures are derived from these by structure-modifying rules. What generative syntax has tacitly developed, then, is a derivational theory of markedness in which marked structures are represented as transformationally complex structures. In contrast, OT o¨ers a radically di¨erent approach to markedness based on the crosslinguistic typology of outputs rather than the derivation of marked outputs from unmarked inputs, and this has led to a fully nonderivational phonological theory of markedness. Can we have a radically nonderivational theory of syntactic markedness similar to current theories of phonological markedness?2 This question is addressed in Bresnan (forthcoming a,b). Bresnan (forthcoming a) proposes a simple markedness theory of bound, free, and zero pronominals that can explain the emergence of the unmarked pronoun in situations where the more marked pronominal forms otherwise required are unavailable. The theory derives the following two crosslinguistic markedness generalizations about pronominals: (1) Markedness relations of pronominal forms (Bresnan, forthcoming a) a. ``No language lacks free forms while some languages may lack bound forms.'' (Carstairs-McCarthy 1992:165±166) b. No language lacks free forms, while some languages may lack zero forms.

114

Joan Bresnan

This theory implies that the free pronoun has the unmarked typological distribution but may also show itself to be unmarked language-internally, in the sense of Jakobson 1984. In languages with both bound and free pronominals, the free pronoun generally is used for focus (Schwartz 1986) in contexts where it contrasts with a bound form, but it may nevertheless ®ll in the gaps in the paradigm of bound pronominals, taking on the nonfocus uses of the latter. In the present chapter I further develop and exemplify this theory and consider possible counterexamples to (1). 5.1

Pronominal Form and Pronominal Content

I will assume without argument that elements that function as de®nite personal pronouns are not structurally uniform across languages but show formal variation, as schematized in (2).3 (2) Range of personal pronominal forms Zero Bound Clitic Weak Pronoun Zero designates pronominals having no expression in morphology or syntax; bound designates morphologically bound pronominals, also called pronominal in¯ections, which are expressed by a½xal structure on a head; clitic refers to elements that have a specialized syntactic position and are phonologically bound to a host (special clitics in Zwicky's (1977, 1985) sense); weak pronouns are free forms, neither phonologically nor morphologically bound to another constituent, but they do not receive primary sentence accents. Pronouns are also free but may receive primary sentence accents. However, the di¨erence between weak and strong free pronouns is not primarily phonological; we distinguish a special category of weak pronouns in a language only if they di¨er from free pronouns in form and syntactic distribution. On this assumption, what universally characterizes a pronoun are its referential role and functions, not its phrase structure category. Some elements that resemble clitic pronouns, such as the indirect object clitic copies in Spanish, are not pronominal in content but simply markers of grammatical agreement (SunÄer 1988; Andrews 1990). They occur with every kind of indirect object, including negative inde®nites, interrogatives, and so on. Likewise, some elements that resemble nonpronouns, such as deictics used anaphorically in many languages or bound agreement morphology on verbs, may actually function as pronominals (Greenberg 1986:xix; Bresnan and Mchombo 1987; Demuth and Johnson 1989). This assumption is in accordance with typologically oriented work from a variety of traditions, including functional syntax (e.g., GivoÂn 1976, 1983, 1984, 1990, 1995; Nichols 1986; Van Valin 1996), lexical functional grammar (e.g., Mohanan 1982; Simpson 1983, 1991; Kameyama 1985; Bresnan and Mchombo 1986, 1987; Andrews 1990; Austin and Bresnan 1996;

Emergence of the Unmarked Pronoun

115

Bresnan 1998b, forthcoming c), OT syntax (Grimshaw and Samek-Lodovici 1998; Samek-Lodovici 1996; Bresnan, forthcoming a,b), and some work in the Minimalist Program (Everett 1996). All this work analyzes the variety of pronominal forms simply as alternative forms of expression (or outputs) not di¨ering in derivational complexity. Pronouns can be identi®ed crosslinguistically by their semantic, informationstructural, and morphosyntactic properties. They are generally de®nable as basic anaphoric expressions characterized by systematically shifting reference to persons within the utterance context. The referents of I, you, and she shift in the sense that they vary systematically with the speaker and addressee: I refers to the speaker and excludes the addressee, you refers to the addressee and excludes the speaker, and she refers to a third party who is neither the speaker nor the hearer. Of course, composite phrasal expressions like ``the speaker of the present utterance,'' ``the addressee of this utterance,'' and ``the woman I told you about yesterday'' could be argued to have the same properties of shifting reference depending on speaker and addressee, but these are not basic expressions syntactically. Finally, anaphoricity distinguishes pronominals from basic expressions that are pure deictics, like this and that: though pronominals often derive historically from deictics (Greenberg 1986:xix), they must have anaphoricity as a synchronic property to be functioning as personal pronouns. (An operational de®nition of anaphoricity is referential dependence on a superordinate pronoun within a sentence: the second I in ``I said that I would come'' shows anaphoricity in this sense, while ``that woman'' in ``I said that that woman would come'' does not, even when that woman is in fact the speaker of this sentence.) Similarly, pronominals sometimes derive historically from common nouns (Sugamoto 1989; Cooke 1968) and from honori®c nominal phrases (MuÈhlhaÈusler and Harre 1990:136± 137), but their referential role and function in the synchronic grammar, not their etymology, determines their pronominality. This issue is discussed further below. The major types of pronominal properties are schematized in (3). (3) Crosslinguistic properties of personal pronouns proÐshifting reference, anaphoricity topÐtopic-anaphoricity (GivoÂn 1976, 1983, 1984, 1990:916¨.) agrÐclassi®cation by person, number, gender (GivoÂn 1984:354±355) PRO stands for the semantic properties shared by all personal pronominals, which include shifting reference and anaphoricity as described above. TOP abbreviates the information-structural functions of personal pronouns such as specialization for reference to topical elements (GivoÂn 1976, 1983, 1984. 1990:916¨.). Many languages have a distinct series of pronominal forms reserved for reference to the topic; in à a, for example, morphologically bound pronominal forms must be used Chichew to refer to a dislocated topic (Bresnan and Mchombo 1986, 1987). AGR represents

116

Joan Bresnan

2

3 top    top 4 pro 5 pro ... agr pro agr Figure 5.1 Representation of pronominal content by feature structures

Zero:

   pro j; top Free:

*

3+ top af ; 4 pro 5 Bound : agr    pro X0; agr 2

Figure 5.2 Representation of pronominals as form-content pairings

the classi®catory dimensions by which personal pronominals are morphologically distinguishedÐperson (allowing for participant deixis and inclusion/exclusion relations among participants), number (singular, dual, paucal, and plural), and gender (classi®cations into kinds) (GivoÂn 1984:354±355); this property is abbreviated by agr in (3). Not all pronouns have agr and top features.4 Personal pronouns can be represented independently of their forms of expression by using feature structures based on these properties, as illustrated in part by ®gure 5.1. The leftmost feature structure in ®gure 5.1 speci®es a pronominal that is specialized for topic anaphoricity and is also classi®ed for person, number, or gender. The rightmost feature structure speci®es a specialized topic-anaphoric pronominal that lacks any agreement classi®cations.5 The pronominal inventory of a language may be de®ned as a set of pairings of instances of structural types from (2) with feature structures representing pronominal content as in ®gure 5.1 (Bresnan, forthcoming a). As shown in ®gure 5.2, for example, bound and free pronouns may be represented as the pairings of a morphological a½x af or a syntactic category X 0 , respectively, with a feature structure representing their pronominal content; and the zero pronoun may be represented as the pairing of pronominal content with no structural expression at all. Null structure is the absence of structure, represented by q. Note that both morphological and syntactic structure are excluded from this de®nition of zero pronouns. Thus zero pronouns here do not include cases of so-called pro-drop in the presence of agreement morphology; the latter are analyzed not as zero pronouns, but as pronominal in¯ections represented as bound in (2) (GivoÂn 1976; Jelinek 1984, 1988, 1990, 1995; Sandoval and Jelinek 1989; Demuth and Johnson 1989; Andrews 1990; Speas 1990; Willie 1989; Sadock 1991; Uyechi 1991; Jelinek and Demers 1994; Bresnan 1998b, forthcoming c); BoÈrjars,

Emergence of the Unmarked Pronoun candidates    top j; pro * 2 3+ top   pro af ; 4 pro 5 ƒƒ! top agr    pro 0 X ; agr

ƒƒƒ ƒƒƒ ƒ!

input

117

output

*

ƒ!

ƒƒ

ƒƒ

3+ top af ; 4 pro 5 agr 2

.. . Figure 5.3 Correspondence model (Bresnan, forthcoming a,b,d)

Vincent and Chapman, 1997; Toivonen 1996, 1997; Everett 1996; and Speas 1997; among many others). (This is a slight simpli®cation. For evidence distinguishing agreement and pronominal in¯ection, see Simpson 1983, 1991; Bresnan and Mchombo 1986, 1987; Austin and Bresnan 1996; Nordlinger 1998; Mereu 1997.) The inventory of pronominals of each language is selected from the possible pairings by evaluating them against universal constraints as prioritized by the language. Within OT morphosyntax, then, the universal content of personal pronominals (which will be the ``input'') will consist of all possible combinations of the pronominal feature types in (3), represented by feature structures as in ®gure 5.1. The universal candidate set of structural analyses of pronouns will include all possible pairings of these feature structures with instances of structural types, including the examples in ®gure 5.2, among many others. This model is illustrated in ®gure 5.3. Bresnan (forthcoming a,b) observes that this model satis®es two fundamental requirements of OT: (1) the universality of the input implied by ``richness of the base'' (Smolensky 1996) and (2) the recoverability of the input from the output, implied by the ``containment'' or ``correspondence'' theories of the input-output relation (Prince and Smolensky 1993; McCarthy and Prince 1995). Because richness of the base implies that the input must be universal, the syntactic gen cannot simply be de®ned as mapping a set of language-particular ``lexical heads'' or morphemes onto structural forms. A more abstract and crosslinguistically invariant characterization of the input is required. Because the recoverability of the input from the output is fundamental to the learnability of OT (Tesar and Smolensky 1998), the input must either be contained in the output or must be identi®able from the output by a correspondence. Hence the candidate set cannot simply consist of surface forms (such as strings of morphemes parsed into phrase structure trees) alone.

118

Joan Bresnan

The theory of representations assumed in this model already exists: lfg (Kaplan and Bresnan [1982] 1995; Dalrymple et al. 1995) provides a mathematically wellde®ned correspondence between parallel feature structures (representing languageindependent content) and categorial structures (representing the variety of surface forms). In lfg morphological and syntactic forms may correspond to complex f(eature)-structures of exactly the same type. Among other results, this allows for a well-de®ned domain for pronominal binding theory over the variety of structural types of pronominal forms shown in (2) (Bresnan 1998b, forthcoming c). The universal input can thus be modeled by sets of f-structures, which provide an abstract and form-independent characterization of content.6 The candidate set can consist of pairs of a c(ategorial)-structure and its corresponding speci®c f-structure, which may be matched to the input f-structure by correspondence (Bresnan, forthcoming d). In this framework, enumeration of the candidate set, formalization of the constraint language, the decidability of the universal parsing problem, and computational complexity issues have all been addressed with clarity (Johnson 1998; Kuhn 1999). Thus, each candidate is a structural expression (whether morphological or syntactic) of speci®ed pronominal content. In this conception of gen, the input represents language-independent ``content'' to be expressed with varying ®delity by the candidate forms, which carry with them their own interpretations of that content (Bresnan, forthcoming a). Faithfulness constraints will require the input feature structures to correspond to the candidate feature structures in order to ensure expressibility of content. (``Harmony'') constraints will also apply to the form-content pairings, as we will see later. Which of the ways of structurally expressing pronouns will appear in the inventory of a given language depends on how the candidates are harmonically ordered by the language, given its prioritizing of the constraints. Crucially, the candidates need not be perfect analyses of the input; as illustrated in ®gure 5.3, they may overparse or underparse the input pronominal content. This is what allows for the emergence of the unmarked pronoun, as we will see. 5.2

Markedness Constraints on Pronominals

If all the pairings of possible structural types with pronominal content actually occurred crosslinguistically, that would mean that the form-function relation is arbitrary, like the Saussurean sign. But this is not what we ®nd. No language has an overt de®nite personal pronoun devoid of any distinctions of person, number, or gender, while many languages have zero pronouns with just this property.7 No language has zero, bound, or clitic personal pronouns used only for emphasis and focus, though many languages have free pronouns with this function (which arises from the presence of other forms specialized for topic anaphora (Bresnan, forthcoming a)).8

Emergence of the Unmarked Pronoun

119

overt z‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚}|‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚{ Zero Bound Clitic Weak Pronoun |‚{z‚} nonovert nonreduced z‚‚‚‚‚‚}|‚‚‚‚‚‚{ Zero Bound Clitic Weak |‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚} Pronoun reduced Figure 5.4 Overt and reduced pronouns

Thus there are generally predictable relations between the functional content of pronominals and their formal expressions. To characterize these, we ®rst observe that the pronominal forms can be classi®ed as in ®gure 5.4 into overt/nonovert forms and reduced/nonreduced forms. Overt pronominals are those with perceptible morphological or syntactic exponents; they are formally expressed by a nonnull structural type, either a phrasal constituent or a morphologically bound form. Only the zero pronominal lacks a perceptible exponent and is categorized as nonovert in this sense. Reduced pronominals are those whose exponents have less phonological or morphological substance than nonreduced pronouns; they may be completely devoid of phonological substance (a zero pronoun), or they may be morphologically dependent, prosodically defective, or incapable of bearing primary sentence accent. Only free (potentially strong) pronouns are categorized as nonreduced in this sense. With this classi®cation of forms we can now formulate two Harmony constraints on possible pairings, shown in (4). (4) Harmony constraints on pronominals a. Reduced , top: Pronominals are reduced if and only if they are specialized for topic anaphoricity. b. Overt , agr: Pronominals are inherently speci®ed for person/number/ gender if and only if they are overt. In functional terms (4a) and (4b) can be viewed as ``naturalness'' constraints on the relation between content and symbolization. The Harmony constraints in (4) are abundantly supported by typological and functional observations as well as detailed studies of particular languages. Constraint (4a) refers to the inventories of grammatical pronominal types speci®ed in (2). In languages that have both reduced and nonreduced pronouns as morphosyntactically distinct grammatical series, the two types contrast in function, with the reduced pronominals being specialized for topic anaphoricity (GivoÂn 1984, 1990:917) and the nonreduced pronouns having focus functions (Schwartz 1986). The general correlation between reduced form and topic anaphoricity is observed by GivoÂn (1984,

120

Joan Bresnan

1990:917) under the name referential iconicity. Haiman, citing Zipf 's law (1935), argues that this generalization has an economy motivation, to minimize e¨ort by reducing expression of frequent, familiar, predictable referents (Haiman 1985:150, 167, 194, 232±232). Van Valin (1996) proposes a scalar representation of the relative markedness of referential coding devices with zero pronominals at the most topical extreme. In addition, many observations of the topic anaphoricity of speci®c types of reduced forms in particular languages have been made. On the topic anaphoricity of zero pronouns see for example Kameyama 1985 for Japanese, Grimshaw and SamekLodovici 1998 and Samek-Lodovici 1996 for Italian, and Lambrecht and Lemoine 1996 for spoken French. On pronominal in¯ections see GivoÂn 1976 for Bantu and à a. On clitics see Lambrecht 1981 for Bresnan and Mchombo 1986, 1987 for Chichew spoken French, GivoÂn 1976 for Spanish, and Cardinaletti 1995 and Cardinaletti and Starke 1996. On weak pronouns see Cardinaletti 1995 and Cardinaletti and Starke 1996 for several Romance and Germanic languages. Constraint (4b), the general correlation between zero pronouns and lack of inherent speci®cation for agr properties, is observed by Bresnan (forthcoming a). It is supported by the fact that in languages that lack verbal agreement morphology, zero pronouns are not restricted as to person and number. This is true for Japanese, Chinese, Malayalam, Jiwarli (Austin and Bresnan 1996:248±250), and many other languages. Contrary to constraint (4b), some languages do appear at ®rst sight to have zero pronouns speci®ed for person, number, or gender. For example, Warlpiri has a de®nite third-person singular zero pronominal object (Hale 1973), and Brazilian Portuguese has a de®nite zero pronominal object that can be used only in the third person (singular or plural) (Farrell 1990:328). However, in these cases the zero pronouns are ®lling paradigmatic gaps in the bound or clitic pronominal system. Thus Warlpiri has overt bound pronominal markers for subject and object on its auxiliary in all persons and numbers, except for the third-person singular object (Hale 1973; Nash 1996:121), which is precisely the gap ®lled by the zero. Similarly, Brazilian Portuguese has a more restricted system of pronominal clitics than other Romance languages; its thirdperson accusative forms, singular and plural, are ``no longer vital'' in the language (Farrell 1990:327), and these are precisely the gaps ®lled by the zero.9 In such cases, the restriction of the zero pronoun to uses requiring speci®c featural content follows from morphosyntactic competition: the bound pronominal forms block the use of the zero wherever their own featural speci®cations apply, leaving the zero to be used elsewhere. We will return to this general e¨ect in discussing Warlpiri. The main point to note is that because the featural values of the zeros in these cases are predictable, it is unnecessary and unexplanatory to specify them as intrinsic properties of a null pronominal form.

Emergence of the Unmarked Pronoun *j *af ‰proŠ ‰proŠ |‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚} Iconicity

121

*cl *Weak ‰proŠ ‰proŠ |‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚} Avoid Allotaxy

Figure 5.5 Structural markedness constraints on pronominals

These facts provide empirical support for (4b). There is also conceptual motivation for the constraint: if we make the plausible assumption that the overt marking of referentially classi®catory contrasts in nonzero pronominals makes them more easily perceptible, constraint (4b) can be motivated by the need to maximize expression of referential contrasts for ease of perception (Haiman 1985:179¨., 191¨.). Of course, not all languages select all the natural candidates available, so further constraints are required. The constraints in ®gure 5.5 assess a mark against the various reduced forms of pronominals; they are a generalization of the constraints proposed by Bresnan (forthcoming a).10 Observe that the constraints in ®gure 5.5 single out the reduced pronominals as marked. Markedness of course has many con¯icting dimensions. Reduced pronominals are unmarked as expressions of topical (frequent, familiar, predictable) referents, for reasons we have already discussed. This generalization is embodied in our Harmony constraint (4a), which may be viewed as a positive markedness constraint. Reduced pronouns are also unmarked in a purely formal sense, having less phonological structure by de®nition. However, there is evidence that reduced forms must be marked syntactically in some ways. We have already noted in (1) that their typological distribution is marked with respect to freestanding pronouns. Although free pronouns contrast with reduced pronominals in languages having both types, the free form is generally used to ®ll in gaps in the systems of reduced pronominals: where reduced pronominals are unavailable, the free pronoun may take on their functions (Bresnan, forthcoming a). In that sense, free pronouns are unmarked in the classical Jakobsonian sense (Jakobson 1984): they are the forms used under neutralization of oppositions within a language. We will see several examples of this phenomenon later. In Haiman's (1985) theory of natural syntax, reduced pronominals have a marked syntactic status, despite their favored use for topic anaphoricity in discourse. Zero and bound pronominals violate a syntactic/semantic iconicity constraint, because they yield a nonisomorphic mapping between syntactic constituents and semantic referents and relationsÐthe zero pronoun because it has semantic content without syntactic constituency, and the a½xal pronoun because it is morphologically part of another constituent (the head) and so noniconically maps a relation and referential

122

Joan Bresnan

role, two distinct semantic constituents, onto a single syntactic constituent. In contrast, the clitic and weak pronouns do not su¨er from this defect, because by de®nition they are syntactic elements that are only prosodically dependent or defective. But clitics and weak pronouns have a di¨erent marked property: they are nonuniform in their syntactic distribution with free (neutral) pronouns. In French, for example, clitic pronouns generally appear preverbally, while free pronouns are postverbal. In West Flemish and Swedish, weak pronouns are attracted to positions (such as that of the complementizer or ®nite verb) from which free pronouns are excluded (Haegeman 1996; Sells 1998). This nonuniformity of independent syntactic expressions of the same semantic roles or grammatical functions is called allotaxy by Haiman (1985:162). (I am assuming that zero and bound pronominals are not allotactic because they are not independent syntactic expressions at all, though nothing crucial hinges on this assumption.) Haiman observes that the avoidance of allotaxy isÐalong with iconicityÐa major source of the syntactic regularity seen in pidgins. The isolating, analytic, uniform syntactic structures of pidgins, he argues, can be explained in terms of their extreme syntactic unmarkedness, embracing iconicity and avoiding allotaxy. It is not surprising that New Guinea Pidgin, derived from English, should have freestanding pronouns, but as Haiman (1985:161±162) notes, the same is true of Kenya Swahili in relation to standard Swahili. Standard Swahili's bound subject and object pronominal a½xes on the verb are replaced in Kenya Swahili by ``independent invariable pronouns in a rigid SVO order.'' Likewise, West African Pidgin Portuguese, according to Naro (1973:444), replaced the various clitic pronominals of Portuguese with the fully stressed, independent strong pronominal forms. From this and other evidence, Haiman (1985:161) concludes: ``In pidgins generally, pronouns are always free-standing words (commonly derived from topicalized forms in the target languages)''Ðthat is, the pidgins adopt the unreduced forms of pronouns, which are the syntactically unmarked forms. Haiman's claim about the prevalence of the free pronoun in pidgins has been widely corroborated for stable, crystallized pidgins. As MuÈhlhaÈusler and Harre (1990:262) observe, ``Pidgins prefer free pronoun forms to bound ones,'' and this preference holds as well for a wide variety of pidgins having non-European lexi®ers (Bresnan 1998c). By instantiating the functional motivations of ``Iconicity'' and ``Avoid Allotaxy'' as the structural markedness constraints of ®gure 5.5, we can derive Haiman's (1985) markedness explanation for pidgin pronominal systems from the initial ranking of markedness constraints above faithfulness in OT. Tesar and Smolensky (1998:253) cite Alan Prince (personal communication) for the idea that in the initial ranking itself, the faithfulness constraints may be lower ranked than the structural markedness constraints. This initial structuring of constraints is proposed as a way to explain

Emergence of the Unmarked Pronoun

123

the acquisition of phonologies consisting of unmarked structures. If unmarked structures incur no marks, they provide no evidence for any particular constraint ranking in OT and so will not lead to convergence on a single grammar. The solution is to hypothesize an initial state of the language learner in which structural markedness constraints outrank faithfulness constraints.11 Accordingly, we have the three families of constraints shown in (5), where Harmony refers to the general form-function harmony constraints ((4a) and (4b)); Struct refers to the syntactic markedness constraints in ®gure 5.5, which penalize reduced pronominal forms; and Faith designates the parsefeature (or Max(feature)) family of constraints that require each attribute of the input to appear in the output f-structure. (5) Harmony g Struct g Faith Because the Harmony constraints appear to be so widely observed, I assume here that they are undominated. But they are likely to be violated in some languages, as suggested in notes 7 and 8. In the present context they simply serve to ®lter out the Sausserian excess of arbitrary pronominal forms from the candidate set. Which of the natural forms that pass the Harmony constraints are actually found in the inventory of a language depends on the relative ranking of Struct and Faith constraints. The ranking of all the structural markedness constraints above the faithfulness constraints means that it is worse to be a reduced form (thus violating iconicity or exhibiting allotaxy) than to be unfaithful to the input. Hence this ranking, by the standard OT logic of markedness, yields only potentially strong pronouns in the pronominal inventory of a language (as in highly analytic languages like English and in pidgins). These unreduced free pronouns will not be specialized for the top property, and hence they will be unfaithful to an input speci®ed for the topicality feature. But that violation will matter less, given the ranking in (5), than the violations incurred by being a syntactically marked form. Tableau T5.1 schematically illustrates these points for a representative sample of the candidate set.12 Thus the ranking shown in (5) gives us a pronominal inventory consisting only of potentially strong pronounsÐ that is, syntactically free pronouns capable of bearing primary sentence accents, having morphological classi®cation for person/number/gender, and being unspecialized for topic anaphoricity. These are syntactically the least marked of the pronominal forms from the point of view of iconicity and the avoidance of allotaxy. As soon as one of the structural markedness constraints is demoted below faithfulness, however, the form it marks enters into the inventory; that form becomes optimal for expressing topical content, as illustrated in tableau T5.2. The tableau has been simpli®ed by omitting all forms violating the Harmony constraints, considering only three representative candidate types, and only the relevant instances of Struct. For nontopical content, the free unreduced pronoun is optimal under the same ranking; see tableau T5.3.

124

Joan Bresnan

Tableau T5.1 Input: [pro, top]

Harmony

Struct

Zero:

[pro, top, agr]

*!

*

Zero:

[pro, agr]

*!

*

Zero:

[pro, top]

Zero:

[pro]

*!

*

[pro, top, agr]

Bound:

[pro, agr]

*!

*

Bound:

[pro, top]

*!

*

Bound:

[pro]

*!

*

Pronoun: [pro, top, agr]

*!

*

*!

f Pronoun:

[pro, agr]

Pronoun:

[pro, top]

*!

[pro]

*!

.. .

*

*!

Bound:

Pronoun:

Faith

*

*

* *

Tableau T5.2 Input: [pro, top] Zero:

Harmony [pro, top]

*j [pro]

Faith

*!

f Bound: [pro, top, agr] Pronoun:

.. .

[pro, agr]

*af [pro]

* *!

Emergence of the Unmarked Pronoun

125

Tableau T5.3 Input: [pro]

Harmony

Zero:

[pro, top]

*j [pro]

Faith

*af [pro]

*!

Bound: [pro, top, agr] f Pronoun:

.. .

*!

[pro, agr]

Tableau T5.4 Input: [pro, top] f Zero:

Harmony

Faith

*j [pro]

[pro, top]

Bound: [pro, top, agr] Pronoun:

*af [pro]

.. .

[pro, agr]

* *! *!

It follows that the demotion of the structural markedness constraint admits the corresponding reduced form into the inventory, but only for topical content; the free, unreduced pronoun remains optimal elsewhere. This result is stated by Bresnan (forthcoming a) and by Carstairs-McCarthy (1992:165±166). (6) Markedness relation among bound and free pronoun inventories (Bresnan, forthcoming a) Free pronouns only (English) Ã a, Navajo, etc.) Both free and bound pronouns (Chichew Bound pronouns only (none) By parity of reasoning, exchanging the ranking positions of the markedness constraints for bound and zero pronouns as shown in tableaux T5.4 and T5.5, we derive the markedness relation among zero and free pronoun inventories shown in (7), also stated by Bresnan (forthcoming a). (7) Markedness relation among zero and free pronoun inventories (Bresnan, forthcoming a) Free pronouns only (English) Both free and null pronouns (Jiwarli, Japanese, etc.) Zero pronouns only (none)

126

Joan Bresnan

Tableau T5.5 Input: [pro]

Harmony

Zero:

Faith

[pro, top]

Bound: [pro, top, agr] f Pronoun:

*af [pro]

.. .

*j [pro] *!

*!

[pro, agr]

Studies of the typology of pronominal systems (Forchheimer 1953; Wiesemann 1986) con®rm that while many languages lack reduced pronominal forms, languages that lack freestanding pronouns are rare.13 5.3

Emergence E¨ects

This theory explains not only the crosslinguistic asymmetries observed in the distribution of pronominal forms, but an important and related language-internal phenomenon: free pronouns may ®ll gaps in the paradigm of bound pronominals. 5.3.1 Pronoun Replacing Bound Pronominal In Bresnan (forthcoming a), I show that free pronouns may ®ll gaps in the paradigm à a. In this language, where both a bound pronomiof bound pronominals in Chichew nal and a free pronoun are available, the free pronoun cannot be used topic anaphorically but is reserved for emphasis or contrast, as Bresnan and Mchombo (1987) observe. This is shown for the preposition ndõ `with, by' in (8). (8) a. ndõ Âõwo with it (class 3) b. naÏwo < *na‡ Ïõwo with ‡ it (cl 3) with it (cl 3) The contracted form is used in topic-anaphoric contexts such as resumption of left dislocations, as illustrated in (9a), while the full pronoun object of the preposition cannot be, as illustrated in (9b). mkaÂngo uwu ndi-na-põÂt-a nawo ku msika lion(3) this I-rm.pst-go-indic with-it(3) to market `This lion, I went with it to market.' b. ?*mkaÂngo uwu ndi-na-põÂt-a ndõ Âõwo ku msika lion(3) this I-rm.pst-go-indic with it(3) to market `This lion, I went with it to market.'

(9) a.

Emergence of the Unmarked Pronoun

127

Tableau T5.6 Input: [to hxi, [pro, top]x ] kwa Zero:

[. . . [pro, top]]

kwa ‡ Bound: [. . . [pro, top, agr]] f kwa Pronoun:

lex

*j [pro]

Faith

*af [pro]

*! *!

[. . . [pro, agr]]

* *

But where a bound form is unavailable, the free form is used topic anaphorically. Thus the preposition kwa `to', unlike ndõ `with, by', does not allow a reduced pronominal object. kwa Âõyo to him (class 3) b. *kwaÏyo < kwa ‡ Ïõyo to ‡ him (cl 3) to him (cl 3)

(10) a.

And with this preposition, the full pronoun object can be used to resume dislocated topics (and in all the other environments normally prohibited to strong pronouns). (11) mfuÂmu iyi ndi-kaÂ-kuÂ-neneÏz-a kwa Âõyo chief(3) this I-go-you-tell.on-indic to him(3) `This chief, I'm going to tell on you to him.' à a, assuming This result follows from the ranking given in tableau T5.2 for Chichew that the lexical gap re¯ects some higher-ranking constraint, which is called lex in tableau T5.6 (For further discussion, see Bresnan, forthcoming a,b). 5.3.2 Pronoun Replacing Zero Pronoun Following the same logic, we should expect similar emergence e¨ects in zero pronoun languages. A fuller investigation is required, but some evidence from Japanese (provided with the assistance of Yukiko Morimoto and Peter Sells, personal communication, March 1997) is suggestive. Example (12) shows a context in which the zero pronoun is used referring to a topic, while the free pronoun is excluded. (12) sono hon-o yonda kedo watashi-wa ?? sore-o/j susume-nai that book-acc read.past but I-top (it) recommend-neg `I read that book but I wouldn't recommend it.' However, as in many languages, zero pronouns in Japanese are restricted to core arguments of the headÐsubjects and objects. A zero postpositional object pronoun is

128

Joan Bresnan

Tableau T5.7 Input: [about hxi, [pro, top]x ] Zero:

[pro, top] tuite

Bound: [pro, top, agr] ‡ tuite f Pronoun:

.. .

[pro, agr] tuite

*j obl

*af [pro]

Faith

*!

*j [pro] *

*! *

lacking. In this situation, the overt free pronoun can be used topic anaphorically, as shown in (13). (13) sono hon-o yonda kedo sore-ni/*j tuite-wa hanasitaku nai that book-acc read.past but (that-dat) about-top talk.want neg `I read that book but I don't want to talk about it.' This fact undoubtedly re¯ects the much broader crosslinguistic generalization that reduced pronominals of all types are distributed according to a hierarchy of argument prominence, being most common with subjects and decreasing with the increasing obliqueness of argument roles (GivoÂn 1976; Siewierska 1999).14 Without attempting a full analysis of this markedness scale in OT terms, we can simply suppose that the absence of a zero postpositional object in Japanese re¯ects a additional high-ranking markedness constraint, abbreviated *j obl in tableau T5.7. (See Bresnan 1998a for a discussion of this contextual markedness property.) The emergence of the unmarked overt pronoun follows.15 5.3.3 Zero Pronoun Replacing Bound Pronominal What happens when more than one structural markedness constraint is demoted below faithfulness? Consider, for example, the ranking in (14). (14)


g Faith g

*j *af g [pro] [pro]

Ranking both the zero and the bound structural markedness constraints below faithfulness means that the marks against these reduced forms are overridden by the importance of faithfully preserving the input content. So both these marked forms could be used to express topic-anaphoric content. But the ranking of the zero markedness constraint above the bound markedness constraint means that it is worse to use the zero pronoun than the bound pronoun. This point is illustrated in à a (tableau T5.2), this ranking tableau T5.8. Like the hypothetical ranking for Chichew would also yield an inventory of bound and free pronominals, with the bound form

Emergence of the Unmarked Pronoun

129

Tableau T5.8 Input: [pro, top] Zero:






Faith

[pro, top]

.. .

[pro, agr]

*af [pro]

*!

f Bound: [pro, top, agr] Pronoun:

*j [pro]

* *!

being specialized for topic anaphoricity. The di¨erence is that in situations where a highly ranked constraint penalizes the bound form, the free pronoun emerges as unà a, but the zero pronoun would emerge here. Under the ranking in marked in Chichew (14), the zero pronoun is relatively unmarked for topical content compared to the free pronoun. This situation might be exempli®ed by Warlpiri, which has bound pronominal markers for subject and object on its auxiliary, as noted (Hale 1973; Simpson 1991). We will assume these are pronominal in¯ections, abstractly characterized as Bound in (2).16 The auxiliary is obligatory in main clauses with verbal predicators, but it is absent in in®nitival clauses, which Simpson shows to have the morphological and categorial structure of nominals. In the latter contexts zero pronouns appear. As we would expect from the absence of agreement morphology, these zero pronouns are not restricted as to person and number (Simpson 1991:141±143). Example (15a) shows a third-person plural use, and (15b) shows a second-person singular use. (15) a. Kurlarda kala-lu-nganpa maja-rninja-rla yu-ngu. spear.abs usit-3plsubj-1plexclobj straighten-infin-serial give-past `They would give us spearsi after straightening (themi ).' b. Ngari ˆ ka-rna-ngku yampi-mi nyuntu just ˆ pres-1sgsubj-2sgobj leave-npast you.sg.abs-euph paka-rninja-wangu-rlu Jungarrayi-kirlangu ngumparna-kurlangu hit-infin-priv-erg Jungarrayi-poss brother.in.law-poss ngajulu-rlu-ju, Japaljarri. I-erg-euph Japaljarri.abs `Japaljarri, I'm just leaving you without beating (you) who belong to Jungarrayi my senior brother-in-law.' If we assume that a high-ranking constraint prevents auxiliaries from appearing as constituents of NPs, the zero emerges as the least marked available topical pronominal, as shown in tableau T5.9. In the context of further constraints on the markedness of pronominal agreement features, the same relative markedness of the

130

Joan Bresnan

Tableau T5.9 Input: [pro, top] f Zero:

*(Aux, NP) [pro, top]

Bound: [pro, top, agr] Pronoun:

Faith

.. .

[pro, agr]

*j [pro]

*af [pro]

* *!

* *!

zero and bound pronominal forms would explain the appearance of the zero as the default form ®lling in gaps in the bound pronominal paradigm. 5.4

Languages without Pronouns?

Finally, let us consider possible counterexamples to the main markedness claim derived from the present theory (1), that no language lacks free (unreduced) pronouns. The claim is that free pronouns are never completely absent from a language. They may be absent in certain grammatical functions. For example, in Southwestern Ojibwe (an Algonquian language of North America) free pronouns have a very restricted distribution, appearing as initially focused elements, as appositions, and in coordinations, bound pronouns being used elsewhere (Schwartz and Dunnigan 1986). Free pronouns may also be absent in certain categories of person. For example, in Dyirbal (a Pama-Nyungan language of Australia), ®rst- and second-person free pronouns exist, but third-person pronouns are supplanted by demonstratives, which commonly co-occur with nominal heads (Dixon 1980:357). Such restrictions can be captured with a richer constraint set allowing for interactions of pronominals with grammatical relation and person (Bresnan 1998a; Aissen, chap. 3, this volume). In the present context, however, we are interested in the question of whether there are languages having no free pronouns in any syntactic function. There are two sources of apparent counterexamples to be considered: (1) languages in which free pronouns have a distinctive syntactic distribution di¨ering from NPs/DPs, and (2) languages in which pronouns are derived from other categories, such as demonstratives, nouns, or in¯ected verb stems. Both types raise the issue of whether the pronominal elements in question are in fact free pronouns and not nonpronominals or pronominals of other types. Taking type (1) ®rst, many languages align free pronouns in prominent edge positions distinct from lexical NP/DP positions. Some examples are given in (16).

Emergence of the Unmarked Pronoun

131

(16) Free pronouns with edge alignment Kichaga: strong object pronouns aligned at right edge of VP (Bresnan and Moshi 1990:166) Warumungu: free pronoun subjects and objects aligned at left edge or second position of clause (Hale 1973:341; Simpson and Heath 1982) Zuni: ®rst- and second-person free pronouns (subjects, objects, possessives) aligned at left edge of IP (Nichols 1997:373) For example, the Bantu language Kichaga has bound pronominal objects pre®xed to the verb stem and strong free pronouns, used for contrast, which are found in a position distinct from lexical NPs at the right edge of the VP (Bresnan and Moshi 1990:166). The Australian language Warumungu (Hale 1973; Simpson and Heath 1982) normally places subject and object pronouns in second position in the clause. Here they may be considered weak pronouns or clitics: they are unstressed, and some combinations of subject and object have a fused portmanteau form (Simpson, 1998). However, subject pronouns may also appear fully stressed in sentence-initial position, as Hale (1973:341) observes:17 Warumungu pronouns, although they normally appear unstressed and cliticized, are in fact independent pronouns in the sense that they can appear as isolated, fully stressed wordsÐas answers to questions, for instance. . . . Furthermore, subject pronouns may (under conditions, perhaps of style, which I do not adequately understand) appear fully stressed in sentence-initial position.

Zuni ®rst- and second-person independent pronouns also take second position in the clause (IP), whether they are subjects, objects, or possessives (Nichols 1997:373). Nichols argues that they are not clitics, however: ``They are not phonologically dependent on another constituent or on one another, they may bear full stress, and other material such as adverbs may intervene between them. . . .'' Similar examples could be multiplied (e.g., nominative and genitive pronouns in Tagalog; see Kroeger 1993:119±123, Nespor 1994, Anderson 1996:167). In e¨ect, then, strong pronouns do exist that are allotactic. Allotaxy is not exclusively a property of weak and clitic pronouns. Even lexical NPs may exhibit it. In some languages, for example, inde®nite NPs have special syntactic positions distinct from other NPs (and often adjacent to the verb). I assume that the special positioning of free (strong) pronouns in some languages is also possible. Thus a language having specialized syntactic positions for (subclasses of ) free pronouns does not constitute a counterexample to the present theory, provided that those pronouns have the functional properties of pronouns and are free unreduced forms, potentially capable of bearing primary sentence accent. What may be involved is merely alignment of pronouns along two distinct dimensions of prominenceÐthe person hierarchy or functional hierarchyÐwith an edge position.18 The violation of

132

Joan Bresnan

allotaxy is overridden by these dominating constraints. The grammatical relations of such edge-aligned pronouns determine their agreement, binding, control, and other properties, and follow in the present framework from the correspondence mappings between c-structure and f-structure at the sentence level (Bresnan, forthcoming a,c,d). The second source of apparent counterexamples to the generalization that no language lacks free pronouns is languages in which pronominal elements are derived from other categories. This situation is not uncommon. As Hale (1973:340) remarks concerning the historical development of the Warlpiri subject and object agreement markers, ``The process of destressing and cliticizing pronouns eventually became an obligatory rule and, subsequently, independent pronouns were re-created from other sources available to the language, such as oblique forms of pronouns like those found in possessives or in other functions not normally subject to cliticization.'' He notes that a Western Australian language (Warnman, see Nash 1996:120) innovated an invariable stem for independent pronouns, which was certainly not a pronoun historically. In their quantitative study of the Amerind personal pronouns, Nichols and Peterson (1996:345±346) observe that some of their sample languages have no distinct independent pronoun roots, but have independent pronouns consisting of ``a generic pronominal root (usually invariant across all person-number categories, and often etymologically a form of the verb `be' or a noun such as `body' or `self ') with nominal or verbal a½xes distinguishing the di¨erent number categories.'' Personal pronouns can also be created from nouns. Examples occur in Romance and Southeast Asian languages, as well as Japanese (Sugamoto 1989).19 Personal pronouns are also commonly created from demonstratives. As Greenberg (1986:xix) observes, It is further well known that many languages do not have a third person pronoun that is distinct from one or more demonstratives. The most common is a distance demonstrative which is also the most common source of purely anaphoric third person pronouns. This is a common diachronic process by which the demonstrative ®rst acquires anaphoric meaning in addition to its demonstrative meaning and then becomes con®ned to anaphoric usage.

à a, This appears to be the case for the nonhuman classes of pronouns in Chichew which are formally identical to the demonstratives. Nevertheless, they behave identically to the distinctive pronominal roots with respect to the topic-anaphoric properties investigated by Bresnan and Mchombo (1987). A similar phenomenon can be found in English. Trudgill and Chambers (1991:8) report that in East Anglian dialects of English it occurs only as an object pronoun, with third-person neuter singular subjects being indicated by that. (17) a. That's raining. b. I don't like itÐthat's no good. (East Anglian dialects of English, Trudgill and Chambers 1991:8)

Emergence of the Unmarked Pronoun

133

(18) A local woman who helps us clean the house here said to me the other day after a long search for the broom, which, like many other things is always being moved around the house by the kids, and had gone missing to be ®nally located down the side of the fridge, ``That's a good place for it. But as soon as you start saying something, that disappears.'' (Louisa Sadler, personal communication, May 6, 1997) In Standard English, demonstrative pronouns cannot be used in these anaphoric contexts. In East Anglian, that has the morphological form of a demonstrative but is functioning as a third-singular neuter personal pronoun. Thus it is incorrect to de®ne pronouns as having distinct pronominal stem forms. It is the synchronic functions of the element that determine its pronominality, not its etymology. Now the Salish languages of the American Northwest have both of these properties at once: special edge positioning for ®rst- and second-person pronouns (which are perhaps weak pronouns or ``clitics''), and a set of independent strong pronouns based on pronominal in¯ections of verbal roots (Hukari 1976; Jelinek and Demers 1994). Jelinek (1997:243) concludes on the basis of these distributional and morphological facts that Straits Salish lacks independent pronouns and remarks that the same holds for Winnebago, a Siouan language (see Lipkind 1945:29; Forchheimer 1953:88). But we now see that this conclusion cannot be drawn solely on the basis of narrowly categorial and morphological conceptions of pronominality, for a great variety of forms of expressions can function as pronouns. In the case of Winnebago, Lipkind (1945:53) writes of the demonstrative su½xes added to the positional verb stems for `sit', `lie', and `stand': ``These . . . form a set of words which function like independent personal pronouns, being used as subject or object or vocatively. Both the verb and the positional take personal pronouns [bound pronominals].'' Of the positional verbs he writes (1945:45): ``The positionals beside their verbal function are also used with demonstratives. . . . Long things are talked of as lying, tall or upright things as standing, and other things as sitting. Clouds `lie', rain `stands,' the sun and moon `sit.' '' In other words, these verbal elements serve as positional classi®ers to the demonstratives. We have, then, a pronominal form morphologically composed of in¯ections for person, a positional classi®er, and a demonstrative su½x. The content conveyed by this formÐnamely, person and kind classi®cation (based on positional type)Ðis fully consistent with pronominality. Finally, we note that Winnebago is a con®gurational language: syntactic position is ``the chief source of grammatical relations'' (Lipkind 1945:12). Thus the appearance of these freestanding person forms in subject and object positions suggests the syntactic function of pronominal arguments. Once again it is the synchronic functions of the element that must determine its pronominality, not its etymology.20

134

Joan Bresnan

In sum, the present theory of pronominal markedness can explain the existence of universal (or near-universal) asymmetries in the typology of pronominal systems, the preponderant use of free pronouns in pidgins of typologically diverse sources, and the emergence of free pronouns to ®ll gaps in the system of reduced pronouns within individual languages. These preliminary results suggest that it is indeed possible to develop a nonderivational theory of markedness for syntax analogous to the OT theories being developed in phonology, but of course much further work is required. Notes Many friends, colleagues, and students have made valuable suggestions that I have incorporated into this work. I owe special thanks to Jane Simpson and Anna Siewierska for ®rst confronting me with evidence of possible languages without free pronouns, to Judith Aissen for helpful comments and probing questions about an earlier draft, and to Paul Smolensky for advice and inspiration. I alone am responsible for any remaining defects in this work. The present version is based in part on work supported by the National Science Foundation under Grant No. BCS-9818077. 1. Explicit markedness proposals in generative syntax have mainly involved imposing preference structures on parameter settings (e.g., Chomsky 1981; Harbert 1986) or formally implementing the elsewhere principle in terms of feature underspeci®cation and rule ordering (as discussed in Bresnan, forthcoming b). See Battistella 1996; Moravcsik and Wirth 1986. 2. In syntax, transformational derivations may be represented in a single structure by coindexing constituents with chains of traces representing their earlier positions in the derivation of the structure. This move enables an OT-style typology of syntactic outputs while maintaining pre-OT assumptions about the inviolable correspondence between syntactic structures and theta roles (Bresnan, forthcoming c). 3. In what follows, the terms pronoun/pronominal and personal pronoun/pronominal will be understood to refer to de®nite personal pronouns/pronominals. Inde®nite pronouns are not included in the present study. 4. A signi®cant feature type omitted here is that of social level or distance; distinctions of social level or distance are elaborately marked by di¨erent pronominal forms in Javanese (Geertz 1960:248¨.), Balinese (Arka 1995), Thai, Burmese, and Vietnamese (Cooke 1968), and are also evident in the use of the Japanese and Korean pronominal systems. MuÈhlhaÈusler and Harre (1990:64) take the major pronominal contrasts to be (i) `` `person' and the features of participant roles'' and (ii) ``distance and proximity (obviative and proximative) both spatial and social.'' 5. As in Bresnan (forthcoming a,b), privative features have been used to represent speci®cation of content. For example, the feature top stands for a privative or monovalent feature, which has only a single (the ``marked'') value. (An opposed monovalent feature such as foc is also available; the opposition is captured by pragmatics rather than the formal opposition of G values.) Such features give rise to benign (``permanent,'' ``inherent,'' or ``trivial'') underspeci®cation in the sense of Steriade (1995). Thus, all the feature structures shown in ®gure 5.1 are possible for both input and output, requiring no further speci®cation or derivation. 6. F-structures can also be read as underspeci®ed semantic structures, either Quasi-Logical Forms or Underspeci®ed Discourse Representation Structures (Genabith and Crouch 1996).

Emergence of the Unmarked Pronoun

135

7. Purely spatial deictics need not carry such ``agr'' information, and a pronominal system derived from such pure deictics could in principle systematically lack these features. Sign languages might provide a case in point, though the status of pronominals in these languages appears controversial. I am grateful to Steve Wechsler for suggesting ASL to me as a possible source of pronouns without person features. 8. Irish is a possible counterexample to this generalization. Its free pronominal objects give indications of being weak pronouns (being unconjoinable without reinforcing morphology), and there are morphological means of strengthening them (McCloskey 1986); the same strengthening morphology also applies to the bound pronominal forms (McCloskey and Hale 1984). 9. Farrell (1990:327) observes, ``For many speakers, they are virtually never used in ordinary speech, although formal registers, particularly formal written registers, continue to demand their use.'' There are also formal restrictions on the use of the clitics (Chagas de Souza 1997). 10. Bresnan's constraints given in (i) are extensionally equivalent to the ®rst two syntactic markedness constraints in (5). (i) Bresnan's (forthcoming a) constraints: a. ProAgr: pro I agr b. jTop: top I j (``I'' designates material implication, and so (a) can be read as asserting that pro implies agr and (b) that top implies zero.) jTop marks nonzero pronominals, but only those that are specialized for topic anaphoricity and that therefore have the top feature. jTop does not mark the (neutral) pronoun, because it is assumed that this pronoun will not have the top property. Similarly, the function of ProAgr is to mark the zero pronominal, which, it is assumed, will always lack inherent agr properties. 11. See Bresnan 1998c on the speci®c relation of this hypothesis to pidginization. 12. Tableau T5.1 only shows candidates having the pro attribute, because we are setting aside (nonpronominal) agreement here. In e¨ect, we are assuming that Max(pro) is ranked above the other members of the Faith family. 13. Possible counterexamples are discussed in section 5.4. 14. There is also an interaction between zero pronouns and the person hierarchy, many languages having a zero pronoun only in the third person (Forchheimer 1953). 15. Japanese attaches constraints of social level to its pronominal system; use of an overt pronoun to designate a person implies social familiarity and is therefore avoided in many situations (Peter Sells and Yukiko Morimoto, personal communication, March 1997). For this reason, an inanimate overt pronoun is used in the examples given here. 16. These are also used as NP-agreement markers (Simpson 1991; Austin and Bresnan 1996), which can be represented by letting the pro value be optionally parsed (see Bresnan and Mchombo 1986, 1987). Several approaches to the representation of optionality that could be used here are discussed by Bresnan (forthcoming b); see also Asudeh, forthcoming. 17. Hale 1973 refers to the language as Warramunga; subsequent to the publication of Hale 1973, it became o½cially designated Warumungu. Hale's spelling has been adjusted accordingly in the quotation (see Hale 1973:340 n. 25).

136

Joan Bresnan

18. This is similar to recent proposals for edge alignment of clitics (Anderson 1996, forthcoming; Legendre, forthcoming), but is not restricted to the clitic forms of pronominals. On harmonic alignment of markedness hierarchies in syntax see Aissen (chap. 3, this volume). 19. Spanish and Portuguese polite second-person pronouns derive from honori®c titles such as Spanish Vuestra Merced and Portuguese Vossa Merce (``Your honor''), which became Usted and Voce(s) respectively; the third-person number agreement required with these forms is a remnant of their nominal origin (MuÈhlhaÈusler and Harre 1990:136±137). In Thai, deferential ®rst-person pronouns often have literal denotations for the head or related parts of the body such as crown or hair, while many deferential second-person pronouns denote the sole of the foot. The historical signi®cance is that ``the inferior speaker places the sole of his hearer's foot, or the dust beneath the foot, on a par with his own head or hairÐthe highest and most respected part of his body'' (Cooke 1968:9). Thai third-person pronouns often have kin-term sources. Burmese and Vietnamese also have nominal sources for many pronouns, although the personal pronouns can be formally distinguished from other nominals, and even in Thai, where ``personal pronouns are not so easily distinguished formally from kintype nouns'' except by limitations on modi®cation, they can be distinguished as ``a class apart'' by their semantic structuring (Cooke 1968:147). 20. Examination of texts is needed to determine the function of such pronominal forms in Winnebago, Straits Salish, and other languages where the presence of free pronouns has been doubted. References Anderson, Stephen R. 1996. How to Put Your Clitics in Their Place, or Why the Best Account of Second-Position Phenomena May Be Something Like the Optimal One. Linguistic Review 13: 165±191. Anderson, Stephen R. Forthcoming. Towards an Optimal Account of Second Position Phenomena. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Andrews, Avery D. 1990. Uni®cation and Morphological Blocking. Natural Language and Linguistic Theory 8: 507±557. Arka, I Wayan. 1995. Social Index and Agreements in Balinese Syntax. Paper presented at ALS'95, Canberra, Australia. Ms., Department of Linguistics, University of Sydney. Asudeh, Ash. Forthcoming. Linking and Optionality in Marathi: An Optimality Theory Analysis. In Peter Sells (ed.), Formal and Empirical Issues in Optimality Theoretic Syntax. Stanford, Calif.: CSLI Publications. Available at http://www.stanford.edu/@asudeh/. Austin, Peter, and Joan Bresnan. 1996. Noncon®gurationality in Australian Aboriginal Languages. Natural Language and Linguistic Theory 14: 215±268. Baker, Mark. 1993. Noun Incorporation and the Nature of Linguistic Representation. In William A. Foley (ed.), The Role of Theory in Language Description, 13±44. Trends in Linguistics, Studies and Monographs 69. Berlin: Mouton de Gruyter. Battistella, Edwin L. 1996. The Logic of Markedness. Oxford: Oxford University Press. BoÈrjars, Kersti, Nigel Vincent, and Carol Chapman. 1997. Paradigms, Periphrases and Pronominal In¯ection: A feature-Based Account. In Geert Booji and Jaap van Marle (eds.), The Yearbook of Morphology, 1±26. Dordrecht: Kluwer.

Emergence of the Unmarked Pronoun

137

Bresnan, Joan. 1998a. Markedness and Morphosyntactic Variation in Pronominal Systems. Paper presented at ``Is Syntax Di¨erent? Common Cognitive Structures for Syntax and Phonology in Optimality Theory,'' CSLI Workshop, Stanford University, December 12±13. Bresnan, Joan. 1998b. Morphology Competes with Syntax: Explaining Typological Variation in Weak Crossover E¨ects. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 59±92. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Bresnan, Joan. 1998c. Pidgin Genesis in Optimality Theory. In Miriam Butt and Tracy Holloway King (eds.), Proceedings of the LFG98 Conference, University of Queensland, Brisbane. Available at CSLI Publications: http://www.csli-publications.stanford.edu/LFG/3/lfg98.html. Ã a PronoBresnan, Joan. Forthcoming a. The Emergence of the Unmarked Pronoun: Chichew minals in Optimality Theory. Proceedings of the Twenty-Third Annual Meeting of the Berkeley Linguistics Society. Berkeley Linguistics Society, University of California, Berkeley. Bresnan, Joan. Forthcoming b. Explaining Morphosyntactic Competition. In Mark Baltin and Chris Collins (eds.), Handbook of Contemporary Syntactic Theory. Oxford: Blackwell. Bresnan, Joan. Forthcoming c. Lexical-Functional Syntax. Oxford: Blackwell. Bresnan, Joan. Forthcoming d. Optimal Syntax. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Bresnan, Joan, and Sam A. Mchombo. 1986. Grammatical and Anaphoric Agreement. In Anne M. Farley, Peter T. Farley, and Karl-Erik McCullough (eds.), CLS 22, Vol. 2, Papers from the Parasession on Pragmatics and Grammatical Theory, 278±297. Chicago Linguistic Society, University of Chicago. Ã a. Bresnan, Joan, and Sam A. Mchombo. 1987. Topic, Pronoun, and Agreement in Chichew Language 63: 741±782. Bresnan, Joan, and Lioba Moshi. 1990. Object Asymmetries in Comparative Bantu Syntax. Linguistic Inquiry 21: 147±185. Reprinted in Sam A. Mchombo (ed.), Theoretical Issues in Bantu Syntax, 47±91. Stanford, Calif.: CSLI Publications. Cardinaletti, Anna. 1995. Pronouns in Germanic and Romance Languages: An Overview. Ms., University of Venice. Cardinaletti, Anna, and Michael Starke. 1996. De®cient Pronouns: A View from GermanicÐ A Study in the Uni®ed Description of Germanic and Romance. In HoÈskuldur ThraÂinsson, Samuel David Epstein, and Steve Peter (eds.), Studies in Comparative Germanic Syntax II, 21± 65. Dordrecht: Kluwer. Carstairs-McCarthy, Andrew. 1992. Current Morphology. London: Routledge. Chagas de Souza, Paulo. 1997. Brazilian Portuguese Pronouns in Competition: An OptimalityTheoretic Approach. Ms., Department of Linguistics, Stanford University, Stanford, Calif. Chomsky, Noam. 1981. Markedness and Core Grammar. In Adriana Belletti, Luciana Brandi, and Luigi Rizzi (eds.), Theory of Markedness in Core Grammar. Pisa: Scuola Normale Supercore di Pisa. Cooke, Joseph R. 1968. Pronominal Reference in Thai, Burmese, and Vietnamese. Berkeley: University of California Press.

138

Joan Bresnan

Dalrymple, Mary, Ronald M. Kaplan, John T. Maxwell III, and Annie Zaenen (eds.). 1995. Formal Issues in Lexical-Functional Grammar. Stanford, Calif.: CSLI Publications. Demuth, Katherine, and Mark Johnson. 1989. Interaction between Discourse Functions and Agreement in Setawana. Journal of African Languages and Linguistics 11: 21±35. Dixon, R. M. W. 1980. The Languages of Australia. Cambridge, England: Cambridge University Press. Everett, Daniel L. 1996. Why There Are No Clitics: An Alternative Perspective on Pronominal Allomorphy. Arlington: Summer Institute of Linguistics and University of Texas, Arlington. Farrell, Patrick. 1990. Null Objects in Brazilian Portuguese. Natural Language and Linguistic Theory 8: 325±346. Forchheimer, Paul. 1953. The Category of Person in Language. Berlin: de Gruyter. Geertz, Cli¨ord. 1960. The Religion of Java. Glencoe, Ill.: Free Press. Genabith, Josef van, and Richard Crouch. 1996. F-structures, QLFs, and UDRSs. Proceedings of the First LFG Conference, Rank Xerox, Grenoble, France, August 26±28, 1996. Available at CSLI Publications: http://www.csli-publications.stanford.edu/LFG/1/lfg1.html. GivoÂn, Talmy. 1976. Topic, Pronoun, and Agreement. In Charles N. Li (ed.), Subject and Topic, 149±188. New York: Academic Press. GivoÂn, Talmy. 1983. Introduction. In Talmy GivoÂn (ed.), Topic Continuity in Discourse, 5±41. Amsterdam: Benjamins. GivoÂn, Talmy. 1984, 1990. Syntax: A Functional-Typological Introduction, 2 vols. Amsterdam: Benjamins. GivoÂn, Talmy. 1995. Functionalism and Grammar. Amsterdam: Benjamins. Greenberg, Joseph. 1986. Introduction: Some Re¯ections on Pronominal Systems. In Ursula Wiesemann (ed.), Pronominal Systems, xvii±xxi. TuÈbingen: Narr. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 193±219. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Haegeman, Liliane. 1996. Object Clitics in West Flemish. In Aaron L. Halpern and Arnold M. Zwicky (eds.), Approaching Second: Second Position Clitics and Related Phenomena, 135±164. Stanford, Calif.: CSLI Publications. Haiman, John. 1985. Natural Syntax: Iconicity and Erosion. Cambridge, England: Cambridge University Press. Hale, Kenneth. 1973. Person Marking in Walbiri. In Stephen Anderson and Paul Kiparsky (eds.), A Festschrift for Morris Halle, 308±344. New York: Holt, Rinehart and Winston. Harbert, Wayne. 1986. Markedness and the Bindability of Subject of NP. In Fred R. Eckman, Edith A. Moravscik, and Jessica R. Wirth (eds.), Markedness, 139±154. New York: Plenum Press. Hukari, Thomas E. 1976. Person in a Coast Salish Language. International Journal of American Linguistics 42: 305±318. Jakobson, Roman. 1984. Structure of the Russian Verb. In Linda R. Waugh and Morris Halle (eds.), Roman Jakobson: Russian and Slavic Grammar, Studies 1931±1981, 1±14. Berlin: Mouton.

Emergence of the Unmarked Pronoun

139

Jelinek, Eloise. 1984. Empty Categories and Noncon®gurational Languages. Natural Language and Linguistic Theory 2: 39±76. Jelinek, Eloise. 1988. The Case Split and Pronominal Arguments in Choctaw. In L. MaraÂcz and P. Muysken (eds.), Con®gurationality; The Typology of Asymmetries, 117±141. Dordrecht: Foris. Jelinek, Eloise. 1990. Grammatical Relations and Coindexing in Inverse Systems. In Katarzyna Dziwirek, Patrick Farrell, and Errapel MejõÂas-Bikandi (eds.), Grammatical Relations: A Cross-Theoretical Perspective, 227±246. Stanford, Calif.: CSLI Publications. Jelinek, Eloise. 1995. Quanti®cation in Straits Salish. In Emmon Bach, Eloise Jelinek, Angelika Kratzer, and Barbara H. Partee (eds.), Quanti®cation in Natural Languages, 487± 540. Dordrecht: Kluwer. Jelinek, Eloise. 1997. Topic and focus in Navajo inverse. In Emily Curtis, James Lyle, and Gabriel Webster (eds.), Proceedings of the Sixteenth West Coast Conference on Formal Linguistics 15: 241±255. Jelinek, Eloise, and Richard A. Demers. 1994. Predicates and Pronominal Arguments in Straits Salish. Language 70: 697±736. Johnson, Mark. 1998. Optimality-Theoretic Lexical Functional Grammar. Paper presented at the Eleventh Annual CUNY Conference on Human Sentence Processing, Special Session on the Lexical Basis of Syntactic Processing: Formal and Computational Issues, Rutgers University, New Brunswick, N.J., March 20. Kameyama, Megumi. 1985. Zero Anaphora: The Case of Japanese. Doctoral dissertation, Stanford University, Stanford, Calif. Kaplan, Ronald M., and Joan Bresnan. [1982] 1995. Lexical-Functional Grammar: A Formal System for Grammatical Representation. In Mary Dalrymple, Ronald M. Kaplan, John T. Maxwell III, and Annie Zaenen (eds.), Formal Issues in Lexical-Functional Grammar, 29±130. Stanford, Calif.: CSLI Publications. [Reprented from Joan Bresnan (ed.), The Mental Representation of Grammatical Relations. Cambridge, Mass.: MIT Press.] Kroeger, Paul. 1993. Phrase Structure and Grammatical Relations in Tagalog. Stanford, Calif.: CSLI Publications. Kuhn, Jonas. 1999. Two Ways of Formalizing OT Syntax in the LFG Framework. Ms. Available at Institut fuÈr maschinelle Sprachverarbeitung, UniversitaÈt Stuttgart: http:// www.ims.uni-stuttgart.de/@jonas/. Lambrecht, Knud. 1981. Topic, Anti-Topic, and Verb Agreement in Non-Standard French. Amsterdam: Benjamins. Lambrecht, Knud, and Kevin Lemoine. 1996. Vers une grammaire des compleÂments zeÂro en FrancËais parleÂ. In Jean Chuquet and Mare Fryd (eds.), Travaux Linguistiques du CERLICO. 9. Absence de Marques et RepreÂsentation de l'Absence, 279±309. Rennes: Presses Universitaire de Rennes et CERLICO. Legendre, GeÂraldine. Forthcoming. Morphological and Prosodic Alignment of Bulgarian Clitics. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Lipkind, William. 1945. Winnebago Grammar. Morningside Heights, N.Y.: King's Crown Press.

140

Joan Bresnan

McCarthy, John J., and Alan S. Prince. 1995. Faithfulness and Reduplicative Identity. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 249±384. GLSA, University of Massachusetts, Amherst. McCloskey, James. 1986. In¯ection and Conjunction in Modern Irish. Natural Language and Linguistic Theory 4: 245±281. McCloskey, James, and Kenneth Hale. 1984. On the Syntax of Person-Number In¯ection in Modern Irish. Natural Language and Linguistic Theory 1: 487±534. Mereu, Lunella. 1997. For a Lexical Functional Representation of Agreement A½xes and Clitics. In Miriam Butt and Tracy Holloway King (eds.), Proceedings of the LFG97 Conference, University of California, San Diego. Available at CSLI Publications: http:// www.csli-publications.stanford.edu/LFG/2/lfg97.html. Mohanan, K. P. 1982. Grammatical Relations and Clause Structure in Malayalam. In Joan Bresnan (ed.), The Mental Representation of Grammatical Relations. Cambridge, Mass.: MIT Press. Moravcsik, Edith, and Jessica Wirth. 1986. Markedness: An Overview. In Fred R. Eckman, Edith A. Moravcsik, and Jessica R. Wirth (eds.), Markedness, 1±11. New York: Plenum Press. MuÈhlhaÈusler, Peter, and Ron HarreÂ. 1990. Pronouns and People: The Linguistic Construction of Social and Personal Identity. Oxford: Blackwell. Naro, Anthony. 1973. The Origin of West African Pidgin. In Claudia Corum, T. Cedric SmithStark, and Ann Weiser (eds.), Papers from the Ninth Regional Meeting of the Chicago Linguistic Society, 442±449. Chicago Linguistic Society, University of Chicago. Nash, David. 1996. Pronominal Clitic Variation in the Yapa Languages. In William McGregor (ed.), Studies in Kimberley Languages in Honour of Howard Coate, 117±138. Munich: Lincom Europa. Nespor, Marina. 1994. The Phonology of Clitic Groups. In Lars Hellan and Henk van Riemsdijk (eds.), Clitic Doubling and Clitic Groups, 5, 67±90. ESF±Eurotyp Working Papers. Department of Language and Linguistics, Tilburg University, The Netherlands. Nichols, Johanna. 1986. Head-Marking and Dependent-Marking Grammar. Language 62: 56± 119. Nichols, Johanna, and David A. Peterson. 1996. The Amerind Personal Pronouns. Language 72: 336±371. Nichols, Lynn. 1997. Logical Structure and Pronominal Movement in Zuni. In Emily Curtis, James Lyle, and Gabriel Webster (eds.), Proceedings of the Sixteenth West Coast Conference on Formal Linguistics 15: 369±380. Nordlinger, Rachel. 1998. Constructive Case: Evidence from Australian Languages. Stanford, Calif.: CSLI Publications. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Sadock, Jerrold. 1991. Autolexical Syntax. Chicago: University of Chicago Press.

Emergence of the Unmarked Pronoun

141

Samek-Lodovici, Vieri. 1996. Constraints on Subjects: An Optimality Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Sandoval, Merton, and Eloise Jelinek. 1989. The bi-Construction and Pronominal Arguments in Apachean. In Eung-Do Cook and Keren D. Rice (eds.), Athapaskan Linguistics: Current Perspectives on a Language Family, 335±377. Trends in Linguistics, State-of-the-Art Reports 15, ed. Werner Winter. Berlin: Mouton de Gruyter. Schwartz, Linda J. 1986. The Function of Free Pronouns. In Ursula Wiesemann (ed.), Pronominal Systems, 405±436. TuÈbingen: Narr. Schwartz, Linda J., and Timothy Dunnigan. 1986. Pronouns and Pronominal Categories in Southwestern Ojibwe. In Ursula Wiesemann (ed.), Pronominal Systems, 285±322. TuÈbingen: Narr. Sells, Peter. 1998. Scandinavian Clause Structure and Object Shift. In Miriam Butt and Tracy Holloway King (eds.), Proceedings of the LFG98 Conference, University of Queensland, Brisbane. Available at CSLI Publications: http://www.csli-publications.stanford.edu/LFG/3/ lfg98.html. Siewierska, Anna. 1999. Reduced Pronominals and Argument Prominence. In Miriam Butt and Tracy Holloway King (eds.), Proceedings of the LFG99 Conference, University of Manchester. Available at CSLI Publications: http://www.csli-publications.stanford.edu/LFG/4/ lfg99.html. Simpson, Jane. 1983. Aspects of Warlpiri Morphology and Syntax. Doctoral dissertation, MIT, Cambridge, Mass. Simpson, Jane. 1991. Warlpiri Morpho-Syntax: A Lexicalist Approach. Dordrecht: Reidel. Simpson, Jane. 1998. Some Features of Warumungu Morphology. In Andrew Spencer and Arnold Zwicky (eds.), Handbook of Morphology, 707±736. Oxford: Blackwell. Simpson, Jane, and Je¨rey Heath. 1982. Warumungu Sketch Grammar. Ms., Australian Institute of Aboriginal and Torres Strait Islander Studies, Canberra. Smolensky, Paul. 1996. The Initial State and ``Richness of the Base'' in Optimality Theory. Technical Report JHU-CogSci-96-4. Department of Cognitive Science, Johns Hopkins University, Baltimore. Speas, Margaret. 1990. Phrase Structure and Natural Language. Dordrecht: Kluwer. Speas, Margaret. 1997. Optimality Theory and Syntax: Null Pronouns and Control. In Diana Archangeli and D. Terence Langendoen (eds.), Optimality Theory: An Overview, 171±199. Oxford: Blackwell. Steriade, Donca. 1995. Underspeci®cation and Markedness. In John Goldsmith (ed.), Handbook of Phonological Theory, 114±174. Oxford: Blackwell. Sugamoto, Nobuko. 1989. Pronominality: A Noun-Pronoun Continuum. In Roberta Corrigan, Fred Eckman, and Michael Noonan (eds.), Linguistic Categorization, 267±291. Amsterdam: Benjamins. SunÄer, Margarita. 1988. The Role of Agreement in Clitic-Doubled Constructions. Natural Language and Linguistic Theory 6: 391±434. Tesar, Bruce, and Paul Smolensky. 1998. The Learnability of Optimality Theory: An Algorithm and Some Basic Complexity Results. Linguistic Inquiry 29: 229±268.

142

Joan Bresnan

Toivonen, Ida. 1996. Finnish Possessive Su½xes in Lexical-Functional Grammar. Proceedings of the First LFG Conference, Rank Xerox, Grenoble, France, August 26±28. Available at CSLI Publications: http://www.csli-publications.stanford.edu/LFG/1/lfg1.html. Toivonen, Ida. 1997. The Possessive Su½xes in Finnish. In Lise M. Dobrini, Lynn Nichols, and Rosa M. Rodriguez (eds.), CLS 27, Vol. 1, The General Session, 389±400. Chicago Linguistic Society, University of Chicago. Trudgill, Peter, and J. K. Chambers (eds.). 1991. Dialects of English: Studies in Grammatical Variation. London: Longman. Uyechi, Linda. 1991. The Functional Structure of the Navajo Third Person Alternation. In Lise M. Dobrini, Lynn Nichols, and Rosa M. Rodriguez (eds.), CLS 27, Vol. 1, The General Session, 434±446. Chicago Linguistic Society, University of Chicago. Van Valin, Robert. 1996. Assumptions and Principles of RRG. Handout. Nijmegen: MaxPlanck-Institut fuÈr Psycholinguistik Conference on Future Developments in Linguistic Theory, June 28. Wiesemann, Ursula (ed.). 1986. Pronominal Systems. TuÈbingen: Narr. Willie, Mary Ann. 1989. Why There Is Nothing Missing in Navajo Relative Clauses. In EungDo Cook and Keren D. Rice (eds.), Athapaskan Linguistics: Current Perspectives on a Language Family, 407±437. Trends in Linguistics, State-of-the-Art Reports 15, ed. Werner Winter. Berlin: Mouton de Gruyter. Zipf, George. 1935. The Psychobiology of Language. Boston: Houghton Mi¿in. Zwicky, Arnold. 1977. On Clitics. Bloomington: Indiana University Linguistics Club. Zwicky, Arnold. 1985. Clitics and Particles. Language 71: 283±305.

Chapter 6 Binding and Discourse Prominence: Reconstruction in ``Focus'' Scrambling

Hye-Won Choi

This chapter examines the ``mixed'' binding property of scrambling in German and the close connection of the reconstruction e¨ect (A0 -property) in binding with a certain ``focality'' restriction. Instead of assuming the A/A0 or operator/nonoperator distinction in movement typology, which has been a dominant approach to the split binding problem of scrambling in the literature, this chapter derives the mixed binding behavior and also the so-called A0 -scrambling's association with focality through Optimality-Theoretic (OT; Prince and Smolensky 1993) interactions among the several competing constraints in syntax and discourse. This new approach provides a solution to the unnecessary complications in movement theory in relation to scrambling, and it also expands the empirical coverage of binding and focusing phenomena by separating focusing from A0 -scrambling. The organization of the chapter is as follows. I ®rst describe the binding and focality facts associated with scrambling in German and brie¯y discuss previous analyses in section 6.1. Then, in section 6.2, I introduce a theory of information structure as the basis for the focality problem. In section 6.3, I propose operator binding constraints and their particular ranking in German and then derive the binding and focality facts through OT constraint competitions between the two components of grammar, syntax and discourseÐthat is, between the binding and phrase-structural constraints in syntax and the information-structural constraints in discourse. Finally, section 6.4 concludes the chapter. 6.1

Mixed Binding and Focality E¨ect

Scrambling shows a mixed property (A- and A0 -) with respect to binding. Notably, the scrambling that has the A0 -binding property is associated with some focality e¨ect. The relevant facts and problems in previous analyses will be discussed in this section.

144

Hye-Won Choi

6.1.1 Facts Binding facts are often used as a diagnostic for the A/A0 distinction. That is, if a movement creates a new binding relation (i.e., feeds a new binding relation or bleeds the existing binding), it is argued to be an A-movement because a typical A-movement like the raising or passive construction in English shows the antireconstruction e¨ect. This is illustrated in (1): the subject raising in (1b) bleeds the binding that obtains in (1a). On the other hand, if a movement maintains the existing binding relation, it is considered an A0 -movement because an A0 -moved element, as in the wh-question or topicalization in English, reconstructs to the base position with respect to binding. This is shown in (2): the wh-movement in (2b) does not create a new binding relation but retains the binding as if the wh-word were in the base position, and thus shows the reconstruction e¨ect.1 (1) a. It seems to every boyi that hisi mother is intelligent. b. *Hisi mother seems to every boyi to be intelligent. (2) a. *Hisi mother loves every boyi . b. *Whoi does hisi mother love? Scrambling exhibits a ``mixed'' behavior with respect to binding in this sense of A/A0 distinction (Bayer and Korn®lt 1994; Dayal 1994; Deprez 1994; Frank, Lee, and Rambow 1992; Lee 1993; Lee and Santorini 1994; Mahajan 1990; Miyagawa 1995; MuÈller and Sternefeld 1993; Neeleman 1994; Saito 1992; Webelhuth 1992; Williams 1994). That is, scrambling sometimes shows the A-property by creating a new binding relation, which does not obtain in the base position. Scrambling also shows the A0 -property in that it maintains the binding relation that obtains in the base position and thus shows the reconstruction e¨ect. The basic facts are illustrated in the German examples (3) and (4).2 Note that only embedded clauses are exempli®ed here to avoid any confusion due to the topicalization and the verb-second e¨ect in the main clauses of German. (3) a. *daû seinei Mutter jedeni that his mother(Nom) everybody(Acc) `that hisi mother likes everybodyi ' b. daû jedeni seinei Mutter that everybody(Acc) his mother(Nom) `that hisi mother likes everybodyi '

mag likes

(4) a.

mag likes

b.

daû jederi seinei Mutter that everybody(Nom) his mother(Acc) `that everybodyi likes hisi mother' daû seinei Mutter jederi that his mother(Acc) everybody(Nom) `that everybodyi likes hisi mother'

mag likes

mag likes

Binding and Discourse Prominence

145

In (3) the quanti®er binder is the object jeden and the pronominal bindee is part of the subject seine Mutter. In the nonscrambled example (3a), the binding does not hold between the quanti®er jeden and the pronominal seine. However, the scrambling of the quanti®er in (3b) feeds a new binding. Compare this example with the weak crossover case in English, illustrated in (2). Unlike the English example, the scrambling in the German example (3b) creates a new binding, making the sentence grammatical. In contrast, in the examples in (4), where the quanti®er binder is the subject jeder and the pronominal bindee is in the object seine Mutter, the binding relation that holds in the nonscrambled sentence in (4a) does not change despite the scrambling of the pronominal phrase seine Mutter in (4b). Scrambling does not bleed binding, and thus binding holds as if the pronoun phrase were in the base position (reconstruction e¨ect). Interestingly, it has often been noted that the so-called reconstruction e¨ect of binding in scrambled sentences, as shown in (4b), is always associated with a certain focality restriction (Bayer and Korn®lt 1994; Neeleman 1994; Finer 1994; Miyagawa 1995). That is, (4b) is grammatical only when seine Mutter bears a high pitch accent, and especially when it is interpreted contrastively.3 Examples (5) and (6) are contexts where (4b) can be used.4 The pitch-accented elements are italicized below. (5) a. Wen glaubst du mag jeder? `Who do you think everybody likes?' b. Ich glaube daû seinei Mutter jederi mag (aber seinen Vater nicht jeder mag). `I think that everybody likes his mother, but not everybody likes his father.' (6) a. Weri glaubst du mag seinei Mutter? `Whoi do you think likes hisi mother?' b. Ich glaube daû seinei Mutter jederi mag (aber seinen Vater nicht jeder mag). `I think that everybody likes his mother, but not everybody likes his father.' In both (5) and (6), seine Mutter `his mother' gets more attention than other elements of the sentence, which is reinforced by the contrast with seinen Vater `his father', and thus gets a high pitch accent. It should be noted here that although it is often called focal or even focus because it bears a stress or pitch accent, the informational status of seine Mutter in (5b) is not equivalent to that in (6b). The former is truly focusÐ that is, the informative part or new informationÐwhereas the latter is given or old information, in that it is not new but mentioned in the previous sentence already. Yet it seems true that (5b) and (6b) share a certain discourse property. This informational property will be further investigated in section 6.2. It is notable that the grammaticality of (4b) contrasts with the ungrammaticality of (3a) in this sense of focality. See the contrast between (7a) and (7b) below. Even a contrastive focality on seine Mutter cannot save (7a) (which is also (3a)) from ungrammaticality.

146

Hye-Won Choi

(7) a. *daû seinei Mutter jedeni mag b. daû seinei Mutter jederi mag

ˆ (3a) ˆ (4b)

On the other hand, this focality e¨ect in the scrambling in (4b) is also in contrast with the lack of such an e¨ect in the scrambling in (3b). (8) a. daû jeden seine Mutter mag b. daû jeden seine Mutter mag c. daû jeden seine Mutter mag

ˆ (3b)

The scrambling of jeden in (3b) is possible irrespective of its focalityÐthat is, whether or not it bears the focal accent. Only this scrambled order and not the nonscrambled order in (3a) enables this binding to hold. This is the reason that the scrambling in (3b) is often distinguished from the scrambling in (4b) (Bayer and Korn®lt 1994; Neeleman 1994; Finer 1994; Miyagawa 1995): the former is called neutral scrambling because it is unrestricted by the focality constraint and the latter called focus scrambling because it is focally constrained (Neeleman 1994). 6.1.2 Previous Analyses Now, let me brie¯y discuss some of the previous analyses of the mixed binding problem introduced above. Most of them crucially rely on the A/A0 or operator/ nonoperator distinction. This dichotomy has caused trouble in explaining the mixed binding property of scrambling. Some have argued that scrambling is neither A- nor A0 -movement, but rather a movement to a third position with mixed properties (Webelhuth 1992; Deprez 1994; Dayal 1994; Ueyama 1994). Others have argued that scrambling is a nonoperator A0 -movement, while wh-movement, for example, is a regular operator A0 -movement (Saito 1992; Frank, Lee, and Rambow 1992; Lee and Santorini 1994). Also, Bayer and Korn®lt (1994), Neeleman (1994), Finer (1994), and Miyagawa (1995) have argued that scrambling is not a uniform phenomenon but that there are actually two di¨erent types of scrambling: one is A-scrambling, which is motivated by case/agreement features;5 the other is A0 -movement scrambling, which is motivated by the need to license the [Focus] operator and it is argued that that is why the reconstruction is associated with foucs. All these approaches face a problem when it comes to wh-movement and topicalization. In German, wh-movement and topicalization, which are typical operator A0 -movements, also exhibit the similar mixed binding behavior (Moltmann 1990:20± 21). This is illustrated in (9) and (10). (9)

a. Weni gebissen? hat seini Hund who(Acc) has his dog(Nom) bitten `Whoi did hisi dog bite?'

Binding and Discourse Prominence

147

b. Jedeni sah seinei Mutter. everybody(Acc) saw his mother(Nom) `Everybodyi , hisi mother saw.' (10) Seinei Mutter liebt jederi . his mother(Acc) loves everybody(Nom) `His motheri , everybodyi loves.' Example (9) exhibits the so-called A-property of wh-movement and topicalization (the antireconstruction e¨ect) and example (10) demonstrates the A0 -property of topicalization (the reconstruction e¨ect). This is exactly what we have observed in the scrambling in (3) and (4). To account for this, one would have to say that the operator movement can also be A-movement as well as A0 -movement. Then, the rationale behind this dichotomous movement classi®cation no longer holds. Also, the parallelism between scrambling and the operator A0 -movement remains unexplained. The approach that treats A0 - and A-scrambling separately faces another problem. In that approach, [Focus] is an operator licensed by the A0 -scrambling. As a matter of fact, however, [Focus] can also be licensed by the ``A-scrambling.'' The ``Ascrambled'' sentence in (3b) can be used in the following ``focal'' context. See (11). (11) a. Weni glaubst du mag seinei Mutter? `Whoi do you think hisi mother likes?' b. Ich glaube daû jedeni seinei Mutter mag. `I think that hisi mother likes everybodyi .' In this context, the scrambled binder jeden is the focal elementÐthat is, [Focus]Ð and thus gets the focal accent. According to the two-scrambling approach, this should be a case of focus scrambling, and therefore must show the A0 -binding property, but (11b) shows the antireconstruction e¨ect. In short, focality does not seem to be an inherent property of the A0 -scrambling. I will argue in the sections to follow that this special focality e¨ect in operator binding of ``focus'' scrambling or the lack of it in ``neutral'' scrambling can be derived with no complications by the OT constraint competition between syntax and discourse pragmatics. Let us ®rst see in section 6.2 how syntax interacts with discourse pragmatics in scrambling in general and then in section 6.3 investigate how the binding facts relate to that general interaction. 6.2

Discourse Competes with Syntax in Scrambling

Independently from the binding facts, Choi (1996, 1999) showed a close connection between syntax and discourse pragmatics in explaining the various topic/focus-

148

Hye-Won Choi

Table 6.1 ‡Prom New ‡New

Prom

Topic, Link

Tail

Contrastive Focus

Completive/Presentational Focus

related e¨ects that have been discussed in Lenerz 1977, Abraham 1986, Moltmann 1990, and Webelhuth 1992. It was argued that scrambling is an ``interface'' phenomenon that results from the OT constraint competitions between various syntactic and discourse constraints. Let us here review the proposal brie¯y, since the constraints introduced in Choi 1996, 1999 also play a crucial role in the current binding problems. 6.2.1 Information Structure Information structure has been proposed as the domain of representation of discourse information such as topic and focus (Lambrecht 1994; Vallduvõ 1992). Choi (1996, 1999) proposes a feature-based information structure as in table 6.1 to capture the crosscutting properties among the di¨erent informational notions introduced in the literature such as topic or link, tail, completive or presentational focus, and contrastive focus (Dik et al. 1981; Rochemont and Culicover 1990; Vallduvõ 1992, 1993; Lambrecht 1994; among others).6 Roughly speaking, a sentence can be divided into the background or given information ([ New]) and the informative or new information ([‡New]). Some given information (e.g., topic) is more important or prominent and hence receives more attention ([‡Prom]) than the other part of the given information (e.g., tail). Likewise, some new information or focus (e.g., contrastive focus) is more urgent or prominent and thus attracts more attention ([‡Prom]) than the ``purely new'' type of information, which simply completes or ®lls in the informational gap (e.g., completive focus). Importantly, topic ([ New,‡Prom]) and contrastive focus ([‡New,‡Prom]) share the prominence feature [‡Prom] in this system, although contrastive focus is new information ([‡New]) while topic is given information ([ New]). (See Choi 1996: chap. 3 for motivations and discussions of the proposal.) In OT syntax, the input usually contains the information indicating predicateargument relation. In this analysis, the input also includes the information-structural information such as [New] and [Prom] as illustrated above.7 An example of input is illustrated in (13) for the sentence given in (12b) in the following discourse context. (Focus [‡New] is underlined in the following examples.) (12) a. Wie steht's mit Hans? Wen glaubst du mag Hans? `What about Hans? Who do you think Hans likes?'

Binding and Discourse Prominence

149

b. (Ich glaube daû) Hans Maria mag. `(I think that) Hans likes Maria.' (13) Input: Hans [Subj;

New,‡Prom]; Maria [Obj; ‡New, Prom]

In the minidiscourse of question-and-answer pair in (12), Hans is presented as the ``topic'' of the sentence. It is the ``center of interest'' in the current discourse, that is, prominent given information, especially with the help of the linguistic clueÐthe ``what about?'' phrase. Therefore, it is marked [ New,‡Prom] in the input. Maria, on the other hand, is the new information, hence [‡New], but with no other linguistic or nonlinguistic clues, its main function is to complete the informational gap that (12a) calls for, not carrying any extra discourse information such as emphasis, counterexpectation, or contrastiveness. Hence, it is not prominentÐthat is, [ Prom].8 In fact, the new or informative part can be presented prominently also. See the context in (14). (14) a. Wen glaubst du mag Hans? `Who do you think Hans likes?' b. (Ich glaube daû) Maria Hans mag. `(I think that) Hans likes Maria!' Here, Maria not only ®lls the informational gap ([‡New]), but also carries some extra information such as ``it is surprising that Hans likes Maria,'' or ``it is Maria (and nobody else) that Hans likes,'' and so on. This way, Maria is marked [‡Prom] in the input. Unlike the situation in (12), Hans is not the center of interest any more but simply given information (i.e., tail), and therefore receives the marking, [ New, Prom]. This information is speci®ed in the input as in (15). (15) Input: Hans [Subj;

New, Prom]; Maria [Obj; ‡New,‡Prom]

Now, let us recall the two ``focal'' contexts in (5) and (6) that allow the scrambled pronominal phrase to be bound by the leftover operator binder. Let me repeat them here as (16) and (17). (16) a. Wen glaubst du mag jeder? `Who do you think everybody likes?' b. Ich glaube daû seinei Mutter jederi mag (aber seinenj Vater nicht jederj mag). `I think that everybodyi likes hisi mother, but not everybodyj likes hisj father.' (17) a. Weri glaubst du mag seinei Mutter? `Whoi do you think likes hisi mother?'

150

Hye-Won Choi

b. Ich glaube daû seinei Mutter jederi mag (aber seinenj Vater nicht jederj mag). `I think that everybodyi likes hisi mother, but not everybodyj likes hisj father.' Seine Mutter in (16b) is ``focus.'' It is new information since it corresponds to the question wen?. Thus, it is [‡New] in the current system of information structure. In addition, it is prominent information. It not only completes the informational gap but also carries extra informationÐfor example, contrastiveness in this caseÐand thus gets more attention ([‡Prom]) than the regular new information does. On the other hand, seine Mutter in (17b) is not ``focus.'' It is given information ([ New]), which was already introduced in the previous sentence (17a). But it is prominent old information ([‡Prom])Ðthat is, the topic of the sentence. In other words, the nature of the information that the scrambled pronominal phrase in (16) and (17) carries is discourse prominence ([‡Prom]) but not discourse newness, which the term ``focus scrambling'' often misleadingly implies. In terms of pitch accent, a discourseprominent element as well as a discourse-new one receives a high pitch accent or is stressed. So in the current system, [‡New] and also [‡Prom] receive stress. In other words, only regular old information ([ New, Prom]) does not get any accent or is destressed. Now, setting aside the binding problem for the time being, let us ®rst see how the discourse-contextual aspect of a sentence, which is instantiated as ``information structure,'' is re¯ected in terms of word order in the scrambling of German within the current OT framework. 6.2.2 Syntax/Discourse Interface Given the input speci®ed as above, candidates compete with each other; the candidate that instantiates the syntactic and discourse-contextual information in the input in the best way possible would be the optimal candidate. The most relevant candidates in the scrambling case are the sentences in di¨erent word orders as in (18).9 (18) Candidates a. (Ich glaube daû) Hans Maria mag. b. (Ich glaube daû) Maria Hans mag. These candidates are evaluated against syntactic and discourse constraints. Choi (1996) proposed the following phrase structure and information structure constraints relevant to scrambling, which are given in (19) and (20) respectively. (19) Phrase structure constraint Subj: Subject should precede nonsubjects. (Subject should c-command nonsubjects.)

Binding and Discourse Prominence

151

(20) Information structure constraints New: [ New] should precede [‡New]. Prom: [‡Prom] should precede [ Prom]. The phrase structure constraint Subj (Canon1 in Choi 1996) requires that subject precede nonsubjects such as object, or that subject c-command nonsubjects provided that German has hierarchical phrase structure.10 This syntactic constraint demands that the constituents be aligned according to their syntactic information and thus be in the so-called canonical or default order where subject precedes object. On the other hand, the information structure constraints New and Prom require that the constituents be aligned according to their informational status. First, New demands that given or old information ([ New]) should precede new or focused information ([‡New]).11 This constraint is rather familiar and has existed in the literature (Praguean and functionalist) in many di¨erent forms. The next constraint, Prom, requires that prominent information, whether topic or contrastive focus, should precede nonprominent information. In short, it says that important and urgent information should be placed earlier in the sentence (GivoÂn 1989). Together with New, Prom demands that topic ([ New,‡Prom]), for example, be sentence initial (cf. TopFirst in Costa, chap. 7, this volume). These constraints potentially con¯ict. The syntactic constraint Subj demands that the subject should precede the object, but the subject would not necessarily be [ New] or [‡Prom], so it may then be the case that subject before object would violate New and/or Prom. The potential con¯icts between these syntactic and discourse constraints are resolved by constraint ranking. I propose the following ranking for German in (21). (21) Constraint ranking in German (to be revised) Prom g Subj g New Given a ranking, each surface word order is taken to be the ``optimal'' output that satis®es these syntactic and discourse constraints in the best way possible within the given informational context.12 Now, let us see how this ranking is established.13 As a matter of fact, the syntactic and discourse constraints are not always in con¯ict. One such case is when a sentence is uttered in a context where every constituent is ``equal'' in its informational status. Suppose it is an answer to a question like ``What happened?'' or ``What's the matter?'' In this case, the whole reply is delivered as new information, [‡New]. In other words, the sentence is not informationally partitioned in a sense, so that each constituent does not get the individual informational marking in the input. Or a sentence can be uttered with no discourse-contextual information at all. It could be uttered out of the blue. In this ``neutral context'' case also, the input would not contain the information-structural information but only the

152

Hye-Won Choi

Tableau T6.1 H[S;

];M[O;

]

Prom

Subj

New

f a. Hans Maria mag b. Maria Hans mag

*!

Tableau T6.2 H[S; N‡P];M[O;‡N P]

Prom

Subj

New

*!

*

*

f a. Hans Maria mag b. Maria Hans mag

syntactic information. See tableau T6.1. In this case, all the candidates vacuously satisfy the information structure constraints Prom and New, so that only the syntactic constraint Subj becomes relevant. Therefore, as illustrated in tableau T6.1, the nonscrambled candidate (a), which does not violate Subj and thus is in the so-called canonical, default, or unmarked order, is the winning candidate in this context. Another example of a ``no con¯ict'' case is the one where syntax and discourse have a ``shared'' interest. A typical case is the subject-topic context. Recall example (12) earlier, repeated in (22). Here, Hans the subject is presented as topic [ New,‡Prom], while Maria the object as (completive) focus [‡New, Prom]. See tableau T6.2. (22) a. Wie steht's mit Hans? Wen glaubst du mag Hans? `What about Hans? Who do you think Hans likes?' b. Ich glaube daû Hans Maria mag. `I think that Hans likes Maria.' Prom requires [‡Prom] to precede [ Prom] so candidate (b), in which the topic Hans is preceded by Maria, violates this constraint. Now, Subj requires subject to precede object, and again candidate (b) violates it because the subject Hans is preceded by the object Maria. Finally, New, which demands [‡New] to follow [ New], is also violated by candidate (b) because there Maria, the focus, [‡New], precedes Hans, [ New]. In short, all the relevant constraints unanimously require the topical subject Hans to precede the focused object Maria, so any ranking of the constraints would result in the same output. However, constraints do come in con¯ict in some contexts and then the ranking becomes crucial. Let's consider a context in which subject is the focus, as in (23).

Binding and Discourse Prominence

153

Tableau T6.3 H[S;‡N P];M[O; N P]

Prom

Subj

f a. Hans Maria mag

New *

b. Maria Hans mag

*!

Tableau T6.4 H[S;‡N P];M[O; N‡P] a. Hans Maria mag

Prom

Subj

*!

f b. Maria Hans mag

New *

*

(23) a. Wer glaubst du mag Maria? `Who do you think likes Maria?' b. I glaube daû Hans Maria mag. `I think Hans likes Maria.' Suppose that neither Hans nor Maria is presented prominently in this context: Hans is purely new-information focus, [‡New, Prom], and Maria is simply old information, [ New, Prom]. Then, Prom is vacuously satis®ed by both the candidates in tableau T6.3. However, Subj and New are con¯icting: candidate (a) satis®es Subj but violates New since the [‡New] subject precedes the [ New] object there; candidate (b) violates Subj but satis®es New since now the [ New] object precedes the [‡New] subject. As shown in (23b), the nonscrambled order is used in this contextÐthat is, it is the optimal output. This means that Subj is ranked higher than New in German (Subj g New). Next, consider a context where the object of a sentence is the topic ([ New,‡Prom]). (24) a. Wie steht's mit Maria? Wer glaubst du mag Maria? `What about Maria? Who do you think likes Maria?' b. Ich glaube daû Maria Hans mag. `I think that Hans likes Maria.' Note that in this context, the scrambled order (i.e., candidate (b)) is used. That is, the topic Maria ([ New,‡Prom]) is scrambled over the subject Hans ([‡New, Prom]). See tableau T6.4. We have seen that Subj is ranked higher than New. As shown in tableau T6.4, candidate (a) is preferred by Subj. In contrast, candidate (b) is pre-

154

Hye-Won Choi

Tableau T6.5 H[S; N P];M[O;‡N‡P] a. Hans Maria mag f b. Maria Hans mag

Prom

Subj

New

*

*

*!

ferred by Prom because there the [‡Prom] Maria is scrambled to the front over the [ Prom] Hans. In other words, Prom and Subj are con¯icting now, and candidate (b), which satis®es Prom, is selected as the optimal output. As a result, we get the ranking of Prom g Subj g New. The above two examples have shown that Subj g New and Prom g Subj. This means that Prom is also ranked higher than New. This ranking predicts that if we have a context where Prom is con¯icting with New, the candidate that violates Prom would be defeated by the candidate that violates New. This is indeed the case. Contrastive focus ([‡New,‡Prom]) is an example of this sort. Prom would put it ®rst because it is [‡Prom], but New would put it later because it is [‡New]. See the example in (25), which is a copy of (14). (25) a. Wen glaubst du mag Hans? `Who do you think Hans likes?' b. (Ich glaube daû) Maria Hans mag. `(I think that) Hans likes Maria!' Here, the object Maria is presented as contrastive focus [‡New,‡Prom]. Unlike the completive focus case, the contrastive focus Maria can scramble over the oldinformation Hans, and thus the scrambled order (i.e., candidate (b)) is chosen as optimal in this context. This is illustrated in tableau T6.5. It is interesting to note that in German, the syntactic and discourse constraints are indeed intermingled with each other and show the active competition between the two components of grammar. Over all, the ®nal ranking of Prom g Subj g New predicts that only prominent informationÐthat is, topic and contrastive focusÐcan scramble over subject in German, which seems to be a desirable result, as illustrated earlier in this section. This prediction is crucial in explaining the focality e¨ect in binding, as we will see in the next section. 6.3

Deriving the Focality E¨ect in Binding

Having illustrated the general interaction between syntax and discourse in German scrambling, we will see in this section how this OT framework resolves the mixed bind-

Binding and Discourse Prominence

155

ing and focality problem raised in section 6.1. The solution crucially lies in the ranked intereaction between a binding constraint in syntax and constraint Prom in discourse. 6.3.1 Operator Binding Constraints in German Operator binding shows crosslinguistic variation, and it has been proposed that this variation can be explained by a couple of simple and universally available constraints. Bresnan (1995, 1998) proposes linear precedence (LP) and syntactic rank (SR) as relevant constraints,14 and shows that LP plays a major role in languages à a and Kiswahili. like Malayalam while SR does so in languages such as Chichew In the OT terminology, this means that Malayalam ranks the LP constraint higher à a or Kiswahili ranks the SR constraint higher. Let me present a simwhile Chichew pli®ed version of the LP and SR constraints in (26) (see Bresnan 1995, 1998 for details). (26) Operator binding constraints LP: A binder should linearly precede its bindee. SR: A binder should outrank its bindee on the functional hierarchy. (Functional Hierarchy:15 Subject > Acc. Object > Dat. Object > Oblique > Adjunct) LP requires that the binder, such as the operator, should linearly precede its bindeeÐ for example, the pronominal. The linear order e¨ect in binding has often been noted in the literature (Mohanan 1983; Haider 1989; Georgopoulos 1991; Williams 1994; Kuno and Kim 1994). So, if the pronominal is placed linearly before the operator, the sentence violates LP. SR, on the other hand, requires that the binder should be ranked higher than the bindee on the functional hierarchy. According to this hierarchy as shown above, if the (accusative) object is the binder and the subject is the bindee, for example, the sentence violates SR. Interestingly, the operator binding in German sometimes appears constrained by LP and other times by SR. Let's look at the operator binding examples in (3) and (4) again, repeated as (27) and (28). (27) a. *daû seinei Mutter jedeni mag b. daû jedeni seinei Mutter mag `that hisi mother likes everybodyi ' (28) a. b.

daû jederi seinei Mutter mag mag daû seinei Mutter jederi `that everybodyi likes hisi mother'

Example (27) is the so-called A-scrambling case. Here the operator binder jeden is the object and the pronominal bindee seine is part of the subject phrase. So obviously, the

156

Hye-Won Choi

object jeden does not outrank the subject of which seine is part. Therefore, both (27a) and (27b) violate SR.16 Example (27a) also violates LP because the binder jeden does not precede the bindee seine. Example (27b), in contrast, satis®es LP since now the binder jeden can linearly precede the bindee seine by means of the scrambling of the quanti®er phrase jeden over the pronominal phrase seine Mutter. In other words, LP but not SR plays a crucial role in this candidate set. The grammaticality of the scrambled order in (27b) seems to suggest that LP is ranked higher than Subj (LP g Subj ). On the other hand, in (28), which is the A0 -scrambling case, SR is satis®ed by both the nonscrambled sentence (a) and the scrambled sentence (b). Here, the binder jeder is the subject and thus outranks the bindee seine, which is part of the object. The di¨erence lies in that (28a) also satis®es LP while (28b) violates it. It should be noted that the scrambled example (28b) is still grammatical despite its violation of LP. The operator binding seems to hold here because SR is satis®ed. In other words, intuitively speaking, the operator binding in German appears to hold unless both LP and SR are violated (Choi 1995; Bresnan 1995, 1998), an example of which is (27a), or (3a). It is, however, not easy to formalize this intuition through the regular constraint ranking. First of all, SR alone, whatever its ranking on the constraint hierarchy, would not make any di¨erence, because this constraint is either invariably violated by all the relevant candidates as in (27) or is satis®ed as in (28). On the other hand, LP (as well as Subj ) would always prefer (28a), in which the subject binder precedes the pronominal bindee. Then how would we get the grammaticality of (28b)? Recall our discussion in section 6.2.1 showing that candidate (28b) is acceptable only when the scrambled pronominal phrase is interpreted as prominent ([‡Prom]) informationÐthat is, topic or contrastive focus. This means that (28b) is the optimal candidate only in that kind of informational context; otherwise (28a) is the optimal candidate. Then we can hypothesize that Prom be ranked higher than LP in German (Prom g LP), so that candidate (28b) wins when it satis®es Prom even though it violates LP. However, this hypothesis is immediately rejected if we consider the ungrammatical example (27a), which never surfaces as optimal output even if it is interpreted as prominent information [‡Prom]. Recall the contrast between (27a) and (28b), which was shown earlier in (7), repeated here as (29). (29) a. *daû seinei Mutter jedeni mag b. daû seinei Mutter jederi mag

ˆ (3a), (27a) ˆ (4b), (28b)

If Prom is ranked higher than LP, then (27a) should win over (27b), just as (28b) wins over (28a), when seine Mutter is interpreted as topic or contrastive focus (i.e., [‡Prom]), but this is not the case. As discussed in section 6.1, the scrambled candi-

Binding and Discourse Prominence

157

dateÐthat is, (3b) or (27b)Ðalways wins regardless of the informational context, which suggests the contradictory ranking, LP g Prom. Therefore, LP by itself cannot capture the binding facts illustrated in (27) and (28). We need to hold on to our original intuition that SR also plays a role and that the binding obtains unless both LP and SR are violated. In fact, this idea can be formalized through local conjunction in OT (Smolensky 1995; Legendre, Smolensky, and Wilson 1998; Legendre, forthcoming a, b; Aissen, chap. 3, this volume). Local conjunction is a formal operation that conjoins existing constraints. (30) Local conjunction The local conjunction of C1 and C2 in domain D: C1 & C2 is violated when there is some domain of type D in which both C1 and C2 are violated. This idea is formalized by imposing a universally ®xed ranking of C1 & C2 with respect to each of its component constraintsÐthat is, ``C1 & C2 dominates C1 , C2 .'' In the current binding problem, LP&SR dominates LP, SR. Since the conjoined binding constraint is ranked higher than each constraint, the violation of one constraint is overruled by the satisfaction of the conjoined constraint as long as the other constraint is satis®ed. This local conjunction property of German operator binding explains the fact that only example (27a) is ungrammatical because here both LP and SR are violated. It also explains why the mixed binding property occurs: the satisfaction of LP explains the antireconstruction e¨ect (A-property) in (27b) and the satisfaction of SR explains the reconstruction e¨ect (A0 -property) in (28b), as demonstrated above. 6.3.2 Discourse Prominence in Binding Now let me derive the special focality e¨ect in the reconstruction example and also the lack of it in the antireconstruction case. This can be derived from the constraint ranking in German between the conjoined binding constraint LP&SR, the information structure constraints Prom and New, and ®nally the phrase structure constraint Subj. I propose that the binding constraint is ranked even higher than the discourse constraint Prom, which renders the ®nal ranking in (31) in German.17 (31) Constraint ranking in German (revised) LP&SR g Prom g Subj g New Given this ranking, both the generalizationsÐthat the so-called neutral scrambling exhibits the A-binding property irrespective of informational status, and that only the ``focus'' scrambling shows the reconstruction e¨ect (Bayer and Korn®lt 1994; Neeleman 1994; Miyagawa 1995)Ðnaturally follow from the ranked interactions between these syntactic (binding and phrase-structural) and discourse constraints, as we will see.

158

Hye-Won Choi

Tableau T6.6 si M[S;

];jni [O;

]

a. seinei Mutter jedeni mag

LP&SR

Prom

Subj

New

*!

f b. jedeni seinei Mutter mag

*

Tableau T6.7 jri [S;

];si M[O;

]

LP&SR

Prom

Subj

New

f a. jederi seinei Mutter mag b. seinei Mutter jederi mag

*!

First, consider the neutral-context case in which no discourse-contextual information is available (or the sentence is not informationally partitioned by sharing the same marking). Compare (27) and (28) again, copied as (32) and (33). (32) a. *daû seinei Mutter jedeni mag b. daû jedeni seinei Mutter mag (33) a. b.

daû jederi seinei Mutter mag mag daû seinei Mutter jederi

As discussed in section 6.1, the competition between (32a) and (32b) is obvious and clear. The scrambled candidate always wins regardless of informational context. However, the competition between (33a) and (33b) is not that clear. Both are acceptable, but (33a) is preferred unless a particular context is provided, namely, when seine Mutter is interpreted as discourse prominent [‡Prom]. In other words, if that discourse condition is not metÐfor example, if no context is providedÐthen candidate (33a) wins. This is exactly what the current analysis predicts. Compare tableaux T6.6 and T6.7. Since the discourse-contextual information is not speci®ed in the input in the neutral-context case, all the candidates in both tableau T6.6 and tableau T6.7 vacuously satisfy the information structure constraints Prom and New. Therefore, only the syntactic constraints, LP&SR and Subj, are crucial in determining the optimal output. Given the conjoined nature of the binding constraint, only candidate (32a) violates LP&SR because it violates both LP and SR, as discussed earlier. Subj, on the other hand, penalizes any candidate not in the default orderÐfor example, the scrambled candidates (32b) and (33b). The higher ranking of the conjoined binding constraint LP&SR over the phrase-structural constraint Subj (LP&SR g Subj ) in tableau T6.6 correctly predicts that candidate (b) is the optimal output because the

Binding and Discourse Prominence

159

scrambling of jeden enables it to satisfy the higher constraint LP&SR even though it violates the lower constraint Subj. In contrast, the conjoined binding constraint LP&SR does not see any di¨erence between the two candidates in tableau T6.7. Remember that both candidates satisfy SR at least. The next relevant constraint Subj penalizes the scrambled candidate (b), and thus the default-order candidate (a) emerges as optimal. Intuitively speaking, the scrambling in tableau T6.6 is motivated to satisfy the binding constraint by placing the binder before the bindee. By contrast, the scrambling in tableau T6.7 cannot be motivated by the binding constraint because the binding already obtains in the base position. Moreover, the scrambling is penalized because it only disrupts the default order and thus violates Subj without satisfying any other constraints. Let us now examine how discourse constraints interact with these syntactic constraints. As summarized in section 6.2, each element of the sentence is marked with discourse features [New] and [Prom] in the input according to its information status. It was also established there that in German Prom is ranked higher than Subj, which in turn is ranked higher than New. We have seen that LP&SR is ranked higher than Subj; if we show LP&SR g Prom, the full ranking in (31) will follow. Let's examine the A-scrambling case (32) ®rst. The discourse contexts that crucially prove the ranking between LP&SR and Prom are those where seine Mutter is interpreted as [‡Prom]Ðthat is, where seine Mutter is interpreted as either topic or contrastive focus. In this situation, seine Mutter gets a high pitch accent by being discourse prominent. Tableau T6.8 is the case where seine Mutter is topic and tableau T6.9 is where it bears contrastive focus. In either competition, candidate (b), the scrambled one, is always considered grammatical. Candidate (a) violates the binding Tableau T6.8 si M[S; N‡P];jni [O;‡N P] a. seinei Mutter jedeni mag

LP&SR

Prom

Subj

New

*

*

*

Prom

Subj

New

*!

f b. jedeni seinei Mutter mag Tableau T6.9 si M[S;‡N‡P];jni [O; N P] a. seinei Mutter jedeni mag f b. jedeni seinei Mutter mag

LP&SR *!

* *

*

160

Hye-Won Choi

constraint LP&SR but satis®es the discourse constraint Prom, whereas candidate (b) satis®es the binding constraint LP&SR but violates Prom. That candidate (b) wins proves that the binding constraint LP&SR is ranked higher than the discourse constraint Prom. This gives us the ®nal ranking (31). List (34) is the illustration of all possible discourse contexts regarding the candidate set in (32). The constraint ranking in (31), especially the top ranking of LP&SR, yields the following result in this candidate setÐthat is, that candidate (b), the scrambled one, always wins. (34) a. b. c. d. e. f. g. h.

INPUT seinei Mutter seinei Mutter seinei Mutter seinei Mutter seinei Mutter seinei Mutter seinei Mutter seinei Mutter

[S; N P]; [S; N‡P]; [S; N P]; [S; N‡P]; [S;‡N P]; [S;‡N P]; [S;‡N‡P]; [S;‡N‡P];

jedeni jedeni jedeni jedeni jedeni jedeni jedeni jedeni

[O;‡N P] [O;‡N P] [O;‡N‡P] [O;‡N‡P] [O; N P] [O; N‡P] [O; N P] [O; N‡P]

) ) ) ) ) ) ) ) )

OUTPUT candidate (b) candidate (b) candidate (b) candidate (b) candidate (b) candidate (b) candidate (b) candidate (b)

In all the possible contexts in (34), candidate (a) never surfaces as the optimal output. In other words, candidate (a) is always ungrammatical. This is precisely because candidate (a) violates the operator binding constraint and the operator binding constraint is higher ranked than any discourse constraint, so that the resulting candidate is never a¨ected by discourse factors such as discourse prominence or discourse newness. Although I cannot illustrate all the tableaux due to space limitations, it is not hard to see the results. The di¨erent input speci®cations in terms of information structure as shown above give di¨erent constraint violations on Prom and New. However, the highest-ranking constraint LP&SR, which is higher than Prom, always prefers candidate (b). Therefore, candidate (b) is always the optimal output whatever the discourse context may be. Note that even the discourse prominence on seine Mutter in (34b) or (34g), which are illustrated in tableaux T6.8 and T6.9, cannot save candidate (a) from ungrammaticality. To summarize, the apparent A-binding property of scrambling in (32b) or (3b) is due to the fact that the scrambling of the operator phrase satis®es the operator binding constraint by virtue of satisfying LP. In this scrambling, no special discourse e¨ect arises because the operator binding constraint LP&SR is ranked higher than the discourse constraints Prom or New. Remember that the scrambling in (3b) is called ``neutral'' scrambling (Neeleman 1994). This scrambling is ``neutral'' not only because it happens in the neutral context, as shown in tableau T6.6, but also because it is ``neutral'' to discourse contextÐthat is, it is compatible with any discourse context.

Binding and Discourse Prominence

161

Tableau T6.10 jri [S;‡N P];si M[O; N‡P]

LP&SR

a. jederi seinei Mutter mag

Prom

Subj

*!

f b. seinei Mutter jederi mag

New *

*

Tableau T6.11 jri [S; N P];si M[O;‡N‡P]

LP&SR

a. jederi seinei Mutter mag

Prom

Subj

New

*

*

*!

f b. seinei Mutter jederi mag

Now, compare (32) with the A0 -binding case in (33). We have seen in tableau T6.7 that the nonscrambled candidate (a) wins in the neutral context when no discoursecontextual information is available. Then what about the contexts in which seine Mutter is interpreted as [‡Prom]Ðthat is, as topic or contrastive focus? See tableaux T6.10 and T6.11. Tableau T6.10 is where seine Mutter is topic, and tableau T6.11 is contrastive focus. In either case, seine Mutter is stressed because it is discourse prominent. Crucially di¨erent from the A-scrambling case demonstrated in tableaux T6.8 and T6.9, the conjoined binding constraint LP&SR does not do much work in this competition, represented by tableau T6.10 or T6.11. LP&SR is satis®ed by both candidates anyway: unlike the situation in (32), SR is satis®ed in (33). Therefore, the crucial constraint is not the binding constraint LP&SR but the discourse constraint Prom. Any candidate that satis®es Prom will win. In both tableau T6.10 and tableau T6.11, candidate (b) satis®es Prom because the [‡Prom] element, seine Mutter, is placed ®rst. In other words, the scrambled candidate (b) wins because it satis®es Prom by scrambling seine Mutter. Now, recall our initial observation in (7) in section 6.1 that some focality or pitch accent on the pronominal phrase makes the binding in (4b) hold, but not the one in (3a). See the contrast again in (35). (35) a. *daû seinei Mutter jedeni mag b. daû seinei Mutter jederi mag

ˆ (3a), (33a) ˆ (4b), (34b)

This is exactly what we have just observed in the contrast between tableaux T6.8 and T6.9 on the one hand and tableaux T6.10 and T6.11 on the other. In the former, candidate (a)Ðthat is, (35a)Ðcan never surface because it violates the highest-ranking LP&SR even though it satis®es Prom. In the latter competition, candidate (b)Ðthat

162

Hye-Won Choi

Tableau T6.12 jri [S; N‡P];si M[O;‡N P]

LP&SR

Prom

Subj

New

*!

*

*

f a. jederi seinei Mutter mag b. seinei Mutter jederi mag

is, (35b)Ðdoes surface as the optimal candidate precisely because it satis®es Prom: LP&SR is irrelevant in this competition because it is satis®ed by both the candidates. Now let's see if any other discourse context yields candidate (b) as optimal in the so-called A0 -scrambling case. See (36). (36) a. b. c. d. e. f. g. h.

jederi jederi jederi jederi jederi jederi jederi jederi

[S; N P]; [S; N‡P]; [S; N P]; [S; N‡P]; [S;‡N P]; [S;‡N P]; [S;‡N‡P]; [S;‡N‡P];

seinei seinei seinei seinei seinei seinei seinei seinei

Mutter Mutter Mutter Mutter Mutter Mutter Mutter Mutter

[O;‡N P] [O;‡N P] [O;‡N‡P] [O;‡N‡P] [O; N P] [O; N‡P] [O; N P] [O; N‡P]

) ) ) ) ) ) ) )

candidate candidate candidate candidate candidate candidate candidate candidate

(a) (a) (b) (a) (a) (b) (a) (a)

Contexts (36c) and (36f ), shown earlier as tableau T6.10 and tableau T6.11 respectively, are the only two cases in which candidate (b) emerges as the optimal candidate. These are the contexts where seine Mutter is interpreted [‡Prom] and thus gets a high pitch accent (recall the relevant contexts in (5) and (6)). In other words, unless the pronominal phrase is [‡Prom], the nonscrambled candidate (a) is preferred. Actually, native speakers report that (a) is preferred in most cases. This preference is easily explained in the current analysis. Candidate (a) is the optimal candidate not only in the neutral context, as illustrated in tableau T6.7, but also in many other contexts with various discourse information with the only exception of the [‡Prom] case, as demonstrated in (36). Of course, if jeder is interpreted as [‡Prom] so that candidate (a) satis®es Prom while candidate (b) violates it, which is the opposite of tableau T6.10 or T6.11, then candidate (a) is the winner, for sure. One such case, (36b), for instance, is demonstrated in tableau T6.12. The fact that candidate (a) is the optimal candidate unless it violates Prom is easily explained by the ranking of Prom g Subj. As shown above, the conjoined binding constraint LP&SR does not distinguish the two candidates in this competition, and the next constraint Prom does not di¨erentiate them unless one of the elements is marked [‡Prom]. Then the competition relies on the next highest constraint Subj.

Binding and Discourse Prominence

163

Tableau T6.13 jri [S;‡N P];si M[O; N P]

LP&SR

Prom

Subj

f a. jederi seinei Mutter mag b. seinei Mutter jederi mag

New *

*!

Subj of course prefers the nonscrambled candidate (a), which maintains the default order. Due to the higher ranking of Subj than New, candidate (a) will be chosen as the optimal candidate regardless of its con®guration in terms of ``discourse newness.'' For example, even when jeder is [‡New], as in (36e), so that candidate (a) violates New, candidate (a) is still the optimal candidate because of its satisfaction of Subj in contrast with candidate (b)'s violation of it. This is illustrated in tableau T6.13. To summarize, the ranked interaction among the operator binding constraint LP&SR, the discourse constraints Prom and New, and the phrase structure constraint Subj nicely captures the fact that (1) German scrambling exhibits the split binding property, and moreover that (2) the A0 -property of ``reconstruction'' takes place only when the scrambled pronominal phrase is interpreted as ``discourse prominent''Ðthat is, [‡Prom]Ðin the current framework. To begin, the split binding property is captured by the conjoined nature of the binding constraint in German. The A-binding propertyÐthat is, the antireconstruction e¨ectÐis explained by LP. The scrambling in (3b), for example, is motivated and licensed by satisfying LP, and thus LP&SR. In other words, by satisfying LP, the scrambling creates the new binding relation, and hence obtains the A-binding property. By contrast, the A0 -binding property is accounted for by SR. In (4), SR is satis®ed by both the nonscrambled and the scrambled candidate. So, scrambling in (4b) is not motivated by the binding constraint at all. In other words, scrambling in (4b) has nothing to do with creating a new binding relation. Next, the A0 -scrambling's association with a certain ``focality'' e¨ect (i.e., discourse prominence in the current framework) is also nicely captured in the current constraint interaction approach. It is actually explained by two interesting aspects of constraint interaction. The ®rst is the conjoined nature of the operator binding constraintÐthat is, candidate (b) in tableau T6.10 or T6.11 is not discarded in the ®rst round of evaluations with respect to the operator binding constraint although it violates LP. The second is the higher ranking of the Prom constraint over the phrase structure constraint Subj: although candidate (b) violates Subj by scrambling the object phrase, it satis®es the higher constraint Prom by doing so. In other words, the only way that the scrambling of the pronominal phrase is allowed is when it is prominentÐthat is, by being topic or contrastive focus. It is exactly the satisfaction of this

164

Hye-Won Choi

constraint Prom that associates this A0 -scrambling with the focality or discourse prominence. 6.4

Conclusion

The current analysis provides a rather simple solution to the split binding problem and also to the so-called A0 -scrambling's association with a certain discourse e¨ect with no complications. The OT constraint interactions between the syntactic constraints (i.e., the binding constraint LP&SR and the phrase structure constraint Subj ) and the discourse constraints (i.e., Prom and New) o¨er systematic answers to the problems raised in section 6.1, as demonstrated in section 6.3. In addition, the current OT approach has a couple of advantages over previous analyses. One advantage of the current analysis is that it is free from the complicated movement typology in relation to scrambling. In the current OT approach, the mixed property is due to the competition of ranked constraints. Because of the top ranking of the conjoined binding constraint, only the scrambled candidate wins in the socalled neutral scrambling, and it thus always shows the A-property. In the so-called focus scrambling, the scrambled candidate can win only in a speci®c discourse context due to the higher ranking of the information structure constraint Prom over the phrase structure constraint Subj. In this way, the current analysis overcomes the complicated problem of movement classi®cation in previous analyses, especially in relation to wh-movement and topicalization, as illustrated in (9) and (10) in section 6.1.2. One other advantage of the current analysis is that focality can be dealt with separately from a certain type of movement. In the current OT analysis, the association of focality with the apparent A0 -binding property is only a byproduct of the interactions between the binding constraint in syntax and information structure constraints in discourse. That is, focality is not inherently associated with a certain movement type, such as A0 -movement or operator movement. It separates focusing, or information structuring in general, from a certain type of syntactic operation, although they do interact. Thus, in principle, it allows the so-called neutral scrambling to involve focusing as well, which we have noted in (11) in section 6.1.2. The current approach solves the problems raised in previous analyses, as in Bayer and Korn®lt 1994, Finer 1994, Neeleman 1994, and Miyagawa 1995, without losing their underlying intuition that there seem to be two types of scramblingÐone ``neutral'' and the other ``focal.'' The highest ranking of the conjoined binding constraint LP&SR explains that the optimal output of this competition is ``neutral'' to any discourse e¨ect, and the interaction between LP&SR and the discourse constraint Prom captures exactly when the ``focal'' element can scramble. Therefore, without proposing two di¨erent types of scrambling, the current OT analysis captures the fact that

Binding and Discourse Prominence

165

scrambling is motivated by two di¨erent factors: one, by binding con®gurations, and the other, by discourse-prominence relations. Notes This chapter is based on the presentation at the Hopkins Optimality Theory Workshop and the University of Maryland Mayfest in May 1997, and an earlier version was also presented at the Sixteenth West Coast Conference on Formal Linguistics in March 1997. I am grateful to Joan Bresnan, Helen de Hoop, Paul Kiparsky, Ivan Sag, Peter Sells, HenrieÈtte de Swart, Thomas Wasow, and Gert Webelhuth for their insightful comments and helpful discussions on this and earlier versions. I also thank JoaÄo Costa, Robert Frank, Jane Grimshaw, Caroline Heycock, GeÂraldine Legendre, Vieri Samek-Lodovici, Paul Smolensky, Sten Vikner, and the audience at the Hopkins/Maryland workshop for their invaluable questions and comments. I am also indebted to Sylvia Ochs, Susanne Riehemann, and Heidi Waltz for their help with German examples. Finally, I owe special thanks to Vieri Samek-Lodovici and GeÂraldine Legendre for their detailed reviews of the chapter at its last stage. Any errors or misinterpretations are of course mine. 1. Nonoperator binding also exhibits similar e¨ects, but I restrict the scope of this chapter to the operator binding of pronominals for simplicity of discussion. See Frank, Lee, and Rambow 1992 among others for the nonoperator binding cases. 2. I only show the operator binding examples here again in parallel with the English examples in (1) and (2). See Moltmann 1990, Neeleman 1994, and Bayer and Korn®lt 1994 for other kinds of evidence of the mixed properties of scrambling, including anaphor binding. 3. If seine Mutter does not bear the high pitch accent or if an appropriate context is not provided in (4b), it is often judged not as acceptable as (4a). That is why this type of sentence sometimes gets a ``?'' for its grammaticality. 4. I thank Susanne Riehemann, Sylvia Ochs, and Heidi Waltz for providing and judging these examples. 5. Neeleman (1994) and Bayer and Korn®lt (1994) actually derive the A-properties of scrambling not from A-movement but from base generation. However, their proposals are parallel to others in that they assume that the mixed binding facts come from the A/A0 typology. 6. See Vallduvõ 1993 for a survey of previous proposals and also for references. 7. Grimshaw and Samek-Lodovici (1995) and Samek-Lodovici (1996) made a similar proposal that discourse information be included in the input (e.g., x ˆ topic; y ˆ focus). 8. The verbÐthat is, magÐwould also get the information-structural marking, probably [ New, Prom], but it is omitted here for the sake of simplicity of discussion since the verb is not scrambled in German (although it may be topicalized). The inability of the verb to be scrambled could also be captured in terms of constraint ranking if we posit a constraint like ``do not adjoin verb'' and rank it higher than the information-structuring constraints introduced here. 9. Among other candidates, we could also get the ones where the verb mag scrambles. For example, (i) (Ich glaube daû) mag Hans Maria; (ii) (Ich glaube daû) Maria mag Hans; etc.

166

Hye-Won Choi

As discussed in note 8, these candidates would fall out of consideration immediately by a higher-ranking constraint like ``do not adjoin verb.'' Again, the most relevant candidates where only arguments scramble are considered in this chapter for simplicity of discussion. 10. Choi (1996) also proposed Canon2, which requires the canonical order between nonsubject constituents. 11. If a sentence were only partitioned into the ground (or presupposition) and focus, this constraint could be considered parallel to AlignFocus in Grimshaw and Samek-Lodovici 1995, Samek-Lodovici 1996, Samek-Lodovici (chap. 11, this volume), and Costa (chap. 7, this volume), which requires a focused constituent to be aligned with the (right) edge of VP. 12. Choi (1996, 1999) also illustrates the semantic in¯uence on scrambling by the Specificity constraints, which exert in¯uence on the feature speci®cation in information structure. These constraints are motivated by the desire to capture the ``de®niteness'' e¨ect in scrambling. 13. The ranking in (21) derives the focus-related e¨ects in German discussed in Lenerz 1977, Abraham 1986, and Moltmann 1990 (see Choi 1996, 1999 for details). 14. Bresnan (1995, 1998) proposes that a constraint based on thematic hierarchy is also relevant to operator binding, but I do not pursue this possibility in this chapter. 15. This notion is comparable to ``Accessibility Hierarchy'' in Keenan and Comrie 1977. 16. See Bresnan 1995, 1998 for a more technical de®nition of ``outranking.'' In simple words, subject outranks object, and any part of object, for example. 17. The ranking of both unconjoined constraints LP and SR with respect to Prom, Subj, and New should be determined by further investigation, except that LP must be ranked lower than Prom. References Abraham, Werner. 1986. Word Order in the Middle Field of the German Sentence. In Werner Abraham and Sjaak de Meij (eds.), Topic, Focus, and Con®gurationality, 15±38. Amsterdam: Benjamins. Bayer, Josef, and Jaklin Korn®lt. 1994. Against Scrambling as an Instance of Move-alpha. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling: Movement and Nonmovement Approaches to Free Word-Order Phenomena, 17±60. Berlin: Mouton de Gruyter. Bresnan, Joan. 1995. Linear Order, Syntactic Rank, and Empty Categories: On Weak Crossover. In Mary Dalrymple, Ronald M. Kaplan, Joan T. Maxwell, and Annie Zaenen (eds.), Formal Issues in Lexical-Functional Grammar, 241±274. Stanford, Calif.: CSLI Publications. Bresnan, Joan. 1998. Morphology Competes with Syntax: Explaining Typological Variation in Weak Crossover E¨ects. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 59±92. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Choi, Hye-Won. 1995. Weak Crossover in Scrambling Languages: Precedence, Rank, and Discourse. Paper presented at the Sixty-Ninth Annual Meeting of the Linguistic Society of America, New Orleans. Choi, Hye-Won. 1996. Optimizing Structure in Context: Scrambling and Information Structure. Doctoral dissertation, Stanford University, Stanford, Calif.

Binding and Discourse Prominence

167

Choi, Hye-Won. 1997. Focus and Binding in Scrambling: An OT Account. In Emily Curtis, James Lyle, and Gabriel Webster (eds.), Proceedings of the Sixteenth West Coast Conference on Formal Linguistics, 81±95. Stanford, Calif.: CSLI Publications. Choi, Hye-Won. 1999. Optimizing Structure in Context: Scrambling and Information Structure. Stanford, Calif.: CSLI Publications. Dayal, Veneeta Srivastav. 1994. Binding Facts in Hindi and the Scrambling Phenomenon. In Miriam Butt, Tracy Holloway King, and Gillian Ramchand (eds.), Theoretical Perspectives on Word Order in South Asian Languages, 237±262. Stanford, Calif.: CSLI Publications. Deprez, Viviane. 1994. Parameters of Object Movement. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling: Movement and Non-movement Approaches to Free Word-Order Phenomena, 101±152. Berlin: Mouton de Gruyter. Dik, Simon, et al. 1981. On the Typology of Focus Phenomena. In Teun Hoekstra, Harry van der Hulst, and Michael Moortgat (eds.), Perspectives on Functional Grammar, 41±74. Dordrecht: Foris. Finer, Daniel. 1994. On the Nature of Two A0 -Positions in Selayarese. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling: Movement and Non-movement Approaches to Free Word-Order Phenomena, 153±183. Berlin: Mouton de Gruyter. Frank, Robert, Young-Suk Lee, and Owen Rambow. 1992. Scrambling as Non-operator Movement and the Special Status of Subjects. In S. Barviers, M. den Dikken, and C. Levelt (eds.), Proceedings of the Third Leiden Conference for Junior Linguists, 135±154. Georgopoulos, Carol. 1991. Canonical Government and the Speci®er Parameter: An ECP Account of Weak Crossover. Natural Language and Linguistic Theory 9: 1±46. GivoÂn, Talmy. 1989. Mind, Code, and Context: Essays in Pragmatics. Hillsdale, N.J.: Erlbaum. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28: 373± 422. Grimshaw, Jane, and Vieri Samek-Lodovici. 1995. Optimal Subjects. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 598±606. GLSA, University of Massachusetts, Amherst. Haider, Hubert. 1989. y-Tracking Systems: Evidence from German. In LaÂszlo MaraÂcz and Pieter Muysken (eds.), Noncon®gurationality: The Typology of Asymmetries, 185±206. Dordrecht: Foris. Keenan, Edward, and Bernard Comrie. 1977. Noun Phrase Accessibility and Universal Grammar. Linguistic Inquiry 8: 63±99. Kuno, Susumu, and Soo-Yeon Kim. 1994. The Weak Crossover Phenomena in Japanese and Korean. In Japanese/Korean Linguistics 4: 1±38. Lambrecht, Knud. 1994. Information Structure and Sentence Form. Cambridge, England: Cambridge University Press. Lee, Young-Suk. 1993. Scrambling as Case-Driven Obligatory Movement. Doctoral dissertation, University of Pennsylvania, Philadelphia. Lee, Young-Suk, and Beatrice Santorini. 1994. Towards Resolving Webelhuth's Paradox: Evidence from German and Korean. In Norbert Corver and Henk van Riemsdijk (eds.),

168

Hye-Won Choi

Studies on Scrambling: Movement and Non-movement Approaches to Free Word-Order Phenomena, 257±300. Berlin: Mouton de Gruyter. Legendre, GeÂraldine. Forthcoming a. Morphological and Prosodic Alignment of Bulgarian Clitics. In J. Dekkers, F. van der Leeuw, and J. van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. New York: Oxford University Press. Legendre, GeÂraldine. Forthcoming b. Optimal Romanian Clitics: A Crosslinguistic Perspective. In V. Motapanyane (ed.), Comparative Studies in Romanian Syntax. The Hague: Holland Academic Graphics. Legendre, GeÂraldine, Paul Smolensky, and Colin Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky, (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Lenerz, JuÈrgen. 1977. Zur Abfolge nominaler Satzglieder im Deutschen. TuÈbingen: Narr. Mahajan, Anoop K. 1990. The A/A-Bar Distinction and Movement Theory. Doctoral dissertation, MIT, Cambridge, Mass. Miyagawa, Shigeru. 1995. Against Optional Scrambling. Ms., MIT, Cambridge, Mass. Mohanan, K. P. 1983. Functional and Anaphoric Control. Linguistic Inquiry 14: 641±674. Moltmann, Friederike. 1990. Scrambling in German and the Speci®city E¨ect. Ms., MIT, Cambridge, Mass. MuÈller, Gereon, and Wolfgang Sternefeld. 1993. Improper Movement and Unambiguous Binding. Linguistic Inquiry 24: 461±507. Neeleman, Ad. 1994. Scrambling as a D-Structure Phenomenon. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling: Movement and Non-movement Approaches to Free Word-Order Phenomena, 387±429. Berlin: Mouton de Gruyter. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Rochemont, Michael, and Peter Culicover. 1990. English Focus Constructions and the Theory of Grammar. Cambridge, England: Cambridge University Press. Saito, Mamoru. 1992. Long Distance Scrambling in Japanese. Journal of East Asian Linguistics 1: 69±118. Samek-Lodovici, Vieri. 1996. Constraints on Subjects: An Optimality Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Smolensky, Paul. 1995. On the Internal Structure of Con, the Constraint Component of UG. Paper presented at a UCLA Linguistics Colloquium, Los Angeles. Smolensky, Paul. 1997. Constraint Interaction in Generative Grammar II: Local Conjunction, or Random Rules in UG. Paper presented at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest 1997, Baltimore.

Binding and Discourse Prominence

169

Ueyama, Ayumi. 1994. Against A/A0 -Movement Dichotomy. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling: Movement and Non-movement Approaches to Free Word-Order Phenomena, 459±486. Berlin: Mouton de Gruyter. VallduvõÂ, Enric. 1992. The Informational Component. New York: Garland. VallduvõÂ, Enric. 1993. Information Packaging: A Survey. Ms., Centre for Cognitive Science and Human Communication Research Centre, University of Edinburgh. Webelhuth, Gert. 1992. Principles and Parameters of Syntactic Saturation. New York: Oxford University Press. Williams, Edwin. 1994. Thematic Structure in Syntax. Cambridge, Mass.: MIT Press.

Chapter 7 The Emergence of Unmarked Word Order

JoaÄo Costa

The sequences SVO, SOV, VSO, and VOS are all well attested basic word orders in di¨erent languages. It is not clear, however, how to formally motivate why this type of variation at the base exists. In this chapter, I intend to look at unmarked word orders in di¨erent languages and explain why there is variation at the base. I will propose an analysis within Optimality Theory (OT; Prince and Smolensky 1993), and suggest that the observed variation is due to the e¨ect of constraints that are normally not active due to the domination of higher-ranked discourse-related constraints. I will therefore suggest that unmarked word orders are a case of the more general phenomenon Emergence of the Unmarked (McCarthy and Prince 1994). The organization of the chapter is as follows: section 7.1 illustrates the word order patterns examined. Section 7.2 spells out theoretical assumptions, and section 7.3 the constraints used and the ranking proposed for Portuguese. In section 7.4, I derive the other word orders considered by reranking the constraints proposed for Portuguese. This work is intended to test the power of OT as a theory of language variation, and to add to the studies done by Grimshaw and Samek-Lodovici (1995, 1998), Samek-Lodovici (1996), Costa (1997c, 1998), and Choi (1996), who suggest that different word orders are not optional but the result of di¨erent functional speci®cations in the input. Here, the reverse situation is explored: if several word orders are legitimate expressions of the functional speci®cation in the input, why and how does a language select one of them as basic? 7.1

Problem

A well-known fact concerning linguistic variation is that di¨erent languages display di¨erent basic word orders, where basic word order means the word order emerging in out-of-the-blue contexts or in sentence-focus contexts. A clear test for this is to check which word order emerges as an answer to a question like ``What happened?'' (among other sources, see Li 1976; Dik 1978).

172

JoaÄo Costa

The following examples show di¨erent base word orders for di¨erent languages. Portuguese is an SVO language, as (1) shows. (1) Portuguese O JoaÄo comeu o bolo. `John ate the cake.' References to the basic word order in Portuguese are made in Ambar 1992, Duarte 1987, and Mateus et al. 1989, among many others. In other Romance languages, the same basic word order can be found. As Rizzi (1982) and others observe, Italian also displays SVO as its basic word order. (2) Italian Andrea ha letto il libro. `Andrea has read the book.' Likewise, Spanish exhibits the same word order, according to Hernanz and Brucart (1987). (3) Spanish Juan ha visto a Maria. `Juan has seen Maria.' It is worth noting that the situation in Spanish is not that simple; authors disagree on what the basic word order in Spanish is. Some studies claim that the basic word order of Spanish is VSO (OrdoÂnÄez and TrevinÄo 1995; Zubizarreta 1995). (4) Spanish Comio Juan los guisantes. ate Juan the peas In this chapter, I will not take a position concerning the correctness of either of these two descriptions. Instead, I will assume that the two descriptions correspond to dialectal variation and that the two word orders are allowed as basic by di¨erent speakers. I will name the SVO variant Spanish A, and the VSO variant Spanish B. Greek is another language with basic VSO order, as Alexiadou and Anagnostopoulou (1995) note. (5) Greek Pandreftike o Petros tin Ilekttra married the Peter-NOM the Electra-ACC `Peter married Electra.' Greek also permits SVO orders, but those are arguably analyzed in terms of left dislocation of the subjects (see Alexiadou and Anagnostopoulou 1995 for details).

Emergence of Unmarked Word Order

173

(6) Greek O Petros pandreftike tin Ilektra. the Peter married the Electra `Peter married Electra.' Arabic is similar to Greek in that its basic word order is VSO (SVO being derived via left dislocation of the subject, as argued in Fassi-Fehri 1989 and Ouhalla 1991). (7) Arabic (from Ouhalla 1991) Sa-ya-shtarii Zayd-un dar-an. fut-3ms-buy Zayd house `Zayd will buy a house.' Other VSO languages include Berber, Chamorro, and Celtic; see examples (8) to (11) (all examples are taken from Ouhalla 1991). (8) Berber Ad-y-segh Moha ijn teddart. fut-3ms-buy Moha one house `Moha will buy a house.' (9) Chamorro Ha-taitai si Maria i lepblu. 3sg-read the Maria the book. `Maria read the book.' (10) Welsh Gwelodd y bechgyn y draig. saw the boys the dragon `The boys saw the dragon.' (11) Irish Deireann siad o paidir say they a prayer `They say a prayer.' Note that in this introduction I am collapsing di¨erent types of VSO languages. In the following sections I will note di¨erences between them that have led some authors to analyze them as belonging to distinct typological groups (see Ouhalla 1991; Alexiadou and Anagnostopoulou 1995). Dutch exempli®es the SOV group. (Though this word order only emerges in embedded context, its status as unmarked is the same with respect to information structure. For details, see Dik 1978.)

174

JoaÄo Costa

(12) Dutch dat Jan de krant leest. that Jan the newspaper reads `. . . that Jan reads the newspaper.' VOS languages are illustrated by Malagasy (see Keenan 1976). In this language, the subject follows both the verb and the object. (13) Malagasy Nahita an-dRabe Rakoto saw ACC-Rabe Rakoto `Rakoto saw Rabe.' In this chapter, I will not consider languages in which the object precedes both the subject and the verb in the unmarked case. Such languages are rare (see Derbyshire 1977 for OVS in Carib), and debate surrounds their status as neutral in the attested cases (see GivoÂn 1984 for discussion). But I will make an attempt to derive their rarity. This variation at the base raises several questions, of which I would like to consider the following subset: 1. What does it (formally) mean to be the unmarked word order? 2. Why is there variation at the base? 3. Is the unmarked word order truly basic or is it in turn derived from a more basic structure? The ®rst question relates to the old question of what it means to be unmarked. There are several intuitive de®nitions of markedness (see Battistella 1996 for a review), but is there a clear formalism for expressing unmarkedness? In this chapter I will argue, along the lines of McCarthy and Prince (1994), that OT provides such a formalism. The second question is of a general nature: If all these word orders are felt to be basic in the respective languages, why should there be di¨erent basic structures? That is, why is there not a uniform basis? The stand taken in this chapter is that there may be a uniform basis from which these basic word orders are derived. Di¨erent orders emerge as unmarked because di¨erent languages correspond to di¨erent constraint rankings; thus di¨erent basic word orders are unmarked with respect to di¨erent constraints. Following the reasoning that underlies the answer to the second question, the position taken in this chapter is that the basic word order does not necessarily correspond to an underived representation. That is, once we assume that the surface base representation may already be the result of compliance with a set of constraints, we may also assume (and in fact, must, in OT) that the identi®ed word orders are themselves derived from a single uniform input. This is in accordance with the tradition and with more recent claims concerning word order, like Kayne 1994.

Emergence of Unmarked Word Order

175

Throughout the chapter, I will not take a strong position with respect to whether there should be only one base word order (like Kayne does). Nevertheless I would like to stress that if this type of approach proves to be on the right track, one should not take the variation at the base as an argument for several basic word orders. 7.2

Theoretical Background: Optimality Theory and Emergence of the Unmarked

According to OT, a particular grammar is a ranking of a set of universal violable constraints. These constraints evaluate the relative well-formedness of a set of candidates generated from a single input. Violation of one constraint does not imply ungrammaticality, provided that the other candidates violate higher-ranked constraints. Tableau T7.1 exempli®es a potential OT diagram. This tableau shows that, in spite of the violation of constraints A and C, candidate 1 is the grammatical one, since the other candidate fatally violates constraint B, which is higher ranked than C. This tableau also shows that is does not matter how many violations there are. Ranking is the crucial factor. This diagram could be larger and could include constraints D, E, and F, all violated by candidate 1 and satis®ed by candidate 2, which would nevertheless be dispreferred. In a potential language where C dominates B, the winning output candidate would be candidate 2. This type of theory is di¨erent from one where constraints are absolute and either active or inactive for each language. In fact, it predicts that the e¨ects of a constraint that are normally invisible because all candidates that satisfy it fatally violate a higher-ranked constraint may be visible in a context where the violation of the higher-ranked constraints is controlled for. That is, the theory predicts that lowranked constraints may be operative in some cases. McCarthy and Prince (1994) call the possibility of seeing the e¨ects of a constraint that normally is violated the Emergence of the Unmarked. Since the Emergence of the Unmarked is a relevant distinctive feature of OT, most works on syntax and phonology within this framework exploit it. This notion is crucial in this chapter, since I will argue that unmarked or base word orders are just a subcase of the Emergence of the Unmarked. The next section shows this for Portuguese. Tableau T7.1 A f Candidate 1

*

Candidate 2

*

B

C *

*!

176

7.3

JoaÄo Costa

Deriving the Unmarked SVO Order in Portuguese

In Costa 1997c, I have proposed that word order variation in Portuguese and its relation to discourse may be captured by assuming the constraints in (14) and the ranking given in (15). (14) Constraints a. AlignFocus: The rightmost constituent in a clause is focused. b. SubjCase: Subjects are Case-licensed in SpecIP. c. ObjCase: Objects are Case-licensed in SpecAgrOP. d. Stay: Do not move. e. TopFirst: Topics are sentence initial (inspired by Li 1976); nontopics cannot be topicalized. Failed by topics that are not sentence initial, and by topicalized nontopics. These constraints are independently motivated and have been used in di¨erent domains, although not necessarily as violable constraints (for AlignFocus, see Grimshaw and Samek-Lodovici 1995, 1998; for Case, Chomsky 1995 and Vikner, chap. 14, this volume; for Stay, Grimshaw 1997). These constraints are in con¯ict, since Stay penalizes movement, and Case and AlignFocus may force a constituent to move. Likewise, AlignFocus and Case are in con¯ict, since AlignFocus requires that a focused subject be rightmost in a sentence, while Case requires that it move leftward to SpecIP. Similarly, TopFirst con¯icts with ObjCase by forcing an object to surface in the topic position instead of in the case-licensing position. In OT, these con¯icts are resolved language-internally, in accordance with the ranking of constraints that constitutes the language's grammar. It may look at ®rst glance as though I am departing from the assumption that different types of movement may feed one another, for example allowing an object to appear in a topicalized position after moving through SpecAgrOP for case purposes. I do not reject the standard view that di¨erent constraints may be satis®ed through chains. However, the crosslinguistic evidence presented here and in Costa 1998 appears to point to the conclusion that compliance with the constraints above implies that XPs surface in given positions. If derivations were enough to satisfy con¯icting constraints, uniform derivations should be found crosslinguistically, and subjects should be able to surface at the rightmost position in the sentence, after passing through SpecIP. Since languages di¨er in having obligatory or discourse-conditioned subject movement, it seems that the derivational approach may not do the whole job alone (see Costa 1998 for a synthesis of the derivational and OT approaches). For cases in which there is no evidence for the need to satisfy a constraint at the surface, chains may be established. Such a case is the theta criterion. Arguments are generated

Emergence of Unmarked Word Order

177

VP-internally for the sake of being assigned theta roles, but need not stay there for the theta criterion to be satis®ed. For Portuguese, I proposed in Costa 1997c that the ranking speci®ed in (15) may characterize the behavior of Portuguese word orders.1 (15) Ranking for Portuguese {AlignFocus, TopFirst} g SubjCase g Stay g ObjCase This ranking determines the following: Subjects only move to SpecIP if they are not focused. Objects never move to SpecAgrOP; they stay in their base position if they are focused and move, adjoining to VP, if they are not focused. That this is the behavior of arguments in Portuguese has been argued in Costa 1997, 1996b. (For other approaches to similar word order alternations, see, among other sources, Ambar 1992; Duarte 1987; Barbosa 1995, 1996; Martins 1994. For details concerning the implementation of the OT approach, see Costa 1997c.) The top ranking of AlignFocus and TopFirst re¯ects the discourse-con®gurational character of the language. One case considered in Costa 1996 is what happens in cases of sentence focus. That is, what word order is observed in contexts answering the question ``What happened?''. As (16) illustrates, the emerging word order is SVO. (16) O que e que aconteceu? What happened a. A Maria partiu um prato. Maria broke a plate b. aPartiu a Maria um prato. c. aPartiu um prato a Maria. d. a*A Maria um prato partiu. e. aUm prato a Maria partiu. f. aUm prato partiu a Maria `Maria broke a plate.' The argument for the emergence of SVO as unmarked goes as follows: from a functional point of view, all these orders are equivalent, provided that the rightmost constituent of the sentence bears the main stress (see Nespor and Vogel 1986; Frota 1994; among others). As long as focus projection is permitted (independently of its formulation in terms of syntactic constituency or in terms of linear order; see Costa 1996 for discussion), the whole sentence may be interpreted as focused independently of the word order displayed. For hearers to interpret the whole sentence as focused, it is enough not to hear any high stress before the end of the sentence. When they hear an SVO sentence with the main prominence on the object, they may interpret it as focus on the object, on the VP, or on the whole sentence, depending on which constituent is taken as the domain of focus. The same holds for a VSO sentence with

178

JoaÄo Costa

main prominence on the object: hearers may interpret the domain of focus as the object alone, the subject and the object, or the whole sentence. Similarly, since there is no topic feature at the input for these sentences, TopFirst is vacuously satis®ed by all word orders that do not involve topicalization, since no sentence-initial element will be the topic. Moreover, if some element is left-dislocated, this constraint is violated once, because there is left dislocation of a nontopic element. The crucial point is that, in sentence-focus contexts, TopFirst is only relevant for ruling out candidates with topicalization of a given constituent. The question is then: Why aren't all these word orders optional actualizations of sentence focus? I would like to argue that this is where the Emergence of the Unmarked plays a crucial role. By virtue of being top ranked, the e¨ects of AlignFocus and TopFirst often make the e¨ects of the other constraints invisible. For instance, every time a subject is in focus, it is right-aligned, and consequently Case has to be violated. If AlignFocus is not violated by any candidate, as is the case in the sentence-focus context, the e¨ects of SubjCase become visible. Since this is the next-highest constraint, it will determine the optimal candidate. The winning candidate will be one that does not violate this constraint. This is exempli®ed in tableau T7.2. The candidates considered in tableau T7.2 include: Candidate Candidate Candidate Candidate Candidate Candidate

(a): SVO sentence with subject in SpecIP and object in situ (b): SVO sentence with subject in SpecIP and object in SpecAgrOP (c): VSO sentence with subject in SpecVP and object in situ (d): VOS sentence with subject in SpecVP and object in SpecAgrOP (e): OVS sentence with subject in SpecIP and object in SpecCP (f ): OVS sentence with subject in SpecVP and object in SpecCP

Tableau T7.2 Portuguese: Input: {V(x, y), Focus ˆ (V, x, y)} AlignFocus

TopFirst

SubjCase

f a. [ IP S V [VP t t O]] b. [ IP S V [ AgrOP O [VP t t t]]]

Stay

ObjCase

**

*

***!

c. [ IP V [VP S t O]]

*!

*

d. [ IP V [ AgrOP O [VP S t t]]]

*!

**

e. [CP O [ IP S V [VP t t t]]]

*!

f. [CP O [ IP V [VP S t t]]]

*!

*

*

***

*

**

*

Emergence of Unmarked Word Order

179

The reader may have noticed that I have deliberately left out of consideration some word orders and representations. SOV order was left out, since I believe that there is an alternation at the base between head-initial and head-®nal languages. I also assume that this alternation should be analyzed at a deeper level than the one proposed here. Since SOV relates to directionality of theta-role assignment (Hoekstra 1984; Travis 1984; among others), and thematic information in OT is partially expressed in the input (see Grimshaw 1997), one has to work out a way of establishing the link between the information in the input and its correlation with the directionality of assignment. Such a goal is beyond the scope of this chapter. Note nevertheless that it is conceivable to incorporate an alternative analysis of SOV word order into the present approach. Zwart (1993), following Kayne (1994), proposes that SOV languages are derived from SVO with overt movement of the object to SpecAgrOP. This would mean compliance with SubjCase, ObjCase, two violations of Stay, and some violation of the constraint that forces V to move to I. For reasons of coherence with previous work (Costa 1997b), where I explain Dutch scrambling resorting to head ®nality, I will not adopt the latter hypothesis, stressing though that this is not crucial for the present chapter. The set of representations left out are all those in which the verb has not moved to I. The representations are not included for the sake of brevity and clarity: this avoids adding constraints on V-to-I to the tableaux, making their interpretation much easier. For an analysis of V-to-I within OT, see Vikner (chap. 14, this volume). Again, including those candidates would not change the argument developed in this chapter. Since I am not questioning the standard evidence for or against V-to-I movement in the languages under discussion, and since the constraints on V-to-I movement (see Vikner, chap. 14, this volume) do not con¯ict with the constraints to be used in the analysis, I will leave correlating the results of Vikner's work and the results of this chapter for future research. Let us then proceed with a detailed analysis of each candidate, starting with the ungrammatical ones. Candidate (b) vacuously satis®es AlignFocus (like all the others). Since the Subject is in SpecIP, SubjCase is satis®ed. Satisfying ObjCase implies that the Object also moves out of VP, incurring one mark for Stay.2 This makes this candidate worse than candidate (a) for Stay, the decision being made by this constraint. Candidate (c) is ruled out, because the subject does not move to SpecIP, violating SubjCase. This is a case of the Emergence of the Unmarked: in general, violation of SubjCase is allowed in Portuguese, if the subject is focus. In that case, the e¨ects of SubjCase are not visible. Here, since AlignFocus is vacuously satis®ed by all candidates, the e¨ects of SubjCase become apparent, and representations that do not cause this constraint to be violated are preferred. Candidate (d) is ungrammatical for the same reason. It fatally violates SubjCase. Candidates (e) and (f ) are slightly di¨erent, since they violate TopFirst. The reason for the mark on TopFirst is that there is an unnecessary topicalization of the object.

180

JoaÄo Costa

This is perhaps also penalized by a constraint not represented in the tableau prohibiting the fronting of elements that are not operators (for constraints of this type, see Grimshaw 1997; Samek-Lodovici 1996; Grimshaw and Samek-Lodovici 1995, 1998). It is not unlikely that this constraint may have to be split up into two. There are several conceivable ways to express the same idea that topicalization of nontopics should not be permitted: not only does it violate TopFirst, at de®ned above, but it also constitutes a violation of *Struc (a constraint penalizing structure; see Grimshaw 1997), since it involves one extra layer of structure. I am calling this constraint *STRUC and not OBHD (obligatory head) as in Grimshaw 1997, because the focus of this chapter is not heads but phrases in general. The two analyses do not make any di¨erent predictions (see Costa 1998 for discussion). In this chapter, I tend to take it for granted that it is excluded because of a constraint that is functional in nature. One advantage of such an approach is that it derives the fact that OS word orders are rare if not nonexistent as unmarked. This follows from the fact that the top-ranked functional/ discourse-related constraints immediately ®lter out such sequences. The only situations where OVS and OSV candidates may be optimal are cases in which there is no movement of the subject, but the object is obligatorily moved to SpecAgrOP. O-initial languages should, according to this line of reasoning, only be possible in languages without V-to-I movement. Since I am not considering V-to-I movement in this chapter, I will leave this prediction unveri®ed. Note that this distinction between TopFirst and *Struc is not a matter of choice between structural and functional approaches, since both constraints are necessary, as will become evident when we look at Celtic. Unless necessary (see the discussion of Celtic below), I will not represent *Struc in the tableaux. In the next section, I will demonstrate that minimal rerankings between constraints of this set derive the unmarked word orders of the other languages discussed in the introduction. I will try to show that the variation at the base discussed above may be the result of the visibility of the e¨ects of di¨erent dominated constraints, depending on the languages. 7.4

Minimal Rerankings, Base Word Orders, and Variation

If the approach advocated here proves valid, it has two important theoretical consequences, since it reinforces two of the premises of OT. First, it con®rms that constraints are violable. In fact, the Emergence of the Unmarked is not expected in any approach defending absolute constraints, since a switched-o¨ principle/constraint/ parameter should remain switched o¨ independently of speci®c constructions. E¨ects of normally inactive constraints are only expected in a framework permitting violability. Moreover, in a model not contemplating constraint interaction as a legitimate

Emergence of Unmarked Word Order

181

possibility, it is di½cult to explain why some principles are sometimes active and sometimes inactive. Second, the analysis here achieves another major goal of OTÐto provide a formalism to capture linguistic variationÐby showing that reranking a universal set of constraints enables us to capture crosslinguistic di¨erences. 7.4.1 Portuguese/Spanish A vs. Spanish B/Greek: SVO vs. VSO The ®rst di¨erence I will examine is the one between SVO and VSO word order of the type described for Spanish by OrdoÂnÄez and TrevinÄo (1995) and for Greek by Alexiadou and Anagnostopoulou (1995). As noted, discrepancies exist among speakers of Spanish with respect to SVO versus VSO word order. Explaining these discrepancies in terms of a reranking of several constraints would be undesirable, since that might have consequences for the overall grammar of the language that would be di½cult to evaluate; it would also predict di¨erences between the two sets of speakers that are most likely untrue. Hence, in the optimal case, we expect the difference between the rankings for Spanish A and Spanish B to be minimal. If this di¨erence can be explained under a minimal reranking, the null hypothesis is that the same di¨erences in the grammar are responsible for the di¨erences between Portuguese and Greek. Before presenting the crucial rankings, one must examine the factors that make these languages di¨er: according to the descriptions in the literature, the di¨erences seem to lie in the behavior of subjects. According to most descriptions of Portuguese (Raposo 1986, 1996; Mateus et al. 1989; Duarte 1987; Ambar 1992; Martins 1994; Costa 1997),3 preverbal subjects in Portuguese are in SpecIP (or SpecAgrSP). The same is argued for Spanish by Hernanz and Brucart (1987), who also describe Spanish as an SVO language. The argument these authors present for these descriptions is based on the behavior of preverbal subjects as A-moved elements (not intervening for A-bar movement and behaving as A-binders, among other properties). Conversely, Alexiadou and Anagnostopoulou (1995) for Greek and OrdoÂnÄez and TrevinÄo (1995) for Spanish have proposed that these languages are VSO, in the sense that subjects remain in SpecVP and preverbal subjects are merely instances of subject left dislocation. They argue in favor of this analysis on the basis of factors like the following: preverbal subjects either block A-bar movement (Spanish), or interact with wh-phrases in the same way topics do, as (17) suggests for Greek; the interpretation of quanti®ers is di¨erent depending on their positionÐonly preverbal QPs have a strong reading (see (18)); the issue of construing pronouns as bound variables, which is only possible with postverbal ones, showing that only the latter are in an A-position (exempli®ed in (19) by data from Catalan, which, according to the authors, behaves like Spanish and Greek).

182

JoaÄo Costa

(17) Greek (from Alexiadou and Anagnostopoulou 1995) a. Pjon (*o Petros) ide (o Petros)? whom (the Peter-NOM) saw (the-Peter-NOM) `Whom did Peter see?' a 0 . *Pjos ton Petro ton ide? who the Peter-ACC Cl-ACC saw `Who saw Peter?' b. Pote (o laos) apofasise (o laos) na andidrasi? when (the people-NOM) decided (the people-NOM) SUBJ react `When did the people decide to react?' 0 b . Pote tin tenia tin provalan ja proti for a when the movie-ACC Cl-ACC showed-3Pl for ®rst time `When did they show the movie for the ®rst time?' (18) Greek (from Alexiadou and Anagnostopoulou 1995) a. Enas heretise ti Maria. Strong (partitive/speci®c) reading one greeted the Maria-ACC `A certain person/one of the people greeted Mary.' b. Heretise enas ti Maria. Weak (inde®nite) reading `Someone greeted Mary.' (19) Catalan (from SolaÁ 1992 and Alexiadou and Anagnostopoulou 1995) a. *Tots els estudiants i es pensen que ells i aprovaran. all the students think that they will-pass `All the students think that they will pass.' b. Tots els jugadors i estan convencus que guanyaran ellsi . all the players are convinced that will-win they `All the players are convinced that they are the ones who will win.' Notice that, as mentioned earlier, Portuguese subjects are able to co-occur with fronted constituents and wh-elements (20); asymmetries like the one in (19) do not exist (see (21)), since both pre- and postverbal subjects may be construed as bound variables. (20) Portuguese a. Que livros e que o JoaÄo leu? which books is that JoaÄo read `Which books did JoaÄo read?' b. Esses livros, o JoaÄo leu. those books, JoaÄo read `Those books, JoaÄo read.'

Emergence of Unmarked Word Order

183

(21) Portuguese a. Todos os estudantes pensam que eles passaraÄo. `All the students think that they will pass.' b. Todos os estudantes pensam que passaraÄo eles. all the students think that will pass they `All the students think that they will pass.' On the basis of these asymmetries between the two sets of languages, I conclude that the distinction is ®ner than what Alexiadou and Anagnostopoulou (1995) propose; they suggest that null-subject languages all instantiate left dislocation whenever a preverbal subject is present. In these languages, it is the verbal agreement of the verb that bears the theta role and is responsible for satisfying EPP features. In light of these data, it seems to me that null-subject languages do not behave uniformly: some have basic VSO word order and their subjects do not move to SpecIP (Greek and Spanish B); others have basic SVO word order and their subjects may move to SpecIP. The conclusion is then that the crucial di¨erence between these two sets of languages rests on whether preverbal subjects are A-bar or A-bar moved. Such a di¨erence has implications for which of the constraints presented above is satis®ed. Given the de®nition of the constraints just presented, whenever a subject moves to SpecIP, it satis®es SubjCase at the expense of violating Stay, since there is one movement operation. In a language in which SubjCase dominates Stay, it is more important to move the subject than to satisfy economy. This is represented in tableau T7.3. As for left dislocation, I will follow the claim of various authors that this phenomenon does not involve movement but base generation of the left-dislocated element in adjunction to CP (see Duarte 1987; Cinque 1990; Raposo 1996, 1997). As mentioned, when there is left dislocation of the subject, the theta role is assigned to the pronominal agreement (pro or clitics), which never occupies SpecIP. In this chapter, I will follow Duarte's (1996) claim that left dislocation must be treated as a case of base generation, in accordance with Cinque (1990), among others. Duarte shows that Portuguese di¨ers from other Romance languages in not using left dislocation (the way Spanish, Italian, and Greek do) as a strategy for topic promotion. The empirical arguments she presents constitute additional reasons for not grouping Tableau T7.3 SubjCase f a. [IP S V [VP t O]] b. [IP V [VP S t O]]

Stay **

*!

*

184

JoaÄo Costa

Portuguese with the other null-subject languages. Since Portuguese does not use left dislocation for promotion of nonsubject topics, it is unlikely to use it for promoting subjects. If one wants to adhere to the subject-left-dislocation analysis for an SVO sentence in Portuguese, one has the burden of explaining why left dislocation is available for subjects only. Accepting the idea that left dislocation involves base generation of the topic, it is possible to generalize over preverbal subjects in Spanish B and Greek, and assume that preverbal subjects in these languages are base generated in a topic position (SpecCP or adjunction to IP, depending on the analysis). That this is indeed the case is defended by Barbosa (1996), who shows that some preverbal subjects must associate with subject clitics overtly realized in SpecIP in some Romance languages. (22) Trentino (from Barbosa 1996) a. EI Mario *(el) parla the Mario he speaks b. Ti *(te) parli. you you speak If preverbal subjects in these languages are base generated in topic position, a crucial di¨erence between them and preverbal subjects in SpecIP is whether or not there is movement involved. Thus, the option of base generating the subject in this position is more economical, which in our terms means that it does not violate Stay. Whether a language chooses between moving the subjects to SpecIP or base generating them in topic position is then a matter of whether it is more important to satisfy Stay (as in (23)) or SubjCase (as in (24)). The two possible situations are given in tableaux T7.4 and T7.5, where candidate (a) is the one with subject left dislocation and candidate (b) has movement of the subject to SpecIP.4 (23) Stay g SubjCase (24) SubjCase g Stay Tableau T7.4 re¯ects the situation in Spanish B and Greek, where it is more important to have an economical representation (without movement) than satisfying Case. Tableau T7.4 Greek and Spanish BÐContext speci®cation in the input: subject is the topic of the sentence

f a. [C=IP S [IP pro V [VP t O]]] b. [IP S V [VP t t O]]

Stay

SubjCase

*

*

**!

Emergence of Unmarked Word Order

185

Tableau T7.5 Portuguese and Spanish AÐContext speci®cation in the input: subject is the topic of the sentence

a. [C=IP S [IP pro V [VP t O]]] f b. [IP S V [VP t t O]]

SubjCase

Stay

*!

* **

In these languages, both candidates violate Stay once because of the movement of the verb. However, the candidate with movement of the Subject to SpecIP involves one more violations of Stay, the second one being fatal. Tableau T7.5 represents what happens in Portuguese and Spanish A: satisfying SubjCase is more important than having an economical representation, hence the winner is the representation with movement of the subject to SpecIP. One may wonder why there is no representation where the subject is base generated in SpecIP, in compliance with the two constraints. Here, I follow Grimshaw's (1997) claim that the only candidates generated by Gen are those that comply with general principles of X-bar structure and the thematic criterion. According to Koopman and Sportiche (1991), the thematic role of the subject is assigned under sisterhood to V-bar, hence base generation of the subject in SpecIP would yield a violation of the theta criterion, precluding generation of such candidates.5 Naturally, this type of analysis raises questions that fall beyond the scope of this chapterÐquestions related to how the thematic interpretation itself of the left-dislocated subject takes place. In principle, these issues may be addressed within the theory, since the thematic information may be not expressed by a local relation between theta-role assigner and assignee in the output, but it will always be represented in the input. If the input contains all the semantic information (for discussion, see Grimshaw 1997; Legendre et al. 1995; Costa 1998), it does not follow in any way that a subject that does not get its theta role in a local relation with the assigner will not be interpreted as the thematic subject of the verb. Also, if Alexiadou and Anagnostopoulou's analysis is correct, the theta role is assigned to the subject agreement marker, not to the leftdislocated item. So far, I have just represented the choice between SVO with movement to SpecIP and SVO with left dislocation of the subject. I still have to explain why for these two language groups the unmarked word order is SVO and VSO, respectively. The crucial part ot the explanation is again the ranking between SubjCase and Stay, though TopFirst also plays a role. Consider the following three candidates: SVO with left dislocation, SVO with subject in SpecIP, and VSO. Since the context is sentence focus, all orders will violate TopFirst once, for the ®rst element of the sen-

186

JoaÄo Costa

tence is not a topic. In addition, the candidate involving left dislocation of the subject will violate it once more, since there is topicalization of a nontopic element. This is similar to the way I excluded OSV and OVS as potential base word orders. The prediction is then that there should be no unmarked word order involving left dislocation of any type. Having controlled the e¨ects of AlignFocus and TopFirst, it is now possible to see whether the choice is made by SubjCase or by Stay. As expected, in Portuguese and Spanish A, it will be more important to satisfy Case than Stay, hence subjects will move to SpecIP. This is represented in tableau T7.6. In Greek and Spanish B, it is more important to satisfy Stay than SubjCase, hence subjects will not move to SpecIP, yielding a more economical representation. This is schematized in tableau T7.7. It is important to note that, in spite of the relevance of TopFirst for the exclusion of the candidate with left dislocation of the subject, it is the ranking between Stay and SubjCase that is crucial. The decision between candidates (b) and (c) is made at the point one of these two constraints (depending on the ranking) evaluates the candidates. In tableau T7.6, SVO is selected, because only in this order has the subject moved to SpecIP, satisfying SubjCase. In tableau T7.7, because of the reverse ranking, Stay prohibits SVO, since there are two traces in the structure as opposed to one in the VSO order. Tableau T7.6 Portuguese and Spanish AÐContext speci®cation in the input: sentence focus AlignFocus a. [CP S [IP V [VP t O]]]

TopFirst

SubjCase

Stay

ObjCase

*!

*

*

*

**

*

*

*

f b. [IP S V [VP t t O]] c. [IP V [VP S t O]]

*!

Tableau T7.7 Greek and Spanish BÐContext speci®cation in the input: sentence focus AlignFocus a. [CP S [IP V [VP t O]]] b. [IP S V [VP t t O]] f c. [IP V [VP S t O]]

TopFirst

Stay

SubjCase

ObjCase

*!

*

*

*

**! *

* *

*

Emergence of Unmarked Word Order

187

Note that one of the purposes in capturing the di¨erence between SVO and VSO has been achieved: given the di¨erence between Spanish A and Spanish B, which is dialectal,6 it is not desirable to propose a substantial di¨erence in constraint pro®les. Indeed, the di¨erence between the two languages is minimal; only SubjCase and Stay need to be reranked with respect to one another. Statements (25) and (26) indicate the rankings for Portuguese and Spanish A, and Greek and Spanish B, respectively.7 (25) Portuguese and Spanish A {AlignFocus, TopFirst} g SUBJCASE g STAY g ObjCase (26) Greek and Spanish B {AlignFocus, TopFirst} g STAY g SUBJCASE g ObjCase In this way, I can control the consequence of reranking these constraints for other aspects of the grammar of the languages under consideration, since SubjCase has in its scope a quite restricted set of elements. If I were to radically modify the constraint pro®le, I would be very likely to predict di¨erences between the two dialects of Spanish not observed in actual data. 7.4.2 Postverbal Subjects: VSO vs. VOS In the preceding section, I have derived the di¨erence between SVO and VSO as unmarked word orders for Portuguese and Spanish A and Greek and Spanish B, respectively. In this section, I would like to develop the analysis presented there in order to accommodate a subtle di¨erence between Portuguese and another Romance language: Italian. Italian is like Portuguese, in that its unmarked word order is SVO, as (27) illustrates.8 (27) Italian (from Pinto 1997) Che cosa eÁ sucesso? What happened? a. Beatrice ha scritto lettere d' amore. Beatrice has written letters of love `Beatrice has written love letters.' b. aHa scritto lettere d' amore Beatrice. has written letters of love Beatrice One crucial way Italian di¨ers from Portuguese is that though VOS orders are allowed when only the subject is in focus, as in (28), VSO orders legitimate in Portuguese are ungrammatical in Italian, independently of the context (see (29); also noted by Rizzi 1982 and Burzio 1986, among others).

188

JoaÄo Costa

(28) Italian (from Pinto 1997) Chi ha scritto lettere d'amore? a. aBeatrice ha scritto lettere d' amore. Beatrice has written letters of love `Beatrice has written love letters.' b. Ha scritto lettere d' amore Beatrice. has written letters of love Beatrice (29) Nessuno ha scritto niente. No one has written anything a. *Ha scritto Beatrice lettere d'amore. b. Ha scritto lettere d'amore Beatrice. Remember that the context in (30) is the one in which VSO orders are legitimate in Portuguese. (30) NingueÂm escreveu nada. a. Escreveu a Beatriz cartas de amor. wrote the Beatriz letters of love b. ?aEscreveu cartas de amor a Beatriz. wrote letters of love the Beatriz `Beatriz wrote love letters.' Here I will follow Pinto's analysis of VOS order. Pinto claims that in spite of the lack of evidence from adverbial placement (see Belletti 1990), VOS orders may be analyzed as an instance of subject in SpecVP with movement of the object to SpecAgrOP, as in the representation in (31). (31) [ IP V [ AgrOP O [VP S t t]]] Such an analysis is also defended in Cinque 1999. Cinque argues for this analysis on theoretical grounds by rejecting right adjunction as a possibility, and by the observation that postverbal subjects in Italian transitive constructions must be absolutely ®nal. Actually, Pinto's observation that inverted subjects may precede VPadjuncts (see (32)) con®rms this hypothesis, since we would otherwise expect to ®nd right-dislocated subjects following all base adjuncts. (32) Italian (from Pinto 1997) a. E' arrivato Dante da Firenze. is arrived Dante from Florence `Dante has arrived from Florence.' b. Ha telefonato Beatrice da Milano. has called Beatrice from Milan `Beatrice has called from Milan.'

Emergence of Unmarked Word Order

189

The only way to rescue the VP-adjoined position for subjects in sentences like (32) would be to assume a multiple right-adjunction structure for the VP, which would nevertheless be problematic, since as Samek-Lodovici (1996) argues, this position (right adjunction to VP) appears reserved for contrastive foci, on which a restriction of uniqueness applies (see Szabolcsi 1981). I will therefore follow Pinto (1997) and Cinque (1999) for the structural representation of VOS order in Italian.9 The question now is whether we can represent the situation in Italian with the set of constraints used earlier. Before trying possible rankings, it is important to understand where the crucial di¨erence between Italian and Portuguese lies. What seems at stake is that Italian has obligatory object movement to SpecAgrOP, while Portuguese does not have it at all.10 As the examples above illustrated, not even context may force a VSO word order in Italian. In our system, this means that the constraint driving the movement of the object is more important than constraints related to discourse. That is, the ranking between ObjCase and AlignFocus has to be the one in (33). (33) ObjCase g AlignFocus Since subjects do not behave this way (their position is dependent on the discourse context), the ranking between AlignFocus and SubjCase has to be kept the same as for Portuguese. (34) AlignFocus g SubjCase Since preverbal subjects are in SpecIP, according to most analyses of Italian (Rizzi 1982; Burzio 1986; Belletti 1990; Samek-Lodovici 1996; Pinto 1997; among many others), it may not be the case that Stay dominates SubjCase (which also explains that scrambling as adjunction to VP is not possible in this language). (35) SubjCase g Stay Since topicalization is possible in Italian (see Rizzi 1995, among others), TopFirst has to be more important than ObjCase. (36) TopFirst g ObjCase We thus arrive at the partial ranking for Italian given in (37). (37) TopFirst g ObjCase g AlignFocus g SubjCase g Stay Though it seems quite di¨erent from the ranking for Portuguese presented in (38), notice that again I only had to rerank one constraint (ObjCase). This is highly desirable, since this word order variation takes place within a single language family. (38) {AlignFocus, TopFirst} g SubjCase g Stay g ObjCase

190

JoaÄo Costa

Tableau T7.8 Portuguese and Spanish AÐContext speci®cation in the input: subject and object are focused AlignFocus

TopFirst

f a. [IP V [VP S t O]]

SubjCase

Stay

ObjCase

*

*

*

b. [IP V [AgrOP O t [VP S t t]]]

*!

*

***

c. [IP V [VP O [VP S t t]]]

*!

*

**

*

Tableau T7.9 ItalianÐContext speci®cation in the input: subject and object are focused TopFirst a. [IP V [VP S t O]]

ObjCase

SubjCase

Stay

*

*

*

*

***

*

*

**

*!

f b. [IP V [AgrOP O t [VP S t t]]] c. [IP V [VP O [VP S t t]]]

AlignFocus

*!

Let us now see how the selection of candidates is done. For this case, it is necessary to consider a context where both subject and object are focused. Tableau T7.8 represents the evaluation tableau for Portuguese, where VSO wins. Tableau T7.9 is the evaluation tableau for Italian, where VOS wins. The VSO candidate and the VOS candidates with movement of the object to SpecAgrOP and with scrambling of the object in adjunction to VP are included in the set of candidates under comparison. As can be seen in the tableaux, the decision regarding which candidate is optimal is made early in the optimization, in the sense that it is done by one of the top-ranked constraints. In this sense, one might think that the ranking of SubjCase and Stay in Italian is irrelevant, since ObjCase immediately ®lters out all but one candidate. However, again the e¨ects of the Emergence of the Unmarked have to be considered. Only the ranking proposed may accommodate the facts of postverbal subjects and the emergence of SVO order in the unmarked case. This is shown in tableau T7.10, where the candidates listed in (39) are considered. (39) Candidate (a): SVO with Subject in SpecIP and Object in SpecAgrOP Candidate (b): SVO with Subject in SpecIP and Object in situ

Emergence of Unmarked Word Order

191

Tableau T7.10 ItalianÐContext speci®cation in the input: sentence focus TopFirst

ObjCase

AlignFocus

SubjCase

f a. [IP S V [AgrOP O t [VP t t t]]]

Stay ****

b. [IP S V [VP t t O]]

*!

**

c. [IP S V [VP O [VP t t t]]]

*!

***

d. [CP S [IP V [AgrOP O t [VP t t]]]

*!

e. [CP S [IP V [VP t O]]]

*!

f. [CP S [IP V [VP O [VP t t]]]

*!

g. [IP V [VP S t O]]

*

***

*

*

*

*

*

**

*!

*

*

*!

***

*

**

h. [IP V [AgrOP O t [VP S t t]]] i. [IP V [VP O [VP S t t]]] Candidate Candidate Candidate Candidate Candidate Candidate Candidate

*!

(c): SVO with Subject in SpecIP and scrambled Object (d): SVO with Subject right dislocated and Object in SpecAgrOP (e): SVO with Subject right dislocated and Object in situ (f ): SVO with Subject right dislocated and scrambled Object (g): VSO with Subject and Object in situ (h): VOS with Subject in SpecVP and Object in SpecAgrOP (i): VOS with Subject in SpecVP and scrambled Object

Let us recap what happens in tableau T7.10. This tableau shows that SubjCase and Stay, though dominated by three constraints, are active. Let us look at the place where candidates fatally violate one constraint: all the candidates with left dislocation of the subject fatally violate TopFirst, since they involve topicalization of a nontopic (similar to what happens in Greek). ObjCase ®lters out all candidates with the object in situ or scrambling, precluding the emergence of VSO order. A decision now has to be made between SVO and VOS (candidates (a) and (h), respectively). AlignFocus does not play any role, hence SubjCase will determine the optimal candidate: SVO. The VOS order fatally violates SubjCase. Note that in this tableau, it is also possible to observe the crucial ranking between SubjCase and Stay. If Stay would dominate AlignFocus, the VOS order would emerge, since it only involves three movement operations, against the four movements involved in representation (a).

192

JoaÄo Costa

Summing up, once again, it is possible to describe a language by virtue of a minimal reranking of the assumed constraints and to show that the e¨ects of dominated constraints are visible in sentence-focus context. 7.4.3 VOS in Italian vs. VOS in Malagasy The analysis proposed for Italian may be extended to accommodate the VOS word order that emerges as unmarked in Malagasy, as (40) illustrates. (40) Malagasy (from Keenan 1976) Nividy mofo ho'an'ny ankizy aho. bought bread for the children I `I bought bread for the children.' Word order in Malagasy, as opposed to Italian, is rigid: complements of the verb may not follow the subject. (41) a. *Nividy bought b. *Nividy bought

aho mofo ho'an'ny ankizy. I bread for the children ho'an'ny ankizy aho mofo. for the children I bread

Likewise, SVO orders are not allowed (unless there is some kind of cleft construction or topicalization that is morphologically marked; see Keenan 1976). (42) *Aho nividy mofo ho'an'ny ankizy. I bought bread for the children Evidence for analyzing this word order in the same terms as the analysis proposed for Italian, repeated in (43), comes from the distribution of question, exclamatory, and ``no longer'' particles. These particles are often used to determine the position of the subject (see, for example, Diesing 1992 for German). In Malagasy, they always precede the subject. (43) [ IP V [ AgrOP O [VP S t t]]] (44) a. Nanome vola an-dRabe ve ianao? gave money acc-Rabe Q you `Did you give money to Rabe?' b. Manasa lamba anie Rasoa! washes clothes EXCL Rasoa `Is Rasoa still washing clothes!' c. Tsy manasa lamba intsony Rasoa. not washes clothes longer Rasoa `Rasoa is no longer washing clothes.'

Emergence of Unmarked Word Order

193

From the preceding examples, one may hypothesize that subjects in Malagasy are VP-internal.11 Expressing this in terms of ranking, it is necessary to capture the fact that VOS is the unmarked word order. This result may be achieved by using the same ranking as for Italian and changing the relative ranking of Stay and SubjCase. For Malagasy, the ranking between these two constraints must be the ranking speci®ed by (45). (45) Stay g SubjCase This predicts that it is more important for a representation to be economical than for an NP to be assigned nominative Case in SpecIP. This, however, cannot be all. The discourse-related constraint AlignFocus must be relatively low ranked in this language, since di¨erent contexts do not permit different word orders, in contrast to the situation in the other languages considered in this chapter. Since subjects are always sentence ®nal, I propose that AlignFocus is ranked below SubjCase. I will not make any proposal concerning TopFirst: since there are topicalizations in this language, TopFirst must be relatively high ranked. However, it is unclear how it interacts with the other constraints. For the sake of exposition, I will leave it top ranked. Additional constraints related to morphologically marked topicalization and focus-related constraints impose requirements on the position of contrastively focused elements. They will most likely be ranked above the constraints discussed here, and trigger morphologically marked topicalization and clefting. Let us now look at the evaluation tableau (tableau T7.11) and see how the ranking proposed derives the word order in Malagasy. Tableau T7.11 shows the irrelevance of SubjCase and AlignFocus for deriving the unmarked word order of this language. The decision is made by TopFirst, which rules out all candidates with left dislocation of the subject; by ObjCase, which rules out all candidates without movement of the object to SpecAgrOP; and by Stay, which decides between candidates (a) and (h), preferring the most economical representationÐthe one without movement of the subject. This case is slightly di¨erent from the other languages discussed, since it is not the case that AlignFocus dominates everything else and it becomes necessary to look at sentence-focus contexts to see the e¨ects of the other constraints. It is nevertheless worth noting that a rearrangement of the constraints derives the unmarked word order of this language. The role played by TopFirst (if any; see the discussion above) is, however, the same as the role played by AlignFocus in the other languages. 7.4.4 More on VSO: Celtic and Arabic The preceding discussion on Malagasy VOS as opposed to Italian VOS enables me to give an account of the opposition between VSO in Celtic versus VSO in Arabic, Cha-

194

JoaÄo Costa

Tableau T7.11 MalagasyÐContext speci®cation in the input: sentence focus TopFirst

ObjCase

a. [IP S V [AgrOP O t [VP t t t]]]

Stay

*!

**

c. [IP S V [VP O [VP t t t]]]

*!

***

d. [CP S [IP V [AgrOP O t [VP t t]]]

*!

e. [CP S [IP V [VP t O]]]

*!

f. [CP S [IP V [VP O [VP t t]]]

*!

***

*

*

*

*

*

**

*

*!

*

*

***

*

**

*

f h. [IP V [AgrOP O t [VP S t t]]] i. [IP V [VP O [VP S t t]]]

AlignFocus

****!

b. [IP S V [VP t t O]]

g. [IP V [VP S t O]]

SubjCase

*!

morro, and Berber. The di¨erence between these languages is, according to Ouhalla (1991), that Berber, Chamorro, and Arabic are VSO languages allowing SVO as a possible word order (namely, when the subjects are topics), while Celtic never allows SVO: it is a rigid VSO language.12 This is similar to the alternation we found between SVO and VOS in Italian versus VOS in Malagasy, though the relation word order/ function is di¨erent. Examples (46) to (48), taken from Ouhalla 1991, illustrate these patterns for Berber, Arabic, and Welsh. (46) Berber a. adysegh Moha ijn teddart. fut (TNS)-3ms (AGR)-buy Moha one house `Moha will buy a house.' b. Moha ad- y- segh ijn teddart. Moha fut 3ms buy a house `Moha will buy a house.' (47) Arabic a. Sayashtarii Zayd-un dar-an fut (TNS) 3ms (AGR)-buy Zayd-nom house-acc `Zayd will buy a house.'

Emergence of Unmarked Word Order

195

b. Zayd-un sayashtarii daar-an. Zayd-nom fut (TNS) 3ms (AGR)-buy house-acc `Zayd will buy a house.' (48) Welsh a. Gwelodd y bechgyn y draig. saw the boys the dragon `The boys saw the dragon.' b. *Y bechgyn gwelodd y draig. `The boys saw the dragon.' Ouhalla's explanation for this di¨erence, though interesting, will not be integrated here. He notes that Welsh di¨ers from the other languages not only in not having SVO, but also in two other features characteristic of SVO languages: Celtic languages have nonin¯ected in®nitives, and the order of in¯ectional morphemes is TnsAgr. Ouhalla suggests that word order alternations (VSO enables SVO, but not vice versa), the order of morphemes, and the availability of nonin¯ected in®nitivals are three consequences of one parametric di¨erence: whether, in the clause structure, AGR selects T or T selects Agr. He shows that all the VSO properties follow from a structure where T selects Agr, and the SVO properties from a structure where Agr selects T. Under Ouhalla's analysis, in a language where T selects Agr, the subject ends up in SpecAgrSP, not needing to move further up, though SpecTP remains a legitimate position for topics. In the languages where Agr selects T, the subject has to move all the way up to Agr for Case purposes, yielding the SVO order. Celtic falls within the latter group of languages, with a language-particular rule stating that subjects are assigned nominative Case in SpecVP. The reason for not integrating this analysis is that it resorts to two mechanisms not available in OT: language-particular rules and parametrization of functional heads. In OT, constraints are universal, and all language variation has to follow from constraint ranking. Furthermore, I have been following the claims put forward in the literature for a dialect of Spanish that is VSO, though the language exhibits the order of morphemes TnsAgr, a counterexample to Ouhalla's generalization.13 Let us ®rst consider Arabic. This language has an analysis similar to the one proposed for Greek and Spanish B: VSO in the unmarked case is a consequence of the dominance of SubjCase by Stay. The dependence on discourse context will follow from the high ranking of TopFirst and AlignFocus. Hence, I propose exactly the same analysis as proposed for Greek and Spanish B, represented in tableau T7.12. Departing from the assumption made above for Malagasy that rigid word order is a consequence of the low ranking of the discourse-related constraints, I can now explain why Celtic has a rigid VSO order. This order will emerge in the context of sentence focus, represented in tableau T7.13.

196

JoaÄo Costa

Tableau T7.12 Arabic and BerberÐContext speci®cation in the input: sentence focus AlignFocus a. [CP S [IP V [VP t O]]]

TopFirst

Stay

SubjCase

ObjCase

*!

*

*

*

b. [IP S V [VP t t O]]

**!

f c. [IP V [VP S t O]]

*

*

*

*

ObjCase

AlignFocus

TopFirst

Tableau T7.13 CelticÐFirst attempt

a. [CP S [IP V [VP t O]]] b. [IP S V [VP t t O]] f c. [IP V [VP S t O]]

Stay

SubjCase

*

*

**! *

*! *

*

*

Note that it is still TopFirst that decides between candidates (a) and (c). However, the low ranking of TopFirst is important to ensure that the same word order will be optimal when the subject is not in focus. This is what happens in tableau T7.14: in this tableau it is necessary to resort to other structural constraints, so that we can see that structural constraints really do determine the VSO word order and that the discourse-related constraints are very low ranked. Hence, in tableau T7.14, I add *Struc, which prohibits generation of phrase structure and is fatally violated by the candidate with CP projected. If this constraint did not play a role, candidate (a) would be optimal. Given the main claim made in this chapter as well as in Grimshaw and Samek-Lodovici 1995 and Costa 1997a, 1997c that optionality arises when discourse constraints dominate syntactic constraints, while rigid word order arises when the converse occurs, it is not surprising that *Struc dominates both AlignFocus and TopFirst. If that were not the case, candidate (a) would emerge as optimal, as in Greek and Spanish B. Note that adding *Struc to the constraint pro®le is not a stipulation to explain the Celtic facts. This constraint is independently used in the work of Legendre and Bresnan, and the role it plays here con®rms the idea that the rigid word order follows from the low rank of AlignFocus and TopFirst.14

Emergence of Unmarked Word Order

197

Tableau T7.14 CelticÐContext speci®cation in the input: subject is topic

a. [CP S [IP V [VP t O]]] b. [IP S V [VP t t O]] f c. [IP V [VP S t O]]

Stay

SubjCase

ObjCase

*

*

*

***!

*

**

*

**

**! *

*

*Struc AlignFocus

*

TopFirst

*

By not adopting the whole of Ouhalla's analysis, it becomes more di½cult to account for the relationship between weak and strong agreement and position of the subject. Ouhalla proposes that subject agreement correlates with the position of the subject at S-structure: Arabic and Celtic have weak agreement because the verb is in SpecVP, while Berber and Chamorro have rich agreement because the verb is in SpecAgrSP. I do not fail to account for the agreement patterns, provided that I adopt the strong lexicalist hypothesis (Chomsky 1995), and as long as I make sure that agreement is explained somewhere else in the model. In fact, morphological alignment constraints may account for agreement (McCarthy and Prince 1994; Grimshaw 1997; Legendre, forthcoming). Since Celtic and Arabic have a similar agreement pattern, it is not clear that there is a correlation between the typological groups Ouhalla proposes and richness of agreement. In that case, I might lose a syntactic account of agreement, but this way I am able to account for the position of the subjects in a uniform way and without resorting to a language-particular rule. More important, if the analysis of Portuguese subjects proposed above is accepted and if Ouhalla's suggestion is accurate, one would expect to ®nd di¨erences in agreement depending on whether subjects are in SpecIP or in SpecVP. This prediction is, however, not borne out. (49) a. Comeram os meninos a sopa. ate-3pl the boys the soup b. Os meninos comeram a sopa. the boys ate the soup 7.5

Conclusion

In this chapter, I have tried to derive the unmarked word orders of several languages in terms of the Emergence of the Unmarked. I have shown that the e¨ect of constraints normally invisible by virtue of the domination by discourse-related constraints becomes visible in the context of sentence focus. For instance, in languages

198

JoaÄo Costa

where the position of the subject is normally determined by the top-ranked constraint AlignFocus, like Portuguese and Greek, the vacuous satisfaction of this constraint in the context of sentence focus makes the e¨ects of the dominated constraints SubjCase and Stay visible. The rankings summarized in (50) were proposed. (50) a. Portuguese and Spanish A {TopFirst, AlignFocus} g SubjCase g Stay g ObjCase b. Spanish B, Greek, Arabic, Berber, and Chamorro {TopFirst, AlignFocus} g Stay g SubjCase g ObjCase c. Italian TopFirst g ObjCase g AlignFocus g SubjCase g Stay d. Malagasy ObjCase g Stay g SubjCase g TopFirst g AlignFocus e. Celtic Stay g SubjCase g ObjCase g *Struc g {TopFirst, AlignFocus} The only word order not considered here is SOV, for the reasons discussed in section 7.1 and because accounting for it would mean taking into account the constraints on V-to-I discussed in Grimshaw 1997 and Vikner (chap. 14, this volume). These constraints are beyond the scope of this chapter. If the analysis presented here is on the right track, it is a step forward in showing the power of OT as a theory of language variation and constraint interaction, since minimal di¨erences between languages may be expressed by means of minimal rerankings between constraints. Notes Thanks to the following people who commented on preliminary versions of this chapter: Sjef Barbiers, Teun Hoekstra, GeÂraldine Legendre, Gereon MuÈller, Vieri Samek-Lodovici, Johan Rooryck, and audiences in Lisbon, Leiden, Utrecht, and Amsterdam. 1. Actually, in Costa 1997c I did not consider TopFirst, since this constraint was not relevant there. As will become evident, this constraint is only necessary when OS orders are considered. 2. The discussion here is simpli®ed to make the general argument clearer. Actually, candidate (b) should have four marks for Stay, since satisfying ObjCase involves projecting AgrOP, which creates one more landing site for the verb. I am not including that mark in order to make the reading of the diagrams easier, and because I am leaving verb movement out of the discussion entirely. 3. See, however, Barbosa 1995 for a di¨erent position, as well as Costa 1997a and Duarte 1996 for counterarguments to her analysis. 4. Note that VSO word order in this context is independently ruled out by AlignFocus, since this order forces the subject to be interpreted as focus, which is not the case: we are not considering contexts in which the subject is the focus.

Emergence of Unmarked Word Order

199

5. Note that there may be candidates with arguments unrealized and theta roles undischarged. This option is permitted under Samek-Lodovici's (1996) formulation of the theta criterion, which imposes requirements on locality relations between assigners and assignees only when the theta role is discharged. 6. This statement must be quali®ed. I do not have information enabling me to tell whether the di¨erence between VSO and SVO in Spanish is dialectal or ideolectal. For the purposes of this chapter, it is enough to note that variation exists among speakers, and that the di¨erences between the rankings capturing it ought to be minimal. See Legendre 1998 for an account of register variation in French in terms of minimal rerankings of constraints. 7. The rankings proposed also predict that all these languages have scrambling of objects as A-bar movement (in adjunction to VP) (see Costa 1997b), if they have scrambling. This is partially con®rmed for Spanish by OrdoÂnÄez (forthcoming), who shows that object movement in Spanish is A-bar. In their comparative work on clitic doubling and scrambling, Alexiadou and Anagnostopoulou (1996) show that Greek objects do not move to SpecAgrOP (instead, there is clitic doubling in this language). What these rankings do not predict is that subjects and objects behave uniformly with respect to Case licensing within the same language, as Chomsky (1995) proposes. The facts of Portuguese do not seem to con®rm such a proposal, hence I do not see this as a weakness in my analysis. 8. In her description of neutral word orders, Pinto notes, following Calabrese (1991) among others, that some inverted structures are felicitous as answers to the question ``What happened''. That is the case with verbs that select some kind of locative argument. Though Pinto's work focuses mainly on such constructions, I will overlook them here. They are not possible with transitive verbs and are thus irrelevant for the present discussion, since it becomes impossible to test the ordering between subject and object. I refer the reader to Pinto 1997 for a description of the contexts of VS order in Italian. The possibility of having VS orders with such verbs may be derived from restrictions on the discourse situation imposed by the lexical meaning of the verbs or by the locative argument itself. 9. Note that I am only partially following Cinque, since he does not include Agr-phrases in the inventory of potential functional projections, following Chomsky (1995). 10. I have not presented the facts here, but there seem to be reasons to believe that Portuguese VOS orders are derived via adjunction to VP instead of movement to SpecAgrOP, a fact explained in Costa 1997c by the domination of ObjCase by Stay. The explanation is as follows: by satisfying ObjCase, a representation is forced to project AgrOP, which creates one more head position for the verb to land in on its way to I, yielding one extra violation of Stay and rendering the representation less economical. 11. One problem with this analysis is how to explain the obligatory movement to AgrOP of nonnominal complements. At this stage, I do not have anything interesting to suggest. Another surprising factor concerning this analysis is the De®niteness restriction presented by Keenan: subjects must be de®nite. This is strange vis-aÁ-vis our conclusion, since it is more common for VP-internal subjects to be inde®nite. Keenan presents data showing that whenever there is an inde®nite subject, a special type of existential construction must be used. This special construction is probably part of the universal candidate set for the relevant input, blocking the use of inde®nites in the canonical subject position. 12. But see Legendre (chap. 9, this volume) on Breton for an example of a Celtic language that also allows SVO and OVS orders.

200

JoaÄo Costa

13. Note that nothing prevents an extension of Ouhalla's analysis from Celtic to Spanish: Spanish might be an Agr-initial language with the language-particular rule assigning nominative Case in SpecVP. 14. One possible analysis not involving *Struc and still deriving the correct results would be to follow Sproat (1985) and assume that VSO in Celtic is an instance of I-to-C, which would make candidate (a) in tableau T7.14 violate Stay twice. McCloskey (1992), Carnie (1995), and Bobaljik (1995) have, however, argued based on the distribution of complementizers that such an analysis is problematic. I will follow the V-to-I analyses, which permit a uni®cation with the Greek and Spanish case. Gereon MuÈller has suggested to me that, if in Celtic there is no option of inserting pro to make sure that the theta role of the subject is discharged, the analysis without *Struc still makes the right predictions. The VSO candidate is still the winner since it is the one involving fewer movement operations. I cannot adopt this suggestion, since topicalization in Celtic is often associated with strong agreement (see Tallerman 1996 for Welsh), not di¨ering from what was observed for null-subject languages in Barbosa 1995 and Alexiadou and Anagnostopoulou 1995. Tallerman (1996) shows that in Welsh, strong agreement occurs in constructions involving pronominals and topicalization, and not in constructions involving movement. Since topicalization (abnormal sentences in Tallerman 1996) is associated with strong agreement, it must not be considered a case of movement, as MuÈller's proposal would suggest. I must therefore resort to *Struc and cannot explain the Celtic behavior in terms of economy of movement alone. References Alexiadou, Artemis, and Elena Anagnostopoulou. 1995. SVO and EPP in Null Subject Languages and Germanic. In FAS Papers in Linguistics, 1±21. Potsdam. Alexiadou, Artemis, and Elena Anagnostopoulou. 1996. Symmetries, Asymmetries, and the Role of Agreement. Paper presented at GLOW XIX, Athens. Ambar, Manuela. 1992. Para uma Sintaxe da InversaÄo Sujeito Verbo em PortugueÃs. Doctoral dissertation, University of Lisbon. Barbosa, Pilar. 1995. Null Subjects. Doctoral dissertation, MIT, Cambridge, Mass. Barbosa, Pilar. 1996. A New Look at the Null Subject Parameter. In JoaÄo Costa, Rob Goedemans, and Ruben van de Vijver (eds.), Proceedings of ConSOLE 3, 375±395. Leiden University. Battistella. 1996. The Logics of Markedness. Oxford: Oxford University Press. Belletti, Adriana. 1990. Generalized Verb Movement. Turin: Rosenberg and Sellier. Bobaljik, Jonathan. 1995. Morphosyntax: The Syntax of Verbal In¯ection. Doctoral dissertation, MIT, Cambridge, Mass. Bresnan, Joan. 1996. Notes on Projections, Heads, and Optimality. Ms., Stanford University, Stanford, Calif. Burzio, Luigi. 1986. Italian Syntax. Dordrecht: Kluwer. Calabrese, Andrea. 1991. Some Remarks on Focus and Logical Structures in Italian. In Harvard Working Papers in Linguistics 1, 91±127. Department of Linguistics, Harvard University, Cambridge, Mass.

Emergence of Unmarked Word Order

201

Carnie, Andrew. 1995. Nonverbal Predication and Head Movement. Doctoral dissertation, MIT, Cambridge, Mass. Choi, Hye-Won. 1996. Optimizing Structure in Context: Scrambling and Information Structure. Doctoral dissertation, Stanford University, Stanford, Calif. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Cinque, Guglielmo. 1990. Types of A-Bar Dependencies. Cambridge, Mass.: MIT Press. Cinque, Guglielmo. 1999. Adverbs and the Universal Hierarchy of Functional Projections. New York: Oxford University Press. Costa, JoaÄo. 1996. Focus in Situ: Evidence from Portuguese. Probus. Costa, JoaÄo. 1997a. Positions for Subjects in European Portuguese. In Proceedings of the West Coast Conference on Formal Linguistics XV. Stanford, Calif.: CSLI Publications. Costa, JoaÄo. 1997b. Scrambling in European Portuguese. In Benjamin Bruening (ed.), MIT Working Papers in Linguistics 31: Proceedings of SCIL 8, 95±114. MITWPL, Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Costa, JoaÄo. 1997c. Word Order and Constraint Interaction. In SeminaÂrios de LinguõÂstica, Universidade do Algarve, Faro. Costa, JoaÄo. 1998. Word Order Variation: A Constraint-Based Approach. Doctoral dissertation, Holland Institute of Generative Linguistics/Leiden University. Derbyshire, D. 1977. Word Order Universals and the Existence of OVS Languages. Linguistic Inquiry 8: 590±599. Diesing, Molly. 1992. Inde®nites. Cambridge, Mass.: MIT Press. Dik, Simon. 1978. Functional Grammar. Dordrecht: Foris. Duarte, IneÃs. 1987. A ConstrucËaÄo de TopicalizacËaÄo na GramaÂtica do PortugueÃs. Doctoral dissertation, University of Lisbon. Duarte, IneÃs. 1996. A topicalizacËaÄo em PortugueÃs Europeu: uma anaÂlise comparativa. In Actas do Congresso Internacional sobre o PortugueÃs, vol 1. Lisbon: APL/Colibri. Fassi-Fehri, A. 1989. Agreement, Incorporation, Pleonastics, and VSO-SVO Order in Arabic. Ms., Faculte des Lettres, Rabat. Frota, SoÂnia. 1994. Is Focus a Phonological Category in Portuguese? In P. Ackema and M. Schoorlemmer (eds.), Proceedings of ConSOLE 1. The Hague: Holland Academic Graphics. Frota, SoÂnia. 1997. Focus and Prosodic Interface. Paper presented at the Primeiro Encontro Informal de Doutorandos em LinguõÂstica, University of Lisbon. GivoÂn, Talmy. 1984. Syntax: A Functional-Typological Introduction. Amsterdam: Benjamins. Grimshaw, Jane. 1997. Projections, Heads, and Optimality. Linguistic Inquiry 28(3): 373±422. Grimshaw, Jane, and Vieri Samek-Lodovici. 1995. Optimal Subjects. In J. Beckman, L. W. Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 598±606. GLSA, University of Massachusetts, Amherst. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Itagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics.

202

JoaÄo Costa

Hernanz, Maria Lluisa, and Jose Maria Brucart. 1987. La sintaxis. Madrid: Editorial CrõÂtica. Huang, James. 1984. On the Distribution and Reference of Empty Pronouns. Linguistic Inquiry 15: 531±574. Kayne, Richard. 1994. The Antisymmetry of Syntax. Cambridge, Mass.: MIT Press. Keenan, Ed. 1976. On Remarkable Subjects in Malagasy. In Charles Li (ed.), Subject and Topic. New York: Academic Press. Koopman, Hilda, and Dominique Sportiche. 1991. The Position of Subjects. Lingua 85: 211± 258. Legendre, GeÂraldine. 1998. PF Cliticization and Complex Inversion in French. Paper presented at the Second Workshop on OT-Syntax, Stuttgart University. Legendre, GeÂraldine. 1999. Why French Stylistic Inversion Is Optimal. Ms., Johns Hopkins University, Baltimore. Legendre, GeÂraldine. Forthcoming. Morphological and Prosodic Alignment of Bulgarian Clitics. In J. Dekkers, F. van der Leeuw, and J. van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Legendre, GeÂraldine, Colin Wilson, Paul Smolensky, Kristin Homer, and William Raymond. 1995. Optimality and Wh-Extraction. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 607± 636. GLSA, University of Massachusetts, Amherst. Li, Charles. 1976. Subject and Topic. New York: Academic Press. Martins, Ana Maria. 1994. Os ClõÂticos na HistoÂria do PortugueÃs. Doctoral dissertation, University of Lisbon. Mateus, Maria Helena, IneÃs Duarte, Isabel Faria, and Ana Brito. 1989. GramaÂtica da LõÂngua Portuguesa. Lisbon: Editorial Caminho. McCarthy, John, and Alan Prince. 1994. The Emergence of the Unmarked: Optimality in Prosodic Morphology. In MerceÁ GonzaÁlez (ed.), NELS 24, vol. 2, 333±379. GLSA, University of Massachusetts, Amherst. McCloskey, James. 1992. On the Scope of Verb Movement in Irish. Ms., University of California, Santa Cruz. Nespor, Marina, and Irene Vogel. 1986. Prosodic Phonology. Dordrecht: Foris. OrdoÂnÄez, Francisco. Forthcoming. Postverbal Asymmetries in Spanish. Natural Language and Linguistic Theory. OrdoÂnÄez, Francisco, and Esthela TrevinÄo. 1995. Los sujetos y objetos preverbales en espanÄol. Paper presented at the Fifth Colloquium on Generative Grammar, CorunÄa, Spain. Ouhalla, Jamal. 1991. Functional Categories and Parametrization. London: Routledge. Pinto, Manuela. 1997. Licensing and Interpretation of Inverted Subjects in Italian. Doctoral dissertation, Utrecht University. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html.

Emergence of Unmarked Word Order

203

Raposo, Eduardo. 1986. On the Null Object Construction in European Portuguese. In Osvaldo Jaeggli and C. Silva-CorvalaÂn (eds.), Studies in Romance Linguistics. Dordrecht: Foris. Raposo, Eduardo. 1996. Towards a Uni®cation of Topic Constructions. Ms., University of California, Santa Barbara. Raposo, Eduardo. 1997. De®nite/Zero Alternations in Portuguese (Towards a uni®cation of Topic Constructions). Paper presented at Linguistic Symposium on Romance Languages. Rizzi, Luigi. 1982. Issues in Italian Syntax. Dordrecht: Foris. Rizzi, Luigi. 1995. The Fine Structure of the Left Periphery. Ms., University of Geneva. Samek-Lodovici, Vieri. 1996. Constraints on Subjects: An Optimality-Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. SolaÁ, Jaume. 1992. Agreement and Subjects. Doctoral dissertation, University of Barcelona. Sproat, R. 1985. Welsh Syntax and VSO structure. Natural Language and Linguistic Theory 3: 173±216. Szabolcsi, Anna. 1981. The Semantics of Topic-Focus Articulation. In J. Groenendijk, T. Janssen, and M. Stokhof (eds.), Formal Methods in the Study of Language, 513±540. Amsterdam: Mathematical Centre. Tallerman, Maggie. 1996. Fronting Constructions in Welsh. In R. Borsley and I. Roberts (eds.), The Syntax of the Celtic Languages. Cambridge, England: Cambridge University Press. Travis, Lisa. 1984. Parameters and E¨ects of Word Order Variation. Doctoral dissertation, MIT, Cambridge, Mass. Zubizarreta, Maria Luisa. 1995. Word Order, Prosody, and Focus. Ms., University of Southern California, Los Angeles. Zwart, C. Jan Wouter. 1993. Dutch Syntax: A Minimalist Approach. Doctoral dissertation, University of Groningen.

Chapter 8 Optimal Clitic Positions and the Lexicon in Romance Clitic Systems

Jane Grimshaw

When pronominal clitics combine in the Romance languages, they occur in ®xed orders, and in di¨erent ®xed orders in di¨erent languages (and dialects), even when attention is limited, as here, to clitic placement in ®nite clauses.1 The research program of explicating the systematicity within these systems, begun within generative grammar in Perlmutter 1971, has been challenged by the complexity of the facts, with many researchers simply concluding that the possible orders must be stipulated in a ``template,'' or surface structure constraint. Monachesi 1995, Bonet 1995, and Miller and Sag 1997 are recent examples. The goal of the present chapter is to show that some fundamental properties of the clitic systems of French, Italian, and Spanish can be understood as a function of a set of ranked, violable morphosyntactic constraints of three core kinds: faithfulness constraints, markedness constraints, and alignment constraints (Prince and Smolensky 1993; McCarthy and Prince 1993). This proposal is developed under the general theoretical assumptions of Optimality Theory (OT; Prince and Smolensky 1993). Interaction among con¯icting constraints is at the heart of the solution I propose for these complex and highly variable systems. The relative order of clitics in Romance ®nite clauses follows from alignment constraints that relate morphosyntactic speci®cations for case and person to the right and left edge of the clause (and hence indirectly the clitic cluster). Alignment interacts with the lexical representations of the clitics (Bonet 1995; Grimshaw 1997a) in ways that predict the positions of the clitics. The alignment constraints also interact with faithfulness constraints, explaining the clitic substitution known as ``spurious se.'' Markedness constraints interact with the faithfulness constraints, and this allows us to derive the inventory of clitic speci®cations seen in the lexicons of the languages under studyÐthat is, to predict which of the universal possibilities for pronominal speci®cations will actually appear in grammatical outputs. The functional lexicon is in this view the output of a set of violable universal constraints in interaction. In other words, it is a function of the grammar of a language, and not an independent set of stipulations concerning the morphemes that can appear in well-formed sentences.

206

8.1

Jane Grimshaw

The Clitic Lexicon

The clitic lexicons of Italian, Spanish, and French are given in (1)±(3). The chart in (1) shows the nonre¯exive third-person forms, which mark distinctions in number, gender, and case. This is illustrated for Italian by table 8.1, which shows the distribution of third-person forms across the paradigm cells for case, person, and number. (1) The Romance lexicon: third-person nonre¯exive forms Italian French Spanish him/it lo le lo her/it la la la them (masculine) li les los them (feminine) le les las to him/it gli lui le to her/it le lui le to them Ð leur les Within ®rst- and second-person forms, however, there are no case or gender distinctions, and all clitics are used for both re¯exive and nonre¯exive forms. (2) The Romance lexicon: ®rstItalian (to) me(self ) mi (to) you(self ) ti (to) us(self ) ci (to) you(self ) vi

and second-person forms French Spanish me me te te nous nos vous os2

Finally, there is a clitic that occurs only as a third-person re¯exive. (3) The Romance lexicon: third-person ``re¯exive'' forms Italian French Spanish self si se se Table 8.1 illustrates the resulting distribution of morphemes across the paradigm in Italian.3 A single clitic form occupies all four slots for ®rst and second persons, while there are multiple forms for the third persons, re¯ecting distinctions in case and re¯exivity. In addressing the issue of how these lexical items are represented, we have two basic possibilities. One option is to posit multiple occurrences of all the morphemes that occur more than once in the table. Thus we posit one clitic, mi, which is ®rst person, accusative, nonre¯exive. Then we posit a second clitic, also mi by accident, which is ®rst person, dative, nonre¯exive, and so on. The second option is to posit only one lexical entry for each morpheme, and treat the identity of forms across certain cells as nonaccidental. In this case, we posit one morpheme mi, with no gender,

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

207

Table 8.1 The Italian clitic system acc dat acc-ref dat-ref

1sg

2sg

3sg

1pl

2pl

3pl

mi mi mi mi

ti ti ti ti

lo/la gli/le si si

ci ci ci ci

vi vi vi vi

li/le Ð si si

case, or re¯exivity speci®cation, rather than multiple clitics, each with a di¨erent speci®cation. While either of these options, or indeed a combination of the two, is descriptively adequate for selecting the correct clitic form in isolation, more fully predicting the properties of Romance clitics depends crucially on the exact nature of their lexical representation. In the course of this chapter, the lexical representations proposed in the present section will have consequences for constraints on clitic ordering, substitution of one clitic for another, and the markedness constraints that derive the clitic inventories themselves by determining which combinations of morphosyntactic features are possible in the grammar of these languages. It turns out to be critical, for example, that the clitic mi has a person speci®cation and a number speci®cation, but no speci®cation for re¯exivity, gender, or case. On the other hand, the clitic lo will have speci®cations for all ®ve properties. The outlined typeface, R for re¯exivity, P for person, N for number, G for gender, and C for case, indicates that the clitic concerned has no speci®cation for that property. Thus a clitic represented as R means that a clitic is neither re¯exive nor nonre¯exive, but occurs in both uses. The Italian lexicon, analyzed in this way, is shown in (4). (4) The Italian lexicon lo R 3 sg m acc la R 3 sg f acc li R 3 pl m acc le R 3 pl f acc gli R 3 sg m dat le R 3 sg f dat mi R 1 sg G C ti R 2 sg G C ci R 1 pl G C vi R 2 pl G C si ‡R P N G C

him/it her/it them (masculine) them (feminine) to him/it to her/it (to) me(self ) (to) you(self ) (to) us(self ) (to) you(self ) self

We can interpret the array in (4) as follows: all speci®cations are preserved in the least marked con®guration, namely, a third-person nonre¯exive clitic. Re¯exivity,

208

Jane Grimshaw

gender, and case are all lost in the marked personsЮrst and secondÐand all distinctions are lost in the (highly marked) re¯exive. This is the consequence of markedness constraints interacting with faithfulness, as we will see in section 8.5, where it forms the basis of the derivation of the inventory of clitics from constraint interaction. 8.2

Selecting the Best Clitic

Given this analysis of the clitic lexicons, what determines which clitic actually occurs? Why, for example, does an Italian speaker select the clitic mi for the ®rst-person accusative feminine form, rather than la, which is both accusative and feminine? The answer is given by a process of optimization, of the kind familiar from OT, which selects from the lexicon the clitic most faithful to the input. (Grimshaw 1997a presents a more detailed discussion, albeit with a slightly di¨erent lexical basis.) The key points here are the following. The input is a morphosyntactic speci®cation constructed from the universal feature inventory for pronounsÐfor example, [ R 3 pl f acc]. (See section 8.5 for further elaboration.) The candidates are the set of pronouns in the language, lexically analyzed as discussed above. The output is the clitic with the lexical representation that best analyzes the input. The selection of the best clitic is conducted by a set of faithfulness constraints that require a match between the input speci®cation and the output speci®cation. For each relevant morphosyntactic property admitted into universal grammar, there is a Faith constraint that is violated if the speci®cation in the input and the output di¨er (in any way). Whenever the input is third person and nonre¯exive, there is a candidate that parses the input perfectly, hence this candidate is always the optimum and is always chosen by the grammar. Consider, for example, tableau T8.1, which exempli®es the crucial comparisons. Since the speci®cations of (b) perfectly match those in the input, it satis®es all the faithfulness constraints, under any ranking. Every other candidate pronoun in the system violates at least one. For example, the candidate in (a) violates FaithRefl, FaithPers, FaithGen, and FaithCase, having no speci®cation for Tableau T8.1 Third-person nonre¯exive inputs Input: [ R 3 sg m acc]

FaithRefl

FaithPers

a. R 1 sg G C

*!

*!

f b.

R 3 sg m acc

c.

R 3 sg m dat

FaithNum

FaithGen

FaithCase

*!

*!

*!

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

209

re¯exivity, gender or case, and the wrong person. The dative counterpart (c) to the optimal candidate violates FaithCase and so is suboptimal even though it satis®es all the other constraints. The situation becomes more interesting as soon as we examine non-third-person inputs, since for these inputs there is never a candidate that o¨ers a perfect parse. The clitic si in candidate (a) violates all faithfulness constraints except FaithRefl. Since the candidate that violates FaithRefl but satis®es FaithPers and FaithNum is in fact the winner, FaithPers or FaithNum must dominate FaithRefl. (Subsequently we will see that parsing of re¯exive is ranked above parsing of number, so the crucial ranking will prove to be FaithPers over FaithRefl with FaithRefl over FaithNum, as shown in the tableaux.) Among the other candidates, the best third-person candidate has better gender and case speci®cations than the winner, but the wrong person, thus FaithPers must dominate FaithGen and FaithCase. In tableau T8.3 we examine the same three candidates as in tableau T8.2, but this time corresponding to a nonre¯exive input. The same candidate is still selected as optimal. (The tableau shows only the candidates relevant for ranking: in the full tableau candidate (a) harmonically bounds all other third-person outputs. Candidate (b) harmonically bounds all other ®rst- and second-person outputs.) Here we have Tableau T8.2 First- and second-person inputs Input: [‡R 2 pl m acc]

FaithPers

a. ‡R P N G C

*!

f b. R 2 pl G C c.

R 3 pl m acc

FaithRefl

FaithNum

FaithGen

FaithCase

*!

*

*

*

*

* *!

*

Tableau T8.3 First- and second-person nonre¯exive inputs Input: [ R 2 pl m acc]

FaithPers

FaithRefl

FaithNum

FaithGen

FaithCase

a. ‡R P N G C

*!

*

*

*

*

*

*

f b. R 2 pl G C c.

R 3 pl m acc

* *!

210

Jane Grimshaw

Tableau T8.4 Third-person re¯exives Input: [‡R 3 pl m acc]

FillRefl

FaithPers

f a. ‡R P N G C

*

b. R 1 pl G C

*

c.

R 3 pl m acc

ParseRefl

*!

FaithNum

FaithGen

FaithCase

*

*

*

*

*

*!

further evidence for the constraint rankings. Candidate (b) is the output, despite the fact that it violates FaithRefl, FaithCase, and FaithGen, all of which are satis®ed by candidate (c). The violation of FaithPers incurred by candidate (c) must be overriding these violations, as will be the case if FaithPers dominates FaithRefl, FaithGen, FaithCase. Finally we come to the case of third-person re¯exive inputs. The winning candidate in tableau T8.4 is faithful to the re¯exivity of the input and to nothing else. The other candidates are more successful with respect to number (candidate (b)) or to person, gender, case, and number (candidate (c)). Hence the rankings established so far will incorrectly select candidate (c), which is faithful to person, over candidate (a), which is faithful to re¯exivity. This problem is solved if we conclude that FaithRefl is really two constraints, one that penalizes absence of a re¯exivity speci®cation (Parse) and one that penalizes presence of the wrong re¯exivity speci®cation (Fill ). (This is in accordance with Grimshaw 1997a, where it is argued that the constraints at work are really the Max and Dep families of McCarthy and Prince 1995, 1999.) FillRefl will penalize the candidate in (c) with a R speci®cation and be satis®ed by si in candidate (a). Provided that FillRefl dominates FaithPers, FaithGen, FaithCase, and FaithNum, the correct output will be selected. If ParseRefl dominates FaithNum, candidate (a) is selected over (b). (Note that if FaithPers is also broken up into FillPers and ParsePers, the ranking of FillPers with respect to FaithNum could be the critical one.) The earlier ranking evidence, we can now see, showed that FaithPers dominates ParseRefl (rather than FaithRefl ), ensuring the choice of a ®rst- or second-person output, which has no re¯exive speci®cation over third-person alternatives that will match the input for re¯exivity. Thus, by ranking the faithfulness constraints, and assuming the lexical representations developed in section 8.1, we can predict which clitic surfaces as the output for each input. Later (section 8.5), we will see that the lexical representations themselves are derived by constraint rankings.

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

8.3

211

Romance Clitic Order: Interaction among Alignment Constraints

When clitics combine, as mentioned above, they do so in ®xed orders, di¨erent for di¨erent linguistic systems. Given the lexicon developed in sections 8.1 and 8.2, the relative order of clitics in combination can be determined by a set of alignment constraints on morphosyntactic speci®cations for person and case. For related work proposing that clitics are subject to alignment constraints, see Anderson 1996, forthcoming; Gerlach 1998; Legendre, 1996, forthcoming, chap. 9, this volume; Van der Leeuw 1995. It is important to note that these constraints govern the order of clitics when they combine with each other, not the well-formedness of the combinations in the ®rst place, which is a matter for markedness constraints (see section 8.4). Since only person and case speci®cations prove important for alignment, we can simplify the lexicon to that given in (5). (5) The Italian lexicon showing person and case only lo 3 acc him/it la 3 acc her/it li 3 acc them (masculine) le 3 acc them (feminine) gli 3 dat to him/it le 3 dat to her/it mi 1C (to) me(self ) ti 2C (to) you(self ) ci 1C (to) us(self ) vi 2C (to) you(self ) si PC self The order of clitic combinations has been taken to be essentially inexplicable, with the restrictions normally stated by means of a positive output constraint, or ``template,'' which stipulates the possible orders and combinations of clitics. Templates proposed in the literature for French, Italian, and Spanish are given in (6)± (8). (6) French (based on Perlmutter 1971:57) me 3-acc 3-dat y en te nous vous se

212

Jane Grimshaw

(7) Italian (Monachesi 1995:122)4 mi ci si lo si ne ti (Ref ) la (imp) gli li le le (acc) ci vi (8) Spanish (based on Perlmutter 1971:45) se 2 1 3 Although the French and Spanish templates as presented by Perlmutter re¯ect the existence of clear morphosyntactic patterns in clitic combinations, the template notation clearly renders the positions of the clitics essentially arbitrary, at least in the absence of a theory of what a template can look like. While the Spanish generalizations might lead one to suspect that person speci®cations are the key, it is clear that person speci®cations alone are not su½cient to determine the ordering of clitics in French or Italian. Third-person forms appear in the ®rst and fourth columns in Italian, for example, and the ®rst column contains ®rst-, second-, and third-person clitics. However, even the Italian clitic template is not random: all third-person accusative clitics are on the right, for instance. It certainly appears that the morphosyntactic properties of the clitics determine where they appear in the template. But how is this accomplished, why should it be so, and what predicts the exact outcome? 8.3.1 French The template in (9) shows the French clitic order temporarily simpli®ed to include only personal clitics. (9) A 1,2 se

B C 3-acc 3-dat

In terms of the lexical analysis given here, the three columns can be summarized as follows: clitics in column B have an accusative case speci®cation, clitics in column C have a dative case speci®cation, and clitics in column A have neither. The clitics in column A are thus the complement of those in B and C. This array and be derived from interacting constraints relating case to position. Speci®cally, we can posit positional constraints DatRt and AccRt, which are satis®ed when an element with the relevant case speci®cation is on the right-hand edge of the clitic cluster, violated otherwise. Such positional constraints are always satis®ed by any clitic in isolation; a third-person accusative by itself necessarily satis-

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

213

®es AccRt (and AccLft; see below), for example. Hence the constraints have no e¨ect except on clitic combinations. The alignment constraints DatRt and AccRt both require that a case-speci®ed clitic occur on the right; hence when only one clitic in a combination has a case speci®cation, that clitic will appear on the right-hand edge. However, when there are two, as there can be if one is dative and one is accusative, provided that the constraint targeting datives is the dominant one, the system will settle on the choice of the dative on the edge. Thus, clitics from column A will precede clitics from column B, as illustrated in tableau T8.5, and clitics from column B will precede clitics from column C, as illustrated in tableau T8.6. When any ®rst- or second-person clitic combines with an accusative, as in tableau T8.5, the constraint that mentions dative is irrelevant, since neither clitic is dative. Hence AccRt is the only relevant constraint and it positions the accusative on the right edge. Tableau T8.6 illustrates the situation when an accusative and a dative co-occur. The dative clitic must be on the right-hand edge to satisfy the dominant DatRt constraint. The position of the third-person accusative violates AccRt, but there is no better con®guration. Note the argument for the crucial ranking: if AccRt g DatRt, then (b) is optimal instead of (a). It is apparent that the lexical representation of the clitics is crucial in predicting clitic order; it determines how the clitics interact with the constraints. The clitic corresponding to a [1 dat] input in tableau T8.5, like all ®rst- and second-person clitics and se, has no case speci®cation, hence it vacuously satis®es both the constraints: it Tableau T8.5 French ®rst person ‡ accusative Input: [1 dat ‡ 3 acc]

DatRt

AccRt

f a. (1C) (3 acc) b. (3 acc) (1C)

*!

Tableau T8.6 French third personÐdative ‡ accusative Input: [3 acc ‡ 3 dat]

DatRt

f a. (3 acc) (3 dat) b. (3 dat) (3 acc)

AccRt *

*!

214

Jane Grimshaw

satis®es them regardless of where it is positioned. When it combines with any clitic with a case speci®cation, that clitic will be positioned by DatRt or AccRt, hence the clitic with no case will end up on the left. All clitics that have no case speci®cation go in the residual position, on the left, as a result of their satisfying the constraints vacuously. This illustrates the role of constraint interaction in picking out the complement of what is targeted by constraints. The clitic corresponding to the [1 dat] input is in the position of an element that is not dative and is not accusative. Se, the thirdperson re¯exive clitic, occurs in the same slot because it also lacks a case speci®cation. This supports the lexical analysis that posits one clitic with no case for each of the ®rst- and second-person forms, rather than two fully speci®ed clitics. If they were speci®ed for accusative or dative they would behave just like third persons, since they would not satisfy the case constraints vacuously. I assume, in accordance with the leading ideas of alignment (see in particular Samek-Lodovici 1998), that the two constraints at work in French are members of a universal pool that consists of all possible combinations of case and person speci®cations with ``Rt'' and ``Lft'' (we will see direct evidence from Italian that alignment constraints like these can target the left edge of the cluster). It follows that DatLft and AccLft are also constraints, and hence also in the grammar of French: they must be ranked below DatRt and AccRt respectively. This will guarantee that it is DatRt and AccRt that make the key decisions in the language. The analysis will thus require the crucial rankings in (10). (10) French ranking DatRt g DatLft, AccRt g AccLft DatRt g AccRt DatLft and AccLft are omitted from the tableaux above, and the same simpli®cation strategy is followed below. Positioning each constraint anywhere to the right of its opposite will leave the selection of the optimum unchanged. See section 8.4.3 for further analysis of ranking among such matched pairs of alignment constraints. Finally, clitics y (a locative) and en (a partitive element) have been omitted from the analysis so far. Since they have no person or case speci®cation, their positioning will not be a¨ected by alignment constraints on positioning these speci®cations. They can be properly positioned by the constraints PartitiveRight and LocativeRight, under a ranking in which PartRt g LocRt, and both constraints dominate all other constraints discussed here. To sum up, the sequencing information in the clitic ``template'' for French can be derived by ranking the constraint placing datives on the right over the constraint placing accusatives on the right. In contrast to standard modern French, Old French had systematic accusative-dative order, and some regional French dialects have dative-accusative (Morin 1979). Reranking the posited alignment constraints derives

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

215

these alternative systems. For example, if AccLft dominates the constraint that leftaligns datives, as well as AccRt, a sequence will occur in accusative-dative order. If AccRt dominates AccLft and the con¯icting dative constraint, a dative-accusative order will emerge. 8.3.2 Italian The template for Italian personal clitics appears in (11). (11) A B 1,2 si dative

C 3-acc

The three columns can be summarized as follows: Clitics in column A have a person speci®cation and are not accusative, clitics in column C have an accusative case speci®cation, and clitics in column B have neither. In describing the columns this way we highlight the fact that they do not all contain natural classes of clitics. In particular, column A in (11) can only be understood as the complement of column C, which is exactly the e¨ect that constraint interaction will derive. The primary generalization is that all accusatives go on the right. Clitics that have a person speci®cation but are not accusativeÐthat is, the remaining person-speci®ed cliticsÐgo on the left, and the still-remaining clitic si, which has no properties, goes in the middle. We can derive this pattern if Italian has AccRt as the dominant member of the alignment constraints for accusative, and PersLft as the dominant member of the alignment constraints for person. The accusative dominates the person constraint, and the Italian pattern emerges. (12) Italian ranking AccRt g AccLft, PersLft g PerRt AccRt g PersLft If an accusative clitic combines with any clitic with no accusative speci®cation, the accusative will occur at the right edge; see tableaux T8.7 and T8.8. The sequencing in tableau T8.8 motivates the ranking: the violation of PersLft in the ®rst candidate is outweighed by the violation of AccRt in the second. If a nonaccusative person clitic Tableau T8.7 ItalianÐaccusative with nonaccusative Input: [1 dat ‡ 3 acc]

AccRt

f a. (1C) (3 acc) b. (3 acc) (1C)

PersLft *

*!

*

216

Jane Grimshaw

Tableau T8.8 ItalianÐaccusative with nonaccusative Input: [‡R 3 ‡ 3 acc]

AccRt

f a. (PC) (3 acc) b. (3 acc) (PC)

PersLft *

*!

Tableau T8.9 ItalianÐnonaccusative with clitic with no case, person Input: [‡R 3 ‡ 1 acc] a. (PC) (1C)

AccRt

PersLft *!

f b. (1C) (PC) combines with a clitic with neither case nor person, the latter will follow the former (see tableau T8.9). In sum, clitics from column A will precede si (in column B) and clitics from both A and B will precede those from column C: this system will guarantee that any accusative clitic appears on the right, and that any clitic with a person speci®cation but no accusative speci®cation will appear on the left. The same ordering of clitics will hold regardless of which accusative clitic occurs, because they all have the same speci®cation for case and person. Similarly, the same result will hold regardless of which of the ®rst- or second-person clitics occurs, since nothing relevant distinguishes them. The third-person dative clitics gli and le pattern with the ®rst- and second-person forms because they all satisfy AccRt vacuously. Hence PersLeft positions them at the left edge, just like mi. Because the alignment constraints relate morphosyntactic speci®cation to position, the lexical properties of the clitics are crucial in sequencing them. Positional e¨ects in turn provide evidence about lexical representation. For example, the same morpheme, le, is used as the feminine dative singular and the feminine accusative plural. It would be incorrect, however, to treat this as a case of underspeci®cation, assigning le no case. This would predict that it would always occur on the left edge of the cluster, even when used as a direct object. (The markedness constraint system to be presented in section 8.5 correctly analyzes le as fully speci®ed.) The full template in (12) also contains positions for a partitive, an impersonal subject, and a locative. Characterizing ci as locative, ne as partitive, and assuming for

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

217

simplicity that the impersonal subject ci has no case or person speci®cation, the full set of orders is derived from the constraint hierarchy in (13), which includes the constraints on personal forms as a subpart. (13) PartRt g ImpRt g AccRt g PersLft g LocLft 8.3.3 Spanish The orders found among clitics in Spanish are given in (14), with combinations of ®rst- and second-person forms temporarily omitted from consideration. (14) A se

B 1/2

C 3

The three columns can be summarized as follows: Clitics in column B have person but no case speci®cation, those in column C have both, and those in column A have neither. Once again, then, we see the e¨ects of positional constraints interacting with the lexical properties of clitics. The Spanish pattern will be the result of right-alignment constraints mentioning case and person, regardless of the ranking between them.5 (15) CaseRt g CaseLft PersRt g PersLft Tableau T8.10 shows the sequencing of a clitic from column A and one from column B: candidate (b) wins under either ranking of CaseRt and PersRt. Tableau T8.11 shows the sequencing of a clitic from column B and one from column C: candidate (a) wins under either ranking. Finally, tableau T8.12 establishes the sequencing of a clitic from column A and one from column C. The case constraint forces a case-marked clitic to the right edge. Of the remaining clitics, one with person goes as far to the right as it can. The remaining clitic then occupies the remaining position. Thus the grammar does not have to include characterizations such as ``has person but not case'': that the facts can be described this way is the result of interacting simple constraints referring only to individual grammatical properties.6 The complex e¨ects are due to interaction. Tableau T8.10 SpanishÐse with ®rst or second person Input: [3 re¯ ‡ 1 dat] a. (1C) (PC) f b. (PC) (1C)

CaseRt

PersRt *!

218

Jane Grimshaw

Tableau T8.11 SpanishЮrst or second person with third person Input: [1 dat ‡ 3 acc]

CaseRt

f a. (1C) (3 acc) b. (3 acc) (1C)

PersRt *

*!

*

Tableau T8.12 SpanishÐse with third person Input: [3 re¯ ‡ 1 dat]

CaseRt

PersRt

f a. (PC) (1C) b. (1C) (PC)

*!

There is no possible con¯ict between CaseRt and PersRt, since the clitics that are a¨ected by CaseRt are a subset of those a¨ected by PersRt, and both constraints position the clitics on the same edge. A clitic that had a case speci®cation but no person speci®cation (nonexistent under the present analysis), combined with a clitic with a person speci®cation but no case speci®cation (which does exist under the present analysis), would allow us to detect the ranking: with the person constraint dominant the second clitic would be on the speci®ed edge; with the case constraint dominant the ®rst clitic would be on the speci®ed edge. Thus, this is an instance where there happens to be no direct evidence for the relative ranking of two constraints, and not an instance where two constraints are crucially unranked. To complete the picture, we need to examine the sequencing of ®rst- and secondperson clitics, which can co-occur in Spanish, although not in French and Italian (see section 8.4.2). Second-person clitics precede ®rst persons in standard Spanish (Perlmutter 1971). This will follow if a constraint aligning ®rst-person speci®cations with the right edge (1Rt) dominates the constraints 1Lft and 2Rt, whose existence follows from the general theory of alignment constraints (see section 8.4.3).7 Since the order in which ®rst- and second-person clitics occur depends only on the ranking of 1Rt and 2Rt, we expect dialect variation in this respect. Tableau T8.13 shows that a (perhaps hypothetical) sequence with se (the clitic with no person or case) in initial position, followed by a second person, followed by a ®rst person will be optimal under this ranking, the revealing comparisons being between the winner and candidate (a) in which the ®rst person precedes the second and between the winning candidate and any candidate (such as (c)) in which a ®rst- and/or second-person clitic is to the left of a clitic with no person marking.

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

219

Tableau T8.13 Spanish Input: [3 re¯exive ‡ 2 acc ‡ 1 dat]

CaseRt

PersRt

1 Rt

2 Rt/Lft

a. (PC) (1C) (2C)

*

*!

**

f b. (PC) (2C) (1C)

*

c. (1C) (2C) (PC)

**!*

** **

**

Tableau T8.14 SpanishÐsecond or ®rst person, gradience of alignment Input: [2 dat ‡ 1 dat ‡ 3 acc]

CaseRt

PersRt

1 Rt

2 Rt/Lft

*!*

***

*

**

f b. (2C) (1C) (3 acc)

***

*

**

c. (1C) (2C) (3 acc)

***

**!

**

a. (3 acc) (1C) (2C)

When ®rst- and second-person clitics combine with a case-marked clitic, as in tableau T8.14 (Perlmutter's (96), p. 51), the second person will again precede the ®rst, and the case-marked clitic will follow, provided that CaseRt g 1Rt. Thus we have shown that the second person will always precede the ®rst in any combination. All candidates in which the case-marked clitic is not at the right edge are eliminated, since there are competitors (such as (b) and (c)) that do satisfy CaseRt. Of the two candidates that satisfy CaseRt, the one that best satis®es 1Rt is optimal. The fact that candidate (c) is ungrammatical and (b) grammatical shows that 1Rt, and the clitic alignment constraints in general, are gradient constraints; see section 8.4.3. Other variation concerns the relative order of se: in Murcian (Heap 1996:231) se follows rather than precedes the ®rst and second persons, while still preceding third persons. This will follow from an analysis in which CaseRt g CaseLft, as in standard Spanish, but PersLft g PersRt, with CaseRt then dominating PersLft.8 Tableau T8.15 illustrates the e¨ect of this in a compact form, for a hypothetical input including a re¯exive, a ®rst person, and an accusative form. It shows only candidates that satisfy CaseRt, by having the accusative on the right edge. Candidate (a) is optimal under the present ranking, while (b) was optimal under the ranking of standard Spanish, where PersRt is the dominant-person alignment constraint.

220

Jane Grimshaw

Tableau T8.15 SpanishÐvariation in the position of se Input: [3 dat ‡ 1 dat ‡ 3 acc] f a. (1C) (PC) (3 acc) b. (PC) (1C) (3 acc)

CaseRt

PersLft ** ***!

8.3.4 Alignment Interacts with Faithfulness: A Clitic Substitution in Spanish Other things being equal, the fewer morphosyntactic properties a clitic has, the fewer alignment constraints it can violate. If S-1 is the set of speci®cations of clitic 1 and S-2 is the set of speci®cations of clitic 2, and S-1 is a subset of S-2, then clitic 1 can satisfy positional constraints violated by clitic 2, if S-1 does not contain the speci®cation targeted by the positional constraint(s). In this way, positional constraints themselves can force the substitution of one clitic for another, otherwise expected, clitic form. The combination of two third-person forms must violate both the dominant-case constraint and the dominant-person constraint. Since each clitic has a speci®cation for both case and person, and only one clitic can be in the position demanded by the highest ranking of the case constraint pair and/or the highest ranking of the person constraint pair, such a combination must violate both constraints. By the same token, a combination of one third-person form and one second- or ®rst-person form will always violate the person constraint, since again only one of the clitics can be on the edge speci®ed for person. The same is true for the combination of one second- or ®rst-person clitic with another second- or ®rst-person clitic. Spanish o¨ers a wellknown example of a system that does not tolerate combinations of third-person clitics. The data in (16) are taken from Bonet 1995. (16) a. E1 premio, lo dieron a Pedro ayer. the prize 3rd-acc gave(3rd-pl) to Pedro yesterday b. A Pedro, le dieron el premio ayer. to Pedro 3rd-dat gave(3rd-pl) the prize yesterday c. A Pedro, el premio, se lo dieron ayer. (*le lo *lo le) to Pedro the prize se 3rd-acc gave(3rd-pl) yesterday `They gave the prize to Pedro yesterday.' Where, based on the forms that occur in isolation, we expect to ®nd both a thirdperson dative and a third-person accusative, what instead surfaces is a third-person accusative accompanied by se. Why doesn't se occur in place of a third-person dative elsewhere? The answer is that the faithfulness constraints will prefer the faithful clitic, which parses person, number, and case. However, as we saw above, with no case or person speci®cation se can improve the performance of a combination on the posi-

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

221

Tableau T8.16 SpanishÐspurious se Input: [ R 3 sg dat ‡ R 3 sg acc] a. ( R 3 sg dat) ‡ ( R 3 sg acc)

CaseRt

PersRt

*!

*!

f b. (‡R P N C) ‡ ( R 3 sg acc) c. (R 1 sg C) ‡ ( R 3 sg acc)

*!

FaithRefl

FaithPers

FaithNum

FaithCase

*

*

*

*

*

**!

*

tional constraints, so se will ``replace'' the expected second-person clitic if the relevant faithfulness constraints are dominated by the relevant positional constraint.9 (See Bonet's work and Grimshaw 1997a for analysis of agreement with spurious se.) Tableau T8.16 shows that le ‡ lo (candidate (a)) violates PersRt and CaseRt (one of them fatally), while se has neither person nor case so it satis®es both vacuously. Thus the candidate that contains se (candidate (b)) satis®es both positional constraints. It violates the faithfulness constraints for re¯exivity, person, number, and case, but provided these are dominated by at least one of CaseRt and PersRt, the right candidate will be selected.10 (17) Either: CaseRt g FaithR, FaithPers, FaithNum, FaithCase Or: PersRt g FaithR, FaithPers, FaithNum, FaithCase Why replace le with se, and not, for example, with a ®rst- or second-person clitic (candidate (c)), which would be faithful to the number in the input, if not to the other speci®cations? Like candidate (b), this option would satisfy CaseRt, since the ®rst clitic in the sequence would have no case. This candidate, however, incurs a violation (one fatal) on both PersRt and FaithPers since it has two clitics bearing person speci®cations, and hence violates PersRt, as well as both the Parse and the Fill form of faithfulness, indicated by two asterisks in the FaithPers column of tableau T8.16. We conclude that at least one of FaithPers and PersRt must dominate FaithNum, so that the violation of FaithNum in candidate (b) is preferred to the extra violations of PersRt and FaithPers in candidate (c). Tableau T8.16 raises an additional question: Why is lo retained instead of le? That is, why is the dative clitic replaced by se and not the accusative clitic? This is explained if we posit a universal markedness hierarchy, in which datives are more marked than accusatives: MarkDat g MarkAcc. (For further discussion of mark-

222

Jane Grimshaw

edness in this system, see sections 8.4 and 8.5; Aissen, chap. 3, this volume; Woolford, chap. 16, this volume.) The two candidates (se le and se lo) are otherwise identical, and se le is eliminated by MarkDat. To complete the picture, one more generalization needs to be addressed. If Spanish prefers se to le in combination with a third-person clitic, why doesn't it also prefer se to the ®rst- and second-person datives? That is, where we expect le ‡ lo, we ®nd se ‡ lo, but where we expect me ‡ lo, that is exactly what we get. Yet me ‡ lo contains a PersRt violation, just as le ‡ lo does. Ranking can account for this observation. While me ‡ lo does violate PersRt it does not violate CaseRt, since the ®rst- and second-person clitics have no case. So if we adopt a ranking in which CaseRt alone dominates FaithRefl, FaithPers, FaithNum, and FaithCase, while PersRt is dominated by the faithfulness constraints, we will derive the result that faithful parsing will be overridden by the case constraint but not by the person constraint. Thus we have evidence for the ®rst of the two ranking possibilities listed in (17). (A minor variation of this solution posits the ranking: Faith1,2 g PersRt, so that faithful parsing of ®rst- and second-person speci®cations in the input takes precedence over elimination of the positional violation for person.)

Figure 8.1 Summary of constraint rankings

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

223

Since there is a crucial constraint ranking involved in deriving spurious se in Spanish, namely, between the faithfulness constraints and the positional constraint(s), it is not a surprise that there are grammars, such as those of French and Italian, that do not have this clitic substitution but simply retain the combination of two thirdperson forms. Of course, every constraint is present in the grammar of each language. Those that play no active role in positioning clitics are ranked below the constraints that are critical: below DatRt and AccRt for French, below AccRt and PersLft for Italian, and below CaseRt PersRt and 1Rt for Spanish. Figure 8.1 summarizes the critical constraint rankings for the clitic arrays analyzed so far. 8.4

Romance Cliticization and Alignment

The positional constraints governing clitics can be understood as gradient alignment constraints, relating morphosyntactic speci®cations to the edges of IP. In interaction with markedness constraints, the alignment constraints limit possible clitic combinations. 8.4.1 The Alignment Constraints The positional constraints are clearly a½liated with the family of alignment constraints (McCarthy and Prince 1993), which require that the edge of one constituent occur at the edge of another. (See the references cited in section 8.3 for related proposals involving the role of alignment constraints in cliticization.) In this view, there are no constraints directly imposing order among clitics. Rather the clitics are competing for positions at the edge of the cluster, just as, for instance, heads and Speci®ers compete for edges of syntactic projections, X 0 and XP, in Grimshaw 1997b:406±409. Much recent work, such as Kayne 1991 and Haverkort 1993, takes the position that Romance clitics left-adjoin to a functional head; see Uriagereka 1995 and Terzi 1999 for more detailed analysis. (For concreteness, I will assume the structure given in ®gure 8.2 for all Romance cliticization in ®nite clauses.11) The notation CLL=R designates a clitic with a speci®cation subject to left or right alignment. A left-alignment constraint requiring coincidence of the left edge of the clitic with the left edge of I 0 would be satis®ed by CLL in ®gure 8.2. Alternatively, and perhaps more interestingly, the alignment may be between the clitic and the left edge of IP, violated in ®gure 8.2 because of the higher-ranking SpecLft constraint (Grimshaw 1997b).12 In the notation of McCarthy and Prince, the constraints are stated as in (18). (18) ALIGN (Speci®er, Left, XP, Left) g ALIGN (case/person, Left, IP, Left) As for rightward alignment, there is again more than one possibility. The alignment could be to the right edge of I, or I 0 , or IP. All these constraints are violated by CLR

224

Jane Grimshaw

Figure 8.2 Structure of ®nite clauses with clitics

in ®gure 8.2. In the interest of maintaining symmetry between the left- and rightalignment constraints, let us posit the right edge of IP as the target for alignment in this case. Why then is alignment violated (by the intervention of I and VP between the clitic and the right edge of IP)? Switching the order of CLR and its sister I in ®gure 8.2 would improve alignment for CLR , so some higher-ranked constraint must prevent this. A candidate for this constraint is AdjunctLft (requiring adjuncts to occur at left edges), which will govern con®gurations where order is not determined by constraints on heads or speci®ers. If adjunct alignment takes priority over the case/person right-alignment constraint, as in (19), clitics will precede what they adjoin to. (19) ALIGN (Adjunct, Left, host, Left) g ALIGN (case/person, Right, IP, Right) Reversing the order of I and its complement VP will also improve right alignment of the clitic to IP: this is prevented by the constraint HdLft (Grimshaw 1997b), which enforces I-VP order rather than VP-I order under I 0 . In accordance with McCarthy and Prince's schema, the interpretation of a clitic alignment constraint such as AccRt is as follows. (20) E accusative-marked X 0 b IP such that the right edge of X 0 and the right edge of IP coincide That these constraints and rankings give the right results is con®rmed by tableau T8.17, based on the analysis of Italian and consistent with the rankings just established, where (in the optimal candidate) the ®rst clitic in the sequence is in the position of CLL in ®gure 8.1, and the second clitic is in the position of CLR . Comparison between candidates (a) and (b) shows that reversing the order of the I and the accusative clitic is ruled out because it violates AdjRt, although it satis®es AccRt. Similarly, reversing the order of the speci®er and the person-marked clitic in

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

225

Tableau T8.17 Alignment of clitics Input: [1 dat ‡ 3 acc]

AdjRt

f a. Spec (1C) (3 acc) I b. Spec (1C) I (3 acc) c. (1C) Spec (3 acc) I

AccRt

SpecLft

**

PersLft ***

*!

**** **

*!

**

(c) violates SpecLft and is thus not possible, even though it improves evaluation with respect to PersLft. One last point of clari®cation. Both the clitics in ®gure 8.2 are separated from the right edge of IP by an I (I 2 for CLL and I 3 for CLR ) and a VP. To be sure that CLR better satis®es right alignment when it is to the right of CLL than in the opposite order, the calculation of alignment violations must be based on the number of terminal elements that intervene between the aligning element and the speci®ed edge. CLR is separated from the right edge of IP by I 3 and VP, if it is the rightmost clitic, but by I 3 , VP, and the other clitic, if another clitic follows it. In this view, the positioning of Romance clitics is accomplished by alignment constraints relating morphosyntactic speci®cations to the right and left edges of syntactic constituentsÐspeci®cally, by hypothesis, IP. This raises two fundamental typological issues. First, if all the morphosyntactic constraints dominate the constraints on speci®ers and heads, the clitics should occur on the edges and the other constituents in the interior of a clause: the sequence CLL ±Spec±I 0 ±CLR satis®es clitic alignment constraints requiring one clitic to be on the left and one on the right of IP. Second, if some, but not all, of the morphosyntactic constraints dominate the constraints on structure, some clitics could precede, and some follow, either the speci®er or the I 0 , a point made by Nicole Nelson (personal communication). Neither of these two scenarios appears likely: both would be eliminated by a general principle requiring that structural alignment constraints (those that mention only items of syntactic theory, such as head, speci®er, X 0 ) must always dominate morphosyntactic alignment constraints (those that mention speci®cations such as person or case). Viewed from a perspective in which clitics are ordered by violable alignment constraints, the sequencing of clitics is not very di¨erent from other word order phenomena; the only characteristic that makes clitics di¨erent from, say, heads or speci®ers is that they are positioned by virtue of their morphosyntactic properties. 8.4.2 Alignment, Markedness, and Well-Formed Clitic Combinations The analysis developed up to this point concerns the order of clitic combinations: nothing predicts which combinations are possible in the ®rst place. It appears, in fact,

226

Jane Grimshaw

that regardless of whether templates or alignment constraints are at work, the wellformedness of a combination on the one hand, and the order in which its elements must occur on the other, are distinct problems. This point is partially disguised in templates, which often seem to contain both kinds of informationÐwhat can occur with what and in what order. The illusory nature of this is clear on inspection, however. For example, the template for French, in (6), properly sequences ®rst and second persons before, and third-person datives after, the third-person accusatives. This leads us to expect that ®rst- and second-person forms should precede third-person datives. Perlmutter (1971:63±65) argues that this combination is impossible, and the only way a third-person indirect object can be realized in combination with a ®rst- or second-person clitic is by a strong form, which has a noncontrastive interpretation in these circumstances. (Simpson and Withgott 1986:160±161 cite examples in which ®rst- or second-person ethical datives co-occur with third-person datives.) Thus, while it may be true that templates can represent the impossibility of some combinations by placing the items that do not co-occur in a single column, they certainly cannot represent all restrictions this way. Once an analysis is developed, then, to eliminate the impossible combinations from di¨erent columns of a template, it is necessary to ask whether it naturally extends to impossible combinations of clitics that can be grouped into a single column. Indeed, this appears to be the case. Alignment constraints alone will never predict that combinations are impossible: just that the best aligned version of a combination will be grammatical. Alignment constraints in combination with others, however, can predict the impossibility of combinations. When alignment constraints interact with faithfulness, as seen in section 8.3.4, a clitic combination can be ruled out completely, with a less faithful clitic replacing one of the expected elements, in order to improve alignment. Other e¨ects, I argue here, concern markedness and are closely related to the markedness factors that underlie the lexicon itself (see section 8.5). If ®rst and second persons, dative, and re¯exive are more marked than, respectively, third person, accusative, and nonre¯exive (see Aissen, chap. 3, this volume; Artstein 1998; Woolford, chap. 16, this volume, for recent OT work on these issues), it turns out that the impossible combinations of clitics are those that involve the marked values for case, person, and/or re¯exivity. Space limitations prevent a full analysis here: a sketch of the analysis of one prohibited combination follows. Italian (Monachesi 1995) and Spanish (Bonet 1994), like French, show prohibitions against combining ®rst and second persons with dative clitics. This is illustrated for Spanish in (21), with data from Bonet 1994. (Note that the Spanish template proposed by Perlmutter cannot account for this generalization.) (21) a. Me recomendaron a eÂl porque era el maÂs in¯uyente. b. *Me le recomendaron porque era el maÂs in¯uyente. `They recommended me to him/her because she/he was the most in¯uential.'

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

227

Tableau T8.18 Realization of input by a strong pronoun Input: [1 ‡ weak dat] a. 1 ‡ dat f b. 1 ‡ strong

Mark1&Dat

Weak

*! *

It appears that the optimal forms in such circumstances are alternative candidates with full pronouns, such as (21a). When an input is speci®ed as weak, by virtue of its discourse status (see Grimshaw and Samek-Lodovici 1998) the best output normally will be a clitic form; I refer to the constraint that mandates this as Weak (Bonet's 1994 ``Avoid Pronoun Strength'' embodies the same idea). However, a clitic combination may violate a markedness constraint that con¯icts with, and dominates, the constraint mandating a weak form. In this circumstance a strong form (a pronoun) will be optimal instead, even as the realization of a weak element in the input. In support of this view, Bonet (1994:43) observes that in Spanish, a strong pronoun is possible without a contrastive focus interpretation and without a doubled clitic, only when the use of the strong pronoun avoids a violation of what she calls the ``PersonCase constraint,'' which bans datives with ®rst- and second-person clitics. The core of the analysis of (21b) is that the ill-formed clitic combination violates a markedness constraint against combining the marked features ®rst/second person and dative (thanks to Ron Artstein, personal communication, for key steps in this analysis). This markedness constraint dominates Weak, and forces the nonclitic realization of one of the arguments, as in (21a). In isolation each argument is realized as a clitic, since the combination constraint is not violated, and combinations involving less marked elements are also realized with clitics: ®rst or second person with accusative, for example. In this analysis, then, the combination of a dative and a ®rst- or second-person form is more marked than the combination of a dative and a third-person form or a ®rst- or second-person form and an accusative. To predict the impossibility of the relevant combinations, the constraint against combining ®rst and second with dative is ranked above Weak; the relevant constraints violated in the other clusters are ranked below Weak. 8.4.3 The Gradience of Alignment Constraints Most of the clitic sequences discussed so far have contained only two clitics. When this is the case, it is impossible to distinguish between two interpretations of the violation of the constraints: that the violations may be absoluteÐa clitic not at the edge violates the constraint exactly once no matter how many elements separate it from

228

Jane Grimshaw

the edge, or they may be gradientÐa clitic not at the edge violates the constraint once for each element that separates it from the edge. In a sequence of two, a misaligned element will violate the relevant alignment constraint once under either interpretation; however, in larger sequences the distinction is crucial. The decision between gradient and absolute violation requires sequences of at least three clitics, returning us to the question addressed in section 8.4.2, of which potential combinations of clitics can in fact occur. Combinations involving highly marked clitics are generally expected to be impossible, and combinations of three clitics are of course more likely to contain marked clitics than combinations of two. It is not clear that the data available from the various languages is strictly comparableÐwhether the same combinations have been examined, the same standards of well-formedness applied, and the same range of interpretations reviewed (see, for example, footnote 6 in Wanner 1977). Perlmutter cites grammatical three-clitic combinations (pp. 50±51, although see note 28), such as the one analyzed in tableau T8.14. As noted previously, the ungrammaticality of candidate (c) and the grammaticality of candidate (b) show that the constraints are gradiently violable. They must be gradient constraints since the violation of 1Rt in (c) is worse than that in (b) only if the separation of the ®rstperson clitic from the right edge by two elements is worse than the separation of the ®rst-person clitic from the right edge by only one element. Under an absolute interpretation of the violations, both candidates would violate 1Rt exactly once, by virtue of containing one element with a ®rst-person speci®cation and not on the right edge. This case illustrates what I take to be the general situation: insofar as combinations of three clitics are allowed in the ®rst place (whether personal clitics, partitives, or locatives), their positioning shows gradience. The general pattern predicted by gradience is for a clitic to appear as close as possible to the targeted edge, when it cannot be the absolute edgemost item. The gradience of the alignment constraints establishes the patterns of constraint violation that they are responsible for. As an example, we can look at the constraints PersLft, PersRt. In (22), violations of PersLft are shown as subscripts to the left of P, where P is a person-marked clitic; violations of PersRt are shown as subscripts to the right of P. (X and Z have no person speci®cation.) What (22) shows is that for an input of four clitics, one of which has a person speci®cation and is thus targeted by the constraints, the total number of violations of PersLft and PersRt added together is constant: there are always a total of three violations. (22) 0 P3 X X X

X 1 P2 Y Y

Y Y 2 P1 Z

Z Z Z 3 P0

ˆ3 ˆ3 ˆ3 ˆ3

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

229

The total number of violations is constant because of a particular property of matched alignment constraints with opposite edge speci®cations: for each clitic with the targeted property, elimination of a violation of one always adds a violation of the other. For example, as a clitic with a targeted speci®cation moves rightward (as P does in (22)), the number of XRt violations it incurs decreases and the number of XLft violations it incurs increases, at exactly the same rate. The patterns of violation in (23) make the same point for a sequence of four clitics, three of them person marked. Here the violations for each person-marked clitic always total 3, and the total number of violations in the sequence is always 9. (23) 0 P3 0 P3 0 P3 0 X0

1 P2

2 P1

0 X0

0 X0

2 P1

3 P0

1 P2 1 P2

0 X0 2 P1

3 P0

3 P0

ˆ9 ˆ9 ˆ9 ˆ9

Where n is the total number of clitics in a sequence, there are n 1 violations of a matched opposite alignment pair for each clitic with the targeted property. For example, each clitic speci®ed for person in (23) incurs 3 violations of the matched opposite alignment pair [PersLft, PersRt], because the total number of clitics in this sequence is 4. For the entire sequence, the total number of violations of the matched opposite alignment pair is t…n 1†, where t is the total number of clitics with the targeted property. In (23), then, the total number of violations per constraint pair is 9Ðthat is, 3  …4 1†. When the choice is between candidates with the same items, which do not di¨er in structure in any way, the lower-ranked constraint of each pair can never make a decisionÐthat is, will be invisible in the grammar. Any candidates still competing by the time they are assessed by the lower ranked of the pair must by de®nition have tied on all higher-ranked constraints, including the higher-ranked member of the pair. But if they tied on the higher ranked of the pair they will tie on the lower ranked of the pair also. Note that the invisibility of the lower-ranked constraint of each pair is not arbitrary. It maintains the absolutely strongest version of the OT universality hypothesis: all constraints are in the grammar of every language. What distinguishes gradient alignment constraints from other types of constraints is that for each pair of constraints, a candidate that incurs n fewer violations on one member of the pair will always incur n more violations on the other member. The only ranking that can ever be crucial, given this argument, is the ranking of the higher member of a matched opposite alignment pair, since the lower member can never be decisive in determining optimality, and hence cannot be crucially ranked with respect to any other constraint.

230

8.5

Jane Grimshaw

Markedness and the Derivation of the Clitic Inventory

At several places in the analysis, the nature of the clitic lexicon plays a critical role. The inventory of cliticsÐthat is, the set of lexical representations that exist in these systemsÐcan itself be derived from ranked markedness and faithfulness constraints; that is, such a ranking predicts the set of morphosyntactic combinations present in, or absent from, these systems. To mention a couple of prominent examples: there are no ®rst- or second-person re¯exive forms, there are no ®rst- or second-person case-marked forms, and there is no re¯exive clitic with gender, case, or number speci®cations. Of course, the grammar (ie., ranked constraints) cannot predict the phonological form of the cliticsÐthat is, the fact that a clitic is la rather than pi. The grammar can, however, predict that a particular combination of morphological speci®cations will exist in the lexicon. I assume that in the unmarked case, each distinct cluster of morphosyntactic speci®cations is associated with a distinct phonological form.13 The derivation of the clitic inventory is very much in the spirit of the derivation of phonemic inventories as proposed in Prince and Smolensky 1993, and pursues the goal suggested in Grimshaw 1997b of deriving the inventory of functional heads in individual grammars from ranked grammatical constraints. The basic idea is that universal grammar speci®es a set of morphosyntactic features that can be encoded in the pronominal systems of natural languages. Clusters of these features constitute the inputs to the constraints that determine the pronominal system. (The analysis presented here assumes that the input always includes a speci®cation for every feature: modi®cation is required to encompass the more complex situation where inputs can be partially speci®ed.) The ranking of markedness constraints (MarkX), which penalize a pronoun for encoding a speci®cation, and of faithfulness constraints (FaithX), which reward a pronoun for encoding input speci®cations, determines which combinations of speci®cations will occur in output forms in a given language. These constraint rankings jointly determine the inventory. We thus posit Faith and Mark constraints for all speci®cations allowed by universal grammar. (I assume that there are Faith and Mark constraints for both the speci®cation and its values. For instance, one constraint in each family targets any number speci®cation, one targets singular, and one targets plural. I will discuss the ranking of the general constraint unless it is the more particular constraint that is important.) For speci®cations that play no role in distinguishing personal clitics from one another (but that can be active in pronominal systems in general) such as the G human distinction, we conclude that the markedness constraint MarkHuman dominates the faithfulness constraint FaithHuman, and thus that all inputs with human speci®cations map to outputs that lack such speci®cations. (This assumes that no other constraint(s) favor preservation of the G human speci®cation.) In cases

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

231

such as this, where a distinction is irrelevant throughout the clitic system, I omit it from lexical representations and from further consideration here altogether. In contrast, for speci®cations that are contrastive in the clitic systems because they are preserved in some output forms, we know that the Faith constraint must dominate the Mark constraint. The Faith constraints for re¯exivity, person, number, gender, and case must all dominate the corresponding Mark constraints. (24) FaithRefl g MarkRefl, FaithPers g MarkPers, FaithNum g MarkNum, FaithCase g MarkCase, FaithGen g MarkGen The interesting challenge posed by the Romance clitic systems is that they are not uniformÐthat is, one cannot say for each speci®cation that it is always, or never, preserved in the output. This can be seen from reexamination of (4) above. Gender, for example, is preserved in the third persons but not elsewhere, and the same is true for case and re¯exivity. This shows that the interaction of simple markedness and faithfulness constraints alone cannot determine the clitic inventory. Such interactions can only generate systems in which a given feature (or value) is always preserved in the output (Faith g Mark ) or systems in which a given feature (or value) is never preserved in the output (Mark g Faith). Characterizing a mixed system will require constraints that assess the markedness not just of single features or values but of combinations of features or values (see section 8.4.2.). These constraints will be referred to as combination constraints.14 Every case where the outline notation has been used in a lexical representation is a case of this kind: the outlined feature, while active in the Romance clitic systems in general, has not been preserved in a particular output. This occurs when a combination markedness constraint con¯icts with a faithfulness constraint, and the combination constraint is dominant. There are three possible types of clitics in the Romance personal clitic lexicon. First, there is a re¯exive clitic, in which re¯exivity does not co-occur with any other speci®cation. Second, there are ®rst- and second-person clitics, which are speci®ed for number but not for re¯exivity, gender, or case. Third, there are nonre¯exive third persons, in which all features are preserved. I illustrate this case ®rst, in tableau T8.19. The optimal output here violates the combination markedness constraints for all speci®cations combined with third persons. The combination constraint can only be satis®ed by eliminating the person speci®cation as in (c), or eliminating all the other speci®cations, as in candidate (b). The ranking of the faithfulness constraints rules out either of these options: no matter which speci®cations is omitted (leading to improvement on the combination constraint), a faithfulness violation is incurred, so that such a candidate cannot be optimal. For third-person inputs, then, all speci®cations are parsed.15 We conclude that the combination markedness constraints that target

232

Jane Grimshaw

Tableau T8.19 All speci®cations parsed in third-person nonre¯exives Input: [ R 3 sg m acc] f a.

FaithRefl/Pers/ Num/Gen/Case

R 3 sg m acc

****

b. R 3 N G C c.

Mark3&Refl/3&Num/ 3&Gen/3&Case

*!***

R P sg m acc

*!

combinations involving third person are all dominated by the con¯icting faithfulness constraints. (25) FaithRefl/Pers/Num/Gen/Case g Mark3 & Refl/Num/Gen/Case For ®rst- and second-person inputs, tableau T8.20 shows how person and number are preserved (the tableau shows ®rst person, but the same basic analysis holds for second persons). The Faith constraint for ®rst and second person, and the combination markedness constraint that bans the combination of ®rst and second person with re¯exive, gender, and case, must both dominate faithfulness for re¯exive, gender, and case. The con¯ict will then be resolved by satisfying Faith1st/2nd and the combination constraint, eliminating re¯exive, gender, and case. Unlike the other speci®cations, number is always preserved in ®rst and second persons, so Faith1st/ 2nd and FaithNum must both dominate the combination constraint that bans them together. Tableau T8.20 First and second persons preserve person and number Input: [‡R 1 sg m acc]

Faith1st/2nd

FaithNum

a. ‡R 1 sg m acc f b. R 1 sg G C c. R P sg G C d. R 1 N G C

Mark 1st/2nd &Num

Mark 1st/2nd &Refl/ Gen/ Case

*

*!**

* *!

FaithRefl/ Gen/ Case

*** ***

*!

***

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

233

Tableau T8.21 Third-person re¯exives preserve just re¯exivity Input: [‡R 3 sg m acc]

FaithRefl

Mark ‡ Refl &Pers/Num/ Gen/Case

f a. ‡R P N G C

****

b. ‡R 3 sg m acc c. R 3 sg m acc

FaithPers/ Num/Gen/ Case

*!*** *!

(26) Faith1st/2nd, Mark1st/2nd & Refl/Gen/Case g ParseRefl/Gen/Case Faith1st/2nd, FaithNum g Mark1st/2nd & Num In sum, for any ®rst- or second-person input, the person and number are parsed and the remaining speci®cations are not. That this is true for nonre¯exive inputs as well as re¯exives can be seen from reexamining tableau T8.20. The tableau is exactly the same if the input is changed to being R, as it is if the input is changed to being plural, or dative, and so on. The remaining case is that of a third-person re¯exive input. The lexicon in (4) represents the output as a clitic that is ‡R but loses all other speci®cations. In the present terms, the combination markedness constraint that targets ‡R in combination with others, Mark ‡ Refl&Pers/Num/Gen/Case, must dominate the constraints that enforce parsing of person, number, gender, and case. These constraints must also be dominated by FaithRefl, to ensure that the re¯exive speci®cation is preserved and the others lost, due to the con¯ict with the combination markedness constraint. Tableau T8.21 illustrates the analysis. (27) FaithRefl, Mark ‡Refl&Pers/Num/Gen/Case g FaithPers/Num/Gen/Case With this ranking, the ‡R speci®cation will be preserved and the others eliminated, resulting in satisfaction of the combination markedness constraint in the optimal candidate. By appeal, then, to markedness constraints against combinations of properties encoded in a single clitic, we can derive the inventory of clitic pronouns used in the languages under study.16 A clitic speci®cation such as ``‡R 2 pl G C'' or `` R 3 sg G C'' is simply impossible, since the constraints will never select it as optimal for any input. Note that there is a di¨erence between the outputs in the case of ®rst and second person and the case of the re¯exive clitic. In the ®rst case both person and number are

234

Jane Grimshaw

Figure 8.3 Ranking diagram for faithfulness and markedness constraints

preserved in the output speci®cation, because the combination constraint involving ®rst/second and number is dominated by FaithNum as well as by Faith1st/2nd. In the second case only re¯exivity is preserved, because only FaithRefl dominates the combination markedness constraint targeting ‡Re¯: it dominates all the other faithfulness constraints. It is worth noting, and suggestive of a course for further research, that the rankings in ®gure 8.3 are broadly consistent with work on markedness in morphosyntactic systems reported in Aissen (chap. 3, this volume) and Artstein 1998. (The same is true of the rankings sketched in section 8.4.2.) Note that the rankings in ®gure 8.3 show that the combination markedness constraints of the form Mark1st/2nd&X dominate the combination markedness constraint Mark3rd&X. Following a line of research developed in Prince and Smolensky 1993, Smolensky 1995, Aissen (chap. 3, this volume), and Artstein 1998, the ranking of a markedness constraint against combinations is partly determined by the ranking of the individual markedness constraints. Mark1st/2nd is univerally ranked higher than markedness constraints for third person: this is what guarantees that ®rst and second persons are more marked than third person. I assume also that re¯exives are more marked than nonre¯exives and that dative is more marked than accusative (this is important in section 8.4.2; see Grimshaw 1997a, and Woolford, chap. 16, this volume, for further evidence). These markedness generalizations follow from the rankings in (28). (28) Mark1st/2nd g Mark3rd MarkDat g MarkAcc Mark ‡ Refl g Mark Refl

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

235

Tableau T8.22 Absence of a second-person form Input: [ R 2 pl m acc] f a.

Mark2nd&Pl

R 3 pl m acc

b. R 2 pl G C

FaithPers

FaithRefl/ Gen/Case

* *!

***

It is no accident that, for example, ®rst- and second-person clitics, which are marked relative to third, do not support further speci®cations for re¯exivity, gender, or case, while the unmarked third-person clitics do. The fundamental prediction of such proposals, transported to the domain of inventory derivation, is that more contrasts will be maintained in less marked forms, because the constraint that bans a marked speci®cation combined with any other will necessarily be ranked above a constraint that bans a less marked speci®cation combined with the same additional property. Thus properties that can combine in lexical representations of Romance clitics (and those that can combine in clitic combinations as sketched in section 8.4.2) are less marked than those not allowed to combine. This general line of explanation is extensible to quite arbitrary-looking problems, suggesting that they are not as arbitrary as they may appear. Note 2 alludes to the gap in Latin American dialects, where there is no second-person plural clitic, a third person being used instead. Tableau T8.22 shows that this state of a¨airs is derived if the relevant combination constraint dominates faithfulness to person. (Note that the Fill version of faithfulness is violated here.) The candidate in (a) is the Latin American optimum; the candidate in (b) is the winner in os dialects. Even the dialect that does not tolerate a second-person plural clitic can preserve a ®rst-person plural clitic and a second-person singular clitic, because Mark1st&Pl and Mark2nd&Sg are both ranked below faithfulness to person. This is again entirely consistent with a general markedness hierarchy in which second person is more marked than ®rst (see Grimshaw 1997a for more evidence for this conclusion from Romance clitics), and plural is more marked than singular; hence the combination markedness constraints against ®rst-person plurals and second-person singulars must necessarily be ranked below the combination markedness constraint against second-person plurals. If a marked form, such as second-person plural or dative plural, is missing from the lexicon, there is no prediction concerning the less marked counterparts, second person or dative singular and ®rst/third person or accusative plural. But if the less marked forms are impossible, their more marked counterparts must be ruled out, since the constraint ranking that eliminates the less marked combination will eliminate the more marked also.

236

Jane Grimshaw

In this way, the clitic lexicons can be generated by faithfulness constraints, together with two kinds of markedness constraints: simple constraints against feature speci®cations and markedness constraints against combinations of feature speci®cations, both ranked in accordance with the markedness hierarchies of universal grammar. The inventory derivation laid out here is, as far as I can see, inconsistent with a theory that either enforces underspeci®cation whenever it is possible or forbids it except when it is absolutely necessary. Instead, whether a property is speci®ed or not is a function of the ranking of the constraints that favor or disfavor it. This means that one cannot simply study the paradigm in table 8.1 and determine the lexical representation of the clitics. The data for clitics in isolation is perfectly consistent, for example, with the hypothesis that se/si has no speci®cations at all (indeed that was the analysis given in Grimshaw 1997a). However, to predict the behavior of this clitic in combinations, and to predict the nonexistence of clitics like se/si with case speci®cations, it is necessary to analyze se/si as re¯exive, which the constraint ranking does. 8.6

Conclusion

The major points of this chapter concern the role of alignment constraints in determining the relative positions of clitics in the three Romance systems studied here, and the role of markedness and faithfulness constraints in determining the lexical representations that form the clitic lexicons. The order of clitics is governed by alignment constraints and is thus subject to the same fundamental principles as any other syntactic ordering, including relative order among heads, speci®ers, and complements (Grimshaw 1997b). The inventory of clitics in each system is determined by the interaction of markedness constraints, which target combinations of morphosyntactic speci®cations, and faithfulness constraints, which mandate preservation of input speci®cations. The derivation of the inventory is a step within the larger program sketched in Grimshaw 1997b, in which the functional lexicons of languages are all derived from the interaction of grammatical constraints. In Grimshaw 1997b, this logic was applied analyzing the distribution of the complementizer that and the auxiliary verb do. In Grimshaw and Samek-Lodovici 1998 it was applied to the English expletive it. These two threads in the proposal made here, together support the view that pronominal clitics in Romance are exemplary of the principles governing functional heads. Their inventory is derived by ranked violable constraints, and their position is determined by syntactic alignment. Notes The research reported here was supported by grant SBR-9511891 from the National Science Foundation to Rutgers University. An earlier version of the alignment constraint analysis was presented at the Hopkins Optimality Workshop, the University of Rochester, Stanford Uni-

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

237

versity, the University of Texas, Harvard University, the ``Lexicon in Focus'' workshop at the UniversitaÈt DuÈsseldorf, the ``Workshop on Clitics'' at the Deutsche Gesellschaft fuÈr Sprachwissenschaft, Konstanz, and the workshop on clitics at the Linguistic Society of America Linguistics Institute, University of Illinois at Urbana±Champaign. Questions, suggestions, and comments from the audiences have sharpened the analysis presented here considerably. Many individuals have contributed to this project: especially Ron Artstein, Eric BakovicÂ, Edward Keer, Nicole Nelson, Alan Prince, Vieri Samek-Lodovici, Bruce Tesar, and Colin Wilson. I would like to thank GeÂraldine Legendre and Sten Vikner for comments on earlier versions of this chapter. 1. See Uriagareka 1995 and Terzi 1999 for discussion of cliticization in non®nite contexts. 2. This clitic does not occur in dialects of Latin America, where a third person replaces it. See section 8.5 for an explanation. 3. See Cardinaletti 1991 for an analysis of Italian loro. 4. ci in the second column of (7) is a locative clitic. Some Italian clitics have a di¨erent phonological shape when they precede a clitic. I will not analyze this e¨ect here. 5. A minor variation of this analysis posits AccRt rather than CaseRt as the dominant constraint. This requires modi®cation of the analysis of spurious se (see section 8.3.4), with DatRt now the constraint that forces replacement of the dative clitic by se. 6. The constraint proposed in Heap 1996 ``Least Leafy Last'' is, in this view, not a constraint at all, but a description of the state of a¨airs resulting from the particular constraint ranking of Spanish, which has the e¨ect of positioning clitics with more person and case speci®cations to the right of clitics with less. 7. Either PersRt or 2Rt must dominate 2Lft, to guarantee that the second-person clitic follows se. This is consistent with, but not included in, the constraint rankings presented here. 8. It also follows if 1Lft g 1Rt, provided that 1Lft g PersLft also. 9. In an earlier analysis (Grimshaw 1997a) I suggested that the clitic combination violated the Obligatory Contour Principle, despite the fact that the clitics are not completely identical. The present alternative makes such an appeal unnecessary. 10. In discussing spurious se, I am assuming that the replacement clitic has to be one drawn from the generally available inventory of Spanish clitics. Thus le is replaced by se. But why couldn't there be an additional clitic that parses every input speci®cation but case, or at least parses more of the input than se does? It would satisfy CaseRt vacuously, and be more faithful to the input than se. This clitic would occur just in spurious se con®gurations. The constraint system laid out in section 8.5 precludes the existence of such a clitic. The reasoning is as follows. The clitic can improve over se only on the Parse constraints that are ranked below ParseRefl, and the combination markedness constraint (see (27)). If the hypothetical clitic parses the ‡R in the input it will fatally violate the combination constraint banning ‡R together with whatever additional feature in the input the hypothetical clitic parses. If it does not parse the ‡R in the input it will fatally violate the ParseRefl constraint. Hence no such specialized clitic can survive, given the rankings that derive the clitic inventory. 11. Alternative structures can be analyzed in a manner that parallels that developed for ®gure 8.2. For example, if the clitics adjoin to I 0 , alignment constraints relating the right and left edges of IP will still give the right orders, with the leftmost clitic separated from the left edge of IP by Spec, and the rightmost clitic separated from the right edge of IP by I and VP.

238

Jane Grimshaw

12. This proposal does raise the possibility that alignment could be (better) satis®ed if the subject were droppedÐthat is, that the presence of clitics could force omission of the subject. However, if deletion in syntax may occur only where there is a speci®c constraint requiring omission, such as DropTopic of Grimshaw and Samek-Lodovici 1998, this problem will not arise. 13. As can be seen in (1) and (2), there are minor di¨erences between the three lexiconsÐfor example, French has les for both genders in the third-person plural accusative, while Italian has li and le. At least for now, I take it that these represent nothing more than arbitrary differences in the mappings between morphosyntactic representations and phonological forms, so that les, for example, accidentally corresponds to two di¨erent morphosyntactic clusters. It is the existence of the clusters themselves that is predictable from the constraint system presented here. 14. This is closely related to Smolensky's 1995 proposal to allow constraint conjunction. See the discussion in Legendre, forthcoming, as well as chap. 1, this volume; Legendre, Smolensky, and Wilson 1998; Aissen, chap. 3, this volume; Choi, chap. 6, this volume. The constraints suggested here ban combinations of properties and may not involve conjoining constraints. 15. There is some evidence that masculine gender and singular number should not be speci®ed even in these forms, based on agreement phenomena (see Grimshaw 1997a for a constraintbased account). 16. Other combination constraints such as Mark Gender & Case, and the general version of the combination constraints targeting person speci®cations (Mark Person & X), must all be ranked below any faithfulness constraints that they con¯ict with and are not shown in ®gure 8.3. Also a crucial part of the system are the constraint rankings established in section 8.2, which guarantee the choice of the appropriate morphosyntactic speci®cation for an input where there is no perfect output to be selected and hence play a critical role in determining the inventory. In a suitably re®ned versionÐfor example, one that distinguishes ®rst and second from third personÐthese rankings, which involve faithfulness constraints, are consistent with ®gure 8.3 but are not shown there. References Anderson, Stephen. 1992. A-Morphous Morphology. Cambridge, England: Cambridge University Press. Anderson, Stephen. 1996. Rules and Constraints in Describing the Morphology of Phrases. In A. Dainora, R. Hemphill, B. Luka, B. Need, and S. Pargman (eds.), CLS 31, Vol. 2, The Parasession on Clitics, 15±31. Chicago Linguistic Society, University of Chicago. Anderson, Stephen. Forthcoming. Toward an Optimal Account of Second Position Phenomena. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Artstein, Ron. 1998. Hierarchies. Ms., Department of Linguistics, Rutgers University, New Brunswick, N.J. Available at http://www.eden.rutgers.edu/@artstein/publications.html. Barbosa, Pilar, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.). 1998. Is the Best Good Enough? Optimality and Competition in Syntax. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics.

Optimal Clitic Positions, Lexicon in Romance Clitic Systems

239

Bonet, EulaÁlia. 1994. The Person Case Constraint: A Morphological Approach. In Heidi Harley and Colin Phillips (eds.), MIT Working Papers in Linguistics 22: The MorphologySyntax Connection, 33±52. MITWPL, Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Bonet, EulaÁlia, 1995. Feature Structure of Romance Clitics. Natural Language and Linguistic Theory 13: 607±647. Cardinaletti, Anna. 1991. On Pronoun Movement: The Italian Dative Loro. Probus 3: 127± 153. Dekkers, Joost, Frank van der Leeuw, and Jeroen van de Weijer (eds.). Forthcoming. Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Gerlach, Birgit. 1998. Restrictions on Clitic Sequences and Conditions on the Occurrence of Clitics in Romance. Arbeiten des Sonderforschungsbereichs 282: Theorie des Lexikons. UniversitaÈt DuÈsseldorf, no. 105. Available as ROA-277-0898 at Rutgers Optimality Archive. Grimshaw, Jane. 1997a. The Best Clitic: Constraint Interaction in Morphosyntax. In L. Haegeman (ed.), Elements of Grammar: A Handbook in Contemporary Syntactic Theory, 169± 196. Dordrecht: Kluwer. Grimshaw, Jane. 1997b. Projection, Heads, and Optimality. Linguistic Inquiry 28: 373±422. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky, (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 193±219. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Haverkort, Marco. 1993. Clitics and Parametrization: Case Studies in the Interaction of Head Movement Phenomena. Doctoral dissertation, Katholieke Universiteit Brabant. Heap, David. 1996. Optimalizing Iberian Clitic Sequences. In Jose Lema and Esthela TrevinÄo (eds.), Theoretical Analyses on Romance Languages. Amsterdam: Benjamins. Kayne, Richard. 1991. Romance Clitics, Verb Movement, and PRO. Linguistic Inquiry 22: 647±686. Legendre, GeÂraldine. 1996. Clitics, Verb (Non)-Movement, and Optimality in Bulgarian. Technical Report JHU-CogSci-96-5. Department of Cognitive Science, Johns Hopkins University, Baltimore. Legendre, GeÂraldine. Forthcoming. Morphological and Prosodic Alignment of Bulgarian Clitics. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Legendre, GeÂraldine, Paul Smolensky, and Colin Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. McCarthy, John, and Alan Prince. 1993. Generalized Alignment. In Geert Booij and Jaap van Marle (eds.), Yearbook of Morphology 1993, 79±153. Dordrecht: Kluwer. McCarthy, John, and Alan Prince. 1995. Faithfulness and Reduplicative Identity. Ms., University of Massachusetts, Amherst, and Rutgers University, New Brunswick, N.J. Available as ROA-60 at Rutgers Optimality Archive.

240

Jane Grimshaw

McCarthy, John, and Alan Prince. 1999. Faithfulness and Identity in Prosodic Morphology. In Rene Kager, H. v. d. Hulst, and W. Zonneveld (eds.), The Prosody-Morphology Interface, 218±309. Cambridge, England: Cambridge University Press. Miller, Philip H., and Ivan Sag. 1997. French Clitic Movement without Clitics or Movement. Natural Language and Linguistic Theory 15: 573±639. Monachesi, Paola. 1995. A Grammar of Italian Clitics. Doctoral dissertation, Tilburg University, Tilburg, The Netherlands. Tilburg Dissertation in Language Studies, TILDIL 1995-3. Morin, Y.-C. 1979. More Remarks on French Clitic Order. Linguistic Analysis 5: 293±312. Perlmutter, David. 1971. Deep and Surface Structure Constraints in Syntax. New York: Holt, Rinehart and Winston. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Samek-Lodovici, Vieri. 1998. Opposite Constraints: Left and Right Focus-Alignment in Kanakuru. Lingua 104: 111±130. Simpson, J., and M. Withgott. 1986. Pronominal Clitic Clusters and Templates. In Hagit Borer (ed.), The Syntax of Pronominal Clitics. Syntax and Semantics Vol. 19. New York: Academic Press. Smolensky, Paul. 1995. On the Internal Structure of the Constraint Component Con of UG. Handout of paper presented at UCLA, Los Angeles. Available as ROA-86-0000 at Rutgers Optimality Archive. Terzi, Arhonto. 1999. Clitic Combinations, Their Hosts, and Their Ordering. Natural Language and Linguistic Theory 17: 85±121. Uriagereka, Juan. 1995. Aspects of the Syntax of Clitic Placement in Western Romance. Linguistic Inquiry 26: 79±123. Van der Leeuw, Frank. 1995. Alignment and Integrity Constraints in Cliticization. In A. Dainora, R. Hemphill, B. Luka, B. Need, and S. Pargman (eds.), CLS 31, Vol. 2, Chicago Linguistic Society, University of Chicago. The Parasession on Clitics, 168±197. Wanner, Dieter. 1977. On the Order of Clitics in Italian. Lingua 43: 101±128.

Chapter 9 Masked Second-position E¨ects and the Linearization of Functional Features

GeÂraldine Legendre

There is a striking word order pattern that has typically been treated as two independent phenomena in the generative literature. Some languages, including all Germanic languages except English, require ®nite verbs to appear in second position (at least in main clauses). Other languages, including many South Slavic languages, require their pronominal, auxiliary, and other clitics to cluster in second position. The standard analysis of the verb-second phenomenon, due originally to den Besten (1983), relies on local head movementÐthat is, a ®nite verb moving from I to C at some point in the derivation (see Vikner 1995:51±64 for a summary of various implementations of this basic idea). On the other hand, analyses of second-position clitics have typically relied on syntactic mechanisms like right adjunction to some head for pronominal clitics (Progovac 1996) and nonlocal head movement across clitic auxiliaries (Rivero 1994), and/or postsyntactic linearization mechanisms like Prosodic Inversion (Halpern 1995) or Morphological Merger (Marantz 1988).1 In general terms, these analyses all exploit the idea that clitics share the syntactic properties of their nonclitic counterparts and are special only to the extent that they are subject to additional prosodically based constraints. In other words, all types of elements subject to second-position requirements are syntactically independent elements, but the mechanisms invoked to handle their exact positioning are, to a large extent, di¨erent. The syntactic approach to clitics has long been challenged by morphologists, including Klavans (1985), Anderson (1992), and others, on the grounds that clitics share fundamental properties with a½xes that are not easily captured within the syntactic approach (see Anderson, 1992, 1993 for an extensive discussion). The latter view has received renewed interest from an Optimality-Theoretic (OT) perspective in the works of Anderson (1996, forthcoming) and Legendre (1996, 1998a, 1999a, forthcoming b,c). This work, which includes detailed analyses of the clitic systems of a number of related and unrelated languages with and without second-position e¨ects, builds on the original claims made in Klavans 1985 and Anderson 1992 but makes crucial use of constraint con¯icts and their OT resolution to explain the

242

GeÂraldine Legendre

diverse patterns observed. Ignoring some di¨erences in terminology and technical details of analysis, Anderson and I have both argued that clitics are the phonetic form (PF) spell-out of morphosyntactic features. That is, clitics are not independent lexical items inserted in phrase structure and moved around on a par with syntactic categories. In Anderson's terms, they are phrasal a½xesÐa½xes that take phrases rather than stems as their hosts. (See the references above for details.) From this perspective, a promising approach to the positioning of clitics exploits the existing alignment-based theory of (word-level) morphology in OT (Prince and Smolensky 1993; McCarthy and Prince 1993b). In this view, the edge-alignment constraints responsible for clitic placement are instantiations of McCarthy and Prince's general schema Align(Category1 , Edge1 ; Category2 , Edge2 ). For example, edge-alignment constraints like Edgemost (Anderson 1996, forthcoming; Legendre 1996, 1998b, 1999a, forthcoming b,c) immediately explain clustering e¨ects. Clitics cluster because they compete for a single position. Clearly, only one clitic may be realized at the very edge of a given domain; others will follow, striving to be as close as possible to that edge. This approach entails (gradiently) violable constraints, since noninitial clitics in the cluster will violate their respective Edgemost constraint. Second-position e¨ects also receive a natural explanation: they result from the interaction of violable Edgemost(Left) with another alignment constraint Noninitial, where Noninitial outranks Edgemost(Left). These interactions will be exempli®ed below.2 Alignment constraints also ®gure prominently in Grimshaw's analysis of possible and impossible combinations of Romance pronominal clitics (Grimshaw, chap. 8, this volume) where they interact with faithfulness to featural speci®cation. For example, positional constraints like DativeRight, AccusativeLeft, and so on determine possible French combinations. Grimshaw assumes that the overall positioning of clitics in a clause is subject to alignment with the edge of IP in the syntactic component, but nothing in her core analysis hinges on that assumption. For Anderson and myself, the striking parallelism between the positioning of ®nite verbs and clitics in second position cannot fail to raise the issue of a common morphologically based approach. Anderson sketches one in his own terms in Anderson (forthcoming). The present chapter takes the hypothesis one step further by arguing that extending the alignment-based analysis of clitics to verb-second phenomena is in fact empirically well motivated, based on a detailed study of constraint interaction in two unrelated non-Germanic languages with distinct word order strategies:3 Breton (VSO) and Macedonian (SVO). The particular proposal made in this chapter is the following. Features traditionally subsumed under ®niteness are universally positioned by violable PF alignment constraints rather than by syntactic movement to a particular functional head position (as is generally assumed for verb-second phenomena since den Besten 1983; see

Masked Second-Position E¨ects, Linearization of Functional Features

243

also Vikner, chap. 14, this volume). Finiteness here refers to the overt morphological marking of distinctions pertaining to the temporal (tense and aspect) and/or person/ number speci®cations of a given predication. Henceforth, ®niteness is referred to as [F]. In this preliminary study, [F] subsumes all component features of ®niteness and does not distinguish tensed verbs from agreeing verbs. In fact, all ®nite verbs discussed in the chapter are both tensed and agreeing. The proposal is grounded in a three-way empirical generalization: (1) clitic auxiliaries are syntactically inert phrasal a½xes positioned at PF, (2) other verbal categories are syntactically active heads positioned in the syntax, (3) yet the positioning of [F]Ðin second positionÐis oblivious to that distinction. A uni®ed analysis of the positioning of all ®nite verbal categories is only possible in PF terms. Verb-second languages allow positional e¨ects of [F] to be easily recovered. The focus here is on masked verb-second e¨ects because they most clearly reveal how surface position results from the interaction of di¨erent constraint families. Because verb-second phenomena are found in verb-initial languages (e.g., Breton), the traditional label V2 is not particularly useful. I will henceforth use the neutral term F2 to refer to ``®niteness in second position.'' The chapter is organized as follows. Section 9.1 presents F2 e¨ects in Breton and Macedonian. Section 9.2 compares possible hosts of [F] in root clauses. Section 9.3 is devoted to the technical details of linearization and a comparative analysis of Breton and Macedonian. Section 9.4 summarizes the proposal. 9.1

Masked F2 E¨ects

The traditional characterization of a given language's basic word order inherited from typological studies (e.g., Greenberg 1966) pertains to lexical categories: S, V, O. As is well known, a considerable number of languages exhibit both relatively free word order and a so-called basic word order. The latter is often characterized as the pragmatically unmarked and statistically most frequent word order in a given language. Despite the fact that these languages have a basic word order, many di¨erent permutations of arguments and V are possible, subject only to discourse considerations. In our terms, this means that constraints pertaining to the linearization of arguments are outranked by discourse constraints pertaining to focus, topics, and so on. The positioning of functional categories like the ones traditionally subsumed under ®niteness is of particular interest because the positioning is typically not governed by discourse considerations. On the other hand, the positioning of functional categories does not necessarily correlate with the basic word order of a language, as we will see shortly. Breton has the characteristic VSO word order of Celtic languages, as exempli®ed in (1).4 An auxiliary encodes tense, person, and number (en `third-person masculine

244

GeÂraldine Legendre

singular', em `®rst-person singular', and so forth). The present tense auxiliary root in the third person is deus (in®nitive endevout `have').5,6 Note that en deus is pronounced as [nos] (and optionally spelled as neus), evidence that it constitutes a phonological unit. Note further that the auxiliary (complex) surfaces in second position, immediately after the non®nite lexical verb. The reverse order aux-V is ungrammatical, despite the fact that VSO word order is maintained. (Some in¯ected verbs look exactly like their nonin¯ected counterpartsÐsee the comment on lenn `read' below. Hence, the ®nite verbal category will be italicizedÐand in some cases also underscored, when necessary for clarityÐin the examples.) (1) Breton a. Lennet en deus Yann al levr. read 3 have Yann the book `Yann has read the book.' b. *En deus lennet Yann al levr. c. Kavet em eus bet al levr. found 1 have been the book `(I) have found the book' d. *Em eus bet kavet al levr. In the absence of a ®nite auxiliary, Breton opts for two strategies. One involves an expletive auxiliary formÐra in (2a)Ðpreceded by the in®nitive (or verbal noun) form of the lexical verb. The resulting word order, again, is VSO. The second strategy involves topicalizing an argument, resulting in SVO or OVS word order, as in (2b) and (2c). Breton has an a½rmative particle a or e, depending on the grammatical function of the DP preceding it (Stump 1989). The verbal form lenn is non®nite in (2a) but ®nite in (2b) and (2c). (2) Breton a. Lenn a ra Anna al levr. read-inf part do-3 Anna the book `Anna reads the book.' b. Anna a lenn a levr. A part read-3 the book `Anna reads the book.' c. Al levr a lenn Anna. the book part read-3 A `Anna reads the book.' The generalization that emerges from (1) and (2) is that Breton simply requires the ®nite verbal category to appear in second position (Ortiz de Urbina 1994; Schafer 1995), a constraint independent of its basic VSO word order. Breton deploys various strategies to ensure that this requirement is met: past participle fronting in (1), an

Masked Second-Position E¨ects, Linearization of Functional Features

245

expletive auxiliary in (2a), and fronting of an argument in (2b) and (2c). A negative particle can also be drafted into securing the second position of an in¯ected auxiliary, as shown in (3). Clausal negation is expressed by means of two negative particles, ne and ket, surrounding the ®nite verbal category (much like French ne . . . pas). (3) Breton a. N 0 en deus ket lennet Yann al levr. neg 3 have neg read Yann the book `Yann has not read the book.' b. Ne lenn ket Yann al levr neg read neg Yann the book `Yann did not read the book.' It is interesting to note that the F2 e¨ect does not always hold in Breton, however. Consider (4), one of several patterns to be analyzed in which F2 is masked. (4) Breton a. E c'halvet en deus Yann. him called 3 have Yann `Yann has called him.' b. *E c'halvas Yann. him called Yann `Yann called him.' Pronominal clitics like e `him' appear in initial position with the result that the ®nite auxiliary complex is in third position in (4a), counting every terminal element that separates it from the clausal edge. Note also that F2 is satis®ed in (4b), yet the result is ungrammatical. Such masking e¨ects, I will argue, simply emerge from constraint interaction. See section 9.3.3 for details. In Macedonian, an F2 e¨ect is found with n/t-participles (also known as verbal adjectives) and adjectival or nominal predicates. In null-subject contexts, the lexical (non®nite) predicate must precede the auxiliary, as shown in (5). The reverse order is ungrammatical. Italicized items are clitics, including present perfect auxiliary sum and copula si, both in¯ected forms of `be'. (5) Macedonian a. Dojden sum. (n/t-participle) come be-1 `(I) have come.' b. *Sum dojden. c. Mil si mi. (adjectival predicate) dear be-2 me-dat `(You) are dear to me.' d. *Si mi mil.

GeÂraldine Legendre

246

In Macedonian, the F2 e¨ect disappears with a class of non®nite verbal forms called l-participles (because of their l-ending, as in zel `taken'). In (6a), the clitic auxiliary sum appears in initial position (in null-subject contexts); it is separated from the l-participle by pronominal clitics with which it forms a cluster. The su½x -l is a modality marker that, in conjunction with the auxiliary `be', conveys secondhand or hearsay knowledge of an event. (6) Macedonian a. Sum mu gi zel parite. (l-participle) be-1 him-dat it-acc taken money `(I) have taken the money from him.' b. *Zel sum mu gi parite. Summing up, despite their obvious basic word order di¨erences, Breton and Macedonian have in common that (1) they are F2 systems and (2) they are not perfect F2 systems. The F2 e¨ect, to some extent, is invisible on the surfaceÐthat is, masked. Masked F2 e¨ects will be argued below to be the direct consequence of constraint interaction. The two languages di¨er, however, with respect to which constraints interact to make them not perfect. In Breton, masked F2 results from the interaction of alignment constraints with information structure constraints or across distinct features. The Macedonian F2 e¨ect results from a competition for second position among distinct features. Note that masked F2 e¨ects are in fact to be expected if violable constraints are involved. That is, a (set of ) constraint(s) that results in second position of a ®nite verbal category may be violated to satisfy a higher-ranked constraint whose e¨ect is, for example, to front a focused argument, thereby a¨ecting the surface position of the ®nite verbal category. This, and other formally similar interactions, will be illustrated below. 9.2

The Status of Finite Verbal Categories

The common F2 pattern is all the more interesting because the status of the aspectual auxiliary (`have' in Breton; `be' in Macedonian) di¨ers in the two languages. In Macedonian, `be' has the status of clitic (Tomic 1996). In Breton, `have' has the status of head of a syntactic projection. Evidence that an auxiliary is a (morphological) clitic rather than a syntactic head in a given language comes from the lack of syntactic properties characteristic of syntactic heads as well as typical lack of prosodic stress. I illustrate the ®rst property with two diagnostic tests belowÐthat is, the absence of subject-auxiliary inversion and the positioning of adverbs. Further testsÐ including the possibility of discourse-neutral long head movement across clitic aux-

Masked Second-Position E¨ects, Linearization of Functional Features

247

iliaries (Rivero 1994) but not other verbal forms, as well as coordination patternsÐ are discussed in Legendre (forthcoming b,c). In languages in which an auxiliary heads a syntactic projection, the auxiliary permits syntactic movement. This is the case in Breton and German. As (7) shows, (wh- and yes-no) questions involve subject-aux inversion in both Breton and German. (7) Breton a. Piv en deus Yann gwelet dec'h? who 3 have Yann seen yesterday `Who did Yann see yesterday?' German b. Wen hat Peter gestern gesehen? who has-3 Peter yesterday seen `Who did Peter see yesterday?' In (7a), for example, the subject Yann intervenes between the auxiliary en deus and the lexical verb gwelet (as in English). Auxiliaries that are clitics, however, systematically fail to undergo subject-aux inversion in (wh- and yes-no) questions. The pattern is not speci®c to Macedonian; it extends to other languages known to have clitic auxiliaries, including Bulgarian, Romanian (Legendre, forthcoming c), and Basque (Legendre, forthcoming a). (8) Macedonian a. KÂe ti ja dade li Penka knigata? will you-dat it-acc gave-3 Q Penka book-the `Will Penka give you the book?' Basque b. Hil da aita? died aux-3 father `Has father died?' As (8a) shows, the inverted subject Penka follows the aux-verb complex kÂe dade (and pronominal clitics in between) rather than simply the auxiliary kÂe.7 In other words, head movement (to a projection higher than VP) operates as if the auxiliary were not present in the structure.8 The absence of subject-aux inversion does not make sense if Basque da and Macedonian kÂe head functional projections of their own, as their counterparts presumably do in Breton and German. Head movement would have to involve two verbal heads and move them in the right order. But if kÂe and da are phrasal a½xesÐthat is, PF exponents of functional featuresÐthey are irrelevant to verb movement motivated by the presence of a syntactic operator, OpSpec and ObligatoryHeads (Grimshaw 1997). One could stipulate the existence of a dominating syntactic constraint requiring all verbal elements to appear together when a

248

GeÂraldine Legendre

clitic auxiliary is present, but that would only recapitulate the empirical generalization and would hardly constitute a genuine explanation. For it should be noted that the absence of subject-aux inversion in Macedonian, Bulgarian, Romanian, Basque, and so on is not a general property of all auxiliaries but a property that distinguishes clitic auxiliaries from their nonclitic counterparts within each language. This will be exempli®ed in Macedonian in (12); see the references given earlier for the other languages. There is substantial independent evidence for the clitic status of (certain) Macedonian auxiliaries. Other clitics (e.g., accusative and dative pronouns) may intervene between the auxiliary and the main verb, but lexical elements may not. (9) Macedonian a. Jas sum mu gi zel parite. I be-1 him-dat it-acc taken money `I have taken the money from him.' b. *Jas sum mu gi parite zel. Macedonian pronominal clitics are best analyzed as agreement markersÐthat is, morphological categories. Macedonian has obligatory clitic doubling with de®nite direct and indirect objects (Tomic 1996). Note that there is no preposition to assign case to the object DP, yet the result is grammatical. (10) Macedonian Goi imam videno Petari . him have-1 seen Peter `(I) have seen Peter.' If these pronominal clitics were active in the syntax, one would expect blocking e¨ects such as strong crossover e¨ects. No such e¨ects, however, are found in Macedonian, as shown in (11).9 (11) Macedonian Ï ovekoti kogoi sÏto goi C vidov ti man whom that him-acc saw-1 `The man whom I saw.' This con®rms the agreement marker status of Macedonian pronominal clitics. Henceforth, I take them to be the PF spell-out of agreement, a proposal reminiscent of the one made in Borer 1984 and SunÄer 1988 on the basis of other languages. See Legendre 1998a for an OT analysis of clitic doubling. Macedonian has two series of perfects, the Proto-Slavonic sum `be' series exempli®ed above and the ima `have' series unique to Macedonian. While sum has the status of clitic, ima has the status of syntactic head of a projection. Evidence comes from

Masked Second-Position E¨ects, Linearization of Functional Features

249

subject-aux inversion in interrogative contexts, as in (12a), and from the fact that adverbs may intervene between ima and the n/t participle, as in (12b). (12) Macedonian a. Go ima li Ivan svrseno? it-acc have-3 Q Ivan solved `Has Ivan solved it?' b. Go imasÏe sam napraveno. it-acc have-3 himself done `(He) has done it himself.' Interestingly, pronominal proclitics ensure that ®nite ima is in second position in (12a), just as they do with a simple ®nite lexical verb like vide in (13a). Note that a single clitic is su½cient to ensure the second position of the element carrying [F].10 If pronominal clitics are the PF spell-out of agreement, as argued above, then the pattern in (12) and (13) is evidence that F2 is a requirement that holds at PF, but not in the syntax. (13) Macedonian a. Go vide. him-acc gave-3 `(She) saw him.' b. Dojden sum. come be-1 `(I) have come' c. Mil si mi. dear be-2 me-dat `(You) are dear to me.' Consider the F2 generalization further, on the basis of (12) and (13). Finiteness is systematically realized in second position, whether its host is a phrasal a½xÐas in (12a) and (12b) as well as (13a)Ðor a syntactic head, as in (13b) and (13c). The status of the auxiliary, either as a PF or a syntactically independent element, is irrelevant to the linearization of ®niteness features. This stands in clear contrast to the fact that the status of the auxiliary is clearly relevant to syntactic movement (e.g., subject-aux inversion in questions) and to the positioning of adverbs. Because F2 applies to auxiliaries that are the PF spell-out of functional features, linearization of [F] must take place at PF. In contrast to the situation in Macedonian, adverbs may separate V from the auxiliary in Breton (Borsley 1990) and German, evidence that con®rms the head status of the auxiliary in these two languages.

GeÂraldine Legendre

250

(14) Breton a. Debret mat en deus d' he goan. eaten well 3 have to his supper `He has eaten well for supper.' German b. Er ist schon gegangen. `He has already left.' Summing up section 9.2, there is substantial crosslinguistic evidence that clitic auxiliaries, contrary to their nonclitic counterparts, are not independent lexical items undergoing movement in the syntax.11 Like the pronominal clitics discussed above,12 they are in fact syntactically inertÐwhich I interpret as evidence that they do not have the status of syntactic head. Rather, clitic auxiliaries are exponents of functional features linearized at PF on the basis of alignment constraints familiar from morphophonological studies (McCarthy and Prince 1993b).13 Crosslinguistically and language internally, the F2 e¨ect is oblivious to status di¨erences among auxiliaries. The fact that it is found with auxiliaries that are not syntactic objects indicates, however, that the F2 e¨ect must be construed primarily as a PF e¨ect. An important generalization would otherwise be missed. 9.3

Linearization of [F]

This section focuses on the OT analysis itself. It opens with some general considerations. 9.3.1 The General Hypothesis The general hypothesis examined in this chapter is that, in terms of its linearization, [F] is subject to the same linearization constraints as clitics. The proposed theory of features relies on two proposals, the ®rst of which is not further explored in the present chapter. The ®rst proposal pertains to the category status of the element bearing a feature. That is, a functional feature may be realized as a (word- or phrase-level) a½x or as a syntactic head. Features are assumed to be listed in the lexicon (Everett 1996). Their category status is derived from a competition between a constraint on realizing features on syntactic heads and other constraints that favor a di¨erent realization. This approach to the lexicon, whereby its grammatical properties are derived from the interaction of constraints rather than simply stipulated, is defended in a number of previous OT studies, including Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998; Grimshaw 1997 as well as chap. 8, this volume; and Grimshaw and Samek-Lodovici 1998. In the interest of space, the present chapter takes a shortcut, however, by simply assuming that the

Masked Second-Position E¨ects, Linearization of Functional Features

251

category status of particular functional features is predetermined (i.e., the relevant competitions are left out). The second proposal is that all morphosyntactic features, including [F], are subject to violable alignment constraints, the precise nature of which is the topic of the next section. Drawn from the lexicon, [F] is present in the input as a property of V (which I interpret as the head of the clause). The positioning of the category that instantiates [F], however, is largely regulated by PF alignment, independently of its category status. In the present analysis, clitics are ``interface phenomena.'' Note that the parallel, rather than sequential, approach to constraint interaction in OT originally formulated in Prince and Smolensky 1993 entails a concept of the interface that pertains to the range of the constraint hierarchy involving the interaction of constraints on di¨erent parts of the linguistic representation. The two parts of the representation are interdependent: syntax (by virtue of encoding important aspects of word order) partially determines the PF representation. However, constraints on hierarchical structure only see items that are present in the syntactic tree. Constraints on the PF part of the representationÐin particular (and possibly only) alignment constraintsÐonly see the morphophonological representation of lexical items; hence they treat heads, word-level a½xes, and clitics alike. 9.3.2 Alignment Constraints Previous work on linearization of clitics in OT (Legendre 1996, 1998b, 1999a, forthcoming b,c; Anderson 1996, forthcoming) has relied on the alignment-based theory of morphology (Prince and Smolensky 1993; McCarthy and Prince 1993b). In particular, second-position e¨ects result from a particular interaction of two con¯icting constraints within a particular domain, Noninitial(X) and Edgemost(X, Left), which I have stated in previous work as follows. (15) a. Noninitial(X): At PF a feature [X] is not realized in intonational phraseinitial position. b. Edgemost(X, Left): At PF a feature [X] is left-aligned with the edge of the nearest projection of the head [X] is associated with. Assuming for the moment a simple null-subject contextÐthat is, a situation where the domains of both constraints completely overlapÐconstraint ranking works as follows. If Edgemost(X) outranks Noninitial(X), a feature [X] violates Noninitial(X) to satisfy Edgemost(X), and [X] is linearized in domain-initial position. Such a ranking yields, for example, the Romanian cliticization pattern (Legendre, forthcoming c) as well as a subset of Bulgarian and Macedonian clitics (Legendre 1998b, 1999a, forthcoming b). If, however, Noninitial(X) outranks Edgemost(X), the latter is violated to ensure satisfaction of the former. Second-position clitics are

252

GeÂraldine Legendre

born out of that con¯ict, so to speak. They are as close to the domain's left edge as they can be without being on the very edge. This ranking applies to particular instantiations of [X], including [F]ÐNoninitial(F) g Edgemost(F)Ðyielding F2 e¨ects. These alignment constraints map the PF realization of features onto a domain that, in principle, could be syntactic or prosodic in nature. As argued in Legendre (forthcoming b), there is substantial crosslinguistic evidence that the domain of Noninitial is prosodic, while that of Edgemost is syntactic. The precise X 0 -theory characterization of the syntactic domain of Edgemost largely depends on one's assumptions about clausal structure. Under the VP-internal subject hypothesis (e.g., Koopman and Sportiche 1991 and others) and the present ``morphological'' analysis of ®niteness, this means that a simple clause containing a simple verb form needs not involve more than a VP (i.e., an alternative IP candidate involves an additional violation of MinimalProjection (MinProj, Grimshaw 1993). Only as much structure is built as required by the input (see also Grimshaw 1997). If null subjects do not exist but result from a syntactic constraint dropping a topic as proposed in Grimshaw and Samek-Lodovici 1998, then a null subject clause is simply a V 0 . This is the hypothesis I adopt below. Consider Macedonian l-participle constructions that permit domain-initial clitics in null-subject contexts. In the presence of an overt subjectÐ(nonclitic) jas `I' in (16b)Ðsum follows the subject rather than preceding it. This is indicative of alignment with a V 0 domain: Edgemost(F) is satis®ed in both (16a) and (16b).14 As the formal competitions displayed as OT tableaux will make clear, violations of Edgemost(F) are measured by the number of morphemes that separate a given clitic from the left edge of V 0 . (16) Macedonian a. [V 0 Sum mu gi zel parite]. `(I) have taken the money from him.' b. [VP Jas [V 0 sum mu gi zel parite]]. `I have taken the money from him.' As is well known, V 0 is not an active level with respect to syntactic movement though it is relevant to coordination and other constituency phenomena. Because phrasal a½xation is not syntactic, the usual syntactic objection to making use of V 0 for PF alignment does not hold. Clitics, in fact, provide evidence for the existence of V 0 , independently of constituency phenomena. Second-position elements, on the other hand, are necessary to recover the domain of Noninitial(F). A V 0 domain speci®cation predicts that auxiliary clitics would appear in second position in a null-subject context but in third position in an overt subject context. The Macedonian F2 pattern with n/t-participles provides evidence that this is not the case.

Masked Second-Position E¨ects, Linearization of Functional Features

253

(17) Macedonian a. [V 0 Dojden sum]. come be-1 `(I) have come.' b. [VP Jas [V 0 sum dojden]]. I be-1 come `I have come.' The position of sum changes, depending on the presence of an overt subject. What remains constant, however, is sum's second position overall in the clause. Hence the domain cannot be V 0 . As shown in (18), parentheticals and dislocations a¨ect the positioning of second-position elements. The following examples from a strict second-position clitic language, Serbo-Croatian, indicate that second position is counted from the prosodic boundary following the parenthetical. (18) Serbo-Croatian a. [ IntP=VP Ja [V 0 sam ti obecÂala igracÏku]]. I be-1 to-you promised toy `I promised you a toy.' b. [VP Ja, tvoja mama, [IntP=V 0 obecÂala sam ti igracÏku]]. promised be-1 you-dat toy I your mom `I, your mom, promised you a toy.' The parenthetical in (18b) separates the subject pronoun ja from the rest of the clause. If the prosodic boundary immediately following the parenthetical were irrelevant to the positioning of clitics, they would be expected to appear right after the pause, so as to best satisfy Edgemost (within V 0 ). Instead, they appear second after the pause, suggesting that the domain is prosodically de®ned (contra Anderson 1996, forthcoming). See RadanovicÂ-Kocic (1996) for further discussion of the relevance of prosodic domains to the positioning of Serbo-Croatian clitics. Going back to the Macedonian pattern, (17) follows under a prosodic characterization of the domain of Noninitial(F). There is only one intonational phrase and sum is second in it in both (17a) and (17b). Based on (17) and (18), I conclude that the domain of Noninitial(X) is the intonational phrase. Ortiz de Urbina (1994) provides additional evidence in Breton suggesting that the prosodic characterization of the domain of Noninitial(F) may in fact be universal. (19) Breton a. *Yann, meus roet al levr deshan. Yann have-1 given the book to him `As for Yann, I've given the book to him.' b. Yann, roet meus al levr deshan.

254

GeÂraldine Legendre

As (19) shows, dislocated constituents that belong to a separate intonational phrase do not count as ®rst elements. Past participle fronting is still required to ensure the second position of [F].15 Note that a clitic may satisfy the second-position requirement by following another prosodically weak element (i.e., another clitic) in some languages. For example, the Bulgarian perfect auxiliary clitic suÆm `be' can be separated from the left edge of the intonational phrase by clitics like the future auxiliary or the negative particle. (20) Bulgarian SÏte/ne suÆm procÏel knigata. fut/neg be-1 read book-the `(I) will have/have not read the book.' In a similar fashion, [F] may satisfy Noninitial by following pronominal clitics in Macedonian and the negative particle in Breton. Relevant examples are repeated in (21). (21) Macedonian a. Ti go dade. `(He) gave it to you.' Breton b. N 0 en deus ket lennet Yann al levr. `Yann has not read the book.' 9.3.3 Breton F2 E¨ects Recall the basic pattern of Breton. Despite its VSO word order, ®nite verbs may not appear in sentence-initial position. In a periphrastic verb construction, the past participle is fronted (so that [F] may satisfy Noninitial ). (22) Breton Lennet en deus Yann al levr. read 3 have Yann the book `Yann has read the book.' In the absence of a ®nite auxiliary, Breton opts for two strategies. One involves an expletive auxiliary formÐra in (23a)Ðpreceded by the in®nitive (or verbal noun) form of the lexical verb. The second strategy involves fronting an argument, resulting in SVO or OVS word order, as in (23b) and (23c). (23) Breton a. Lenn a ra Anna al levr. read-inf part do-3 Anna the book `Anna reads the book.'

Masked Second-Position E¨ects, Linearization of Functional Features

255

b. Anna a lenn a levr. Anna part read-3 the book `Anna reads the book.' c. Al levr a lenn Anna. the book part read-3 Anna `Anna reads the book.' Interestingly enough, the two strategies lead to di¨erent interpretations. Press (1986) emphasizes the neutral discourse status of (23a), while he and Schafer (1995) emphasize the nonneutral status of (23b) and (23c). In the latter but not the former, the fronted argument is focalized by virtue of encoding new information (a case of what Kiss (1998) calls ``information focus''; see also the contributions by Choi, Costa, and Samek-LodoviciÐchaps. 6, 7, and 11, respectivelyÐin the present volume). In Schafer's words, a ``subject-initial sentence is a classic answer to a question asking who did something'' (1995:151n.). This is an important clue that the input speci®cations underlying (23a) on the one hand and (23b) and (23c) on the other are di¨erent. The inputs to (23b) and (23c) include di¨erent constituents marked with the discourse feature [focus]; the input to (23a) includes no such information. As I will demonstrate below, past participle fronting in Breton has the dual function of enforcing the second-position requirement on [F] and matching the information structure of the output with the speci®cations of the input. That is, past participle fronting also ensures that no argument is interpreted as new information if the input does not contain a [focus] feature. Previous analyses have recognized the ®rst motivationÐthat is, F2 (Ortiz de Urbina 1994; Schafer 1995; Borsley, Rivero, and Stephens 1996)Ðbut not the second. Consider ®rst the formal optimization underlying the discourse-neutral auxiliary structure in (22). The auxiliary en deus is not a clitic; hence it is base generated in the head position of an extended projection of V (Grimshaw 1991). Incorporating Grimshaw and Samek-Lodovici's 1998 claim that null subjects do not exist, I assume that projections of V are simply V's in the absence of an overt speci®er. (Nothing hinges on this assumption, however.) The least costly way of achieving F2 is by fronting the non®nite verb, in violation of economy of structure and movement. As formally expressed in tableau T9.1, alternative options are more costly (including an expletive auxiliary for reasons that will be become clear with the next competition, tableau T9.2). In the interest of clarity and space, the candidate set in tableau T9.1 includes only the best of the lot: (a) fronting V, (b) not fronting V, and (c) fronting an argument, say O. The main constraints that interact to produce the basic Breton pattern are Noninitial(F) and Edgemost(F)ÐNI(F) and E(F) for shortÐand Faith, to be discussed shortly. (Finiteness is indicated by a subscript for easy identi®cation.) Candidate (b) is elimi-

GeÂraldine Legendre

256

Tableau T9.1 Breton VauxSO Input: V(S, O), [perf ], [F]

Faith

NI(F)

f a. [V 0 Vi [V 0 auxF [VP S [V 0 ti O]]]

E(asp)

*t Y

b. [V 0 auxF [VP S [V 0 V O]] c. [VP Oi [V 0 auxF [VP S [V 0 V ti ]]

E(F)

*! *!

*

nated by Noninitial(F). In each candidate, the ®nite auxiliary is leftmost in its V 0 , hence Edgemost(F) is satis®ed. The auxiliary also encodes [aspect], here perfective, hence the constraint Edgemost(asp). Its violations are parallel to those of Edgemost(F); they will be omitted in other competitions. Fronting V (candidate (a)) requires an additional projection whose head is the landing site of V-movement. V-movement (candidate (a)) and fronting of O (candidate (c)) entail violations of *t/Stay (Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998; Grimshaw 1997), but these violations of economy of structure and movement are compensated for by the fact that higher-ranked constraints are satis®ed. Consider the comparative cost of fronting an argument in discourse-neutral contexts, candidate (c) in tableau T9.1. I propose that candidate (c) is suboptimal because fronting either S or O entails nonrecoverability of the input (i.e., a violation of input-output faithfulness or Faith).16 That is, given an input where both S and O are discourse neutral, the optimal output, with either fronting of S or O, entails an interpretation in which the fronted element is not discourse neutral. This is clear from Press 1986 and Schafer 1995, as reported above. Their comments can easily be recast in OT terms, based on a constraint like AlignFocusLeft that is satis®ed when an element marked [focus] in the input is leftmost in the clause (Samek-Lodovici 1998).17 The focused XP presumably is also subject to a general constraint that places XPs in speci®er position. In tableau T9.1, there is no focus feature in the input, hence AlignFocusLeft is vacuously satis®ed and left out of the competition. Note, however, that AlignFocusLeft does not allow a listener to assign a focus feature to an element in speci®er position. This is because the constraint has the form: if ‡[focus] then left-aligned. Rather, comprehension requires its converse. If a left speci®er position is ®lled, the ®ller is interpreted as focalized: if left-aligned, then ‡[focus]. A Faith violation results from a constraint motivated by requirements of comprehension. Candidate (c) is eliminated. If this analysis is on the right track, then fronting both V and an argument should be ungrammatical in the presence of an auxiliary because a fatal violation of Faith

Masked Second-Position E¨ects, Linearization of Functional Features

257

arises. Indeed, various scholars have noticed the ungrammaticality of structures like (24). (24) Breton *Al levr lennet en deus Yann. the book read 3 have Yann `Yann has read the book.' When the second position of [F] is in peril, due to the fact that there is no aspectual auxiliary to carry [F], Breton resorts to an expletive auxiliary, as in (23a). In OT terms, expletive elements are costly because they violate FullInterpretation, which penalizes elements whose lexical conceptual structure is unparsed (Grimshaw 1997; Grimshaw and Samek-Lodovici 1998).18 The fact that Breton opts for an expletive auxiliary rather than fronting an argument shows that Faith must outrank FullInterpretation (FI in tableau T9.2 for short).19 The relative ranking of FullInterpretation and Edgemost(F) and *t is indeterminate. Because no argument is marked with a [focus] feature in tableau T9.2, candidates (c) and (d) fare poorly with respect to Faith. Note that candidate (c)'s violation of Faith is crucial to the outcome of the competition: (c) better satis®es all other constraints. Yet the subject is in the top speci®er position leading to its interpretation as focalized, in violation of input-output faithfulness. Candidate (d) is in fact harmonically bound by candidate (a) ((d) violates every constraint violated by (a) plus one more constraint), hence it can never win under any ranking (Prince and Smolensky 1993). Candidate (b) is eliminated because the ®nite verb is in initial position. Expletive ober `do' (candidate (a)) appears as a last resort, to prevent a fatal violation of Noninitial(F). This is, formally, very similar to do-support in English, though the violation that expletive do obviates is an ObligatoryHeads (ObHd ) violation (see Grimshaw 1997 for details). In fact, Breton further uses expletive ober in the context of VP topicalization, arguably to obviate an ObHd violation. Tableau T9.2 Breton expletive auxiliary Input: V(S, O); [F]

Faith NI(F)

f a. [V 0 Vi [V 0 explF [VP S [V 0 ti O]]]] b. [V 0 VFi [VP S [V 0 ti O]]]

FI

E(F)

*t

Y

Y

*

*

*!

c. [VP S [V 0 VF O]]

*!

d. [VP Sj [V 0 explF [VP tj [V 0 V O]]]]

*!

258

GeÂraldine Legendre

(25) Breton Lenn al levr a ra Yann. read the book part do-3 Yann `Read the book, Yann does.' Assuming that the constituent lenn a levr has been topicalized to the speci®er position of an extended projection of V (under compulsion from a dominating appropriate information structure constraint), there is a head that must be ®lled, otherwise ObHd is violated. Expletive ober comes to the rescue. The present analysis of F2 makes the prediction that the focalization of an argument and an expletive auxiliary are incompatible with each other. Thus, the present analysis automatically accounts for the ungrammaticality of (26), noted in Borsley 1990 and Borsley, Rivero, and Stephens 1996. (26) Breton a. *Al levr a ra lenn Anna. the book part do-3 read Anna b. *Anna a ra lenn al levr. Consider why. If neither subject nor object is marked with the feature [focus] in the input, both (26a) and (26b) incur fatal violations of Faith (in addition to violations of FullInterpretation). In contrast, the optimal counterpart (23a) incurs only a violation of FullInterpretation. If the object O contains the feature [focus] in the input, O must then be fronted to satisfy AlignFocusleft. O, in turn, automatically provides a shield for [F], which can then satisfy Noninitial(F). There is absolutely no need for an expletive element. The result is a simple OVS structure (23c). When S has the feature [focus] in the input, S can satisfy AlignFocusleft without movement in lower SpecVP. There again, [F] automatically satis®es Noninitial(F) and yields an SVO structure (23b). On either reading, (26a) and (26b) are suboptimal candidates. The present analysis of F2 further predicts that Breton will exploit other strategies whenever the input makes them available. One of them involves wh-phrases. (27) Breton Piv en deus Yann gwelet dec'h? who 3 have Yann seen yesterday `Who did Yann see yesterday?' The presence of a wh-phrase in the speci®er of the phrase headed by the auxiliary o¨sets the need to move V to get [F] in second position, and thereby avoids an additional violation of *t. (I am assuming that piv `who' moves from its underlying VP-internal object position to the available top speci®er position to satisfy OpSpec (Grimshaw 1997.)

Masked Second-Position E¨ects, Linearization of Functional Features

259

Another avenue is available when the input contains a [neg] feature. According to the literature on Breton, the negative particle ne is a clitic. This means that it is realized at PF. Yet it prevents [F] from violating Noninitial(F). This is evidence that the second-position requirement on the ®nite verb is a PF requirement. The same pattern holds in South Slavic (Legendre 1998b, 1999a, forthcoming b). Adding an expletive auxiliary and, as a result, a violation of FullInterpretation is simply unnecessary. (28) Breton Ne lenn ket Anna al levr. neg read-3 neg Anna the book `Anna does not read the book.' The negative adverb ket is assumed to be adjoined to VP in tableau T9.3 (p. 260). Violations of *t are incurred by all candidates, hence they can be ignored. The fact that the optimal candidate (a) violates Egdemost(F) allows its relative ranking to be recovered: FullInterpretation g Edgemost(F). Otherwise candidate (b) or (c) would win. Note, however, that the two Edgemost constraints are unranked one relative to the other. Either ranking will favor candidate (a). Interestingly, Press (1986:128) notes in his descriptive grammar that expletive ober can only be used emphatically in the negative, yielding grammatical structures like candidate (c). That is, in a negative context ober has the same (emphatic) function English do has in a½rmative contexts: he did lie (vs. he lied ). The present analysis o¨ers a simple and straightforward explanation for this. The negative particle ne is su½cient to shield [F]. Thus, candidate (c)Ðwith the ®nite expletive auxiliary in third positionÐis the optimal candidate for a distinct input, presumably one in which emphasis is placed on negation. Leaving the details to be worked out, whatever information structure constraint applies when emphasis is placed on negation overrides FullInterpretation. As a consequence, the ®nite expletive auxiliary surfaces in third position and the F2 e¨ect is masked. As noted early in the chapter, masking of F2 also occurs in the presence of clitic pronouns. The examples are repeated here for convenience. (29) Breton a. E c'halvet en deus Yann. him called 3 have Yann `Yann has called him.' b. *E en deus c'halvet Yann. My proposal is this. Both [acc] and [F] want to be Edgemost in V 0 . If there is only one V 0 projection above the lexical projection of V, that of aux, only one feature can be satis®ed (as in candidate (b) in tableau T9.4). Breton o¨ers another option, namely

b. [V 0 acc auxFi [VP S [V 0 V O]]]

f a. [V 0 acc Vi [V 0 auxF [VP S [V 0 ti O]]]]

Input: V(S, O); [F] [acc] [perf ]

Tableau T9.4 Masked F2 e¨ects

d. [V 0 VFi neg [VP neg [VP S [V 0 ti O]]]]

E(acc)

NI(F)

*!

E(F)

*!

c. [V 0 Vi [V 0 neg explF [VP neg [VP S [V 0 ti O]]]]

FI

*!

*!

NI(F)

b. [V 0 neg Vi [V 0 explF [VP neg [VP S [V 0 ti O]]]]

f a. [V 0 neg VFi [VP neg [VP S [V 0 ti O]]]]

Input: V(S, O); [F] [neg]

Tableau T9.3 Breton negVSO

*

Y

NI(acc)

*

E(neg)

*

Y

Y

*t

E(F)

*

Y

NI(neg)

*

*

*

Y

*t

260 GeÂraldine Legendre

Masked Second-Position E¨ects, Linearization of Functional Features

261

fronting the past participle, which it routinely does to satisfy Noninitial(F). Fronting the past participle incurs a *t violation but it provides an additional V 0 . In return for this, both Edgemost constraints can be simultaneously satis®ed. Each feature is properly aligned, albeit in di¨erent V 0 projections.20 In this analysis, no con¯ict arises because the building up of structure is relatively ``cheap'' in Breton. Economy of structure (MinProj; Grimshaw 1993) and *t are subordinated to Edgemost constraints. Thus, Edgemost(X) can be satis®ed in separate projections. But not all features are ``equal,'' hence [F] cannot exchange its place with [acc]. If it did, it would fatally violate Noninitial(F). The consequence is that the ®nite auxiliary is relegated further away from the clausal edge, in third position. Note that candidate (a) in tableau T9.4 violates the HeadMovement constraint (HMC, Travis 1984)ÐV is fronted across an aux head. In fact, the HMC is also violated by optimal candidates in tableaux T9.1 and T9.2. This is not surprising since, from an OT perspective, the HMC is violable like any other constraint (see also Vikner, chap. 14, this volume). In Breton, it must be subordinated to Edgemost constraints. For present purposes, we may assume that the HMC ranks as low as *t. Note that the competition in tableau T9.4 does not eliminate a third candidate, (c), in which the ®nite auxiliary and V exchange positions. If one assumes base generation of aux in the lower verbal projection and V in the higher one, (c) in fact fares better than (a), since (c) only violates Noninitial(acc). The reason such a candidate is suboptimal is arguably Case, if Nominative Case assignment in VSO languages relies on government from I to the subject in SpecVP (McCloskey 1991). The formal analysis in tableau T9.4 makes a prediction. Clitic pronouns should be ungrammatical in initial position just in case there is only one available verbal projection above the one containing the overt subject. As noted in Borsley, Rivero, and Stephens 1996:62, simple tense structures like (30a) and (30b) are indeed ungrammatical. (30) Breton a. *E c'halvas Yann. him called-3 Yann `Yann called him.' b. *O gwelis. them saw-1 `I saw them.' Note that the ®nite verb is in second position in (30), yet the result is ungrammatical. In the present analysis, the problem is that only one feature, [acc], is able to satisfy Edgemost, and that is not good enough. A further prediction is that, all other things being equal, (30a) and (30b) should lose to counterparts with a ®nite expletive auxiliary. The grammatical sequence

262

GeÂraldine Legendre

should be: accusative, pro-V-®nite ober (subject). I have not been able to check this prediction with a native speaker; Borsley, Rivero, and Stephens do not provide the grammatical counterparts to (30a) and (30b). In an analysis relying on inviolable principles, the explanation for the contrast between (29) and (30) must rely on an added stipulation. For example, Borsley, Rivero, and Stephens (1996:62) resort to stipulating that ``the combination of a clitic and a ®nite verb counts as a ®nite verb.'' This solves the problem in (30) but it immediately creates another one. As illustrated in (31), the noninitial requirement on ®nite verbs in Breton is not absolute. The data is from Borsley, Rivero, and Stephens 1996. (31) Breton a. EmanÄ Yann war an hent. is Yann on the road `Yann is on the road.' b. EmanÄ Anna o lenn al levr. is Anna prog read the book `Anna is reading the book.' In (31a), ®nite emanÄ `be' is used as a copula. In (31b), emanÄ is used to express the progressive aspect, in conjunction with the particle o. Breton shares with English (and other languages) the fact that the progressive aspect is expressed by means of a discontinuous relation between two morphemes, which by themselves may be used for other purposesÐfor example, copula be and gerund -ing. Both are necessary to express the progressive. How to best capture the discontinuous relation is beyond the scope of the present chapter. What exact feature label to give to a multipurpose auxiliary is not obvious either. (I will provisionally call it [aspect] or [modality], or [copula], depending on its use.) Instead, I focus here on the fact that once an auxiliary has been coopted for a particular function, it has to be positioned. The other morphemeÐfor instance, progressive o in BretonÐis subject to its own alignment constraint, which I will ignore here. In our terms, the [aspect] feature's preference is clearly one for Edgemost(asp) over Noninitial(F) in Breton. The ®nite aux/copula emanÄ violates Noninitial(F) in order to satisfy the Edgemost requirement on the other feature it carries, [aspect]. It is not the case that Tense is licensed in two di¨erent ways in Breton, one for emanÄ and one for all other verbs and auxiliaries, as proposed in Borsley, Rivero, and Stephens 1996:63. Rather, the two features carried by emanÄ (aspect, F) con¯ict with one another. Alignment of [aspect] is dominant and thus dictates the outcome, masking the F2 e¨ect in Breton. Of course, this does not imply that emanÄ must always surface in clause-initial position. Whether it does or not depends on constraint interaction. Thus under VP topicalization triggered by a high-ranking information structure constraint, emanÄ

Masked Second-Position E¨ects, Linearization of Functional Features

263

surfaces as expected following the topicalized constituent and the negative particle, if one is available. The example is from Borsley, Rivero, and Stephens 1996:56. (32) Breton O lenn al levr n' emanÄ ket Yann. prog read the book neg is neg Yann `Reading the book, Yann isn't.' To sum up, I have argued that Breton exploits a number of strategies to achieve linearization of [F] in second position, including V fronting or the use of an expletive auxiliary. While the conclusion that Breton is an F2 language is in accord with previous claims (Ortiz de Urbina 1994; Schafer 1995), both the overall analysis of F2 and that of expletive ober are quite di¨erent.21 First, the analysis is grounded in economy of structure and movement. Structure is built up only as much as is needed to meet the demands of high-ranked constraints. Second, the analysis relies on OT's unique input-output faithfulness to explain the relationship between word order and information structure status. That relationship is completely ignored in the generative literature. Third, there is an expletive auxiliary (ober) that comes into play as a last resort to save structures that otherwise would fatally violate di¨erent constraints in di¨erent contexts. Fourth and last, constraints are violable. For example, FullInterpretation is violated by the optimal candidate in tableau T9.2, while violating it is fatal to candidates with expletive ober in tableau T9.3. This fundamental property of OT, by itself, explains the grammaticality of expletive ober in some contexts but not others. F2 is masked in Breton to the extent that it is invisible on the surface. Formally, this is the result of two types of constraint interactions. First, F2 is masked whenever some information structure constraint overrides the constraint interaction that yields second position (e.g., VP topicalization). Second, F2 is masked when several features are realized as separate phrasal a½xes. Under the Breton constraint ranking, no con¯ict arises. Edgemost(X) can be satis®ed in separate projections. Additional structure is built up with the consequence that the ®nite auxiliary is relegated further away from the clausal edge. Thus, F2 can be masked by the relatively ``cheap'' building up of structure in Breton. When the constraint Noninitial(F) is active (i.e., it eliminates suboptimal candidates) as in tableaux T9.1 to T9.3, the unmarked pattern, F2, emerges from the competition. This is an instance of the Emergence of the Unmarked (McCarthy and Prince 1994). 9.3.4 Macedonian F2 E¨ects We ®nally turn to Macedonian and a di¨erent constraint interaction underlying masked F2 e¨ects. I propose that, in Macedonian, masked F2 e¨ects derive from

264

GeÂraldine Legendre

[F] competing with other features for the same (second) position. In other words, Macedonian is both an F2 language like Breton and a second-position clitic language like related South Slavic languages, Bulgarian and Serbo-Croatian. To start, consider the basic empirical facts again. On the one hand, pronominal clitics precede ®nite lexical verbs but follow non®nite ones. This is shown in (33). (33) Macedonian a. Ti go dade. you-dat it-acc gave-3 `(He) gave it to you.' b. Davaijki mi ja. give-ger me-dat it-acc `Giving it to me . . .' On the other hand, participles display two patterns. Like non®nite gerunds, verbal and predicate adjectives (as well as predicate nouns) precede auxiliary and pronominal clitics, as (34a) and (34b) show. Non®nite l-participles used to report unwitnessed events, however, follow, rather than precede, the same set of clitics, as in (34c).22 (34) Macedonian a. Dojden sum. come be-1 `(I) have come.' b. Mil si mi. dear be-2 me-dat `(You) are dear to me.' c. Sum ti go dal. be-1 you-dat it-acc given-evid `(I) have (supposedly) given it to you.' The idea is this. On the one hand, the ®nite verbal category seeks a noninitial position; see (33a) as well as (34a) and (34b). Translated into our terms, this follows from the ranking: Noninitial(F) g Edgemost(F). On the other hand, pronominal clitics also seek a noninitial position, as shown in (33b) and (34c). This suggests the additional ranking: Noninitial(acc, dat) g Edgemost(acc, dat). The outcome is of course a competition for noninitial position. If present in the input, [F] prevails, as shown in (33a); hence [F] is realized on the verb in noninitial position, while object clitics violate Noninitial(F). This reveals the relative ranking of the constraints pertaining to [F] and those pertaining to other features: Noninitial(F) g Noninitial (acc, dat). Finite verbs appear after clitics because the Noninitial constraint on [F] outweighs that on [acc], [dat].

Masked Second-Position E¨ects, Linearization of Functional Features

265

Tableau T9.5 Macedonian ®nite verbs Input: V(S, O); [F] [dat] [acc]

NI(F)

f a. [V 0 dat acc VF ] b. [V 0 VF dat acc] c. [V 0 dat VF acc]

NI(case)

E(dat)

Y

*!

* *

E(acc)

E(F)

Y

YY

** **!

*

The formal competition underlying (33) is displayed in tableau T9.5. In the interest of space, [dat] and [acc] are subsumed under [case] in Noninitial(case). The ®nite verb ends up in third position if two pronominal clitics are present because Edgemost(dat, acc) favors the two features [acc] and [dat] over [F]. That is, the relative ranking of Edgemost constraints favors candidate (a) over candidate (c). The relative position of the two pronominal clitics is determined by the dominant Edgemost(dat). A number of predictions immediately follow. One is that, in the absence of the ®nite feature [F], object clitics cluster in noninitial position. That is, when Noninitial(F) is vacuously satis®ed, encapsulated Noninitial(case) dominates, ruling out the suboptimal candidate that contains domain-initial clitics. The ranking Noninitial(case) g Edgemost(case), in turn, ensures second position, for example, for the feature [accusative] in the presentative construction with clitic doubling in (35). (35) Macedonian Ene go cÏovekot. here him-acc man `Here is the man.' The Macedonian future is a periphrastic construction in which the future auxiliary is an invariable clitic. The lexical verb, on the other hand, appears in the perfective present and thereby instantiates [F]. In the present analysis, the future auxiliary is predicted to precede the ®nite verb. Again, this is because given the choice between placing a ®nite verb or a clitic in noninitial position, Macedonian opts for the ®rst alternative. This prediction is correct, as shown in (36). (36) Macedonian KÂe dojdam. will come-1 `(I) will come.' Thus the rankingÐNoninitial(F) g Noninitial(case) g Edgemost(case) g Edgemost(F)Ðentails procliticization with ®nite verbs and encliticization with non®nite verbs, the core pattern of Macedonian identi®ed in Joseph 1983 and Tomic 1996.

266

GeÂraldine Legendre

Further predictions pertain to the intermixing of Edgemost constraints. In particular, the present analysis of [F] predicts that a ®nite verb (i.e., when [F] is realized on a head) can break up a clitic cluster if Edgemost(F) is ranked among other Edgemost constraints. Such patterns are actually found in Macedonian imperatives: [negVimp,F -dat-acc]. (37) Macedonian Ne davajte mu go! neg give-imp dat acc `Don't give it to him.' Ne is a clitic (Tomic 1996). I have argued elsewhere that imperatives involve two features, [imp] and [F]. [F] may or may not be parsed in a given language, giving rise to non®nite imperatives. The Balkan imperatives display agreement with their understood subject. Hence they are ®nite, as discussed in Legendre (1998b, 1999a). Dative and accusative clitics are inseparable in Macedonian. The present analysis, however, predicts that this is not a universal pattern. Reranking of Edgemost constraints, in particular, predicts that pronominal clitics can be split by [F]. The prediction is correct. For example, in Franco-ProvencËal dialects of Romance, object clitics (i.e., accusative, dative, and partitive) can be split and separated by a ®nite auxiliary or a complex verb. Examples (38a) and (38b) are from Olszyna-Marzys 1964:48, while example (38c)Ðcited in Kayne 1991:661Ðis originally from Chenal 1986. I have retained the quasi-phonetic spelling of the authors. (38) a. Si mots ei va se metr kontr a dzuta. this ¯y 3-dat fut-3sg re¯-acc put against the cheek `This ¯y will land on his cheek.' b. Me chei pa PnchwenaP. re¯-dat be-1-sg not part-remember `I did not remember (of ) it.' c. T 0 an teÈ deut- lo? 2-dat have they said 3-acc `Have they told it to you?' We return to our main interest, deviations from the core pattern in MacedonianÐ predicative constructions and the two past participle patterns, repeated in (39). (39) Macedonian a. Dojden sum. come be-1 `(I) have come.' b. Mil si mi. dear be-2 me-dat `(You) are dear to me.'

Masked Second-Position E¨ects, Linearization of Functional Features

267

c. Sum ti go dal. be-1 you-dat it-acc given-evid `(I) have (supposedly) given it to you.' Despite the fact that all participles are non®nite (see note 22), predicative constructions and n/t-participles show encliticization, while l-participles show procliticization. The basic idea is the following. First, note that clitic auxiliaries like si, sum are special in the sense that they instantiate two separate features, [aspect/copula] and [F]. Now suppose that [aspect/copula] is basically a domain-initial clitic in an otherwise second-position clitic language. In terms of constraint ranking, this means that Edgemost(asp/cop) outranks Noninitial(asp/cop). We have a con¯ict: on the one hand, sum wants to be in domain-initial position because it instantiates [asp]; on the other, sum wants to be in second position because it instantiates [F] and [F] is subject to Noninitial(F) g Noninitial(asp/cop). This con¯ict can be resolved by having Edgemost(asp/cop) tie with Noninitial(F). As a consequence, Edgemost(asp/cop) and Noninitial(F) violations cancel out and the optimal candidate is determined by lower-ranked constraints. (See my introductory chapter in this volume for further discussion of ties.) In the case of predicate adjectives and -n/-t participles, the next-lower constraints on the hierarchy are Noninitial(asp/cop) and Noninitial(dat, cop), unranked with respect to one another, which preclude domain-initial position. There is one way and one way only to satisfy them: by encliticization. The competition is displayed in tableau T9.6 (p. 268) for predicate adjectives. Note that candidate (e) (reversing the relative order of [cop] and [dat] from the optimal candidate (a)) is eliminated by two violations of gradient Edgemost(cop). The claim that the F2 e¨ect in Macedonian is purely positional receives additional support from the fact that other elements besides the past participle or the predicative adjective can serve as hosts of clitics and [F]. Emphatic subject pronouns can, and so can modi®ers that always precede the element they modify (Legendre 1998b). (40) Macedonian a. Ti si mi mil. you be-2 me-dat dear `You are dear to me.' b. Mnogu si mi mil. very be-2 me-dat dear `(You) are very dear to me.' Returning to l-participles (39c), recall that they express evidential modality in conjunction with the clitic auxiliary be. I will label the aux as encoding [modality] and the participle as bearing the feature [evidential], interpreted here as subject to the same alignment constraints as other features. The formal competition underlying (39c) is depicted in tableau T9.7. Note that all elements carry at least one feature,

*

*

e. [V 0 modF dat Vevid acc]

f. [V 0 dat modF acc Vevid ]

*

*

c. [V 0 Vevid modF dat acc]

Y

*

*

*!

*

*!

*!

*

**!

**!

Y

*

*!

*

Y

NI (md)

NI(dat)

E (dat)

NI(cop)

NI (case)

NI(F)

NI (F)

d. [V 0 modF Vevid dat acc]

**!

b. [V 0 dat acc modF Vevid ]

f a. [V 0 modF dat acc Vevid ]

Input: V(S,O); [F] [mod] [dat] [acc] [evid]

E (md)

**!

f. [V 0 dat pred copF ]

Tableau T9.7 Macedonian l-participles

**!

*

Y

E(cop)

e. [V 0 pred dat copF ]

d. [V 0 copF pred dat]

c. [V 0 dat copF pred]

b. [V 0 copF dat pred]

f a. [V 0 pred copF dat]

Input: V(S, IO); [F], [cop], [dat]

Tableau T9.6 Macedonian predicate adjectives

**

***!

***

***

*

YY

E (acc)

*

**

*

YY

E(dat)

*

*

**

E (F)

**

**

*

Y

E(F)

***

**

*

***

YYY

E (ev)

268 GeÂraldine Legendre

Masked Second-Position E¨ects, Linearization of Functional Features

269

hence a Noninitial constraint is violated by all candidates. Edgemost(mod) and Noninitial(F) tie, hence single violations of either one count alike. The decision falls to Edgemost constraints. Candidate (c), for example, shows the same pattern of violations of the three highest-ranked constraints (given the equal ranking of Edgemost(mod) and Noninitial(F)). But fronting of the participle entails one additional violation of Edgemost(dat), compared to not fronting the participle (candidate (a)). Hence candidate (a) is optimal. Breaking the cluster of clitics, as in candidates (d), (e), and (f ), has a similar e¨ect. It entails additional violations of Edgemost(acc, dat), compared to the optimal candidate (a) or even a fatal violation of Noninitial(case) (as in candidate (f )). To sum up, Macedonian exempli®es a di¨erent type of masked V2 e¨ect because [F] is borne by a clitic auxiliary. In the absence of other e¨ects, the competition is just among features that compete with one another for linearization in second position. Once the nature of the competition is understood in these terms, the complex Macedonian pattern of pro- versus encliticization is a natural consequence of optimizing their con¯icting alignment requirements at PF. Viewed from the perspective of OT, the basic and elegant generalization proposed in Joseph 1983 and Tomic 1996Ð procliticization with ®nite verbs and encliticization with non®nite verbsÐcan be retained without its analytical complications because the deviations are the direct consequence of the optimization process. 9.4

Conclusion

The overall claim of this chapter is that the positioning of functional elements like [F] is not a deep syntactic phenomenon, contrary to the prevailing view in generative circles. Rather, it is a fairly super®cial process involving a lot of jostling at PF. [F], like other features commonly realized as clitics, is universally positioned by violable alignment constraints. Depending on the languageÐin particular, whether [F] is realized as a clitic or notÐand the input, the competition may be strictly among PF constraints (as in the Macedonian examples), or it may involve the interaction of alignment constraints with syntactic and information structure constraints (as in the Breton examples). Either kind of interaction may, in some contexts, mask the basic second-position requirement that the two languages impose on their ®nite verbal elements. F2 is masked by other e¨ects derived from higher-ranked constraints, including some pertaining to information structure (Breton) or to competing features (Macedonian and Breton). The very nature of constraint interaction in OT has been shown to allow such e¨ects to be recovered when higher-ranked constraints are controlled for. Under these conditions, the unmarked pattern emerges, namely F2. [F] is oblivious to the status distinction among clitic auxiliaries and nonclitic verbal elements (including lexical verbs). Traditionally, this has been interpreted as evidence

270

GeÂraldine Legendre

that all verbal categories have the same status of syntactic heads. I have argued against this view, on the basis that it does not explain the crosslinguistic syntactic inertness of clitic auxiliaries. The uni®cation can therefore not be in syntactic terms. I have proposed that the uni®cation takes place at PF via [F], which, in some cases in some languages, may drive the syntax of verbal heads. The present analysis achieves a further uni®cation of all second-position phenomena, whether they pertain to clitics or to verbal categoriesÐand thereby bolsters an earlier claim made in Anderson (1993, forthcoming). In particular, languages are not partitioned into V2 languages, LHM languages, and so on. In the present analysis, they are all subject to the same constraints. What is observable on the surface is the product of constraint interaction, where constraints are construed as violable. The present proposal is a radical departure from approaches to in¯ectional morphology that rely on movement in the syntax, either via head adjunction (e.g., Baker 1988; Pollock 1989) or via checking of strong features (Chomsky 1995). As is well known, the head-adjunction approach has led to a proliferation of abstract functional projections that cannot be assumed to be present in all languages, opening the door to further parametrization. Checking Theory, at least for Chomsky (1995), relies on abstract rather than morphological features, raising some serious questions about the explanatory value of formal features, once they are divorced from any morphological grounding. There is a clear alternative in an optimizing system where structure does not come for free. The present chapter shows that the latter approach is not only viable but empirically well motivated; it o¨ers a promising avenue for addressing the nature of intra-Germanic complications that are a challenge to the traditional V-to-C analysis of F2 e¨ects.23 It also ®nds a natural place in a general alignment-based theory of surface word order, aspects of which are explored elsewhere in this volume. Notes This chapter has bene®ted from discussions with Paul Smolensky, comments by Luigi Burzio, Jane Grimshaw, and Sten Vikner, and elicitations from Olga TomicÂ. I thank and hereby absolve all from any remaining errors. 1. Prosodic Inversion moves a clitic (generated in the leftmost position in its syntactic domain) immediately to the right of its prosodic host at PF. The host is assumed to be a prosodic word, since the original proposal was made for (western) Serbo-Croatian clitics, which appear in second position after the ®rst prosodic word. Morphological Merger is a similar operation, but movement is to the right of a lexical head (rather than a prosodic word). Use of such mechanisms largely amounts to undoing at PF what was previously done in the syntactic component. See Legendre (1996) for further discussion. 2. The existence of a Noninitial constraint ®nds corroboration in one of the better-known empirical generalizations on clitics, called the Tobler-Mussa®a Law in Old Romance studies (see, for example, BenincaÁ 1995).

Masked Second-Position E¨ects, Linearization of Functional Features

271

3. Space precludes discussion of SOV languages like Basque, which show strikingly similar e¨ects. See Legendre (forthcoming a) for details. 4. All the Breton data reported in this chapter is drawn from the generative literature (Stump 1989; Borsley 1990; Ortiz de Urbina 1994; Schafer 1995; Borsley, Rivero, and Stephens 1996) and a descriptive grammar (Press 1986). The Macedonian data comes from Tomic 1996 and supplementary elicitations from Olga TomicÂ. 5. Elsewhere, this clitic functions as an object clitic. Historically, the root expresses `be': en deus `x is to him' (Stump 1989). 6. Some complications pertaining to ®niteness in Breton need to be put to rest before proceeding with the main issue. In Breton, there is near complementarity between the presence of subject agreement in person and number on verbal forms and null subjects. Typically, verbs agree with null subjects, but a default third-person singular form called analytic in the literature is found with overt subjects (Stump 1989): (1) LevriouÁ a lennont `(They) read books (lit. books particle read-3pl)', (2) LevriouÁ a lenn (*lennont) ar vugale `The children read books (lit. books particle read-3sg the children)'. Complementarity of subject agreement and null subjects is not absolute, however, because the past tense auxiliary endevout `have' does not follow this distribution. Despite the fact that the ®nite status of these auxiliary forms en deus `3sg masc', he deus `3sg fem', o deus `3pl', and so on is controversialÐsee, for example, Borsley and Stephens 1989 versus Stump 1989Ðthere is, in my opinion, substantial theory-external evidence for their ®nite status. First, the forms clearly di¨er with respect to the clitic preceding the root, depending on the person/number of the subject of predication. Second, these forms occur in null-subject contexts, which according to Stump (1989), require an agreeing form. Henceforth, I will use forms of endevout to document the position of ®nite verbs and leave aside the more complex issue of Breton agreement, which does not a¨ect the points made here. Further complications arise with preverbal subjects and negation, which we may ignore here altogether (see Stump 1989 and Borsley and Stephens 1989 for discussion). 7. The Macedonian future (clitic) auxiliary is used in the example below because the auxiliary be is restricted to reported modality in the speech of my informant, hence incompatible with questions (Olga TomicÂ, personal communication). The fact that the future auxiliary is invariable is irrelevant (other ®nite clitic auxiliaries behave like (8a)). 8. Despite super®cial similarity with the French Stylistic Inversion construction OuÁ est parti Pierre? `Where has Peter gone?', the pattern in (8) is not a case of Stylistic Inversion. As is well known since Kayne 1972, the French construction is triggered, for example, by a fronted whphrase. Note also that Stylistic Inversion is ungrammatical in yes-no questions, whereas the absence of subject-auxiliary inversion is found in both wh- and yes-no questions. See Legendre 1999b for an OT analysis of Stylistic Inversion. 9. Unfortunately, I do not have corresponding data in Breton. 10. The fact that Macedonian clitic pronouns and auxiliaries can occur in initial position has traditionally been interpreted as evidence that they are not second-position clitics (Joseph 1983; Tomic 1996). But they only do so in some contexts. We have already seen that the auxiliary be alternates between initial and second position. Pronominal clitics (underscored in the examples below) do the same. They ``cluster'' in second position in some contexts, including l-participles and non®nite verbs: (1) sum ti go kazal `I have told it to you (lit. have to-you it told)', (2) davaijki mi ja `giving it to me (lit. giving to-me it)'. See Legendre 1998b for further discussion.

272

GeÂraldine Legendre

11. Walpiri (Central Australia) provides additional evidence for the present analysis of auxiliary clitics as PF objects. Walpiri is a free word order language, meaning that the lexical categories S, V, O may appear in any order in relation to one another (Hale 1973, 1992). The only requirement is that an auxiliary encoding tenseÐka in the example belowÐappear in second position. Thus, example (i) exempli®es only one of six possible well-formed sentences incorporating the same lexical elements. (i) Karnta-ngku ka yarla karla-mi. woman-ergative pres yam dig-nonpast `The/a woman is digging yams.' Besides encoding present tense, ka forms the base to which are a½xed morphemes encoding person, number (and case, to some extent) of arguments, as well as complementizers. The base may be phonologically nullÐin which case only person markers (and complementizers) are realized. Note that the second-position requirement applies only if the base of the auxiliary is monosyllabic or if it is phonologically null. This restriction alone is strongly suggestive of a PF phenomenon. 12. The fact that pronominal clitics satisfy subcategorization properties of verbs is not necessarily evidence that they are syntactically independent items, contra the view held since Kayne 1975. The e¨ect could be derivative, arising from the absence of clitic doubling in a given language. A theory that takes clitic doubling to be the default and derives its absence in some languages from the interaction of the need to mark case versus economy is sketched out in Legendre 1998a. Note further that ethical dative clitics share the same positional restrictions as other pronominal clitics, despite the fact that they do not satisfy subcategorization frames. See Sportiche 1996 for further discussion. 13. As extensively discussed in Anderson's (1992) book and subsequent work, there is a strong universal tendency for clitics to behave like word-level a½xes despite the fact that clitics are less choosy with respect to their host. 14. Alternatively, the domain of Edgemost(F) is VP and (16b) results from a dominant constraint on speci®ers relevant to the positioning of subjects, possibly SpecLeft (Grimshaw 1997). This implies that Edgemost(F) is violated in (16b). The advantage of construing the domain as VP (as pointed out by Jane Grimshaw, personal communication) is that it exploits the basic competition mechanism of OT whenever an overt subject surfaces. Yet construing the domain as VP makes a wrong prediction. Constraint reranking predicts that a language is possible in which the subject may split clitics. Where clitic splitting occurs, it only involves verbal elements, as far as I know. See (38) for some examples and discussion. 15. Ortiz de Urbina (1994) also notes that the clause-initial ban on ®nite verbal forms does not extend to embedded clauses. In other words, Breton is an asymmetric F2 language. (i) Lavaret en deus [ he deus desket Anna he c'henteliouÂ]. said 3 have 3 have learned Anna her lessons `(He) has said that Anna has learned her lessons.' Why is that? Note that the subordinate clause contains no clause-initial complementizer. I tentatively propose that (a) the subordinate clause is not intonationally marked as a separate Intonational Phrase, and (b) this state of a¨airs is related to the absence of an overt complementizer. (I was unable to verify this with native speakers.) Embedded [F] automatically satis®es Noninitial(F) if the whole clause constitutes one Intonational Phrase domain. Past

Masked Second-Position E¨ects, Linearization of Functional Features

273

Participle fronting is suboptimal because it entails unnecessary additional Edgemost(F) violations, in violation of economy. Germanic languages complicate the picture but, to some extent, provide supportive evidence. German has embedded F2 in the absence of a complementizer: Er sagt, die Kinder haben diesen Film gesehen `lit. he says the children have this ®lm seen'. As con®rmed by Sten Vikner (personal communication), there is an intonational break right after the main verb. Thus, embedded [F] is second in a separate Intonational Phrase, as expected under the present analysis. Moreover, the fact that Icelandic and Yiddish display embedded V2 in the presence of a complementizerÐa problem under the V-to-C analysisÐis consistent with the claim that a complementizer ``signals'' a separate Intonational Phrase. Yet the fact that the Mainland Scandinavian languages display V2 in the presence of a complementizer only in the presence of certain matrix verbs (Vikner 1995) suggests that additional factors (and dominating constraints) are relevant. 16. Input-Output Faithfulness plays a crucial role in a number of studies, including Legendre et al. (1995), Legendre, Smolensky, and Wilson (1998), Grimshaw (chap. 8, this volume), Bresnan (chap. 5, this volume), as well as Bakovic and Keer (chap. 4, this volume). 17. I am ignoring typological issues relevant to left versus right alignment of focused elements. See both Costa and Samek-Lodovici (chaps. 7 and 11, respectively, this volume) for relevant discussion. 18. What determines whether a language makes use of an expletive auxiliary as opposed to an expletive subject (which could equally ensure second position of the verb) is presumably the relative ranking of constraints pertaining to subjects, SpecIP, and heads. 19. Note that, in the interest of limiting constraints to a more manageable number, the particle a is assumed to be part of expletive ober. It can easily be reinstated with its own Edgemost constraint, a¨ecting the number of Edgemost(F) violated by the expletive auxiliary. This does not a¨ect the outcome in any way. 20. Additional candidates are eliminated by other constraints, provided they outrank Noninitial(acc). For example, placing the pronominal clitic in the speci®er of the lowest VP in a candidate similar to (a) saves a violation of Noninitial(acc), but it entails (assuming a *doubly ®lled Spec constraint) that the subject has to move somewhere and violate either *Adjunction (Legendre 1999b) or Faith. 21. Schafer (1995) adopts the standard analysis of Germanic V2Ðthat is, verb movement from I to C. Note, however, that I to C is string-vacuous in a VSO language. Hence, Schafer only shows that Breton is compatible with the Germanic analysis, not that it must be analyzed as an instance of I-to-C movement. Borsley, Rivero, and Stephens (1996) propose a Long Head Movement (LHM) analysis insteadÐthat is, head movement across another head. Because Borsley, Rivero, and Stephens (1996) also appeal to tense as a motivation behind LHM, their analysis might, at ®rst glance, be characterized as very similar to the present one. In their analysis, languages with F2 e¨ects are characterized as being either LHM or verbsecond languages. Two separate mechanisms of Tense licensing are proposed in LHM languages. Tense may be licensed by a V adjoined to Tense (i.e., in its checking domain) or in the internal domain of C. The latter is proposed for Breton and for Slavic auxiliaries allowing LHM (i.e., clitic auxiliaries). The former (Tense licensing in its checking domain) is proposed for auxiliaries that allow VP topicalization (i.e., nonclitic auxiliaries). Two homophonous (expletive) ober auxiliaries with di¨erent Tense-licensing properties are invoked because, ober, in

274

GeÂraldine Legendre

their terms, allows both LHM and VP topicalization. In the present analysis, there is no such thing as an LHM language since LHM is intimately tied to an analysis of clitic auxiliaries as syntactic heads. This is not to say that the HeadMovement constraint (HMC, Travis 1984) is never violated. It can be, under compulsion of higher-ranked constraints. This is the case in Bulgarian under compulsion of information structure constraints (Legendre 1996) and Breton (when V is fronted across an auxiliary to satisfy Noninitial(F), as in tableau T9.1). There is no need for a separate concept of ``functional auxiliary'' that subsumes clitic auxiliaries on the one hand and other auxiliaries that violate HMC on the other (Rivero 1994; Borsley, Rivero, and Stephens 1996) because the latter is simply the consequence of constraint interaction. Other proposals, like the present one but focusing on Germanic V2 languages, appeal to ®niteness features. These include many di¨erent versions, including Holmberg and Platzack (1995) arguing for a ®niteness operator [‡F] located in C in V2 languages and Tomaselli (1990) favoring tense and agreement in C. Vikner (1995:51±64) o¨ers critiques of both analyses couched in standard Government-Binding terms. 22. Some comments on the ®niteness of l-participles are called for here. Because of the contrast just mentioned, some scholars, in particular Joseph (1983) and Tomic (1996), have proposed the following generalization: Macedonian clitics procliticize to ®nite verbs but encliticitize to non®nite verbs. This leads to treating l-participles as ®nite. Legendre (1998b), however, provides a number of arguments against this conclusion, including the following. First, l-participles, like non®nite -n/-t participles, are in¯ected for gender and number but not for person. Second, they co-occur with an auxiliary that is ®nite (with tense ‡ person/number morphology). Treating l-participles as ®nite would amount to claiming that two verbal elements in a single sentence are ®nite, despite a lack of corresponding morphology for one of them. Finally, l-participles in other South Slavic languages behave exactly like their non®nite participles, including with respect to cliticization. 23. As is well known, V2 is limited to root clauses in some Germanic languages (Dutch, German) but not others (Icelandic, Yiddish). See, for example, Vikner 1995 for a clear presentation of the crosslinguistic facts and issues. The present approach suggests investigating masking e¨ects in asymmetric languages like Dutch and German. References Anderson, Stephen R. 1992. A-Morphous Morphology. Cambridge, England: Cambridge University Press. Anderson, Stephen R. 1993. Wackernagel's Revenge: Clitics, Morphology, and the Syntax of Second Position. Language 69: 68±98. Anderson, Stephen R. 1996. How to Put Your Clitics in Their Place. Linguistic Review 13: 165±191. Anderson, Stephen R. Forthcoming. Towards an Optimal Account of Second Position Phenomena. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Baker, M. 1988. Incorporation: A Theory of Grammatical Function Changing. Chicago: University of Chicago Press. BenincaÁ, Paola. 1995. Complement Clitics in Medieval Romance: The Tobler-Mussa®a Law. In A. Battye and I. Roberts (eds.), Clause Structure and Language Change, 325±344. Oxford: Oxford University Press.

Masked Second-Position E¨ects, Linearization of Functional Features

275

Den Besten, Hans. 1983. On the Interaction of Root Transformations and Lexical Deletive Rules. In W. Abraham (ed.), On the Formal Syntax of Westgermania, 47±131. Amsterdam: Benjamins. Borer, Hagit. 1984. Parametric Syntax: Case Studies in Semitic and Romance Languages. Dordrecht: Foris. Borsley, Robert D. 1990. A GPSG Approach to Breton Word Order. In R. Hendrick (ed.), Syntax and Semantics 23: The Syntax of Modern Celtic Languages, 81±95. New York: Academic Press. Borsley, Robert D., Maria-Luisa Rivero, and Janig Stephens. 1996. Long Head Movement in Breton. In R. D. Borsley and I. Roberts (eds.), The Syntax of Celtic Languages, 53±74. Cambridge, England: Cambridge University Press. Borsley, Robert D., and Janig Stephens. 1989. Agreement and the Position of Subjects. Natural Language and Linguistic Theory 7(3): 407±427. Chenal, AndreÂ. 1986. Le franco-provencËal valdoÃtain. Aoste: Musumeci. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Greenberg, Joseph H. 1966. Some Universals of Grammar with Particular Reference to the Order of Meaningful Elements. In J. H. Greenberg (ed.), Universals of Language, 73±113. Cambridge, Mass.: MIT Press. Grimshaw, Jane. 1991. Extended Projections. Ms., Brandeis University, Waltham, Mass. Grimshaw, Jane. 1993. Minimal Projection, Heads, and Optimality. Ms., Rutgers University, New Brunswick, N.J. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28(3): 373±422. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 193±219. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Hale, Ken, 1973. Person Marking in Walbiri. In S. R. Anderson and P. Kiparsky (eds.), A Festschrift for Morris Halle, 308±344. New York: Holt, Rinehart and Winston. Hale, Ken. 1992. Basic Word Order in Two ``Free Word Order'' Languages. In D. L. Payne (ed.), Pragmatics of Word Order Flexibility, 63±82. Amsterdam: Benjamins. Halpern, Aaron. L. 1995. On the Placement and Morphology of Clitics. Stanford, Calif.: CSLI Publications. Holmberg, Anders, and Christer Platzack. 1995. The Role of In¯ection in Scandinavian Syntax. Oxford: Oxford University Press. Joseph, Brian. 1983. The Synchrony and Diachrony of the Balkan In®nitive. Cambridge, England: Cambridge University Press. Kayne, Richard. 1972. Subject Inversion in French Interrogatives. In J. Casagrande and B. Saciuk (eds.), Generative Studies in Romance Languages, 70±126. Rowley, Mass.: Newbury House. Kayne, Richard. 1975. French Syntax. Cambridge, Mass.: MIT Press. Kayne, R. S. 1991. Romance Clitics, Verb Movement, and PRO. Linguistic Inquiry 22(4): 647± 686.

276

GeÂraldine Legendre

Kiss, Katalin. E. 1998. Identi®cational Focus versus Information Focus. Language 74: 245±273. Klavans, Judith. 1985. The Independence of Syntax and Phonology in Cliticization. Language 61(1): 95±120. Koopman, Hilda, and Dominique Sportiche. 1991. The Position of Subjects. Lingua 85: 211±258. Legendre, GeÂraldine. 1996. Clitics, Verb (Non)-Movement, and Optimality in Bulgarian. Technical Report JHU-CogSci-96-5. Department of Cognitive Science, Johns Hopkins University, Baltimore. Legendre, GeÂraldine. 1998a. PF Cliticization and Complex Inversion in French. Handout of a talk delivered at the Second OT Workshop, University of Stuttgart, November. Legendre, GeÂraldine. 1998b. Second Position Clitics in a V2 Language: Con¯ict Resolution in Macedonian. In J. Austin and A. Lawson (eds.), Proceedings of the 1997 ESCOL Meeting, 139±149. CLC Publications, Cornell University, Ithaca, N.Y. Legendre, GeÂraldine. 1999a. Morphological and Prosodic Alignment at Work: The Case of South Slavic Clitics. In S. J. Blake, E.-S. Kim, and K. N. Shahin (eds.), Proceedings of the West Coast Conference on Formal Linguistics XVII, 436±450. Stanford, Calif.: CSLI Publications. Legendre, GeÂraldine. 1999b. Why French Stylistic Inversion is Optimal. Ms., Johns Hopkins University, Baltimore. Legendre, GeÂraldine. Forthcoming a. Evidence for an OT Conception of a Parallel Interface. Proceedings of NELS 1999. GLSA, University of Massachusetts, Amherst. Legendre, GeÂraldine. Forthcoming b. Morphological and Prosodic Alignment of Bulgarian Clitics. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Legendre, GeÂraldine. Forthcoming c. Positioning Romanian Verbal Clitics at PF: An Optimality-Theoretic Analysis. In Birgit Gerlach and Janet Grijzenhout (eds.), Clitics in Phonology, Morphology, and Syntax. Amsterdam: Benjamins. Legendre, GeÂraldine, P. Smolensky, and C. Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Legendre, GeÂraldine, C. Wilson, P. Smolensky, K. Homer, and W. Raymond. 1995. Optimality and Wh-Extraction. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 607±636. GLSA, University of Massachusetts, Amherst. Marantz, Alec. 1988. Clitics, Morphological Merger, and Mapping. In M. Hammond and M. Noonan (eds.), Theoretical Morphology: Approaches in Modern Linguistics, 253±270. New York: Academic Press. McCarthy, John, and Alan Prince. 1993a. Generalized Alignment. Yearbook of Morphology, 79±153. McCarthy, John, and Alan Prince. 1993b. Prosodic Morphology I: Constraint Interaction and Satisfaction. Technical Report TR-3, Center for Cognitive Science, Rutgers University, New Brunswick, N.J. [Forthcoming, MIT Press, Cambridge, Mass.] Available at http://www.ruccs. rutgers.edu/publicationsreports.html.

Masked Second-Position E¨ects, Linearization of Functional Features

277

McCarthy, John, and Alan Prince. 1994. The Emergence of the Unmarked: Optimality in Prosodic Morphology. In MerceÁ GonzaÁlez (ed.), NELS 24, vol. 2, 333±379. GLSA, University of Massachusetts, Amherst. McCloskey, Jim. 1991. Clause Structure, Ellipsis, and Proper Government in Irish. Lingua 85(2±3): 259±302. Olszyna-Marzys, Zygmunt. 1964. Les pronoms dans les patois du Valais Central. Berne: Editions Francke. Ortiz de Urbina, Jon. 1994. Verb-Initial Patterns in Basque and Breton. Lingua 94: 125±153. Pollock, Jean-Yves. 1989. Verb Movement, Universal Grammar, and the Structure of IP. Linguistic Inquiry 20: 365±424. Press, Ian. 1986. A Grammar of Modern Breton. Berlin: Mouton de Gruyter. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Progovac, Ljiljana. 1996. Clitics in Serbian/Croatian: Comp as the Second Position. In A. L. Halpern and A. M. Zwicky (eds.), Approaching Second: Second Position Clitics and Related Phenomena, 411±428. Stanford, Calif.: CSLI Publications. RadanovicÂ-KocicÂ, Vesna. 1996. The Placement of Serbo-Croatian Clitics: A Prosodic Approach. In A. L. Halpern and A. M. Zwicky (eds.), Approaching Second: Second Position Clitics and Related Phenomena, 429±445. Stanford, Calif.: CSLI Publications. Rivero, Maria-Luisa. 1994. Clause Structure and V-Movement in the Languages of the Balkans. Natural Language and Linguistic Theory 12: 63±120. Samek-Lodovici, Vieri. 1998. Opposite Constraints: Left and Right Focus-Alignment in Kanakuru. Lingua 104(1±2): 111±130. Schafer, Robin. 1995. Negation and Verb Second in Breton. Natural Language and Linguistic Theory 13(1): 135±172. Sportiche, Dominique. 1996. Clitic Constructions. In J. Rooryck and L. Zaring (eds.), Phrase Structure and the Lexicon, 213±276. Dordrecht: Kluwer. Stump, Gregory T. 1989. Further Remarks on Breton Agreement. Natural Language and Linguistic Theory 7(3): 429±471. SunÄer, Margarita. 1988. The Role of Agreement in Clitic-Doubled Constructions. Natural Language and Linguistic Theory 6(3): 391±434. Tomaselli, Anna. 1990. La sintassi del verbo ®nito nelle lingue germaniche. Padua: Unipress. TomicÂ, Olga M. 1996. The Balkan Slavic Clausal Clitics. Natural Language and Linguistic Theory 14: 811±872. Travis, Lisa. 1984. Parameters and E¨ects of Word Order Variation. Doctoral dissertation, MIT, Cambridge, Mass. Vikner, Sten. 1995. Verb Movement and Expletive Subjects in the Germanic Languages. Oxford: Oxford University Press.

Chapter 10 Order Preservation, Parallel Movement, and the Emergence of the Unmarked

Gereon MuÈller

In many cases, there is no ambiguity in movement rule application: the movement operation can target only one designated position and can a¨ect only one designated item. A typical example of this situation is wh-movement in simple questions in English. However, sometimes ambiguity does show up in the application of a movement operation, either because the movement operation may (or must) apply more than once, or because more than one item in the clause can in principle be moved. A curious but nevertheless prevalent feature of many of the latter instances of rule application is that movement must be order preserving. So far, this property of movement operations does not seem to have been explained in a simple and uni®ed way, and it is the main goal of this chapter to give an account of why it holds. The account is based on the general constraint Parallel Movement (henceforth ParMove) in (1), which I will try to justify in what follows. (1) ParMove If a c-commands b at level Ln , then a c-commands b at level Ln‡1 (where a, b are arguments). The basic idea is that c-command relations between arguments must be preserved from one level of representation to the next one. Assuming strict binary branching, c-command relations between arguments are necessarily asymmetric. Assuming furthermore (contra Chomsky 1995) a model of grammar as in Chomsky 1981, 1986 that recognizes three levels of syntactic representation (D-structure, S-structure, and LF), (1) demands that asymmetric c-command relations between arguments at Dstructure must be preserved at S-structure, and that asymmetric c-command relations between arguments at S-structure must be preserved at LF.1 I will show that ParMove permits a straightforward account of order preservation with various instances of ambiguous movement (superiority e¨ects in English, multiple wh-movement in Bulgarian, object shift in Danish and Icelandic, Case-driven NP raising, pronoun fronting in German, and quanti®er raising in German).

280

Gereon MuÈller

Before I turn to an illustration of how ParMove accounts for order-preservation e¨ects with otherwise ambiguous movement rule application, a fundamental problem should be noted that the constraint appears to raise for nonambiguous movement operations a¨ecting objectsÐlike, for example, simple wh-movement or topicalization in languages like English (see (2)).2 (2) a. (I don't know) [CP what2 [TP she1 said t2 ]] b. [CP John2 [TP she1 doesn't really like t2 ]] Both examples clearly violate ParMove because NP2 is asymmetrically c-commanded by NP1 at D-structure but ends up asymmetrically c-commanding NP1 at S-structure. In view of this, various steps could be taken. First, ParMove could be rejected immediately. Second, ParMove could be split up in such a way that it is formulated only for speci®c movement operations, or only for certain kinds of items.3 Since this would mean giving up a uni®ed explanation and potentially missing a generalization, I will opt for a third possibility and assume that ParMove's apparent violability does in fact not disqualify it from being a fundamental constraint that restricts all movement operations. To execute this idea, I will develop an Optimality-Theoretic (OT) approach (cf. Prince and Smolensky 1993) to order-preserving movement that covers both the cases in which (1) seems to be relevant and the cases in which it seems to be irrelevant. The main idea is that Par-Move belongs to the class of faithfulness constraints that demand identity of input and output of a derivation. The constraint has the following properties: it is violable, universal, and typically ranked low. The last property ensures that its e¨ects are often blurred by higher-ranked constraints and can therefore be observed only under favourable conditions. Thus, the analysis involves a prototypical instance of what McCarthy and Prince (1994) call the ``emergence of the unmarked.'' On a more general note, what follows can be viewed as an argument for an OT organization of grammar: (i) Di¨erent constructions exhibit the same property P. (ii) P can be accounted for by a general constraint C, or by several constructionspeci®c constraints c1 , c2 , and so on. (iii) Other things being equal, C is to be preferred over c1 , c2 , and so forth, for reasons of explanatory adequacy. (iv) Since C must be assumed to be violable and ranked, the concepts of violability and ranking must play a role in syntactic theory. 10.1

Background Assumptions

An OT grammar has two subcomponents (see Prince and Smolensky 1993). One part of the grammar (called Gen) more or less conforms to standard assumptions: it contains only inviolable constraints, and its main task is to generate the candidates that compete for well-formedness. As for the constraints that make up Gen, I will assume

Order Preservation, Parallel Movement, Emergence of the Unmarked

281

that they include X-bar theoretic restrictions, constraints on y-assignment, and others. The candidates themselves can be viewed as output representations (such as Grimshaw's (1997) S-structure representations), as n-tuples of representations, or as complete derivations (cf. Chomsky 1995). Here I will adopt the last view.4 The derivations created by Gen are then subjected to a process of ``harmony evaluation'' in the second, OT part of the grammar. This grammatical subcomponent consists of a set of constraints that are universal, violable, and ranked. Depending on how well it satis®es the ranked constraints, a derivation may or may not be optimal (ˆ grammatical) in its candidate set. Optimality can be de®ned as in (3). (3) Optimality A derivation Di is optimal (ˆ grammatical) i¨ there is no derivation Dj in the same candidate set that has a better constraint pro®le. Dj has a better constraint pro®le than Di i¨ there is a constraint Conk such that: a. Dj satis®es Conk better than Di . b. There is no constraint Conl dominating Conk on which Di and Dj di¨er. Another background notion in need of clari®cation is that of a candidate set. Although matters turn out to be a little more complex on closer inspection, it may su½ce for present purposes to adopt the concept of candidate set developed in Chomsky 1995 that is based on identical numerations (i.e., simplifying somewhat, on identity of lexical material). (4) Candidate Set Two derivations Di and Dj are in the same candidate set i¨ they have an identical numeration. Two types of constraints can be distinguished that impose con¯icting requirements on candidates (which must then be resolved by a ranking of the constraints). First, there are faithfulness constraints that ensure that the input and the output of a derivation di¨er as little as possible, and that thereby minimize the e¨ects of syntactic operations. Among these I take to be a constraint that blocks S-structure movement, namely, Stay.5 (5) Stay S-structure movement is not allowed. Another faithfulness constraint is ParMove; like Stay, it minimizes the e¨ects of movement on a given input. In contrast, a second class of markedness constraints demand that the input and the output of a derivation di¨er; these constraints ensure that movement operations occur in syntax. Particularly relevant in this context will be XCrit (``X-Criterion'') constraints that trigger S-structure movement of a designated element to an X-position,

282

Gereon MuÈller

such as WhCrit (triggering wh-movement) and PronCrit (triggering pronoun movement). On the basis of these assumptions, I will now turn to various instances of orderpreservation e¨ects, and I will show that they can be accounted for in a uni®ed way by invoking ParMove as a violable and ranked constraint. The structure of the argument will look as follows. Given a ranking XCrit g Stay, two cases of orderpreservation e¨ects induced by ParMove can be distinguished. Suppose ®rst that ParMove is ranked lower than XCrit. In that case, a candidate may violate ParMove if this is necessary to ful®ll XCrit, but the violation must be kept minimal: if XCrit can be satis®ed equally well by more than one derivation, ParMove forces the choice of the derivation that minimally violates it. Second, it might be that ParMove is ranked higher than some XCrit constraint. Now, ParMove cannot be violated anymore by X-movement so as to ful®ll XCrit. Rather, if X-movement cannot be order preserving, a derivation will be optimal that avoids X-movement in favor of X±in situ. As we will see, both these cases are attested, and even in a single language (with di¨erent movement types, i.e., di¨erent XCrit constraints). 10.2

Order Preservation

In this section, I show how ParMove explains order-preservation e¨ects with di¨erent movement types triggered by various XCrit constraints: wh-movement in English and Bulgarian, object shift in Danish and Icelandic, Case-driven movement in general, and pronoun fronting and quanti®er raising in German. 10.2.1 Superiority E¨ects in English Superiority e¨ects as they show up with wh-movement in English can be viewed as a typical example of a restriction on otherwise ambiguous rule application. In cases in which there is more than one possible wh-phrase that might undergo wh-movement, the rule of wh-movement can in fact not apply ambiguously to either of the whphrases; rather, the highest wh-phrase has to undergo movement (see Chomsky 1973). This is shown by ``standard'' superiority e¨ects as in (6a) and (6b) (i.e., e¨ects that have been shown to be reducible to the ECP in Aoun, Hornstein, and Sportiche 1981 and Chomsky 1981), and by ``pure'' superiority e¨ects as in (6c) and (6d) (i.e., e¨ects that involve two objects and are therefore not reducible to the ECP in an obvious way; cf. Hendrick and Rochemont 1982 and Pesetsky 1982). (6) a. I wonder [CP who1 C [TP t1 bought what2 ]] b. *I wonder [CP what2 C [TP who1 bought t2 ]] c. Whom1 did John persuade t1 [CP to visit whom2 ] ? d. *Whom2 did John persuade whom1 [CP t20 to visit t2 ] ?

Order Preservation, Parallel Movement, Emergence of the Unmarked

283

Various proposals have been suggested in the literature to account for these e¨ects. However, ParMove o¨ers a new perspective on the phenomenon: the candidates in (6a) and (6c) maintain D-structure order; their competitors in (6b) and (6d) reverse D-structure order. If it can be shown that this ParMove violation is fatal in (6), but tolerable in simple questions like (2a), the superiority e¨ect is derived. To this end, I will now introduce some constraints that regulate wh-movement. I assume that wh-movement is triggered by the Wh-Criterion (WhCrit) (see May 1985; Lasnik and Saito 1984, 1992; and Rizzi 1991, among others).6 WhCrit forces wh-phrases to overtly show up in the domain of C‰‡whŠ (where the notion of ``domain'' is that of Chomsky's (1993) ``checking domain''), either via substitution in the speci®er position of a C node that bears a ‰‡whŠ feature, or via adjunction in the C domain. (7) WhCrit XP‰‡whŠ is in the domain of C‰‡whŠ at S-structure. Clearly, Stay and Wh-Crit impose con¯icting requirements on derivations. As shown in MuÈller 1997b, Ackema and Neeleman 1998, and Legendre, Smolensky, and Wilson 1998, the relative ranking of Stay and (a version of ) WhCrit yields the contrast between wh±in situ type languages such as Korean, Japanese, and Chinese, and wh-movement languages like English and German: a language with the ranking Stay g WhCrit is predicted to lack overt wh-movement; in a language with the ranking WhCrit g Stay, wh-movement is forced, in (minimal) violation of Stay. To distinguish Bulgarian-type languages that exhibit multiple wh-movement (see Rudin 1988) from English-type languages that have only single wh-movement even in multiple questions, another constraint is necessary. Assuming that the speci®er of C is always unique, it is clear that a maximal satisfaction of WhCrit in multiple questions requires adjunction in the C domain. The constraint that we are looking for must therefore be one that blocks this latter operation. Such a constraint has indeed been argued for on independent grounds by Grimshaw (1997); she dubs it PureEP (``Purity of Extended Projection''). The following, modi®ed version of PureEP may su½ce for present purposes (most important, Grimshaw's notion of ``highest subordinate extended projection'' is here replaced by ``domain of C'').7 (8) PureEP Adjunction is prohibited in the domain of C at S-structure. Under a ranking PureEP g WhCrit, question formation of the English type is predicted: One wh-phrase can undergo overt movement (violating the lower-ranked Stay), but all the others must stay in situ, so as to ful®ll the higher-ranked PureEP (which would be violated after wh-adjunction in the C domain). If, on the other hand, WhCrit g PureEP, this implies that the violation of PureEP incurred by wh-

284

Gereon MuÈller

Tableau T10.1 Simple wh-movement in English Candidates

PureEP

WhCrit

f D1 : [CP what2 C John1 . . . t2 ] D2 : [CP Ð C John1 . . . what2 ]

ParMove

Stay

*

*

*!

adjunction is permitted if it leads to maximal satisfaction of WhCrit; thus, multiple overt wh-movement of the Bulgarian type is derived. The question now is how ParMove is ranked in English. Clearly, Wh-Crit is in con¯ict not only with Stay (since it triggers overt movement), but also typically with ParMove (since it may force a reversal of the D-structure order of arguments). The fact that (2a) is still the optimal candidate (as opposed to a candidate with wh±in situ at S-structure, which violates neither Stay nor ParMove) therefore shows that WhCrit must outrank the two faithfulness constraints in English. We can thus postulate the following ranking for English (where the relative ranking of ParMove and Stay is irrelevant for the cases considered here). (9) Ranking in English PureEP g WhCrit g ParMove g Stay Tableau T10.1 illustrates the competition from which the derivation that generates the S-structure representation in (2a) emerges as the optimal candidate in its candidate set. Next, consider the pair of examples in (6a) and (6b) that illustrates the superiority e¨ect in English. The derivations that generate (6a) and (6b) have an identical numeration and thus compete, together with other derivations that, for instance, leave both wh-phrases in situ, or that move both wh-phrases overtly. The competition is illustrated in tableau T10.2. Given the ranking PureEP g WhCrit, multiple overt movement as in D3 and D4 is blocked, and given the ranking WhCrit g Stay, a complete wh±in situ strategy as in D2 is excluded. All these candidates incur violations of high-ranked constraints. The competition of D1 and D5 is of particular importance in the present context. The candidates have an identical constraint pro®le except for ParMove, which by itself is ranked fairly low and can in principle be violated by wh-movement in English (see tableau T10.1). However, given that, all other things being equal, D1 respects ParMove and D5 does not, D5 is blocked by D1 , and the superiority e¨ect is accounted for, as the consequence of a fatal violation of the low-ranked constraint ParMoveÐa violation of a type permitted in other

Order Preservation, Parallel Movement, Emergence of the Unmarked

285

Tableau T10.2 Multiple wh-movement in English Candidates

PureEP

f D1 : [CP who1 C t1 . . . what2 ]

WhCrit

ParMove

*

D2 : [CP Ð C who1 . . . what2 ]

Stay *

**!

D3 : [CP who1 what2 C t1 . . . t2 ]

*!

D4 : [CP what2 who1 C t1 . . . t2 ]

*!

D5 : [CP what2 C who1 . . . t2 ]

**

*

*

**

*!

*

candidate sets, where there is no initial ambiguity (i.e., no otherwise identical constraint pro®le) involved.8 10.2.2 Wh-Movement in Bulgarian Rudin (1985, 1988) observes that Bulgarian exhibits multiple overt wh-movement, in the sense that all wh-phrases must be in the domain of C‰‡whŠ at S-structure. This is shown by the contrast in (10). [CP Koj1 kogo2 C [TP t1 vizÏda t2 ]] ? sees whonom whomacc b. *[CP Koj1 C [TP t1 vizÏda kogo2 ]] ? sees whomacc whonom

(10) a.

Interestingly, the moved wh-phrases must show up in a ®xed order subject  object. C [TP t1 vizÏda t2 ]] ? (11) *[CP Kogo2 koj1 sees whomacc whonom As Rudin (1988) shows, given certain additional assumptions, this e¨ect can be treated more or less on a par with the superiority e¨ect in English in a framework that incorporates the ECP. However, the phenomenon is more general (see Rudin 1985). Multiple wh-movement of three wh-arguments in double object constructions also typically results in a ®xed order. This order happens to be the D-structure one; see (12).9 (12) a.

[CP Koj1 kogo2 kakvo3 C [TP t1 e pital t2 t3 ]] ? asked who whom what b. *[CP Koj1 kakvo3 kogo2 C [TP t1 e pital t2 t3 ]] ? c. *[CP Kakvo3 koj1 kogo2 C [TP t1 e pital t2 t3 ]] ?

286

Gereon MuÈller

d. *[CP Kakvo3 kogo2 koj1 C [TP t1 e pital t2 t3 ]] ? e. *[CP Kogo2 koj1 kakvo3 C [TP t1 e pital t2 t3 ]] ? f. *[CP Kogo2 kakvo3 koj1 C [TP t1 e pital t2 t3 ]] ? This state of a¨airs strongly suggests a constraint like ParMove at work in both (10a)/(11) and (12). However, in order to derive these data from ParMove, an additional assumption seems necessary: In contrast to what is postulated by Rudin (1988), fronting of wh-phrases must not proceed via right adjunction of one NP1 to another NP2 located in SpecC (or to another NP2 at an earlier stage in the derivation, when NP2 is still in situ, as suggested by Ackema and Neeleman (1998) and Grewendorf and Sabel (1996)), but rather via left adjunction to CP, as in (13). Only this way does the linear order of fronted wh-phrases correspond to a simple notion of asymmetric c-command.10 (13) [CP Koj1 [CP kogo2 [CP kakvo3 C [TP t1 e pital t2 t3 ]]]] ? Indeed, adopting (13) seems to be compatible with the main bulk of evidence that Rudin (1988) presents in support of right adjunction to SpecC (most of which suggests that wh-fronting in Bulgarian moves all wh-phrases to the left of C, unlike what is the case in, e.g., Polish). Furthermore, PureEP penalizes both adjunction to CP and adjunction to SpecC, and therefore does not discriminate between the two options. In addition, closer inspection reveals that the kind of wh-cluster formation envisaged by Rudin (1988) raises a number of conceptual and empirical problems.11 Therefore, I will assume that adjunction to CP is the only way permitted by Gen for WhCrit to be ful®lled in multiple questions. Suppose now that the only parametric di¨erence from English relevant in this context concerns the ranking of WhCrit; this constraint dominates PureEP in Bulgarian, inducing multiple wh-movement at S-structure in a multiple question. (14) Ranking in Bulgarian WhCrit g PureEP g ParMove g Stay In simple wh-questions, where PureEP can be vacuously ful®lled, we expect Bulgarian to behave exactly like English. In (15), a wh-object is moved across a non-whsubject, in violation of ParMove. (15) [CP Kakvo2 praviV [TP Ivan1 tV t2 ]] ? Ivan what does The derivation generating (15) is optimal for the very same reasons that the derivation generating (2a) in English is; tableau T10.3 mirrors tableau T10.1. In multiple questions, things begin to di¨er. Due to the rankings WhCrit g PureEP and WhCrit g Stay, all wh-phrases must undergo overt wh-movement in

Order Preservation, Parallel Movement, Emergence of the Unmarked

287

Tableau T10.3 Simple wh-movement in Bulgarian Candidates

WhCrit

PureEP

ParMove

Stay

*

*

PureEP

ParMove

f D1 : [CP kakvo2 pravi Ivan1 t2 ] D2 : [CP Ð pravi Ivan1 kakvo2 ]

*!

Tableau T10.4 Multiple wh-movement in Bulgarian Candidates

WhCrit

f D1 : [CP koj1 kogo2 kakvo3 . . . t1 t2 t3 ]

**

D2 : [CP koj1 . . . t1 kogo2 kakvo3 ]

*!*

D3 : [CP koj1 kogo2 . . . t1 t2 kakvo3 ]

*!

D4 : [CP Ð . . . koj1 kogo2 kakvo3 ]

Stay *** *

*

**

*!**

D5 : [CP koj1 kakvo3 kogo2 . . . t1 t2 t3 ]

**

*!

***

D6 : [CP kakvo3 koj1 kogo2 . . . t1 t2 t3 ]

**

*!*

***

Bulgarian, as they in fact do in (12). The optimal candidate in a candidate set that includes the derivations generating the S-structure representations in (12) therefore inevitably violates PureEP twice (because two instances of CP-adjunction must occur, triggered by WhCrit), violates Stay three times (WhCrit forces movement of all three wh-phrases), and does not violate WhCrit at all. Thus, with respect to these three constraints, all the candidates in (12) have an identical pro®le. It is here that the low-ranked ParMove becomes relevant again; it discriminates between the candidates and chooses as the sole optimal derivation the one that preserves the D-structure order. All this is shown in tableau T10.4. Here, D2 is a derivation that fatally violates WhCrit but would have won under the English ranking. Similarly, D4 is a derivation that is excluded by WhCrit; this candidate would have been classi®ed as optimal under a Korean-type ranking. D3 is a candidate in which two wh-phrases undergo movement, and one stays behind in its in situ position; this candidate is predicted to be suboptimal under any ranking of the constraints adopted here. Finally, D5 and D6 are (some of the) derivations that exhibit wh-movement of all three wh-phrases but

288

Gereon MuÈller

fail to maintain the D-structure order, in contrast to D1 , which therefore emerges as optimal. In sum, under a ranking WhCrit g ParMove it is correctly predicted that whmovement must preserve order if it can (as in multiple questions in English and Bulgarian), but if there is no way wh-movement can do so, ParMove can selectively be violated (as in simple questions in English and Bulgarian). In the next subsection, I turn to a manifestation of the other conceivable situationÐa movement type either respects ParMove or cannot apply at all. 10.2.3 Pronominal Object Shift in Danish Object shift is a clause-internal A-movement operation restricted to weak (unstressed) pronouns in Mainland Scandinavian languages like Danish. Object shift is dependent on leftward raising of the main verb to a higher position (i.e., on V/2 movement in Danish, which lacks overt V-to-I movement; see Vikner 1990). Furthermore, if V raising has applied, object shift is normally obligatory for weak pronouns (see, among others, Holmberg 1986; Vikner 1990, 1994; Deprez 1994; Roberts 1995). A pair of examples that illustrates the obligatoriness of object shift in Danish is given in (16) (see Vikner 1994). (16) a. *Hvorfor why b. Hvorfor why

kobteV bought kobteV bought

Peter Ð ikke tV den1 ? it Peter not Peter den1 ikke tV t1 ? not Peter it

If two (weak) object pronouns show up in a double-object construction, both have to undergo object shift. Interestingly, multiple object shift preserves the D-structure order of arguments (see Vikner 1990 and Johnson 1991). This is shown by the data in (17). (17) a. b. c. d. e. f.

Peter viste hende1 den2 jo t1 t2 Peter showed her it indeed *Peter viste den2 hende1 jo t1 t2 *Peter viste Ð Ð jo hende1 den2 *Peter viste Ð Ð jo den2 hende1 *Peter viste hende1 Ð jo t1 den2 *Peter viste Ð den2 jo hende1

This pattern is exactly what we would expect, given ParMove. Interestingly, there is a further restriction: Simple object shift of a pronominal direct object across a lexical indirect object NP is blocked, in contrast to object shift of a pronominal indirect object in the presence of a nonpronominal direct object; see (18) (from Vikner 1990).

Order Preservation, Parallel Movement, Emergence of the Unmarked

(18) a. *Peter Peter b. Peter Peter

viste showed viste showed

289

den2 jo Marie1 t2 it indeed Marie hende1 jo t1 bogen2 her indeed the book

Again, this suggests ParMove at workÐthe S-structure representation in (18b) preserves the argument order established at D-structure, which the S-structure representation in (18a) does not. However, in contrast to what we have seen in the case of simple wh-movement in English and Bulgarian, here it looks as though ParMove has a somewhat stronger e¨ect on movement: either pronominal object shift respects ParMove, or it may not apply at all. That this di¨erence is related to di¨erent movement types, and not to di¨erent languages, is shown by the fact that simple whmovement and topicalization can freely violate ParMove, as shown in (19) (from Vikner 1990, personal communication).12 (19) a. Hvilken bog2 har Peter1 lñst t2 ? which book has Peter read b. Denne bog2 viste Peter ikke Marie1 t2 this book showed Peter not Marie To account for order preservation with object shift, I would like to suggest that object shift of pronouns in the Scandinavian languages is triggered by a constraint analogous to WhCrit. This constraint will be referred to in what follows as the Pronoun Criterion, or PronCrit. PronCrit forces weak pronouns to show up in the domain of a functional head p at S-structure. I will assume that the functional projection pP intervenes between TP and vP/VP.13 (20) PronCrit Weak pronouns must be in the domain of p at S-structure. Now the data in (17) can be analyzed in complete analogy to what was said about order preservation in Bulgarian multiple questions in the previous subsection. To ful®ll PronCrit, all pronouns in a double object construction undergo overt raising to pP, with one pronoun substituting in Specp and the other adjoining to pP.14 PronCrit must be ranked higher than Stay in Danish, but whether ParMove dominates PronCrit or is dominated by it does not play a role for the data in (17). The situation is di¨erent in (18a), repeated here as (21a) and augmented by (21b) (see Vikner 1990). (21) a. *Peter Peter b. ??Peter Peter

viste den2 jo Marie1 t2 showed it indeed Marie viste Ð jo Marie1 den2 shows indeed Marie it

290

Gereon MuÈller

Tableau T10.5 Multiple object shift in Danish Candidates

ParMove

PronCrit

f D1 : V [ hende1 [ den2 p . . . t1 t2 ]] D2 : V [ den2 [ hende1 p . . . t1 t2 ]]

** *!

D3 : V [ Ð [ Ð p . . . hende1 den2 ]] D4 : V [ Ð [ Ð p . . . den2 hende1 t2 ]]

** *!*

*!

D5 : V [ Ð [ hende1 p . . . t1 den2 ]] D6 : V [ Ð [ den2 p . . . hende1 t2 ]]

Stay

*!

**

*

*!

*

*

*

Example (21a) corresponds to the wh-movement cases (2a) in English and (15) in Bulgarian in the sense that it involves unambiguous movement violating ParMove. However, in this case, ungrammaticality results, and this implies that ParMove must in fact be ranked higher than PronCrit in Danish. Consequently, the optimal candidate violates PronCrit to ful®ll ParMove, and not vice versa. Hence, it is (21b) (with the direct object pronoun in situ despite V raising), rather than (21a), that turns out to be optimal.15 Under these assumptions, we end up with a ranking like (22) for Danish. (22) Ranking in Danish PureEP g WhCrit g ParMove g PronCrit g Stay Tableau T10.5 then illustrates the situation with multiple object shift in Danish (see (17)). D3 ±D6 incur violations of PronCrit because they fail to move both pronouns. D1 and D2 , in contrast, both respect PronCrit via multiple object shift. The constraint pro®le of these two candidates is identical, except for the fact that D1 respects ParMove and D2 violates it; the latter violation therefore becomes fatal. The competition in cases of simple object shift of a direct object pronoun in double object constructions in Danish is shown in tableau T10.6 (see (21)). Here, the optimal candidate is one that violates PronCrit to ful®ll the higher-ranked ParMove (but see the quali®cation in note 15). Thus, we have derived the fact that object shift can never change the order of arguments in DanishÐthe movement type respects Dstructure order or does not apply at all.16 Note ®nally that the ranking WhCrit g ParMove g PronCrit yields the result that simple wh-movement may alter the D-structure order of arguments, in contrast

Order Preservation, Parallel Movement, Emergence of the Unmarked

291

Tableau T10.6 Simple object shift in Danish Candidates

ParMove

D1 : V [ den2 p . . . Marie1 t2 ]

PronCrit

Stay

*!

*

f D2 : V [ p . . . Marie1 den2 ]]

*

Tableau T10.7 Simple wh-movement in Danish Candidates

PureEP

WhCrit

f D1 : [CP hvilken bog2 C Peter1 . . . t2 ] D2 : [CP Ð C Peter1 . . . hvilken bog2 ]

ParMove

Stay

*

*

*!

to simple object shift; see (19a). Thus, tableau T10.7 is identical in all relevant aspects to tableau T10.1 from English. This basically exhausts the number of ways ParMove can interact with a con¯icting constraint XCrit that triggers movement: either XCrit is ranked higher, in which case it permits a violation of ParMove if there is no other way for it to be satis®ed (this holds for WhCrit in all the languages discussed here), or XCrit is ranked lower, in which case the optimal candidate is one without Xmovement in the case of con¯ict (this holds for PronCrit in Danish). All other pieces of evidence for a constraint like ParMove that will be presented below follow one of the two patterns that have now been established. 10.2.4 Object Shift of Lexical NPs in Icelandic The situation with multiple object shift of nonpronominal NPs in Icelandic is similar to what we have seen with pronominal object shift in Danish. However, unlike pronominal object shift, lexical object shift in Icelandic is optional; see Holmberg 1986, Vikner 1990, and Collins and ThraÂinsson 1996, among others. The following data show that multiple object shift of nonpronominal NPs in double object constructions is order preserving (from Collins and ThraÂinsson 1996). EÂg I b. *EÂg I

(23) a.

laÂna lend laÂna lend

MarõÂu1 bñkurnar2 Maria the books bñkurnar2 MarõÂu1 the books Maria

ekki t1 t2 not ekki t1 t2 not

292

Gereon MuÈller

Interestingly, even though object shift is in principle optional for full NPs in Icelandic, it is blocked if the shifted item is the direct object, and the indirect object stays in situ. The indirect object, on the other hand, can shift even if the direct object stays in situ. This is shown in (24). (24) a. *EÂg I b. EÂg I

laÂna lend laÂna lend

bñkurnar2 ekki MarõÂu1 t2 the books not Maria MarõÂu1 ekki t1 bñkurnar2 Maria not the books

Again, the contrasts in (23) and (24) suggest an account in terms of ParMove. The main di¨erence between the two object shift types that is relevant in the present context is that nonpronominal object shift in Icelandic is an optional movement operation. As is typical for all syntactic theories employing an economy constraint like Stay, permitting true optionality of rule application initially poses a problem. In line with the standard approach taken in view of this situation (see, e.g., Chomsky 1995), I will assume that the apparent optionality of object shift of lexical NPs in Icelandic is to be reanalyzed as being due not to a genuine optionality of rule application, but to the optional presence of movement-triggering features on lexical items Ðthe movement operation as such is obligatory. Thus, suppose that there is a constraint that obligatorily triggers object shift of lexical NPs bearing a certain feature F (``FCrit''), and that F is optionally instantiated on a lexical item in the numeration. Exactly what this feature F looks like, and what position F-Crit forces F-bearing NPs to move to, depends on further assumptions that I do not have the space to pursue here, but a clari®cation of these issues is not necessary for the account of order preservation developed in what follows.17 The empirical evidence suggests that FCrit (unlike, e.g., WhCrit) is ranked lower than ParMove in Icelandic, in complete analogy to the ranking ParMove g PronCrit in Danish. From such a ranking it then follows that object shift of lexical NPs in Icelandic either respects ParMove or does not apply at all.18 (25) Ranking in Icelandic PureEP g WhCrit g ParMove g FCrit g Stay To account for the data in (23) and (24), we have to consider three candidate sets. In one candidate set (see tableau T10.8), F is present on both the direct and indirect object. Here, the derivation generating (23a) (D3 ) is the optimal candidate. The competing derivation generating (23b) (D4 ) is blocked due to a fatal ParMove violation, and the derivations that underlie (24a) and (24b) (i.e., D5 and D2 , respectively) involve fatal violations of FCrit, since F is present on both objects, by assumption. But of course the string in (24b) is well formed as such, so there must be a derivation

Order Preservation, Parallel Movement, Emergence of the Unmarked

293

Tableau T10.8 Nonpronominal object shift, with F on indirect object & direct object Candidates

ParMove

FCrit

D1 : Ð Neg NP1 NP2

*!*

D2 : NP1 Neg t1 NP2

*!

Stay

*

f D3 : NP1 NP2 Neg t1 t2

**

D4 : NP2 NP1 Neg t1 t2

*!

D5 : NP2 Neg NP1 t2

*!

** *

*

Tableau T10.9 Nonpronominal object shift, with F on indirect object only Candidates D1 : Neg NP1 NP2 f D2 : NP1 Neg t1 NP2

ParMove

FCrit

Stay

*! *

that generates it that is optimal. Indeed, the derivation generating (24b) emerges as a winning candidate in a second candidate set that contains derivations in which the object shift feature is present only on the indirect object; see tableau T10.9. The relevant derivations that need to be considered here are D1 , in which the indirect object fails to undergo object shift, and D2 , in which it is moved. Clearly, the latter candidate is the only one that ful®lls FCrit, and it is therefore optimal.19 However, (24a) is blocked as suboptimal in all candidate sets. In particular, it is blocked as involving a fatal ParMove violation in a third candidate set that contains the relevant feature F only on the direct object. Here the optimal derivation is one with both object NPs in situ; see tableau T10.10.20 Thus, ParMove restricts object shift of lexical NPs in Icelandic in such a way that it must not apply if D-structure argument order cannot be preserved, in complete analogy to pronominal object shift in Danish. Note that D1 emerges as optimal not only in the candidate set underlying tableau T10.10 (with F instantiated on the direct object only), but also in yet another candidate set in which F is not instantiated at all. Hence we end up with the result that a sentence with both objects in situ is derivationally ambiguous, in the sense that it can be the winner of two di¨erent competitions (i.e., we have another instance of neutralization of input speci®cations; see note 15).

294

Gereon MuÈller

Tableau T10.10 Nonpronominal object shift, with F on direct object only Candidates

ParMove

f D1 : Neg NP1 NP2 D5 : NP2 Neg NP1 t2

FCrit

Stay

* *!

*

10.2.5 Case-Driven NP Raising According to some analyses (see note 17), optional object shift of lexical NPs in Icelandic is already an instance of Case-driven NP raising. However, quite independently of this speci®c issue, the problem of accounting for order preservation with Case-driven NP raising shows up in general in approaches that postulate multiple Case-driven movement within a single clause. Assuming with Chomsky (1995) that NPs are base generated VP-internally and must raise to an external position to check structural Case, a well-known problem arises: Why does NP raising (whether overt or covert) of, for example, a subject NP and an object NP in a simple transitive structure result in asymmetric c-command of the object by the subject, and not vice versa? For the sake of concreteness, suppose, basically as in Chomsky 1995, that objects are base generated VP-internally, and that subjects are base generated in the speci®er of a light verb v, as in (26). The question then is why the target of Case-driven subject movement must be SpecT, and why the target of Case-driven object movement must be a position in the domain of v (a second speci®er of v, or, as I will assume here, a vP-adjoined position). (26) [TP Ð [T0 T [vP Adj2 [vP [Spec1 NPSubj ] [v0 v [VP V NPObj ]]]]]] It is clear that order-preservation e¨ects with Case-driven NP raising can be explained as the result of a violable and low-ranked ParMove constraint that interacts with the relevant markedness constraint (FCrit or, perhaps more speci®cally, the Case Filter). NPSubj raising to SpecT accompanied by NPObj raising to Adj2 of v in (26) respects ParMove, whereas NPSubj raising to Adj2 accompanied by NPObj raising to SpecT does not, and the former option therefore blocks the latter option as suboptimal.21 10.2.6 Pronoun Fronting in German Movement of weak (unstressed) pronouns to the Wackernagel domain results in a ®xed order in German (see, e.g., Lenerz 1977, 1992; Haftka 1981).22 As (27) shows, displaced subject pronouns obligatorily precede displaced object pronouns.

Order Preservation, Parallel Movement, Emergence of the Unmarked

(27) a.

daû that b. *daû that

sie1 es2 she it es2 sie1 it she

wahrscheinlich t1 t2 gelesen probably read wahrscheinlich t1 t2 gelesen probably read

295

hat has hat has

Displaced direct object pronouns precede displaced indirect object pronouns: (28) a.

daû that b. *daû that

es1 ihm2 it him ihm2 es1 him it

der ART der ART

Fritz t1 t2 gegeben Fritz given Fritz t1 t2 gegeben Fritz given

hat has hat has

Like pronominal object shift in Scandinavian, pronoun fronting to a Wackernagel position in German is obligatory; weak pronouns that show up to the right of VPadjoined adverbials invariably create ungrammaticality. (29) *daû sie1 wahrscheinlich t1 es2 gelesen hat that she probably it read has If all three arguments in a double object construction are weak pronouns, the order after pronoun movement to the Wackernagel position is subject  direct object  indirect object; all other permutations are ill formed. (30) a. b. c. d. e. f.

daû sie1 es2 ihm3 wahrscheinlich zum Geburtstag t1 t2 t3 schenken wird that she it him probably for the birthday give will *daû sie1 ihm3 es2 wahrscheinlich zum Geburtstag t1 t2 t3 schenken wird *daû es2 sie1 ihm3 wahrscheinlich zum Geburtstag t1 t2 t3 schenken wird *daû es2 ihm3 sie1 wahrscheinlich zum Geburtstag t1 t2 t3 schenken wird *daû ihm3 sie1 es2 wahrscheinlich zum Geburtstag t1 t2 t3 schenken wird *daû ihm3 es2 sie1 wahrscheinlich zum Geburtstag t1 t2 t3 schenken wird

These facts are strongly reminiscent of the situation with pronominal object shift in Danish (the movement operation targets a VP-external position, it results in a ®xed order, and it a¨ects only weak pronouns), and therefore suggest a uni®ed approach in terms of ParMove. However, there are some di¨erences. First, at ®rst glance it looks as though the landing sites of the two operations are not identical: they must be vP/VP external in both cases, but Scandinavian object shift targets a position following the subject, while German pronoun fronting ends up in the Wackernagel position, which can either precede or follow the position of subjects in German; see (31). (31) a.

daû that b. *daû that

der ART der ART

Fritz3 Fritz Fritz3 Fritz

es1 ihm2 t3 t1 t2 gegeben it him given ihm2 es1 t3 t1 t2 gegeben him it given

hat has hat has

296

Gereon MuÈller

There is independent evidence that the coexistence of the data in (28) and (31) is due to an optionality concerning the surface position of subjects in German, and not to an optionality in pronoun fronting: not only is it impossible for a weak pronoun to follow a VP-adverbial (see (29)); it cannot follow a nonpronominal indirect object NP either, even if it follows the subject. (32) a.

daû that b. *daû that

der ART der ART

Fritz3 Fritz Fritz3 Fritz

es1 der Maria2 t3 t1 t2 gegeben it ART Maria given der Maria2 es1 t3 t1 t2 gegeben ART Maria it given

hat has hat has

Accordingly, I would like to suggest that this di¨erence between object shift in Danish and Wackernagel fronting in German is unrelated to the issue of pronoun movement: both movements are triggered by Pron-Crit and end up in the p domain, which is the Wackernagel position. In this view, the di¨erence between Danish and German is simply that German subjects may optionally stay in situ or raise to SpecT (see Grewendorf 1989, among others), whereas Danish subjects must raise to SpecT, strictly precluding adjacency of C and a fronted pronoun.23 A second (and more fundamental) di¨erence is that the order of the fronted pronouns is direct object  indirect object in German, but indirect object  direct object in Danish. If both S-structure orders result from the activity of a low-ranked ParMove constraint, this presupposes that the D-structure order with double object verbs in German is direct object  indirect object, in contrast to what I have assumed so far for Icelandic and Danish (where the base order must be indirect object  direct object). Indeed, there are some di¨erences between double object constructions in German and, for example, Danish that support such a view.24 To name just one, it has been noted by Grewendorf (1988) that a direct object NP can bind an indirect object anaphor that it precedes in German but not vice versa; see (33). (33) a.

daû that b. *daû that

der the der the

Arzt doctornom Arzt doctornom

den Patienten1 sich1 im Spiegel zeigte the patientacc himselfdat in the mirror showed dem Patienten1 sich1 t1 im Spiegel zeigte the patientdat himselfacc in the mirror showed

As shown in MuÈller and Sternefeld 1994, this follows if the direct object is base generated above and to the left of the indirect object. In this view, (33a) is well formed because the indirect object re¯exive can be A-bound at S-structure by the direct object NP, whereas (33b) is ill formed because it involves an illicit crossover con®guration after scrambling of the indirect object to a position c-commanding the direct object re¯exive. The situation is di¨erent in Danish, however. An indirect object can bind a direct object anaphor that follows it; see Vikner 1985.

Order Preservation, Parallel Movement, Emergence of the Unmarked

297

(34) Jeg viste Jon1 ham selv1 i spejlet I showed Jondat himselfacc in the mirror This suggests that the D-structure order is indirect object  direct object in Danish, in contrast to the situation in German. I believe that further di¨erences between German and the Scandinavian languages (related, e.g., to locality, weak crossover, passivization, and word order) point to the same conclusion (see MuÈller 1995). For the time being, though, I will leave it at that, and turn to a third crucial di¨erence. Scandinavian object shift can never violate ParMove, whereas German pronoun fronting can. This implies that PronCrit must be ranked higher than ParMove in German, in contrast to what we have seen in Danish. Hence, if pronoun fronting in German leads to a reversal of D-structure argument order with nonpronominal NPs, which are not subject to PronCrit, this is permitted; see (28).25 Thus, I would like to suggest the following ranking in German, which treats PronCrit on a par with WhCrit, as far as the relation to ParMove is concerned.26 (35) Ranking in German PureEP g WhCrit g PronCrit g ParMove g Stay Given the above assumptions, the order-preservation e¨ects with pronoun fronting in German follow from this ranking. For instance, tableau T10.11 shows why (28a) emerges as the winner of the competition in (28). Both D1 and D2 incur two ParMove violations because two objects are moved across the subject. The violation that is fatal for D2 is the third ParMove violation that results from a reversal of the order of the two objects.27 Similarly, ParMove yields the correct results for more complex examples, in which all three arguments in a double object construction are pronouns susceptible to PronCrit, as in (30). Tableau T10.12 shows that the derivation that generates (30a), which maximally respects ParMove, is the sole optimal candidate in Tableau T10.11 Pronoun fronting in German with a nonpronominal subject Candidates

PronCrit

f D1 : es1 ihm2 der Fritz t1 t2 . . . D2 : ihm2 es1 der Fritz t1 t2 . . .

ParMove

Stay

**

**

***!

**

D3 : es1 der Fritz t1 ihm2 . . .

*!

*

*

D4 : ihm2 der Fritz es1 t2 . . .

*!

**

*

D5 : der Fritz es1 ihm2 . . .

*!*

298

Gereon MuÈller

Tableau T10.12 Pronoun fronting in German with a pronominal subject Candidates

PronCrit

ParMove

f D1 : sie1 es2 ihm3 . . . t1 t2 t3

Stay ***

D2 : sie1 es2 . . . t1 t2 ihm3

*!

D3 : sie1 ihm3 . . . t1 es2 t3

*!

** *

**

D4 : sie1 ihm3 es2 . . . t1 t2 t3

*!

***

D5 : es2 sie1 ihm3 . . . t1 t2 t3

*!

***

D6 : es2 ihm3 sie1 . . . t1 t2 t3

*!*

***

D7 : ihm3 sie1 es2 . . . t1 t2 t3

*!*

***

D8 : ihm3 es2 sie1 . . . t1 t2 t3

*!**

***

its candidate set, and blocks the derivations that generate the remaining S-structure representations in (30). Finally, to conclude the discussion of pronoun fronting in German, a few remarks on scrambling are appropriate. As (36) shows, the strict order-preservation e¨ects with pronoun fronting do not show up with nonpronominal NPs. On the contrary, in a double object construction, all permutations derivable by scrambling are well formed (although not equally unmarked; note that the candidate that maximally respects ParMove by exhibiting D-structure order is (36b), whereas the unmarked order is the order in (36a)). (36) a. daû der Fritz der Maria das Buch gab that ART Fritznom ART Mariadat the bookacc gave b. daû der Fritz das Buch der Maria gab c. daû der Maria der Fritz das Buch gab d. daû der Maria das Buch der Fritz gab e. daû das Buch der Fritz der Maria gab f. daû das Buch der Maria der Fritz gab Thus, the way out of ParMove e¨ects that is evidently available for lexical NPs in German must not be available for weak pronouns. In fact, it has been argued by a number of researchers, and on di¨erent grounds, that pronoun fronting in German cannot be analyzed as an instance of scrambling; see in particular Thiersch's (1978:84) rule C1 (which a¨ects pronouns but not lexical NPs), and also Cardinaletti and Roberts 1991, Schmidt 1992, Zwart 1993, and MuÈller 1998. I will assume that

Order Preservation, Parallel Movement, Emergence of the Unmarked

299

whereas pronoun movement is triggered by PronCrit and ends up in the p domain, scrambling in German is a movement rule that a¨ects only lexical NPs (PPs, CPs) and applies within the vP/VP domain. Given Stay and ParMove, there must be a higher-ranked trigger for scrambling. It has been argued that this trigger should not be encoded as a (criterion-like) standard constraint that forces the localization of an XP in a designated structural domain, but rather as a set of various interacting linearization constraints that are themselves violable and ranked and that center around notions like focus, topic, speci®city, de®niteness, animacy, and so on. A predecessor of such an analysis can be found in Uszkoreit 1984. OT approaches of this general type include Choi 1996; Costa, chap. 7, this volume; BuÈring 1999; and MuÈller 1999. The last paper can be viewed as complementing the present chapter; it focuses on the inherent tension between the requirement to make scrambling maintain D-structure order imposed by ParMove, and the requirement to make scrambling change Dstructure order imposed by the higher-ranked linearization constraints. The loophole created by scrambling for an explanation of order-preservation e¨ects with pronoun fronting in German is now closed, and a proper account of the interaction of ParMove and order-changing scrambling has been hinted at. I will leave it at that, concluding that a violable ParMove constraint makes it possible to develop a uni®ed approach that treats pronoun fronting in German the same way as object shift in Danish.28 10.2.7 Relative Scope and QR in German At least since Lako¨ 1971, linguists have generally assumed that relative scope at the level of logical form is crucially determined by surface word order in the unmarked case. If scope reversal nevertheless takes place in a sentence, this can then viewed as a consequence of the presence of one of various intervening factors. Kroch (1974) explicitly suggests handling the impact of these intervening factors in terms of repair strategies (called scope readjustment rules). Factors that can create a relative scope that di¨ers from S-structure order include speci®c intonation patterns, inherent properties of quanti®ers (like ``strength''), and so on; see, for example, Kroch 1974; Huang 1982; Reinhart 1983; Liu 1997. As far as German is concerned, there is agreement that S-structure word order is indeed highly relevant for determining relative scope (see Frey 1989; Moltmann 1991; Pafel 1993; BuÈring 1996; Beck 1996; Diesing 1997; Heck 1999). Abstracting away from certain potentially intervening factors (i.e., assuming that neutral intonation is present, that the quanti®ed expressions involved are not inherently prone to wide scope readings, and so on), it seems that by far the most natural (and perhaps the only available) reading for the sentences in (37) that involve a subject quanti®er and an object quanti®er is one that corresponds to the S-structure order of the quanti®ed items.

300

Gereon MuÈller

(37) a. daû mindestens ein Gast1 viele Geschenke2 mitbrachte that at least one guestnom many presentsacc brought b. daû viele GaÈste1 mindestens ein Geschenk2 mitbrachten that many guestsnom at least one presentacc brought c. daû viele Geschenke1 mindestens ein Gast2 t1 mitbrachte that many presentsacc at least one guestnom brought d. daû mindestens ein Geschenk1 viele GaÈste2 t1 mitbrachten that at least one presentacc many guestsnom brought In (37a) and (37b), the subject NP1 precedes the object NP2 , and this order determines relative scope. More interesting in the present context are the examples in (37c) and (37d). Here, the object NP1 is scrambled across the subject NP2 , but relative scope can still be read o¨ the S-structure representation. Due to scrambling of the object in front of the subject, the subject does not take scope over the object anymore (assuming as before that a neutral intonation pattern is present, and not one that tends to trigger scope inversion, such as the so-called I-intonation pattern; see Jacobs 1982 and BuÈring 1995, among others). Next, consider the relative scope of direct and indirect object in a double object construction in German, as in (38). (38) a. daû man that one b. daû man that one c. daû man that one d. daû man that one

mindestens einem Gast1 viele Geschenke2 gab at least one guestdat many presentsacc gave vielen GaÈsten1 mindestens ein Geschenk2 gab many guestsdat at least one presentacc gave mindestens ein Geschenk1 vielen GaÈsten2 gab at least one presentacc many guestsdat gave viele Geschenke1 mindestens einem Gast2 gab many presentsacc at least one guestdat gave

Again, relative scope corresponds to the S-structure order, irrespective of the base position of the two arguments, and irrespective of how the S-structure order is derived from D-structure. Thus, if an indirect object precedes a direct object, it takes scope over it (cf. (38a) and (38b)), and the opposite is the case if the direct object precedes the indirect object at S-structure (see (38c) and (38d)); note that this way readings can be forced that are not necessarily the pragmatically most plausible ones.29 As a sketch of an analysis, suppose that quanti®ers obligatorily undergo quanti®er raising (QR) at LF in order to avoid a type mismatch that would block type-driven interpretation in situ (see Stechow 1996 and Heim and Kratzer 1998, among others); this requirement can be called QCrit (Quanti®er-Criterion). It then follows that the data in (37) and (38) show that, in the unmarked case, QR is an order-preserving movement operation, in the sense that it does not change the c-command relations among quanti®ers that hold at S-structure. This, of course, immediately follows from

Order Preservation, Parallel Movement, Emergence of the Unmarked

301

ParMove, in the same way that the order-preservation e¨ects that were discussed in the preceding subsections do. The only fundamental di¨erence is that in this case, the evidence exclusively concerns the relationship between S-structure and LF, not between D-structure and S-structure. QCrit dominates ParMove, which implies that the relative S-structure order of di¨erent quanti®ed NPs must be preserved with multiple QR at LF, but that intervening nonquanti®ed NPs (e.g., proper names) can be crossed by QR if necessary. Furthermore, in this approach intervening factors that create scope reversal (like intonation or inherent quanti®er strength) can be represented as constraints that outrank ParMove and thus blur its e¨ects in the case of con¯ict.30 10.3

Parallel Movement and the Minimal Link Condition

The last section has shown that a variety of movement types exhibit orderpreservation e¨ects that are immediately amenable to an account in terms of ParMove, on the basis of an OT approach to syntax. These movement types di¨er substantially with respect to a number of properties. All may face an initial ambiguity concerning the items they can apply to, but some movement types can a¨ect more than one item per clause, whereas others can a¨ect only one item (wh-movement in Bulgarian vs. wh-movement in English). Also, some movement types are obligatory, whereas others are optional (pronominal object shift in Danish vs. object shift of lexical NPs in Icelandic). Moreover, some movement types show A-bar properties, whereas others show A-properties (pronoun fronting in German vs. pronominal object shift in Danish). Finally, some movement types are overt, whereas others are covert (wh-movement in Bulgarian, some instances of Case-driven raising vs. quanti®er raising in German, other instances of Case-driven raising). Taken in isolation, each of these order-preservation e¨ects with a given movement operation can certainly be accounted for in one way or another without resort to a constraint like ParMove. However, given that order preservation is a recurring pattern among movement operations with otherwise radically di¨erent properties, and among various types of languages, such a strategy would seem to miss a simple underlying generalization that can be captured directly by a violable and ranked constraint like ParMove. The question arises of whether there might be another constraint with comparable empirical range in a standard, non-OT model of grammar. An obvious candidate is Chomsky's (1995) Minimal Link Condition in (39) (the notion of closeness here is ideally to be understood in terms of c-command). (39) Minimal Link Condition (MLC) K attracts a only if there is no b, b closer to K than a, such that K attracts b.

302

Gereon MuÈller

Since the MLC is essentially a feature-based version of the Superiority Condition, it comes as no surprise that it accounts for superiority e¨ects in English (see (6)). The MLC as such does not account for order preservation with multiple wh-movement in Bulgarian, though (see (10a)/(11), (12)). Richards (1997:114) and Mulders (1997) suggest economy-based accounts of the phenomenon and add what is essentially a transderivational shortest-path condition to the standard formulation of the MLC; this, however, means that the two instances of order preservation with wh-movement do not receive a uni®ed account. Chomsky (1995:355±367) discusses in detail the e¨ect of the MLC on Case-driven NP raising as in (26), and observes that the MLC does not predict these movements to be possible without additional assumptions: since a subject NP asymmetrically c-commands an object NP in the premovement structure, the MLC in its standard formulation blocks raising of the object NP to a higher position, crossing the subject NP. In view of this, Chomsky proposes two possible solutions. First, one can de®ne closeness in (39) not in terms of the structural notion of c-command, but in terms of the more liberal notion of minimal domain (equidistance). Second, one can assume that the base position of the subject NP is higher than the Case position of the object NP, which leads to extremely short paths with Case-driven movement and arguably threatens to render the idea of Case-driven movement itself vacuous. Collins and ThraÂinsson (1996) and Collins (1997: chap. 3) develop an MLC-based account of order preservation with object shift. They presuppose that object shift is uniformly an instance of Case-driven NP raising, and that the de®nition of closeness in (39) is not simply de®ned in terms of c-command, but incorporates the concept of minimal domain/equidistance. Because of equidistance, a derivation of double object shift constructions such as (23a) in Icelandic is permitted in which ®rst the direct object moves to a position in the same minimal domain as the indirect object, and then the indirect object moves to a higher position. If, however, the indirect object is moved to the lower target position ®rst, and the direct object then raises to a higher position (crossing both the indirect object and its trace), as in (23b), the MLC is violated because the second movement operation is not su½ciently local anymore. This approach works well for cases like (23) in Icelandic and can be extended to (17) in Danish. However, it does not account for (24) in Icelandic and (18) in Danish. The problem is that the ill-formed examples (24a) and (18a) correspond to intermediate steps in the derivation of the well-formed examples (23a) and (17a), respectively, and hence cannot violate the MLC for principled reasons. To rule out examples of this type, Collins and ThraÂinsson (1996:420±424) introduce additional assumptions unrelated to the MLC, which again means a uni®ed account is impossible. In view of the fact that the MLC has a signi®cantly smaller range of empirical coverage in these domains than ParMove that can only be enlarged by additional, unrelated assumptions, and given that MLC-based accounts of order preservation

Order Preservation, Parallel Movement, Emergence of the Unmarked

303

with pronoun fronting and quanti®er raising have, to the best of my knowledge, not been proposed (and do not suggest themselves in any obvious way), I would like to contend that an account of order-preservation e¨ects in terms of a violable ParMove constraint is superior to one in terms of an inviolable MLC. 10.4

Conclusion

Let me summarize the main ®ndings of this chapter. Evidence from a variety of languages suggests that in the unmarked case, all movement operations are order preserving. In view of this, I have argued that a violable Par-Move constraint that is part of an OT grammar can account for that property in a uni®ed way, in contrast to other constraints like the MLC.31 The extent to which ParMove has an e¨ect on a given movement type depends on how the constraint that triggers the movement type (XCrit) is ranked with respect to ParMove. Under a ranking XCrit g ParMove, it follows that X-movement can selectively violate ParMove; in this case, ParMove becomes important only if two candidates behave identically with respect to XCrit (and otherwise). This situation holds in the case of wh-movement in Bulgarian, English, Danish, Icelandic, and so on, pronoun fronting in German, Case-driven NP raising in general, and quanti®er raising in German. However, under a reverse ranking ParMove g XCrit, we obtain the result that X-movement can never violate ParMove: in the case of con¯ict, ParMove blocks X-movement altogether. This more drastic e¨ect has been argued to show up with object shift of pronouns in Danish, and object shift of lexical NPs in Icelandic. Of course, there is no principled reason why such an e¨ect should not show up with wh-movement, topicalization, or a related movement type. Indeed, it has been noted that only subjects can undergo relativization in Malagasy (see Keenan and Comrie 1977). Similarly, only subjects can undergo wh-movement in Lango (see Legendre, Smolensky, and Wilson 1998 and the references cited there). These facts follow from rankings like ParMove g RelCrit and ParMove g WhCrit: under such rankings, relativization and wh-movement maintain asymmetric c-command at D-structure or do not apply at all. To conclude, given the observation that order preservation is a common and recurring pattern among movement types that otherwise di¨er substantially, and that most of these movement types in fact do not preserve order under certain circumstances, I believe that a case can be made for postulating an underlying constraint that is violable and usually ranked quite low, as I have tried to do here with ParMove. There seems no way to maintain a general, non-construction-speci®c constraint like ParMove in a grammar that recognizes only inviolable constraints. Therefore, to the extent that the preceding discussion has made the existence of such a constraint plausible, it can be viewed as an argument for an OT organization of grammar.

304

Gereon MuÈller

Still, many questions remain. Some have been discussed or at least alluded to in the preceding sections; here, I will con®ne myself to pointing out one obvious question that demands further investigation. Since ParMove often predicts crossing (rather than nesting) paths with instances of multiple movement, the status of e¨ects that have sometimes been attributed to a nestedness (or path containment) condition (see Fodor 1978; Pesetsky 1982; May 1985) is unclearÐat ®rst glance, it looks as though the two constraints are incompatible. It is not obvious to what extent the two concepts (order preservation/crossing on the one hand, and nesting on the other) can or should be reconciled. However, I think that three observations are worth bearing in mind when this issue is properly addressed. First, surprising as this may seem at ®rst, it has turned out that some e¨ects analyzed in terms of a nestedness condition now actually follow from ParMove. For instance, this holds for superiority e¨ects in English, derived in Pesetsky 1982 from an illicit crossing of paths. Second, some of the standard nestedness e¨ects involve di¨erent movement types that, consequently, are triggered by di¨erent criteria (e.g., topicalization and wh-movement, wh-movement and tough-movement, and so on), about which ParMove says nothing. In these cases, there is no ambiguity in rule application (topics must be moved to topic positions, and wh-phrases must end up in wh-positions, irrespectively of whether these movements preserve D-structure order or not). Finally, it should be noted that, under present assumptions, the existence of a constraint like ParMove in a grammar does not imply that there cannot be another constraint in the same grammar that demands the opposite in certain contextsÐafter all, violability of constraints is one of the crucial assumptions of OT. Notes For helpful comments and discussion, I would like to thank Birgit Alber, Artemis Alexiadou, Chris Collins, Gisbert Fanselow, John Frampton, Werner Frey, Hans-Martin GaÈrtner, GuÈnther Grewendorf, Brigitta Haftka, Hubert Haider, Hans Kamp, Shin-Sook Kim, GeÂraldine Legendre, Andre Meinunger, JuÈrgen Pafel, Inger Rosengren, Joachim Sabel, Vieri SamekLodovici, Uli Sauerland, Peter Staudacher, Arnim von Stechow, Markus Steinbach, Wolfgang Sternefeld, Sten Vikner, Chris Wilder, Ilse Zimmermann, and the audiences of the SFB 340 workshop on clause structure at TuÈbingen University (1996), the ZAS colloquium at Humboldt University, Berlin (1997), and the MLC workshop at Potsdam University (1997). The ®rst version of this chapter (MuÈller 1997a) was written during my stay at Potsdam University in the winter of 1996±97, and I am grateful to the Innovationskolleg ``Formale Modelle kognitiver KomplexitaÈt'' for its generous support. 1. Note that since ParMove correlates two nonadjacent steps in the derivation (here encoded as levels), this constraint belongs to the class of ``global rules'' (in the sense of Lako¨ 1971), just as Chomsky's (1981) Projection Principle does. 2. Throughout this chapter, I assume as background the clause structure argued for in Chomsky 1995, with CP dominating TP, TP dominating vP, and vP dominating VP. This clause structure will later be extended by adding a functional projection related to pronouns.

Order Preservation, Parallel Movement, Emergence of the Unmarked

305

3. This second strategy does indeed roughly correspond to construction-speci®c predecessors of ParMove that can be found in the literature. Thus, Lako¨ (1971), Kroch (1974), Reinhart (1983), Huang (1982), and Lasnik and Saito (1992) (also see Huang 1995:165±166 for further references) develop constraints on relative scope assignment that in one way or another incorporate the gist of ParMove. Their proposals are con®ned to the interaction of quanti®ed XPs (see section 10.2.7), and can be interpreted as requiring multiple quanti®er raising to maintain S-structure order. (Lako¨ 's analysis is formulated in the framework of generative semantics, where D-structure plays the role of logical form, and accordingly, his version of the constraint requires that relative scope at D-structure must be preserved at surface structure.) Clearly, these approaches are not in con¯ict with the existence of order-changing movement operations as in (2), but they do not generalize to other instances of order-preserving movement either. Similar conclusions apply in the case of Watanabe's (1992) constraint regulating (anti-) superiority, and in the case of Haegeman's (1995) and Meinunger's (1995) constraints on relation preservation in A-chains. 4. Not much depends on this in the present context, though. The derivational constraints that will be discussed below, including ParMove, do of course presuppose that the candidates they apply to are derivations, but these constraints could all straightforwardly be reformulated as representational constraints. However, see MuÈller 1997b for some arguments in support of a truly derivational view of competing candidates. 5. See Grimshaw 1997; Legendre, Smolensky, and Wilson 1998; and others. This constraint is an amalgamation of Chomsky's (1995) economy constraints Last Resort and Procrastinate. 6. Also compare Grimshaw's (1997) OpSpec and Legendre, Smolensky, and Wilson's (1998) *Q, which yield partially similar e¨ects. Note furthermore that the formulation of WhCrit here is a simpli®cation; I believe that the constraint must be decomposed into two conjunctive statements (one about wh-phrases, and one about C‰‡whŠ nodes) if more intricate whconstructions are taken into account; see MuÈller 1997b. 7. Note that PureEP not only blocks adjunction to a maximal projection, but also adjunction to C. Ultimately, the latter consequence is responsible for blocking V raising to C via adjunction in embedded questions in English in Grimshaw's approach. Assuming that V raising to C can be substitution in root clauses (where C is arguably radically empty), in contrast to embedded clauses (where C bears selection features), V raising in root questions does not have to violate PureEP. This can ultimately account for the well-known root/embedded asymmetry concerning V raising in wh-questions that Grimshaw (1997) captures by restricting PureEP to embedded contexts. See Rizzi and Roberts 1989 for related discussion of C in root and embedded contexts; Vikner (chap. 14, this volume) for a further application of a constraint like PureEP; Grimshaw 1999 for a factorization of her PureEP into more elementary subconstraints; and Legendre 1999 for the postulation of a whole class of *Structure constraints of which PureEP in its present formulation would be just one instantiation. 8. The question arises of how the lack of superiority e¨ects in a language like German (see, e.g., Haider 1993) is to be explained in this approach; cf. (i). (i) a. (Ich weiû I know b. (Ich weiû I know

nicht) [CP not nicht) [CP not

C was2 gelesen hat ] wer1 whonom whatacc read has t2 gelesen hat ] was2 C wer1 whatacc whonom read has

306

Gereon MuÈller

The problem is that German is otherwise similar to English with respect to the constraints regulating wh-movement (but see MuÈller 1997b for discussion of some di¨erences, mainly pertaining to partial wh-movement in German). I will not try to develop a full-¯edged account of the lack of superiority e¨ects in German here. It may su½ce to point out that this problem arises under most approaches to superiority in English, including one in terms of the original Superiority Condition (or a variant thereof, like Chomsky's (1995) MLC). Fanselow (1997) argues that German does not exhibit superiority e¨ects with arguments that are clause-mates because wh-movement of, say, the object NP in (ib) may apply from a scrambling position in front of the subject. Such an approach, which essentially ties the absence of regular superiority e¨ects in German to the relative freedom of clause-internal word order, could presumably be reconciled with the ParMove account developed here. 9. There are some apparent exceptions to this generalization where the order of the moved whphrases is not completely ®xed (re¯ecting D-structure optionality), or additional factors seem to intervene. See Billings and Rudin 1994 for further discussion. 10. Of course, more intricate notions of c-command could technically do the job under the NP-adjunction hypothesis; see Watanabe 1992 and Kayne 1994, among others. (Note incidentally that whereas SpecC is occupied by the leftmost wh-phrase in Rudin 1988, it is occupied by the rightmost wh-phrase on Kayne's (1994) assumptions.) 11. First, adjunction of one argument NP2 to another argument NP1 (that does not dominate NP2 in the premovement structure) invariably violates the Strict Cycle Condition of Chomsky (1995) as an instance of syntactic lowering, irrespective of whether NP1 is in situ or in SpecC at the point at which adjunction applies (technically, the reason is that an embedded feature is checked by overt movement). A related problem is posed by the fact that adjoining NP2 to NP1 yields a con®guration in which c-command of t2 in the strict sense is not available anymore (cf., e.g., Koster 1987 and Stechow and Sternefeld 1988). Furthermore, Stechow (1996) and Beck (1996) show that genuine wh-cluster formation by NP-adjunction poses problems for a strictly compositional interpretation of multiple questions, in contrast to the situation with (13). Finally, it turns out that assuming this type of acyclic adjunction operation to be possible leads to substantial overgeneration problems with (non-wh) scrambling in free word order languages (see MuÈller 1998: chap. 4). 12. Note that (19b) also shows the futility of arguing that the presence of the indirect object in (18a) turns VP into an island for extraction. 13. Here and in what follows, I abstract away from the fact that Scandinavian object shift depends on movement of the main verb to a higher position. This requirement could be integrated into PronCrit, but such a move might raise problems for a uni®ed account of object shift in Danish and pronoun movement in German, which does not seem to depend on V raising (see below). The reason for this di¨erence will have to be left open in this chapter. 14. Note that PureEP only rules out adjunction to CP, not adjunction to pP, so the fact that Danish does not exhibit multiple wh-movement is unproblematic here. 15. As observed by Vikner (1990), (21b) is the only way a sentence can be created on the basis of this lexical material, but it is not completely acceptable. The question arises of how the deviance of (21b) is to be accounted for. The key to solving this problem might be Vikner's (1990: sec. 4.3.3.1) observation that (21b) improves signi®cantly if the pronoun is stressed. Under present assumptions, stressing the pronoun renders it irrelevant for PronCrit. Thus, the optimal candidate here might in fact not be one that violates PronCrit, but one that respects

Order Preservation, Parallel Movement, Emergence of the Unmarked

307

both ParMove and PronCrit at the cost of the violation of a lower-ranked faithfulness constraint that prohibits pronoun ``strengthening''Ðthat is, turning an initially weak pronoun into a strong pronoun in the course of the derivation. (The question marks in (21b) should then not be interpreted as signaling intermediate acceptability, but rather as signaling ill-formedness if den is weak (unstressed), and complete well-formedness if den is strong.) In this view, the optimal candidate of the competition underlying (21) (with a weak pronoun in the numeration) would be identical to the optimal candidate of a minimally di¨erent competition in which the pronoun was strong to begin with. See Legendre, Smolensky, and Wilson 1998 and Bakovic and Keer (chap. 4, this volume) for discussion of this concept of syntactic neutralization of di¨erent input speci®cations. 16. Another derivation D3 that generates the ill-formed sentence (ii) must also be ruled out. t1 t 2 (ii) *Peter viste Marie1 den2 jo indeed Peter showed Marie it Here, Marie1 and den2 both undergo object shift, with only two violations of Stay arising (and no violation of either PronCrit or ParMove). However, such a derivation violates a general and high-ranked constraint (which we may call FMatch) that demands that a can be moved to a position b only if a is equipped with features that match those of b; perhaps this constraint should in fact be viewed as inviolable and belonging to Gen. Among other things, FMatch ensures that there is no movement of [ wh] phrases to SpecC‰‡whŠ , and it also implies that there can be no object shift of nonpronouns to the p domain in Danish. Alternatively, FMatch could be built into the formulation of PronCrit itself, by adding a statement such as ``. . . and only weak pronouns can be in the domain of p.'' The choice between the two options depends on a number of further assumptions, and the di¨erence is in any case a subtle one. 17. A proposal concerning the trigger of Icelandic object shift that can straightforwardly be reconciled with this general view is made by Collins and ThraÂinsson (1996). They suggest that object shift of nonpronominal NPs in Icelandic is an instance of optional A-movement to a Case position, and that the pertinent feature for object shift can be either strong or weakÐstrength of this feature would then correspond to the presence of F, in the framework adopted here. 18. As an alternative to optionality of feature instantiation, the optionality of object shift of lexical NPs in Icelandic might also be accounted for by an appropriately (globally) de®ned constraint tie that would involve FCrit and Stay. See Prince and Smolensky 1993 for the concept, and Pesetsky 1998, Ackema and Neeleman 1998, Legendre, Smolensky, and Wilson 1998, and MuÈller 1997b for some applications in syntax. 19. Note that due to a high-ranked (or inviolable) constraint like FMatch, we can disregard from the start derivations that involve an additional and unmotivated movement of the direct object, as in the candidates D3 , D4 (both with multiple object shift), and D5 in T8 . Similar considerations apply with respect to tableau T10.10. 20. Further corroboration of this analysis is provided by Vikner's (1990/1995:198±200) and Bobaljik and Jonas's (1996:212±214) observation that object shift can never cross the in situ subject in transitive expletive constructions in Icelandic; see (iii). (iii) a.

paD there b. *paD there

lauk ®nished lauk ®nished

einhver1 verkefninu2 someone the assignment verkefninu2 einhver1 t2 the assignment someone

308

Gereon MuÈller

Given that the subject NP1 asymmetrically commands the object NP2 at D-structure, the contrast in (iii) follows directly from ParMove, along the same lines as (23b). Having said that, it seems that the empirical evidence is not quite as clearcut as one might wish. For instance, a di¨erent assessment of data involving object shift in transitive expletive constructions in Icelandic is reported in Collins 1997:18. 21. This analysis can be straightforwardly extended to Case-driven movement in double object constructions. If all three arguments undergo overt movement to a Case position, ParMove requires them to reassemble in their D-structure oder. If Case-driven movement at S-structure does not apply to one (or two) of the objects (because the relevant feature is not present), as argued for German in MuÈller and Sternefeld 1994, a ranking FCrit/Case Filter g ParMove correctly ensures a minimal violation of ParMove with the moved argument(s). 22. All generalizations in this subsection hold for weak, unstressed pronouns only; both strong (stressed) and clitic pronouns behave di¨erently. Simplifying somewhat, strong pronouns behave like nonpronominal NPs, as in Danish. They do not undergo movement to a Wackernagel position, and they do not exhibit ®xed-order e¨ects, like nonpronominal NPs (see (36); i.e., all the data in this subsection are acceptable if the pronouns are stressed). Clitic pronouns, on the other hand, must phonologically attach to a lexical host by de®nition, but the landing site does not have to be in the Wackernagel domain, and the order of the pronouns becomes much less rigid if cliticization is involved. 23. Thus, a subject NP in German will only optionally have a feature speci®cation that forces raising to SpecT, by a Case- or EPP-related markedness constraint. Alternatively, a tie of this constraint and Stay could be envisaged. 24. It has sometimes been claimed on the basis of markedness considerations that the relative D-structure order of direct and indirect object in German depends on the choice of verb (see, e.g., Haider 1992), and that verbs like geben (`give') can induce an order indirect object  direct object, whereas verbs like aussetzen (`expose') must induce the reverse order direct object  indirect object. In MuÈller 1999 I show that this assumption is untenable, and that issues of relative markedness do in fact support the view adopted here, namely, that direct object  indirect object is the sole D-structure order in German, with di¨ering unmarked S-structure orders resulting from scrambling operations triggered by optimization needs. 25. There are two further di¨erences that I will not pursue here. As noted, Scandinavian object shift depends on V movement, which German pronoun fronting does not seem to do; and pronoun fronting in German and pronominal object shift in Scandinavian behave di¨erently with respect to the status as A-bar or A-movement (e.g., the former movement licenses parastic gaps, the latter does not). 26. Note that the ranking WhCrit g PronCrit is not actually forced by the empirical evidence discussed in this chapter. 27. By assumption, the subject NP Fritz stays in situ in the derivations of this candidate set; it lacks the relevant feature that triggers overt raising to SpecT. If there is such a feature in a derivation D6 , D6 will belong to a di¨erent candidate set, and therefore cannot block D1 even if it fares better with respect to ParMove. If, however, Fritz is moved in accordance with ParMove to a clause-initial position without bearing an appropriate feature, a fatal violation of the high-ranked (or inviolable) FMatch will result. Note also that although D5 is excluded as an ill-formed derivation, the S-structure string as such is indeed well formed (see (31));

Order Preservation, Parallel Movement, Emergence of the Unmarked

309

however, it is generated by derivation D6 and therefore involves three argument traces (i.e., three Stay violations). 28. Needless to say, some problems still have to be left open under this analysis. For instance, as noted by Werner Frey (personal communication), there are some constructions that exhibit the same ®xed order of fronted pronouns, but where it is less obvious that this order is the D-structure one. This holds, for example, for coherent in®nitive constructions. Consider (iv). daû es2 ihm3 that itacc himdat b. ?*daû ihm3 es2 that himdat itacc

(iv) a.

keiner t3 [a t2 zu lesen ] empfohlen no-onenom to read recommended keiner t3 [a t2 zu lesen ] empfohlen no-onenom to read recommended

hat has hat has

Given that the indirect object NP3 is base generated in the matrix clause in (iv), and the direct object NP2 in the embedded in®nitive a, it seems that ParMove would incorrectly predict the ill-formed S-structure order in (ivb) rather than the well-formed S-structure order in (iva). Thus, under present assumptions, a ParMove account of the data in (iv) seems to presuppose a base-generation approach to coherent in®nitives in German that does not postulate the presence of an a constituent in (iv) (see Haider 1993, among others), so that NP3 does not asymmetrically c-command NP2 in the base. An alternative strategy would be to alter certain premises so far taken for granted. For instance, it might be that the data in (iv) actually indicate the activity of an output/output faithfulness constraint, such that the pronoun order in complex clauses must re¯ect the pronoun order in simple clauses, which in turn is derived from ParMove. For reasons of space, I will not pursue these matters here. 29. This assessment of the data is in line with the ®ndings of Moltmann (1991), Pafel (1993), Beck (1996), and BuÈring (1996). Frey (1989) and Heck (1999) systematically permit an additional reading in sentences that involve quanti®er scrambling, which corresponds to the relative order of quanti®ers after reconstruction (also see Aoun and Li 1993). This more liberal variety may require some additional assumptions, but does not undermine the claim that order preservation is a prevalent property of quanti®er scope that lends itself to an OT analysis; see Heck 1999. 30. Of course, a thorough execution of this general approach will unravel various complications, will have to integrate further assumptions, and will therefore gain complexity. Still, I think that the basic proposal of a low-ranked and violable constraint requiring order preservation with relative scope would be una¨ected by these complications. Also see Vikner 1997 and Heck 1999 for elaborate optimality-based approaches to relative scope that incorporate the gist of a constraint like ParMove. 31. Note that the movement types discussed in this chapter do not exhaust the list of phenomena that might suggest an approach in terms of ParMove; see MuÈller 1997a for some further possible applications. References Ackema, Peter, and Ad Neeleman. 1998. Optimal Questions. Natural Language and Linguistic Theory 16: 443±490. Aoun, Joseph, Norbert Hornstein, and Dominique Sportiche. 1981. On Some Aspects of Wide Scope Interpretation. Journal of Linguistic Research 1: 69±95. Aoun, Joseph, and Audrey Li. 1993. Syntax of Scope. Cambridge, Mass.: MIT Press.

310

Gereon MuÈller

Beck, Sigrid. 1996. Wh-Constructions and Transparent Logical Form. Doctoral dissertation, UniversitaÈt TuÈbingen. Billings, Loren, and Catherine Rudin. 1994. Optimality and Superiority: A New Approach to Overt Multiple Wh Ordering. In JindrÏich Toman (ed.), Formal Approaches to Slavic Linguistics 3: The College Park Meeting. Ann Arbor: Michigan Slavic Publications. Bobaljik, Jonathan, and Dianne Jonas. 1996. Subject Positions and the Roles of TP. Linguistic Inquiry 27: 195±236. BuÈring, Daniel. 1995. The 59th Street Bridge Accent. Doctoral dissertation, UniversitaÈt TuÈbingen. BuÈring, Daniel. 1996. Interpretation and Movement: Toward an Economy-Theoretic Treatment of German ``Mittelfeld'' Word Order. Ms., UniversitaÈt zu KoÈln. BuÈring, Daniel. 1999. Let's Phrase It! Ms., University of California, Santa Cruz. [Forthcoming in Gereon MuÈller and Wolfgang Sternefeld (eds.), Competition in Syntax. Berlin: Mouton de Gruyter.] Cardinaletti, Anna, and Ian Roberts. 1991. Clause Structure and X-Second. Ms., University of Venice and University of Geneva. Choi, Hye-Won. 1996. Optimizing Structure in Context. Doctoral dissertation, Stanford University, Stanford, Calif. Chomsky, Noam. 1973. Conditions on Transformations. In A Festschrift for Morris Halle, ed. Stephen Anderson and Paul Kiparsky, 232±286. New York: Academic Press. Chomsky, Noam. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, Noam. 1986. Barriers. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1993. A Minimalist Program for Linguistic Theory. In The View from Building 20, eds. Kenneth Hale and Samuel Jay Keyser, 1±52. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1995. Categories and Transformations. Chap. 4 of The Minimalist Program, 219±394. Cambridge, Mass.: MIT Press. Collins, Chris. 1997. Local Economy. Cambridge, Mass.: MIT Press. Collins, Chris, and HoÈskuldur ThraÂinsson. 1996. VP-Internal Structure and Object Shift in Icelandic. Linguistic Inquiry 27: 391±444. Deprez, Viviane. 1994. Parameters of Object Movement. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling: Movement and Non-Movement Approaches to Free Word-Order Phenomena, 101±152. Berlin: Mouton de Gruyter. Diesing, Molly. 1997. Yiddish VP Order and the Typology of Object Movement in Germanic. Natural Language and Linguistic Theory 15: 369±427. Fanselow, Gisbert. 1997. The Proper Interpretation of the Minimal Link Condition. Ms., UniversitaÈt Potsdam. Fodor, Janet. 1978. Parsing Strategies and Constraints on Transformations. Linguistic Inquiry 9: 427±473. Frey, Werner. 1989. Syntaktische Bedingungen fuÈr die Interpretation. Doctoral dissertation, UniversitaÈt Stuttgart. Grewendorf, GuÈnther. 1988. Aspekte der deutschen Syntax. TuÈbingen: Narr.

Order Preservation, Parallel Movement, Emergence of the Unmarked

311

Grewendorf, GuÈnther. 1989. Ergativity in German. Dordrecht: Foris. Grewendorf, GuÈnther, and Joachim Sabel. 1996. Multiple Speci®ers and the Theory of Adjunction. Ms., UniversitaÈt Frankfurt/Main. Sprachwissenschaft in Frankfurt Working Papers 17. Institut fuÈr deutsche Sprache und Literatur. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28: 373±422. Grimshaw, Jane. 1999. Clause Structure and Competition in Syntax. Ms., Rutgers University, New Brunswick, N.J. Haegeman, Liliane. 1995. The Syntax of Negation. Cambridge, England: Cambridge University Press. Haftka, Brigitta. 1981. Reihenfolgebeziehungen im Satz. In an Autorenkollektiv led by KarlErich Heidolph, Walter FlaÈmig, and Wolfgang Motsch (eds.), GrundzuÈge einer deutschen Grammatik, 702±764. Berlin: Akademieverlag. Haider, Hubert. 1992. Branching and Discharge. Ms., UniversitaÈt Stuttgart. Haider, Hubert. 1993. Deutsche Syntax, generativ. TuÈbingen: Narr. Heck, Fabian. 1999. Quanti®er Scope in German and Cyclic Optimization. Ms., UniversitaÈt TuÈbingen [Forthcoming in Gereon MuÈller and Wolfgang Sternefeld (eds.), Competition in Syntax. Berlin: Mouton de Gruyter.] Heim, Irene, and Angelika Kratzer. 1998. Semantics in Generative Grammar. Oxford: Blackwell. Hendrick, Randall, and Michael Rochemont. 1982. Complementation, Multiple Wh, and Echo Questions. Ms., University of North Carolina, Chapel Hill, and University of California, Irvine. Holmberg, Anders. 1986. Word Order and Syntactic Features in the Scandinavian Languages and English. Doctoral dissertation, Department of General Linguistics, University of Stockholm. Huang, Cheng-Teh James. 1982. Logical Relations in Chinese and the Theory of Grammar. Doctoral dissertation, MIT, Cambridge, Mass. Huang, Cheng-Teh James. 1995. Logical Form. In Gert Webelhuth (ed.), Government and Binding Theory and the Minimalist Program, 125±175. Oxford: Blackwell. Jacobs, Joachim. 1982. Syntax und Semantik der Negation im Deutschen. MuÈnchen: Fink. Johnson, Kyle. 1991. Object Positions. Natural Language and Linguistic Theory 9: 577±636. Kayne, Richard. 1994. The Antisymmetry of Syntax. Cambridge, Mass.: MIT Press. Keenan, Ed, and Bernard Comrie. 1977. Noun Phrase Accessibility and Universal Grammar. Linguistic Inquiry 8: 63±99. Koster, Jan. 1987. Domains and Dynasties. Dordrecht: Foris. Kroch, Anthony. 1974. The Semantics of Scope in English. Doctoral dissertation, MIT, Cambridge, Mass. Lako¨, George. 1971. On Generative Semantics. In Danny Steinberg and Leon Jakobovits (eds.), Semantics, 232±296. Cambridge, England: Cambridge University Press. Lasnik, Howard, and Mamoru Saito. 1984. On the Nature of Proper Government. Linguistic Inquiry 15: 235±289. Lasnik, Howard, and Mamoru Saito. 1992. Move a. Cambridge, Mass.: MIT Press.

312

Gereon MuÈller

Legendre, GeÂraldine. 1999. Why French Stylistic Inversion Is Optimal. Ms., Johns Hopkins University, Baltimore, Md. Legendre, GeÂraldine, Paul Smolensky, and Colin Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Lenerz, JuÈrgen. 1977. Zur Abfolge nominaler Satzglieder im Deutschen. TuÈbingen: Narr. Lenerz, JuÈrgen. 1992. Zur Syntax der Pronomina im Deutschen. Sprache und Pragmatik 29, University of Lund. Liu, Feng-Hsi. 1997. Scope and Speci®city. Amsterdam: Benjamins. May, Robert. 1985. Logical Form. Cambridge, Mass.: MIT Press. McCarthy, John, and Alan Prince. 1994. The Emergence of the Unmarked: Optimality in Prosodic Morphology. In MerceÁ GonzaÁlez (ed.), NELS 24, vol. 2, 333±379. GLSA, University of Massachusetts, Amherst. Meinunger, AndreÂ. 1995. Prominence Hierarchy and Phrase Ordering. In Artemis Alexiadou, Nanna Fuhrop, Paul Law, and Sylvia LoÈhken (eds.), Forschungsschwerpunkt Allgemeine Sprachwissenschaft, Sprachtypologie und Universalienforschung der FoÈrderungsgesellschaft Wissenschaftliche Neuvorhaben mbH (FAS) Working Papers in Linguistics 2, 95±121. Berlin: Zentrum fuÈr Allgemeine Sprachwissenschaft (ZAS). Moltmann, Friederike. 1991. Scrambling in German and the Speci®city E¨ect. Ms., MIT, Cambridge, Mass. Mulders, Iris. 1997. Mirrored Speci®ers: Minimal Link E¨ects in Multiple Speci®ers. Ms., Utrecht University. MuÈller, Gereon. 1995. A-Bar Syntax. Berlin: Mouton de Gruyter. MuÈller, Gereon. 1997a. Parallel Movement. In Franz d`Avis and Uli Lutz (eds.), Zur Satzstruktur im Deutschen, 171±214. Sonderforschungsbereich 340 Working Papers 90. TuÈbingen. MuÈller, Gereon. 1997b. Partial Wh-Movement and Optimality Theory. Linguistic Review 14: 249±306. MuÈller, Gereon. 1998. Incomplete Category Fronting: A Derivational Approach to Remnant Movement in German. Dordrecht: Kluwer. MuÈller, Gereon. 1999. Optimality, Markedness, and Word Order in German. Linguistics 37: 777±818. MuÈller, Gereon, and Wolfgang Sternefeld. 1994. Scrambling as A-Bar Movement. In Studies on Scrambling, ed. Norbert Corver and Henk van Riemsdijk, 331±385. Berlin: Mouton de Gruyter. Pafel, JuÈrgen. 1993. Scope and Word Order. In Syntax, vol. 1, ed. Joachim Jacobs, Arnim von Stechow, Wolfgang Sternefeld, and Theo Vennemann, 867±879. Berlin: de Gruyter. Pesetsky, David. 1982. Paths and Categories. Doctoral dissertation, MIT, Cambridge, Mass. Pesetsky, David. 1998. Some Optimality Principles of Sentence Pronunciation. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 337±383. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics.

Order Preservation, Parallel Movement, Emergence of the Unmarked

313

Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Reinhart, Tanya. 1983. Anaphora and Semantic Interpretation. London: Croom Helm. Richards, Norvin. 1997. What Moves Where When in Which Language? Doctoral dissertation, MIT, Cambridge, Mass. Rizzi, Luigi. 1991. Residual V/2 and the Wh-Criterion. Ms., Universite de GeneÁve. Rizzi, Luigi, and Ian Roberts. 1989. Complex Inversion in French. Probus 1: 1±30. Roberts, Ian. 1995. Object Movement and Verb Movement in Early Modern English. In Studies in Comparative Germanic Syntax, ed. Hubert Haider, Susan Olsen, and Sten Vikner, 269±284. Dordrecht: Kluwer. Rudin, Catherine. 1985. Aspects of Bulgarian Syntax: Complementizers and Wh-Constructions. Columbus, Ohio: Slavica. Rudin, Catherine. 1988. On Multiple Questions and Multiple Wh Fronting. Natural Language and Linguistic Theory 6: 445±501. Schmidt, Claudia Maria. 1992. Eine mediale Klitiposition. Ms., UniversitaÈt zu KoÈln. Stechow, Arnim von. 1996. Against LF Pied-Piping. Natural Language Semantics 4: 57±110. Stechow, Arnim von, and Wolfgang Sternefeld. 1988. Bausteine syntaktischen Wissens. Opladen: Westdeutscher Verlag. Thiersch, Craig. 1978. Topics in German Syntax. Doctoral dissertation, MIT, Cambridge, Mass. Uszkoreit, Hans JuÈrgen. 1984. Word Order and Constituent Structure in German. Doctoral dissertation, University of Texas, Austin. Vikner, Sten. 1985. Parameters of Binder and of Binding Category in Danish. Working Papers in Scandinavian Syntax 23: 1±61. Vikner, Sten. 1990. Verb Movement and the Licensing of NP Positions in the Germanic Languages. Doctoral dissertation, Universite de GeneÁve. Revised version appeared 1995 as Verb Movement and Expletive Subjects in the Germanic Languages. Oxford: Oxford University Press. Vikner, Sten. 1994. Scandinavian Object Shift and West Germanic Scrambling. In Norbert Corver and Henk van Riemsdijk (eds.), Studies on Scrambling, 487±517. Berlin: Mouton de Gruyter. Vikner, Sten. 1997. The Interpretation of Object Shift, Optimality Theory, and Minimalism. Working Papers in Scandinavian Syntax 60, 1±24. Department of General Linguistics, University of Lund. Watanabe, Akira. 1992. Wh-in-Situ, Subjacency, and Chain Formation. Ms., MIT. MIT Occasional Papers in Linguistics 2. Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Zwart, Jan Wouter. 1993. Dutch Syntax: A Minimalist Approach. Doctoral dissertation, University of Groningen.

Chapter 11 Crosslinguistic Typologies in Optimality Theory

Vieri Samek-Lodovici

A striking feature of Optimality Theory (OT) is its intrinsic crosslinguistic scope: any language-speci®c analysis always entails a crosslinguistic typology through constraint reranking (Prince and Smolensky 1993). This work investigates the structure of the entailed typology. I will begin by examining how in®nite candidate sets can be reduced to ®nite typologies through the elimination of all harmonically bounded candidatesÐthat is, of candidates that are suboptimal on all rankings (Prince and Smolensky 1993; Samek-Lodovici and Prince 1999). Next, I will investigate the role of inputs, showing how the related typologies grow monotonically as the number of constraints that apply nonvacuously with respect to an input increases. Finally, I will consider OT typologies across inputs, and the role of ranking consistency in determining them. For the sake of concreteness, I will develop the above concepts and results with respect to the syntax of new-information focus, deriving the typology shown in table 11.1. This involves languages with SVO and VSO basic word order that either lack structural focus, or realize structural focus at the left or right edge of VP, or ®nally restrict structural focus to nonsubject constituents. The basic word order of a language, the presence or absence of structural focus, and its potentially restricted nature will all emerge as epiphenomenal e¨ects from the interaction between the constraints Subject and Stay, governing phrase structure and movement (Grimshaw 1993, 1997), and the focus-alignment constraints AF-L and AF-R, governing the syntactic realization of focused phrases (Grimshaw and Samek-Lodovici 1995, 1998; Samek-Lodovici 1996). 11.1

Constraints and Structural Assumptions

To keep the typology su½ciently simple, the analysis is restricted to languages showing SVO or VSO basic word order, where ``basic word order'' stands for the order found in a matrix clause whose constituents other than the clause itself are neither foci nor discourse topics. This is the order displayed in ``out-of-the-blue''

316

Vieri Samek-Lodovici

Table 11.1 Structural focus typology None

Right

Left

Partial

SVO

Kera

Italian

Aghem

English (optionally)

VSO

Scottish Gaelic

Spanish (VSO varieties)

Podoko

Ð

utterances, or as an answer to the question ``What happened?'' (Vallduvõ 1992; Zubizarreta 1992; Fassi Fehri 1993; Pinto 1997; Costa, chap. 7, this volume). I will also assume that when the whole clause is focused its constituents are unmarked for focus (for a di¨erent assumption, see Costa, chap. 7, this volume). The restriction to SVO and VSO languages simpli®es the analysis by excluding all head-®nal languages with OV order, whose derivation requires additional constraints, including constraints governing X 0 -theory such as HeadLeft, HeadRight, and SpecifierLeft (Grimshaw 1997; Keer 1996; Keer and Bakovic 1997), or caserelated constraints, such as those proposed in Costa (1998b; chap. 7, this volume), Legendre 1998, or Samek-Lodovici 1996, 1998.1 Following Vikner (chap. 14, this volume) and contra Grimshaw (1993, 1997), I will assume that IP is always projected, and, in addition, that V raises to Io in all languages examined, except English. Therefore, in all syntactic representations used here, the VP always begins right after the verb V, and any constituent following V lies within VP, while any constituent preceding it is outside VP. In particular, an SVO clause has the structure shown in (1a), with the subject S in SpecIP and the object O in situ, whereas a VSO clause has the structure in (1b), with the subject in SpecVP and the object in situ. (1) a. [ ip S V [vp t t O ]] b. [ ip V [vp S t O ]] Any argument outside or adjoined to VP is assumed to have moved from its basegenerated VP-internal position (but see Neeleman and Weerman 1999 for an alternative view). Traces are represented by t and are easily identi®able as subject, object, or verbal traces according to their position. For example, the VP-initial trace in (1a) is in SpecVP and is left by the subject, whereas the second trace, in the middle of the VP, is left by the raising verb. A third trace will appear in object position whenever the object is moved. This interpretation holds for all the structures considered in this chapter. The choice between the SVO and VSO basic word order is governed through the constraints Stay and Subject (Grimshaw 1993, 1997; see also MuÈller and CostaÐ chaps. 10 and 7 respectively in this volumeÐfor alternative analyses of basic word order).

Crosslinguistic Typologies in Optimality Theory

317

The constraint Stay penalizes movement, and is violated once by each trace (cf. movement economy, Chomsky 1992, 1995). See (2). (2) Stay: Traces are not allowed. Failed by traces. The Subject constraint has its roots in the extended projection principle (Chomsky 1982) and requires that the highest A-position of the clause be structurally realized; see (3). In this chapter this position is always SpecIP, and Subject is violated whenever SpecIP is not ®lled by an overt subject or a trace. (3) Subject: The highest A-speci®er of a clause must be structurally realized. Failed when it is left structurally unrealized. Objects and indirect objects are assumed unable to satisfy Subject even when moved to SpecIP, although in a more ®ne-grained model the suboptimal status of these structures should follow from the interaction with additional constraints governing thematic linking and case assignment. (See for example the constraints proposed in Legendre, Raymond, and Smolensky 1993; Legendre 1998; Samek-Lodovici 1996; Costa 1998b and chap. 7, this volume; Bresnan, 2000.) As for structural focus, I will exclusively consider constituents focused through question-and-answer (QA) pairs and examine the syntactic position they have there (the relevant data follow in the appendix). This operational criterion ensures that we are always dealing with the same kind of focus in all the languages investigated here. This kind of focus is most often considered an instance of presentational or newinformation focus (Rochemont 1986; Dik et al. 1981; see also Choi and CostaÐ chaps. 6 and 7 respectivelyÐin this volume). Rooth, however, analyzes QA pairs in terms of the same semantic representation proposed for contrastive focus (Rooth 1985:10±15). I will leave open this important issue and simply assume that the focusalignment constraints introduced below apply to whichever notion of ``focus'' will turn out to be the correct one for QA pairs. Focused constituents are directly targeted by the constraints AF-L(eft) and AFR(ight), where AF stands for ``Align Focus.'' These constraints require focused phrases to align respectively with the left and right edge of VP, as stated in (4) (Grimshaw and Samek-Lodovici 1995, 1998; Samek-Lodovici 1996, 1998). (4) AF-L, AF-R: Align focused constituents with the left/right edge of VP. Failed by misaligned foci. As alignment constraints, however, AF-L and AF-R are more precisely de®ned along McCarthy and Prince's (1993) theory of general alignment, as in (5). The de®nition of AF-L, for example, states that for any focused phrase XPf , there is a VP such that the left edge of XPf matches the left edge of VP.

318

Vieri Samek-Lodovici

(5) AF-L ˆ Align (XPf , Left, VP, Left) AF-R ˆ Align (XPf , Right, VP, Right) Note that a focused subjectÐhenceforth represented as S f Ðalready satis®es AF-L when lying in SpecVP, because its left edge is already aligned with the left edge of VP, as shown in (6a). A focused object insteadÐrepresented as Of Ðcan satisfy AF-L only by left-adjoining to VP, as shown in (6b) and (6c), because either a subject or its trace intervenes between the in situ position of the object and the left VP edge. O] (6) a. V [vp S f t b. V [vp Of [vp S t t]] c. S V [vp Of [vp t t t]] Likewise, to satisfy AF-R in a transitive clause, a focused subject must right-adjoin to VP to avoid an intervening object O, as in (7a). The same holds for focused objects when they are followed by an in situ indirect object IO, as shown in (7b). (7) a. V [vp [vp t t O ] Sf ] b. S V [vp [vp t t t IO ] Of ] The right-adjunction representations shown in (7) are challenged by works appearing in the wake of Kayne's (1994) antisymmetry monograph, such as Reinhart 1995; OrdoÂnÄez 1995, 1997; Costa 1997a as well as chap. 7, this volume; Pinto 1997; Zubizarreta 1998. Under an antisymmetric perspective, a SpecVP subject S f in a ditransitive construction could satisfy AF-R only through leftward scrambling of all the internal arguments separating it from the right edge of VP, as shown in (8) (traces would of course not count as obstacles to alignment). (8) . . . O IO [vp S f t t t ] A full discussion of this alternative would require a separate publication. The interested reader will ®nd some of the reasons I opted for a right-adjunction representation explained at length in note two. For a critical review of the antisymmetry hypothesis, see also BuÈring and Hartmann 1994 and Ackema and Neeleman 1999. Another aspect of AF-L and AF-R worth additional comment concerns their specifying the syntactic category targeted by focus alignment (Grimshaw, personal communication; Costa, forthcoming, personal communication). Ideally, this should follow from constraint interaction, letting it vary if necessary from language to language. Indeed, under the above de®nition, AF-R and AF-L cannot cover languages where focus aligns with the edges of IP and CP. Any solution beyond parameterizing the relevant category requires a careful study of the factors that make VP alignment suboptimal in favor of IP or CP alignment. These factors could include case adjacency, preventing leftward focus from placing phrases between the verb and the object, or a preference for alignment with the edge of extended projections (Grimshaw 1991),

Crosslinguistic Typologies in Optimality Theory

319

favoring alignment with IP and CP rather than VP. Both alternatives appear compatible with the analysis developed here, which should survive as a subcase of the larger typology determined by these solutions. Finally, note that none of the four constraints proposed above regulates the presence versus absence of structural focus, or which constituents should focus structurally and which not. Both these issues will follow from constraint interaction. 11.2

Finite Typologies from In®nite Candidate Sets

We may now use the constraints just de®ned to see how a ®nite typology emerges from an in®nite candidate set once harmonic bounding is taken into account. To reach this goal, we ®rst need to de®ne what structures compete and relative to what inputs. I will follow Grimshaw (1993, 1997) and view inputs as argument structure speci®cations, with focused arguments marked as such in the input itself (Grimshaw and Samek-Lodovici 1995, 1998). For example, a clause projected by a transitive verb where neither subject nor object is focused has the input in (9a), whereas a ditransitive clause with a focused object has the input in (9b). (9) a. hV(x, y)i b. hV(x, y, z), y ˆ focusi For each input, Gen generates an in®nite set of structures; all structures are assumed to parse all the arguments speci®ed in the input, and, as mentioned, all raise the verb to Io . The generated candidate set includes all the possible X 0 -projections that can be generated above VP, and all the possible movement operations involving the verb or its arguments. An example of the kind of candidates generated with respect to the input hV(x, y)i follows in (10). These include the structure for the VSO basic word order in (10a), that for the SVO basic word order in (10b), structures with left and right VP adjunction of the subject in (10c) and (10d), structures with left adjunction of the object in (10e), a structure where the verb further raises from Io to Co in (10f ), and any other conceivable structure along these lines, as symbolized by (10g). (10) a. V [vp S b. SV [vp t c. V [vp [vp t d. V [vp S [vp t e. V [vp O [vp S f. V [ ip S t [vp t g. ... ... ...

t t t t t t

O] O] O] S] O]] t]] O]] ...

320

Vieri Samek-Lodovici

Whether a constituent is speci®ed as focused in the input or not does not a¨ect the composition of the candidate set. Inputs di¨ering in the number of arguments, on the other hand, yield distinct candidate sets, since all members of the same candidate set share the arguments speci®ed in the input. De®ning an invariant candidate set across inputs is possible, but requires additional constraints to govern the relation between the number of arguments in the input and those overtly realized by each competing structure. See, for example, Parse and FullInterpretation in Grimshaw 1993, 1997; Grimshaw and Samek-Lodovici 1995, 1998; Samek-Lodovici 1996. We may now examine the crosslinguistic typology associated with any given input. Following Samek-Lodovici and Prince 1999, let us call any structure selected as optimal under at least one constraint ranking a potential winner for that input, and any harmonically bounded structure a loser for that input (a brief explanation of harmonic bounding follows below). The linguistic typology associated with an input is then equal to the set of potential winners for that input3 (Prince and Smolensky 1993; Samek-Lodovici and Prince 1999). Consider for example the crosslinguistic typology for the input hV(x, y)i with respect to the candidate set examined in (10). This is shown in tableau T11.1. The constraints AF-R and AF-L apply only vacuously, because no argument is focused; therefore only Subject and Stay play a role in determining the potential winners. Crucially, all structures except (a) and (b) are harmonically boundedÐthat is, their violation pro®le includes all the violations of some other candidate plus some additional violations. For example, (b) and (c) both violate Stay twice, but (c) also violates Subject, which (b) satis®es through its SpecIP subject. Therefore, (b) harmonically bounds (c), which entails that (c) will lose against (b) under any constraint Tableau T11.1 Harmonic bounding for input I1 ˆ hV(x, y)i

a.

V

[VP S t O]

b.

S V

[VP t t O]

Stay

*

* **

*

**

V [VP S [VP t t O]]

*

**

V [VP O [VP S t t ]]

*

**

7 c.

V [VP

7 d. 7 e.

[VP t t O] S ]

Subject

7 f. V [ IP S t

[VP t t O]]

7 g.

...

...

*** (*)

*(*. . .)

AF-L

AF-R

Crosslinguistic Typologies in Optimality Theory

321

ranking (Prince and Smolensky 1993). The same holds for all structures marked with ``8,'' including the in®nite conceivable structures under (g): they all in fact necessarily involve at least one more movement operation than (a) and (b), and therefore one additional Stay violation, which is su½cient to make them harmonically bounded by either (a) or (b) (potential additional violations are shown in parentheses). The set of potential winners for the input hV(x, y)i thus includes only the two structures in (11), respectively expressing VSO and SVO basic word order.4 (11) a. V [vp S t O ] b. S V [vp t t O ] The size of the crosslinguistic typology associated with the input hV(x, y)i is thus much smaller than the theoretical upper bound of 4! ˆ 24 languages and is determined only by the constraints applying nonvacuously. This does not imply that the set of potential winners associated with an input is necessarily ®nite. For example, if Stay were not part of the constraints under consideration, any structure with the subject in SpecIP would be a potential winner in tableau T11.1, no matter how many movement operations it involved. However, even in the case of in®nite potential winners, harmonic bounding permits us to derive the set of structures predicted to be grammatical relative to a speci®c input and independently from any speci®c constraint ranking. Harmonic bounding thus makes it possible to test the crosslinguistic implications of OT analyses and fully exploit the predictive power of the theory. 11.3

Monotonic Typological Enrichment

Let us now consider the input hV(x, y), x ˆ focusi, which focuses subjects of transitive verbs and di¨ers minimally from the previously examined input. The presence of focus allows for the nonvacuous application of the two focus constraints. As tableau T11.2 shows, the set of potential winners determined by harmonic bounding now includes three structures. The ®rst two, in (a) and (b), are the same structures selected as potential winners for the previous input. The third is structure (c). This structure satis®es AF-R by right-aligning the focused subject, whereas (a) and (b) fail AF-R. Therefore, under this input, (a) and (b) no longer harmonically bound (c). All the remaining structures are harmonically bounded by (a), (b), or (c), and are therefore losers. The overall typology for the input hV(x, y), x ˆ focusi thus consists of the three potential winners in (12). (12) a. V [vp S f t O ] b. S f V [vp t t O ] c. V [vp [vp t t O ] S f ]

322

Vieri Samek-Lodovici

Tableau T11.2 Harmonic bounding for input I2 ˆ hV(x, y), x ˆ focusi

a.

V

[VP Sf t O]

b.

Sf V

[VP t t O]

c.

V [VP

[VP t t O] Sf ]

Subject

Stay

*

*

AF-L

AF-R *

**

*

*

**

*

*

7 d.

V [VP Sf [VP t t O]]

*

**

*

7 e.

V [VP O [VP Sf t t ]]

*

**

*

7 f. V [ IP Sf t

[VP t t O]]

7 g.

...

...

(*)

***

*

*

*(*. . .)

(*)

(*)

The new typology is a superset of the previous one. The focus speci®cation in the new input caused the nonvacuous application of AF-L and AF-R, and this increased the typology monotonicallyÐthat is, without excluding the structures already present in the previous typology. As we will see, monotonic enrichment is a general property of OT typologies, as stated in the theorem in (13). (13) Monotonic enrichment: Let I and I 0 be two inputs such that (i) a set of constraints C apply vacuously relative to I and nonvacuously relative to I 0 , (ii) the violations for all other constraints remain identical for both inputs, and the potential winners for I have distinct violation pro®les. Then the set of potential winners associated with I 0 is a superset of that associated with I. Let us prove this result by showing that for any two inputs I and I 0 of the appropriate kind, any potential winner W for I remains a potential winner when evaluated with respect to I 0 . By de®nition, a potential winner W for I is optimal for some ranking R. By hypothesis, R includes a set of constraints V that apply vacuously to all candidates. Each of these constraints can be freely reranked without a¨ecting the optimal status of W, because they are satis®ed by all candidates. Let then R/V be a reranking of R where all the constraints in V are demoted below all other constraints. The ranking R/V has the shape R/V ˆ C1 . . . Cn V1 . . . Vm , where each constraint Ci applies nonvacuously and each Vk vacuously. Candidate W is optimal in R/V, because the reranking of the constraints in V does not a¨ect optimization. Note that W is also optimal with respect to the subranking R 0 ˆ C1 . . . Cn forming the initial section of R/V, since all the constraints in V make no contribution to its selection as optimal in

Crosslinguistic Typologies in Optimality Theory

323

R/V. This means that W already outperforms all other competing candidates in the initial sequence R 0 ˆ C1 . . . Cn , and that by the time the ®rst constraint in V is evaluated there are no unbeaten competitors left. Now consider the second input I 0 . Some of the constraints in V now apply nonvacuously. By hypothesis, however, the set of violations incurred by each candidate on each constraint Ci remains unaltered. It follows that the violations incurred by all structures in the initial section R 0 ˆ C1 . . . Cn in R/V remain invariant when evaluated with respect to I 0 . Therefore, W is optimal under the subranking R 0 ˆ C1 . . . Cn even with respect to I 0 . But then W is also optimal relative to the ranking R/V ˆ C1 . . . Cn V1 . . . Vm , because W already outperforms every competitor in R 0 ˆ C1 . . . Cn , independently of the new violations assigned by each nonvacuous constraint Vk . The ranking R/V then constitutes a ranking on which W is optimal relative to I 0 , and therefore W retains its potential winner status under I 0 , proving the theorem. An interesting consequence of the monotonicity theorem is that inputs that maximize the number of nonvacuous constraints give rise to more informative typologies than other inputs. The theorem in fact ensures that only these inputs can yield richer typologies. But richer typologies require more stringent constraint-domination relations to select a speci®c potential winner as optimal, making its observation by learners and analysts alike more informative toward the reconstruction of the associated grammar. Under speci®c circumstances, richer typologies can also be informative about the inputs that determine them, with potential consequences for language acquisition. For example, consider again the two inputs examined so far and their typologies, repeated in (14). Other things being equal, a child observing structure (14e) could also infer the input she ought to consider, since this structure is a potential winner only under input I2 (this property, however, does not extend to the other structures, because they are possible under both inputs). (14) Input I1 ˆ hV(x, y)i a. V [vp S t O ] b. S V [vp t t O ] Input c. d. S f e.

I2 ˆ hV(x, y), x ˆ focusi V [vp S f t O ] V [vp t t O ] V [vp [vp t t O ] S f ]

Obviously, not all inputs comply with the conditions required by the monotonicity theorem. For example, the typologies associated with inputs focusing di¨erent arguments, such as hV(x, y), x ˆ focusi and hV(x, y), y ˆ focusi, will not show a subsetsuperset relation. This relation is nevertheless predicted whenever two inputs di¨er

324

Vieri Samek-Lodovici

only for the presence in one of the two inputs of optional features relevant for only some constraints. A quick survey of the OT literature easily shows how frequently this is the case, with constraints speci®cally targeting focus, topic, prominence, agreement, speci®city, and wh-features (see Legendre et al. 1995; Grimshaw and Samek-Lodovici 1995, 1998; Choi 1996; Samek-Lodovici 1996; MuÈller 1997; Speas 1997; Ackema and Neeleman 1998; Costa 1998b; Keer and Bakovic 1997; Legendre 1998; Woolford 1995). 11.4

Harmonic Bounding and Typological Upper Bounds

Putting optionality aside, the grammatical structures that characterize a language are each optimal relative to a distinct input. A typology of the possible languages predicted by an OT analysis must thus contemplate optimal forms across a plurality of inputs. At ®rst glance, uncovering a cross-input typology of this kind appears daunting. A universal grammar consisting of N constraints could in principle distinguish N! optimal candidates per input, and therefore potentially determine (N!)M distinct languages for M distinct inputs. For example, the four constraints introduced above, once evaluated relative to three distinct inputs, would appear to predict (4!)3 ˆ 13.824 distinct languages. In practice, however, harmonic bounding and ranking consistency across inputs dramatically limit the size of cross-input typologies. In the following, we will examine how these factors slash the size of the cross-input typology for our four constraints and three inputs to just 8 possible languages. The role of harmonic bounding in drastically delimiting the size of input-speci®c typologies was already examined in section 11.2, where it respectively reduced the input-speci®c typologies for the inputs hV(x, y)i and hV(x, y), x-focusi to the two sets of potential winners repeated in (15). (15) Input I1 ˆ hV(x, y)i a. V [vp S t O ] b. S V [vp t t O ] Input c. d. S f e.

I2 ˆ hV(x, y), x ˆ focusi V [vp S f t O ] V [vp t t O ] V [vp [vp t t O ] S f ]

Let us extend the cross-input typology by considering the set of potential winners for the input hV(x, y, z), y ˆ focusi, involving a ditransitive verb and a focused object. The presence of an indirect object will help us determine whether the object focuses in situ, in which case it occurs to the left of the indirect object, or in right VP-

Crosslinguistic Typologies in Optimality Theory

325

Tableau T11.3 Harmonic bounding for input I3 ˆ hV(x, y, z), y ˆ focusi Subject

Stay

AF-L

AF-R *

a.

V

[VP S t Of IO]

*

*

*

b.

V [VP

[VP S t t

IO] Of ]

*

**

*

c.

V [VP Of [VP S t t

IO]]

*

**

*

d. S V

[VP t t Of IO]

**

*

e. S V [VP

[VP t t t

IO] Of ]

***

*

IO]

***

f. S V [VP Of [VP t t t 7 g. IO V

[VP S t Of t ]]

7 h. IO V

[VP S t t

7 i. . . . . . .

t]

... ...

Of ]

*

*

*

**

*

*

***

*

(*)

**(*. . .)

(*)

*

(*)

adjoined position, and hence to the right of the indirect object. The evaluation of the corresponding candidate set, generated in accord with the assumptions laid out in section 11.1, follows in tableau T11.3. There are six distinct potential winners, divided into two groups according to the position of the subject, which can remain in SpecVP, as in (a) to (c), or raise to SpecIP, as in (d) to (f ). In each group, the focused object may occur in situ, in rightward VP-adjoined position, or in leftward VP-adjoined position. All other generated candidates are harmonically bounded by one of these six potential winners. Consider for example candidate (g), which raises the indirect object IO to SpecIP position, leaving the focused object Of in situ. By assumption IO cannot satisfy Subject. Therefore, (g) is harmonically bounded by (b), which parallels (g) on all constraints, but also satis®es AF-R, whereas (g) violates it. Candidate (g) cannot be enhanced by focusing the object rightward either. The resulting structure, shown in (h), satis®es AF-R but violates Stay one additional time, and is thus harmonically bounded by (e). The three inputs examined so far,5 henceforth referred to as inputs I1 , I2 , and I3 , are thus respectively associated with a set of two, three, and six potential winners, as shown in (16) (brackets represent VP boundaries). (16) Input I1 ˆ hV(x, y)i V [ StO ] S V [ t tO ]

326

Vieri Samek-Lodovici

Input I2 ˆ hV(x, y), x ˆ focusi V [ Sf t O ] Sf V [ t tO] V [[ t t O ] S f ] Input I3 ˆ hV(x, y, z), y ˆ focusi V [ S t Of IO ] V [[ S t t IO ] Of ] V [ Of [ S t t IO ]] S V [ t t Of IO ] S V [[ t t t IO ] Of ] S V [ Of [ t t t IO ]] The largest set of languages that can be formed by picking one structure from each of the above sets is equal to the cartesian product of the three sets, yielding a maximum of 2  3  6 ˆ 36 distinct languages. Harmonic bounding alone thus delimits the size of the predicted typology from the theoretical upper bound of 13.824 to just 36 languages, yielding a far more precise estimate of its complexity and identifying the actual structures involved in it. This result is encoded in the de®nition of the harmonic boundary of a typology, given in (17). (17) Harmonic boundary: Given a set of constraints S and a set of inputs I, the harmonic boundary for the crosslinguistic typology determined by S relative to I is equal to the cartesian product of the sets of potential winners associated with each input in I. Though immensely more precise, even the harmonic boundary of a typology constitutes just an upper bound for the size and articulation of actual crosslinguistic typologies. The harmonic boundary in fact counts as a possible language every ntuple of potential winners hW1 . . . Wn i, formed by picking one structure from each input-speci®c potential-winner set. This overlooks the fundamental hypothesis of OT that constraint hierarchies are invariant across inputs. Under this hypothesis, potential winners derived from distinct inputs but requiring inconsistent rankings for their selection can never be members of the same language, and therefore any n-tuple including them should not be considered a possible language. In our case, taking ranking consistency into account eliminates 28 impossible languages, reducing the typology from the 36 predicted by the harmonic boundary to just 8 possible grammars, as the next two sections will show. 11.5

Cross-Input Consistency: SVO Languages

Ranking inconsistencies across inputs are best highlighted through comparative tableaux (Prince 1999), which will therefore be used throughout the following discus-

Crosslinguistic Typologies in Optimality Theory

327

Tableau T11.4 C1 A@B

B

C2

C3

C4

A

B

Tableau T11.5 Input I1 ˆ V(x, y)

A ˆ S V [t t O] @ B ˆ V [S t O]

Subject

Stay

A

B

sion. Each row of a comparative tableau refers to the competition among two structures, such as the hypothetical structures A and B in tableau T11.4. Each constraint slot is marked with the name of the structure beating the other for that speci®c constraint and is left blank if there is no winner among the two: here B beats A on C1 and C4 but loses against A on C3 . Both structures perform equally well on C2 . This kind of tableau permits us to easily check the ranking conditions allowing either structure to beat the other across a constraint hierarchy. In the case at hand, A wins only if C3 outranks C1 and C4 , whereas B wins only if either C1 or C4 outranks C3 . Note that optimal status requires beating all alternative potential winners, not just one competitor: each beaten competitor thus necessitates an additional row in the tableau. Let us begin investigating the role of ranking consistency by considering the ranking relations associated with the two potential winners for the input I1 ˆ hV(x, y)i. As we saw in section 11.2, the two focus constraints apply vacuously in this case and are therefore irrelevant. As tableau T11.5 shows, the SVO candidate A beats the VSO candidate B on Subject, because it realizes SpecIP, whereas B wins on Stay, because the subject does not move. The SVO order is thus possible only when Subject outranks Stay, while the VSO order requires the reverse ranking. We may now investigate which of the 36 languages in the harmonic boundary of the focus typology are consistent with each of these two ranking conditions. We will ®rst examine the 18 SVO-patterns compatible with the ranking Subject g Stay, and then turn to the remaining 18 VSO-patterns compatible with the reverse ranking. To facilitate exposition, I will only consider structural representations like those examined in the previous sections and postpone all the data justifying these structures to the appendix. 11.5.1 Lack of Structural Focus in SVO Languages: Kera The ®rst SVO language predicted to be possible by the above constraints lacks structural focus altogetherÐthat is, it focuses its arguments according to the same

328

Vieri Samek-Lodovici

Tableau T11.6 Input I2 ˆ hV(x, y), x ˆ focus) A ˆ Sf V [t t O]

Subject

A @ B Ð B ˆ V [[t t O] Sf ]

A

A @ C Ð C ˆ V [ Sf t O]

A

Stay

AF-R

AF-L

B C

C

SVO basic word order followed by unfocused constituents. The corresponding pattern, attested in the Chadic language Kera (Ebert 1979; see appendix), is given in (18). Each structure is the optimal form selected with respect to the inputs I1 , I2 , and I3 examined in the previous sections. For reasons of space, traces are henceforth omitted and square brackets represent VP boundaries. (18) SVO and lack of structural focus (Kera) SV[O], S f V[O], SV[Of IO] Lack of structural focus requires Stay to outrank both focus constraints, to prevent focused constituents from aligning with the left or right edge of the VP. Combined with the condition for SVO order, this yields the ranking in (19). (19) Lack of structural focus (SVO) Subject g Stay g {AF-R, AF-L} Let us check whether the above ranking precisely determines the pattern in (18). We do not need to examine the focusless input I1 , because under this input both focus constraints apply vacuously, and the Subject g Stay ranking already ensures the selection of the SVO basic word order, as we saw at the beginning of this section. The same holds for all the languages examined in the following paragraphs. Tableau T11.6 shows how the ranking in (19) predicts the correct structure for focused subjects under input I2 . The optimal structure A, which focuses the subject in SpecIP and matches the SVO order, is compared with the two alternative potential winners for this input. For A to beat B, Subject must outrank AF-R, as in the ranking above. Structure A in fact beats B on Subject, because B leaves SpecIP unrealized, but A loses against B on AF-R, because it realizes the subject preverbally, thus violating AF-R. The same ranking also ensures that A beats candidate C, which leaves the subject in situ. Having only one trace, structure C violates Stay one time less than A and satis®es AF-L, whereas A violates it. But both these constraints are lower ranked than Subject, which A satis®es and C does not. Therefore A beats C relative to the overall ranking.

Crosslinguistic Typologies in Optimality Theory

329

Tableau T11.7 Input I3 ˆ hV(x, y, z), y ˆ focusi A ˆ S V [ t t Of IO]

Subject

Stay

AF-R

A @ B Ð B ˆ S V [[ t t t IO] Of ]

A

B

A @ C Ð C ˆ S V [Of [t t t IO]]

A

A @ D Ð D ˆ V [S t Of IO]

A

A @ E Ð E ˆ V [[S t t IO] Of ]

A

A @ F Ð F ˆ V [Of [S t t IO]]

A

AF-L

C

D E F

Note, incidentally, that B and C can still be made optimal by reranking above Subject either one of the focus constraints. This would not contradict the ranking between Subject and Stay. Focusing the subject in VP-aligned position for input I2 , as in B and C, is therefore compatible with the basic SVO-order selected for input I1 , a fact to which we will return to in the following sections. The ranking in (19) also guarantees lack of structural focus in relation to focused objects in input I3 . It in fact selects as optimal structure A, which focuses the object in situ and follows the SVO order. As tableau T11.7 shows, in each competition the constraints favoring A outrank all the constraints favoring each alternative potential winner. In particular, the only two other SVO structuresÐB and CÐbeat A over the focus constraints, but they lose against A over the higher-ranked Stay, which A violates one time less than them. The remaining structures D, E, and F all follow the VSO word order and therefore all fail the highest-ranked constraint Subject. Note that D, E, and F cannot be made optimal without also a¨ecting the ranking for the optimal structures selected for inputs I1 and I2 . Structures E and F would in fact require either of the focus constraints to outrank Subject, which is incompatible with the selection of S f V[O] as optimal for input I2 in tableau T11.6. Structure D requires Stay above Subject, but this is incompatible with the SVO basic order selected as optimal for input I1 . Therefore, the three patterns in (20a), which keep the optimal forms SVO and S f V[O] ®xed for inputs I1 and I2 but select structure D, E, or F for input I3 , require contradictory rankings across the three inputs under scrutiny and thus characterize impossible languages. (20a) Impossible languages SVO, S f V[O], V [S Of IO] SVO, S f V[O], V [[S IO] Of ] SVO, S f V[O], V [Of [S IO]]

330

Vieri Samek-Lodovici

The remaining two structures, B and C, can be made optimal by promoting either of the focus constraints higher than Stay but lower than Subject, without necessarily disrupting the ranking conditions necessary for the optimal status of SVO and S f V[O] for inputs I1 and I2 . The following two languages are thus still possible and will be considered in section 11.5.3. (20b) Still possible (see section 11.5.3) SVO, S f V[O], SV [[ IO ] Of ] SVO, S f V[O], SV [ Of [IO ]] 11.5.2 Uniform Rightward Focus in SVO Languages: Italian The SVO basic word order is also compatible with uniform leftward and uniform rightward focus. In the rightward-focus case, focused constituents are realized at the right edge of the VP, as in the pattern below, instantiated by Italian (Antinucci and Cinque 1977; Calabrese 1982, 1992; Saccon 1993; Belletti and Shlonsky 1995; SamekLodovici 1996; Pinto 1997; see appendix). (21) SVO and rightward focus (Italian) SVO, V [[O] S f ], SV [[IO] Of ] For the subject to focus rightward, AF-R must outrank both Stay and Subject: the ®rst because otherwise the subject would be prevented from moving to its focus position, the second because otherwise the subject would raise to SpecIP position. In addition, AF-R must outrank AF-L, or else the subject would focus leftward. Once these conditions are joined with the ranking for SVO order, we obtain the ranking conditions speci®ed in (22). (22) Rightward focus (SVO) AF-R g Subject g Stay AF-R g AF-L Promoting AF-R to the highest rank ensures the proper rightward realization of focused subjects with respect to input I2 , as in structure B in tableau T11.8. The alternative potential winners A and C are beaten by B on AF-R, and beat B only on the lower-ranked constraints, thus losing across the overall ranking. Tableau T11.8 Input I2 ˆ hV(x, y), x ˆ focus) B ˆ V [[t t O] Sf ]

AF-R

Subject

B @ A Ð A ˆ Sf V [t t O]

B

A

B @ C Ð C ˆ V [ Sf t O]

B

Stay

AF-L

C

C

Crosslinguistic Typologies in Optimality Theory

331

The same ranking focuses objects in rightward position while keeping subjects in SpecIP, as in structure B in tableau T11.9. In all comparisons with the other competitors, the constraints where B wins against them are ranked higher than those where it loses. Structure B is also the only structure consistent, rankingwise, with the SVO and V[O]S f structures selected as optimal for input I1 and I2 . Making D and E optimal would in fact require reversing the condition Subject g Stay necessary for the SVO order in input I1 . Structures A, C, and F, on the other hand, would require demoting AF-R below either Stay or AF-L, with the e¨ect of selecting V[S f O] rather than V[[O]S f ] as the optimal realization for input I2 in tableau T11.8. We thus have ®ve additional languages incompatible with the requirement for ranking consistency across inputs. The corresponding patterns are listed in (23). (23) Impossible languages SVO, V [[ O ] S f ], SVO, V [[ O ] S f ], SVO, V [[ O ] S f ], SVO, V [[ O ] S f ], SVO, V [[ O ] S f ],

S V [Of IO] S V [Of [IO]] V [S Of IO] V [[S IO] Of ] V [Of [S IO]]

11.5.3 Partial Rightward Focus: English A particular version of rightward focus, whose pattern is shown in (24), realizes all focused constituents in rightward position except focused subjects, which occur in SpecIP. (24) Partial rightward focus SVO, S f V[O], SV [[ IO ] Of ] Tableau T11.9 Input I3 ˆ hV(x, y, z), y ˆ focusi B ˆ S V [[ t t t IO] Of ]

AF-R

B @ A Ð A ˆ S V [ t t Of IO]

B

B @ C Ð C ˆ S V [ Of [t t t IO]]

B

B @ D Ð D ˆ V [ S t Of IO]

B

B @ E Ð E ˆ V [[S t t IO] Of ] B @ F Ð F ˆ V [ Of [ S t t IO]]

B

Subject

Stay

AF-L

A C B

D

B

E

B

F

F

332

Vieri Samek-Lodovici

This pattern constitutes an interesting hybrid, because it parallels languages lacking structural focus with respect to focused subjects, but languages with rightward structural focus with respect to focused objects. An example of partial rightward focus is provided by English, although this language does not perfectly match the above pattern because it lacks V-to-I movement. English obligatorily focuses subjects preverbally but may optionally focus objects to the right of indirect objects (Rochemont and Culicover 1990:24; but see Costa 1996, 1997b, 1998a for an alternative view). As Rochemont and Culicover show, in fact, a heavy object following an indirect object is grammatical when focused, as in (25), but ungrammatical when unfocused, as in (26), showing that focus is one of the factors licensing the rightward realization of heavy objects. (Focus in italics.) (25) Q: A:

What did John purchase for his wife? John purchased for his wife a new fur coat.

(26) Q: For whom did John purchase a new fur coat? A: *John purchased for his wife a new fur coat. For partial rightward focus to occur, AF-R must dominate Stay, or else focused objects would remain in situ. However, AF-R must in turn be dominated by Subject, or else focused subjects would focus rightward too. This yields the ranking conditions in (27), including the ranking AF-R g AF-L required by any instance of rightward focus. (27) Partial rightward focus (SVO) Subject g AF-R g {Stay, AF-L} Under this ranking, the basic word order selected for input I1 remains the SVO one, because Subject still outranks Stay. Moreover, the optimal structure for input I2 correctly realizes focused subjects in preverbal position, as in structure A in tableau T11.10. Structure A in fact beats all competing structures on the highest-ranked constraint Subject. The same ranking, however, properly focuses objects rightward, as in structure B in tableau T11.11. Again, all competing structures lose to B on the two highestranked constraints and are therefore suboptimal. Tableau T11.10 Input Iˆ hV(x, y), x ˆ focusi A ˆ Sf V [ t t O]

Subject

AF-R

A @ B Ð B ˆ V [[ t t O] Sf ]

A

B

A @ C Ð C ˆ V [ Sf t O]

A

Stay

AF-L

C

C

Crosslinguistic Typologies in Optimality Theory

333

Tableau T11.11 Input I3 ˆ hV(x, y, z), y ˆ focusi B ˆ S V [[ t t t IO] Of ]

Subject

AF-R

Stay

B @ A Ð A ˆ S V [ t t Of IO]

B

A

B @ C Ð C ˆ S V [Of [t t t IO]]

B

B @ D Ð D ˆ V [ S t Of IO]

B

B @ E Ð E ˆ V [[S t t IO] Of ]

B

B @ F Ð F ˆ V [ Of [S t t IO]]

B

B

AF-L

C D E

B

F

F

Symmetric relations and competitions occur when AF-L and AF-R are switched, yielding the ranking in (28). (28) Partial leftward focus (SVO) Subject g AF-L g {Stay, AF-R} This ranking predicts the partial leftward focus pattern in (29), which is still unattested. This language would focus subjects in SpecIP, and objects and indirect objects in leftward VP-adjoined position. Note that only focused indirect objects would alter the basic word order of the languages by occurring between the verb and the object, as in the structure SV[IOf [O]]. This characteristic makes it easier to overlook the pattern. (29) Partial leftward focus (unattested) SVO, S f V[O], SV [Of [IO]] The analysis correctly predicts the unattested status of two other conceivable partial focus patterns. The ®rst concerns partial rightward focus within VSO languages Ðthat is, a pattern that focuses subjects in SpecVP but the object rightward, as in the pattern hVSO, V[S f O], V[[S IO] Of ]i. Focusing the object rightward requires AF-R to outrank both AF-L and Stay, or else the object could not move in its rightadjoined position. But focusing the subject in SpecVP requires AF-L or Stay to outrank AF-R, reversing the ranking just established. Partial rightward focus is thus properly predicted to occur only among SVO languages (partial leftward focus in VSO languages is not distinguishable from uniform leftward focus, examined in section 11.6.3). The second prediction concerns the absence of languages where subjects are focused structurally but internal arguments are focused in situ, as in the three patterns in (30) (the missing fourth pattern, focusing the subject leftward as in

334

Vieri Samek-Lodovici

hVSO, V[S f O], V[SOf IO]i, coincides with the pattern for VSO languages lacking structural focus, discussed in section 11.6.1). (30) Impossible languages SVO, V[ S f O ], S V [Of IO] SVO, V[[ O ] S f ], S V [Of IO] VSO, V[[ O ] S f ], V [S Of IO] These patterns are impossible because to focus the internal arguments in situ, Stay must outrank both focus constraints. But under this condition, subjects can no longer focus structurally, since this requires one or the other focus constraint to outrank both Stay and Subject. The corresponding languages are therefore predicted to be impossible, and to my knowledge are indeed unattested. Before we conclude this section, note that the impossible patterns considered here recur as speci®c cases in the sets of impossible languages identi®ed in other sections, and should therefore not be counted twice toward the 36 patterns of the harmonic boundary examined in section 11.4. 11.5.4 Uniform Leftward Focus in SVO Languages: Aghem Uniform leftward focus, attested in Aghem (Hyman 1979) and represented in (31), is analogous to uniform rightward focus, except that the highest-ranked constraint is AF-L rather than AF-R. The constraint AF-L must in fact outrank Subject, or else focused subjects would raise to SpecIP, and it must also outrank STAY and AF-R, or else focused objects would either remain in situ or focus rightward. Once combined with the ranking responsible for SVO basic word order, we obtain the ranking conditions speci®ed in (32). (31) SVO and uniform leftward focus (Aghem) SVO, V[S f O], SV [ Of [IO]] (32) Uniform leftward focus (SVO) AF-L g Subject g Stay AF-L g AF-R The correctness of the above conditions relative to inputs I2 and I3 can be checked in the two following comparative tableaux. In the case of input I2 , the optimal structure, focusing subjects leftward, is C. This structure beats its alternatives on both AF-L and Stay. One of these constraints must thus dominate Subject and AF-R, on which C loses. Since ranking Subject above Stay is required by the SVO order optimal relative to input I1 , the constraint dominating SubjectÐand hence all other constraintsÐmust be AF-L, con®rming the ranking in (32). (See tableau T11.12.)

Crosslinguistic Typologies in Optimality Theory

335

Tableau T11.12 Input I2 ˆ hV(x, y), y ˆ focusi C ˆ V [ Sf t O]

AF-L

Subject

Stay

C @ A Ð A ˆ Sf V [ t t O]

C

A

C

C @ B Ð B ˆ V [[ t t O] Sf ]

C

AF-R

C

B

Tableau T11.13 Input I3 ˆ hV(x, y, z), y ˆ focusi C ˆ S V [ Of [t t t IO]]

AF-L

Subject

Stay

C @ A Ð A ˆ S V [ t t Of IO]

C

C @ B Ð B ˆ S V [[ t t t IO] Of ]

C

C @ D Ð D ˆ V [ S t Of IO]

C

C

D

C @ E Ð E ˆ V [[S t t IO] Of ]

C

C

E

C

F

C @ F Ð F ˆ V [ Of [ S t t IO]]

AF-R

A B

E

The same ranking properly focuses objects leftward in input I3 , as shown in structure C in tableau T11.13. Structure C in fact beats all its alternatives on either one or both of the two highest-ranked constraints. Here too we ®nd a host of patterns whose realization requires inconsistent rankings. Let us examine them in turn. Reranking Stay above AF-L, as required if A and D are to win, is inconsistent with the SVO order selected for input I1 , because Stay would then also outrank Subject, contradicting the conditions for the SVO order selected in input I1 . This could be obviated by promoting Subject as well, but then Subject would outrank AF-L, and in the previous competition for input I2 , the optimal C would lose against candidate A. Thus, selecting A and D as optimal for input I3 is incompatible with the SVO and SV[Of [IO]] structures selected as optimal for inputs I1 and I2 . Structures B and E, on the other hard, could beat C just by promoting AF-R above AF-L. Even this reranking, however, would cause the optimal C to lose against structure B in input I2 in tableau T11.12. Finally, for F to be optimal, Stay must outrank Subject, again against the SVO order selected for input I1 . The ®ve patterns speci®ed in (33) are thus impossible.

336

Vieri Samek-Lodovici

(33) Impossible languages SVO, V[S f O], S V [Of IO] SVO, V[S f O], S V [[IO ] Of ] SVO, V[S f O], V [S Of IO] SVO, V[S f O], V [[S IO] Of ] SVO, V[S f O], V [Of [S IO ]] 11.5.5 Summary The above discussion has shown that only 5 of the distinct 18 SVO patterns in the harmonic boundary involve noncontradictory rankings. The remaining 13 patterns were shown to represent impossible languages. Constraint consistency across inputs thus plays a role second only to harmonic bounding in delimiting the typology entailed by a set of constraints. 11.6

Cross-Input Consistency: VSO Languages

The second half of the focus typology concerns languages with VSO basic word order, thus reversing the relation between Subject and Stay into Stay g Subject. As discussed in section 11.5.3, this ranking is incompatible with partial focus, which to the best of my knowledge is in fact unattested in VSO languages. Lack of structural focus, and uniform leftward and rightward focus, are instead possible and attested, as the following sections will show. 11.6.1 Lack of Structural Focus in VSO Languages: Scottish Gaelic As in the corresponding SVO pattern, structural focus is absent in VSO languages whenever the focus constraints are lower ranked than Stay. The invariant pattern in (34), found in Scottish Gaelic (Ward, personal communication; see appendix) and according to Moutaouakil also in Modern Arabic (1989:1±66), is thus determined by the ranking conditions in (35). (34) VSO and lack of structural focus (Scottish Gaelic) VSO, V[S f O], V[S Of IO] (35) Lack of structural focus (VSO) Stay g {Subject, AF-R, AF-L} The highest-ranked constraint Stay freezes all focused constituents in situ, and therefore focused subjects are realized in SpecVP, as in structure C in tableau T11.14. Any alternative realization involves additional movement operations, and therefore loses against C with respect to Stay. However, the ambiguous status of SpecVP subjects, which count both as in situ and as left VP-aligned, makes it also possible to

Crosslinguistic Typologies in Optimality Theory

337

Tableau T11.14 Input I2 ˆ hV(x, y), x ˆ focusi C ˆ V [ Sf t O]

Stay

Subject

C @ A Ð A ˆ Sf V [ t t O]

C

A

C @ B Ð B ˆ V [[ t t O] Sf ]

C

AF-R

AF-L C

B

C

Tableau T11.15 Input I3 ˆ hV(x, y, z), y ˆ focusi D ˆ V [ S t Of IO]

Stay

Subject

D @ A Ð A ˆ S V [ t t Of IO]

D

A

D @ B Ð B ˆ S V [[ t t t IO] Of ]

D

B

D @ C Ð C ˆ S V [ Of [t t t IO]]

D

C

D @ E Ð E ˆ V [[S t t IO] Of ]

D

D @ F Ð F ˆ V [ Of [S t t IO]]

D

AF-R

AF-L

B C E F

select C as optimal just by promoting AF-L higher than Subject and AF-R. This input is thus insu½cient to con®rm the ranking relations proposed in (35). Note that focusing the subject in preverbal position, as in A, requires Subject to outrank Stay, contradicting the condition Stay g Subject underlying the VSO basic order assumed optimal for input I1 . It follows that any pattern combining VSO order with preverbal focused subjects in input I2 is impossible, independently of how objects are focused. This rules out the six patterns in (36). (36) Impossible languages VSO, S f V[O], S V [Of IO] VSO, S f V[O], S V [[ IO ] Of ] VSO, S f V[O], S V [Of [IO]] VSO, S f V[O], V [S Of IO] VSO, S f V[O], V [[S IO] Of ] VSO, S f V[O], V [Of [S IO ]] Going back to the lack of structural focus in VSO languages, let us consider focused objects. As tableau T11.15 shows, the highest-ranked constraint Stay forces their in situ realization, as in structure D. The ranking ambiguity between Stay and AF-L that held relative to input I2 is resolved here: D can be optimal only if Stay is

338

Vieri Samek-Lodovici

ranked highest. Any other ranking would see D lose against one of the other potential winners. We can also derive the additional four impossible patterns listed below (the pattern involving structure F is possible, as we will see in section 11.6.3). The ®rst three patterns, for A, B, and C, realize the unfocused subject in SpecIP position. This requires Subject g Stay, in direct contradiction with the selection of VSO as the optimal structure for input I1 . The last pattern, selecting structure E for input I3 , represents partial rightward focus within VSO languages, and its impossibility has already been examined in section 11.5.3. (37) Impossible languages VSO, V[S f O], S V [Of IO] VSO, V[S f O], S V [ IO ] Of ] VSO, V[S f O], S V [Of [IO ]] VSO, V[S f O], V [[S IO] Of ] 11.6.2 Uniform Rightward Focus in VSO Languages: Spanish Like SVO languages, VSO languages allow for uniform rightward focus whenever AF-R is ranked highest in the hierarchy, as in (38). (38) Uniform rightward focus (VSO) AF-R g Stay g Subject AF-R g AF-L The corresponding pattern, shown in (39), is partially instantiated in regional varieties of Spanish, except for the raising of topic-marked subjects to SpecIP in QA pairs concerning focused objects (Zubizarreta 1992; OrdoÂnÄez and TrevinÄo 1995; OrdoÂnÄez 1997; see appendix). I will assume that the raising of topic-marked subjects is caused by independent constraints, such as for example TopFirst, as proposed in Costa (1998b; chap. 7, this volume). (39) VSO and uniform rightward focus (Spanish) VSO, V [[O] S f ], V [[S IO] Of ] As tableau T11.16 shows, ranking AF-R highest is necessary to realize focused subjects rightward, as in B. Under any other ranking B would in fact lose against SpecIP subjects, shown in A, or focus in situ, shown in C. The same ranking ensures that objects focus rightward as well, as in structure E below (see tableau T11.17). Structure E already beats all other competitors except B on the highest-ranked constraint AF-R. As for B, the higher rank of Stay relative to Subject, required by VSO order, ensures that it too loses against E under the proposed ranking.

Crosslinguistic Typologies in Optimality Theory

339

Tableau T11.16 Input I2 ˆ hV(x, y), x ˆ focusi B ˆ V [[ t t O] Sf ]

AF-R

B @ A Ð A ˆ Sf V [ t t O]

B

B @ C Ð C ˆ V [ Sf t O]

B

Stay

Subject

AF-L

A C

C

Tableau T11.17 Input I3 ˆ hV(x, y, z), y ˆ focusi E ˆ V [[S t t IO] Of ] E @ A Ð A ˆ S V [ t t Of IO]

AF-R

Stay

E

E @ B Ð B ˆ S V [[ t t t IO] Of ]

Subject

AF-L

A E

B C

E @ C Ð C ˆ S V [ Of [t t t IO]]

E

E

E @ D Ð D ˆ V [ S t Of IO]

E

D

E @ F Ð F ˆ V [ Of [S t t IO]]

E

C

F

Tableau T11.17 also clearly shows that none of the suboptimal structures can form a pattern with the VSO and V[[O]S f ] structures selected as optimal for inputs I1 and I2 . Candidate B is incompatible with VSO order, as we just saw, and structures A, C, D, and F can be made optimal only by demoting AF-R from its highest rank, contradicting the condition found necessary in tableau T11.16 for focused subjects. All the patterns in (40) thus represent impossible languages. (40) Impossible languages VSO, V [[O] S f ], SV[Of IO] VSO, V [[O] S f ], SV [[ IO] Of ] VSO, V [[O] S f ], SV [Of [IO]] VSO, V [[O] S f ], V[S Of IO] VSO, V [[O] S f ], V [Of [S IO]] 11.6.3 Uniform Leftward Focus in VSO Languages: Podoko Finally, let us consider uniform leftward focus under VSO basic order, present in the Chadic language Podoko (Jarvis 1981; Tuller 1992). The pattern is speci®ed in (41). (41) VSO and uniform leftward focus (Podoko) VSO, V [S f O], V [Of [S IO]]

340

Vieri Samek-Lodovici

Tableau T11.18 Input I2 ˆ hV(x, y), x ˆ focusi C ˆ V [ Sf t O]

AF-L

Stay

Subject

C @ A Ð A ˆ Sf V [ t t O]

C

C

A

C @ B Ð B ˆ V [[ t t O] Sf ]

C

C

AF-R

B

Tableau T11.19 Input I3 ˆ hV(x, y, z), y ˆ focusi F ˆ V [ Of [S t t IO]]

AF-L

F @ A Ð A ˆ S V [ t t Of IO]

F

F @ B Ð B ˆ S V [[ t t t IO] Of ]

F

F @ C Ð C ˆ S V [ Of [t t t IO]] F @ D Ð D ˆ V [ S t Of IO]

F

F @ E Ð E ˆ V [[S t t IO] Of ]

F

Stay

Subject

AF-R

A F

B

F

C

B

D E

To focus objects leftward, AF-L must outrank AF-R and Stay. Once combined with the condition for VSO order, we obtain the ranking shown in (42). (42) Uniform leftward focus (VSO) AF-L g Stay g Subject AF-L g AF-R The highest rank assigned to AF-L ensures that subjects are focused in SpecVP, as in structure C in tableau T11.18, where the subject is aligned with the left edge of the VP. As we saw in section 11.6.1, the same result could be achieved by ranking Stay higher than Subject and AF-R, because in VSO languages leftward focusing and focusing in situ in the case of subjects coincide. Ranking AF-L highest, however, is necessary to focus objects leftward, as in structure F in tableau T11.19. Any reranking would in fact see F lose to one of the alternative potential winners. In particular, F would lose against D if AF-L were switched with Stay, against A if were switched with Subject, or ®nally against E if the switch involved AF-R. This concludes the discussion of VSO languages. Of the 18 VSO patterns in the harmonic boundary, only 3 survived as possible, while 15 turned out inconsistent rankingwise.

Crosslinguistic Typologies in Optimality Theory

11.7

341

Conclusions

Two main results emerge from this study, the ®rst speci®c to the syntax of structural focus, and the second, more general, concerning OT typologies. Structural focus involves by de®nition an alteration of the canonical word order of a language, leaving the impression that the principles governing its syntax are unconstrained by those determining basic word order. The model developed here took the opposite approach, exploiting OT to force maximal interaction between focus and word order constraints. The single most visible manifestation con®rming this interaction was o¨ered by partial structural focus, where focus according to basic word order and focus-induced word order coexist in the same language depending on which focused constituent we look at. In the model proposed here, this mixed pattern, together with its restriction to SVO languages and the other related properties examined in section 11.5.3, follow straightforwardly from the interaction of the proposed constraints. Note furthermore that whether structural focus applies uniformly across syntactic constituents, or only partially so, or does not apply at all, is always an epiphenomenal property derived from the interaction of focus and word order constraints, not one encoded in the constraints themselves. The best evidence for the proposed model, however, lies in the overall typology it derives, with leftward, rightward, uniform, partial, and lack of structural focusÐin SVO and VSO languagesÐall following from, and more important, exhausting the interaction of the constraints AF-R, AF-L, Stay, and Subject. This study also supports a more explicit use of harmonic bounding in OT analyses. Harmonic bounding can reduce an in®nite number of candidates to a ®nite number of potential winners, providing an e¨ective tool to exploit the crosslinguistic predictive power of OT. It also permits us to investigate the relation between the typologies of di¨erent inputs, an aspect of which was captured by the theorem on monotonic enrichment in section 11.3. Using harmonic bounding, this study also showed how the theoretical upper bound associating N constraints with N! optima per input, and therefore …N!† M patterns across M distinct inputs, may greatly overestimate the actual size of a typology, predicting, for the focus typology examined here, 13.824 distinct paradigms against the only 8 logically possible. A signi®cantly more precise upper bound is provided by the harmonic boundary of a typology, which in our case cut the estimated size of the focus typology to just 36 languages, leaving it to constraint consistency to bring the total size down to the ®nal 8 possible focus patterns examined in this chapter.

342

11.8

Vieri Samek-Lodovici

Appendix

11.8.1 SVO and Lack of Structural Focus: Kera (Ebert 1979) In her detailed grammar for this East Chadic language, Ebert argues for an SVO basic word order (p. 206). In the following data, the su½x ``-n'' marks past tense when su½xed to verbs (p. 77), as well as de®niteness when su½xed to nouns ending in a vowel (p. 165). Interrogatives are marked through clause-®nal ``moÂ'' (p. 223). For possessive ``-aÁ'' and ``-i,'' see pp. 130±132. For the pronominal gloss for ``a,''  ,'' and ``kPÂna,'' see pp. 130, 145, 165, and 191, respectively. ``moÂtoÁ,'' ``huÁluÁm (43) Basic order: S V O IO (Ebert 1979: example 1, p. 175; example ``Kl1,'' p. 78; example 18, p. 211) a. HPÁ lgP moÂtoÁ-n aÂyaÂ-n kPÂsaÂw a geÁdr-aÁ-nÂ. wife bigfem -the gave sauce to man-her-the `The main wife gave the sauce to her man.' Â. b. WP fPÂlaÂ-n huÁluÁm he found man-the `He found the man.' c. HaÁrgaÂ-n aÂwaÂ-n dPÂr kPÂna. goat-the fall water to-here `The goat fell in the water.' (44) Focused subjects precede verb: S f V [ t t O] (Ebert 1979: example 4, p. 224) Q: Mintõ lPÂtaÂ-n nPwr-i moÂ? who hit sister-your Q-mark `Who hit your sister?' A: AgeÁleÁm lPÂtaÂ-n nPwr-i. Agelem hit sister-your `Agelem hit your sister.' (45) Focused objects precede indirect objects: S V [ t t Of IO ] (Ebert 1979: example, p. 224) Q: AgeÁleÁm lPÂtaÂ-n mintõ moÂ? Agelem hit who Q-mark `Who did Agelem hit?' A: AgeÁleÁm lPÂtaÂ-n nPwr-i. Agelem hit your sister `Agelem hit your sister.'

Crosslinguistic Typologies in Optimality Theory

343

(46) Focused indirect objects follow objects: S V [ t t O IOf ] (Ebert 1979: example 6, p. 225) Q: A aÂyaÂ-n kuÂsuÂkõÂ-n a mintõ moÂ? she gave meat-the to whom Q-mark `Whom did she give the meat to?' A: A aÂyaÂ-n kuÂsuÂkõÂ-n a hPÁlgPÂ-nÂ. she gave meat-the to woman-the `She gave the meat to the woman.' 11.8.2 SVO and Uniform Rightward Focus: Italian (Antinucci and Cinque 1977; Calabrese 1982, 1992; Saccon 1993; Belletti and Shlonsky 1995; Samek-Lodovici 1996; Pinto 1997) Focus in QA pairs occurs clause ®nally in Italian, optionally followed by rightdislocated constituents. Focused neg-phrases can be licensed by a c-commanding negative marker in Io , showing that focus is VP-adjoined (not IP-adjoined). Northern Italian varieties replace past tense with present perfect, which involves an auxiliary, not properly matching the V-to-I pattern examined in the previous sections; therefore, the following examples use the auxiliary-free future tense. (47) Basic word order: S V O IO Q: Cosa succederaÁ? what happen-FUT `What will happen?' A: Maria regaleraÁ dei ®ori a Marco. Mary donate-FUT some ¯owers to Mark `Mary will donate some ¯owers to Mark.' (48) Focused subjects follow objects: V [[ t t O] S f ] Q: Chi chiameraÁ la polizia? who call-FUT the police `Who will call the police?' A: ChiameraÁ la polizia Maria. call-FUT the police Mary `Mary will call the police.' (49) Focused objects follow indirect objects: S V [[ t t t IO] Of ] Q: Cosa regalerai a Gianni? what (you) donate-FUT to John `What will you give to John?' A: RegaleroÁ a Gianni un libro. (I) donate-FUT to John a book `I will give John a book.'

344

Vieri Samek-Lodovici

11.8.3 SVO and Uniform Leftward Focus: Aghem (Hyman 1979) Hyman discusses a variety of criteria that identify SVO as the canonical order for Aghem, a Bantu language of Cameroun. In particular, SVO order is described as the only possible order to answer clause-focusing questions such as (50) below. This QA pair has been split into two distinct examples but is described as forming a single QA pair in the main text (pp. 143±145). (``P2 '' expresses distant past; see note on p. 192.) (50) Basic order: VS (Hyman 1979: examples 5 and 6a, pp. 144±145) Á mOÁ Q: A koÂm kwOÁ? it P2 -past happen what `What happened?' A: EÂnaÂ" mOÁ nÄ2n noÏ. Inah P2 -past run focus-marker `Inah ran.' In Aghem, focused subjects, objects, and indirect objects occur immediately after the verb (Hyman 1979:144±148). In the data below, I have consistently glossed the subject particle ``aÂ'' as the expletive there, whereas in the original examples it occurs as either it, or Dummy Subject, or Subject Marker; see Hyman 1979:xi, 146. (51) Focused subjects occur between the verb and the object: V [Sf t O], and V [Sf t] (Hyman 1979: example 16b, p. 146; for the QA pair in (51b), see examples 6d and 6b as discussed on pp. 144±145) Á a. Q: A mOÁ z2 nduÂghOÁ bEÂ-=kOÂ? there P2 -past eat who fufu `Who ate the fufu?' Á z2 aÂ-f2n bEÂ-=kOÂ. A: A mOÁ there P2 -past eat friends fufu `The friends ate the fufu.' Á b. Q: A mOÁ nÄ2n nduÂghOÁ? there P2 -past run who `Who ran?' Á A: A mOÁ nÄ2n EÂnaÂ". there P2 -past run Inah `Inah ran.' (52) Focused objects follow the verb: S V [ Of [ t t t ]] (Hyman 1979: examples 15a and 15b, p. 146) Q: F2l aÁ mOÁ z2 kwOÂ? friends it P2 -past eat what `What did the friends eat?'

Crosslinguistic Typologies in Optimality Theory

345

A: F2l aÁ mOÁ z2 k2-bEÂ. friends it P2 -past eat fufu `The friends ate the fufu.' (53) Focused adjuncts precede objects: S V [ Adjunctf [ t t O ]] (Hyman 1979: examples 19b and 19c, p. 147. Answer A below contrasts with the object-adjunct order of its focusless correspondent ``f2l aÁ mOÁ z2 k2-bE aÂn =soÂm''; see example 19a, p. 147) Q: F2l aÁ mOÁ z2 ghE bEÂ-=kOÂ? friends it P2 -past eat where fufu `Where did the friends eat fufu?' z2 aÂn =soÂm bEÂ-=kOÂ. A: F2l aÁ mOÁ friends it P2 -past eat in farm fufu `The friends ate the fufu in the farm.' 11.8.4 VSO and Lack of Structural Focus: Scottish Gaelic These data were kindly provided by Cecil Ward (personal communication). Structurally identical data for Irish Gaelic were provided by Caoimhin P. Odonnaile (personal communication). (54) Basic word order: VSO Chuir duine soir air a0 bhoÁrd. put[‡past] man vase on art[‡prep] table `A man put a vase on the table.' (55) Focused subjects remain in situ: V [ Sf t O] Q: Co a sgrõÁobh litir? who REL wrote[‡past] letter `Who wrote a letter?' A: SgrõÁobh duine litir. wrote[‡past] man letter `A man wrote a letter.' (56) Focused objects remain in situ: V [S t Of IO] Q: De a chuir duine air a0 bhoÁrd? what REL put[‡past] man on art[‡prep] table `What did a man put on the table?' A: Chuir duine soir air a0 choÁrd. put[‡past] man vase on art[‡prep] table `A man put a vase on the table.' A similar analysis is proposed by Moutaouakil (1989:1±66) for Modern Standard Arabic. According to this author, Arabic is a VSO language (p. 55) with in situ newinformation focus (p. 56).

346

Vieri Samek-Lodovici

(57) Basic order: VSO (Moutaouakil 1989: example 20 and discussion, p. 25) Q: MaÅ l-habaru? what the-news? `What is happening?' A: Rajac a Zaydun min s-safari. returned Zaydnom from the-journeygen `Zayd has returned from his journey.' (58) Focused subjects remain in situ: V [ Sf t O] (Moutaouakil 1989: example 76, p. 40) Q: Man "allafa kitaÅban? who wrote bookacc `Who wrote a book?' A: "allafa Zaydun kitaÅban. wrote Zaydnom bookacc `Zayd wrote a book.' (59) Focused objects remain in situ: V [S t Of IO] (Moutaouakil 1989: example 77, p. 40) Q: MaÅdaÅ "akala c Amrun? c what ate Amrnom `What did c Amr eat?' A: "akala c Amrun dajaÅjan. c ate Amrnom chickenacc c ` Amr ate chicken.' 11.8.5 VSO and Uniform Rightward Focus: Spanish (Zubizarreta 1992; OrdoÂnÄez and TrevinÄo 1995; OrdoÂnÄez 1997; Costa, chap. 7, this volume) Unlike Italian, Spanish allows for unfocused SpecVP subjects. Like Italian, on the other hand, it shows clause-®nal focus in QA pairs. In the example for focused objects, the subject is a topic and therefore raises to SpecIP. (60) Basic order: VSO (OrdoÂnÄez 1997:31) Q: aQue paso ayer? `What happened yesterday?' A: Ayer gano Juan la loterõÂa. yesterday won John the lottery `Yesterday John won the lottery.'

Crosslinguistic Typologies in Optimality Theory

347

(61) Focused subjects follow in situ objects: V [[ t t O] Sf ] (Zubizzareta 1992:22±24) Q: aQuien comio una manzana? `Who ate an apple?' A: Comio una manzana Juan. ate an apple John `John ate an apple.' (62) Focused objects follow indirect objects: S V [[ t t IO] Of ] (Zubizzareta 1992:22±24) Q: aQue puso MarõÂa sobre la mesa? what put Mary on the table `What did Mary put on the table?' A: MarõÂa puso sobre la mesa el libro. Mary put on the table the book `Mary put a book on the table.' 11.8.6 VSO and Uniform Leftward Focus: Podoko (Jarvis 1981; Tuller 1992) Jarvis identi®es VSO as the canonical order for Podoko. The alternative SVO order may occur but has a very restricted distribution. The VOS order is attested, too, but is restricted to speci®c dialogue-oriented registers of the grammar; see Jarvis 1981 for discussion and for evidence supporting VSO as the basic word order. (63) Basic word order: V S O PP (Jarvis 1981: example 11, p. 159) TalPda malP slPbP akP bala. cook-for-him mother-my meat for father-my `My mother cooked meat for my father.' (64) Focused subjects follow the verb and precede objects: V [ Sf t O] (Jarvis 1981: examples 19a and 19b, p. 161) Q: A tPla wa slPbP na " register-mark cook who meat Q-mark `Who cooked the meat?' A: A tPla malP slPba. register-mark cook-for-him mother-my meat `My mother cooked meat.' (65) Focused objects follow the verb and precede subjects: V [ Of [ S t t ]] (Jarvis 1981: examples 14a and 14b, p. 160) Q: A tPla tawP ndi na " register-mark cook what one Q-mark `What did one cook?'

348

Vieri Samek-Lodovici

A: A tPla slPbP nda. register-mark cook meat one `One cooked meat.' (66) Focused Adjuncts follow the verb and precede subjects: V [ PPf [ S t O]] (Jarvis 1981: examples 21a and 21b as well as 22a and 22b, p. 161) Q: A tPla ta wa malP slPbP na? register-mark cook for whom mother-my meat Q-mark `For whom did my mother cook the meat?' A: A tPla ta balP malP slPba. register-mark cook for father-my mother-my meat `My mother cooked the meat for my father.' Q: A tPla hawP malP slPbP na? register-mark cook where mother-my meat Q-mark `For whom did my mother cook the meat?' A: A tPla dP ykwPbPgP malP slPba. register-marker cook in kitchen mother-my meat `My mother cooked the meat in the kitchen.' Notes This work was generously supported by the National Science Foundation grant NSF SBR 95 11891, by Rutgers University, and by Konstanz University. I am particularly grateful for comments and suggestions to Peter Ackema, Eric BakovicÂ, Miriam Butt, JoaÄo Costa, Viviane Deprez, Jane Grimshaw, Edward Keer, Takeo Kurafuji, GeÂraldine Legendre, Ad Neeleman, Caoimhin Odonnaile, Alan Prince, Paul Smolensky, Hubert Truckenbrodt, Sten Vikner, Cecil Ward, Colin Wilson, as well as audiences at the University of DuÈsseldorf, the University of KoÈln, the University of Manchester, the University of Marburg, Rutgers University, Stuttgart University, and the University of Utrecht. 1. Case-related constraints may enrich the predicted typology for left-headed languages as well. Extending the grammar to case adjacency, for example, allows the derivation of languages like Kanakuru (Newman 1974; Tuller 1992), where leftward and rightward focus occur in complementary distribution within the same language (Samek-Lodovici 1998). 2. Consider ®rst the Italian data in (i) below, where the rightward-focused subject binds into the object at its left. Under a leftward scrambling analysis this should not be possible because the object scrambles to an A-position higher than the subject, as shown in (iia). The same data, however, follow straightforwardly if the subject has right VP-adjoined as in (iib), because in this case the in situ object is bound by the subject trace in SpecVP. (i) Q: Chi incoraggia gli alunni in questa scuola? Who encourages the students in this school? A: Incoraggia [i suoii alunni] [ogni insegnante] i . encourages [the his students] [each teacher] `Each teacher encourages his own students.'

Crosslinguistic Typologies in Optimality Theory

349

(ii) a. . . . O [vp Sf t t ] b. . . . [vp [vp t t O ] Sf ] Next, consider the analyses where objects move (overtly) leftward to get case in SpecAgro , leaving an in situ focused subject at its right (Reinhart 1995; Costa 1997a; Pinto 1997; Cinque 1999; Zubizarreta 1992, 1998). As Costa observes, these analyses predict that a focused subject will precede a VP-internal prepositional phrase, because PPs do not need to get case. Pinto (1997:235), however, points out that this prediction does not hold in Italian: as (iii) shows, postverbal focused subjects follow prepositional arguments (the reverse order is possible only under right dislocation of the PP). (iii) Q: Chi eÁ arrivato da Firenze? who is arrived from Florence `Who arrived from Florence?' A: E 0 arrivato da Firenze Dante. is arrived from Florence Dante `Dante arrived from Florence.' Finally, consider the important evidence for leftward scrambling provided in OrdoÂnÄez 1995, 1997 for Spanish, where indirect objects bind subjects in V-IO-S patterns. The reverse binding relation is not possible, precisely as predicted by a leftward-scrambling analysis. (Similar data hold for Italian.) (iv) Q: aQue le regalo a cada ninÄoi sui amigo para su compleanÄos? what cl-gave to each boy his friend for his birthday `What did his friend give to each child for his birthday?' A: Este libro se lo regalo a cada ninÄoi sui amigo. this book cl-gave to each boy his friend `His friend gave this book to each child.' Note that the focused constituent is here the object este libro. The subject is mentioned in the question and thus cannot be focused. Therefore these data support the availability of object leftward scrambling in Spanish and Italian but do not directly refute focus-based right adjunction. Obviously, they do suggest that the same scrambling analysis might apply to the focus case. But as the data in (i) showed, when the same paradigm is applied to a focused subject in Italian the opposite binding facts hold, again in favor of the right-adjunction analysis. 3. Samek-Lodovici and Prince (1999) show that losers also include collectively bounded structuresÐthat is, structures that never win because for each ranking there is a structure beating them, though not necessarily always the same one. These candidates cannot be excluded on the basis of classical harmonic bounding alone. However, their existence requires the presence of more than one gradient constraint, and they therefore cannot occur in this study, where Stay is the only gradient constraint available. 4. As mentioned, basic word order occurs in answers to questions like ``What happened?'', which focus the entire answer. A more precise input for basic word order should thus focus the entire underlying argument structure, as in hV(x, y) ˆ focusi. It is natural to assume that under this input all extended projections of V count as focused, starting with VP. But VP aligns with itself. Therefore under this input all candidates in Gen satisfy both focus constraints, determining the exact same set of potential winners described for the unfocused hV(x, y)i in the main text.

350

Vieri Samek-Lodovici

5. Examining additional inputs, with focused indirect objects and focused adjuncts, complicates the exposition without adding signi®cant content to it. The focus pattern followed by these constituents is in fact the same for focused objects in input I3 , thus focusing leftward when objects focus leftward, focusing in situ when objects focus in situ, and so on. A more interesting extension concerns inputs with multiple foci, but a lack of relevant data for some of the languages considered here does not allow a proper investigation of this topic. References Ackema, Peter, and Ad Neeleman. 1998. WHOT? In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 15±33. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Ackema, Peter, and Ad Neeleman. 1999. Parsing Rightward Movement. Ms., University of Groningen and University College London. Antinucci, Francesco, and Guglielmo Cinque. 1977. Sull'ordine delle parole in Italiano: l'emarginazione. Studi di Grammatica Italiana VI: 121±146. Belletti, Adriana, and Ur Shlonsky. 1995. The Order of Verbal Complements: A Comparative Study. Natural Language and Linguistic Theory 13: 489±526. Bresnan, Joan. 2000. Optimal Syntax. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Phonology, Syntax, and Acquisition. Oxford: Oxford University Press. BuÈring, Daniel, and Katarina Hartmann. 1994. The Kayne Mutiny. In D. Beerman, D. LeBlanc, and H. van Riemsdijk (eds.), Rightward Movement. Linguistik Aktuell±Linguistics Today, vol. 17. Amsterdam: Benjamins. Calabrese, Andrea. 1982. Alcune ipotesi sulla struttura informazionale della frase in Italiano e sul suo rapporto con la struttura fonologica. Rivista di Grammatica Generativa 7: 3±78. Calabrese, Andrea. 1992. Some Informal Remarks on Focus and Logical Structures in Italian. In Susumu Kuno and HoÈskuldur ThraÂinsson (eds.), Harvard Working Papers in Linguistics 1, 91±127. Department of Linguistics, Harvard University, Cambridge, Mass. Choi, Hye-Won. 1996. Optimizing Structures in Context: Scrambling and Information Structure. Doctoral dissertation, Stanford University, Stanford, Calf. Chomsky, Noam. 1982. Some Concepts and Consequences of the Theory of Government and Binding. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1992. A Minimalist Program for Linguistic Theory. MIT Occasional Papers in Linguistics 1. Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Cinque, Guglielmo. 1999. Adverbs and Functional Heads. Oxford: Oxford University Press. Costa, JoaÄo. 1996. Word Order and Constraint Interaction. In Seminarios de Linguistica. University of Algarve. Costa, JoaÄo. Forthcoming. Focus in Situ: Evidence from Portuguese. Probus. Costa, JoaÄo. 1997a. Position of Subjects in European Portuguese. In Brian Agbayani (ed.), Proceedings of the West Coast Conference on Formal Linguistics XV. Stanford, Calif.: CSLI Publications.

Crosslinguistic Typologies in Optimality Theory

351

Costa, JoaÄo. 1997b. Word Order and Constraint Interaciton. Ms., Holland Institute of Generative Linguistics/Leiden University. Costa, JoaÄo. 1998a. Parameters vs. Violable Constraints in the Analysis of DiscourseCon®gurationality. In M. Catherine Gruber, Derick Higgins, Kenneth Olson, and Tamra Wysocky (eds.), CLS 34: The Panels, 17±30. Chicago Linguistic Society, University of Chicago. Costa, JoaÄo. 1998b. Word Order Variation: A Constraint-Based Approach. The Hague: Holland Academic Graphics. Dik, Simon, Maria E. Ho¨mann, Jan R. de Jong, Sie Ing Djiang, Harry Stroomer, and Lourens de Vries. 1981. On the Typology of Focus Phenomena. In Teun Hoekstra, Harry van der Hulst, and Michael Moortgat (eds.), Perspectives on Functional Grammar, 41±74. Dordrecht: Foris. Ebert, Karen. 1979. Sprache und Tradition der Kera (Tschad), Teil III: Grammatik. Berlin: Verlag von Dietrich Reimer. Fassi Fehri, Abdelkader. 1993. Issues in the Structure of Arabic Clauses and Words. Dordrecht: Kluwer. Grimshaw, Jane. 1991. Extended Projections. Ms., Brandeis University, Waltham, Mass. Grimshaw, Jane. 1993. Minimal Projections, Heads, and Optimality. Ms., Rutgers University, New Brunswick, N.J. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28: 373±422. Grimshaw, Jane, and Vieri Samek-Lodovici. 1995. Optimal Subjects. In Jill Beckman, Laura Walsh Dickey, and Suzanne Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 589±605. GLSA, University of Massachusetts, Amherst. Grimshaw, Jane, and Vieri Samek-Lodovici. 1998. Optimal Subjects and Subject Universals. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 193±219. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Hyman, Larry. 1979. Aghem Grammatical Structure. University of Southern California Occasional Papers in Linguistics 7. Department of Linguistics, University of Southern California, Los Angeles. Jarvis, Elisabeth. 1981. Some Considerations in Establishing the Basic Word Order of Podoko. Studies in African Linguistics 12(2): 155±167. Kayne, Richard. 1994. The Antisymmetry of Syntax. Cambridge, Mass.: MIT Press. Keer, Edward. 1996. Anti-*that-Trace E¨ects in Norwegian. Ms., Department of Linguistics, Rutgers University, New Brunswick, N.J. Available as ROA-187 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Keer, Edward, and Eric BakovicÂ. 1997. Have Faith in Syntax. In Emily Curtis, James Lyle, and Gabriel Webster (eds.), Proceedings of the Sixteenth West Coast Conference on Formal Linguistics, 255±269. Stanford, Calif.: CSLI Publications. Available as ROA-200 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Legendre, GeÂraldine. 1998. Focalization in French Stylistic Inversion. Paper presented at the Symposium on Inversion in Romance, University of Amsterdam. Legendre, GeÂraldine, William Raymond, and Paul Smolensky. 1993. An Optimality-Theoretic Typology of Case and Grammatical Voice Systems. In Joshua S. Guenther, Barbara A. Kauer,

352

Vieri Samek-Lodovici

and Cheryl C. Zoll (eds.), Proceedings of the Nineteenth Annual Meeting of the Berkeley Linguistics Society, 464±478. Berkeley Linguistics Society, University of California, Berkeley. Legendre, GeÂraldine, Colin Wilson, Paul Smolensky, Kristin Homer, and William Raymond. 1995. Optimality and Wh-Extraction. In Jill Beckman, Laura Walsh Dickey, and Suzanne Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 607±636. GLSA, University of Massachusetts, Amherst. McCarthy, John, and Alan Prince. 1993. Generalized Alignment. Yearbook of Morphology, 79±153. Moutaouakil, Ahmed. 1989. Pragmatic Funcitons in a Funcional Grammar of Arabic. Dordrecht: Foris. MuÈller, Gereon. 1997. Partial Wh-Movement and Optimality Theory. Linguistic Review 14: 249±306. Neeleman, Ad, and Fred Weerman. 1999. Flexible Syntax: A Theory of Case and Arguments. Dordrecht: Kluwer. Newman, Paul. 1974. The Kanakuru Language. West African Languages Monograph Series 9. Ilkley: Scolar Press. OrdoÂnÄez, Francisco. 1995. Post-Verbal Asymmetries in Spanish. Ms., CUNY, New York. OrdoÂnÄez, Francisco. 1997. Word Order and Clause Structure in Spanish and Other Romance Languages. Doctoral Dissertation, CUNY, New York. OrdoÂnÄez, Francisco, and Esthela TrevinÄo. 1995. Los sujetos y objetos preverbales en espanÄol. Paper presented at the Fifth Colloquium on Generative Grammar, CorunÄa, Spain. Pinto, Manuela. 1997. Licensing and Interpretation of Inverted Subjects in Italian. Utrecht: UiL OTS Dissertation Series. Prince, Alan. 1999. A Proposal for the Reformation of Tableaux. Ms., Rutgers University, New Brunswick, N.J. Available as ROA288 at Rutgers Optimality Archive: http://www. ruccs.rutgers.edu/roa.html. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Reinhart, Tanja. 1995. Interface Strategies. OTS Working Papers, OTS-WP-TL-95-002. Research Institute for Language and Speech, Utrecht. Rochemont, Michael. 1986. Focus in Generative Grammar. New York: Cambridge University Press. Rochemont, Michael, and Peter Culicover. 1990. English Focus Constructions and the Theory of Grammar. New York: Cambridge University Press. Rooth, Mats. 1985. Association with Focus. Doctoral dissertation, University of Massachusetts, Amherst. Saccon, Graziella. 1993. Postverbal Subjects. Doctoral dissertation, Harvard University, Cambridge, Mass.

Crosslinguistic Typologies in Optimality Theory

353

Samek-Lodovici, Vieri. 1996. Constraints on Subjects: An Optimality Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Samek-Lodovici, Vieri. 1998. Opposite Constraints: Left and Right Focus-Alignment in Kanakuru. Lingua 104: 111±130. Samek-Lodovici, Vieri, and Alan Prince. 1999. Optima. Technical Report TR-57, Center for Cognitive Science, Rutgers University, New Brunswick, N.J. Available as ROA-363 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Speas, Peggy. 1997. Constraints on Ranked Constraints. Paper presented at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest 1997, Baltimore. Tuller, Laurice. 1992. The Syntax of Postverbal Focus Constructions in Chadic. Natural Language and Linguistic Theory 10: 303±334. VallduvõÂ, Enric. 1992. The Informational Component. New York: Garland. Woolford. Ellen. 1995. Object Agreement in Palauan: Speci®city, Economy, and Optimality. In Jill Beckman, Laura Walsh Dickey, and Suzanne Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 655±700. GLSA, University of Massachusetts, Amherst. Zubizarreta, Maria Luisa. 1992. Word Order in Spanish and the Nature of the Nominative Case. Ms., University of Southern California, Los Angeles. Zubizarreta, Maria Luisa. 1998. Prosody, Focus, and Word Order. Cambridge, Mass.: MIT Press.

Chapter 12 Form and Function in the Typology of Grammatical Voice Systems

Peter Sells

This chapter looks at the question of the inventory of voice systems in an Optimality Theory (ot) approach. By considering the relative members in a given inventory, we can also make substantive claims about the range of typological variation in voice systems. The notion of ``form'' here concerns the actual voice markers, while the notion of ``function'' concerns the abstract levels of discourse prominence that such markers represent, and the contrasts in prominence that di¨erent forms express. One leading idea that I will argue for, in section 12.1, is that form and function do not necessarily match. In section 12.2, I discuss the notion of prominence that underlies the function of voice markers, and present a system of representation inspired by Payne's (1994) work on Cebuano. This concludes the ®rst part of the chapter, which I hope will provide the basis for more detailed work on the notion of prominence in the input. The second part of the chapter is devoted once more to form: section 12.3 looks at what information the output forms represent, and how they do that; this is the familiar domain of ``linking theory,'' where thematic roles are associated with grammatical functions and levels of prominence. The usual approach to determination of outputs in ot is that individual candidate forms are evaluated in terms of markedness and faithfulness with respect to an input. In section 12.4, introducing ideas from Dispersion Theory, I move from this notion of outputs to one where the output is an inventory of voice forms, di¨erentiated by contrasting prominence pro®les, and where di¨erent inventories (sets of voice forms) are evaluated relative to each other. For a given language, its voice system will be the best inventory of forms relative to a ranking of constraints for evaluation. A di¨erent language will have a di¨erent ranking of the constraints, and hence a di¨erent inventory. Section 12.5 presents examples of voice inventories that correspond to the major typological variants recognized in the literature on voice. The chapter is somewhat programmatic, but nevertheless I hope that the reader will be able to get an idea of how the relation between form and function in voice systems should be conceived, and how we might develop an ot analysis based on that conception.

356

12.1

Peter Sells

Voice in Optimality Theory

Legendre, Raymond, and Smolensky (1993; hereafter lrs) present a wide-ranging proposal to account for typological variation in case and voice systems. While lrs is a ground-breaking work for ot morphosyntactic typology, it perhaps idealizes too far from the surface forms that it aims to describe, and con¯ates aspects of case marking with aspects of argument-to-function linking. In this chapter, I will just concentrate on linking, ignoring case assignment completely, and I will con®ne my discussion to transitive verbs with notional agent (a) and patient ( p) arguments. The lrs system assigns a notion of ``prominence'' to an argument via its grammatical function and case properties, providing three kinds of linking for all languages, as shown in (1). (1) a. Active (agent and patient have equal prominence; ``AP'' in lrs) b. Passive (patient is asymmetrically more prominent than agent; ``aP'' in lrs) c. Antipassive (agent is asymmetrically more prominent than patient; ``Ap'' in lrs) These three types constitute the inputs in the ot system; the outputs are the manifestations of these prominence relationships in terms of grammatical function and case. In the next section I will address the complex issue of prominence, a topic on which lrs say no more than is shown in (1). In the remainder of this section, I consider more the question of determining in an abstract sense what the ot outputs in a theory of voice should be, and what relationship to the inputs they should have. 12.1.1 Outputs I: Parsing the Input I believe that the lrs system fails to relate to the overt voice forms that languages have in two ways, which I address in this and the following subsection, considering input-output relations and the role of voice morphology respectively. In this subsection, I discuss the fact that the relation between prominence as de®ned in discourse-pragmatic terms and the prominence relations provided by voice systems is somewhat indirect. Languages do not necessarily provide voice forms that exhaustively and perfectly parse the prominence relations shown in (1), as is presumed in lrs. As stated, the lrs analysis presumes that each language has a canonical transitive linking, with alternatives for passive and antipassive (for inputs ``aP'' and ``Ap'' respectively). Antipassive in English, for example, is taken to be instantiated by examples with dropped objects such as John ate. Yet it seems inappropriate, if not misleading, to consider that English has antipassive in the same way that, say, West Greenlandic has. One response to this charge might be a statement along the following lines: Yes, it is true that antipassive manifests itself in the grammar in a much more thoroughgoing way in West Greenlandic than in English. However, this should

Form, Function in Typology of Grammatical Voice Systems

357

not be viewed as a consequence of some di¨erence in the abstract properties of the voice systems of the two languages, but rather as a consequence of the functional load that antipassive must bear in an ergative-absolutive language. In fact, I will take the view expressed in this last statement very seriously, and make it a central part of the analysis I will developÐthat the key to understanding voice systems lies precisely in how languages choose to share out the ``functional load'' among elements in an inventory of voice forms. Languages clearly di¨er in what the inventory of their given voice system is, and we ®nd a range of inventories Ðfrom languages with no voice alternations whatsoever to languages with severalway voice systems, such as languages of the Philippines. In fact, in developing a linguistic analysis, the recognition of just one language with no voice alternations leads us to an important point, glossed over in lrs: namely, that voice systems may fail to represent the input information with full faithfulness. The canonical transitive output expression for any language in lrs is A1 P2 : the A argument is expressed as the least marked in terms of grammatical function and case, and the P argument is expressed as the next least marked form. For English, this would correspond to a nominative subject and an accusative object. Now, it seems uncontroversial that English uses this form for both canonical inputs AP and also antipassive inputs Ap; the notion of antipassive morphology, or an antipassive construction, just does not seem to ®gure in the grammar of English. So while expression of an Ap input as an A1 P2 output might be somewhat unfaithful, it also might be the best the language can do, given other constraints in the morphosyntax. The limiting case would be a language that had no voice morphology, which would of necessity have to express all three inputs with a single surface form.1 Following lrs and Aissen (1998b), I assume that the input itself contains no more information relevant to the determination of voice than prominence informationÐ for example, I take it that the input does not contain a piece of ``passive'' information for a Passive output, and an ``antipassive'' for Antipassive, and so on. This would seem to rob the ot account of most of its predictive power, by building properties of the form of the intended output into the input. In the view I adopt here, the input is purely abstract information about predicate-argument structure and argument prominence, and the outputs should show how that information is structurally realized morphosyntactically in a given language. The usual approach of an ot analysis is to model a compromise between markedness constraints and faithfulness constraintsÐfull weight given to faithfulness provides as many output forms as there as distinctions in the input, and full weight to markedness gives a single output for all inputs (full neutralization). However, I will go beyond this basic system in my proposal for voice systems, introducing the notion of an inventory of outputs, rather than a single output, where the inventory is determined largely by markedness constraints. The bulk of the chapter looks at output

358

Peter Sells

properties, properties of the candidates in the output inventory, without fully determining how faithfulness is evaluated. In practice, determining faithfulness with respect to input prominence is extremely di½cult, for two main reasons: ®rst, as outlined in section 12.2, the nature of the input in terms of prominence is not well understood; and second, the overt expression of prominence relationships involves not only voice marking, but also case marking, word order, and possibly intonation. Nevertheless, despite the serious issue of faithfulness to inputs with respect to the output expression of prominence information, I hope to map out some of the territory that an ot approach (perhaps any approach) would have to cover. 12.1.2 Outputs II: Relative Markedness Another question on which lrs say little concerns the markedness relations between the voice forms in a given language. For example, in English, active voice is morphologically unmarked, and passive is a marked variant. In an ergative language, antipassive is the marked variant. Any theory of voice should express these markedness relationships, and an ot analysis ought to be well suited to the task, in that ot is essentially a theory of markedness. In fairness, lrs do imply that each linking form that gives a case-marking pattern di¨ering from the canonical transitive pattern will have a morphological mark on the verb. But this implies that English should have antipassive morphology, for example. In contrast to this view, I will interpret the morphosyntactic evidence strongly, as the guiding factor in determining the inventory of voice forms. Because it has no distinguishable antipassive form, I will conclude that English really does not have an antipassive form, and therefore it has no output form that perfectly expresses an Ap input, though of course it will have a form that can best express that Ap input. In fact, this is precisely what one would expect in an ot systemÐa perfect relationship between inputs and outputs does not necessarily have to exist. I should make explicit a deliberate con¯uence of terms in my discussion: when a linking pattern is marked in the sense of markedness theory, in other words when the pattern of prominence it expresses is noncanonical, the verbal form expressing that pattern bears a morphological mark, to signal its particular status. As noted by Aissen (1998b), this is precisely the kind of situation that markedness theory is intended for. So, for English, as compared to the passive, a canonical transitive verb will be assumed to be the unmarked form not only in terms of what discourse-pragmatic properties of its arguments it expresses, but also in terms of its own morphological marking. This seems uncontroversial to me, but the direct connection between these two senses of ``being marked'' can be nicely exploited in an ot approach. I will refer to this idea as the Marking of Contrasting Forms.

Form, Function in Typology of Grammatical Voice Systems

359

(2) Marking of Contrasting Forms A marked form has a contrasting function. Since marked and contrasting presuppose an unmarked standard of comparison, I use the notion of an inventory of candidate outputs to give formal substance to (2): a set of basic constraints give a single output with a certain function (here, the ability to express certain prominence relationships). Then I allow other outputs to be added in along with this ®rst element into the output inventory, so long as they (a) di¨er in function and (b) bear a morphological mark (see section 12.4.1). Exactly how a language expresses voice alternations is not just a matter of morphology, of course. In English, there is no dedicated passive morpheme, but rather ``passive'' is expressed by the past participle form of the verb, in conjunction with the other syntactic elements that form the passive construction. Nevertheless, it is uncontroversial that the passive is ``marked'' in terms of its expression compared to the active. I will assume that all relevant voice alternations are distinguished by at least some morphological marking, possibly accompanied by di¨erences in syntactic structure compared to the canonical transitive. 12.1.3 Prominence Relationships: Demotion and Promotion Most approaches to voice systems have assumed that there is a canonical basic form, with other forms derived via demotion of arguments, and that morphosyntactic demotion of an argument corresponds to an expression of a lower level of prominence by that argument. However, research on the languages of the Philippines has consistently shown that voice alternations typically do not demote any argument (see Mulder and Schwartz 1981; Shibatani 1988; de Wolf 1988; Kroeger 1993), and I believe that this may be a more general feature of the whole Austronesian group. I will propose here that such languages have morphology that ``boosts'' the prominence of the subject without altering the prominence of the nonsubject. This proposal will be supported by evidence from the literature on discourse studies (e.g., Payne 1994), and I suggest that it means that all Philippine voice forms potentially show some asymmetry in prominence relations among the two arguments. Exactly how the Philippine-type languages boost the prominence of the subject is not crucial to the discussion here. In Sells (1998, 2000) I have argued that various types of evidence point to the function of the voice markers as linking an argument as the subject, and simultaneously instantiating it as a null pronominal. Given that part of the analysis, the overt nominative NP seen in Philippine clauses is not the subject as such, but rather an adjunct coindexed with the null pronominal subject. This is like the accounts of ``pronominal argument'' languages given in Jelinek 1984 and Baker 1991, or the pronominal instantiation of agreement marking in Bresnan and Mchombo 1987, though restricted to the subject argument. I will assume that this is

360

Peter Sells

how subjects can have boosted prominence but will not address the particulars of the analysis further. 12.1.4 Summary In summary, the ot analysis will build on inputs that express abstract predicateargument and argument prominence, and the outputs will be morphosyntactic expressions that express that information, to some degree of faithfulness. Compared to the basic transitive output form, which every language has, there might be other voice forms derived by argument demotion and/or promotion (boosting). All such derived forms will express di¨erent prominence relationships from the basic transitive form, and these marked prominence relationships will be mirrored by overt morphosyntactic marking in the output, in accordance with (2). 12.2

Prominence

In this section I introduce some of the empirical and theoretical considerations concerning the notion of prominence that will form the basis of the subsequent analysis. 12.2.1 Distinctions in Levels of Prominence: Inputs The ideas here are informed by a large body of work in the functional-typological literature, which has consistently shown evidence for ®ner distinctions in prominence, culminating in a four-way classi®cation of voice based on discourse-pragmatic features of agent and patient.2 This is shown in table 12.1; prominence here is determined in terms of Scalar Topicality of agent and patient (a and p respectively). Scalar Topicality is itself determined by the discourse measures of Referential Distance and Topic Persistence. Referential Distance is the measure of how far back in discourse the previous mention of a referent is, from a given point; the lower the measure of Referential Distance, the more topical the entity is. Topic Persistence measures how many times in succeeding discourse a referent is mentioned, from a given point; hence the higher the measure, the more frequent and topical the entity in question is. Table 12.1 Scalar topicality (from Payne 1994) Classi®cation

Prominence

Active Inverse Antipassive Passive

a-High a-High/Med a-High a-Low

Topicality p-High/Med p-High p-Low p-High

p topical, but a more so. a topical, but p more so. p nontopical. a nontopical.

Form, Function in Typology of Grammatical Voice Systems

361

Here the labels such as ``Active,'' ``Inverse,'' and so on classify properties of the ongoing discourse, as determined in terms of argument prominence, and are classi®cations of inputs. Since inputs may be unfaithfully parsed, a one-to-one correspondence with the overt verbal voice morphology is not necessary. To represent the inherent scalarity of this discourse information, we could use the notational system in (3) to express the relative prominence of agent and patient. These would be the ot inputs representing the actual properties of prominence of arguments in a full analysis. (3) Prominence relations (OT inputs) a. Antipassive: a g p; a is much more prominent than p. b. Active: a > p; a is more prominent than p. c. Inverse: a < p; a is less prominent than p. d. Passive: a f p; a is much less prominent than p. 12.2.2 Distinctions in Levels of Prominence: Outputs Now, based on the discourse-pragmatic work summarized in table 12.1 and (3), we could take the relational notations in (3) as representations of the input. The lrs system of inputs assumes that an argument is either prominent (represented in lrs by uppercase) or that it is not (represented by lowercase). Let us represent these as levels of prominence of 1 and 0 respectively. Clearly, without further elaboration, levels of just 1 versus 0 will not cover the four types of relations shown in (3), and the obvious way to extend the numerical system I am introducing here is to have three levels of prominence, represented by the integers 2, 1, 0. Intuitively, the di¨erence in prominence between one argument with level 2 prominence and another argument with level 0 prominence is that the ®rst is ``much more'' prominent than the second. If the di¨erence in prominence levels were only 1, the ®rst argument would only be ``somewhat more'' prominent. Let us now use this numerical system for the output expressions. We can relate these prominence levels to overt expression, beginning with the parsing equivalences shown in (4), where x ranges over a and p. For the purposes of this chapter, I assume that the two arguments agent and patient can in principle be realized as subject, object, or oblique, and that these functions correspond to di¨erent levels of prominence. (4) a. obl ˆ x0: Obliques have low prominence. b. s, o ˆ x1: Subject and object have mid prominence. c. s ˆ x2: Boosted subjects have high prominence.

362

Peter Sells

These equivalences embody the standard assumptions that term arguments indicate positive prominence, while obliques correspond to low prominence.3 Their role in the overall system is described shortly. I add in here the idea of high prominence from boosted subjects, an idea motivated further in the following subsection. The recognition of two levels of distinction in prominence is crucial to my approach, as will become clear in section 12.4. Table 12.2 will be a part of GEN, the function that generates candidates; hence, when a given argument is linked as an oblique argument, it will thereby express a low level of prominence. What we are interested in now is a system for representing outputs, and for ease of reference, a way of naming those outputs. I will use the four classi®cations of voice in table 12.1, and augment them with the pre®x ``Super'' when the subject has a prominence of 2. To preserve continuity with the discourse-pragmatic term ``Inverse,'' but to avoid confusion with the voice-related notion of ``inverse'' (in opposition to ``direct''),4 I will use the term ``Inverse-gr'' from now on. This term is based on the notion of ``Inverse Grammatical Relations,'' argued for in Manning 1996 for forms in which agent is linked to object and patient to subject. Although this approach is reminiscent of the proposal in Marantz 1994, it is rather di¨erent. Marantz (1984:196) suggests a much more radical inversion in which the Agent role is assigned by the verb, while the Patient is an external argumentÐe¨ectively it inverts the argument structure in ergative languages. In Manning's approach, all languages crucially have the same argument structures, but di¨er only in the linking relations to surface expression. Universal binding phenomena, for example, refer to properties of argument structure, not surface expression, and hence are largely constant across all languages (Manning 1996:50¨.). Manning argues that linking pattern of Agent to object and Patient to subject is the correct analysis for the majority of ``ergative'' languages, and I will adopt his analysis for canonical ergative languages.5 Roughly speaking, any language considered ergative could have a canonical voice form corresponding to any of the Inversegr or Passive forms belowÐa grammatical form with a patient subject and agent nonsubject. Given these terminological conventions, the range of output forms is presented in table 12.2, with the linking properties as de®ned by the equivalences in (4). Note that Active and Inverse-gr have the same prominence relationships, di¨ering only in the linking of arguments to functions. The prominence relationships a0-p0 and a2-p2 are not listed in table 12.2; they cannot be derived by this system, as described in section 12.3.3. The ``Expressed Prominence'' column in table 12.2 is simply the prominence information due to the linking pattern interacting with (4), and is shown in the ``Linking'' column; the ``Voice type'' is the name we give to that particular array.

Form, Function in Typology of Grammatical Voice Systems

363

Table 12.2 Output forms Voice type

Expressed prominence

Linking

LRS equivalent

Active Inverse-GR Antipassive Passive SuperActive SuperInverse-GR SuperAntipassive SuperPassive

a1-p1 a1-p1 a1-p0 a0-p1 a2-p1 a1-p2 a2-p0 a0-p2

aS 1-pO 1 aO 1-pS 1 aS 1-pOBL 0 aOBL 0-pS 1 aS 2-pO 1 aO 1-pS 2 aS 2-pOBL 0 aOBL 0-pS 2

AP Ap aP

Let me clarify what this represents. The output candidates are the items under ``Linking,'' and their properties are derived as follows: GEN generates candidates with arguments linked to S and O, and (5) determines what prominence values can be associated with those linked arguments. The goal in the later stages of this chapter is then to show which candidate(s) from table 12.2 will actually be available as the output forms of a given language. Starting at the top of table 12.2, the voice type called ``Active'' is a form whose two arguments both have some degree of prominence, and roughly equal prominence. This is achieved by linking the agent to subject and patient to object, which (4) determines will give the a1-p1 prominence pattern. The next line shows the Inversegr form, which has the same prominence relationships but the inverse linking of arguments to functions. The Antipassive and Passive forms have one argument with no prominence (prominence of 0), since that argument is linked as an oblique (see (4a)). The next four forms repeat the pattern, except that, by assumption, they have boosted subjects, whose prominence is of level 2. The forms in table 12.2 are possible outputs that any language may have. Every language does not have to express the same idealized prominence relationships through its voice system, and to study the typology of voice systems, I will speci®cally concentrate on points of di¨erence. Further, these prominence relationships in output forms are idealized; the numerical system adopted in table 12.2 is useful for expressing the linguistic distinctions that will be central to the Dispersion Theory analysis in section 12.4, though I will not specify exactly how these output forms relate to inputs. 12.2.3 Interlude on Input-Output Relations: Cebuano How would the voice markers in a given languageÐsome subset of the forms in table 12.2Ðexpress the prominence relationships in (3)? This is a question that I will not

364

Peter Sells

answer directly in this chapter; in this subsection, I will try to illustrate what factors need to be brought to bear in considering it. To do this, I will summarize the results in Payne 1994 for the Philippine language Cebuano. Philippine languages play a central role in my approach because the primary function of their voice systems does not involve argument demotion. The main voice forms that all Philippine languages display are known as ``Actor Focus'' and ``Goal Focus.'' ``Actor Focus'' is the usual Philippine linguistics term for a verb whose agent-like argument is the subject. In terms of thematic relations, ``Actor'' covers roughly agent, perceiver, and experiencer (see, e.g., Schwartz 1976). ``Goal Focus'' refers to a form whose patientlike argument is the subject. More strictly, this argument is considered to be the ``end point of the action'': hence the term ``goal.''6 The term ``Focus'' is potentially somewhat misleading. While there has been considerable debate as to whether what I am calling subject here is really subject or topic, or an element with some characteristics of both,7 the argument in question is uncontroversially not a focus. Hence we should view a term like ``Actor Focus'' as indicating a verbal form that gives some sort of centrality to the Actor (agent) argument, a centrality realized as subjecthood in terms of linking theory. A ``Goal Focus'' form will accord similar centrality to the patient(-like) argument. I will use the abbreviations af and gf below. There is less controversy in Philippine studies about the fact that the nonsubject argument (the argument ``not in focus'') is not demoted to oblique, but rather retains properties of a core argumentÐa ``term'' in Relational Grammar terminology. I will simply assume here that such arguments are objects; the arguments for this position can be found in Kroeger 1993:40±46 and Foley and Van Valin 1984:139, among others. Kroeger presents speci®c arguments that the nonsubject agent has properties that show it is a term and not an oblique. Foley and Van Valin (1984:136±139) cite data from Walton's (1983) work on Sama, which shows that the agent is marked di¨erently in gf and in passive constructionsÐas a bare np when it is the nonsubject in gf, but preceded by a preposition in passive (as shown in (6) and (7)). Other Philippine languages of this type appear to be Northern Kankanay (Porter 1979) and Umiray Dumaget (MacLeod 1972). In passive, the agent phrase is fundamentally optional, in contrast to its relatively obligatory status in gf. These kinds of arguments show the termhood of the nonsubject agent in gf constructions.8 There is also general, though not total, consensus in the Philippine linguistics community that the gf form is either unmarked, or the least marked, in terms of morphological marking, and it typically appears more frequently in text counts than the af form does (for discussion see Foley and Van Valin 1984:136¨. as well as Kroeger 1993:55±56). The voice markers express both argument linking and aspectual information, and hence even the least marked gf form may have some overt

Form, Function in Typology of Grammatical Voice Systems

365

expression (e.g., in the Cebuano example (5a)). The view of gf as unmarked with respect to af is reinforced by Payne's results in table 12.3 below: the gf can express any voice type, once word order is factored in. Some simple Cebuano data, taken from Bell 1979, illustrate the basic voice alternation. Philippine sentences are verb initial, and nonpronominal NPs are typically preceded by a case clitic or preposition. With the gf form, in (5a), the patient argument is the nominative subject, preceded by the nominative clitic ang, and the agent bears genitive case. Because this is not a true passive, I have not translated it as a passive. In (5b), the af voice form coincides with a nominative agent. (5) a. luto'-on sa babaye ang bugas sa lata cook-GF gen woman nom rice obl can `The woman will cook the rice in the can.' (rice ˆ subj ) b. mag-luto' ang babaye ug bugas sa lata AF-cook nom woman obl rice obl can `The woman will cook the rice in the can.' (woman ˆ subj ) As noted, there seems to be good evidence that the genitive agent in (5b) and the ``oblique'' patient in (5b) are term arguments; here I have simply followed Bell's notations for the cases. Now let us look at some Sama data from Walton 1983:121, 131. The examples in (6) illustrate the contrast between the GF and AF forms, and those in (7) illustrate the contrast between the GF and passive forms. (6) a. j-b'lli ku taumpa' GF-buy lsg.NON-NOM shoes `I bought the shoes.' (shoes ˆ subj ) b. N-b'lli aku taumpa' AF-buy lsg.NOM shoes `I bought shoes.' (I ˆ subj ) (7) a. j-b'lla d'nda kiyakan kami GF-cook girl food 1pl `The girl cooked our food.' (girl ˆ obj ) b. j-b-i-lla uk d'nda kiyakan kami GF-PASS-cook OBL girl food 1pl `Our food was cooked by the girl.' (girl ˆ obl ) In terms of linking, then, the af and gf voice markers represent the constellations shown in (8), where no argument is demoted. (8) a. Goal Focus: patient is subject, agent is object. b. Actor Focus: agent is subject, patient is object.

366

Peter Sells

These correspond, respectively, to the Super-Inverse-gr and SuperActive forms in table 12.2. Even though the af and gf forms do not di¨er via argument demotions, they do not express the same prominence relationships: this will be a crucial part of my account in section 12.4. Rather, they express asymmetric prominence by boosting the subject argument. By the equivalences in (4), and the linking in (8), the hypothesized expression of a gf form will be a1-p2. The prominence expression of the af form will be a2-p1, which di¨ers from (and hence contrasts with) the gf form. Let us now look at the discussion in Payne 1994 to see how the Philippine language Cebuano actually expresses the four input categories in (3) in terms of voice marking plus word order. His ®ndings are shown in table 12.3. For simple transitives, there are just the two voice forms, af and gf; Payne shows that word order is also relevant, and in his notation, ``ap'' indicates that the agent precedes the patient in the surface string, and vice versa for ``pa.'' We see from table 12.3 that Cebuano also uses word order to indicate how prominence is parsed. This is true in many languages, perhaps most: the voice system gives a basic pattern of prominence, with other surface features of the language, notably case, position, prosodic features, and (non)pronominal expression providing extra degrees of di¨erentiation. In largely noncon®gurational languages, such as those of the Philippines, word order can play a direct role in the expression of prominence. We see from table 12.3 that with gf, it is the relative order of a and p that indicates prominence (the ®rst is more prominent); order is less important with af, since af always indicates that agent is more prominent than patient. Consequently, any complete theory of how prominence and voice are related must be able to provide a system on the order of complexity of table 12.3, allowing also for the e¨ects of word order. Evidence such as that summarized in table 12.3 should be the key empirical underpinning for any account of voice: from an extensive and careful study of texts, Table 12.3 Cebuano expression of voice types: Voice markers plus relative order of a and p (simpli®ed from Payne 1994) Input (see (3))

Surface expression

Antipassive

AF plus ap GF plus ap

Active

AF plus ap/pa GF plus ap

Inverse

GF plus pa

Passive

GF plus pa

Note: AF ˆ agent subject; GF ˆ patient subject; ap ˆ agent precedes patient in surface order; pa ˆ patient precedes agent in surface order.

Form, Function in Typology of Grammatical Voice Systems

367

Payne isolated the four input types, and then noted which surface forms are used for each input. His work shows clearly that parsing input prominence must involve various components of the surface expression. The prominence of voice forms that I am proposing here will provide part of the information that must be put into a complex calculation to predict the relationships we see in table 12.3. I will not attempt to provide such a calculation in this chapter. Rather, I devote the rest of the chapter to showing how the more abstract notion of contrast as expressed in the numerical system of output prominences plays a crucial role in the selection of the members of a voice system. 12.2.4 Summary In this section, I have tried to motivate the idea that there are three levels of prominence in the input, and that demotion and promotion (boosting) are both available as means for altering morphosyntactic expression (in the output). I have developed a numerical system (in (6)) for representing the output forms. Finally, I illustrated the complexity that we ®nd in a thoroughly researched case study: the two voices of Cebuano bear a complex relation to the input prominences, and do not obviously represent those inputs faithfully. Hence, I have used well-established results from the functionalist literature to provide the bounds within which prominence-based accounts to voice should be constrained. Rather than the three abstract voice types proposed by lrs based on two levels of prominence, there are three levels of prominence, giving the full range of voice types shown in table 12.2. 12.3

An Approach to Linking

This second part of the chapter focuses just on the inventory of output forms. In this and the following section, I develop a proposal for the ot constraints that will be used to characterize individual voice forms as well as their relationships to other voice forms. 12.3.1 Lexical Mapping Theory The ®rst part of my proposal is that there are four linking constraints that give the underlying ``character'' of the language. These constraints were derived by taking the underlying ideas of Lexical Mapping Theory (lmt; see Bresnan and Kanerva 1989) and interpreting them as negative constraints, leaving argument linking to subject to be an emergent property of the tableau.9 In lmt, the elements of argument structure, expressed as thematic roles, are linked to the four grammatical functions subject, object, (thematically) restricted object, and oblique. To express generalizations about linking, the four functions are de®ned in terms of two features, [Go(bjective)] and [Gr(estricted)], as shown in (9).

368

Peter Sells

(9) Function decomposition in LMT a. subj ˆ [ o, r] b. obj ˆ [‡o, r] c. objy ˆ [‡o, ‡r] d. obl ˆ [ o, ‡r] I will ignore the thematically restricted objects in my discussion here. In English, agent arguments alternate between subjects and obliques (passive by-phrases), but they are never objects; hence, in the lmt classi®cation, agents alternate between the [ o] functions only. Similarly, patients typically only alternate between direct objects (active) and subjects (passive), so patients can be designated as intrinsically [ r]. Another way of thinking of this is that agents are never objects, and patients are never obliques. Generalizing to all languages, the former is the one ``hard'' constraint in all lmt work so far, yet even this is apparently directly contradicted by the evidence from Philippine languages, in which agents seem to function as objects quite freely (see Kroeger 1993; Manning 1996). As discussed in Sells 1998, this poses a very strong challenge to lmt in its original conception. The natural ®rst move to make in an ot system would be to interpret the linking generalizations as violable constraints, rather than as absolute maxims. Generalizing from the negative constraints on agents and patients just mentioned, we could hypothesize the four markedness constraints shown in (10).10 (10) a. b. c. d.

agent is not object: *a ˆ obj patient is not object: *p ˆ obj agent is not oblique: *a ˆ obl patient is not oblique: *p ˆ obl

We could then rank these constraints, to predict what kinds of linking would be less or more marked in a given language. For example, for English, (10a) would be the highest-ranked constraintÐthat is, the worst situation is to have agents linked as objects. In a true ergative language, patients are never objects (so constraint b is highest): patients are either subjects (absolutives), or obliques via the operation of antipassive. In lmt, the linking of a particular argument to subject is not directly built in to the system; rather, there are overall mapping constellations, and every verbal form is expected to link one argument as subject, respecting principles such as those in (9). In contrast, the constraints in (10) keep the spirit of this approach but require no extra stipulation about some argument being subject. Because the constraints are negative constraints on linking to the other functions, the least marked linkingÐregardless of the ranking of constraintsÐwould be to link all arguments as subject: then no constraints would be violated. In standard Lexical-Functional Grammar (see Bresnan 1982), in the context of which lmt has been developed, the unrestricted functions

Form, Function in Typology of Grammatical Voice Systems

369

subject and object can only be instantiated once per clause nucleus. This means that each verbal form can only have one subject, and so the best linking for a twoargument predicate would be to have one argument as subject and one as nonsubject. This is true regardless of the ranking of linking constraints, a property that this approach shares with lrs. Aissen (1997, 1999) has shown that typological implication and explanation can follow from an ot account grounded in strong empirical and functional observations, developed theoretically in the notion of a constraint subhierarchy. My approach here also uses the notion of subhierarchies, though I do not derive them from the alignment of prominence scales the way Aissen does, following the ideas in Prince and Smolensky 1993. Speci®cally, I simply claim that the linking constraints in (10) can only admit certain rerankings, and that (11) must be respected in a grammar (any ranking of constraints). (11) *a, p ˆ obl g *a, p ˆ obj That is, obliques are always more marked than objects, regardless of thematic role (subjects are even less marked than objects). To remind readers of this, I will put a double line between the obl and obj markedness constraints in the tableaux below. 12.3.2 Faithfulness in the Determination of Voice The account I am about to develop does not let faithfulness to input prominence directly determine what voice forms a language has. Rather, my approach is to determine an inventory of voice types in advance of any faithfulness considerations (which are complex; see section 12.2.3). This is motivated by the approach to morphological markedness that I have taken. Let us take the case of English, with Active and Passive forms, and let us suppose an input in which agent is less prominent than patient; hence Passive should be the output form. In a faithfulness-driven approach, Passive will be the winning candidate, since it is more faithful than Active to the input. Now, if faithfulness is to determine the Passive form as the output, Active and Passive must be tied with regard to all higher-ranking constraints, because only in this case can faithfulness make the crucial decision between the two. Hence, no matter which markedness constraints dominate the faithfulness constraint in question, the Active and Passive must be equally (un)marked. At the column in the tableau where the faithfulness constraint applies, the Active form will incur a violation, but the Passive form will not and will be the winner. Now, in what sense is the Passive here a ``marked'' form? According to this scenario this question has no obvious answer, for the Passive is no more marked, and more faithful, than the Active. The approach here is to ``pre®lter'' candidatesÐthe dispersion theory constraints introduced below cause a certain number of candidates to be available as viable

370

Peter Sells

options, and these are the voice-alternating forms that the language has, some of which are marked in the sense of (2). In this approach, Passive must be more marked than Active since Passive links one argument as an oblique, while Active does not, and linking to oblique is always more marked than linking to a term, by (11). Aissen (1998, 1999) presents a di¨erent approach to markedness, by using local conjunction of *j (Avoid zero expression) with constraints relating elements from two domains. For example, the fact that passive is marked but active is not would be expressed by the following rankings in her system; see (12). (12) f*Su/Pat & *jg g . . . *struc . . . g f*Su/Agt & *jg These constraints say that a construction with a patient subject (i.e., passive) is dispreferred, and even more so if there is no morphological mark (expressed by *j), where a domain for that marking (e.g., the verb) can be speci®ed. The lower-ranking constraint *struc penalizes all morphological marks. Hence the constraints shown in (12) have the e¨ect that there is no morphological mark except when there is a patient subject. So, passive is marked but active is unmarked. This system assumes that the choice of agent or patient subject is determined by faithfulness to prominence in the input, but I am not sure that this is straightforward, as discussed above. Additionally, the system has to formally rule out local conjunction of *struc with linking constraints, for that in principle could give exactly the opposite of (12)Ða language with an unmarked passive and marked active. 12.3.3 Unmarked Linking Assuming the prominence equivalences from (4), we can now set up a tableau to determine the unmarked linking in a given circumstance. Leaving boosted subjects out of the discussion for now, four candidates from table 12.2 are potentially well formed, as shown in tableau T12.1. For expository purposes, I rank the four constraints as shown. Consider the ranking of constraints shown here. Candidate [x] is the winning candidate. In terms of markedness in the tableau, [x] is the least marked and hence the Tableau T12.1 *p ˆ OBL

*a ˆ OBL

[w] aS 1 pO 1

[z] aOBL 0 pS 1

*a ˆ OBJ

*

[x] aO 1 pS 1 [y] aS 1 pOBL 0

*p ˆ OBJ

Active *

*

Inverse-GR Antipassive

*

Passive

Form, Function in Typology of Grammatical Voice Systems

371

best, but then we can also say that candidate [w] is the next best, and so on. My proposal will be that the winnerÐthe least marked candidate in terms of the tableau Ðis the morphologically unmarked voice form that a language provides. So (16) characterizes a language in which basic underived nona½xed transitive verbs link as Inverse-gr forms, with agent as object and patient as subject. Naturally, other rankings of constraints would give one of the other candidates as the unmarked form. In the evaluation, I ignore candidates in which no argument is subject: these would induce more violations than any in tableau T12.1. Only a candidate with both arguments linked as subjects would be better than any of the four shown in the tableau, but this is ruled out by the fact discussed in section 12.3.1 that lfg does not allow two subjects in the same nucleus (i.e., GEN does not generate candidates with two subjects). Additionally, I ignore candidates in which both arguments are linked as obliques, since these would be more marked than candidates with at least one term argument. From these considerations, it follows that the nonoccurring prominence relationships of a0-p0 and a2-p2 (i.e., the possible prominence relationships not listed in table 12.2) are not available as viable options; the former would correspond to a verb with two obliques (and no subject), and the latter to a verb with two (boosted) subjects. We can impose further restrictions on the overall descriptive power of the system. In terms of the overall typology, we do not want or need twelve di¨erent constraint rankings. I have proposed that the ranking of the constraints *a ˆ obj, *p ˆ obj and *a ˆ obl, *p ˆ obl must respect (11), reducing to four the number of possible rankings of those constraints with respect to each other. Now let us add in the constraints that boost the subject, or not, which I will represent as s ˆ x2 and s ˆ x1, respectively. The candidates now are forms that parse level 2 or level 1 prominence on the subject, and so the candidate set doubles compared to tableau T12.1, to match table 12.2. The ranking shown in tableau T12.2 is ®nally a ``real'' one, for a Philippine-type language. The winning candidate is the SuperInverse-gr form, and this would correspond to a morphologically unmarked gf form. These two new constraints are added at the beginning of the constraint ranking; they can be reranked with respect to each other but cannot be ranked lower than the linking constraints. That is, I propose that the constraints given so far fall into the ranking hierarchy shown in (13). Hence, the number of candidates that can be single outputs of the basic linking system is four: Active, SuperActive, Inverse-gr, and SuperInverse-gr, any of the upper four candidates in (17); the lower four candidates are all more marked and could not be single outputs. (13) s ˆ x2, s ˆ x1 g *a, p ˆ obl g *a, p ˆ obj Each potential single output in tableau T12.2 will form the basis of the inventory of output candidates for a given language, as described in the following section.

aOBL 0 pS 1

aOBL 0 pS 2

aS 1 pOBL 0

aS 2 pOBL 0

aO 1 pS 1

f aO 1 pS 2

aS 1 pO 1

aS 2 pO 1

Tableau T12.2 Philippine type

*!

*!

*!

*!

S ˆ x2

*

*

*

*

S ˆ x1

*

*!

*p ˆ OBL

*

*!

*a ˆ OBL

*

*!

*p ˆ OBJ

*

*

*a ˆ OBJ

Passive

SuperPassive

Antipassive

SuperAntipassive

Inverse-GR

SuperInverse-GR

Active

SuperActive

372 Peter Sells

Form, Function in Typology of Grammatical Voice Systems

12.4

373

Dispersion Theory

So far, I have illustrated a system of basic linking that provides exactly one output for any given language, on the continuing assumption that the constraints discussed so far outrank any faithfulness constraints. In a language without voice alternations, this would be precisely the single form that the language has at its disposalÐregardless of the input, this single form would have to be the output. However, for most languages, we need to understand how to expand the inventory. One possibility would be to rerank constraints for di¨erent voice alternations.11 That is, we would treat each voice-marking morpheme as indicating a di¨erent ranking from that giving the unmarked linking described above, so that marked morphology would indicate a marked ranking. This would expand the inventory of outputs for the language as a whole in the following way: each ranking of constraints would give one output, as is standard in ot; but if a language allowed two rankings, there would be two outputs, and so on. This would work, if the particular rankings could somehow be tied to the nonexistence of voice morphology in one case (the unmarked linking above), and to the existence of voice morphology in the otherÐ though exactly how this would be achieved is not clear to me. Such a reranking account would therefore have to stipulate the presence or absence of morphological marking, and so does not relate the notions of marked linking and marked expression. Additionally, as I will show below, the e¨ect achieved by simple reranking is not in fact the only force at work in a voice system. I will argue that the nonbasic voice forms must express prominence relationships that contrast with those of the basic form. Since some di¨erent voice forms do not contrast in prominence (i.e., Active and Inverse-gr), these two forms will never appear in the same language, even though there is a trivial reranking of constraints that will map between these two outputs. This shows us that it is not simply mechanical reranking that gives di¨erent inventories with these then representing the typology of voice systems. Rather, there must be some functional contrast between members of an inventory. Once we take this view, we can adopt a single constraint ranking for all voice forms in a given language. 12.4.1 Marking Contrasts What we need is an approach that deals in inventories from the beginning. Fortunately one already exists, in Dispersion Theory (dt), as introduced in Flemming 1996. Flemming provides an account of various phenomena in the domain of phonology where the central notion is one of a form occupying a certain region of analytical space in an inventory. He argues that some general principles govern the overall system; here I focus on (i) ``Maximize the number of contrasts'' and (ii) ``Maximize the distinctiveness of contrasts.''

374

Peter Sells

To the best of my knowledge, dt has not been applied outside the domain of phonology, but it provides a natural perspective for the analysis of morphology. The idea of elements existing that contrast along various dimensions but that nevertheless constitute a single system dividing up a certain space is as familiar in morphology as it is in phonology. In the domain of voice systems, constraints (i) and (ii) have direct application in the present contextÐthe idea being that a contrast in the expression of di¨erent prominence relationships is exactly the notion of contrast relevant in voice systems. Most languages will specify that the system of voice should provide a certain small number of contrasts in expressions of prominence, and they will specify criteria (thresholds) for when two forms are considered distinctive enough. Suppose there is a language with both passive and antipassive morphology, in addition to its basic transitive form. Then that language will be one that requires two contrasts (one fewer than the number of elements in the inventory), and each of the three forms will have di¨ering prominence relationships. In the forms we call passive and antipassive, one argument will be linked as oblique and hence express low prominence; these two forms will be less symmetric in their prominence relationships than the basic transitive. My interpretation of the ideas from dt for voice systems is shown in (14).12 (14) a. Provide a given number (n) of prominence contrasts …0 U n U 3†. b. Provide forms that express distinct prominences. Here, a form will be considered to have ``distinct prominence'' from another form if there is any prominence contrast between the two, and I will just consider systems that have one or two contrasts. Rather than maximizing the number and extent of contrasts, as in (i) and (ii) above, it seems that the voice system just pays attention to the existence of some number of contrasts to some degree. In the previous section, I discussed a basic system of linking that gives a single output for a given language, an output we can refer to as the ``unmarked'' one. Suppose we require, in addition to this unmarked output, that a language gives us one more voice-marking form, with di¨erent prominence assignments, and with a morphological mark. In other words, we force the language to ``Mark one contrast,'' as described in (15a). This is part of a family of constraints, as described in the remainder of (20). (15) Prominence contrast a. ``Mark one contrast'' requires the overall system to provide one extra output form in addition to the basic unmarked linking, such that the prominence relations in the two forms di¨er. (Represented below as mark-1.) b. ``Mark two contrasts'' requires the overall system to provide two extra output forms in addition to the basic unmarked linking, such that the

Form, Function in Typology of Grammatical Voice Systems

375

prominence relations in the three forms di¨er. (Represented below as mark-2.) c. ``Mark three contrasts'' has the same character; given the four basic linking constraints above, a maximum of three contrasts is possible. (Not discussed further here.) To get two outputs, the idea is not to rerank constraints; rather, with a single constraint ranking, we require there to be a contrast, hence two forms. The tableau will be interpreted to provide us with the best two outputs, but we have now added another dimension to the evaluation: not only the best two outputs, but the best two that contrast with respect to each other in the prominence relationships they express. Let's consider ®rst how contrasts would appear, in the sample Philippine-type tableau presented above (see tableau T12.3). Here, all the candidates with nonboosted subjects lose immediately, and of the remaining ones, the prominence relations in the optimal candidate, indicated as [1], are a1-p2. This would be the morphologically unmarked form, corresponding to gf in a real Philippine language, as in tableau T12.2. If we add something to the system to force us to look for another candidate, the next best is indicated as [2], and it has contrasting prominence relations a2-p1. Hence, this candidate could also be an output, and would be morphologically marked, by (2). For Philippine languages, this second output would be the af form, and the language would be ``marking one contrast'' with respect to the gf form. After that, if two contrasts were required, the next best form would be the one marked [3], with relations a0-p2, which would be a Philippine passive form, and so on. The logic of the system, then, is summarized in (16). (16) a. The most optimal candidate will provide the basic, morphologically unmarked voice form of the language in question. b. The next most optimal candidate with contrasting prominence relations will also be in the output set if the language chooses to mark one contrast, and this form will be morphologically marked with respect to the form generated as in (a). c. The linking properties of the voice-marking morpheme predicted to exist by (b) will be determined by the properties of the tableau. d. And so on, for more contrasts in the system. To determine these contrasts, I will add the constraints mark-1 and mark-2, and these will dictate how many elements will be in the voice inventory. 12.4.2 Comparison of Inventories In tableau T12.3, [2] is the least marked form that satis®es mark-1, when compared to the unmarked output [1]. Though tableau T12.3 represents a mechanism for

aOBL 0 pS 1

[3] aOBL 0 pS 2

aS 1 pOBL 0

aS 2 pOBL 0

aO 1 p S 1

[1] aO 1 pS 2

aS 1 pO 1

[2] aS 2 pO 1

Tableau T12.3 Philippine type

*

*

*

*

S ˆ x2

*

*

*

*

S ˆ x1

*

*

*p ˆ OBL

*

*

*a ˆ OBL

*

*

*p ˆ OBJ

*

*

*a ˆ OBJ

Passive

SuperPassive

Antipassive

SuperAntipassive

Inverse-GR

SuperInverse-GR

Active

SuperActive

376 Peter Sells

Form, Function in Typology of Grammatical Voice Systems

377

Tableau T12.4 Expressing one voice contrast: w ˆ aS 2 pO 1; x ˆ aO 1 pS 2; y ˆ aS 2 pOBL 0; z ˆ aOBL 0 pS 2 MARK- MARK1 2 {w}

*!

*

{x}

*!

*

{y}

*!

*

{z}

*!

*

f {w, x}

*

{w, y}

*

{w, z}

*

{x, y}

*

{x, z}

*

{y, z}

*

*p ˆ OBL

*a ˆ OBL

*p ˆ OBJ *

* * * * *! *! *!

* *

*! *!

* *

*

*

{w, x, y}

*!

*

{w, x, z}

*!

{w, y, z}

*!

*

*

{x, y, z}

*!

*

*

*

{w, x, y, z}

*!

*

*

*

*

*

*a ˆ OBJ

*

*

*

* *

*

choosing an output inventory, I still refer to it as a tableau, and I will use the tabular notation as shorthand for a system of inventory comparison; the actual intent of the theory is presented in this section. Strictly speaking, the interpretation that we want for the constraints in (16) is that an inventory containing forms [1] and [2], and no others, is a better inventory than any other possible inventory, if the goal is to mark one contrast. That is, [1] and [2] are the least marked forms that contrast. To illustrate the system fully, in tableau T12.4 I show the actual comparison of inventories, inventories with one to four members. Suppose that we have the four relevant linking constellations listed as w, x, y, and z, and to show e¨ects of the dt contrast constraints, I include mark-1 and mark-2. To simplify things, I leave out the s ˆ x2 and s ˆ x1 constraints.

378

Peter Sells

What tableau T12.4 shows is that only inventories with two members survive the ®rst constraint. The ®rst block of inventories do not have enough members to express one contrast, and the last two blocks have too many members, and hence express too many contrasts.13 Following the contrast constraints, we evaluate each member of each inventory with regard to the linking constraints. The least marked inventory is the one consisting of x and w, and within that inventory form x is the least marked, and so is the morphologically unmarked one. Let us look at this more carefully, considering the output inventory {w, x}. In this inventory, candidate w violates *p ˆ obj and candidate x violates *a ˆ obj. Hence the inventory violates these two constraints. However, these two violations are less severe than the violations of any of the other two-member inventories. So {w, x} is the output. Reversing the ranking of mark-1 and mark-2 will require the inventory to have three members, as shown in tableau T12.5. Here, the best output consists of SuperInverse-gr, SuperActive, and SuperPassive. This, then, is the proposed analysis for a system like that in Sama, with SuperPassive in addition to gf and af forms. Probably there are languages in the Austronesian group that have basic gf and af forms with SuperAntipassive as the next form; this would involve reranking *p ˆ obl and *a ˆ obl in tableau T12.5. Tableaux T12.4 and T12.5 represent the constraints that languages may have to give a certain number of voice contrasts. In the case of tableau T12.5, the three forms {w, x, z} will then be the competitors for faithful expression of any given input. To avoid the rather cumbersome representations in these tableaux, in the rest of this chapter I will use the more familiar compact tabular representations, ranking candidates relative to each other but allowing more than one candidate to be classi®ed as an output. 12.4.3 Nonboosted Subjects I have illustrated the system so far with boosted subject candidates playing the crucial role. However, in contrast to tableau T12.3, in a language with simple unboosted Inverse-gr as the unmarked form, Active will not be the next alternative, since these two forms do not contrast in prominence relations. This is illustrated formally in tableau T12.6. For additional compactness, I collapse the two constraints on subject prominence into s ˆ x2 g s ˆ x1 or s ˆ x1 g s ˆ x2, and just show the candidates that satisfy this requirement. The key di¨erence with tableau T12.3 is that the uppermost candidate in tableau T12.6 is not available, because its prominence relations are not distinct from the unmarked candidate marked [1]. This is indicated for now by the [*] notation. Active is not a candidate with prominence relations distinct from Inverse-gr (and vice versa). If we look for the next best candidates, they are [2] and [3], in that order. From these considerations it follows that any language without boosted subjects will

Form, Function in Typology of Grammatical Voice Systems

379

Tableau T12.5 Expressing two voice contrasts: w ˆ aS 2 pO 1; x ˆ aO 1 pS 2; y ˆ aS 2 pOBL 0; z ˆ aOBL 0 pS 2 MARK- MARK2 1 {w}

*!

*

{x}

*!

*

{y}

*!

*

{z}

*!

*

{w, x}

*!

{w, y}

*!

{w, z}

*!

{x, y}

*!

{x, z}

*!

{y, z}

*!

*p ˆ OBL

*a ˆ OBL

*p ˆ OBJ *

* * * * * *

*

*

* *

*

*

*

*

f {w, x, z}

*

{w, y, z}

*

*!

*

{x, y, z}

*

*!

*

*

*

*

*

*

*a ˆ OBL

*p ˆ OBL

*p ˆ OBJ

*a ˆ OBJ

*!

* *

{w, x, y}

{w, x, y, z}

*a ˆ OBJ

*! *

*

*

*

* *

*

Tableau T12.6 Nonboosted patient subject S ˆ x1 g S ˆ x2 [*] aS 1 pO 1

*

[1] aO 1 pS 1

*

[2] aS 1 pOBL 0 [3] aOBL 0 pS 1

Active

* *

Inverse-GR Antipassive Passive

380

Peter Sells

have Passive or Antipassive if it has one contrast enforced, and it will have the other voice form if it requires two contrasts. One crucial e¨ect of the contrast constraints is that we do not need radical reranking of the four linking constraints to ensure that Antipassive is the ®rst alternative candidate in tableau T12.6. That is, we could e¨ectively eliminate the [*] candidate in tableau T12.6 by making it the most marked; to do that, we could take the tableau and move *p ˆ obj to be the highest constraint. From such a ranking, Antipassive would indeed emerge as the ®rst alternate, and Passive as the second (see (17a))Ðthe four voice forms would be ordered in terms of markedness as shown in (17b). In contrast to the dt approach, the more radical reranking approach considered here would of necessity give Active as the fourth possible output in a hypothetical language with four voice forms: Active is a possible output in the language's inventory of voice forms. As shown in (17b), this would allow the constraint ranking to describe a language in which Active is the most marked voice form, a language type that I do not believe exists. (17) a. *p ˆ obj g *a ˆ obl g *p ˆ obl g *a ˆ obj b. Inverse-gr (unmarked), Antipassive, Passive, Active (most marked) Allowing the object linking constraints to be reranked with respect to the oblique linking constraints in this way would open the door to twelve-way typological variation, in e¨ect allowing any (re)ranking of the four linking constraints. In the dt approach with the contrast constraints, we can eliminate the [*] candidate without moving the *a ˆ obj constraint, and in this way signi®cantly limit the range of variation that the system allows. Further, we predict that Active could not possibly be an occurring voice form in a language characterized by tableau T12.6, in contrast to the language described by (17a). 12.5

Voice Systems

Now I discuss brie¯y the di¨erent variants generated by the system that I have outlined. 12.5.1 Argument-Term Languages The ®rst class of languages I look at are those with boosted subjects, and with SuperInverse-gr as the unmarked form (see tableau T12.7). This is the classic Philippine type, with SuperActive as the ®rst alternate form; in either voice, both arguments are linked as terms. As in tableau T12.6, the unmarked form is indicated by [1], the next best form by [2], and so on. For simplicity I do not include mark-2. We see that such a language has an unmarked gf (SuperInverse-gr) form; the af (SuperActive) form would be the next best and available candidate, and would bear a

[3] aOBL 0 pS 2

[4] aS 2 pOBL 0

[1] aO 1 pS 2

[2] aS 2 pO 1

Tableau T12.7 Argument term S ˆ x2 g S ˆ x1

MARK-1

*

*p ˆ OBL

*

*a ˆ OBL *

*p ˆ OBJ

*

*a ˆ OBJ

SuperPassive

SuperAntipassive

SuperInverse-GR

SuperActive

Form, Function in Typology of Grammatical Voice Systems 381

382

Peter Sells

morphological mark due to (2). These would be the voice forms in a two-way voice system, as in the af and gf forms of Cebuano described in section 12.2.3. If two contrasts were enforced, as is found in Sama, the next form would be SuperPassive, according to tableau T12.7, and this would also be morphologically marked with respect to the unmarked winner [1]. It should be noted that there are no morphemes designated as having SuperActive and SuperPassive properties in this view; rather, the morphemes just mark the fact of a distinction, and it is the tableau that dictates what the properties of the morphemes must be. Thus tableau T12.7 is to be interpreted as describing a language in which output [1] is the unmarked form, and output [2] is a form with a morphosyntactic mark of its distinct character, that character being the di¨erent prominence relationships as shown. For [2], it is not that the SuperActive morpheme is endowed with special morphosyntactic properties that defeat the normal scheme of the language linking the agent as subject and the patient as object. (Consider the idea in GovernmentBinding theory that the passive morpheme suppresses the external argument and robs the verb of its ability to assign case.) Rather, SuperActive is an abstract linking pattern that the language allows, and the SuperActive morpheme registers the fact that that pattern is not the only linking pattern in the language, and in fact not the least marked pattern. The symmetry of the system also allows for languages of the Philippine type in which SuperActive is the unmarked form, with SuperInverse-gr as the ®rst alternate. I am not sure if such languages have been documented; if they do not exist, I have no explanation. 12.5.2 Patient-Object Languages with Passive Now we move to languages without boosted subjects, and begin with a language like English. In English, the canonical case-marking patterns are Nominative-Accusative, but there is no reason why a ``syntactically accusative'' language with these voicemarking relations may not have ergative-absolutive surface case-marking patterns (see section 12.5.3). In terms of the linking constraints, tableau T12.8 di¨ers from tableau T12.7 in that the ranking of the lowest two is reversed. This leaves a patient object in the least marked form. The key here is that the second candidate in tableau T12.8 is not available, since its prominence relations are not distinct from the unmarked candidate marked [1], as shown by the * under the ``Mark one contrast'' constraint: Inverse-gr is not a candidate with distinct prominence relations. Strictly, what this represents is the idea that an inventory consisting of the upper two candidates in the table would contain the least marked forms, but would not satisfy mark-1. The inventory consisting of [1] and [2] contains the least marked forms that satisfy mark-1.

[2] aOBL 0 pS 1

aS 1 pOBL 0

aO 1 pS 1

[1] aS 1 pO 1

S ˆ x1 g S ˆ x2

Tableau T12.8 Patient-Object with Passive

* vs.[1]

MARK-1

*

*p ˆ OBL

*

*a ˆ OBL

*

*a ˆ OBJ *

*p ˆ OBJ

Passive

Antipassive

Inverse-GR

Active

Form, Function in Typology of Grammatical Voice Systems 383

384

Peter Sells

From this it follows that a language as in tableau T12.8 will have a Passive morpheme if it has one contrast enforced, and it will have an Antipassive morpheme if it requires two contrasts. Again, the properties of the passive morphosyntax are not inherent in the morphemes and structures used but are dictated by the tableau. Moreover, the fact that Passive is the ®rst alternative candidate does not require the last four linking constraints to be radically reranked compared to the ranking(s) for the other typological variants. If we interpret the functionalist aspects of ``Mark one contrast'' to be even stronger, namely, ``Link a di¨erent argument as subject,'' then in fact the order of *p ˆ obl and *a ˆ obl becomes irrelevant. Consider English, as shown in tableau T12.8: we start out with Active, and then we need to ®nd one form that provides a di¨erent set of prominence relations and a distinct subject. The next best form, Inverse-gr, does not provide a prominence contrast. Suppose now, di¨erent from tableau T12.8, that the next best is Antipassive. This provides a prominence contrast, but no new subject (still the agent). Hence, the only functionally useful form would be Passive. The constraint ``Link a di¨erent argument as subject'' can be used to give the usual typological link that languages with unmarked Active have Passive while languages with unmarked Inverse-gr have Antipassive. Placing particular weight on this constraint could further limit a given language's inventory to just two forms, since there are only two arguments that can be linked to subject. However, I suspect that it is not universally true that all languages with Active have Passive before or instead of Antipassive, and similarly for Inverse-gr languages.14 Regardless of the syntactic voice system, in some languages the surface case marking seems most closely related to the need for Passive or Antipassive. Languages can certainly mark more than one contrast: many examples can be found of languages, from di¨erent families, that have both Passive and Antipassive forms. For example, both are found in the Mayan language Mam (England 1988), the Austronesian language Chamorro (Cooreman 1988), and Labrador Inuttut (Grimshaw and Mester 1985); Dixon (1994:149¨.) presents several other languages. Rude (1988) discusses the form and function of Passive and Antipassive in Nez Perce, observingÐas expectedÐthat ``Ergative'' (my Inverse-gr here) is used when agent is more prominent than patient. The Antipassive is used when the agent is much more prominent, and the Passive is used when the patient has prominence but the agent has none. In summary, ``Link a di¨erent argument as subject'' could not be a requirement on every marked voice form, but plausibly can be viewed as a requirement on the ®rst marked form in some languages. However, there are languages in which even ``Mark one contrast'' does not mean ``Link a di¨erent argument as subject,'' a situation to which I now turn.

Form, Function in Typology of Grammatical Voice Systems

385

12.5.3 Patient-Object Languages with Antipassive Following a suggestion by Rachel Nordlinger, I will assume that there are some languages characterized by tableau T12.9, exempli®ed at least by some languages of Australia. These languages typically have Ergative-Absolutive case marking, and would then have Antipassive as a morphologically marked voice form. The ranking of constraints only di¨ers from that in tableau T12.8 in that *a ˆ obl and *p ˆ obl are ¯ipped. Bittner and Hale (1996) present Warlpiri as an instance of a language of this type. Basque also seems to have the relevant properties, although it has a Passive as well (see Jacobsen 1985), which would be the third form in tableau T12.9.15 Clearly a full analysis of such languages would need to take account of the relations between case marking and discourse prominence, for in some languages it seems that the choice between Ergative and Absolutive case carries information relating to discourse prominence, or at least topicality, and the need for Antipassive comes from the need to get Absolutive marking on agent arguments. Hence the functional role of ``Mark one contrast'' would be to get a di¨erent argument in Absolutive case, which marks the lowest term (the patient object is Absolutive in the transitive, and the agent subject is absolutive in the intransitive antipassive). The ``need'' for this Antipassive comes from the way case marking expresses prominence, rather than being due to some inherent property of the voice system per se. 12.5.4 Agent-Object Languages with Antipassive These are languages in which Inverse-gr is the unmarked form; they are nonPhilippine-type ``ergative'' languages of the familiar variety, canonically with a casemarking array of Ergative-Absolutive for transitive verbs. Many languages may fall into this category, including classic Eskimo-Aleut languages such as Inuit, and various Mayan languages (see Manning 1996:11±33). These ®t the pattern of ``Inversegr'' mapping that lies at the heart of Manning's proposal; the dt contrast constraint means that Active is not an option for the voice system, as shown in tableau T12.10. As with the languages mentioned in the previous subsection, we expect languages in this class to have Antipassive, which interacts with the case-marking patterns and the prominences associated with case. On the assumption that Absolutive marks subjects, the Inverse-gr±Antipassive alternation will link di¨erent arguments as subject, as desired. The linking system also allows for a class of Agent-Object languages with Passive, though it is unlikely that such languages exist. The linking patterns would be ao1 ps1 and aobl0 ps1, respectively, and hence would have functional value only if the case-marking system systematically grouped the transitive object (agent) and the intransitive passive subject (patient). In turn, this would mean that patient subjects

aOBL 0 pS 1

[2] aS 1 pOBL 0

[1] aO 1 pS 1

aS 1 pO 1

S ˆ x1 g S ˆ x2

Tableau T12.10 Agent-Object languages

aOBL 0 pS 1

[2] aS 1 pOBL 0

aO 1 pS 1

[1] aS 1 pO 1

S ˆ x1 g S ˆ x2

Tableau T12.9 Patient-Object with Antipassive

* vs.[1]

MARK-1

* vs.[1]

MARK-1

*

*a ˆ OBL

*

*a ˆ OBL

*

*p ˆ OBL

*

*p ˆ OBL

*

*p ˆ OBJ

*

*a ˆ OBJ

*

*a ˆ OBJ

*

*p ˆ OBJ

Passive

Antipassive

Inverse-GR

Active

Passive

Antipassive

Inverse-GR

Active

386 Peter Sells

Form, Function in Typology of Grammatical Voice Systems

387

of transitives were marked di¨erently from patient subjects of intransitives (the passives), which would be a very unusual case-marking system.16 12.6

Conclusion

There are two key predictions of what I am proposing, following from the fact that the linking constraints are in the subhierarchy (11). Putting aside languages with boosted subjects, the system ensures that one of Inverse-gr and Active is always the ``®rst form'' a language must have. However, without boosted subjects, these two forms do not contrast in prominence relationships, and so at most one could be a surface output. The prediction is then that languages without boosted subjects must have Active or Inverse-gr as the ®rst form, and then must have a demoting Passive or Antipassive as the ®rst marked alternative; if there is a second marked form it will be the other from the demoting forms. The second prediction is the opposite. In a Philippine-type language with boosted subjects, SuperPassive or SuperAntipassive could not be the only marked form. Such a language must have both SuperInverse-gr and SuperActive ®rst, since these do indeed contrast in prominence relationships, and both are less marked than SuperPassive or SuperAntipassive. As far as I know, there are many Philippine(-type) languages that have gf and af forms and no others, and some like Sama that have passive forms as well. But there appear to be none that, say, have gf and an antipassive but no af form, suggesting this implicational hierarchy among the various voice forms. Hence the demoting voice types are available, but only ``after'' the non-argument-demoting SuperActive and Super-Inverse-gr forms have found their expression. In addition to the prediction about the typology of voice systems, since it is the availability of subjects with boosted prominence that gives rise to a Philippinetype voice system, there is a further prediction that such languages will show observably distinct morphosyntactic properties from systems where demotion is the only device for altering prominence. I have argued this to be so in Sells (1998, 2000). At a more general level, we can isolate the following conclusions. First, the properties of voice-marking morphosyntax can be derived from the system of tableaux presented here: the need for contrast motivates what the properties of output expressions such as passive must be. As a result, there is no need to stipulate any of the linking properties of passive. Second, morphologically marked forms can be insightfully related to unmarked forms via the abstract linguistic markedness properties expressed by the tableaux, and the implicational relationships that emerge from those markedness relationships lead to more speci®c predictions about voice inventories, as described in the preceding paragraph. Of course, a complete account must include a better understanding of input-output relationships, and extend to case marking (as lrs attempted) and word order. Never-

388

Peter Sells

theless, I hope to have shown that the line of analysis developed here can provide the basis for a typologically sound account of voice systems. Notes This is a slightly expanded and somewhat revised version of the paper I presented at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest in Baltimore, May 1997. I would like to thank Arto Anttila, Edward Flemming, Vivienne Fong, Yukiko Morimoto, MarõÂa-Eugenia NinÄo, and Rachel Nordlinger for comments on an earlier draft of this chapter, and to Jane Grimshaw and an anonymous reviewer for comments on the pre®nal version. I also owe special thanks to Joan Bresnan for providing invaluable feedback and suggestions during the development of the approach presented here. 1. Some languages seem to have no voice morphology, and so prominence distinctions must be expressed by other means, such as word order or case. Such languages include Sinhala (spoken in Sri Lanka; see Inman 1993), many Australian languages, and many polysynthetic languages. Aissen (1998b) also mentions languages that lack voice distinctions. 2. Key works in the development of this classi®cation include Shibatani 1985 and GivoÂn 1990. My discussion here is based on that in Payne 1994. 3. For example, in Aissen's (1999) work, her proposal is that the unmarked prominence for subjects is high prominence, and the unmarked prominence for all nonsubjects low prominence, based on the two levels of prominence in lrs. 4. For an ot analysis of this, see Aissen 1997; she refers to the phenomenon as ``person-based inverse-marking'' and extends the analysis using markedness constraints in Aissen 1998b. 5. Manning argues that many more languages show patient prominence (or lack of agent prominence, in the terms of Wierzbicka 1981:72) than have been traditionally recognized as ``ergative,'' once argument-linking patterns are included in addition to surface ergativeabsolutive case-marking patterns. I largely adopt Manning's view here, with some re®nements. 6. As is well known, many Philippine languages have more than two voice forms; Cebuano has four, Tagalog has ®ve, and none of them are like true passive or antipassive in involving demotion to oblique of an agent or patient. I would assume that these other voices are e¨ectively applicative varieties of Goal Focus, but I ignore them here. Foley and Van Valin (1984) label Goal Focus ``UF,'' for ``Undergoer Focus,'' which is probably a more accurate term. 7. The classic statement of the issues here is found in Schachter 1976. Kroeger (1993), whose position I adopt, argues that there are many viable ``subjecthood'' properties in Tagalog, once phenomena sensitive only to argument structure (regardless of surface expression) are factored out (see also Manning 1996). Cena (1995) and Schachter (1995) sound some notes of caution. 8. Statistical data that bears on this issue is presented in Shibatani 1988:93. Various counts have shown that the agent in English passives is absent at least 70% of the time, and similar proportions are found in Japanese. In contrast, in Cebuano counts, the agent of the gf form is absent only around 20% of the time. Cooreman (1982) observes that in Chamorro the agent is absent only 20% with the gf form, yet is absent 85% with the passive. 9. These constraints are also inspired in part by notes on linking in OT due to MarõÂa-Eugenia NinÄo. 10. These are similar to the representations in Aissen 1998a. It should be noted that in her approach, ``Agt'' indicates the oblique expression of a passive agent, not simply the agent role.

Form, Function in Typology of Grammatical Voice Systems

389

There are probably also (low-ranked) markedness constraints against linking to subj, but these play no role here. 11. I am not aware of a proposal for this with regard to voice systems, but constraint reranking within case systems is explored in Nakamura 1997. For more discussion, see Woolford (chap. 16, this volume). 12. I discuss the idea that the notion of contrast looks not only at prominence, as stated here, but also at which thematic role is mapped to subject, in section 12.5.2. 13. Strictly speaking, this would be an oversimpli®cation when forms with nonboosted subjects are considered, since not all members will show prominence contrasts (speci®cally, Active and Inverse-gr). Then an inventory with two members may in fact show no prominence contrasts between the two. Adding in one more member would give an inventory of three members, marking only one contrast. Because each candidate member violates one basic linking constraint, an inventory with three members showing just one contrast would lose out to an inventory with just two members, showing the same contrast. Hence, in practice, only inventories with the minimum number of members needed to express the desired number of contrasts will be highly valued. 14. Manning (1996:73±74) discusses and appears to accept the idea that only syntactically ergative languages (my Inverse-gr) have Antipassive forms. 15. lrs discuss as a separate class ``active-stative'' languages, such as Lakhota. I would imagine that these languages are syntactically all of the Patient-Object type (or, possibly, AgentObject), with surface case marking sensitive to thematic properties rather than grammatical function (see, e.g., Woolford 1997; chap. 16, this volume). Since I am not attempting to classify case-marking systems here, I will o¨er no new typological linking variant for such languages. 16. I have assumed that the two pairs of linking constraints obey the subhierarchy in (11). If we relaxed this, the tableau could be set up to give preference to obliques over objects, and would characterize a language with no transitive (underived) verbs. It is not clear if a language of this type exists; in terms of surface properties, it would be a radical ergative-absolutive language with only intransitive verbs. The classic example that may ®t this pro®le is Dyirbal, and Yidiny may be another (see the discussion in Andrews 1985:131¨. and Dixon 1994:175¨.) Without taking this too far, I note that Manning (1996:66±70), following Dixon (1994), questions whether it is in fact true that Dyirbal only allows intransitive clauses, presenting several di¨erent considerations that would lead one to think that the basic unmarked form is Inversegr, rather than Passive. References Aissen, Judith. 1997. The Syntax of Obviation. Language 73: 705±750. Aissen, Judith. 1998. Harmonic Alignment in Morphosyntax. Paper presented at the workshop Is Syntax Di¨erent?, Stanford University, Stanford, Calif., December. Aissen, Judith. 1999. Markedness and Subject Choice in Optimality Theory. Natural Language and Linguistic Theory 17: 673±711. Andrews, Avery. 1985. The Major Functions of the Noun Phrase. In Timothy Shopen (ed.), Language Typology and Syntactic Description, I: Clause Structure, 62±154. Cambridge, England: Cambridge University Press.

390

Peter Sells

Baker, Mark. 1991. On Some Subject/Object Non-asymmetries in Mohawk. Natural Language and Linguistic Theory 9: 537±576. Bell, Sarah. 1979. Cebuano Subjects in Two Frameworks. Doctoral dissertation. MIT, Cambridge, Mass. [Distributed by the Indiana University Linguistics Club.] Bittner, Maria, and Ken Hale. 1996. Ergativity: Towards a Theory of a Heterogeneous Class. Linguistic Inquiry 27: 531±604. Bresnan, Joan (ed.). 1982. The Mental Representation of Grammatical Relations. Cambridge, Mass.: MIT Press. Ã a: A Case Study of Bresnan, Joan, and Jonni Kanerva. 1989. Locative Inversion in Chichew Factorization in Grammar. Linguistic Inquiry 20: 1±50. Ã a. Bresnan, Joan, and Sam Mchombo. 1987. Topic, Pronoun, and Agreement in Chichew Language 63: 741±782. Cena, Resty M. 1995. Surviving without Relations. Ms., Edmonton, Alberta. Cooreman, Ann. 1982. Topicality, Ergativity, and Transitivity in Narrative Discourse: Evidence from Chamorro. Studies in Language 6: 343±374. Cooreman, Ann. 1988. The Antipassive of Chamorro: Variations on the Theme of Transitivity. In Masayoshi Shibatani (ed.), Passive and Voice, 561±593. Amsterdam: Benjamins. de Wolf, Charles. 1988. Voice in Austronesian Languages of Philippine Type: Passive, Ergative, or Neither? In Masayoshi Shibatani (ed.), Passive and Voice, 143±193. Amsterdam: Benjamins. Dixon, R. M. W. 1994. Ergativity. New York: Cambridge University Press. England, Nora C. 1988. Mam Voice. In Masayoshi Shibatani (ed.), Passive and Voice, 525± 545. Amsterdam: Benjamins. Flemming, Edward. 1996. Evidence for Constraints on Contrast: The Dispersion Theory of Contrast. In Chai Shune Hsu (ed.), UCLA Working Papers in Phonology 1, 86±106. Department of Linguistics, UCLA, Los Angeles. Foley, William A., and Robert D. van Valin. 1984. Functional Syntax and Universal Grammar. New York: Cambridge University Press. GivoÂn, Talmy. 1990. Syntax: A Functional-Typological Introduction. vol. 2. Amsterdam: Benjamins. Grimshaw, Jane, and Armin Mester. 1985. Complex Verb Formation in Eskimo. Natural Language and Linguistic Theory 3: 1±19. Inman, Michael V. 1993. Semantics and Pragmatics of Colloquial Sinhala Involitive Verbs. Doctoral dissertation, Stanford University, Stanford, Calif. Jacobsen, William H., Jr. 1985. The Analog of the Passive Transformation in Ergative-Type Languages. In Johanna Nichols and Anthony Woodbury (eds.), Grammar Inside and Outside the Clause, 176±191. New York: Cambridge University Press. Jelinek, Eloise. 1984. Empty Categories and Non-con®gurational Languages. Natural Language and Linguistic Theory 2: 39±76. Kroeger, Paul. 1993. Phrase Structure and Grammatical Relations in Tagalog. Dissertations in Linguistics. Stanford, Calif.: CSLI Publications. Legendre, GeÂraldine, Paul Raymond, and Paul Smolensky. 1993. An Optimality-Theoretic Typology of Case and Grammatical Voice Systems. In Proceedings of the Nineteenth Annual

Form, Function in Typology of Grammatical Voice Systems

391

Meeting of the Berkeley Linguistics Society, 464±478. Berkeley Linguistics Society, University of California, Berkeley. MacLeod, Thomas R. 1972. Verb Stem Classi®cation in Umiray Dumaget. Philippine Journal of Linguistics 3: 43±74. Manning, Christopher D. 1996. Ergativity: Argument Structure and Grammatical Relations. Dissertations in Linguistics. Stanford, Calif.: CSLI Publications. Marantz. Alec. 1994. On the Nature of Grammatical Relations. Cambridge, Mass.: MIT Press. Mulder, Jean, and Arthur Schwartz. 1981. On the Subject of Advancements in the Philippine Languages. Studies in Language 5: 227±268. Nakamura, Wataru. 1997. A Constraint-Based Typology of Case Systems. Doctoral dissertation, State University of New York, Bu¨alo. Payne, Thomas E. 1994. The Pragmatics of Voice in a Philippine Language: Actor-Focus and Goal-Focus in Cebuano Narrative. In Talmy GivoÂn (ed.), Voice and Inversion, 317±364. Amsterdam: Benjamins. Porter, Doris. 1979. Northern Kankanay Morphology. Studies in Philippine Linguistics 3: 20±62. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint interaction in generative grammar. Ms., Rutgers University, New Brunswick, N.J. Available at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Rude, Noel. 1988. Ergative, Passive, and Antipassive in Nez Perce: A Discourse Perspective. In Masayoshi Shibatani (ed.), Passive and Voice, 547±560. Amsterdam: Benjamins. Schachter, Paul. 1976. The Subject in Philippine Languages: Actor, Topic, Actor-Topic, or None of the Above. In Charles N. Li (ed.), Subject and Topic, 491±518. New York: Academic Press. Schachter, Paul. 1995. The Subject in Tagalog: Still None of the Above. Ms., UCLA, Los Angeles. Schwartz, Arthur. 1976. On the Universality of the Subject: The Ilocano Case. In Charles N. Li (ed.), Subject and Topic, 519±543. New York: Academic Press. Sells, Peter. 1998. The Function of Voice Markers in the Philippine Languages. In Steven G. Lapointe, Diane K. Brentari, and Patrick M. Farrell (eds.), Morphology and Its Relation to Phonology and Syntax, 111±137. Stanford, Calif.: CSLI Publications. Sells, Peter. 2000. Raising and the Order of Clausal Constituents in the Philippine Languages. In Ileana Paul, Vivianne Phillips, and Lisa Travis (eds.), Formal Issues in Austronesian Linguistics, 117±143. Dordrecht: Kluwer. Shibatani, Masayoshi. 1985. Passives and Related Constructions: A Prototype Analysis. Language 61: 821±848. Shibatani, Masayoshi. 1988. Voice in Philippine Languages. In Masayoshi Shibatani (ed.), Passive and Voice, 85±142. Amsterdam: Benjamins. Walton, Charles. 1983. Sama Verbal Semantics: Classi®cation, Derivation, and In¯ection. Master's thesis, Temple University, Philadelphia. Wierzbicka, Anna. 1981. Case Marking and Human Nature. Australian Journal of Linguistics 1: 43±80. Woolford, Ellen. 1997. Four-Way Case Systems: Ergative, Nominative, Objective, and Accusative. Natural Language and Linguistic Theory 15: 181±227.

Chapter 13 Constraints on Null Pronouns

Margaret Speas

Within the empirical domain of null pronouns, research in Optimality Theory (OT) is now at a point where we can begin to compare the predictions of competing accounts of the same phenomena. Grimshaw and Samek-Lodovici (1995), Samek-Lodovici (1996), Bresnan (forthcoming; chap. 5, this volume), and Speas (1995, 1997) have all pointed out unexplained properties of null pronouns that lead to the conclusion that syntax involves universal constraints that may be violated and whose ranking varies crosslinguistically. Each author takes a di¨erent approach and suggests di¨erent constraints. The ultimate goal of such work is a theory that can account for the distribution and interpretation of null pronouns with a minimum of stipulation, and that lends insight into the nature of the language faculty and its interaction with other cognitive faculties. In this chapter, I will lay out what I take to be the central facts to be accounted for in a theory of null pronouns and will propose an account in terms of ranked, violable constraints, which eliminates the stipulations necessary if constraints are inviolable. I will then brie¯y discuss how my account compares and contrasts with those of Samek-Lodovici and Bresnan. Since the three accounts focus on slightly di¨erent empirical domains, the coverage of data is overlapping but not identical. Certain problems having to do with the nature of control and of discourse prominence arise if we attempt to merge the three theories. I will suggest that resolving these problems requires that we investigate the role of point of view in the mapping from syntax to semantics and discourse. 13.1

The Licensing and Interpretation of Null Pronouns

Standard accounts of null pronouns distinguish those in subject position of in®nitives and gerunds (PRO) from those in other positions (pro). Many authors have pointed out problems inherent in maintaining this distinction (see especially Bouchard 1984; Sportiche 1983; Manzini 1983; Borer 1989; Huang 1984). In my account, I will adopt the view of Borer and Huang, that the apparent distinction is simply an artifact of the

394

Margaret Speas

interaction of several di¨erent principles. I will assume as they do that there is just one null pronoun, which I will write as Pro. I will argue that the crosslinguistic distribution and interpretation of Pro follows from the di¨erent rankings of the violable constraints. Before proceeding with an account, it is useful to review the facts about the distribution and interpretation of null pronouns. Although these may be well known, a simple review reveals the extent to which covert violability has been written into standard treatments. 13.1.1 Null Pronouns in English In English, null pronouns are possible only in the subject position of non®nite clauses. In general, overt NPs are excluded from this position. This is shown in (1). In ®nite subject position and in object position of verbs or prepositions, Pro is excluded and overt NPs must appear, as shown in (2). (1) a. Mary hopes Pro/*she/*her to see Bill. b. Mary hired Bill Pro/*her/*she to impress her competitors. c. Pro/*One to err is human. (2) a. Mary hopes she/*Pro will see Bill. b. Mary hired him/*Pro c. Mary talked to him/*Pro The complementarity of overt and null pronouns in English is not total, however. With certain verbs, such as expect and prefer, the subject of the non®nite complement can be either Pro or an overt pronoun, and with certain others, such as believe, the subject of the non®nite complement must be overt. (3) a. Mary expects Pro to have a good time. b. Mary expects her to have a good time. c. *Mary believes Pro to have good taste. d. Mary believes her to have good taste. Bouchard (1984:172±189) argued that Pro in English occurs in positions to which no Case is assigned but that have an independent theta role. I will adopt this view and will consider Pro to be a pronoun that lacks features altogether (aside from categorial identity). This means that it lacks Case as well as person-number-gender features and phonological features. I will show how this assumption leads to a uni®ed account of null arguments crosslinguistically, if we take constraints to be violable and ranked di¨erently in di¨erent languages. 13.1.2 On Null Case In the Minimalist Program, PRO is said to bear Null Case, as proposed by Martin (1992), Chomsky and Lasnik (1993), and BosÏkovic (1995, 1997).1 PRO can only

Constraints on Null Pronouns

395

appear in a position from which its features can be attracted to check Null Case, speci®cally the speci®er of a [‡tense, ®nite] in®nitive. Control predicates select such in®nitives, while ECM predicates select in®nitives that are [ tense, ®nite]. It is argued following Stowell (1982) that this di¨erence in tense features shows itself in the di¨ering temporal semantics of the two types of in®nitives. While it seems clear that there are intriguing correlations between the temporal properties of in®nitives and their ability to take a null subject, it is less clear that the attribution of these properties to a null Case is right. The two strongest arguments in favor of the proposal that PRO bears some sort of Case come from Chomsky and Lasnik (1993) (reprinted as chap. 1 of Chomsky 1995; see pp. 107±199), following Martin (1992). They say that PRO must have some Case because it cannot remain in a position from which Case has been removed, (4), and the positions in which PRO is licit cannot be launching sites for Case-driven movement of an overt NP (5). (4) a. *It was arrested PRO. b *my belief [PRO to be clever] (5) a. It is illegal [PRO to be noisy] b. *John is illegal [t to be noisy] Constructions like a (4a) can be ruled out in another way as well. If we assume following Huang (1984) that null pronouns are obligatorily coindexed with the closest nominal element, then PRO here is obligatorily coindexed with the pronoun in subject position, in violation of Binding Principle B. Example (4b) is supposed to show that since nouns can never assign Case, PRO cannot occur as the speci®er of a noun's complement. This is based on the assumption that the noun belief retains the selection properties of the verb, selecting a [ T F] complement, in which there is no assignor/attractor of null Case. However, the NPs in (6) are also ungrammatical, although they are headed by nouns whose verbal counterparts select an in®nitive in which null Case can be assigned. (6) a. *Mary's hope [PRO to be clever] b. *Mary's trial [PRO to win] c. *Mary's persuasion of John [PRO to tell the truth] This suggests that the ill-formedness of (4) cannot be straightforwardly attributed to the noun's inability to assign Case. Rather, it appears that belief simply does not select an in®nitive (like hope, trial, persuasion, but unlike promise, e¨ort, desire). Thus, (4b) does not provide an argument for the existence of null Case. The examples in (5) can be accounted for by theta theory. The external argument role in (5a) is assigned to the expletive, which in turn must be coindexed with the in®nitival clause. Alternative (5b) is ungrammatical because this coindexing, and hence theta assignment, fails.

396

Margaret Speas

The proposal that PRO bears a null Case introduces several conceptual problems as well. First, this is the only example of a Case that must be realized in a speci®c phonological form, which turns out to be identical in all languages. Clearly some generalization about the relationship between [‡tense, ®nite] and the possibility of a null subject is being missed. Second, it must be stipulated that PRO but not pro bears Null Case. If there can be two null nominals di¨erentiated only in that one bears null Case and the other can check any Case, it is not clear why languages do not have null nominals speci®ed for, say, Dative Case. This also ignores the convincing evidence presented by Borer (1989) and Huang (1984) that the distribution of the two null pronouns can be uni®ed. For these reasons, I will assume that it is not necessary to posit a null Case, and will consider Pro devoid of Case and agreement features. 13.1.3 Null Pronouns Crosslinguistically Although English restricts Pro to Caseless positions, many languages do not. In Subject position of ®nite clauses, Pro occurs in languages with either very rich agreement morphology, such as Spanish, or no agreement at all, such as Mandarin (Jaeggli and Sa®r 1989).2 (7) a. Pro vio esa pelõÂcula. saw:3S that movie `She/he saw that movie.' b. Zhangsan shuo [Pro] kanjianle Lisi Zhangsan says saw Lisi `Zhangsani says hei saw Lisi.'

Spanish

Mandarin

In many languages, Pro cannot occur in object position. (8) a. *Mary says John saw Pro. b. *Ana vio Pro3 Ana saw-3sg c. Zhangsani shuo Lisi kanjianle 0  i=j . Zhangsan says Lisi saw `Zhangsan says Lisi saw him/her other than Zhangsan or Lisi.'

English Spanish Mandarin

Huang (1984, 1991) argued that null objects in Mandarin are variables bound by a null operator, not null pronouns.4 He proposed the reason that the null object in sentences like (8c) cannot corefer with the matrix subject is that it is a variable and is thus subject to Binding Condition C. He proposed that null pronouns are ruled out in object position by the interaction between Binding Principle B and a Generalized Control Rule that forces all null pronouns to be controlled by the closest nominal element (an NP, or AGR). His theory predicts that null pronouns should be impossible in object position in all languages.

Constraints on Null Pronouns

397

This prediction appears incorrect, as shown in investigations of various languages that lack object agreement yet, by Huang's tests, allow null pronominal objects. The languages investigated include Brazilian Portuguese (Farrell 1990), Korean (Yoon 1985), and Thai (Hoonchamlong 1991; Pingkarawat 1989). A summary and discussion of these types of data can be found in Cole 1987. I use the following Thai examples (from Hoonchamlong 1991:238±240)5 to illustrate, and refer the reader to the above sources for complete argumentation. (9) a. Pro bOOk 2 waa3 [nit4 hen5 nOy2]. speak say [Nit see Noy] `He/she/they said that Nit saw Noy.' b. Nit bOOk2 wa3 [NOy2 hen5 Pro]. Nit speak say [Noy see ] `Nit said that Noy saw him/her/them.'

Thai

In summary, languages fall into three types with respect to the distribution of null pronouns; see (10). (10) The distribution of null pronouns Type 1: Pro ˆ subject of non®nite clause Type 2: Pro ˆ subject of non®nite6 clause and subject of ®nite clause Type 3: Pro ˆ subject of non®nite clause, subject of ®nite clause, and object Nonexistent: Pro ˆ object only Languages that allow Pro outside of non®nite clauses are those with either ``rich'' agreement, like Italian and Spanish, or with no agreement at all, like Mandarin and Thai. I consider these to be the central facts to be accounted for in a theory of null pronouns. My claim will be that ranked violable constraints can account for them in a straightforward way, and that this approach allows us to simplify the relevant constraints by removing ``covert violability'' written into the principles within a principlesand-parameters type approach. 13.1.4 Covert Violability and Control It is well known that Pro in English usually must be controlled within a certain domain. In (11), the matrix verbs determine which NP will control Pro; in (12), Pro is within an adjunct but must still be controlled by an NP in the sentence. For example, (12a) cannot mean that Mary resigned after people in general (Pro-arb) ®led a protest. (11) a. Maryi hopes Proi to have a good time. b. Maryi promised John [Proi to mow the lawn]. c. Maryi persuaded Johnj [Proj to mow the lawn].

398

Margaret Speas

(12) a. Maryi resigned [after Proi ®ling a protest]. b. Maryi called John [Proi to tell him about the test]. In languages with rich agreement, Pro gets its interpretive features from the agreement. As Huang (1984) and Borer (1989) have pointed out, this coindexing is parallel to the obligatory control coindexing just seen. However, the one case in which this requirement of coindexing does not hold is when there is no local antecedent for Pro. In (13a), the only possible controller does not c-command Pro; in (13b) there is no antecedent at all, and in (13c), the available antecedent is outside the relevant domain. In all these cases, Pro can receive either an arbitrary interpretation or a discourse antecedent. (13) a. Pro to behave better would enhance Bill's reputation. b. Pro to err is human. c. Pat wondered how Pro to make ¯owers out of Kleenex. In short, Pro must be locally controlled if and only if it can be. Within the GB literature, this fact was captured by writing the Control principle so that it would apply vacuously to cases where there was no local controller. For example, Huang (1989) gives the Generalized Control Rule spelled out in (14). (14) Generalized Control Rule (GCR) (Huang 1989:193) An empty pronominal is controlled in its control domain (if it has one) Control Domain: D is the Control Domain for p i¨ it is the minimal category that satis®es both (a) and (b): a. D is the lowest S or NP that contains (i) p, or (ii) the minimal maximal category containing p. b. D contains a SUBJECT accessible to p. In the sentences in (13), Pro has no Control Domain, so the GCR is satis®ed vacuously. Huang's if-clause writes violability into the rule. Moreover, the reason Pro has no control domain in these cases is that the minimal category containing it has no accessible Subject (read: potential antecedent). From an OT perspective, it seems obvious that the GCR is a violable constraint, and that candidates that violate it will be grammatical exactly when no competitors can obey it (and no higher constraints are violated). In the following section, I will show how taking an OT perspective on the GCR leads to a theory that simpli®es the GCR and gives a uni®ed account of null pronouns in both ®nite and non®nite contexts. 13.2

Optimality Theory as the Mapping from Syntax to Interface Levels

I assume the general OT architecture, in which the grammar includes a function GEN and a universal set of ranked constraints. I also assume that the constraints

Constraints on Null Pronouns

399

adhere to some sort of economy principles, which constitute the restrictions on possible constraints. A thorough exploration of the nature of these restrictions is beyond the scope of this chapter. However, I have formulated my constraints so that they contain no disjunction of conjunction, use only features or categories independently necessary in any framework, and add no new grammatical requirements over and above what is well established within the principles-and-parameters/Minimalist framework. I hope to show that a theory that assumes constraints are violable and ranked rather than inviolable and parameterized accounts for a broader range of data, while simplifying the constraints themselves and eliminating ad hoc properties such as feature ``strength.'' Inputs in the syntactic component di¨er from inputs in phonology. Determining what the input should be is more di½cult for syntax than for phonology. All structure building in phonology is done by Gen, so derivations can be dispensed with. If the inputs for syntax are unstructured strings of syntactic primitives, outputs will include sentences with di¨erent meanings, and it becomes di½cult to determine the limits of comparison.7 Therefore, most work in syntax within OT posits structured inputs. Grimshaw (1997) takes the input to be a verb and its argument structure and its extended projections, with assignment of lexical items to arguments. In a similar vein, Bresnan (forthcoming) proposes that the input is ``a (possibly underspeci®ed) feature structure representing some given morphosyntactic content independently of its forms of expression.'' If inputs already have structure, one must ask how that structure is generated and constrained. If a syntactic theory is supposed to describe and explain language structure, the OT model of syntax is incomplete without an explanation of the origin of the input structures. It may be possible to say that argument structure is generated by some nonlinguistic conceptual component, but this explanation is not su½cient for sentences more complex than a single verb with arguments that happen to be single lexical items. Consider for example the sentence in (15). (15) The author of this story signed the book that I bought. The verb signed has the argument structure in (16a), and in this example the argument positions map to the complex phrases with representations something like those in (16b) and (16c). (16) a. [‡V sign hx, yi] b. x ˆ [‡N, ‡de®nite author hR, zi 8 ; [z ˆ [‡N, ‡de®nite,‡proximate book]]] c. y ˆ [[‡N, ‡de®nite book ] [‡V buy hw, ui w ˆ we u ˆ book(rel)]] This minimal representation of the input to sentence (15) includes several levels of hierarchical structure that would generally be taken to be linguistic rather than just conceptual. Even if the structures are semantic/conceptual, we need a theory of possible structures, and I take this to be within the domain of a theory of syntax.

400

Margaret Speas

For these reasons, I take the theory I am elaborating here to be a theory of the mapping from syntactic structures to the semantic and phonological interface levels. I assume that syntactic structures are composed as in the Minimalist framework (Chomsky 1995), by means of the operation Merge, which combines lexical items, phrase markers, and copies of items that have already been so composed. Successive applications of Merge create phrase markers, in which features are checked, and these phrase markers are the input to Gen, as well as to the semantic component. In other words, I am suggesting that ranked violable constraints determine how syntactic constituents are mapped to their pronunciation and interpretation. Thus, I am following Grimshaw (1997) and Samek-Lodovici (1996) in the idea that the input to Gen is a structured phrase marker, but I take this to be a complete syntactic representation rather than a set of lexical entries. From the Minimalist Program, I adopt the idea that the language faculty includes minimal operations that construct syntactic representations from lexical items. These structures form the input to Gen, and then the candidate set generated by Gen is the input to pronunciation and interpretation. To be grammatical, a sentence must be the optimal candidate in its set. This does not mean that nonoptimal candidates cannot be pronounced or interpreted. The levels with which the syntax interacts have access to the entire candidate set, but only the optimal one is well formed. Thus, I am adopting the architecture of the Minimalist Program, but proposing that crosslinguistic variation is due not to strong versus weak features, but to alternate rankings of violable constraints. Instead of saying that a given feature is strong in a certain language, I would say that the requirements of some features are ranked more highly than the requirements of others. For example, the distinction between overt and covert wh-movement would result from con¯icting requirements that operators must be in a scope position and that arguments and modi®ers must be in positions that structurally realize thematic relations. This view allows us to constrain the relationship between candidate structures and their interpretations without stipulating that all candidates have the same Logical Form in a given competition. Gen yields a set of possible ways to pronounce a given structure, and it yields a structure that is then interpreted. Since the input is already structured and the same output is both pronounced and interpreted, all candidates will generally have the same interpretation, but there are limited ways candidates may vary in interpretation, as we will see. This proposal is not a mere terminological change, nor does it add apparatus not already necessary in the Minimalist Program. For one thing, as soon as we look at data other than movement versus nonmovement of wh-phrases and Case-marked DPs, it is not clear that feature strength can account for all language variation.9 Moreover, the Minimalist theory requires some theory of the ``pied piping of phonetic material.'' My claim is that if we work out that theory in terms of ranked violable constraints, the notion of feature strength becomes super¯uous.

Constraints on Null Pronouns

13.3

401

Optimal Null Arguments

13.3.1 The Functions Gen and Eval Gen takes a syntactic structure and generates a set of candidates for pronunciation and interpretation. Gen can add phonological material, but cannot in general add syntactic or semantic entities. The one exception to this is that Gen can add indices. This exception occurs because indices are not directly syntactic, have no phonetic content, and do not have a meaning in and of themselves. They serve the purpose of mapping from the syntax, in which pronouns are just pronouns, with no reference as yet assigned, and interpretation, in which the reference of pronouns is determined.10 I assume Gen can delete anything. Deletion is subject to faithfulness conditions whereby candidates with deletion are more marked that candidates without the deletion. I will not be examining candidates with deletion here. The function Eval evaluates each output candidate with respect to the ranked constraints. These candidates can then be pronounced and interpreted, although only the optimal candidate will be grammatically well formed. 13.3.2 Constraints on Pronouns I take Pro to be a minimal nominal, with no features (except its identity as a nominal). Thus, Pro has no phi, Case, or phonological features, but it is a syntactic entity that can be part of a numeration. Gen can produce candidates in which Pro remains featureless, and also candidates in which person/number and phonological features are added (i.e., it can insert an overt pronoun). Pronouns with features will incur a faithfulness violation. My analysis makes use of the six constraints shown in (17). (17) Control(CTL): A featureless pronoun must be coindexed with the closest c-commanding nominal element in its C-Domain. Binding Theory Principle B (BTB): A pronoun must be free in its Binding Domain. Max (PRO): If Pro occurs in the input, then its output correspondent is Pro. Agr(x): ‡Tense has AGR features. No Phi-Fts: Avoid AGR features Case: An NP with phi features must receive Case. (18) C-Domain: a is the C-Domain for b i¨ a. a dominates b and b. a dominates some set of checked feature(s) g and c. g c-commands b and d. no other set of checked features c-commands b

402

Margaret Speas

(19) a is the Binding Domain for b i¨ a is the least complete functional complex containing b, a governor for b, and a SUBJECT accessible to b. The constraints Agr(X) and No Phi-Fts are adapted from Samek-Lodovici 1996. Samek-Lodovici uses constraints with these names that have the same e¨ect as the constraints stated here. I have just changed the wording to avoid adding anything not already necessary in the Minimalist Program. I have adopted a standard but anachronistic de®nition of Binding Domain just to indicate that the Binding Domain is to be determined in the same way it would be determined in the Minimalist framework. The de®nition of C-Domain restricts Control so that it will be blocked by an intervening set of checked features. The relevant features include ‡WH, ‡AGR, and ‡T, and these will be checked in those cases where the Minimalist Program would consider them strong. In English the relevant checked features would be ‡WH and ‡AGR. In the tableaux below, I indicate the C-Domain with square brackets. Because these constraints are violable, they can be greatly simpli®ed. Control needs no clause about cases where there is no Domain. In such cases, grammaticality will be predicted by the fact that no indexing can satisfy the constraint, so candidates that violate it can be optimal. Principle B can now apply to Pro in all contexts, with no need to stipulate that the subject of non®nite clauses is also ‡anaphor. Max(PRO) is a standard faithfulness constraint.11 Agr(x) simply states a descriptive fact that must be stated in any framework that assumes that all languages have the same feature inventory. No Phi-Fts corresponds to a subcase of the Minimalist principle of Full Interpretation. Finally, Case applies to all and only NPs with phi features, and hence not to Pro, which has no such features. This avoids having to distinguish two types of null pronouns. There is nothing that says that null pronouns cannot be in Case positions, but null pronouns in Case positions are subject to the other constraints, which rule out null pronouns in languages in which those representations are not optimal. 13.3.3 Null Pronouns in English If Control and BTB both outrank Max(PRO), overt pronouns rather than Pro will appear in both subject and object position of ®nite clauses. This is shown below in tableau T13.1 for subjects and in tableau T13.2 for objects.12 In subject position of an embedded ®nite clause, overt pronouns violate Max(PRO) and obey Control and BTB, whether the pronoun is coindexed with a higher NP or not. Candidates like those in tableaux T13.1 and T13.2 with all NPs in Case positions will never violate Case. Furthermore, since Pro has no features, representations with Pro never violate Case. Only overt NPs in Caseless positions violate Case. If we assume that Case dominates the other constraints here, overt NPs will be ruled out in Caseless positions, as we see in tableau T13.3. Since Pro must be controlled, we also see now that Control must dominate Principle B in English.

Constraints on Null Pronouns

403

Tableau T13.1 Control, BTB g Max(Pro) (Finite Subject) Mary hopes Pro will see Bill

Control

a. Mary hopes [Pro will see Bill]

*!

b. Maryi hopes [Proi will see Bill]

*!

Max(Pro)

BTB

f c. Mary hopes [she will see Bill]

*

f d. Maryi hopes [shei will see Bill]

*

Tableau T13.2 Control, BTB g Max(PRO) (Object) Mary promoted Pro

Control

a. [Mary promoted Pro]

BTB

Max(PRO)

*!

b. [Maryi promoted Proi ]

*!

f c. [Mary promoted her]

*

d. [Maryi promoted heri ]

*!

*

Tableau T13.3 Non®nite clause Mary hopes Pro to see Bill

Case

a. [Mary hopes Pro to see Bill]

Control

BTB

Max(PRO)

*!

f b. [Maryi hopes Proi to see Bill]

*

c. [Mary hopes her to see Bill]

*!

d. [Maryi hopes heri to see Bill]

*!

* *

*

404

Margaret Speas

Tableau T13.3 illustrates the OT execution of Huang's (1984, 1989) observation that if the control domain and binding category coincide, the principles corresponding to Control and Btb con¯ict. Huang concluded that null pronouns must be excluded from such contexts, such as direct object position. However, in my analysis, Pro can emerge just in case some higher constraint (here, Case) excludes the otherwise preferred overt pronoun. This analysis solves a longstanding problem in e¨orts to treat the subject of non®nite clauses as a pronoun. When alternative Case assigners allow an overt pronoun to surface as subject of a non®nite clause, that pronoun must be disjoint from the higher subject, whereas the null pronoun must be bound (controlled). (20) a. Maryi hopes Proi=  j to win. b. Maryi hopes for her  i=j to win. Here the entire clause is the C-Domain, since the closest checked features to Pro are the features of the matrix INFL. The whole clause is also the Binding Domain for Pro, so coindexed representations violate BTB. When Case can be obeyed in a candidate with an overt NP, it will win over all candidates with Pro, because overt NPs vacuously obey Control. Since the indexed representation violates BTB, the representation without coindexing is the winner, as shown in tableau T13.4.13 This contrasts with the candidate set in (22), where Pro won out over all overt NPs, but coindexed Pro won out over unindexed Pro because Control is ranked higher than BTB. When a verb optionally Case-marks the spec of its complement, as I assume is the case with expect, the embedded subject must be controlled if it is null and disjoint if it is overt. (21) a. Maryi expects Proi to have a good time. b. Maryi expects herj to have a good time. Since expect is lexically speci®ed to assign Case optionally, there are two di¨erent inputs: one with a Case-assigning verb, as shown in tableau T13.5, and one with a Tableau T13.4 Mary hopes for Pro to see Bill a. [Mary hopes for Pro to see Bill] b. [Maryi hopes for Proi to see Bill]

Case

Control

BTB

*! *!

f c. [Mary hopes for her to see Bill] d. [Maryi hopes for heri to see Bill]

Max(PRO)

* *!

*

Constraints on Null Pronouns

405

Tableau T13.5 Caseless expect in input Mary expects[ Case] Pro to see Bill

Case

a. [Mary expects Pro to see Bill]

Control

BTB

Max(PRO)

*!

f b. [Maryi expects Proi to see Bill]

*

c. [Mary expects her to see Bill]

*!

d. [Maryi expects heri to see Bill]

*!

* *

*

BTB

Max(PRO)

Tableau T13.6 Case-marking expect in input Mary expects[‡Case] Pro to see Bill a. [Mary expects Pro to see Bill] b. [Maryi expects Proi to see Bill]

Case

Control *!

*!

f c. [Mary expects her to see Bill] d. [Maryi expects heri to see Bill]

* *!

*

verb that does not assign Case, as shown in tableau T13.6. As we saw above, when the verb does not assign Case the candidates with overt embedded subjects are ruled out in favor of the null subject, and the controlled Pro wins the competition between Control and Btb. When Case-marking expect is in the input, the candidate with the overt embedded subject is optimal, since overt pronouns satisfy Control vacuously. Notice that the distribution of Pro follows here without any stipulation that Pro cannot have Case. Pro can have Case in principle, but Case-marked Pro will only be optimal if candidates with overt NPs are ruled out independently. We will see how this works in the analysis of Spanish. Finally, when there is no possible local antecedent for Pro, candidates with Pro will violate Control whether they are indexed with some nonlocal NP in the sentence, as in tableaux T13.7 and T13.8, or are unindexed, as in tableau T13.9 (p. 406). Indexed and unindexed candidates will tie and both be grammatical, since candidates with overt pronouns violate the higher-ranked Case. Coindexed candidates will corefer with the NP sharing their index; unindexed candidates can get an antecedent from discourse or can be arbitrary.

406

Margaret Speas

Tableau T13.7 The only antecedent fails to c-command Pro to behave badly will damage Bill's reputation

Case

Control

f a. [Pro to behave . . .] Bill's]

*

f b. [Proi to behave . . .] Billi 's]

*

BTB

Max(PRO)

c. [Him to behave . . .] Bill's

*!

*

d. [Himi to behave . . .] Billi 's]

*!

*

Tableau T13.8 No antecedent within control domain Pat wonders how Pro to make ¯owers . . .

Case

Control

f a. Pat wonders [how Pro to make ¯owers . . .]

*

f b. Pati wonders [how Proi to make ¯owers . . .]

*

c. Pat wonders [how her to make ¯owers . . .]

Btb

Max(PRO)

*!

*

Tableau T13.9 No antecedent at all It's fun Pro to dance

Case

f a. [It's fun Pro to dance] b. [It's fun her to dance]

Control

BTB

Max(PRO)

* *!

*

Constraints on Null Pronouns

407

In summary, the interaction of the constraints Control, Binding Principle B, Case, and Max(PRO) can account for the distribution of null subjects as well as objects, and subjects of non®nite as well as ®nite clauses. It does so with no ad hoc licensing condition speci®c to null pronouns, and without positing a dubious entity such as ``null Case.'' In this system, no syntactic features refer to the phonological properties of a pronoun, nor do they refer to discourse properties. In the following sections, we will see how a di¨erent ranking yields the facts of a language like Thai, which allows null subjects and objects. I will show how the OT approach yields a simple account of variation in the status of embedded null subjects in Thai and Mandarin. We will also see how the conditions on agreement capture the di¨erences between Spanish-type languages and English, without recourse to a special licensing condition. 13.3.4 Null Pronouns in Thai As discussed above, the absence of object Pro in many languages led Huang to propose that Control and Binding principles interact in a way that makes object Pro universally impossible. However, we apparently do get object Pro in languages like Thai. It might be possible to stipulate, as does Rizzi (1986), that verbs can license object Pro in some languages. But that sort of account loses the uni®cation of PRO and pro captured in Huang's theory. Cole (1987) retains Huang's Generalized Control Rule, but adds a parameter whereby it fails to apply to Case-marked Pro in some languages. Such a parameter greatly undermines the appeal of the GCR account by making the Control rule Generalized in only some languages. Cole's parameter amounts to a hedge added to an otherwise general account. In OT, we can turn Cole's account into an explanatory one, because in OT the fact that a certain constraint fails to be satis®ed in certain languages is just what we expect. In particular, Control is violable, and sentences with Control violations are grammatical in just those languages in which other constraints outrank Control. In Thai, those constraints are Btb and Max(PRO). Tableaux T13.10 and T13.11 express Cole's insight that Control can be violated in some languages. Below I show inputs containing Pro. Thai also has overt pronouns, and these will arise from inputs with overt pronouns, where there will be no Max(PRO) violation. This same constraint ranking also allows Pro in subject position, as shown in tableau T13.12.14 See tableau T13.13 as well. In an embedded clause in Thai, Pro is also allowed and may be either coindexed with the higher subject or not, as shown in (22). Also see tableau T13.14 (p. 409). In such cases, the tensed embedded clause is the C-Domain as well as the Binding Domain, and there is no c-commanding antecedent in that domain. (22) Nit4 bOOk2 waa3 [Pro hen5 NOy2]. Nit speak say [ see Noy] `Nit said that he/she/they saw Noy.'

408

Margaret Speas

Tableau T13.10 Thai input with Pro object . . . Nit4 hen5 Pro N saw

Btb

Max(PRO)

Control

f a. . . . Nit4 hen5 Pro

*

b. . . . Nit4i hen5 Proi

*!

c. . . . Nit4 hen5 khaw4 `. . . Nit saw5 him/her'

*!

Tableau T13.11 Thai input with overt pn. object . . . Nit4 hen5 khaw4 `N saw him/her

Btb

Max(PRO)

Control

a. . . . Nit4 hen5 Pro

*!

b. . . . Nit4i hen5 Proi

*!

f c. . . . Nit4 hen5 khaw4 `. . . Nit saw him/her' Tableau T13.12 Thai input with Pro subject Pro hen5 Nit4

Btb

Max(PRO)

f a. Pro hen5 Nit4

Control *

b. khaw4 hen5 Nit4 `s/he saw Nit'

*!

Tableau T13.13 Thai input with overt subject khaw4 hen5 Nit4 a. Pro hen5 Nit4 f b. khaw4 hen5 Nit4 `s/he saw Nit

Btb

Max(PRO)

Control *!

Constraints on Null Pronouns

409

Tableau T13.14 Thai embedded subject Nit4 bOOk2 waa3 Pro hen5 NOy2

Btb

Max(PRO)

Control

f a. Nit4 bOOk2 waa3 [Pro hen5 NOy2]

*

f b. Niti bOOk waa3 [Proi hen5 NOy]

*

c. . . . [khaw4 hen5 NOy] `. . . [He saw Noy]

*!

Tableau T13.15 Thai embedded in®nitive [Nuan1 yaak2 Pro kin1 ?aahaan5 ciin1]

Btb

Max(PRO)

f a. [Nuan1 yaak2 Pro kin1 ?aahaan5 ciin1] b. [Nuan1i yaak2 Proi kin1 ?aahaan5 ciin1] c. [Nuan1 yaak2 khaw4 kin1 ?aahaan5 ciin1] `. . . He eat Ch. food'

Control *

*! *!

This ranking predicts that in embedded in®nitives, where the C-Domain and the Binding Domain overlap, the optimal candidate should be one with Pro that is not coindexed with the subject. (See tableau T13.15.) Finite complements in Thai are characterized by the presence of the complementizer ``waa3'' and the possibility of having auxiliaries. At ®rst glance, this prediction seems to be wrong. In (23), the embedded clause is non®nite, and the null subject must be coreferent with the matrix subject. (23) [Nuan1 yaak2 Pro kin1 ?aahaan5 ciin1] Nuan want eat food Chinese `Nuani wants Proi=  j to eat Chinese food.' However, Pingkarawat (1989) argues that this apparent control is actually determined by lexical semantics rather than syntactic coindexing.15 She points out that some verbs take ®nite complements that require coreference of the embedded and matrix subjects; See (24). (24) Nuanli khit4 waa3 [Proi=  j =Phim1 tON3 pay1 rooN1rian1 wan1nii4] Phim must go school tomorrow Nuan think say `Nuani thought that shei /Phim had to go to school today.' (cited in Hoonchamlong 1991:259)

410

Margaret Speas

Pingkarawat divides verbs into four classes according to their control properties, and argues that these properties do not correlate with the ®nite-non®nite distinction. If she is right, the OT approach presented here makes the right prediction. I have claimed that unindexed Pro need not necessarily be disjoint in reference from a higher NP. Unindexed Pro simply does not have its coreference determined by syntactic principles. In an in®nitive in Thai, the optimal candidate contains unindexed Pro, and the reference of this Pro is determined by lexical semantic properties. 13.3.5 Finite Subject Pro in Languages That Do Not Allow Object Pro We have seem that null objects are disallowed when Control and BTB are higher than Max(PRO), as in English. But this ranking also rules out null subjects, which do not have an antecedent in their C-Domain (the embedded ®nite clause). If there were some alternative way that Control could be satis®ed in a language, we would expect null subjects to emerge in this position. I will develop an account of null subjects in ®nite clauses that is based on the claim of Huang (1984) that rich agreement can count as a controller. Another way Control could be satis®ed by an embedded subject in a ®nite clause is if the C-Domain were the higher clause. We expect this to be the case in languages in which the features that de®ne the C-Domain are not checked features. Thus, when Control and BTB outrank Max(PRO), null objects will be ruled out (unless there is rich object agreement), and null subjects will be allowed just in case Control can be satis®ed. Spanish satis®es Control due to rich agreement, and Mandarin satis®es Control due to the features that delimit Domains. Following Samek-Lodovici (1996), I adopt the view that the presence of agreement for some feature X is due to the constraint Agr(X) being ranked higher than No PhiFeatures. In Spanish, Agr(person) is higher than No Phi-Features, so person agreement is present. In English, the ranking is reversed, and so as proposed by Kayne (1989), English agreement does not include person. As shown in tableaux T13.16 and T13.17, these rankings yield null subjects for Spanish but not for English. As in cases discussed above, the grammatical sentence with the overt pronoun in Spanish arises from an input with an overt pronoun. Like Spanish, Mandarin allows Pro subjects and not Pro objects. However, Mandarin has no agreement morphology at all. Moreover, according to Huang (1984, 1989), Pro subjects are only possible in embedded clauses and must be coreferent with the higher subject. (25) a.

Zhangsan shuo Pro bu renshi Lisi. Zhangsan say not know Lisi `Zhangsani says that hei does not know Lisi.'

Constraints on Null Pronouns

411

Tableau T13.16 Subject of ®nite clause in Spanish Pro vio esa pelõÂcula

Control

Max(PRO)

Agr(prs)

a. Pro vi esa pelõÂcula

*!

*

b. Proi vii esa pelõÂcula

*!

*

c. Pro vio[‡prs] esa pelõÂcula

*!

NoFts

*

f d. Proi vio[‡prs]i esa pelõÂcula

*

e. Ella vio esa pelõÂcula

*!

*

Tableau T13.17 Subject of ®nite clause in English Pro saw that ®lm

Control

NoFts

Max(PRO)

Agr(prs)

a. Pro saw that ®lm

*!

*

b. Proi sawi that ®lm

*!

*

c. Pro saw[‡prs] that ®lm

*!

*

d. Proi sawi [‡prs] that ®lm

*!

*

f e. She saw that ®lm

*

*

b. *Pro bu renshi Lisi.16 (cf. Ta bu renshi Lisi) `He/she does not know Lisi.' c. *Zhangsan bu renshi Pro. (cf. Zhangsan bu renshi ta) `Zhangsan does not know him/her.' Because Pro subjects can (and must) be controlled by a subject in the next higher clause, it must be the case that embedded subjects but not matrix subjects can have a candidate that obeys Control. This will be the result if the matrix S rather than the embedded S is the C-Domain. It is well known that Mandarin can have long-distance binding of re¯exives. Cole, Hermon, and Sung (1990) and Cole and Sung (1994) propose that INFL in Mandarin moves up into a higher clause, extending the Binding Domain for a re¯exive. The Binding Domain for pronouns is not extended in this way; pronouns must still be free only in the embedded clause. If it is correct that

412

Margaret Speas

Tableau T13.18 Mandarin embedded subject Zhangsan shuo Pro bu renshi Lisi

Control

a. [Zhangsan shuo {Pro bu renshi Lisi}]

Btb

Max(PRO)

*!

f b. [Zhangsani shuo {Proi bu renshi Lisi}] c. [Zhangsan shuo {ta bu renshi Lisi}]

*!

Tableau T13.19 Mandarin matrix subject cannot be Pro Pro bu renshi Lisi

Control

a. [{Pro bu renshi Lisi}]

*!

b. [{Proi bu renshi Lisii }]

*!

f c. [{Ta bu renshi Lisi}]

Btb

Max(PRO)

*

INFL moves, the features that would delimit a C-Domain are in the higher clause, and so the C-Domain for an embedded sentence is extended as well. If the C-Domain in Mandarin is the higher clause, while the Binding Domain for pronouns (indicated with curly brackets) is the lower clause, we get the right results. (See tableaux T13.18 to T13.20.) Thus, Pro is possible only as the subject of an embedded clause in Mandarin. Mandarin di¨ers from Thai, in which Pro can appear in any position. The di¨erence follows from a combination of a di¨erent ranking and a di¨erent position for the features that delimit the C-Domain.17 In this way we maintain the view of Cole, Hermon, and Sung (1990) that a theory in which the de®nition of a Domain is the same for all languages is more restrictive than one that allows parameterization of Domains. The di¨erences follow from the interaction of the de®nition of the Domains with other principles. Further, these authors claim that the type of head movement observed in Mandarin only takes place in languages that lack agreement (Cole and Sung 1994:364). If this is right, we have an explanation for the tendency of null subjects to occur in languages either with rich agreement or no agreement. Null subjects can occur in languages that either have the Thai ranking,18 or can satisfy Control due to having either rich agreement or no agreement.

Constraints on Null Pronouns

413

Tableau T13.20 Mandarin Object cannot be Pro Zhangsan bu renshi Pro a. {{[Zhangsan bu renshi Pro}] b. [{Zhangsani bu renshi Proi }] f c. [{Zhangsan bu renshi ta}] 13.4

Control

BTB

Max(PRO)

*! *! *

Predictions and Typology

We have established that there are three types of languages with respect to the occurrence of subject and object Pro in ®nite clauses: (1) those like Thai and Korean, which allow both subject and object Pro; (2) those like English, which disallow both subject and object Pro; (3) those like Spanish and Mandarin, which allow subject Pro but not object Pro. These three types are accounted for by the rankings we have proposed for the Control, BTB, and Max(PRO) constraints. If BTB and Max(PRO) outrank Control, Pro is permitted as both subject and object of ®nite clauses. If Control and Btb outrank Max(PRO), Pro cannot occur in these positions unless the language has either rich agreement or no agreement at all, in which case Pro is licensed as subject of at least some ®nite clauses. However, none of the rankings we have considered licenses object Pro but disallows subject Pro in ®nite clauses. In fact, we predict that such a language is impossible. If BTB and Max(PRO) outrank Control, null subjects will be possible in both positions. If Control and Btb outrank Max(PRO), object Pro will be ruled out by the interaction of Control and Btb. Objects do not have the option of satisfying control while obeying Btb, unless there is rich object agreement. The situation is summarized in table 13.1. However, table 13.1 does not show any language type in which Btb is dominated by Control and Max(PRO). If we consider only the constraints discussed so far, such a ranking would give rise to a language type in which object Pro is licensed but must be controlled by the subject, as in tableau T13.21. Moreover subject Pro is licensed in such a language only if Max(PRO) outranks Control. Ranking Control g Max(PRO) g Btb licenses coindexed object Pro but disallows subject Pro in ®nite clauses. Such a language could have overt pronouns resulting from inputs with overt pronouns, as shown in tableau T13.22. The only possibility for a language with null objects but overt subjects would be where the null object obligatorily coreferred with the subject. But surely this situation will be ruled out by the independently needed constraints on re¯exive predicates. Presumably the relevant constraint is universally higher than Btb. For example,

414

Margaret Speas

Table 13.1 Distribution of Pro in ®nite clauses Language

Subject Pro

Object Pro

Account

Thai, Korean

yes

yes

Btb, Max(Pro) g Control

English

no

no

Control, Btb g Max( Pro)

Spanish, Mandarin

yes

no

Control, Btb g Max( Pro); subject Pro is controlled by Agreement or by higher subject

***

no

yes

Impossible, because object Pro can't satisfy Control without violating Btb

Tableau T13.21 If Control g Max(PRO) g Btb Mary promoted Pro

Control

a. [Mary promoted Pro]

Max(Pro)

Btb

*!

f b. [Maryi promoted Proi ]

*

c. [Mary promoted her]

*!

d. [Maryi promoted heri ]

*!

*

Tableau T13.22 If Control g Btb g Max(PRO) Pro promoted Bill a. [Pro promoted Bill] f b. [She promoted Bill]

Control

Max(PRO)

*! *

Btb

Constraints on Null Pronouns

415

Burzio (1998) proposes a treatment of coreference that uses a universally ranked family of Binding Constraints. If we add such a constraint, the system of constraints I have proposed will predict that there are no languages with object Pro but without subject Pro. 13.5

Unifying the Accounts: The Role of Point of View

The focus of this chapter has been on trying to unify all occurrences of null pronouns under one general theory. Thus, I have shown how simpli®ed versions of familiar principles can yield the range of language types, if we assume that these principles are OT-type constraints that can be violated. The treatments of null pronouns by Samek-Lodovici (1996) and Bresnan (chap. 5, this volume) have a slightly di¨erent focus and hence capture a slightly di¨erent range of facts. They account only for null pronouns in subject position and do not deal with PRO or with null objects. On the other hand, they give an account of cases where null pronouns are not only licensed but obligatory. Further, Samek-Lodovici accounts for the interaction of Subject null pronouns with NPs in postverbal position, and Bresnan accounts for the range of pronoun types (null, bound, clitic, weak, full pronoun). Since the range of data covered by the various theories is di¨erent, we would expect that they could be combined to cover the entire range. For example, I could add a constraint ``Reduce Topic'' to my account. If this constraint were higher than the constraints that rule out null pronouns (Control and Agr(X)), topics could be null despite violating one of the other constraints. Similarly, Bresnan and/or SamekLodovici could add constraints ``Drop Caseless Pn'' and/or ``Drop Object Pn,'' creating a family of ``Drop'' constraints. However, much of the data covered by their constraints on the pronunciation of topics is already covered in my theory by the interaction of Control, Btb, and Max(Pro). The attempt to decide which way the merger should go requires that we address questions having to do with the nature of topic and the dropped pronouns, the nature of Control, and the ways syntax should be allowed to interact with phonology and semantics. To take the last and most general issue ®rst, I consider it a virtue of my account that the syntactic constraints used to account for the distribution of null pronouns make no speci®c reference either to phonological properties or to semantic or discourse properties. I take Pro to be a purely syntactic object, composed only of category features and no others. The fact that it is featureless means that it can violate Control, but none of the constraints mention its pronunciation or its meaning or discourse function. Eval employs these constraints to determine which candidates will be syntactically well formed, and these candidates are then mapped to pronunciation

416

Margaret Speas

and interpretation. In this way, my syntactic system is autonomous in the same way that OT systems of phonology are: those do not make reference to syntax or semantics. They take phonological inputs and yield the optimal pronunciation, which is mapped to morphology, syntax, and semantics. If there is indeed a syntactic component with OT properties, the most restrictive theory will be one in which the constraints of that component are purely syntactic. An alternative view would be that no OT constraints are purely syntactic. Perhaps the only purely syntactic principles are those minimal principles that create phrase structures, with all other constraints specifying the mapping between these structures and other levels. This is in essence the position taken by Bresnan (forthcoming; chap. 5, this volume). Presumably the choice between these two options will become clearer as work in OT syntax progresses. Either of these two options gives us a more restrictive theory than an approach that mixes purely syntactic conditions with interface conditions. Turning now to the questions of the nature of topic and control, there are vexing problems in our e¨orts to understand these clearly, but I would like to suggest that there may be a common direction toward answers to both questions. The accounts of both Samek-Lodovici and Bresnan rely on a correlation between reduced (or null) pronouns and topic anaphoricity. A pronoun that refers to the topic will be reduced unless the interaction of other constraints causes the overt (strong) form to emerge as optimal. This makes it di½cult for them to use the same principles to account for null objects and for subjects of non®nite clauses. Subjects of non®nite clauses can clearly occur in sentences where some other NP must be topic, as shown in (26). It may be the case that subjects of in®nitives are null for some independent reason, but intuitively, both types of subject are null because their content can be recovered, so it is worth looking at a theory in which no speci®c Topic feature needs to be mentioned in the constraints. Since standard de®nitions of topic imply that a sentence has only one topic, languages like Thai where both subjects and objects can be null are unexpected. Of course the standard assumption could be wrong, but if it is, we lack any clear de®nition of what class of pronouns it is that must be null in order to obey the relevant constraint. Although it seems intuitively clear that some type of discourse prominence is involved, a constraint will be vacuous unless it has a de®nition of the relevant type of prominence. (26) a. A: B: b. A: B:

What happened to that sofa? John tried Pro to move it. Did John persuade anyone Pro to contribute to the campaign? No, but he[+reduced] persuaded Mary Pro to vote for him.

Constraints on Null Pronouns

(27) A: Khun1 chOOp3 duu1 naN5 may4? you like see movie Q `Do you like to watch movies?' B: chOOp3. like

417

Thai

(Hoonchamlong 1991:104)

Moreover, it is possible to ®nd examples of Case-marked Pro in contexts that Samek-Lodovici (1996) de®nes as nontopic contexts. He claims that a null subject is ruled out in (28b) (his (3b) on p. 31) because the potential antecedent, Gianni, is the agent of a passive, which is a nontopic. (28) a. Questa mattina, la mostra e stata visita da Giannii this morning the exhibition was visited by Gianni b. Piu tardi, *proi / eglii /luii ha visitato l'universita more late (he) visited the university However, if we change the predicate in the second sentence to one that has a closer semantic connection to the ®rst sentence, the null pronoun becomes possible.19 (29) a. Questa mattina, la mostra e stata vista da Giannii this morning the exhibition was visited by Gianni b. Piu' tardi, proi /luii non si ricordava gia' piu' che cosa aveva visto. Later, pro not remember what things has seen `Later, he could not remember what things he saw.' (30) a. Questa mattina, la mostra e stata visita da Giannii this morning the exhibition was visited by Gianni b. Prima di allora, proi /luii non aveva mai pensato/immaginato che l'arte before that pro/he not had thought/imagined that art fosse cosi interessante. could-be so interesting. `Before that, he had never thought/imagined that art could be so interesting.' Similarly, in Spanish we ®nd that the NP in a by-phrase can be the antecedent of pro when the event expressed by the predicate is closely related to the event expressed in the context sentence. In both (31b) and (31b), pro may easily be coreferent with the NP in the by-phrase.20 (31) a. Esta manyana, nuestro restaurante fue visitado por el inspectori de salud this morning our restaurant was visited by the health inspector b. Durante dos horas, proi examino el sitio tratando de encontrar for two hours, pro examined the place trying to ®nd cucarachas cockroaches

418

Margaret Speas

(32) a. Esta manyana, nuestro restaurante fue visitado por el inspectori de salud this morning our restaurant was visited by the health inspector b. Manyana, proi podria aparecer por el vuestro. tomorrow pro might visit at yours Suppose we say that the NP in a by-phrase tends not to be a topic, but may be elevated to topic status under certain conditions. Then it seems that we have no independent test for topichood, and so we are left unsure of what sort of discourse prominence is needed to satisfy the relevant constraint. Notice that in (32b) it must be understood that el vuestro `yours' means `your restaurant', suggesting that the restaurant is still the topic. As noted, if we choose to give up the claim that a sentence has only one topic, we are in need of a new de®nition of the relevant type of discourse prominence. The system I have proposed captures the distribution of null pronouns without making mention of the notion ``Topic.'' Because of this restriction, it does not as it stands account for the fact that certain discourse-prominent pronouns are obligatorily null. My analysis relies instead on control, which I have characterized as a general requirement that featureless categories be bound. A central problem lurking within this account has to do with whether control really is syntactic at all. It has often been argued that control is actually a semantic phenomenon. (See in particular Bach 1979; Chierchia 1984; Ladusaw and Dowty 1988.) For example, it is well known that implicit arguments may control, and that the choice of controller seems in some cases to best be described in terms of the semantics of the predicate involved. Purpose clauses always seem, for instance, to be controlled by the bene®ciary of the action, as (33) suggests. (33) a. Max bought a dog for Mary Pro to play with. Pro ˆ Mary b. Max bought a dog Pro to play with Pro ˆ Max c. Max bought glue Pro to use for this project Pro ˆ whoever does the project Further, Ladusaw and Dowty (1988) argue convincingly for a semantic solution to the problem known as ``Bach's paradox.'' In (34), the controller of ``ask'' seems to change when the embedded predicate is passivized. Ladusaw and Dowty point out that the choice of controller depends on the sense of ask: if it means simply ``request,'' this entails that the NP denoted by the object will most plausibly be the one responsible for carrying out the action. If it means ``request permission,'' the subject will most plausibly be the one responsible for the action. (34) Bach's paradox a. John asked Pro to leave. Pro ˆ subject b. John asked Mary Pro to leave. Pro ˆ object c. John asked his mother Pro to be allowed to leave.

Pro ˆ subject

Constraints on Null Pronouns

419

It has also been frequently pointed out that control predicates and their properties tend to be the same from language to language. This is unexpected if control properties are simply stipulated as some kind of syntactic feature (but see Larson 1991 for an attempt to ground these properties in syntactic di¨erences). Thus, although I have tried to restrict my account to purely syntactic constraints, it is not at all clear that control should really be treated as purely syntactic. Yet the distribution of Pro is correlated with Case, and the interpretation of Pro seems to require syntactic binding in many cases, which suggests that the role of syntax cannot be eliminated. I would like to suggest a path of research that may lead to answers to the questions raised above about both discourse prominence and control. Speci®cally, I will show a series of ways in which point of view, as de®ned by Cantrall (1974), Mitchell (1986), Kuno (1987), and Tenny (1998), plays a role in the licensing and interpretation of null pronouns in both ®nite and non®nite contexts. First of all, Mitchell (1986) notes that Morgan (1970) showed that controlled PRO (in Morgan's terminology, the output of Equi NP Deletion), unlike overt pronouns, must be a quasi-indicator. A quasi-indicator, following CastanÄeda (1968), is a thirdperson pronoun that can stand in for a person's internal ``I.'' Morgan uses the example of an amnesiac baseball player to illustrate. (35) Suppose the baseball player Ernie Banks gets beaned, develops amnesia, and is taken to the hospital . . . he reads in the newspapers about a baseball player named Ernie Banks. He decides that he likes Ernie Banks, and would like him to leave Chicago and go play for the Mets. Under this scenario, one could say (36a) but not (36b). The null embedded subject in (36b) is a quasi-indicator, and as such can be said to be ®rst person, from the subject's point of view. The overt pronoun in (37), on the other hand, is ambiguous. (36) a. Ernie Banks wants Ernie Banks to leave Chicago. b. Ernie Banks wants Pro to leave Chicago. (37) Ernie Banks hopes he will leave Chicago. Perhaps it is the quasi-indicator status of the null subject of control complements that leads us to think that syntactic binding is involved. Control complements do seem to involve the Subject's point of view. For example, Mitchell shows that seem ‡ AP constructions normally must be from the speaker's point of view, and therefore the AP must be one that is subjectively evaluated by the speaker; see (38).

420

(38) John seems

Margaret Speas

tired friendly intelligent ?tall ?married ?unemployed

However, when we embed a seem ‡AP construction under a control predicate, this restriction is removed, as (39) suggests. (39) John wants to seem tired/tall/married/unemployed Notice that the restriction is not lifted if the embedding is under a noncontrol predicate; see (40). (40) John thinks that he seems tired/friendly/ ?tall/ ?unemployed These facts would follow if the control predicate introduced the subject's point of view in place of the speaker's. Another indication that point of view is crucial in interpreting Pro is that when Pro is not controlled, point of view is at least partly responsible for determining which antecedent will be chosen from discourse. This is shown by the fact that in order to get an arbitrary reading, the sentence must be from what Tenny (1998) calls ``public point of view.'' Suppose a mother is reminding her children to behave. She can say (a) but not (b). In (b), the point of view is that of the speaker rather than the public point of view; therefore Pro arb is ill formed, as shown in (41). (41) a. Pro to behave oneself would be a good idea. b. *Pro to behave oneself would help me get this job. Turning to Pro in ®nite clauses, Kitagawa (1998) argues that Pro in certain Japanese narratives serves to bring the reader in to the subject's point of view. These narratives contain what Kitagawa calls introductory pro: a null pronoun used near the beginning of a narrative, before the referent has been introduced. Introductory Pro is not necessarily the topic of the narrative. For example, the sentences in (42) are the initial sentences of one such narrative. The topic of the narrative is a woman who is very ill. The null pronoun in (42b) refers to her husband. (42) a. Itsu-made inochi-ga motsu noka when-til life-NOM hold Q `How long would (her) life hold?' b. Sore-wa hProi kowakute kikenakatta that-top fearing could-not-ask `Pro was too afraid to ask.' Kitagawa shows that introductory pro occurs only in the context of certain emotive words, which are obligatorily evaluated from the point of view of the referent to the

Constraints on Null Pronouns

421

null pronoun. Introductory pro is used when the writer wishes the reader to take the point of view of the referent. Kitagawa further suggests that the subject must be null because null pronouns are the most underspeci®ed in person, number, and gender features, and as such are the most conducive to empathetic identi®cation. Another way of putting this is if the person features are overt in Japanese, these must be evaluated from the point of view of the speaker, and this inhibits taking the subject's point of view in interpreting the sentence. Thus, in at least one type of narrative we see the importance of point of view in assigning a referent to pro subjects of ®nite clauses. In Mitchell's terms, we can describe Kitagawa's introductory pro as a type of quasi-indicator standing in for the reader's ``I'' in an abstract sense. Thus, point of view seems important for null pronouns in both ®nite and non®nite clauses. I suggest that further investigation of the relationship between person and point of view can be a fruitful route to resolving the twin problems of the nature of control and the typology of discourse prominence. 13.6

Conclusion

I have presented an account of null pronouns that attempts to capture their distribution and crosslinguistic variation in all syntactic contexts. I have tried to restrict the constraints in my account to those that are purely syntactic, and that add nothing to the grammar that is not already necessary in the Minimalist Program. The resulting system, which relies on the interaction of constraints on control, binding, and Case along with faithfulness constraints, allows for a uni®cation of the di¨erent types of null pronouns and for a simpli®cation of each of the relevant principles. The data accounted for by my system and those of Bresnan (forthcoming; chap. 5, this volume) and Samek-Lodovici (1996) are overlapping but not identical. I have pointed out that any attempt to unify the theories must deal with certain questions about the nature of topic and control. Finally, I presented some evidence that these questions may ®nd common answers in an investigation of the role of point of view in the licensing and interpretation of null pronouns in both in®nitives and ®nite clauses. Notes For many helpful comments on this material, I am grateful to Diana Archangeli, Terry Langendoen, Sten Vikner, Chisato Kitagawa, Lyn Frazier, Eva Juarros, Elisabeth Villalta, Maria Nella Carminati, Robin Schafer, Ellen Woolford, Jane Grimshaw, Vieri Samek-Lodovici, and the participants in the HOT workshop. 1. The idea that PRO has some sort of Case was suggested by Chomsky (1986:104) and SigurDsson (1989:183). 2. See Speas 1995 for a discussion of apparent counterexamples to this generalization).

422

Margaret Speas

3. I follow Campos (1986) and Authier (1989) in assuming that arbitrary null objects in Spanish and Italian involve quanti®cation and are not pronominal. 4. Huang's analysis of null objects as traces (see note 16) is disputed by Langendoen and Xu (1985), who essentially analyze all types of null elements in Mandarin as instances of pro, as in the analyses of Thai and Korean presented here. 5. I use the following orthography for Thai examples: O ˆ low back V, 1 ˆ mid tone, 2 ˆ low tone, 3 ˆ falling tone, 4 ˆ high tone, 5 ˆ rising tone. 6. Rizzi (1986) made clear that the crucial feature is AGR rather than tense, but to keep the description simple, I will set this distinction aside. 7. I am grateful to an anonymous reviewer for drawing my attention to empirical problems with an earlier attempt to allow the input to be an unstructured set of words. For accounts in which candidates with di¨erent meanings are compared, see Legendre, Smolensky, and Wilson 1998; Keer and Bakovic 1997; Bakovic and Keer, chap. 4, this volume. 8. I assume following Williams (1987) that nouns have a referential argument position, but nothing hinges on this assumption. 9. Examples include scrambling, licensing of null pronouns, licensing of null complementizers, Binding theory, multiple wh-movement, and ellipsis. These are all domains in which it is either unclear what feature could be involved, or unclear that the relevant feature is not already checked. 10. I think that movement chains also are indexed by Gen, with faulty indexings either ruled out by principles regarding vacuous quanti®cation and full interpretation, or not generated due to something about the architecture of Gen. This will mean that all cases of Merge are the same, and a given phrase can be in the numeration more than once, but there is no special operation of copying. 11. I assume there is a parallel constraint that applies to overt pronouns, but it is not crucial for the data being examined here. 12. Note that if the input contained an overt pronoun, the result would be the same in all the English cases. 13. There could be accidental coreference, such as when one person says ``No one hopes for Mary to win'' and another says ``Well, MARY hopes for her to win.'' I assume these readings do not involve coindexing. 14. Here I leave out the candidate where the NP object is coindexed with the subject. These will violate BTB if the object is a pronoun, and some other condition on R-Expressions if the object is a full NP. 15. See Hoonchamlong (1991:253±263) for arguments against Pingkarawat's proposal. 16. Huang treats null matrix subjects as variables bound by the null operator. 17. The feature di¨erence could either be a lexical di¨erence, or a di¨erence having to do with ranking of constraints on the realization of functional features. 18. We ought to ®nd languages with agreement that allow null pronouns, if they have the Thai constraint ranking. McCloskey and Hale (1984) argue that Irish is a language with weak agreement and null subjects.

Constraints on Null Pronouns

423

19. Thanks to Maria Nella Carminati for these examples. Although she reports that the null pronoun is ®ne in all three of these contexts, she reports a preference for the overt pronoun in (28) and (29), but no preference in (30). 20. Thanks to Elizabeth Villalta for these examples, and for bringing such examples to my attention. References Authier, J-M. 1989. Arbitrary Null Objects and Unselective Binding. In Oswaldo Jaeggli and Ken Sa®r (eds.), The Null Subject Parameter, 45±67. Dordrecht: Kluwer. Bach, Emmon. 1979. Control and Montague Grammar. Linguistic Inquiry 10: 515±531. Borer, Hagit. 1989. Anaphoric AGR. In Osvaldo Jaeggli and Ken Sa®r (eds.), The Null Subject Parameter, 69±109. Dordrecht: Kluwer. BosÏkovicÂ, ZÏeljko. 1997. The Syntax of Non®nite Complementation: An Economy Approach. Cambridge, Mass.: MIT Press. BosÏkovicÂ, ZÏeljko. 1995. Principles of Economy in Non®nite Complementation. Doctoral dissertation, University of Connecticut, Storrs. [Distributed by MITWPL, Department of Linguistics and Philosophy, MIT, Cambridge, Mass.] Bouchard, Denis. 1984. On the Content of Empty Categories. Dordrecht: Foris. Bresnan, Joan. Forthcoming. Optimal Syntax. In Joost Dekkers, Frank van der Leeuw, and Jeroen van de Weijer (eds.), Optimality Theory: Syntax, Phonology, and Acquisition. Oxford: Oxford University Press. Bresnan, Joan. 1982. Control and Complementation. Linguistic Inquiry 13: 343±434. Burzio, Luigi. 1998. Anaphora and Soft Constraints. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 93±113. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Campos, Hector. 1986. Inde®nite Object Drop. Linguistic Inquiry 17(2): 354±359. Cantrall, William. 1974. Viewpoint, Re¯exives, and the Nature of Noun Phrases. The Hague: Mouton. CastanÄeda, Hector-Neri. 1968. On the Logic of Attributions of Self-Knowledge to Others. Journal of Philosophy 65: 439±456. Chierchia, Gennaro. 1984. Topics in the Syntax and Semantics of In®nitives and Gerunds. Doctoral dissertation, University of Massachusetts, Amherst. Chomsky, Noam. 1986. Knowledge of Language: Its Nature, Origins, and Use. New York: Praeger. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Chomsky, Noam, and Howard Lasnik. 1993. The Theory of Principles and Parameters. In J. Jacobs, A. von Stechow, W. Sternefeld, and T. Vennemann (eds.), Syntax: An International Handbook of Contemporary Research, 506±569. Berlin: de Gruyter. [Reprinted as chap. 1 of Chomsky, The Minimalist Program.] Cole, Peter. 1987. Null Objects in Universal Grammar. Linguistic Inquiry 18: 597±612

424

Margaret Speas

Cole, Peter, Gabriella Hermon, and Li-May Sung. 1990. Principles and Parameters of LongDistance Re¯exives. Linguistic Inquiry 21: 1±22. Cole, Peter, and Li-May Sung. 1994. Head Movement and Long-Distance Re¯exives. Linguistic Inquiry 25: 355±406. Farrell, Patrick. 1990. Null Objects in Brazilian Portuguese. Natural Language and Linguistic Theory 8(3): 325±346. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28(3): 373±421. Grimshaw, Jane, and Vieri Samek-Lodovici. 1995. Optimal Subjects. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 598±606. GLSA, University of Massachusetts, Amherst. Hoonchamlong, Yuphaphann. 1991. Some Issues in Thai Anaphora: A Government and Binding Approach. Doctoral dissertation, University of Wisconsin, Madison. Huang, James. 1984. On the Distribution and Reference of Empty Pronouns. Linguistic Inquiry 15: 531±574. Huang, James. 1989. Pro-Drop in Chinese. In Oswaldo Jaeggli and Ken Sa®r (eds.), The Null Subject Parameter, 185±214. Dordrecht: Kluwer. Huang, James. 1991. Remarks on the Status of the Null Object. In Robert Freiden, (ed.), Principles and Parameters in Comparative Grammar, 56±76. Cambridge, Mass.: MIT Press. Jaeggli, Osvaldo, and Kenneth Sa®r. 1989. The Null Subject Parameter. Dordrecht: Kluwer. Kayne, Richard. 1989. Facets of Romance Past Participle Agreement. Paola BenincaÁ (ed.), Dialect Variation and the Theory of Grammar, 85±103. Dordrecht: Foris. Keer, Edward, and Eric BakovicÂ. 1997. Have Faith in Syntax. In Emily Curtis, James Lyle, and Gabriel Webster (eds.), Proceedings of the Sixteenth Annual West Coast Conference on Formal Linguistics, 225±270. Stanford, Calif.: CSLI Publications. Kitagawa, Chisato. 1998. Introductory pro in Japanese Journalistic Narratives. Ms., University of Massachusetts, Amherst. Kuno, Susumu. 1987. Functional Syntax. Chicago: University of Chicago Press. Ladusaw, William, and David Dowty. 1988. Toward a Nongrammatical Account of Thematic Roles. In Wendy Wilkins (ed.), Syntax and Semantics 21: Thematic Relations, 62±74. New York: Academic Press. Langendoen, Terence, and Xu Liejiong. 1985. Topic Structures in Chinese. Language 61(1): 1±27. Larson, Richard. 1991. Promise and the Theory of Control. Linguistic Inquiry 22: 103±139. Legendre, GeÂraldine, Paul Smolensky, and Colin Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Manzini, M. Rita. 1983. On Control and Control Theory. Linguistic Inquiry 14: 421±446. Martin, Roger. 1992. On the Distribution and Case Features of PRO. Ms., University of Connecticut, Storrs. McCloskey, James, and K. Hale. 1984. On the Syntax of Person-Number In¯ection in Modern Irish. Natural Language and Linguistic Theory 1(4): 487±534.

Constraints on Null Pronouns

425

Mitchell, Jonathan. 1986. The Formal Semantics of Point of View. Doctoral Dissertation, University of Massachusetts, Amherst. Morgan, Jerry. 1970. On the Criterion of Identity for Noun Phrase Deletion. CLS 6, 380±389. Chicago Linguistic Society, University of Chicago. Pingkarawat, N. 1989. Empty Noun Phrases in Thai and the Theory of Control. Doctoral dissertation, University of Illinois, Champagne-Urbana. Rizzi, Luigi. 1986. Null Objects in Italian and the Theory of pro. Linguistic Inquiry 17: 501± 558. Samek-Lodovici. 1996. Constraints on Subjects: An Optimality-Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Â rmann. 1989. Verbal Syntax and Case in Icelandic. Doctoral dissertaSigurDsson, HalldoÂr A tion, University of Lund. Speas, Margaret. 1995. Generalized Control and Null Objects in Optimality Theory. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 637±654. GLSA, University of Massachusetts, Amherst. Speas, Margaret. 1997. Optimality Theory and Syntax: Null Pronouns and Control In D. Archangeli and D. T. Langendoen (eds.), Optimality Theory: An Overview, 171±199. Oxford: Blackwell. Sportiche, Dominique. 1983. Structural Invariance and Symmetry. Doctoral dissertation, MIT, Cambridge, Mass. Stowell, Tim. 1982. The Tense of In®nitives. Linguistic Inquiry 13: 561±570. Tenny, Carol. 1998. Person and Evidentiality in Japanese. Paper presented at the Symposium on Diachronic and Synchronic Studies of Syntax of East Asian Languages, University of Southern California, Los Angeles, November 6±8. Williams, Edwin. 1987. Implicit Arguments, the Binding Theory, and Control. Natural Language and Linguistic Theory 5: 151±180. Yoon, James. 1985. On the Treatment of Empty Categories in Topic Prominent Languages. Ms., University of Illinois, Champaign-Urbana.

Chapter 14 Vo -to-Io Movement and do-Insertion in Optimality Theory

Sten Vikner

The empirical phenomena to be discussed in this chapter are the position of the ®nite verb in embedded clauses (preceding or following a sentential adverbial, i.e., in Io or in Vo ) and the possibility or impossibility of do-insertion. For reasons of space, ways that the analysis may be extended to cover cases (mostly main clauses) where the verb may or may not move to a position to the left of the subject (i.e., to Co ) will only be brie¯y sketched at the end of the chapter. To a large extent, I will base my analysis on Grimshaw's (1997) analysis of English, and attempt to modify and extend it to also cover the variation found between English, Danish, French, and Icelandic. One of my goals is to illustrate how Optimality Theory (OT) can account for language variation in terms of di¨erences in constraint rankings between languages. A central feature of OT (see the introductory chapter of the present volume and references there) is that constraints are taken to be relative (``soft'') rather than absolute (``hard''); (1) makes the distinctions clear. (1) a. Absolute: ``If a sentence violates constraint C, it is ungrammatical.'' b. Relative: ``That a sentence violates constraint C may have a certain cost, but this cost is smaller than the cost of violating constraint B, which again is less costly than a violation of constraint A.'' In other words, although there is a price to be paid every time a constraint is violated, the price is not always the grammaticality of the sentence in question. As will become clear, my particular version of OT owes a heavy debt to the principles-and-parameters framework (see Chomsky 1986; Rizzi 1990; and, e.g., Vikner 1995:3±35), although violability of constraints was not possible within this framework. Furthermore, although violability is a feature of the Minimalist framework (Chomsky 1995), this is only the case to a certain extent (see, e.g., the discussion in Vikner 1997a:12±13), although certain implementations of minimalism di¨er from Chomsky 1995 precisely in assigning a more central role to violability (see, e.g., Bobaljik 1995:351 and Holmberg 1997:214).

428

Sten Vikner

However, violability is not the only central feature of OT. Following Prince and Smolensky (1993) and Grimshaw (1997:373), I take the four ideas presented in (2) to sum up the essence of OT. (2) a. Constraints may be violated. b. Constraints are ranked in a hierarchy (a grammar is a particular ranking of constraints). c. Constraints are universal, i.e., in all languages, the same constraints apply, except that they are ranked di¨erently from language to language (language variation is variation in the constraint hierarchy). d. Only the optimal version of a sentence is grammatical; all nonoptimal versions are ungrammatical (the optimal candidate of two is the one with the fewest violations of the highest constraint on which the two candidates di¨er). By way of illustration of (2), consider the variation in position of the ®nite verb in English and French, as in (3) and (4), as originally discussed in Emonds 1978 and Pollock 1989. This particular di¨erence is one of the central di¨erences to be discussed in this chapter. The same di¨erence, although restricted to embedded clauses, is found between Danish and Icelandic, as will be shown in more detail later. (3) English really saw the ®lm. a. The actor * the ®lm. b. The actor saw really (4) French vraiment voit le ®lm. a. *L'acteur le ®lm. b. L'acteur voit vraiment (sees) the ®lm The actor (sees) really I am assuming (at least) two kinds of movement to take place here: the ®nite in¯ectional a½x of the verb moves from Io to Vo in (3a)/(4a), whereas the ®nite verb moves from Vo to Io in (3b)/(4b). This variation can be seen as resulting from the interaction between two constraints, which will be discussed in more detail in connection with (13) in section 14.2: LXMV (No Movement of a Lexical Head: violated when a verb leaves Vo ) and PRBD (Proper Binding: violated when a traceÐhere in Io Ðc-commands its antecedent, here in Vo ). In English, the price to be paid when a trace c-commands its antecedent Ðthat is, when something is moved downward in the treeÐis smaller than the price to be paid when a lexical category (here the main verb) is moved, because the ranking is LxMv g PrBd. In French, on the other hand, the ranking is PrBd g LxMv, and so the cost of moving the verb out of VP is smaller than the cost of moving an element downward in the tree. (See (5) and (6).)

Vo -to-Io Movement and do-Insertion

429

(5) English (and embedded clauses in Danish) LxMv g PrBd a. The actor vj really sawj the ®lm. violates PRBD ˆ (10d) b. *The actor sawj really vj the ®lm. violates LXMV ˆ (10e) (6) French (and embedded clauses in Icelandic) PrBd g LxMv a. *L'acteur vj vraiment voitj le ®lm. violates PRBD ˆ (18a) b. L'acteur voitj vraiment vj le ®lm. violates LXMV ˆ (18b) As we see, constraints are violable, (2a); note that both (5a) and (6b) violate a constraint and yet they are grammatical. Constraints are ranked in a hierarchy, (2b), and they are universal, (2c); note that the same constraints are found in English and French, though in a di¨erent ranking. Finally, only the optimal version of a sentence is grammatical, (2d). Both (5b) and (6a) are ungrammatical, because they are less optimal than their competitors (5a) and (6b). On the highest-ranked constraint on which there is a di¨erence, (5a) has fewer violations than its competitor (5b), and (6b) has fewer violations than its competitor (6a). 14.1

Preliminary Assumptions

Before I begin the discussion of the analysis itself, I want to set out a few preliminary assumptions. The competing versions of a sentence (the ``candidates'') are di¨erent realizations of the same input. I will follow Grimshaw (1997:375±376) and assume that ``the input for a verbal extended projection is a lexical head plus its argument structure and an assignment of lexical heads to its arguments, plus a speci®cation of the associated tense and aspect.'' A typical set of competing candidates, in this case for an embedded negated clause, contains at least the examples in (7). (7) English a. *Although b. *Although c. Although d. *Although

the ®lm Peter saw not the ®lm not saw Peter see the ®lm Peter did not not did see the ®lm Peter

... ... ... ...

Now it is up to the OT linguist to ®gure out which constraints are relevant here and how they are ranked, so that only (7c) is optimal in English (and only (7b) in Danish, and only (7a) in French and Icelandic). When do is used as a so-called dummy or light verb, as in (7c), it is not part of the input. This contrasts with the use of do as a normal verb (e.g., in Peter did his homework) and also with auxiliaries like have, be, will, can, . . . , all of which I take to be part of the input.

430

Sten Vikner

I assume that only lexical Xo s/XPs are selected, and that functional Xo s/XPs are projected, in the sense of Grimshaw 1991. I furthermore assume that Io is projected (in the above-mentioned sense) by Vo and separately from Vo , and I take the subject to be base generated in SpecVP (see, e.g., Koopman and Sportiche 1991:212). Sentential adverbials are taken to left-adjoin to VP, as in (8c). In Grimshaw 1997:383, where Io is not always projected, the typical English declarative clause may consist only of a VP, (8a). If this were also the case in clauses containing a sentential adverbial (Grimshaw 1997:382, note 6), then the sentential adverbial (which I take to be an XP) would have to be adjoined to a V-bar (a nonmaximal category), (8b), which would clash with standard assumptions concerning X-bar theory (also incompatible with, e.g., Alexiadou 1997 and Cinque 1999, where the ®xed order of adverbials is due to their being in the speci®er position of di¨erent XPs between IP and VP). (See (8).) (8) English a. I think that [VP the actor saw the ®lm]. b. I think that [VP the actor [V0 actually [V 0 saw the ®lm]]]. c. I think that [IP the actor [VP actually [VP t saw the ®lm]]]. In the analysis suggested below, adverbials are not taken to adjoin to V-bar; the word order is analyzed with the sentential adverbial being (left-)adjoined to VP and the subject in SpecIPÐthat is, as (8c). The constraint responsible for forcing the subject out of SpecVP is Case, since the subject needs to move to SpecIP to be assigned nominative (see Grimshaw 1997:390). Finally, I will be using the notational conventions; see (9). (9) a. b. c. d. e. f. g. h. i.

f ! t v Ð e Su Ob A

marks the optimal candidate(s), i.e., the grammatical version(s). marks the fatal violation in all nonoptimal candidates. is a trace of a maximal projection. is a trace in the verb chain (a trace of the verb or of the in¯ection). is an empty maximal projection. is an empty Xo (either only abstract features or radically empty). means ``subject.'' means ``object.'' means ``adverbial.''

Note that regardless of whether the example in question is English or not, the lexical items in the corresponding tableau are always English. 14.2

The Six Basic Constraints

The basic phenomena to be considered are the various possible positions of the ®nite verb in Danish, English, French, and Icelandic, as frequently discussed in the

Vo -to-Io Movement and do-Insertion

431

literatureÐfor example, Emonds 1978; Pollock 1989; Holmberg and Platzack 1995; Vikner 1995, 1997b. I will begin the analysis with the six constraints that I take to be involved in the derivation of embedded declaratives (see (10) and tableau T14.1). (10) Danish Jeg tror (at) skuespilleren virkelig saÊ ®lmen. English I think (that) the actor actually saw the ®lm. Tableau T14.1 shows an analysis valid for both Danish and English. In both languages the optimal candidate is (d). Candidates (a) to (u) make up a rather extensive set of candidates, which vary with respect to three di¨erent properties: presence or absence of do, the position of the subject (SpecVP, SpecIP, or SpecCP), and the position of the ®nite verb (Vo , Io , or Co ). Tableau T14.1

432

Sten Vikner

Tableau T14.2 (Repeated from tableau T14.1(a) and (d)) Case a. *C [IP Ð v A [VP Su saw Ob f d.

C [IP Su v A [VP t

ObHd

LxMv

*!

saw Ob

PrBd

Stay

*

*

*

**

Tableau T14.3 (Repeated from tableau T14.1(d) and (g)) Case f d.

C

ObHd

[IP Su v A [VP t saw Ob

g. *C [CP Su e [IP t

v A [VP t saw Ob

*!

LxMv

PrBd

Stay

*

**

*

***

The ®rst of the ®ve constraints in tableau T14.1 is CASE. (Every DP or DP-chain must have case. See Grimshaw 1997:390 and the case ®lter in Chomsky 1981:49; see also Costa, chap. 7, this volume, Samek-Lodovici 1996:169, and Woolford, chap. 16, this volume, for more elaborate OT implementations of the case ®lter.) Case makes the subject leave SpecVP and move to SpecIP, on the assumption that nominative is assigned to SpecIP. Compare (a) to (d); see tableau T14.2. OBHD (Obligatory Heads; an Xo may not be emptyÐsee Grimshaw 1997:377 and Haider 1988:101) ensures that there are no radically empty heads. A head containing a trace is not taken to be empty. Compare (g) to (d); see tableau T14.3. I am assuming that X-bar theory forces the existence of a head (Co in (g)) whenever the relevant speci®er position is ®lled (SpecCP in (g), ®lled by the subject). LXMV (No Movement of a Lexical Head; see Grimshaw 1997:385) punishes saÊ/saw for leaving Vo . Compare (e) to (d); see tableau T14.4. PRBD (Proper Binding; traces must not c-command their antecedent, an updated version of Fiengo 1977:45,53) is the (low) cost of having an `upside-down'' chain, where the trace in Io c-commands the antecedent in Vo . This constraint is not actually necessary for verb movement in Danish and English, since no candidates are ruled out by it, but it will be necessary for verb movement in French and Icelandic; see (11) and tableau T14.9 (p. 436). PrBd is probably also relevant for other kinds of movement in Danish and English, which like verb movement in French and Icelandic go upward, not downward.

Vo -to-Io Movement and do-Insertion

433

Tableau T14.4 (Repeated from tableau T14.1(d) and (e)) Case f d.

C [IP Su v

ObHd

LxMv

PrBd

Stay

*

**

A [VP t saw Ob

e. *C [IP Su saw A [VP t v

Ob

*!

**

Tableau T14.5 (Repeated from tableau T14.1(d) and (p)) Case f d.

C [IP Su v A [VP t saw Ob

p. *C [IP Su v A [VP t did [VP t see Ob

ObHd

LxMv

PrBd

Stay

*

**

*

***!

STAY (traces should be avoided; see Grimshaw 1997:375, and also ``*trace'' in Legendre et al. 1995:611 and economy of derivation in Chomsky 1991:432, 1995:145) punishes every step of every movement (``movement as last resort''). Compare (p) to (d); see tableau T14.5. Here do is inserted under Vo (for an alternative possibility, direct insertion of do under Io , see section 14.4). The ranking of the ®ve constraints in tableau T14.1 thus causes (d) to be optimalÐ that is, better than its competitors (a) to (c) and (e) to (u). This is why (d) gives rise to a grammatical sentence, whereas the competitors either result in ungrammatical sentences or in a sentence that might also be analyzed as (d) (i.e., (g)). However, (d) is not the only possible grammatical version of the sentence in question: it is also possible to leave out the complementizer that. The candidates in tableau T14.6 are all further candidates competing with (and parallel to) the ones already encountered above, in tableau T14.1, except that they do not contain the complementizer that. This illustrates one way of allowing for optionality in OT: the two winning candidates, (d) in tableau T14.1 and (d) in tableau T14.6, have the same constraint pro®le (as suggested by Grimshaw 1997:410; see Bakovic and Keer, chap. 4, this volume, and MuÈller 1999 for further discussion of various ways of deriving optionality). The only di¨erence between the two tableaux containing candidates from the same competition, tableaux T14.1 and T14.6, is the absence versus presence of that. This is relevant for PRPR (Projection Principle), which is violated when there is movement into the highest Xo of an embedded clause (following Grimshaw's

434

Sten Vikner

Tableau T14.6 (Continues and completes tableau T14.1)

1997:393 suggestion, based on Rizzi and Roberts 1996:106; see also McCloskey 1992 and MuÈller 1997:262, and notice that in Grimshaw 1997:393 and Bakovic 1998:40, PrPr is called ``Pure-EP''). PrPr was not shown in tableau T14.1, because it was not violated by any of the candidates there. For PrPr to be violated, there has to be movement from Vo to Io ; compare (b) and (c) in tableau T14.7. Io has to be the highest Xo of the embedded clauseÐthat is, Co has to be absent; compare (a) and (c) in the same tableau. The introduction of PrPr has no consequences in this Danish/English example because the optimal candidate does not involve Vo -to-Io movement. In tableau T14.8, it does not a¨ect (b), the version of (a) without the complementizer at/that, which thus has exactly the same constraint pro®le as (a). Given that this constraint pro®le is the optimal one, both candidates are optimal and thus grammatical.

Ob

A [VP t saw Ob

[IP Su saw A [VP t v

[IP Su v

Ob

f b.

*

PrBd

PrBd

*

LxMv

*

*!

LxMv

[IP Su v A [VP t saw Ob

ObHd

ObHd

*

PrPr

*!

PrPr

f a. C [IP Su v A [VP t saw Ob

Case

Tableau T14.8 (Repeated from tableaux T14.1(d) and T14.6(d))

c. *

f b.

a. *C [IP Su saw A [VP t v

Case

Tableau T14.7 (Repeated from tableaux T14.1(e) and T14.6(d) and (e))

**

**

Stay

**

**

**

ˆ (15e)

ˆ (15d)

ˆ (10e)

ˆ (15d)

ˆ (10d)

Stay

Vo -to-Io Movement and do-Insertion 435

436

Sten Vikner

In fact, none of the candidates in tableau T14.1, violate PrPr, irrespective of whether they have movement of the ®nite verb, due to the presence of the complementizer at/that, which will always count as the highest head of the embedded clause. This whole line of argument presupposes that embedded clauses with and without that may both be winning candidates in the same competition (``candidates realizing the same input''), although they do not contain exactly the same lexical elements (in minimalist terms, they do not have the same ``numeration''; see Chomsky 1995:225). For an alternative OT analysis that takes clauses with and without that to be winners of di¨erent competitions, see Legendre et al. 1995; Legendre, Smolensky, and Wilson 1998 and Bakovic & Keer (chap. 4, this volume). 14.3

Optional vs. Obligatory Complementizers

Consider now embedded declaratives in French and Icelandic, which are derived by means of the same six constraints. In these two languages, the ®nite verb precedes the adverbial instead of following it, and there is also no complementizer optionality such as the one found in Danish and English: que=aD `that' is obligatory (see (11) and tableau T14.9). (11) French Je crois *(que) l'acteur voit vraiment le ®lm. Icelandic EÂg tel *(aD) leikarinn sjaÂi aÂreiDanlega myndina. I think (that) the actor sees really the ®lm As brie¯y discussed at the beginning of the chapter, I propose a di¨erence in constraint ranking between Danish/English and French/Icelandic: Danish and English have LxMv g PrBd; French and Icelandic have PrBd g LxMv (see tableau T14.10). In French/Icelandic, violating LxMv (moving the verb out of VP) is thus less costly than violating PrBd (leaving the verb in Vo where it is c-commanded by a Tableau T14.9 Case a. *C [IP Su v f b.

PrPr

A [VP t saw Ob

C [IP Su saw A [VP t v

c. *

[IP Su v

d. *

[IP Su saw A [VP t v

ObHd

PrBd *!

Ob *! *!

Stay **

*

A [VP t saw Ob Ob

LxMv

** **

*

**

Vo -to-Io Movement and do-Insertion

437

Tableau T14.10 Danish and English:

Case

PrPr

ObHd

French and Icelandic:

Case

PrPr

ObHd

LxMv

PrBd

PrBd

LxMv

Stay

Stay

coindexed trace in Io ), exactly the opposite of Danish/English. Hence in tableau T14.9 (b) is optimal and (a) is not, whereas in tableau T14.1 (d) (the Danish/English parallel to (a) in the earlier tableau) was optimal and (e) (the Danish/English parallel to (b) in the earlier tableau) was not. As I originally observed some years ago (see Grimshaw 1997:413), it now follows that the French/Icelandic complementizer que=aD `that' is obligatory. Leaving the complementizer out of the optimal (b) in tableau T14.9 results in (d), which violates PrPr, since it has movement into the highest Xo of an embedded clause (parallel to a suggestion in Deprez 1994:124). 14.4

Insertion of Auxiliaries in Io and Economy of Projection

In Danish, French, and Icelandic, ®nite auxiliary verbs behave exactly like their nonauxiliary counterparts. All ®nite verbs follow the adverbial in Danish, and all ®nite verbs precede the adverbial in French and Icelandic (see (12) and (13)). (12) Danish Jeg tror French Je crois Icelandic EÂg tel I think (13) Danish Jeg tror French Je crois Icelandic EÂg tel I think

(at) skuespilleren

virkelig

®lmen.

que l'acteur

voit

aD leikarinn that the actor

myndina. sjaÂi aÂreiDanlega (sees) the ®lm (sees) really

(at) skuespilleren

vraiment

ser

virkelig vraiment

le ®lm.

har set ®lmen.

que l'acteur

a

vu

le ®lm.

aD leikarinn that the actor

seÂD myndina. ha® aÂreiDanlega (has) seen the ®lm (has) really

438

Sten Vikner

This is not the case in English, however: ®nite main verbs follow the adverbial (like Danish), but ®nite auxiliaries precede the adverbial (like French and Icelandic) (see (14)). (14) English a. I think (that) the actor actually saw the ®lm. b. I think (that) the actor has actually seen the ®lm. English even has one more complication, namely, that main verb be behaves like an auxiliary rather than like a main verb. To capture this distributional di¨erence inside one and the same language, the means used in sections 14.2 and 14.3, reranking of constraints, is not available; constraint ranking is language speci®c, not construction speci®c. The solution to be suggested here (which is an attempt to derive the analysis stipulated in Grimshaw 1997:382; see also Emonds 1994:157±164) is that the English ®nite auxiliary is inserted directly under Io , in contrast to English ®nite main verbs and to both main and auxiliary ®nite verbs in Danish, French, and Icelandic, all of which are inserted under Vo . I would like to suggest that two con¯icting constraints exist: EcVP and VVo . Depending on which is more highly ranked, (auxiliary) verbs are inserted in Vo or in Io . ECVP (Economy of VPs) punishes every occurrence of a VPÐthat is, the fewer VPs used in a derivation, the better. If this constraint is the crucial one, insertion of a verb directly under Io is preferred, because it uses one less VP than when a verb is inserted under Vo . See (15) and tableau T14.11, and compare (b) and (c) in the tableau. I formulate this constraint as EcVP for expository reasons only; it might as well have been formulated as a more general EcXP (Economy of Maximal Projections). VV o (Verb in Vo ) punishes chains that contain a verb but that do not include a Vo (all verbs should be inserted into their own Vo and project their own VP), so if this constraint is the crucial one, insertion of a verb under Vo is preferred. Again compare (b) and (c) in tableau T14.11. Notice that I take both main verbs and auxiliary verbs to be verbs in the sense of this constraint. This di¨erence between (a) and (b) on the one hand and (c) on the other is that (c) has insertion of has directly under Io and therefore has one VP less than the corresponding candidate with has inserted under Vo , (b). (15) English I think (that) the actor has actually seen the ®lm. Candidates (d) to (f ) correspond to (a) to (c) but lack the complementizer that. The presence or absence of that makes no di¨erence here; the constraint pro®les of (c)

[VP t seen Ob

[IP Su has A

f f.

[VP t seen Ob

[IP Su has A [VP t v

e. *

A [VP t has [VP t seen Ob

[VP t seen Ob

[IP Su v

C [IP Su has A

d. *

f c.

[VP t seen Ob

A [VP t has [VP t seen Ob

b. *C [IP Su has A [VP t v

a. *C [IP Su v

Tableau T14.11 English, ex. (15) yFl

Case

*!

PrPr

ObHd

*

**

**!

*

**!

**!

EcVP

*

*

LxMv

*

*

VVo

*

*

PrBd

*

***

***

*

***

***

Stay

Vo -to-Io Movement and do-Insertion 439

440

Sten Vikner

and (f ) are identical. Therefore both are optimal, accounting for the optionality of that. The reason there is no violation of PrPr in (f ) is that although the ®nite verb is in the highest Xo of the embedded clause, it was not moved there. If the ®nite auxiliary had been moved to Io , the relevant candidates would be (b) and (e), and since (e) will always be less optimal than (b), the optionality of that would have been lost. Tableau T14.11 also contains a third new constraint, which is ranked very highly. yFL (Theta Filter) punishes every argument not assigned a thematic role. I take thematic roles to be assigned by Vo and Po inside their maximal projections (following Chomsky 1981 and Koopman and Sportiche 1991) and so if a main verb is inserted directly under Io , its arguments will not be assigned their thematic roles (indeed, the sentence will not contain a VP at all). To see how this actually works, we have to consider a case where the ®nite verb is a thematic verb (one that assigns one or more thematic roles). The example corresponding to (c) in tableau T14.11Ðthat is, that contains a ®nite main verb inserted under Io (see (c) in tableau T14.12)Ðwill thus lose out to its rival candidate ((a) in tableau T14.12), due to the e¨ects of yFl (neither the subject nor the object is assigned a thematic role). Consider (16). (16) English I think that the actor (actually) saw the ®lm. Notice that yFl only punishes arguments that do not receive a thematic role, and not thematic roles that are not assigned to an argument. I think that the latter, unassigned thematic roles, may indeed exist in all four languages under discussionÐ for example, when have and do are used as auxiliary verbs. Therefore, a constraint punishing unassigned thematic roles may have to be ranked lower than yFl, which punishes the opposite: arguments that do not receive a thematic role. Until now, the constraints discussed were part of the evaluation component of the grammar, Eval (see the introductory chapter), and so whether or not the optimal candidate violates a given constraint depends on the ranking of that particular constraint. Even if a given constraint is unviolated in a particular grammar, there is no guarantee that this is also the case in a di¨erent grammar. Another possibility is that a particular constraint is part of the candidate-generating component of the grammar, Gen. In this case it cannot possibly be violated by the optimal candidate, because Gen will not generate any candidates that violate it. If a constraint (e.g., yFl ) is part of Gen, it is never violated in any languages, because Gen does not vary between languages; language variation results from variation in the constraint ranking in Eval. yFl prefers main (thematic) verbs to be inserted under Vo , but says nothing about auxiliary verbs. VVo prefers all verbs to be inserted under Vo , and EcVP prefers all

Ob Ob

c. C [IP Su saw A

A [VP t saw Ob

b. C [IP Su saw A [VP t v

f a. C [IP Su v

Tableau T14.12 English, ex. (16)

*!*

yFl

Case

PrPr

ObHd

*

*

EcVP

*!

LxMv

*

VVo *

PrBd

**

**

Stay

ˆ (10e)

ˆ (10d)

Vo -to-Io Movement and do-Insertion 441

442

Sten Vikner

verbs to be inserted under Io . If it is assumed that yFl is unviolable (i.e., part of Gen), then main verbs are generated in Vo in all languages, and only the point of insertion of auxiliary verbs (i.e., verbs that do not assign a thematic role) is subject to language variation. Either VVo is ranked above EcVP, in which case all verbs are inserted under Vo (Danish, French, and Icelandic), or EcVP is ranked above VVo , in which case nonthematic verbs are inserted under Io (English). Only if nonthematic verbs may be inserted directly under Io (i.e., only under the English ranking) is do-support possible, because under the Danish/French/Icelandic ranking, nonthematic verbs like dummy do would also have to be inserted under Vo , in which case such an insertion would not achieve anything. The reason English has do-support with negation is that both Vo -to-Io movement and Io -to-Vo movement across negation may be avoided; see sections 14.5 and 14.6 (do-support with movement to Co behaves in a parallel fashion; see Vikner 1998). A reason for linking the auxiliary/main verb di¨erence in English to thematic roles may be found in the following di¨erence between auxiliary have/do and main verb have/do; see (17) and (18). Only the auxiliaries occur in Io (as can be seen from their appearance left of negation, left of a sentential adverbial, left of a subject quanti®er, in tag questions, and in VP-ellipses; Scholten 1988:3±7). (17) English a. He b. *He did c. He did d. *He did e. *He did f. He did

actually did the dishes. the dishes. actually do the dishes. not the dishes. not not do see the ®lm. see the ®lm. not

(main verb in V ) (main verb in I ) (main verb in V ) (main verb in I ) (auxiliary verb in V ) (auxiliary verb in I )

(18) English (main verb in V ) actually had a lot of money. a. She a lot of money. b. *She had actually (main verb in I ) have a lot of money. c. She did not (main verb in V , US) a lot of money. d. *She had not (main verb in I ) (``somewhat uncommon,'' Quirk et al. 1985:131) (verb: ``get,'' UK, informal) got a lot of money. e. She had not actually had earned a lot of money. (auxiliary verb in V ) f. ?She earned a lot of money. (auxiliary verb in I ) g. She had actually have earned a lot of money. (auxiliary verb in V ) h. *She did not earned a lot of money. (auxiliary verb in I ) i. She had not There are no such di¨erences between auxiliary be and main verb be; both occur in Io . (See (19).)

Vo -to-Io Movement and do-Insertion

(19) English a. ?She b. She c. *She d. She e. ?He f. He g. *He h. He

actually actually is does not not is actually actually is does not not is

is very nice. very nice. be very nice. very nice. is working. working. be working. working.

443

(main verb in Vo , not stressed) (main verb in Io ) (main verb in Vo ) (main verb in Io ) (auxiliary verb in Vo , not stressed) (auxiliary verb in Io ) (auxiliary verb in Vo ) (auxiliary verb in Io )

Roberts (1985:30) and Scholten (1988:160) suggest that only verbs that do not assign thematic roles may be inserted under Io . This gives the right prediction concerning main verb be, which presumably does not assign a thematic role (Haegeman 1994:68; Parsons 1995:645; Williams 1994:44), unlike main verb have/do, but like auxiliary have/be/do. This is captured here by taking yFl to be inviolable (i.e., part of Gen), which will ensure that verbs that assign thematic roles are universally generated in Vo . Consider now what happens in Danish, where VVo is ranked above EcVP; see (20) and tableau T14.13. (20) Danish Jeg tror (at) skuespilleren virkelig har set ®lmen. I think (that) actor-the really has seen ®lm-the In tableau T14.13, candidate (c)Ðwhich was optimal in EnglishÐloses out to (a), because it is more important in Danish that every verb chain contains a Vo (VVo ) than it is to keep the number of VPs to an absolute minimum (EcVP). PrPr is not relevant here: the optimal candidate without at `that' does not violate it, since there is no movement into the highest Xo of an embedded clause. Candidates (a) and (d) are thus both optimal. Consider ®nally what happens in French and Icelandic in (21) and tableau T14.14. Here VVo is ranked above EcVP (like Danish but unlike English). (21) French Je crois *(que) l'acteur a vraiment vu le ®lm. Icelandic EÂg tel *(aD) leikarinn ha® aÂreiDanlega seÂD myndina. I think (that) the actor has really seen the ®lm In tableau T14.14, (c)Ðwhich was optimal in EnglishÐloses out to (b), because it is more important in French and Icelandic (as in Danish) that every verb chain contains a Vo (VVo ) than it is to keep the number of VPs as low as possible (EcVP).

[VP t seen Ob

[IP Su has A

f. *

C [IP Su has A [VP t v

[VP t seen Ob [VP t seen Ob

[IP Su has A [VP t v

[IP Su has A

e. *

f. *

A [VP t has [VP t seen Ob

[IP Su v

d. *

[VP t seen Ob

[VP t seen Ob

A [VP t has [VP t seen Ob

c. *C [IP Su has A

f b.

a. *C [IP Su v

Tableau T14.14 French/Icelandic, ex. (21)

[VP t seen Ob

A [VP t has [VP t seen Ob

[IP Su has A [VP t v

[IP Su v

e. *

f d.

[VP t seen Ob

c. *C [IP Su has A

A [VP t has [VP t seen Ob [VP t seen Ob

C [IP Su v

b. *C [IP Su has A [VP t v

f a.

Tableau T14.13 Danish, ex. (20)

yFl

yFl

Case

Case

*!

PrPr

*!

PrPr

ObHd

ObHd

*!

*!

VVo

*!

*!

VVo

*

**

**

*

**

**

EcVP

*

**

**

*

**

**

EcVP

*!

*!

PrBd

*

*!

LxMv

*

*

LxMv

*

*

PrBd

*

***

***

*

***

***

Stay

*

***

***

*

***

***

Stay

444 Sten Vikner

Vo -to-Io Movement and do-Insertion

445

PrPr is relevant in tableau T14.14; any version of the optimal candidate that lacks que=aD `that' would violate PrPr, since there would be movement into the highest Xo of an embedded clause. Compare (e) to (b). As shown in tableau T14.9, in French and Icelandic, PrBd is ranked above LxMv, whereas in Danish and English, LxMv is ranked above PrBd. When PrBd is ranked above LxMv, verbs inserted under Vo move from Vo to Io rather than stay in Vo . Compare (a) and (b) in tableau T14.14. So far I have discussed the ranking between EcVP and VVo independently of the ranking between LxMv and PrBd. The interaction among all four constraints is as follows. If VVo outranks EcVPÐthat is, if all verbs are inserted under Vo Ðthere are two relevant possibilities: the result is French/Icelandic (all ®nite verbs undergo Vo -to-Io movement) if PrBd outranks LxMv, and it is Danish (no ®nite verbs undergo Vo -to-Io movement) if LxMv outranks PrBd. If EcVP outranks VVo Ðthat is, if nonthematic verbs are inserted under Io and thematic verbs under Vo Ðno variation is found when the input contains a nonthematic verb, because this verb will occur in Io regardless of the ranking of the other constraints. Variation is found with EcVP outranking VVo when there is no nonthematic verb in the input. There are three relevant possibilities: The result is English if PrBd is outranked by the other three constraints (i.e., the ®nite thematic verb occurs in Vo ). The result would be a variant of English with Vo -to-Io movement of thematic verbs if LxMv were outranked by the other three constraints. Finally, the result would be another variant of English with do-insertion with thematic verbs in all contexts if VVo were outranked by the other three constraints. Until now two di¨erences in constraint ranking have been assumed: 1. English/Danish di¨er from French/Icelandic as far as PrBd is concerned (absence vs. presence of Vo -to-Io movement). 2. English di¨ers from the other three as far as VVo is concerned (di¨erence vs. no di¨erence between thematic and nonthematic verbs, and also presence vs. absence of do-insertion); see tableau T14.15. 14.5

Negation, the HMC, and Insertion of Empty do in Io

For negated clauses, no further ranking di¨erences between the languages are necessary, just one additional constraint, the Head Movement Constraint (HMC), which may be taken as being ranked immediately below EcVP in all four languages. In earlier versions of this chapter, I attempted to explain the di¨erent e¨ects of negation in the four languages by saying that Nego could not be part of the verb

yFl

yFl

yFl

English:

Danish:

French/Icelandic:

Tableau T14.15

Case

Case

Case

PrPr

PrPr

PrPr

ObHd

ObHd

ObHd

EcVP

EcVP

VVo VVo

EcVP

PrBd

LxMv

LxMv

LxMv

VVo

PrBd

PrBd

Stay

Stay

Stay

446 Sten Vikner

Vo -to-Io Movement and do-Insertion

447

chain in English (not being an Xo in Nego ), but the verb could move through Nego in the other three (Danish ikke, French pas, Icelandic ekki being XPs in SpecNegP). This could be achieved by two constraints NegX o and NegXP, which would be ranked one way in English and the other way in the other three languages. However, such an approach does not seem particularly insightful. I now want to show that it is possible to use only constraints independently necessary: VV o (as seen above) and the HMC (see Travis 1984:131; Baker 1988:53; Rizzi 1990:11). I would like to suggest that the HMC is violated whenever Nego intervenes in the verb chain in any of the four languages (i.e., when Io c-commands Nego and Nego c-commands Vo ). This is least controversial for English, where not has often been taken to be in Nego , and therefore to block the formation of a chain between Vo and Io (Pollock 1989:397; Roberts 1993:338, note 21). (See (22) and tableau T14.16.) (22) English I think (that) the actor did not see the ®lm. HMC punishes every Xo that intervenes in an Xo chain (i.e., that is c-commanded by a link of an Xo chain and itself c-commands another link in the same Xo chain) unless it is bound by the antecedent (the lexical element) of the chain. Nego is taken to be such a potential intervening element when the verb moves across it. It does not matter whether the movement would have been downward (compare (a) and (e) in tableau T14.16) or upward (compare (b) and (e)). If tableau T14.16 had not contained a negation, there would have been no do-insertion and (a) would have been more optimal than (e). EcVP (Economy of VPs) means the fewer VPs, the better, hence the preference for a do inserted under Io rather than under Vo . Compare (d) and (e). I assume that did is in Io in (e) and ( j), an assumption based on the fact that it precedes not, and I also assume that it has been inserted directly under Io , since otherwise its presence in Io would violate PrPr in ( j). This is how the optionality of that is achieved: with did inserted under Io , the two candidates with and without that have the same constraint pro®le; see (e) and ( j). If did had been inserted under Vo , the two candidates with and without that would not have had the same constraint pro®le because of PrPr; see (d) and (i). Insertion of do under Io violates VVo , just as insertion of have or be under Io does. In all three cases, the verb will neither be in Vo nor have a trace in Vo (see also section 14.7). In English, however, this violation of VVo has no consequences: the candidate that violates it, (e) in tableau T14.16, is still optimal, because the competing candidates all have more violations of higher-ranking constraints than (e). I also take HMC to be violated for Danish, French, and Icelandic when Nego intervenes in the verb chain. However, in these three languages, such a violation is preferable to a violation of VVo Ðthat is, a verb chain across Nego is preferable to

[VP t see Ob

[IP Su did not [VP t v

[IP Su did not

i. *

f j.

[VP t see Ob

[IP Su v

h. *

not [VP t did [VP t see Ob

Ob

[IP Su saw not [VP t v

g. *

not [VP t saw Ob

[VP t see Ob

[IP Su v

C [IP Su did not

f. *

f e.

[VP t see Ob

not [VP t did [VP t see Ob

d. *C [IP Su did not [VP t v

c. *C [IP Su v

Ob

not [VP t saw Ob

b. *C [IP Su saw not [VP t v

a. *C [IP Su v

Tableau T14.16 English, ex. (22) yFl

Case

*!

*!

PrPr

ObHd

*

**

**!

*

*

*

**!

**!

*

*

EcVP

*

*

*

*!

*

*

*!

*!

HMC

*

*

*

*

LxMv

*

*

VVo

*

*

*

*

PrBd

*

***

***

**

**

*

***

***

**

**

Stay

448 Sten Vikner

Vo -to-Io Movement and do-Insertion

449

do-insertion. This is due to the independently motivated higher ranking of VVo in these three languages than in English; see section 14.4. The HMC itself is ranked the same in all four languages, below EcVP and above LxMv. This analysis presupposes that something like the stray a½x ®lter (see Baker 1988:140) applies and is either very highly ranked or inviolableÐthat is, that the verbal a½x in Io has to be a½xed to a verb. (See (23) and tableau T14.17.) (23) Danish Jeg tror (at) skuespilleren ikke saÊ ®lmen. I think (that) actor-the not saw ®lm-the VVo (Verb in Vo ) punishes verb chains that do not include a Vo . In tableau T14.17, compare (e)Ðwhich was optimal in EnglishÐto (a), which lost in English because it violates HMC. LxMv (No Movement of a Lexical Head) punishes saÊ `saw' for leaving Vo . Compare (b) and (a). EcVP (Economy of VPs) means the fewer VPs, the better. Compare (c) and (a). PrPr is not relevant here. The version of the optimal candidate lacking at `that' does not violate it, since there is no movement into the highest Xo of an embedded clause. Candidates (a) and (f ) are thus both optimal. In French and Icelandic, the picture is only slightly di¨erent, due to the di¨erent ranking of PrBd, which also entails a di¨erence with respect to the complementizerless version of the optimal candidate; see (24) and tableau T14.18. (24) French Je crois *(que) l'acteur ne voit pas le ®lm. Icelandic EÂg tel *(aD) leikarinn sjaÂi ekki myndina. I think (that) the actor sees not the ®lm VVo (Verb in Vo ) punishes verb chains that do not include a Vo . Compare (e) in tableau T14.18 (which was optimal in English) to (b) (which lost in English because it violates HMC and LxMv). PrBd (Proper Binding) punishes voit/sjaÂi `sees' for staying in Vo , which forces Vo to-Io movement. Compare (a) and (b). EcVP (Economy of VPs) means the fewer VPs , the better. Compare (d) and (b). PrPr is relevant here. Any version of the optimal candidate lacking que=aD `that' would violate it, because there would be movement into the highest Xo of an embedded clause. Compare (g) and (b). Let me ®nally point out that I do not necessarily want to suggest that the negative elements pas, ikke, and ekki have to be Xo elements that c-command the VP. Rather, I only want to suggest that irrespective of the actual XP/Xo status of the lexical ele-

C [IP Su v

[VP t see Ob [VP t see Ob

[IP Su did not [VP t v

[IP Su did not

i. *

j. *

not [VP t did [VP t see Ob

[IP Su v

h. *

Ob

[IP Su saw not [VP t v

not [VP t saw Ob

g. *

[IP Su v

[VP t see Ob

e. *C [IP Su did not

f f.

[VP t see Ob

not [VP t did [VP t see Ob

d. *C [IP Su did not [VP t v

c. *C [IP Su v

Ob

not [VP t saw Ob

b. *C [IP Su saw not [VP t v

f a.

Tableau T14.17 Danish, ex. (23) yFl

Case

*!

*!

PrPr

ObHd

*!

*!

VVo

*

**

**!

*

*

*

**!

**!

*

*

EcVP

*

*

*

*

*

*

*

*

HMC

*

*

*

*!

LxMv

*

*

*

*

PrBd

*

***

***

**

**

*

***

***

**

**

Stay

450 Sten Vikner

[VP t see [VP t see

[IP Su v

[IP Su did not [VP t v

[IP Su did not

h. *

i. *

j. *

not [VP t did [VP t see

Ob

[IP Su saw not [VP t v

g. *

not [VP t saw Ob

[IP Su v

[VP t see

e. *C [IP Su did not

f. *

[VP t see

d. *C [IP Su did not [VP t v

not [VP t did [VP t see

Ob

not [VP t saw Ob

C [IP Su saw not [VP t v

c. *C [IP Su v

f b.

a. *C [IP Su v

Tableau T14.18 French/Icelandic, ex. (24) yFl

Case

*!

*!

PrPr

ObHd

*!

*!

VVo

*

**

**!

*

*

*

**!

**!

*

*

EcVP

*

*

*

*

*

*

*

*

HMC

*

*!

*

*!

PrBd

*

*

*

*

LxMv

*

***

***

**

**

*

***

***

**

**

Stay

Vo -to-Io Movement and do-Insertion 451

452

Sten Vikner

ment of negation, Nego (like SpecNegP) may block the formation of a verb chain across it in any of the four languages. It has often been assumed about NegP-spec that irrespective of whether SpecNegP is lexically ®lled by a negation, it cannot be part of an A-bar-chain. Note that English has negative islands even though not/n't is (or could be) an Xo (Cinque 1990:80; Rizzi 1990:15). The same can be said about Nego ; it cannot form part of the verb chain. Where exactly not, pas, ikke, and ekki actually are, in SpecNegP, Nego , or somewhere else, is a di¨erent question (for a discussion of the possibilities with respect to Italian and related languages, see Zanuttini 1997:23, 101). This means that the above analysis is compatible with Roberts's (forthcoming: note 12) suggestion that Xo can block XP-movements and vice versa as long as they have the same type of features (in this case operator features). 14.6

Negation and Insertion Under Io of Auxiliaries or of Empty do

To account for clauses with negation and auxiliaries, no further constraints and no further di¨erences in constraint ranking are necessary. It follows from the rankings already suggested that neither in English nor in the other three languages is do-insertion possible. Consider ®rst English; see (25) and tableau T14.19. (25) English I think (that) the actor has not seen the ®lm. Adding do serves no purpose here (following Grimshaw 1997:384), since have may itself be base generated in Io , so that the HMC is not violated anyway. This means that adding do will merely cost an extra VP, namely, the VP of non®nite have; compare (e) to (d) in tableau T14.19. Notice that this presupposes that auxiliaries like have in the tableau are part of the input, as mentioned in section 14.1. The complementizer is optional; any version of the optimal candidate without that would not violate PrPr, because there is no movement into the highest Xo of an embedded clause. Candidate (d) and a version of (d) without that would thus have the same (optimal) constraint pro®le. EcVP punishes (a) and (b) (optimal in Danish and in French/Icelandic respectively), which have more VPs than (d). Because have does not assign any thematic roles, base generating it in Io , as in (d), will not violate yFl, but only VVo . Next consider Danish, where there are no relevant changes compared to the negated ®nite main verb in tableau T14.17; see (26) and tableau T14.20. (26) Danish Jeg tror (at) skuespilleren ikke har set ®lmen. I think (that) actor-the not has seen ®lm-the

C [IP Su has not

C [IP Su v

[VP t seen Ob [VP t have [VP t seen Ob

e. *C [IP Su did not

not [VP t did [VP t have [VP t seen Ob

d. *C [IP Su has not

c. *C [IP Su v

[VP t seen Ob

not [VP t has [VP t seen Ob

b. *C [IP Su has not [VP t v

f a.

Tableau T14.20 Danish, ex. (26)

[VP t have [VP t seen Ob

[VP t seen Ob

not [VP t did [VP t have [VP t seen Ob

e. *C [IP Su did not

f d.

c. *C [IP Su v

[VP t seen Ob

not [VP t has [VP t seen Ob

b. *C [IP Su has not [VP t v

a. *C [IP Su v

Tableau T14.19 English, ex. (25)

yFl

yFl

Case

Case

PrPr

PrPr

ObHd

ObHd

*

***!

**

*

**

*!

*

**

*

HMC

EcVP

*!

VVo

*!

LxMv

*

*

**!*

*

**!

*

**!

VVo

*

*

**!

LxMv

*

HMC

EcVP

*

*

PrBd

*

*

PrBd

**

*

****

***

***

Stay

**

*

****

***

***

Stay

Vo -to-Io Movement and do-Insertion 453

454

Sten Vikner

VVo punishes (d) in tableau T14.20 (optimal in English) because har `has' has no trace in Vo , as opposed to (a). LxMv punishes (b) (optimal in French/Icelandic) because har `has' leaves Vo , as opposed to (a). EcVP punishes (c), which has more VPs than (a). The complementizer is optional, since any version of the optimal candidate lacking at `that' would not violate PrPr, because there is no movement into the highest Xo of an embedded clause. Candidate (a) and a version of (a) without at would have the same (optimal) constraint pro®le. Finally, consider French and Icelandic. Here there are also no relevant changes compared to the negated ®nite main verb in tableau T14.18; see (27) and tableau T14.21. (27) French Je crois *(que) l'acteur n' a pas vu le ®lm. Icelandic EÂg tel *(aD) leikarinn ha® ekki seÂD myndina. I think (that) the actor has not seen the ®lm VVo punishes (d) in tableau T14.21 (optimal in English) because a/ha® `has' has no trace in Vo , as opposed to (b). PrBd punishes (a) (optimal in Danish), because a/ha® `has' stays in Vo , as opposed to (b). EcVP punishes (c), which has more VPs than (b). The complementizer is obligatory. Any version of the optimal candidate without que=aD `that' would violate PrPr, since there would be movement into the highest Xo of an embedded clause. 14.7

Constraints Are Never Switched O¨

An important di¨erence between OT and the principles-and-parameters framework (Chomsky 1986; Rizzi 1990) is that nothing in OT directly corresponds to parameters (see also section 1.3 in chap. 1). In a certain sense, one might say that ranking V-Vo above EcVP corresponds to a particular setting of a particular parameter, namely, one that makes thematic and nonthematic verbs behave alike in Danish, French, and Icelandic. Correspondingly, one might expect that ranking VVo below EcVP corresponds to a di¨erent setting of the same parameter, namely, the one that forces a difference between thematic and nonthematic verbs in English. But whereas in principlesand-parameters terms, this might have meant switching o¨ the parameter that forces a di¨erence between thematic and nonthematic verbs, there is no switching o¨ constraints in OT: even in English where VVo is ranked below EcVP, VVo is still active. (See (28) and tableau T14.22) (28) English I think that the actor (actually) saw the ®lm.

PrPr

*

EcVP

*

ObHd

c. C [IP Su did A [VP t see Ob

PrPr

Case

*

Case

yFl

Ob

A [VP t saw Ob

b. C [IP Su saw A [VP t v

f a. C [IP Su v

yFl

[VP t have [VP t seen Ob

e. *C [IP Su did not

Tableau T14.22 English, ex. (28)

[VP t seen Ob

d. *C [IP Su has not

not [VP t did [VP t have [VP t seen Ob

[VP t seen Ob

not [VP t has [VP t seen Ob

C [IP Su has not [VP t v

c. *C [IP Su v

f b.

a. *C [IP Su v

Tableau T14.21 French/Icelandic, ex. (27)

HMC

ObHd

*!

LxMv

*!

VVo

*

PrBd

*

***!

**

*

**

*!

*

**

*

HMC

EcVP

*!

VVo

*

**

**

Stay

*

*!

PrBd

**

*

****

***

***

Stay

ˆ (10e)/(24b)

ˆ (10d)/(24a)

*

LxMv

Vo -to-Io Movement and do-Insertion 455

456

Sten Vikner

VVo must be ranked below EcVP to allow insertion directly under Io (of auxiliaries or dummy do); see for example (c) in tableau T14.11 or (e) in tableau T14.16. Nevertheless, VVo is still active, in that it serves to exclude (c) in tableau T14.22, because nothing is gained by inserting do in Io here. It would not mean using one less VP, nor would it prevent movement out of Vo , since (a) also has only one violation of EcVP and no violation of LxMv. In addition to punishing the insertion of verbs directly under Io , VVo thus also does the work here that Grimshaw's (1997:381) ``Full Interpretation'' (FI) does in her analysis, in that it punishes the use of dummy do (which she calls ``light verb'' do). Corresponding to Grimshaw's (1997:386±387) suggestion for FI, VVo may also have to be a gradient constraint, so that it would be violated to a lesser extent by dummy do than by dummy divulge or dummy domesticate. Dummy do, dummy divulge, and dummy domesticate would all violate VVo because (some of ) their lexical/categorial properties would be ignored if they were not inserted under Vo . However, the violation incurred by dummy do would be smaller than the violations incurred by other verbs, because do has fewer lexical/categorial properties than other verbs, and so when do is used as a dummy or light verb, fewer lexical/categorial properties have to be ignored (exactly how to quantify this is anything but straightforward). This view would also necessitate that do be seen as having fewer lexical/ categorial properties than even, for example, be and have, to ensure that the light verb used with negation (and in questions and other subject-auxiliary inversion contexts) is do rather than be or have. The reason for the well-formedness of a string corresponding to (c) in tableau T14.22 with emphatic/contrastive stress on do (i.e., . . . that the actor DID see the ®lm) is precisely that something is gained by inserting do here (by allowing a violation of VVo ), namely, that a violation of the HMC is avoided, under the assumption (see Trachtenberg 1996) that emphasis is an independent functional head between Io and Vo , much like Nego . (See (a) and (e) in tableau T14.16.) In Danish, for instance, where V-Vo is ranked higher than HMC, the analysis correctly predicts that emphasis, like negation, does not make any di¨erence for the position of the verb, (the example is . . . at skuespilleren virkelig SAÊ ®lmen, `. . . that the actor really SAW the ®lm'); see (a) and (e) in tableau T14.17. Finally, another shortcoming of a putative ``dummy do-parameter'' should be noted, namely, that it would lead us to expect that languages either do or do not have a dummy do. This is not the case. Although Danish, French, and Icelandic do not have do-insertion the way English does, they all do have a so-called verbum vicariumÐthat is, a verb that substitutes for other verbs under certain circumstances. Furthermore, these verbs are the straightforward translations of do: Icelandic gera, Danish gùre, and French faire. (See (29).)

Vo -to-Io Movement and do-Insertion

(29) Icelandic Haltu a Danish Hold paÊ English Hold French Tiens

457

blyÂantinum eins og eÂg

geri.

blyanten

ligesom jeg

gùr.

the pencil

as

do

le crayon

comme je le fais.

I

(it).

Exactly how to analyze these examples is not completely clear, however, since all of them would seem to involve optionality, in that ``full'' versions are also possibleÐfor example, Hold the pencil as I hold it. Presumably it would be possible to take the price of using do as a light verb (i.e., the price of ignoring some of its lexical/categorial properties) to be more than o¨set by the gain of not having to use a full VP-structure containing an object coreferent with an object in the matrix clause. 14.8

A Brief Sketch of the Analysis of Main Clauses

Until this point, only two di¨erences in constraint ranking have been assumed, namely, the two already shown in tableau T14.15. In fact, the only di¨erence compared to T14.23 is the introduction of HMC between EcVP and LxMv in all four languages. Unfortunately, space does not permit a close examination of main clauses, which is also the reason we will not see any syntactic di¨erences between French and Icelandic. Such a close examination (see Vikner 1998) would reveal two important di¨erences with respect to verb second: whereas all four languages have V2 main clauses when the initial element is a wh-element, only Danish and Icelandic have V2 in non-whinitial main clauses. Another di¨erence concerns subject-initial main clauses, where English never has V2 (Peter actually saw the ®lm), not even in questions (Who actually saw the ®lm?). To account for these di¨erent variations concerning V2, three more constraints would have to be introduced (following Grimshaw 1997 and Bakovic 1998): OPSP (operators in speci®er position), WHSP (wh-operators in speci®er position), and OPSC (operators in scope position). Operators here include wh-operators and topicalized constituents. A scope position is any position that c-commands the IP. V2 in main clause questions is driven by WhSp, which forces nonsubject whelements to move to SpecCP; see also the wh-criterion (Rizzi 1996:64; MuÈller 1997: 263). V2 in non-wh main clauses is driven by OpSp, which forces nonsubject operators (including both wh-operators and topicalized constituents) to move to SpecCP. WhSp would thus be ranked highly in all four languages, whereas OpSp is ranked highly in Danish and Icelandic but low in English and French.

yFl

yFl

yFl

English:

Danish:

French/Icelandic:

Tableau T14.23

Case

Case

Case

PrPr

PrPr

PrPr

ObHd

ObHd

ObHd

EcVP

EcVP

VVo

VVo

EcVP

HMC

HMC

HMC

PrBd

LxMv

LxMv

LxMv

VVo

PrBd

PrBd

Stay

Stay

Stay

458 Sten Vikner

Vo -to-Io Movement and do-Insertion

459

The movement of subjects to a position outside IP is forced in Danish, French, and Icelandic by the high ranking of OpSc, and avoided in English by a low ranking of the same constraint. The V2 verb movement itself is driven by a di¨erent constraint, ObHd, which also drives many other types of Xo -movement; see, for example, Grimshaw 1997 and Bakovic 1998. This is di¨erent from the wh-criterion, which drives both wh-movement and Vo -movement. Whether the verb undergoing V2 is do or a verb also present in the input depends on the ranking, so that only English (with LxMv g VVo ) will insert do in such contexts. In the other three languages (with VVo g LxMv), it is less expensive to move a verb out of VP than to insert a dummy verb. In accounting for V2 in OT by means of the two constraints WhSp and OpSp, it is possible to avoid the trap of binary parameters, whereby languages either have to have a particular property (e.g., V2) or not to have it. Languages may have a lot of V2Ð that is, ``real'' V2 (Danish/Icelandic)Ðor have just a little V2, namely, ``residual'' V2 (English/French). Furthermore, by having WhSp be a more restricted version of (i.e., a subset of ) OpSp, the (hopefully correct) prediction is made that there are no languages that have V2 in topicalizations but not in questions. A more complete picture would thus have four di¨erent variations among the languages (as shown in tableau T14.24): 1. The variation concerning PRBD accounts for the variation in Vo -to-Io movement. 2. The variation concerning VV o accounts for the variation concerning do-insertion and also for whether nonthematic verbs behave like thematic ones or not. 3. The variation concerning OPSP accounts for the variation in non-wh V2. 4. The variation concerning OPSC accounts for the variation in subject questions. 14.9

Conclusions

OT syntax allows the existence of constraints that are violated by the optimal candidate but still rule out other candidates. Examples are ECVP (an implementation of economy of projection), Stay (an implementation of economy of derivation), and VV o . OT syntax also allows the existence of constraints that are both violated by the optimal candidate in some language and nevertheless still rule out other candidates in another language. Examples are HMC, LXMV, and PRBD. This treatment of the HMC makes possible a uni®ed treatment of NegP across the languages, in particular of the blocking e¨ects of Nego and NegP-spec. Also made possible are principled accounts of the position of the ®nite verb in embedded clauses, of why do and other nonthematic verbs are inserted under Io in English, though not in the other languages, of why there is no do-insertion with

yFl

yFl

yFl

yFl

English:

Danish:

Icelandic:

French:

Tableau T14.24

Case

Case

Case

Case

PrPr

PrPr

PrPr

PrPr

OpSp

OpSp

WhSp

WhSp

WhSp

WhSp

OpSc

OpSc

OpSc

ObHd

ObHd

ObHd

ObHd

EcVP

EcVP

EcVP

VVo

VVo

VVo

EcVP

HMC

HMC

HMC

HMC

PrBd

PrBd

LxMv

LxMv

LxMv

LxMv

VVo OpSc

PrBd

PrBd

Stay

Stay

Stay

Stay

OpSp

OpSp

460 Sten Vikner

Vo -to-Io Movement and do-Insertion

461

auxiliaries in English, and of when the complementizer that/que/aD/at is optional and when it is obligatory. The interaction between the four crucial constraints is as follows. If VVo outranks EcVPÐthat is, if all verbs are inserted under Vo Ð there are two relevant possibilities: the result is French/Icelandic (all ®nite verbs undergo Vo -to-Io movement) if PrBd outranks LxMv, and it is Danish (no ®nite verbs undergo Vo -toIo movement) if LxMv outranks PrBd. If EcVP outranks VVo Ðthat is, if nonthematic verbs are inserted under Io and thematic verbs under Vo Ðno variation is found when the input contains a nonthematic verb, since this verb will occur in Io regardless of the ranking of the other constraints. Variation is found with EcVP outranking VVo when there is no nonthematic verb in the input. There are three relevant possibilities: The result is English if PrBd is outranked by the other three constraints (i.e., the ®nite thematic verb occurs in Vo ). The result would be a variant of English with Vo -to-Io movement of thematic verbs if LxMv were outranked by the other three constraints. Finally, the result would be another variant of English with do-insertion with thematic verbs in all contexts if VVo were outranked by the other three constraints. Because the facts discussed in this chapter (the position of ®nite main verbs and of ®nite auxiliary verbs in embedded clauses with and without negation in Danish, English, French, and Icelandic) may now be seen as connected and deriving from common sources, and because their derivation in earlier (principles-and-parameters) accounts (e.g., Vikner 1995) involved a number of stipulations no longer necessary in the present analysis, I hope to have shown that OT syntax allows us to take a signi®cant step forward in the analysis of verb movement. Note I would like to thank Eric BakovicÂ, Fabian Heck, Gunnar Hrafn Hrafnbjargarson, JoÂhannes GõÂsli JoÂnsson, Takeo Kurafuji, Edward Keer, Gereon MuÈller, Nicole Nelson, Alan Prince, Luigi Rizzi, Ian Roberts, Ramona RoÈmisch-Vikner, Vieri Samek-Lodovici, Tanja Schmid, Â rmann SigurDsson, Paul Smolensky, Melissa Trachtenberg, Carl Vikner, Ralf Vogel, HalldoÂr A Colin Wilson, and Heike Zinsmeister, and in particular Jane Grimshaw and GeÂraldine Legendre. Thanks also to participants in classes at the Universities of Stuttgart, Geneva, Iceland, Lund, and Rutgers University, and to audiences at talks given at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest (Baltimore), at the Minimal Link Condition Workshop (Potsdam), and at the Universities of Amsterdam, Cologne, Groningen, Leiden, Tilburg, Tromsù, TuÈbingen, Utrecht, Venice, Johns Hopkins University, the Netherlands Institute for Advanced Study (Wassenaar), and the Arbeitsgruppe Strukturelle Grammatik (Max-Planck-Gesellschaft, Berlin). The research reported here was supported by the Deutsche Forschungsgemeinschaft as part of the project The Optimality-Theoretic Syntax of German from a Comparative Germanic Perspective (Grant MU 1444/2-1, Principal Investigators: G. MuÈller, S. Vikner).

462

Sten Vikner

References Alexiadou, Artemis. 1997. Adverb Placement. Amsterdam: Benjamins. Baker, Mark. 1988. Incorporation. Chicago: University of Chicago Press. BakovicÂ, Eric. 1998. Optimality and Inversion in Spanish. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 35±58. Cambridge, Mass.: MIT Press. Bobaljik, Jonathan. 1995. Morphosyntax: The Syntax of Verbal In¯ection. Doctoral dissertation, MIT, Cambridge, Mass. Chomsky, Noam. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, Noam. 1986. Barriers. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1991. Some Notes on Economy of Derivation and Representation. In Robert Freidin (ed.), Principles and Parameters in Comparative Grammar, 417±454. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Cinque, Guglielmo. 1990. Types of A-Bar-Dependencies. Cambridge, Mass.: MIT Press. Cinque, Guglielmo. 1999. Adverbs and Functional Heads: A Cross-Linguistic Perspective. New York: Oxford University Press. Deprez, Viviane. 1994. A Minimal Account of the That-t E¨ect. In Guglielmo Cinque, Jan Koster, Jean-Yves Pollock, Luigi Rizzi, and Ra¨aella Zanuttini (eds.), Paths Towards Universal Grammar: Studies in Honor of Richard Kayne, 121±135. Washington, D.C.: Georgetown University Press. Emonds, Joseph. 1978. The Verbal Complex of V 0 -V in French. Linguistic Inquiry 9: 151±175. Emonds, Joseph. 1994. Two Principles of Economy. In Guglielmo Cinque, Jan Koster, JeanYves Pollock, Luigi Rizzi, and Ra¨aella Zanuttini (eds.), Paths Towards Universal Grammar: Studies in Honor of Richard Kayne, 155±172. Washington, D.C.: Georgetown University Press. Fiengo, Robert. 1977. On Trace Theory. Linguistic Inquiry 8(1): 35±61. Grimshaw, Jane. 1991. Extended Projection. Ms., Brandeis University, Waltham, Mass. Grimshaw, Jane. 1997. Projection, Heads, and Optimality. Linguistic Inquiry 28(3): 373±422. Haegeman, Liliane. 1994. Introduction to Government and Binding Theory. 2nd ed. Oxford: Blackwell. Haider, Hubert. 1988. Matching Projections. In Anna Cardinaletti, Guglielmo Cinque, and Giuliana Giusti (eds.), Constituent Structure: Papers from the 1987 GLOW Conference, 101± 121. Dordrecht: Foris. Holmberg, Anders. 1997. The True Nature of Holmberg's Generalization. In Kiyomi Kusumoto (ed.), NELS 27, 203±217. GLSA, University of Massachusetts, Amherst. Holmberg, Anders, and Christer Platzack. 1995. The Role of In¯ection in Scandinavian Syntax. New York: Oxford University Press. Koopman, Hilda, and Dominique Sportiche. 1991. The Position of Subjects. Lingua 85: 211± 258. Legendre, GeÂraldine, Paul Smolensky, and Colin Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha

Vo -to-Io Movement and do-Insertion

463

McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Legendre, GeÂraldine, Colin Wilson, Paul Smolensky, Kristin Homer, and William Raymond. 1995. Optimality and Wh-Extraction. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 607± 636. GLSA, University of Massachusetts, Amherst. McCloskey, James. 1992. Adjunction, Selection, and Embedded Verb Second. Linguistics Research Report LRC-92-07. University of California, Santa Cruz. MuÈller, Gereon. 1997. Partial Wh-Movement and Optimality Theory. Linguistic Review 14: 249±306. MuÈller, Gereon. 1999. Optionality in Optimality-Theoretic Syntax. Glot International, 4(5): 3±8. Parsons, Terence. 1995. Thematic Relations and Arguments. Linguistic Inquiry 26(4): 635±662. Pollock, Jean-Yves. 1989. Verb Movement, Universal Grammar, and the Structure of IP. Linguistic Inquiry 20(3): 365±424. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Quirk, Randolph, Sidney Greenbaum, Geo¨rey Leech, and Jan Svartvik. 1985. A Comprehensive Grammar of the English Language. London: Longman. Rizzi, Luigi. 1990. Relativized Minimality. Cambridge, Mass.: MIT Press. Rizzi, Luigi. 1996. Residual Verb Second and the Wh-Criterion. In Adriana Belletti and Luigi Rizzi (eds.), Parameters and Functional Heads, 63±90. New York: Oxford University Press. Rizzi, Luigi, and Ian Roberts. 1996. Complex Inversion in French. In Adriana Belletti and Luigi Rizzi (eds.), Parameters and Functional Heads, 91±116. New York: Oxford University Press. Roberts, Ian. 1985. Agreement Parameters and the Development of English Modal Auxiliaries. Natural Language and Linguistic Theory 3(1): 21±58. Roberts, Ian. 1993. Verbs and Diachronic Syntax. Dordrecht: Kluwer. Roberts, Ian. Forthcoming. Head Movement. In Mark Baltin and Chris Collins (eds.), Handbook of Syntactic Theory. Oxford: Blackwell. Samek-Lodovici, Vieri. 1996. Constraints on Subjects: An Optimality Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Scholten, Christina. 1988. Principles of Universal Grammar and the Auxiliary Verb Phenomenon. Doctoral dissertation, UCLA, Los Angeles. Trachtenberg, Melissa. 1996. Emphatic A½rmation in English and Danish. Ms., Rutgers University, New Brunswick, N.J. Travis, Lisa. 1984. Parameters and E¨ects of Word Order Variation. Doctoral dissertation, MIT, Cambridge, Mass.

464

Sten Vikner

Vikner, Sten. 1995. Verb Movement and Expletive Subjects in the Germanic Languages. New York: Oxford University Press. Vikner, Sten. 1997a. The Interpretation of Object Shift, Optimality Theory, and Minimalism. In Working Papers in Scandinavian Syntax 60, 1±24. University of Lund. Vikner, Sten. 1997b. Vo -to-Io Movement and In¯ection for Person in All Tenses. In Liliane Haegeman (ed.), The New Comparative Syntax, 189±213. London: Longman. Vikner, Sten. 1998. V2, Vo -to-Io Movement, do-Insertion, and Negation in Optimality Theory. Ms., University of Stuttgart. Williams, Edwin. 1994. Thematic Structure in Syntax. Cambridge, Mass.: MIT Press. Zanuttini, Ra¨aella. 1997. Negation and Clausal Structure. New York: Oxford University Press.

Chapter 15 Bidirectional Optimization and the Theory of Anaphora

Colin Wilson

Research in Optimality Theory (OT; Prince and Smolensky 1993) and the Minimalist Program (Chomsky 1995) claims that the status of a sentence with respect to a grammar is determined by competition. Informally, the claim holds that a sentence is well formed if and only if it is better than (or at least as good as) all the alternatives. In this chapter, I bring evidence from anaphora to bear on a fundamental question raised by this work: What properties must one sentence have in order to qualify as an ``alternative'' to another? More simply, Which sentences compete?1 Building on Burzio (1989, 1992, 1998), I argue that the theory of anaphora must incorporate at least two de®nitions of ``alternative'' sentence, and hence at least two types of competition. In this section, I ®rst introduce the two competitions and the binding phenomena used to motivate them. I then sketch the main contribution of this chapter: a solution to the problem of combining the two competitions into a coherent and empirically adequate theory of the interface between syntax and semantics. This section ends with an outline of the rest of the chapter. Certain minimality conditions on anaphor binding, illustrated by the Marathi (Indo-Aryan) sentences in (1), motivate an interpretive competition: a competition among sentences that share a common syntactic structure, but di¨er with respect to the semantic interpretation of that structure. (1) Relativized minimality in Marathi (data from Dalrymple 1993:19±20; see also Wali 1976) a. Tomi mhanat hota [ ki Suej ni swataahlaa  i=j maarle ]. Tom said that Sue ERG anaphor-ACC hit `Tom said that Sue hit him.' b. Janei mhanaali [ ki [ swataacii parikshaa ] sampli ]. Jane said that anaphor-GEN test ®nished `Jane said that her test was over.' According to Dalrymple (1993), the Marathi anaphor swataah must be bound within the minimal ®nite clause that contains it, as in (1a), except when that clause does not

466

Colin Wilson

contain a potential antecedent for the anaphor, as in (1b).2 That is, in the terminology adopted in this chapter, the domain in which swataah must have an antecedent is not absolute, or rigid, but relativized to the presence of a potential antecedent; thus (1) illustrates the anaphora-speci®c case of relativized minimality (Rizzi 1990). Relativized minimality e¨ects emerge from interpretive competition in essentially the following way. A sentence of the form [Ai . . . [d Bj . . . anaphori . . . ]], in which a binding relation crosses a potential antecedent (B), loses to the alternative [Ai . . . [d Bj . . . anaphorj . . . ]], which better satis®es a rigid locality constraint on binding that is sensitive to domains of type d. But a sentence of the form [Ai . . . [d . . . anaphori . . . ]] does not face such an alternative (because there is no potential antecedent for the anaphor within d), and therefore survives the competition despite the fact that it violates the aforementioned locality constraint. Thus in (1a) the interpretation in which the anaphor is bound by Tom loses to the interpretation in which it is bound by Sue. But the syntactic structure of (1b) does not have an alternative interpretation in which the anaphor is bound within its minimal ®nite clause; therefore binding by Jane is well formed. (In (1b), swataah appears within the embedded subject, which is therefore not a potential antecedent for it, by the prohibition on circular referential dependencies, or ``i-within-i'' condition (Chomsky 1986).) These examples reveal that the grammar of binding is in some cases sensitive to whether an anaphor could be bound within a certain syntactic domain. This basic observation underlies many existing theories of binding (e.g., Burzio 1992; Chomsky 1986; Huang 1983; Pollard and Sag 1992).3 Interpretive competition takes the interface between syntax and semantics to have a particular ``direction'': as in the principles-and-parameters framework and the Minimalist Program, the interface is taken to be the locus of mapping from syntactic structures to semantic structures, not vice versa. But there is another logical ``direction'' of the interfaceÐthat is, the interface can be viewed as the locus of mapping from semantic structures to syntactic ones, a perspective similar to that adopted in Burzio 1998, Legendre, Smolensky, and Wilson 1998, and Samek-Lodovici 1996, among other sources. This alternative direction leads to another kind of competition, here called expressive: a competition among sentences that share a semantic structure (or intended meaning), but di¨er with respect to how that structure is syntactically expressed. In this chapter, I motivate expressive competition with economy conditions of the type illustrated in (2). (2) Referential economy in Icelandic (data from Maling 1984:212, 1986:284; see also ThraÂinsson 1979, 1997) a. Haralduri skipaDi meÂr aD raka *hanni /sigi . Harold ordered me to shave(in®nitive) him/anaphor `Harold ordered me to shave him.'

Bidirectional Optimization and the Theory of Anaphora

467

b. JoÂni veit aD MarõÂa elskar hanni /*sigi . Jon knows that Maria loves(indicative) him/anaphor `Jon knows that Maria loves him.' The examples above are representative of the (partial) complementarity of Icelandic third-person pronouns (here, hann `he') and the anaphor sig. When a binding relation is su½ciently local, as in (2a), the bound element must be realized as an anaphor, not as a pronoun.4 But when the binding relation is nonlocal, as in (2b), the anaphor is excluded and the bound element must be a pronoun. Such distributions can be accounted for with expressive competition basically as follows. As a consequence of the lexical distinction between (certain) anaphors and pronouns (see section 15.2), the former are preferred to the latter by a referential economy constraint. Therefore, a sentence of the form [Ai . . . pronouni . . . ] loses the expressive competition to a sentence of the form [Ai . . . anaphori . . . ] except when the binding relation in [Ai . . . anaphori . . . ] is excluded by a minimality condition (rigid and/or relativized). In other words, referential economy is satis®ed when minimality permits, but violated by grammatical sentences when the intended antecedent is too distant to license an anaphor. Returning brie¯y to the Icelandic examples above, the ungrammaticality of the bound pronoun hann in (2a) follows from referential economy, given that the anaphor sig need not have an antecedent within the minimal in®nitival clause that contains it (i.e., given that Haraldur is su½ciently local to license the anaphor). In contrast, the bound pronoun in (2b) is grammatical, despite violating referential economy, because sig must generally ®nd an antecedent within the minimal ®nite clause that contains it (i.e., given that JoÂn is not su½ciently local to license the anaphor). Given that interpretive and expressive competition are both well motivated, the question immediately arises of how they can be combined into a coherent theory of the syntax-semantics interface. An attractive possibility is that the two can be collapsed into a single competition among sentences that di¨er semantically and syntactically. But the main claim of this chapter is that interpretive and expressive competition in fact must apply separately, and in the order that they have been introduced. This claim is embodied in the following bidirectional model of the interface, in which interpretive and expressive competition are formalized as OT optimizations. (In table 15.1, I is the set of sentences, or hsyntactic structure, semantic interpretationi pairs, allowed by universal grammar, and I 0 is the set of outputs of interpretive optimization.) The rest of the chapter builds up the bidirectional optimization model according to the following outline. Section 15.1 formalizes the account of relativized minimality e¨ects on anaphor binding that was introduced above. This section demonstrates that such e¨ects can be derived from a rigid minimality constraint operating under

468

Colin Wilson

Table 15.1 Bidirectional optimization model Optimization

Input

Output

interpretive

Given any pair hS, Mi in I

. . . ®nd the most harmonic pair hS, Ni in IÐthat is, maximize the harmony of the semantic interpretation of S.

expressive

Given any pair hS, Mi in I 0

. . . ®nd the most harmonic pair hT, Mi in I 0 Ðthat is, maximize the harmony of the syntactic expression of M.

interpretive optimization. Section 15.2 employs expressive optimization to formalize the account of referential economy e¨ects. It then goes on to show that rigid minimality e¨ects on anaphor binding can be accounted for with a particular ranking of the referential economy constraint and the minimality constraint of section 15.1. Section 15.3 considers a model of the syntax-semantics interface that con¯ates interpretive and expressive optimizationÐthat is, a model in which alternative sentences are allowed to di¨er simultaneously with respect to their syntactic and semantic components. This model makes a prediction about anaphoric systems that seems generally incorrect and that is speci®cally discon®rmed by the anaphoric system of Chamorro (Western Austronesian; Chung 1981, 1989). Formal analysis of this problem leads to the bidirectional optimization model in table 15.1, which is introduced and applied to Chamorro in section 15.4. Section 15.5 concludes the chapter with some remarks on proposals related to the bidirectional model and on alternative answers to the question of what sentences compete. 15.1

Relativized Minimality and Interpretive Optimization

Every OT grammar maps a universal set of inputs to a set of outputs. The set of inputs generally contains more representational distinctions (or contrasts) than are observed in any given language. That is, the input-output mapping is often neutralizing: two (or more) inputs are mapped to the same output, thus eliminating a universally possible contrast from the language under consideration. In this chapter, I account for relativized minimality e¨ects on anaphor binding with the neutralization shown abstractly in table 15.2. For this neutralization to be possible, a candidate output must be able to diverge from the input with respect to which potential antecedent (here, A or B) actually binds a given anaphor. This is the nontrivial aspect of the mapping from table 15.2(b) to the output shown. For the purposes of this chapter, I take binding to be a semantic relation, one regulated by syntactic structure but not directly encoded in it (see Jackendo¨ 1990). Consequently, the neutralization

Bidirectional Optimization and the Theory of Anaphora

469

Table 15.2 Neutralization for relativized minimality Input

Output

a. [Ai ... [d Bj ... anaphorj ... ]] ƒƒƒƒƒ! [Ai ... [d Bj ... anaphorj ... ]] ! b. [Ai ... [d Bj ... anaphori ... ]] ƒƒƒƒƒ Table 15.3 Interpretive optimization Input

Output

Given any pair hS, Mi in I

. . . ®nd the most harmonic pair hS, Ni in IÐthat is, maximize the harmony of the semantic interpretation of S.

above falls within the space of mappings allowed by (what I call) interpretive optimization: an optimization over candidate outputs that share the syntactic structure, but not necessarily the semantic structure, of the input. Representing inputs and candidate outputs as pairs consisting of a syntactic component and a semantic component (in that order), we have the formal de®nition of interpretive optimization given in table 15.3. As in the introductory section of the chapter, I stands for the inventory of hsyntax, semanticsi pairs allowed by universal grammar; this is the set of universal inputs, as well as the set from which candidate outputs are drawn. Given the limited scope of this chapter, I do not attempt to de®ne the inventory I. However, I do assume that binding relations in this universal inventory are restricted only by a c-command condition (or some other more adequate con®gurational condition; see Reinhart 1983). Throughout, I simplify the candidate sets for interpretive optimization by considering only candidate outputs that are either identical to the input or that diverge from it just with respect to binding relations. For concreteness, binding is taken to be represented by quanti®er-variable structures (as in Reinhart 1983; see Fiengo and May 1994), although nothing in the analysis hinges on this assumption. Having established a framework in which the neutralization in table 15.2 is possible, I now turn to the question of why the distinction between inputs like 15.2(a) and 15.2(b) should ever be lost, and of exactly how this relates to the relativized minimality e¨ect illustrated above. The answer comes in two parts. First, as many have argued, certain anaphors have a lexical de®ciency that makes them unable to refer independently. I adopt the particular characterization of this de®ciency proposed by Bouchard (1984), Burzio (1991:86¨.), and Reinhart and Reuland (1993:658±659), according to which the relevant anaphors lack lexical (or

470

Colin Wilson

``inherent'') speci®cation for the j-features (person, gender, number, Case). And I follow these researchers, among many others, in assuming that the underspeci®cation of an anaphor causes it to actively ``seek out'' an antecedent (see also Chomsky 1981:211¨.). Second, this search for an antecedent is guided by the following two violable constraints, along with a few others to be introduced later. (3) Interpretive constraints a. Local Antecedent (abbreviated LocalAnt[d]) If a syntactic domain of type d contains an anaphor a, then it also contains an antecedent for a. b. Binding Faithfulnesss (abbreviated BindFaith) If a binding relation holds between two elements in the input, then it also holds in the output. The Markedness constraint in (3a) embodies rigid minimality: it requires an anaphor to be bound within a syntactic domain of type d even if that domain does not contain a potential antecedent for the anaphor. According to a very extensive literature (e.g., Anderson 1986; Burzio 1992; Harbert 1982; George and Korn®lt 1981; Giorgi 1984; Manzini and Wexler 1987; Progovac 1994; Reinhart and Reuland 1993), the relevant type of domain is de®ned by the existence of certain in¯ectional features (e.g., tense, mood, agreement). To a ®rst approximation, a syntactic domain d is opaque to anaphor binding in direct proportion to the morphophonological richness and semantic independence of the in¯ectional features that it hosts. Rather than trying to account for this relationship between in¯ectional content and opacity for binding (see the works cited), I simply take it as given, and call on particular instantiations of (3a) as needed. The Faithfulness constraint in (3b) requires an element bound in the input to have the same antecedent in the output, and therefore con¯icts with (3a) when an anaphor's input antecedent lies on the other side of the boundary of some d domain. Because the focus of this part of the chapter is on minimality e¨ects, (3b) will almost always be dominated by (3a), and will thereby be rendered inactive. But, from a broader perspective, Faithfulness to binding is necessary to account for cases in which a given anaphor can take multiple antecedents, some of which are more local than others. 15.1.1 Relativized Minimality Applying the theory developed above to the Marathi contrast in (1), I ®rst invoke the version of (3a) that is sensitive to ®niteness: if a ®nite clause contains an anaphor, then it must also an antecedent for that anaphor. Ranking this Markedness constraint, call it LocalAnt[‡®n], above BindFaith yields a neutralization of the form

Bidirectional Optimization and the Theory of Anaphora

471

Tableau T15.1 Nonlocal binding mapped to local binding Input: Tomi mhanat hota [ ki Suej ni swataahlaai maarle ] T. said that S. ERG anaphor hit `Tom said that Sue hit him (ˆ Tom)' Candidate Outputs

LocalAnt[‡®n]

f a. Tomi . . . [ Suej . . . swataahlaaj . . . ] `Tom said that Sue hit herself ' b. Tomi . . . [ Suej . . . swataahlaai . . . ] `Tom said that Sue hit him (ˆ Tom)'

BindFaith

* *!

Note: Given the representational assumptions adopted above, the optimal candidate in this tableau could be more explicitly represented as the pair h[Tom mhanat hota [ki Sue ni swataahlaa maarle]], lx[say(Tom, hit(x,x))](Sue)i. In this tableau, as in most tableaux in the chapter, only the constraint violations that di¨erentiate the candidates considered are shown. in table 15.2 when the anaphor in question has a potential antecedent within the minimal ®nite clause that contains it. Even if the input designates a ``target'' antecedent that lies outside this clause, the anaphor is bound by the more local potential antecedent in the optimal output. The ranking given in tableau T15.1 does not tolerate violation of LocalAnt[‡®n] when there is a candidate output that satis®es the constraint. But now consider an input in which an anaphor does not have a potential antecedent within the minimal ®nite clause that contains it. Given that the lexical underspeci®cation of an anaphor drives it to ®nd a syntactic antecedent, as assumed above, and given the fundamental claim of OTÐthat grammars are hierarchies of violable constraintsÐwe expect that violation of the Markedness constraint will be grammatical under such circumstances. And, of course, this is exactly what we observe for syntactic structures like that of (1b), which has the interpretation predicted by tableau T15.2. (For now, we ignore the shaded candidate, in which the anaphor swataah has been unfaithfully replaced by the form aapan.5 This type of unfaithfulness is not allowed by the de®nition of interpretive optimization, but is allowed under expressive optimization, as discussed in section 15.2; see especially section 15.2.2.) Generalizing the results of these two tableaux, candidate outputs in which the anaphor swataah is bound from outside the minimal ®nite clause that contains it will

472

Colin Wilson

Tableau T15.2 Nonlocal binding optimal as a last resort Input: Janei mhanaali [ ki [ swataacii parikshaa ] sampli ] J. said that anaphor-GEN test ®nished `Jane said that her (ˆ Jane's) test was over' Candidate Outputs f a. Janei . . . [ . . . swataacii . . . ]

LocalAnt[‡®n]

BindFaith

*

b. Janei . . . [ . . . aapani . . . ] always be nonoptimal, regardless of what antecedent the anaphor has in the input, except when the syntactic component of the input does not provide a potential antecedent (i.e., an argument that c-commands the anaphor) within that domain. Thus the system of interpretive optimization captures the relativized minimality e¨ect, and has the clear advantage of doing so without positing a relativized minimality constraint. Nothing in the grammar refers to the presence or absence of a ``potential'' (or ``accessible'') antecedent. Rather, the fact that the grammaticality of long-distance binding depends on how locally the anaphor could be bound follows from the competition among interpretationally distinct candidates. As discussed earlier, the conceptual di¨erence between the interpretive optimization introduced in this section and the optimization to be introduced in the next section lies in the ``direction'' assigned to the syntax-semantics interface. Here we have taken the direction to be from syntax to semantics, with the role of optimization being to ®nd the best interpretation of a syntactic structure (given an input, or target, interpretation). In section 15.2, we reverse direction, using optimization to ®nd the best syntactic ``expression'' of a semantic structure (again, given an input/target expression). Controlling the interaction between these two directions of the interface then forms the project for the rest of the chapter. 15.2

Referential Economy and Expressive Optimization

I begin this section by brie¯y reviewing some evidence given by Burzio (1989) for an economy constraint on the speci®cation of bound elements (see also Richards 1997). Recall the contrast between (2a), in which the matrix subject (JoÂn) can grammatically bind an anaphor (sig) but not a pronoun (hann), and (2b), which illustrates exactly the opposite pattern. As already discussed in section 15.1.1, I follow many researchers in assuming that certain anaphors (henceforth referred to as ``SE'' anaphors) have impoverished lexical entries relative to pronouns (let alone lexical nouns). Speci®-

Bidirectional Optimization and the Theory of Anaphora

473

cally, I assume that SE anaphors lack any lexical speci®cation for the j-featuresÐ features that all pronouns are speci®ed for, at least to some degree, and that an argument must have in order to independently establish reference to some entity in the domain of discourse. Given these representational assumptions, contrasts like the one in (2) follow from a particular interaction between minimality (rigid or relativized) and a constraint that I call referential economy, which requires arguments to be lexically devoid of j-features (i.e., essentially, to be SE anaphors). When minimality allows a given bound-variable interpretation to be expressed with a SE anaphor, as in (2a), referential economy can render a pronoun ungrammatical under the same interpretation. But when minimality precludes the anaphor, as in (2b), referential economy yields, and a bound pronoun is grammatical. A theory of anaphora that did not incorporate referential economy could not relate the grammaticality of a bound pronoun to the availability of an anaphor in this case or any of the others discussed by Burzio (1989) and Richards (1997). Because SE anaphors must (barring the alternative interpretive mechanisms discussed in section 15.1.2) be bound, in the technical sense, we correctly predict that (2b) is grammatical when coreference without binding is intended (ThraÂinsson 1979). Another example of the referential economy e¨ect is provided by Vikner (1985), who explicitly addresses the bound-variable versus coreference distinction (see also Anderson 1986; Reninhart 1983; Richards 1997; ThraÂinsson 1976). (I have provided the word-by-word glosses in the following examples; the translations are Vikner's.) (4) Referential economy in Danish (Vikner 1985: note 10, based on examples in ThraÂinsson 1976:237) a. Kun Peteri bad Susanj om [ PROj at elske sigi ]. only Peter asked Susan about to love anaphor `Only one person asked Susan to love him, and that was Peter.' b. Kun Peteri bad Susanj om [ PROj at elske hamk ]. only Peter asked Susan about to love him `Only one person asked Susan to love Peter, and that was Peter.' When a binding relation is su½ciently local (e.g., as here, when it crosses only the boundary of an in®nitival clause), the SE anaphor sig must be used, to the exclusion of a pronoun. But when the intended interpretation of the same structure involves conference without binding, as in (4b), the pronoun is grammatical. 15.2.1 Expressive Optimization Languages vary with respect to the structural contexts in which they disallow boundvariable interpretations of pronouns (e.g., compare the Icelandic and Danish sentences discussed above with English sentences like Everyonei asked Mary to help himi , in which binding is perfectly grammatical). In OT, this surface variation indicates

474

Colin Wilson

Table 15.4 Neutralization for referential economy Input

Output

a. [Ai ... anaphori ... ] ƒƒƒƒƒ! [Ai ... anaphori ... ] ! b. [Ai ... pronouni ... ] ƒƒƒƒƒ Table 15.5 Expressive optimization Input

Output

Given any pair hS, Mi in I

. . . ®nd the most harmonic pair hT, Mi in IÐthat is, maximize the harmony of the syntactic expression of M.

that the set of universally possible sentences containing bound pronouns is relatively unrestricted (or ``rich''). Taking it to be quite rich, I assume that only a c-command condition restricts pronoun binding in the set of universal inputs (which, as in section 15.1, I assume to be identical to the set of universal candidates). It follows from this assumption that the ungrammaticality of sentences like (2a) must be accounted for by unfaithfuness in the input-output mapping. In particular, I propose the abstract neutralization shown in table 15.4 to account for referential economy e¨ects. The nontrivial part of table 15.4, of course, is the mapping from input (b), which contains a bound pronoun, to the output shown, in which the same binding relation terminates in an anaphor. As discussed in section 15.1, interpretive optimization absolutely prohibits breaches of faithfulness to the syntactic (including morphosyntactic) component of the input. Therefore, we need another type of optimization to account for referential economy e¨ects. For the moment, we simply assume that this second optimization, formalized in table 15.5, must be kept separate from the ®rst; this point is argued at length in section 15.3. This formulation of expressive optimization incorporates two provisional assumptions. First, candidate outputs have been required to share the semantic component of the input. Thus table 15.5 is the mirror image of interpretive optimization, an assumption that will be supported in section 15.4. Second, the universal inventory of inputs and candidates for expressive optimization has been identi®ed with the inventory of inputs/candidates for interpretive optimization (i.e., I). The consequences of this assumption will be examined in detail in section 15.3, leading to its ultimate rejection in section 15.4. Continuing to adopt both assumptions for present purposes, I place referential economy (5a) in con¯ict with a Faithfulness constraint on feature speci®cation (5b).

Bidirectional Optimization and the Theory of Anaphora

475

Tableau T15.3 Bound pronoun mapped to bound anaphor Input: Haralduri skipaDi meÂr aD raka hanni Harold ordered me to shave(in®nitive) him `Harold ordered me to shave him (ˆ Harold)' Candidate Outputs

RefEcon

f a. Haralduri skipaDi meÂr aD raka sigi b. Haralduri skipaDi meÂr aD raka hanni

FtrFaith *

*!

(5) Expressive constraints a. Referential Economy (abbreviated RefEcon) An argument does not have any lexical j-feature speci®cation. b. Feature Faithfulness (abbreviated FtrFaith) An argument speci®ed for feature [F] in the input is speci®ed for [F] in the output. Ranking RefEcon above FtrFaith gives rise to the unfaithful mapping of table 15.4 when anaphor binding satis®es the relevant minimality condition, as illustrated in tableau T15.3, based on the Icelandic structure in (2a). (The interaction between economy and minimality deserves thorough analysis, which I take up momentarily.) Note that the same output would be optimal for an input exactly like the one above except that the pronoun is replaced by an anaphor; thus we have a concrete example of the schematic neutralization in table 15.4. For future reference, I dub RefEcon g FtrFaith the ``ranking schema'' for referential economy e¨ects. Note that the opposite ranking would seek to preserve the input contrast between a bound pronoun and a bound anaphor. This suggests an account for reported cases of overlap in the distribution of bound pronouns and anaphors (Hestvik 1990; Huang 1983), although further empirical investigation is required to establish that such cases actually involve homophony. More important for the facts at hand, note that it is possible for a bound pronoun to be optimal even when RefEcon outranks FtrFaith. If a minimality condition removes the structure with an anaphor from the set of candidates evaluated by the expressive constraints (see the shaded candidate in tableau T15.4), then surface-violation of RefEcon becomes unavoidable. This is illustrated in tableau T15.4, which is based on (2b). The discussion so far has left open the major question of exactly how candidates like (b) in tableau T15.4 are rendered nonoptimal under expressive optimization. Clearly, ranking RefEcon below the appropriate instantiation of LocalAnt[d] (here,

476

Colin Wilson

Tableau T15.4 Bound pronoun optimal as a last resort Input: JoÂni veit aD MarõÂa elskar hanni Jon knows that M. loves(indicative) him `Jon knows that MarõÂa loves him' Candidate Outputs f a. JoÂni veit aD MarõÂa elskar hanni

RefEcon

FtrFaith

*

b. JoÂni veit aD MarõÂa elskar sigi LocalAnt[indicative]) will settle the issue. But allowing such rankings requires taking a conceptual step that should be made explicit: the interpretive constraints (3) and expressive constraints (5), originally proposed for di¨erent optimizations, must be placed in the same hierarchy. In other words, a particular grammar must consist of a single total ordering of all the constraints (even if, to borrow a term from Smolensky 1996, that ordering is ``used'' di¨erently by each optimization). In the following subsection, I show that this single-hierarchy assumption leads to a novel uni®cation of rigid and relativized minimality e¨ects. 15.2.2 Minimality E¨ects and the Locus of Competition The constraints RefEcon and FtrFaith cannot a¨ect the outcome of the interpretive optimization in table 15.3, because for that optimization candidate outputs do not di¨er with respect to their (morpho)syntactic components. But, as just noted, LocalAnt[d] constraints can in¯uence expressive optimization: such constraints target anaphors, not pronouns, and therefore prefer the latter when evaluating a binding relation that crosses any signi®cant syntactic boundary (e.g., the boundary of an indicative clause). More speci®cally, given an input with a binding relation that crosses the boundary of a domain of type d, LocalAnt[d] con¯icts with RefEcon and, if the bound element in the input is an anaphor, with FtrFaith. For an example of this con¯ict, we return brie¯y to the analysis of relativized minimality in Marathi (section 15.1). For the analysis to survive expressive optimization, LocalAnt[‡®n] must be outranked by at least one of RefEcon and FtrFaith; both of these constraints have been inserted at the top of the hierarchy in tableau T15.5, which is otherwise identical to tableau T15.2. Here I follow Wali and Subbarao (1991) and analyze Marathi aapan as a pronoun (see Dalrymple 1993). In point of fact, this form exhibits properties of both pronouns (ability to take split antecedents, no locality condition on binding) and anaphors (requires an antecedent within the sentence).

Bidirectional Optimization and the Theory of Anaphora

477

Tableau T15.5 Relativized minimality in Marathi Input: Janei mhanaali [ ki [ swataacii parikshaa ] sampli ] J. said that anaphor-GEN test ®nished `Jane said that her(ˆ Jane's) test was over' Candidate Outputs

RefEcon

FtrFaith

f a. Janei . . . [ . . . swataacii . . .] b. Janei . . . [ . . . aapan i . . .]

LocalAnt[‡®n]

BindFaith

* *!

*!

This mixed character strongly suggests that the anaphor-pronoun dichotomy assumed here should be replaced by a more ®ne-grained scale of lexical speci®cation (hence, in this theory, of referentiality; see Burzio 1991 and Richards 1997 for relevant proposals). All that is necessary for the present account is that aapan lies closer on the scale to full pronouns than swataah does. Whichever of RefEcon and FtrFaith is higher ranked excludes the pronoun candidate, in violation of LocalAnt[‡®n]. The item in (6) records the ranking in tableau T15.5 for future reference (braces or curly brackets around a set of constraints stands for ``at least one of '' that set). (6) Ranking schema for relativized minimality (with respect to domain d) {RefEcon, FtrFaith} g LocalAnt[d] g BindFaith. In OT, the major source of crosslinguistic variation is reranking, and every possible total ordering of the universal constraints de®nes a possible natural language. From this typological perspective, consider the hierarchy in which, contrary to the schema above, some LocalAnt[d] outranks both the expressive constraints. As the outcome of the competition in tableau T15.6 shows, this hierarchy generates a rigid minimality e¨ect: an anaphor contained in a domain d (here, a ®nite clause) cannot be bound from outside d even if d does not contain a potential antecedent for the anaphor. This hierarchy generates a pattern of binding found in another dialect/idiolect of Marathi discussed by Dalrymple (1993:19±20). Judgments of speakers of this variety, here called Marathi2, are given in (7), which should be compared with (1). (Note that the sentence exactly like (7b) except that aapan replaces swataah is grammatical in all dialects/idiolects, suggesting that FtrFaith is the active constraint in tableau T15.5.) (7) Rigid minimality in Marathi2 (Dalrymple 1993:19¨.) a. Tomi mhanat hota [ ki Suej ni swataahlaa  i=j maarle ]. Tom said that Sue ERG anaphor-ACC hit `Tom said that Sue hit him.'

478

Colin Wilson

Tableau T15.6 Rigid minimality (Marathi input) Input: Janei mhanaali [ ki [ swataacii parikshaa ] sampli ] J. said that anaphor-GEN test ®nished `Jane said that her (ˆ Jane's) test was over' Candidate Outputs a. Janei . . . [ . . . swataacii . . .] f b. Janei . . . [ . . . aapani . . .]

LocalAnt[‡®n]

RefEcon

FtrFaith

*

*

BindFaith

*!

b. *Janei mhanaali [ ki [ swataacii parikshaa ] sampli ]. Jane said that anaphor-GEN test ®nished (`Jane said that her/someone else's test was over.') To summarize, in the present system rigid and relativized minimality e¨ects arise from di¨erent rankings of the same constraints. This way of unifying the two e¨ects is supported by the fact that they are sensitive to the same syntactic domains, as illustrated by the dialectal/idiolectal variation in Marathi just discussed. And the success of this particular case of typology by reranking in turn supports one of the major claims of OT. Before closing this subsection, I compare the account of minimality e¨ects proposed here, which derives relativized minimality from the interaction of a rigid minimality constraint (LocalAnt[d]) and other constraints on anaphora, with a possible alternative account that would simply posit a relativized minimality constraint, say, the version given in (8). (8) Most Local Antecedent[d] An anaphor is bound by the most local potential antecedent (i.e., by the potential antecedent that is closest to the anaphor with respect to d domains). The internal structure of this constraint mirrors that found in de®nitions of ``binding category'' that refer to an accessible subject (Chomsky 1986; Huang 1983); in Rizzi's Relativized Minimality condition on government (Rizzi 1990); and in the Minimal Link Condition on movement (more recently, on attraction) in the Minimalist Program (Chomsky 1995). Conditions in this family, (8) included, perform (what I will call) an internal competition among alternative structures. That is, they judge a given relation (binding, government, chain link) as well formed if and only if it is as short as it could be, given the distribution of other elements that could potentially enter into the same type of relation (other potential antecedents, other potential governors,

Bidirectional Optimization and the Theory of Anaphora

479

other categories with the same features). Consequently, these conditions are not sensitive to the di¨erence between two relations whose absolute lengths (in terms of d domains) are not the same, but that are both as short as possible. This contrasts with LocalAnt[d], which, being a rigid minimality constraint, always records the absolute length of a binding relation (i.e., this constraint assesses a mark to a binding relation that crosses the boundary of a domain d even if that domain does not contain a potential antecedent). Restricting the discussion to OT, there are two arguments for the present approach over an alternative that would incorporate (8). First, MostLocalAnt has a degree of complexity that LocalAnt lacks. More important, the type of complexity built into the MostLocalAnt constraintÐthe ability to compare alternative binding relations to determine which is most localÐis of exactly the same type that OT accords to grammars. All OT grammars perform (what I will call) external competition: harmonic ordering of alternative structures, based on their constraint violations and a hierarchy. If constraints like (8) are allowed, the need for external competition (competition adjudicated by a hierarchy) will be mitigated by internal competition (competition built into the logic of a constraint). Given external competition, internal competition introduces a redundancy, and weakens the theory accordingly. Second, the account in which LocalAnt has been replaced by MostLocalAnt cannot use reranking to account for the dialectal/idiolectal variation noted above in Marathi. The reason is clear: MostLocalAnt does not assess a violation to binding relations like the one shown in candidate (a) in tableau T15.6; therefore, RefEcon and/or FtrFaith will demand an anaphor in the output regardless of how the three constraints are ranked. To account for Marathi2, the MostLocalAnt theory would have to be supplemented with a rigid minimality constraint, namely LocalAnt, which would then render the former constraint super¯uous. Therefore, both conceptual and empirical considerations lead to the conclusion that the locus of competition should be external to the constraints, not buried within them. 15.3

Relativized Minimality and Referential Economy in the Same Grammar

This section begins the second part of the chapter, in which I analyze the interaction between the two optimizations (interpretive and expressive) proposed in sections 15.1 and 15.2 respectively. Brie¯y reviewing the de®nitions of the optimizations, they both take pairs consisting of a syntactic structure and a semantic structure as inputs and return such pairs as outputs. But they are directly opposed with respect to which of the universally possible pairs qualify as ``alternatives'' to one another. Interpretive optimization holds the syntactic component of the input ®xed (i.e., every alternative structure, or candidate output, must share that component), while leaving the semantic component free (i.e., a candidate output need not share that component). In contrast,

480

Colin Wilson

expressive optimization holds the semantic component ®xed across the candidate set, while leaving the syntactic component free to vary. The question facing us now is how to integrate these two optimizations; how to reconcile their complementary de®nitions of the candidate set; how, in the terminology of the opening paragraphs of the chapter, to formulate a coherent model of grammar that recognizes both ``directions'' of the syntax-semantics interface. In section 15.3.1, I consider the simplest possible answer to this question, and derive from it a surprising prediction. The prediction is then tested against the anaphoric system of Chamorro (Chung 1981, 1989) and shown to be incorrect (section 15.3.2). This result motivates a more structured integration of interpretive and expressive optimization, which I develop and apply to Chamorro in section 15.4. 15.3.1 A Single-Optimization Model The candidate sets for interpretive and expressive optimization were de®ned conservatively, in the following sense. The empirical e¨ect used to motivate interpretive optimization (i.e., relativized minimality) only requires candidate outputs to di¨er from the input with respect to their semantic components; therefore it was assumed that the syntactic component of the input is held ®xed across the candidate set. Symmetrically, the main empirical e¨ect that motivated expressive optimization (i.e., referential economy) only requires candidate outputs to di¨er from the input with respect to their syntactic components; thus the semantic component was assumed to be held ®xed. Faced with the apparent problem of integrating the two optimizations, a natural hypothesis is that the de®nitions of the two optimizations were excessively parochial. Given that both e¨ects are observed, perhaps the correct conclusion to draw is that a candidate output can di¨er from the input on both components. Under this hypothesis, the distinction between interpretive and expressive optimization collapses Ðtherefore the problem of integrating them disappearsÐand we are left with the single-optimization model pictured in table 15.6. In the following analysis of table 15.6, we focus on the fact that it embodies the least restrictive theory of the interaction between the neutralization for relativized minimality (table 15.2) and the neutralization for referential economy (table 15.4). The unfaithful mappings within those two neutralizations have the abstract forms speci®ed in (9). Table 15.6 Single-optimization model Input

Output

Given any pair hS, Mi in I

. . . ®nd the most harmonic pair hT, Ni in IÐthat is, maximize the harmony of the syntactic and semantic components of the output simultaneously.

Bidirectional Optimization and the Theory of Anaphora

481

(9) Unfaithful maps a. [Ai . . . [d Bj . . . anaphori . . . ]] q [Ai . . . [d Bj . . . anaphorj . . . ]] (rel. min.) b. [Ai . . . pronouni . . . ] q [Ai . . . anaphori . . . ] (ref. econ.) Thinking about these mappings for the moment as processes, it should be clear that the single-optimization model in principle allows the one in (9b) to feed the one in (9a). That is, replacing a bound pronoun in the input with an anaphor in the output can trigger a change from nonlocal binding to local binding. This chain of processes, or combination of mappings, is pictured in (10). (In the picture, the mapping previously attributed to expressive optimization is on the horizontal axis, while the mapping previously attributed to interpretive optimization is on the vertical axis.) (10) Combination of the unfaithful maps [Ai . . . [d Bj . . . pronouni . . . ]] q [Ai . . . [d Bj . . . anaphori . . . ]] (input) # [Ai . . . [d Bj . . . anaphorj . . . ]] (output) Any grammar that combined the unfaithful mappings as in (10) would generate a radically degenerate anaphoric system. For not only would anaphor binding be subject to a relativized minimality e¨ect, as indicated by the second ``step'' of (10), but pronoun binding would also be subject to exactly the same e¨ect, by the ®rst ``step'' and the feeding relation. In other words, structures of the form [Ai . . . [d Bj . . . Xi . . . ]] (in which A binds X across the boundary of a syntactic domain d that contains a potential antecedent, namely B) would be absolutely ill formed according to such a grammar. Such a structure would be ungrammatical regardless of whether X is instantiated as an anaphor or a pronoun. I assert that the single-optimization model in fact predicts such an impoverished system whenever the relativized minimality e¨ect (with respect to anaphors) and the referential economy e¨ect are observed in the same language. Somewhat more formally, table 15.7 shows an implicational universal of the model. Before assessing the empirical validity of this prediction, let us ®rst understand why it holds. The rankings required by (a) in table 15.7 are given by the schemas in tableau T15.3 and (6) and repeated in (11). Table 15.7 Absolute ungrammaticality of binding across a potential antecedent a. If language L exhibits a referential economy e¨ect and a relativized minimality e¨ect with respect to domains of type d, b. then the single-optimization model predicts that [Ai ... [d Bj ... Xi ... ]] is ungrammatical in L regardless of whether X is a pronoun or an anaphor.

482

Colin Wilson

(11) Ranking schemas a. RefEcon g FtrFaith (ref. econ.) b. {RefEcon, FtrFaith} g LocalAnt[d] g BindFaith (rel. min.) The total orderings consistent with both schemas all have the property that RefEcon dominates LocalAnt[d] (i.e., RefEcon is the highest-ranked constraint; FtrFaith can be ranked anywhere below RefEcon). Consider now the input shown in (10): [Ai . . . [d Bj . . . pronouni . . . ]]. The fact that RefEcon dominates both FtrFaith and LocalAnt[d] implies that taking the ®rst ``step'' of (10) is more harmonic that having an output that is identical to the input. The mapping from [Ai . . . [d Bj . . . pronouni . . . ]] to [Ai . . . [d Bj . . . anaphori . . . ]] increases harmony by trading a violation of the highest-ranked constraint for violations of two lower-ranked constraints. But the system can increase the harmony of the input/output mapping even further by taking the second ``step'' of (10). Moving from nonlocal binding of the anaphor, as in [Ai . . . [d Bj . . . anaphori . . . ]], to local binding, as in [Ai . . . [d Bj . . . anaphorj . . . ]], trades a violation of LocalAnt[d] for a violation of the lower-ranked BindFaith. In other words, once the pronoun in the input has been replaced by an anaphor, the locality constraint on binding becomes active and forces local binding over nonlocal binding, as it does whenever possible in the language (by hypothesis). This discussion is summarized by tableau T15.7, in which the candidates have been ordered as in (10). The next subsection (section 15.3.2) brings evidence from Chamorro to bear on the prediction in table 15.7. Before undertaking this somewhat involved project, let me brie¯y point out that the literature contains no mention of any condition banning bound pronouns that are insu½ciently local to their binders. On the contrary, increasing the syntactic distance between a pronoun and its binder improves the binding relation in many cases (see section 15.2 and the various antilocality conditions in the literature, including the Principle B of Chomsky 1981 and subsequent Tableau T15.7 Nonlocal pronoun binding reduces to local anaphor binding Input: [Ai . . . [d Bj . . . pronouni . . . ]] Candidate Outputs

FtrFaith

LocalAnt…d†

b. [Ai . . . [d Bj . . . anaphori . . . ]]

*

*!

f c. [Ai . . . [d Bj . . . anaphorj . . . ]]

*

a. [Ai . . . [d Bj . . . pronouni . . . ]]

RefEcon

BindFaith

*!

*

Bidirectional Optimization and the Theory of Anaphora

483

research; the Chain Condition of Reinhart and Reuland 1993; the Overt Pronoun Constraint of Montalbetti 1984; and the Minimal Disjointness condition of Aoun and Li 1990). These general considerations lead us to expect that table 15.7, under which the relativized minimality condition on anaphor binding extends to bound pronouns, will not withstand empirical scrutiny. 15.3.2 Chamorro Anaphora In this section, I discuss evidence for three factual claims. First, Chamorro anaphors are subject to a relativized minimality e¨ect. Second, the distribution of Chamorro pronouns is sensitive to the distribution of anaphors in the particular way that I have been calling the referential economy e¨ect. And third, Chamorro pronouns can be bound in structures that, given the two e¨ects just mentioned, the single-optimization model predicts to be ungrammatical. My discussion of this language closely follows the insightful analysis of Chung (1981, 1989), to which the reader is referred for further details about Chamorro grammar (see also Chung 1983, 1987; Topping 1973). The distinction between anaphors and pronouns is phonologically marked for the inanimate ``proforms'' in Chamorro, but not for the animate ones, which will be the focus of our discussion. (I use the term ``proform'' as Chung 1989 does, to refer to pronouns and anaphors as a class.) That the anaphor-pronoun distinction indeed extends to the animate proforms can be perceived through the interaction of anaphora with certain surface ®lters that apply to Chamorro transitive clauses (Chung 1981). One of these surface ®lters, which bans transitive clauses with third-person plural subjects, will be especially relevant for establishing a relativized minimality e¨ect in the language. Some basic data illustrating this ®lter is provided in (12). (12) *Third-person plural subject of transitive clause a. Man-hugandu i famagu"un. [Chung 1981:9bÐhenceforth, C81:9b] AGR-play the children `The children played.' b. Ha-bisita si Juan si nana-nÄa. [C81:18] AGR-visit UNM John UNM mother-his `John visited his mother.' c. *Ma-bisita i manatungo"-hu si Juan gi hospitat. AGR-visit the friends-my UNM John LOC hospital `My friends visited Juan at the hospital.' [C81:27a] d. Ma-bisita si Juan gi hospitat ni manatungo"-hu. PASS-visit UNM John LOC hospital OBL friends-my `Juan was visited at the hospital by my friends.' [C81:38a] In (12a), the subject (i famagu"un) is third-person plural, but the clause is intransitive; therefore this sentence escapes the ®lter. In (12b), the clause is transitive, but the

484

Colin Wilson

subject (Juan) is third-person singular, so this sentence also escapes the ®lter. But in (12c) the clause is transitive and the subject (i manatungo"-hu) is third-person plural; therefore the ®lter rules this sentence out. Example (12d) is included to show that passive clauses do not inherit the ungrammaticality of the corresponding actives. Although generally true of transitive clauses in Chamorro, the surface ®lter at work in (12) has a set of systematic exceptions, as uncovered by Chung (1981). Here, the relevant exception is that the ®lter is suspended whenever the direct object either is or contains a proform that is (1) interpreted as bound by the subject, and (2) su½ciently close to the subject to be an anaphor. For our purposes, the signi®cant contrast is illustrated in (13). (13) Interaction between bound anaphora and the surface ®lter a. Ma-fa"tinas [i statue nu sihai=  k ] ni kle i famalao"ani . INFL-make the statue OBL them(selves) OBL clay the women `The womeni made statues of themselvesi /*themk out of clay' [C1989:25d] b. *Ma-li"i" [i litratu-n Jose put sihai=k ] i famalao"ani . INFL-see the picture-L Jose about them(selves) the women `The womeni saw Jose's picture of themi=k ' [C1989:26b] Both the examples in (13) contain a transitive clause with a third-person pural subject (i famalao"an `the women'). Starting with (13a), observe that the sentence is ungrammatical, as expected, when the proform siha within the direct object is not bound by the subject. However, the sentence is grammaticalÐin violation of the surface ®lterÐwhen the proform is bound by the subject. Example (13b) contrasts minimally with (13a), having the same syntactic structure except for the fact that the direct object of the clause contains a possessor (Jose). Describing this minimal contrast in terms that should by now be familiar, in (13a) the closest potential antecedent for the proform siha is the matrix subject. But in (13b) the proform has a potential antecedent closer than the matrix subject, namely Jose. This small syntactic change makes all the di¨erence as far as the surface ®lter is concerned: (13b) is ungrammatical regardless of whether siha is interpreted as bound by the matrix subject. The fact that (13a) and (13b) di¨er with respect to the position of the closest potential antecedent for siha strongly suggests that the grammaticality contrast between them is related to a relativized minimality e¨ect of the kind analyzed in sections 15.1 and 15.2. To formalize this idea, we have to ®rst accept that there is a morphosyntactic distinction between anaphoric and pronominal siha (and other animate proforms; Chung 1989). As discussed at the end of section 15.3.1, it would be typologically anomalous for pronoun binding to be subject to relativized minimality. And, more conclusively, we will see cases of pronoun binding that defy both relativized and rigid minimality. Thus, under certain circumstances an argument realized as siha does not

Bidirectional Optimization and the Theory of Anaphora

485

have lexical, or ``inherent,'' j-features (i.e., it is an SE anaphor). (The fact that the form in¯ects for such features must therefore be due to a ``spelling-out'' of the binding relation, or else to a transfer of features from the antecedent to the anaphor. Although this may appear problematic, I later discuss a crosslinguistic typology within which Chamorro represents an expected, if extreme, case.) If the grammaticality di¨erence between (13a) and (13b) relates to relativized minimality, the relevant instantiation of the minimality constraint LocalAnt[d] is violated when anaphor binding crosses the boundary of a noun phrase, call it LocalAnt[NP]. Following the ranking schema for relativized minimality (6), this constraint must outrank BindFaith and be outranked by RefEcon and/or FtrFaith. (Looking ahead, I take the dominant constraint to be RefEcon and suppress FtrFaith for the moment.) As shown in tableau T15.8, ranking RefEcon above whatever Markedness constraint accounts for the surface ®lter yields the contrast in (13). Provisionally, I dub the latter constraint ``No3plSubject'' (see Aissen, chap. 3, this volume and Artstein 1998 for formal theories of related constraints). And I use ``SE'' to stand for the Chamorro anaphor regardless of its phonological realization. In this tableau, candidate outputs are allowed to di¨er from the input on three dimensions: the features of arguments (e.g., an anaphor in the input can be exchanged for a pronoun in the output); the interpretation of arguments (e.g., an anaphor bound by one antecedent in the input can be bound by another antecedent in the output, or even replaced by a free pronoun); and voice (e.g., clauses that are active in the input can be realized as passive in the output). All three types of input-output disparity are sanctioned under the single-optimization model (table 15.6), which we are provisionally adopting. And the only type that has not been previously discussed in this chapter, change in voice, has been independently proposed in the literature (see, for example, Legendre, Raymond, and Smolensky 1993). The important points to draw from tableau T15.8 are as follows. For the ®rst input, the proform contained in the input direct object can be licensed as an anaphor only if the output clause is active. Only then will the proform be c-commanded by a potential antecedent (i.e., the subject i famalao"an). The ranking RefEcon g ``No3plSubject'' therefore guarantees that the output is active, despite the fact that doing so creates a transitive clause with a third-person plural subject. In particular, the dominance of RefEcon ensures that the passive candidate ((b) in tableau T15.8), in which the surface ®lter is satis®ed but the object-internal argument is a pronoun, is nonoptimal. Again, passivization and binding are incompatible in this candidate set, because the oblique argument of a passive does not c-command into the subject. In the candidate set for the second input, however, passivization and anaphor binding can co-occur. Speci®cally, if the argument bound by the subject in the input

b. INFL-PASS-V [ NP . . . sihak ] . . . famalao"ani Passive: `Statues of themk were made by the womeni . . .'

f a. INFL-V [ NP . . . SEi ] . . . famalao"ani Active: `The womeni made statues of themselvesi . . .'

Candidate Outputs

Input: Ma-fa"tinas [i statue nu SEi ] ni kle i famalao"ani INFL-make the statue OBL OBL clay the women Active: `The womeni made statues of themselvesi out of clay'

Tableau T15.8 Relativized minimality in Chamorro

*!

RefEcon

*No3plSbj *LocalAnt(NP)

No3plSubj, LocalAnt(NP)

*

BindFaith

486 Colin Wilson

f. INFL-PASS-V [NP . . . Josej . . . sihak ] . . . OBL famalao"ani Passive: `Josej 's picture of themk was seen by the womeni '

e. INFL-V [NP . . . Josej . . . SEi ] . . . famalao"ani Active: `The womeni saw Josej 's picture of themselvesi '

d. INFL-V [NP . . . Josej . . . SEj ] . . . famalao"ani Active: `The womeni saw Josej 's picture of himselfj '

f c. INFL-PASS-V [NP . . . Josej . . . SEj ] . . . OBL famalao"ani Passive: `Josej 's picture of himselfj was seen by the women'

Candidate Outputs

Input: Ma-li"i" [i litratu-n Josej put SEi ] i famalao"ani INFL-see the picture-L J. about the women Active: `The womeni saw Josej 's pictures of themselvesi '

Tableau T15.8 (continued )

*!

RefEcon

*No3plS! *LocalAnt(NP)!

*No3plS!

No3plSubj, LocalAnt(NP)

*

*

BindFaith

Bidirectional Optimization and the Theory of Anaphora 487

488

Colin Wilson

remains an anaphor, but changes its antecedent to the possessor (Jose), then passivization can save a violation of the surface ®lter while satisfying the (assumed inviolable) c-command condition on binding. (Jose and the anaphor are constituents of the same core argument, and therefore their c-command relationship does not change under passivization.) This simultaneous satisfaction of RefEcon, ``No3plSubject,'' and the c-command condition is achieved by the optimal candidate, (c) in tableau T15.8. This candidate is more harmonic than (d) because the latter spuriously violates the surface ®lter, more harmonic than (e) because (c) contains a more local binding relation, and more harmonic than (f ) because the latter contains an additional argument with independent reference. To summarize, an anaphor-pronoun distinction in the animate pronouns of Chamorro, and furthermore a relativized minimality e¨ect on the anaphors, is visible through the interaction of anaphora with the surface ®lter on transitive clauses with third-person plural subjects. When the subject of a transitive clause is the closest potential antecedent for an argument contained within the direct object, that argument can be realized as an anaphor bound by the subject, in which case referential economy overrides the surface ®lter. But when the object-internal argument has a closer potential antecedent (speci®cally, the possessor of the direct object), binding by the clausal subject is blocked by relativized minimality, hence the surface ®lter is free to apply. The preceding discussion establishes half of what is necessary to (nonvacuously) apply the prediction derived in section 15.3.1 to Chamorro. Before establishing the other half, namely, that Chamorro exhibits a referential economy e¨ect in addition to the relativized minimality e¨ect discussed aboveÐI pause brie¯y to consider the status of anaphors that spell-out the j-features of their antecedents. Recall that SE anaphors have been de®ned here, as elsewhere in the literature, as elements that lack lexical speci®cation for the j-features. Although this lexical underspeci®cation often leads to morphophonological invariance (i.e., the form of the anaphor does not re¯ect the features of its antecedent), the anaphors of Chechen and Ingush (Caucasian; Nichols 1994a, 1994b) and of Northern Pomo (Pomo; O'Connor 1992) show that such invariance is not a necessary property of SE anaphors. For example, in Chechen the anaphors in¯ect for the person and number of their antecedents (see Nichols 1994a). In their nonlogophoric use, these forms obey familiar syntactic conditions on SE anaphors: they must be bound by a subject; they must have an antecedent within the ®rst indicative clause that contains them, a minimality condition we also observed in section 15.2.1; and the third-person forms are in complementary distribution with the corresponding pronouns. Similarly, the ti-anaphors of Northern Pomo in¯ect for the number of their antecedents (see O'Connor 1992: 287). When not interpreted logophorically, these anaphors must be bound by a sub-

Bidirectional Optimization and the Theory of Anaphora

489

ject within the same clause, they are restricted to third-person antecedents, and they are in complementary distribution with the third-person pronouns. If languages have the option of realizing the j-features of an SE anaphor's antecedent on the anaphor itself, one possibility is for such in¯ected anaphors to come out homophonous with minimally di¨erent forms that have j-features inherently (i.e., pronouns). As already discussed, I take this (possibly marked) state of a¨airs to hold within the set of Chamorro animate third-person proforms, essentially following Chung (1989). The question of what mechanism allows anaphors to be spelled out with the features of their antecedents will not be examined here, although agreement seems a promising candidate (see Anderson 1992; Burzio 1992; Reinhart and Reuland 1993). Understanding the argument for a referential economy e¨ect in Chamorro requires some background knowledge about how certain wh-questions are formed. One relevant contrast is provided in (14). (14) Interaction between referential economy and wh-questions a. Hayii ti paÈra u-gu-in-aiya ti [i opinian Maria giya guiyai ]? who? not will INFL-PASS-love the opinion.L Maria LOC him `Whoi is Maria's opinion of himi not going to be liked by?' [C1989:45b] b. *Hayii paÈra u-ni-naÈ"na" ti [i minagahit put guiyai ]? who? will INFL-PASS-hide the truth about him `Whoi is the truth about him(self )i going to be hidden by?' [C1989:46b] c. Hayii paÈra u-na"na" ti [i minagahit put guiyai ]? who? will INFL-hide the truth about him `Who will hide the truth about himselfi ?' [C1989:47b] In Chamorro, ``the normal way to question the external argument of a transitive verb in the irrealis is to form a Wh-question on the agent of the corresponding passive'' (Chung 1989:159). Thus the passive sentence (29a) is grammatical, whereas its agent counterpart would not be (Chung 1989:159). But this general strategy for forming questions leads to ungrammaticality in examples like (14b). To express the interpretation in which the wh-phrase and the proform within the patient argument are coindexed, the active (14c) must be used instead. As observed by Chung, the di¨erence between the two passive sentences (14a) and (14b) lies in precisely how local the trace of the wh-phrase would be to the proform guiya (animate third-person singular) in the corresponding active sentence. In our terms, the active counterpart of (14a) would not place the trace close enough to satisfy relativized minimality; the possessor Maria stands as a closer potential antecedent for the proform, just as Jose dose for siha in example (14b). But in the active counterpart of (14b), namely (14c), the trace is close enough to satisfy relativized

490

Colin Wilson

minimalityÐimportantly, the noun phrase containing guiya does not have a possessor (or other argument) that could potentially bind the proform. Given this observation, we can analyze the contrast between (14a) and (14b) as follows. First, the form guiya realizes both an animate third-person singular pronoun and an anaphor bound by an argument with those features. Second, there is a preference for using an anaphor instead of a pronoun to express a binding relation, and this preference can override the factor that drives passivization when the external argument of an irrealis clause is questioned. Third, the preference for an anaphor is itself overridden by the relativized minimality e¨ect discussed earlier. Thus the passive (14a) is grammatical because even in the corresponding active the wh-trace is not close enough to license an anaphor in the position of guiya; hence the proform is a true pronoun in this example. But the passive (14b) is ungrammatical because in the active (14c) the wh-trace is the closest potential antecedent for the proform, thus licensing it as an anaphor; guiya in this last example realizes an anaphor together with the j-features of its antecedent. The crucial point for us to draw from this analysis is that the contrast between (14b) and (14c) reduces to a referential economy e¨ect of the type discussed in section 15.2. The possibility of using an anaphor, as in (14c), makes the pronoun ungrammatical under the same interpretation, as in (14b). To formalize this portion of the account, we invoke the ranking schema for referential economy: RefEcon g FtrFaith. Additionally, it must be the case that RefEcon outranks whatever constraint prefers passivization under extraction of the external argument. I have no de®nite proposal to make concerning this constraint, and therefore simply assume that the passive is optimal when the constraints considered here do not favor the active. Tableau T15.9 accounts just for the contrast between (14b) and (14c); further discussion of (14a) is provided later. The competition in this tableau is between an active sentence in which the trace of the wh-phrase binds an anaphor (as before, written ``SE'') and a passive sentence in which the bound argument is (faithfully) realized as a pronoun. ReferentialEconomy decides the competition in favor of the ®rst structure, despite the fact that it is unfaithful to the input (*FtrFaith), has a binding relation that violates locality (*LocalAnt[NP]), and transgresses whatever constraint generally enforces passivization under external-argument extraction in the irrealis. As in the ®rst competition of section 15.3.2 (see (28)), passivization is incompatible with anaphor binding, because the noun-phrase internal proform would not be c-commanded by a coindexed element in an A-position (speci®cally, the wh-trace would not c-command into the subject from the oblique-argument position). Having reviewed evidence for the claim that both relativized minimality and referential economy play an active role in the anaphoric system of Chamorro, I now return to the single-optimization model (20) and its predictions about such languages. The implicational universal derived from the model in section 15.3.1 is repeated in

b. Hayii . . . INFL-PASS-V ti [ . . . guiyai ] Passive: `Whoi is the truth about himi going to be hidden by?'

f a. Hayii . . . INFL-V ti [ . . . SEi ] Active: `Whoi is going to hide the truth about himselfi ?'

Candidate Outputs

*!

RefEcon

ti [i minagahit put guiyai ]? Input: Hayii paÈra u-ni-naÈ"na" who? will INFL-PASS-hide the truth about him Passive: `Whoi is the truth about himi going to be hidden by?'

Tableau T15.9 Referential economy in Chamorro

*

FtrFaith *

LocalAnt(NP)

BindFaith

Bidirectional Optimization and the Theory of Anaphora 491

492

Colin Wilson

Table 15.8 Absolute ungrammaticality of binding across a potential antecedent a. If language L exhibits a referential economy e¨ect and a relativized minimality e¨ect with respect to domains of type d, b. then the single-optimization model predicts that [Ai ... [d Bj ... Xi ... ]] is ungrammatical in L regardless of whether X is a pronoun or an anaphor.

table 15.8 from table 15.7. This prediction is discon®rmed by examples such as (14a), repeated below as (15a), in which a pronoun is bound by a wh-operator. (15) Binding by a wh-operator a. Hayii ti paÈra u-gu-in-aiya ti [i opinian Maria giya guiyai ]? who? not will INFL-PASS-love the opinion.L Maria LOC him `Whoi is Maria's opinion of himi not going to be liked by?' [C1989:45b] b. Hayii paÈra u-in-agradesi ti [i inangokku-n Maria nu guiyai ]? who? will INFL-PASS-appreciate the trust-L Maria OBL him `Whoi is Maria's trust in himi going to be appreciated by?' [C1989:45a] c. Hayii paÈra u-k-in-astiga ti [i paÈtgun [ OPj nik-um-aÈssi tj gui"i ]]? him who? will INFL-PASS-punish the child INFL-tease `Whoi is the child who teased himi going to be punished by?' [C1989:45c] Given that Chamorro has a relativized minimality e¨ect sensitive to noun phrase boundaries, all three examples above ®t the structure that the single-optimization model predicts to be ungrammatical. In each case, a pronoun (guiya or gui" ) is bound by an operator (here, hayi `who') across the boundary of a noun phrase that contains a potential antecedent for the pronoun (namely, Maria in (15a) and (15b) and the trace of the relative operator in (15c)). Note that, in these examples, the pronoun must in fact be bound by the operator in the strict sense. Because the wh-phrase does not refer to a particular individual in the domain of discourse, the pronoun cannot link to the wh-phrase indirectly (i.e., by so-called accidental coreference). Similar remarks apply to binding by a quanti®ed noun phrase, as in the example in (16).6 (16) Binding by a quanti®ed noun phrase in-pe"lu kaÈda unu na rabbiti gi [kasiyas ni INFL-put each one L rabbit LOC cage COMP in-fahani paÈra guiyai ]. WH[obj].agr-buy.for for it `We put each rabbiti in the cage that we had bought for iti ' Here the bound pronoun guiya is contained within the relative clause modifying kasiyas `cage'. The potential antecedent within the noun phrase is the subject of the

Bidirectional Optimization and the Theory of Anaphora

493

relative clause. The grammaticality of this example, as well as those discussed earlier, therefore provides direct evidence against the single-optimization model. A concrete demonstration of how the single-optimization model incorrectly rules out these grammatical stuctures is provided by tableau T15.10, in which (15a) is the input. In the next section, I prevent such undesired input-output mappings by limiting the interaction between the constraints responsible for referential economy and those responsible for relativized minimality. 15.4

Bidirectional Optimization

Recalling the process-oriented discussion of section 15.3.1, the single-optimization model (20) gives rise to the prediction just discon®rmed by allowing the unfaithful mappings associated with referential economy and relativized minimality to stand in the feeding relation shown in (10), repeated in (17). (17) Improper combination of the unfaithful maps [Ai . . . [d Bj . . . pronouni . . . ]] q [Ai . . . [d Bj . . . anaphori . . . ]] (input) # [Ai . . . [d Bj . . . anaphorj . . . ]] (output) The new model of the syntax-semantics interface I propose excludes this feeding relation by (i) maintaining the distinction between the interpretive optimization of table 15.3 and the expressive optimization of table 15.5, which the single-optimization model obliterates, and (ii) applying the two optimizations in the order just mentioned. This bidirectional model is de®ned as in table 15.9, where I stands for the universal base of hsyntax, semanticsi pairs, as before, and I 0 stands for the subset of I that survives interpretive optimization. A member hS, Mi of I is in I 0 i¨ it is the optimal output for some input during interpretive optimization. The ®rst stage of this model ®lters the universal base (I) to a (typically smaller) inventory of structures (I 0 ). The second stage can ®lter this inventory further, and typically will, but it cannot regain distinctions lost during the ®rst stage. This is the sense (and the only sense) in which interpretive optimization precedes expressive optimization in the bidirectional model. The set-theoretic picture in table 15.10 shows how these optimizations-qua-®lters apply in succession to de®ne a language-particular inventory. An informal understanding of how the bidirectional optimization model avoids mappings like the one in (17) can be gained by considering what type of neutralization is possible at each stage of the model. During the interpretive stage, candidate outputs

c. Hayii . . . INFL-PASS-V ti [ . . . Mariaj . . . guiyai ] Passive: `Whoi is Mariaj 's opinion of himi not going to be liked by?'

b. Hayii . . . INFL-V ti [ . . . Mariaj . . . SEi ] Active: `Whoi is not going to like Mariaj 's opinion of himi ?'

f a. Hayii . . . INFL-PASS-V ti [ . . . Mariaj . . . SEj ] Passive: `Whoi is Mariaj 's opinion of herselfj not going to be liked by?'

Candidate Outputs

*!

RefEcon

Input: Hayii ti paÈra u-gu-in-aiya ti [i opinian Mariaj giya guiyai ]? who? not will INFL-PASS-love the opinion.L M. LOC him Passive: `Whoi is Mariaj 's opinion of himi not going to be liked by?'

Tableau T15.10 Improper reduction of the anaphoric system

*

*

FtrFaith

*!

LocalAnt(NP)

*

BindFaith

494 Colin Wilson

Bidirectional Optimization and the Theory of Anaphora

495

Table 15.9 Bidirectional optimization model Optimization

Input

Output

interpretive

Given any pair hS, Mi in I

. . . ®nd the most harmonic pair hS, Ni in IÐthat is, maximize the harmony of the semantic interpretation of S.

expressive

Given any pair hS, Mi in I 0

. . . ®nd the most harmonic pair hT, Mi in I 0 Ðthat is, maximize the harmony of the syntactic expression of M.

Table 15.10 Serial order of optimizations universal base (I)

K intermediate inventory (I 0 )

K

language-particular inventory (set of all grammatical structures in language L)

must take their syntactic components (including the morphosyntactic feature speci®cations of arguments) from the input. Therefore, the neutralization that accounts for referential economy (table 15.4) is not possible during this stageÐeven in a language that ranks RefEcon above the feature-faithfulness constraint (FtrFaith). On the other hand, the neutralization that accounts for relativized minimality e¨ects (table 15.2) is possible during this stage, subject to the constraint hierarchy of the language at hand. During the expressive stage, exactly complementary conditions hold. The neutralization for relativized minimality is no longer possible, because candidate outputs must share the semantic component (in particular, the binding relations) of the input. But now the neutralization for referential economy is a live option, and will occur unless the constraint hierarchy of the language at hand prevents it. Given that interpretive optimization precedes expressive optimization, this complementary distribution of the two neutralizations yields the following result: the particular combination of mappings in (17) cannot arise in the bidirectional optimization model, because the optimization that permits the ®rst ``step'' is ordered after the optimization that permits the second.7 To summarize, formal and empirical analysis has uncovered a deep defect in the single-optimization model of the syntax-semantics interface. In general terms, the defect amounts to an inability to constrain the combination of unfaithful inputoutput mappings that are individually necessary. The bidirectional model overcomes this problem by making the relationship between syntactic expressions and semantic interpretations more structured. It leaves the individual mappings intact, but properly regulates their combination by keeping the two logical ``directions'' of the interface

496

Colin Wilson

Table 15.11 An abstract base a. [Ai ... [d ... anaphori ... ]] b. *[Ai ... [d ... pronouni ... ]]

c. [Ai ... [d Bj ... anaphorj ... ]] d. *[Ai ... [d Bj ... pronounj ... ]] e. *[Ai ... [d Bj ... anaphori ... ]] f. [Ai ... [d Bj ... pronouni ... ]]

separate and applying them sequentially. The next subsection (section 15.4.1) applies the bidirectional model to abstract structures of the kind shown in (17). We then return to the problematic Chamorro data in section 15.4.2. 15.4.1 Abstract Application of the Model For the purposes of this demonstration, the universal base of structures (I) is de®ned by table 15.11. Note that the ``*'' diacritics above mark the goal of the demonstration, not its starting point. That is, we want to derive the pattern of grammatical and ungrammatical structures from an initial inventory in which all the permutations of binder and bound element shown in table 15.11 are equally legitimate. The desired pattern, of course, combines a referential economy e¨ect and a relativized minimality e¨ect with respect to domains of type d. The necessary rankings have been deduced already; see (11) and tableaux T15.11 and T15.12. The two optimizations of the bidirectional model sort the structures of table 15.11 into di¨erent piles, or equivalence classes. For interpretive optimization, which applies ®rst, two structures are in the same class i¨ they share the same syntactic component. This immediately implies that (a) and (b), which do not share their syntactic components with any other members of the base, survive the ®rst stage of the bidirectional model (i.e., each one is mapped to itself by interpretive optimization). Tableau T15.11 shows how interpretive optimization applies when candidates (c)±(f ), each of which belongs to an equivalence class that contains another member of the base, are taken as inputs. (In this subsection and the next, I adopt the conventions of placing the input ®rst in each candidate set and using double lines to separate candidate sets. In tableau T15.11, the letter to the left of each candidate indexes it with respect to table 15.11.) The intermediate inventory (I 0 ) is de®ned as the set of optimal candidates in tableau T15.11, together with (a) and (b) in table 15.11, which are optimal by default. Note that structure (e) in table 15.11, in which an anaphor is bound from outside a d domain that contains a potential antecedent, has been excluded from the inventory by neutralization with (c). Note also that even the highest-ranked constraint is violated by optimal forms when no alternative structure satis®es it (see the third and fourth candidate sets immediately above).

Bidirectional Optimization and the Theory of Anaphora

497

Tableau T15.11 Interpretive optimization applied to the abstract base Inputs and Candidate Outputs

RefEcon

FtrFaith

LocalAnt…d†

BindFaith

e. [Ai . . . [d Bj . . . anaphori . . .]]

*!

*

e. [Ai . . . [d Bj . . . anaphori . . .]]

*!

f c. [Ai . . . [d Bj . . . anaphorj . . .]]

f c. [Ai . . . [d Bj . . . anaphorj . . .]]

*

f d. [Ai . . . [d Bj . . . pronounj . . .]]

*

f. [Ai . . . [d Bj . . . pronouni . . .]]

*

f f. [Ai . . . [d Bj . . . pronouni . . .]]

*

d. [Ai . . . [d Bj . . . pronounj . . .]]

*

*!

*!

Expressive optimization takes the members of I that are left in I 0 and groups them into new piles/equivalence classes. Now two structures are in the same class i¨ they share the same semantic component. Under this grouping, every remaining structure has a competitorÐexcept, importantly, candidate (f ). Tableau T15.12 presents three of the relevant optimizations (the remaining two optimizations, which yield the identity maps (a) q (a) and (c) q (c), can be easily deduced from those shown). After this second stage, only the three optimal structures of tableau T15.12 remain in the inventory of grammatical forms, and these are exactly the three that are grammatical according to the pattern in table 15.11. Noneconomical (b) and (d) have been eliminated by referential economy. But crucially, (f )Ðthe structure that the singleoptimization model predicts to be ungrammatical given the same hierarchyÐexists in the ®nal inventory, despite the fact that it violates RefEcon. The only alternative structure that could beat (f ) during expressive optimization is the anaphor candidate (e). But, as already noted, (e) was removed from the set of legitimate competitors by the ®rst optimization. (Note in particular that the structure to which (e) was neutralized, namely (c), does not contain the same binding relation as (f ); therefore (c) and (f ) do not compete.) Thus the bidirectional optimization model allows referential economy to play an active role in shaping an anaphoric system while preventing it from reducing that system to an unnaturally limited state. Table 15.12 summarizes the optimizations performed in this section.

498

Colin Wilson

Tableau T15.12 Expressive optimization applied to the intermediate inventory Inputs and Candidate Outputs

RefEcon

b. [Ai . . . [d . . . pronouni . . .]]

FtrFaith

LocalAnt…d†

*

*

BindFaith

*!

f a. [Ai . . . [d . . . anaphori . . .]] d. [Ai . . . [d Bj . . . pronounj . . .]]

*!

f c. [Ai . . . [d Bj . . . anaphorj . . .]] f f. [Ai . . . [d Bj . . . pronouni . . .]]

* *

Table 15.12 Summary of optimizations Universal base …I† [Ai [Ai [Ai [Ai [Ai [Ai

[d [d [d [d [d [d

Interpretive optimization

Intermediate inventory …I 0 †

anaphori ]] ƒƒƒƒƒƒƒƒƒƒƒ! [Ai pronouni ]] ƒƒƒƒƒƒƒƒƒƒ ƒ! [Ai Bj anaphorj ]] ƒƒƒƒƒƒƒƒ! [Ai ! Bj anaphori ]] ƒƒƒƒƒƒƒƒ Bj pronounj ]] ƒƒƒƒƒƒƒƒ! [Ai Bj pronouni ]] ƒƒƒƒƒƒƒƒƒ! [Ai

Expressive optimization

Inventory of grammatical forms

[d anaphori ]] ƒƒƒƒƒƒƒƒƒƒƒƒƒ ƒ! [Ai [d anaphori ]] ƒ! [d pronouni ]] ƒƒƒƒƒƒƒƒƒƒƒƒ [d Bj anaphorj ]] ƒƒƒƒƒƒƒƒƒƒƒ! [Ai [d Bj anaphorj ]] ƒƒƒ! ƒƒƒƒ ƒ ƒ ƒ [d Bj pronounj ]] ƒƒ [d Bj pronouni ]] ƒƒƒƒƒƒƒƒƒƒƒ! [Ai [d Bj pronouni ]]

15.4.2 Application to Chamorro The analysis of the previous subsection carries over straightforwardly to the Chamorro facts discussed in section 15.3.2. For concreteness, I illustrate the account using the irrealis wh-questions in (14c) and (15a). Taking ®rst (14c), in which there is no potential antecedent within the noun phrase that contains the bound proform, the relevant interpretive and expressive tableaux are provided below; see tableaux T15.13 and T15.14. As in section 15.3.2, I make the standard assumption that an anaphor must be A-bound; therefore the counterpart of candidate (14c) with an anaphor is impossible, because the trace of the wh-phrase does not c-command the direct object from the oblique position. (Following the conventions adopted in section 15.4.1, the input for each competition below is the ®rst candidate of that competition and competitions are separated by double lines.) So far, the bidirectional model produces the same results as the single-optimization model, correctly accounting for the contrast between the active (14c) and the passive

* *

f c. Hayii . . . INFL-PASS-V ti [ . . . guiyai ] Passive: `Whoi is the truth about himi going to be hidden by?'

RefEcon

f b. Hayii . . . INFL-V ti [ . . . guiyai ] Active: `Whoi will hide the truth about himi ?'

f a. Hayii . . . INFL-V ti [ . . . SEi ] Active: `Whoi will hide the truth about himselfi ?'

Inputs and Candidate Outputs

Meta-input: Hayii paÈra u-na"na" ti [i minagahit put Xi ]? who? will INFL-hide the truth about Active: `Whoi will hide the truth about him(self )i ?'

Tableau T15.13 Interpretive optimization

FtrFaith *

LocalAnt(NP)

BindFaith

Bidirectional Optimization and the Theory of Anaphora 499

*! *!

c. Hayii . . . INFL-PASS-V ti [ . . . guiyai ] Passive: `Whoi is the truth about himi going to be hidden by?'

RefEcon

b. Hayii . . . INFL-V ti [ . . . guiyai ] Active: `Whoi will hide the truth about himi ?'

f a. Hayii . . . INFL-V ti [ . . . SEi ] Active: `Whoi will hide the truth about himselfi ?'

Candidate Outputs

Input: Hayii paÈra u-na"na" ti [i minagahit put guiyai ]? who? will INFL-hide the truth about him

Tableau T15.14 Expressive optimization

*

FtrFaith *

LocalAnt(NP)

BindFaith

500 Colin Wilson

Bidirectional Optimization and the Theory of Anaphora

501

Tableau T15.15 Interpretive optimization

(14b). (See tableau T15.9 for the single-optimization account.) The predictions of the two models diverge, in favor of the bidirectional alternative, with respect to examples like (15a), in which there is a potential antecedent within the noun phrase that contains the bound proform. The single-optimization model predicts such binding relations to be absolutely ungrammatical. But the bidirectional optimization model sanctions them, in accord with the facts of Chamorro, as shown in tableaux T15.15 and T15.16. (Here I simplify the exposition by omitting some candidates in the active voice; recall that, in this case, passive is the grammatical voice regardless of the interpretation of the proform. Nothing changes if the other active-voice candidates are included. As in section 15.3.2, I do not show the constraint that enforces passive over active, all other things being equal.)

f e. Hayii . . . INFL-PASS-V ti [ . . . Mariaj . . . guiyai ] Passive: `Whoi is Mariaj 's opinion of himi not going to be liked by?'

d. Hayii . . . INFL-V ti [ . . . Mariaj . . . SEj ] Active: `Whoi is not going to like Mariaj 's opinion of herselfj ?'

f c. Hayii . . . INFL-PASS-V ti [ . . . Mariaj . . . SEj ] Passive: `Whoi is Mariaj 's opinion of herselfj not going to be liked by?'

b. Hayii . . . INFL-V ti [ . . . Mariaj . . . guiyaj ] Active: `Whoi is not going to like Mariaj 's opinion of herselfj ?'

a. Hayii . . . INFL-PASS-V ti [ . . . Mariaj . . . guiyaj ] Passive: `Whoi is Mariaj 's opinion of herselfi not going to be liked by?'

Inputs and Candidate Outputs

*

*!

*!

RefEcon

Input: Hayii ti paÈra u-gu-in-aiya ti [i opinian Mariaj giya X ]? who? not will INFL-PASS-love the opinion.L M. LOC Passive: `Whoi is Mariaj 's opinion of himi /her(self )j not going to be liked by?'

Tableau T15.16 Expressive optimization

*

*!

FtrFaith

LocalAnt(NP)

BindFaith

502 Colin Wilson

Bidirectional Optimization and the Theory of Anaphora

503

The key result of the analysis appears in the ®nal optimization of tableau T15.16. Binding of a pronoun (here, guiya) across a potential antecedent (here, Maria) is optimal, even though it violates the highest-ranked constraint, because there simply is no alternative structure left to block it. All the structures that contain an anaphor and that remain after interpretive optimization (tableau T15.15) have a di¨erent binding relation, and are therefore unable to compete against (e) in tableau T15.16. Note in particular that tableau T15.15(a), which would be more harmonic than tableau T15.16(e) if it were in the candidate set, was removed from the inventory of possible forms by the earlier optimization (speci®cally, by neutralization to local binding). Thus the bidirectional model, unlike the single-optimization model, correctly accounts for the status of RefEcon in the anaphoric system of ChamorroÐits strength, as in the ®rst competition of tableau T15.16, and its violability, as in the second competition. More generally, the new model preserves the simple constraint systems of sections 15.1 and 15.2 by using the two logical directions of the syntax-semantics interface to temper the interaction between them. 15.5

Conclusion

Due to space limitations, I must restrict this conclusion to a couple of brief remarks. First, I should acknowledge a conceptual relation between the proposal made here and Prince and Smolensky's (1993: chap. 9) Lexicon Optimization. Given Lexicon Optimization, OT phonology has the same ``bidirectional'' character that I have ascribed to the syntax-semantics interface. (For further development of Lexicon Optimization, see Tesar and Smolensky 1998. See also Smolensky 1996 for a related account of comprehension/production mismatches.) Second, this chapter has investigated the notion of ``alternative structure'' from a relatively high level: the de®nitions of the candidate sets for interpretive and expressive optimizations each require input-output identity on one of two major structural components (syntactic and semantic, respectively). Intriguing research suggests some possible avenues of re®nement. For example, it might be the case that the equivalence classes for interpretive optimization are de®ned, not by shared syntactic structure per se, but rather by a common lexical basis (Chomsky 1995; see also Keer and Bakovic 1997) or perhaps by common clusterings of interpretable features (Nakamura 1997). Similarly, it could turn out that the semantic structures of this chapter are more properly identi®ed with LFs (Legendre, Smolensky, and Wilson 1998), in which case the candidate sets for expressive optimization might be de®ned by criteria other than truth-functional equivalence (see also Richards 1997: note 7). Particularly relevant to the present work are the proposals of Reinhart (1995:75), Fox (1995:105), and Golan (1993), according to which members of an equivalence class must share both lexical material and certain aspects of semantic interpretation. Comparison of the predic-

504

Colin Wilson

tions of the bidirectional optimization model with the predications of these alternative models and extension of the bidirectional model to empirical domains other than anaphora are projects for future research. Notes I would like to thank Eric BakovicÂ, Laura Benua, Luigi Burzio, Edward Flemming, Jane Grimshaw, Ed Keer, GeÂraldine Legendre, Peter Sells, Paul Smolensky, and Sten Vikner for very helpful feedback on this work. Special thanks to Luigi, GeÂraldine, and Sten for extensive comments on a previous draft. I would also like to thank the audiences and organizers of the Workshop on Optimality Theory and Cognition (Stanford, 1996), the HOT Conference/ Maryland Mayfest (1997), and the Second Workshop on Optimality Theory Syntax (University of Stuttgart, 1998). This work was partially supported by NSF grant IRI-9213894 to Legendre and Smolensky. 1. For expository convenience, in this introductory section I use the term sentence to refer, informally, to a syntactic structure together with its semantic interpretation. More speci®c representational assumptions are introduced later. 2. The judgments in (1) represent only one of the dialects/idiolects of Marathi that Dalrymple (1993) discusses. See section 3.2 for analysis of the dialectal/idiolectal variation. 3. The relativized minimality e¨ect on anaphor binding is also attested in Italian (Giorgi 1984), Chinese (Huang and Liu 1997), Icelandic (Pica 1987), and various other languages (Burzio 1992); see also Chung 1989 and section 15.3 on Chamorro. 4. Throughout this chapter, I use the terms binding and bound in the technical sense of bound anaphora (Reinhart 1983). 5. Here [n] spells the retro¯ex nasal, a point on which my transcription di¨ers from that of Dalrymple (1993). Note also that apaan is subject to case in¯ection and allomorphy, just like swataah, but I ignore this variation in form here. 6. I am very grateful to Sandy Chung for providing this example. 7. Rather, the unfaithful mapping associated with referential economy can only counterfeed the unfaithful mapping associated with relativized minimality, a point that will be important for connecting the present proposal with McCarthy's (1997) Sympathy Theory. Note that the alternative ordering of the optimizations (expressive before interpretive) does not solve the problem addressed in this chapter. References Anderson, S. R. 1986. Types of Dependencies in Anaphora: Icelandic (and Other) Re¯exives. In L. Hellan and K. Koch Christensen (eds.), Topics in Scandinavian Syntax, 65±88. Dordrecht: Reidel. Anderson, S. R. 1992. A-Morphous Morphology. In B. Comrie, R. Huddleston, R. Lass, D. Lightfoot, J. Lyons, P. H. Matthews, R. Posner, S. Romaine, N. V. Smith, and N. Vincent (eds.), Cambridge Studies in Linguistics. Cambridge, England: Cambridge University Press. Aoun, J., and Y.-H. Li. 1990. Minimal Disjointness. Linguistics 28: 189±204. Artstein, R. 1998. Hierarchies. Ms., Rutgers University, New Brunswick, N.J.

Bidirectional Optimization and the Theory of Anaphora

505

Bouchard, D. 1984. On the Content of Empty Categories. Dordrecht: Foris. Burzio, L. 1989. On the Non-Existence of Disjoint Reference Principles. Rivista di Grammatica Generativa 14: 3±27. Burzio, L. 1991. The Morphological Basis of Anaphora. Journal of Linguistics 27: 81±105. Burzio, L. 1992. The Role of the Antecedent in Anaphoric Relations. In R. Friedin (ed.), Current Issues in Comparative Grammar, 1±45. Dordrecht: Kluwer. Burzio, L. 1998. Anaphora and Soft Constraints. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 93±113. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, N. 1986. Knowledge of Language: Its Nature, Origin, and Use. Westport, Conn.: Praeger. Chomsky, N. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Chung, S. 1981. Transitivity and Surface Filters in Chamorro. In J. Hollyman and A. Pawley (eds.), Studies in Paci®c Languages and Cultures in Honour of Bruce Biggs, 311±332. Auckland: Linguistics Society of New Zealand. Chung, S. 1983. The ECP and Government in Chamorro. Natural Language and Linguistic Theory 1: 207±244. Chung, S. 1987. The Syntax of Chamorro Existential Sentences. In E. Reuland and A. ter Meulen (eds.), The Representation of (In)de®niteness, 191±225. Cambridge, Mass.: MIT Press. Chung, S. 1989. On the Notion ``Null Anaphor'' in Chamorro. In O. Jaeggli and K. Sa®r (eds.), The Null Subject Parameter, 143±184. Dordrecht: Kluwer. Dalrymple, M. 1993. The Syntax of Anaphoric Binding. CSLI Lecture Notes, No. 36. Stanford, Calif.: CSLI Publications. Fiengo, R., and R. May. 1994. Indices and Identity. Cambridge, Mass.: MIT Press. Fox, D. 1995. Condition C E¨ects in ACD. In R. Pensal®ni and H. Ura (eds.), MIT Working Papers in Linguistics 27: Papers on Minimalist Syntax, 105±119. Department of Linguistics and Philosophy, MIT, Cambridge, Mass. George, L., and J. Korn®lt. 1981. Finiteness and Boundedness in Turkish. In F. Heny (ed.), Binding and Filtering, 105±127. Cambridge, Mass.: MIT Press. Giorgi, A. 1984. Towards a Theory of Long Distance Anaphors: A GB Approach. Linguistic Review 3: 307±361. Golan, Y. 1993. Node Crossing Economy, Superiority, and D-Linking. Ms., Tel-Aviv University. Harbert, W. 1982. In Defense of Tense. Linguistic Analysis 19: 1±18. Hestvik, A. 1990. LF-Movement of Pronouns and the Computation of Binding Domains. Doctoral dissertation, Brandeis University, Waltham, Mass. Huang, C.-T. J. 1983. A Note on the Binding Theory. Linguistic Inquiry 14: 554±561. Huang, C.-T. J., and C.-S. L. Liu. 1997. Anaphoricity and Logophoricity in Mandarin Chinese. Paper presented at the Long-Distance Re¯exives Workshop, LSA Linguistic Institute, Cornell University, Ithaca, N.Y.

506

Colin Wilson

Jackendo¨, R. 1990. Semantic Structures. Cambridge, Mass.: MIT Press. Keer, E., and E. BakovicÂ. 1997. Have Faith in Syntax. In E. Curtis, J. Lyle, and G. Webster (eds.), Proceedings of the Sixteenth West Coast Conference on Formal Linguistics, 255±269. Legendre, G., W. Raymond, and P. Smolensky. 1993. An Optimality-Theoretic Typology of Case and Grammatical Voice Systems. In Joshua S. Gventer, Barbara A. Kaiser, and Cheryl C. Zoll (eds.), Proceedings of the Nineteenth Annual Meeting of the Berkeley Linguistics Society, 464±478. Berkeley Linguistics Society, University of California, Berkeley. Legendre, G., P. Smolensky, and C. Wilson. 1998. When Is Less More? Faithfulness and Minimal Links in Wh-Chains. In Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky (eds.), Is the Best Good Enough? Optimality and Competition in Syntax, 249±289. Cambridge, Mass.: MIT Press and MIT Working Papers in Linguistics. Maling, J. 1984, Non-Clause-Bounded Re¯exives in Modern Icelandic. Linguistics and Philosophy 7: 211±241. Maling, J. 1986. Clause-Bounded Re¯exives in Modern Icelandic. In J. Maling and A. Zaenen (eds.), Modern Icelandic Syntax, 277±287. Syntax and Semantics 24. San Diego: Academic Press. Manzini, M. R., and K. Wexler. 1987. Parameters, Binding Theory, and Learnability. Linguistic Inquiry 18: 413±444. McCarthy, J. 1997. Sympathy Theory: An Approach to Phonological Opacity in OT. Paper presented at the Hopkins Optimality Theory Workshop/University of Maryland Mayfest 1997, Baltimore. Montalbetti, M. 1984. After Binding. Doctoral disseration, MIT, Cambridge, Mass. Nakamura, Masanori. 1997. Object Extraction in Bantu Applicatives: Some Implications for Minimalism. Linguistic Inquiry 28: 252±280. Nichols, J. 1984a. Chechen. In R. Smeets (ed.), The Indigenous Languages of the Caucasus, vol. 4.2, 1±78. Delmar, N.Y.: Caravan Books. Nichols, J. 1984b. Ingush. In R. Smeets (ed.), The Indigenous Languages of the Caucasus, vol. 4.2, 79±145. Delmar, N.Y.: Caravan Books. O'Connor, M. C. 1992. Topics in Northern Pomo Grammar. In J. Hankamer (ed.), Outstanding Dissertations in Linguistics. New York: Garland. Pica, P. 1987. On the Nature of the Re¯exivization cycle. In Joyce McDonough and Bernadette Plunkett (eds.), NELS 17, 483±499. GLSA, University of Massachusetts, Amherst. Pollard, C., and I. Sag. 1992. Anaphors in English and the Scope of the Binding Theory. Linguistic Inquiry 23: 261±304. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. Progovac, L. 1994. Negative and Positive Polarity: A Binding Approach. In J. Bresnan, B. Comrie, W. Dressler, R. Huddleston, R. Lass, D. Lightfoot, J. Lyons, P. H. Matthews, R. Posner, S. Romaine, N. V. Smith, and N. Vincent (eds.), Cambridge Studies in Linguistics. Cambridge, England: Cambridge University Press.

Bidirectional Optimization and the Theory of Anaphora

507

Reinhart, T. 1983. Coreference and Bound Anaphora: A Restatement of the Anaphora Questions. Linguistics and Philosophy 6: 47±88. Reinhart, T. 1995. Interface Strategies. OTS Working Papers (OTS-WP-TL-95-002). Utrecht: Utrecht Institute of Linguistics, Utrecht University. Reinhart, T., and E. Reuland. 1993. Re¯exivity. Linguistic Inquiry 24: 657±720. Richards, N. 1997. Competition and Disjoint Reference. Linguistic Inquiry 28: 178±187. Rizzi, L. 1990. Relativized Minimality. Cambridge, Mass.: MIT Press. Samek-Lodovici, V. 1996. Constraints on Subjects: An Optimality Theoretic Analysis. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Smolensky, P. 1996. On the Comprehension/Production Dilemma in Child Language. Linguistic Inquiry 27: 720±732. Tesar, B. and P. Smolensky. 1998. Learnability in Optimality Theory. Linguistic Inquiry 29(2): 229±268. ThraÂinsson, H. 1976. Re¯exives and Subjunctives in Icelandic. In NELS 6, 225±239. GLSA, University of Massachusetts, Amherst. ThraÂinsson, H. 1979. On Complementation in Icelandic. New York: Garland. ThraÂinsson, H. 1997. Long-Distance Re¯exives: Division of Labor between Morphology, Syntax, and Discourse. Paper presented at the Long-Distance Re¯exives Workshop, LSA Linguistic Institute, Cornell University, Ithaca, N.Y. Topping, D. 1973. Chamorro Reference Grammar. Honolulu: University of Hawaii Press. Vikner, S. 1995. Parameters of Binder and Binding Category in Danish. In Working Papers in Scandinavian Syntax 23: 1±58. Wali, K. 1976. Two Marathi Re¯exives and Their Implication for Causative Structure. Doctoral dissertation, Syracuse University, Syracuse, N.Y. Wali, K., and K. V. Subbarao. 1991. On Pronominal Classi®cation: Evidence from Marathi and Telugu. Linguistics 29: 1093±1110.

Chapter 16 Case Patterns

Ellen Woolford

Standard Case-licensing principles (see Chomsky 1981, 1986, 1995) are designed to handle situations where the Case-licensing feature of the relevant local head alone determines what Case an argument will have. However, there are a range of problematic situations wherein relying solely on the features of the local head to determine Case is insu½cient. These problematic situations, which are the focus of this chapter, can be handled if violable markedness and faithfulness constraints are added to Case Theory. These problematic phenomenon can be classed into two types: dependency and valency e¨ects. Situations where the Case of one argument depends on the Case of another are dependency e¨ects. For example, the Case of objects in Icelandic and Hindi depends on the Case of the subject (Yip, Maling, and Jackendo¨ 1987; Mahajan 1990). Burzio's (1986) generalization describes another dependency e¨ect. In valency e¨ects, the Case of an argument depends on whether the verb is transitive. Many ergative languages limit ergative subjects to transitive clauses, and a similar valency restriction can occur with dative subjects. Previous attempts to solve such problems within standard Case theory have involved placing restrictions on accusative Case licensing to explain why some objects cannot be accusative (e.g., Burzio 1986; Woolford 1993). An alternate approach proposed by Yip, Maling, and Jackendo¨ (1987) is to order the principles that associate Cases with arguments, so that nominative is assigned before accusative. Legendre, Raymond, and Smolensky (1993) and Nakamura (1997, 1999) take the latter strategy within Optimality Theory (OT), varying the ranking among the principles that associate Cases with arguments to account for di¨erences in Case patterns across languages. These initial OT results provide a simple account of certain dependency and valency e¨ects, without complicating the Case-licensing principles. Nevertheless, ranking the Case-licensing principles is not the only means of accomplishing this goal. The approach proposed here maintains the principles of the Case theory of Chomsky (1981, 1986, 1995) as universal and inviolable, accounting for dependency and valency e¨ects, and other crosslinguistic di¨erences in Case patterns

510

Ellen Woolford

with a supplementary set of ranked violable markedness and faithfulness constraints. Whenever there is a choice of licensed Cases for a particular argument (a situation that occurs frequently), the violable constraints determine which Case will surface. Under this approach, dependency e¨ects follow from markedness. All the contexts in which it previously appeared necessary to restrict accusative Case licensing (by Burzio's (1986) generalization or Woolford's (1993) Accusative Case Blocking) turn out to be contexts in which nominative Case is potentially licensed on the object as well as accusative, and nominative is selected as the less marked Case. Valency e¨ects and other dependency e¨ects involve the interaction of various markedness and faithfulness constraints, including faithfulness constraints that are contextually restricted.1 These same constraints account for the surface inventory of Cases used in a language.2 This chapter is organized as follows. Section 16.1 lays out the universal principles of Case Theory retained from Chomsky (1981, 1986, 1995). Because the entire burden of determining which Case will surface is no longer carried by these Case-licensing principles, they can be formulated in a maximally simple and restrictive form. If they overgenerate, the overgeneration is curbed by the violable constraints. This section discusses the kinds of universal generalizations that these Case-licensing principles capture, motivating the decision to keep them as inviolable principles. Section 16.2 presents an overview of the violable constraints used in the chapter. Markedness constraints require avoidance of marked Cases (e.g., *Ergative, *Dative, *Accusative, and so on). There are some universally ®xed ranking relations among these constraints: inherent Cases are universally more marked than structural Cases, and accusative is universally more marked than nominative. An additional markedness constraint prohibits multiple occurrences of the same Case. Faithfulness constraints require lexically determined inherent Case-licensing features of verbs to be checked. (There is no literal faithfulness to the Case of the input, because arguments have no Case in the input; Gen adds Case in the output candidates.3) The faithfulness constraints have variant contextually restricted forms, requiring faithfulness to be obeyed only in particular contexts, such as transitives or perfective aspect. Section 16.3 begins the analysis of various valency, dependency, and inventory problems, focusing on those involving the dative Case. These e¨ects, as well as the behavior of unaccusatives that motivated Burzio's Generalization, are shown to follow from the interaction of the violable markedness and faithfulness constraints. Section 16.4 deals with the same range of problems with respect to ergative Case, showing that the same solution works for ergative constructions. This section also deals with ergative-dative patterns, arguing that these provide empirical evidence for the approach proposed here, and against any approach that ranks Case-licensing

Case Patterns

511

rules. Section 16.5 is a brief look at how to deal with other sorts of dependency e¨ects. 16.1

The Universal Case-Licensing Principles

This approach retains much of the Case theory of Chomsky (1981, 1986, 1995), including the requirement that every argument have a licensed Case. Any candidate violating these principles will be eliminated from the candidate set before the violable constraints apply. 16.1.1 Licensing Structural Case Structural Case is licensed on an argument in the proper structural relationship with the appropriate licensing head. Nominative is licensed by Tense and accusative is licensed by V. The proper structural relationship includes Spec-head relationships, plus any situation where the head c-commands the argument in a su½ciently local relationship to allow Case checking by feature raising (Chomsky 1995). (Covertly raising just the features of an argument for checking replaces LF movement of the entire argument in Chomsky 1986.) The feature-raising option allows nominative Case to be licensed on objects, as well as on subjects, which is necessary for nominative objects in dative subject constructions (e.g., Icelandic).4 The standard assumption that Tense can only license one nominative prevents the subject and object of a single clause from both being nominative.5 Accusative Case cannot be licensed on an external subject, but it can be licensed on a VP-internal subject. Previously, it appeared necessary to restrict accusative Case licensing, to prevent accusative licensing of unaccusative subjects (e.g., Burzio 1986; Woolford 1993), but now no such restrictions are needed. The violable constraints will prevent accusative Case from surfacing on an unaccusative subject. The universal Case-licensing principles express certain apparently universal restrictions on nominative and accusative Case licensing, and that fact motivates retaining these principles as inviolable. In addition to the impossibility of structural accusative on external subjects (outside ECM constructions), we do not ®nd (to my knowledge) that nominative Case occurs inside PPs in any language, even though nominative is the least marked Case in other situations. It is not possible to move a nominative Case feature out of a PP up to Tense for Case checking because of locality considerations that prevent feature raising out of a PP. These structural Caselicensing principles also prevent nominative and accusative Case from occurring in NPs (unless there is a clausal origin), because NPs have no Tense and nouns are [ V]. There have also been claims in the literature for structurally licensed inherent Cases such as dative; this possibility is discussed in the next section, on inherent Case licensing.

512

Ellen Woolford

16.1.2 Licensing Inherent Case Inherent Cases (also called lexical or quirky Cases) can only be licensed by heads that assign a theta role to the argument in question (Chomsky 1995). There is a fairly strong correlation between particular theta roles and particular inherent Cases, although this correlation is far from perfect; see (1). (1) Inherent cases and thematic roles loosely correlates with dative loosely correlates with ergative loosely correlates with inherent acc

goals and experiencers agents themes

For arguments that ergative Case belongs on the list of inherent Cases, see Woolford 1997. It has been suggested that dative Case is structurally licensed in certain situations (Czepluch 1988; Broekhuis and Cornips 1994; Webelhuth 1995).6 We will thus consider the possibility of structurally licensed datives and ergatives in the candidate sets discussed in this chapter, although it is not clear exactly how such Cases would be licensed. Although inherent Cases are licensed in conjunction with theta-role assignment, the mere ability to assign a thematic role to an argument does not necessarily give a verb the ability to license an inherent Case. For example, like takes a dative subject in Icelandic but love does not, and eat takes an ergative subject in Hindi but bring does not (Comrie 1984). Thus the ability to license inherent Case must be part of a verb's lexical entry (Mohanan 1982; Zaenen and Maling 1984; Zaenen, Maling, and ThraÂinsson 1985).7 Nevertheless, it is assumed here that all languages are alike in having a class of verbs that can license dative and ergative Case, even though the membership of those classes may di¨er somewhat across languages. If those Cases do not surface in a particular language, it is due to the violable constraints. 16.2

Overview of the Violable Constraints

When the Case-licensing principles do not uniquely determine what Case a particular argument will have, the violable constraints determine which licensed Case will surface. The idea that one argument can be potentially licensed for more than one Case has long been part of Case theory, but only with respect to potentially licensing both an inherent Case and a structural Case on a single argument. In Chomsky 1981, 1986, the overlap in inherent and structural Case-licensing principles was handled by postulating a serial order in the Case assignment process, so that inherent Cases were assigned earlier (in D-structure), before the structural Case-licensing principles had a chance to apply (at S-structure). Under the assumption that D-structure Case could never be changed, that approach guaranteed that a licensed inherent Case

Case Patterns

513

would always surface. That approach is less appealing now that D-structure has disappeared from the theory (Chomsky 1995). Moreover, there is evidence that lexically licensed Cases are not actually guaranteed to surface. Under the approach advocated here, there are many situations in which more than one Case is potentially licensed on the same argument and the violable constraints determine which of these licensed Cases will surface. In a contest between an inherent and a structural Case, the relative ranking of faithfulness constraints (favoring inherent Cases) versus markedness constraints (favoring structural Cases) determines the outcome. In situations in which two di¨erent structural Cases are licensed on one argument, markedness constraints make the decision. 16.2.1 Markedness Constraints Paralleling markedness constraints in phonology, such as *voice, it is proposed here that there is a series of markedness constraints pertaining to Case, with the form of *Accusative, *Dative, *Ergative, and so on.8 Among these constraints, there are certain universally ®xed rankings. Nominative is generally agreed to be the least marked Case (e.g., Dixon 1979; Tsunoda 1981), and nominative does appear to surface in all languages.9 Thus *Nominative is always ranked below all other marked Case constraints.10 Inherent Cases are much more likely to be morphologically marked than structural Cases, suggesting that inherent Cases are universally more marked than structural Cases. This view is consistent with Grimshaw's suggestion (chap. 8, this volume) that *Dative is universally ranked above *Accusative.11 Among the inherent Cases, there do not appear to be any universal markedness rankings. An OCP-type markedness constraint prohibits two instances of the same Case in the same clause. Such constraints have existed in the literature for years as languagespeci®c constraints (e.g., the Double-o constraint of Japanese (Harada 1973; Kuroda 1992)); these can now be elevated to the status of universal but violable constraints. (See Yip 1995 for a general survey of OCP constraints in syntax, morphology, and phonology.) (2) OCP

case :

prohibits two (adjacent?) instances of the same Case

16.2.2 Faithfulness Constraints With no Cases present in the input, there is no possibility of literal faithfulness to the input Case of an argument. However, we need some kind of faithfulness constraint to allow the more marked inherent Cases to surface when the verb carries a lexical feature licensing them. That is, we need a faithfulness constraint to do the job that was done in Chomsky 1981 by claiming that the inherent Cases were assigned at D-structure,

514

Ellen Woolford

giving them priority over the structural Cases assigned at S-structure. Under this approach, there is a family of faithfulness constraints to do this job that require faithfulness to (checking of ) any lexically speci®ed inherent Case-licensing features carried by verbs. (3) FaithLex: A lexically speci®ed inherent Case-licensing feature must be checked. In phonology, faithfulness constraints have variants that hold only in restricted contexts, such as roots, syllable onsets, or stressed syllables (see Beckman 1998). In syntax, inherent Cases often surface only in certain contexts, such as transitive constructions or perfective constructions. Such contextual restrictions will be accounted for here by means of contextually restricted versions of FaithLex.12 (4) Contextually restricted Faithfulness Constraints FaithLextrans : FaithLex holds only in transitive clauses. FaithLexperf : FaithLex holds only in perfective clauses. To simplify the calculation of the typological predictions of this chapter, it will be assumed here that there is a universally ®xed ranking such that the restricted versions of FaithLex are always ranked above the general version. While the opposite ranking will make no empirical di¨erence here, some work in phonology shows that it can make an empirical di¨erence under certain circumstances (Prince 1997), and that the opposite ranking of general and speci®c constraint variants may be needed in some languages (Keer 1999). Let us now turn to the analysis of the valency, dependency, and inventory e¨ects that require these violable constraints. 16.3

Two Types of Dative Subject Languages

Languages that allow dative subjects fall into two types: one type shows a valency e¨ect, prohibiting dative subjects in intransitive clauses, while the other type does not, allowing dative subjects in intransitive clauses as well as in transitives (Woolford 1997). Japanese is an example of the ®rst type (Shibatani 1977), while Icelandic is an example of the second. (5) Japanese: prohibits dative subjects in intransitive clauses13 a. *Akatyan-ni moo arukeru. baby-dat already walk-can `The baby can walk already.' (Shibatani 1977:807) b. Akatyan-ga moo arukeru. baby-nom already walk-can `The baby can walk already.' (Shibatani 1977:807)

Case Patterns

515

(6) Japanese: allows dative subject in transitive clauses a. Taroo-ni eigo-ga hanaseru. Taro-dat English-nom (*acc) speak-can `Taro can speak English.' (Shibatani 1977:806) b. Sensei-ni okane-ga ir-u. teacher-dat money-nom need-pres `The teacher needs money.' (Shibatani 1977:799) (7) Icelandic: allows dative subjects in intransitives and transitives a. BaÂtnum hvolfdi. boat-dat capsized `The boat capsized.' (Levin and Simpson 1981 (1b)) b. Barninu batnaDi veikin. child-dat recovered-from disease-nom (*acc) `The child recovered from the disease.' (Yip, Maling, and Jackendo¨ 1987:223) The Icelandic pattern, where all lexically licensed Cases are preserved on the surface, is the pattern that we expect under the Case theory of Chomsky 1981, but Japanese is problematic. To deal with that problem, Shibatani (1977:807) proposes a language-speci®c constraint for Japanese, requiring a nominative Case to be present in every matrix clause. Shibatani speci®es that a dative Case on an intransitive subject is changed to a nominative. Shibatani's view can be translated into OT terms as a claim that faithfulness to lexically licensed Cases is violable. Tsunoda (1981) reformulates Shibatani's constraint to capture the intuition that what is going on in Japanese is caused by markedness. In addition, Tsunoda elevates this constraint from a language-speci®c constraint for Japanese to a universal constraint (although he notes that there are exceptions to it). Tsunoda's reformulation, the Unmarked Case Constraint, requires that every sentence in every language have an NP in the unmarked Case (nominative or absolutive). Tsunoda's contribution thus takes two important steps in the right direction that we want to preserve. First, he explains why nominative is selected over dative, because nominative is the least marked Case, and second, he elevates this constraint to a universal (with exceptions). Translating these insights into OT terms, we want the Japanese valency restriction on intransitives to follow from universal, but violable, markedness constraints. The markedness constraints we need for this data, and their universally ®xed ranking, are given in (8). (8) *Dative g *Accusative g *Nominative

516

Ellen Woolford

The distribution of lexically licensed dative Case in intransitive constructions in Japanese and Icelandic follow from the two possible relative rankings of *Dative and FaithLex, the faithfulness constraint that requires lexically licensed inherent Cases to surface.14 (9) a. Japanese: *Dative g FaithLex b. Icelandic: FaithLex g *Dative Before we look at the tableaux illustrating how the above rankings produce the desired results, let us brie¯y discuss the candidate set to which these constraints apply. 16.3.1 The Candidate Set Although Gen will generate candidates containing arguments with no Case, as well as candidates with unlicensed Cases, these candidates are removed from the candidate set before the violable constraints apply, because they violate the universal principle requiring all arguments to have a licensed Case. With respect to the intransitive constructions we are considering here, the sole argument can be licensed for nominative Case or for accusative Case (since the argument is not an agent) and for dative, if the input verb carries an inherent Case-licensing feature for dative (or if there is a structurally licensed dative). 16.3.2 Datives in Intransitive Constructions Tableau T16.1 shows the three possible output candidates with these licensed Cases. In Icelandic, all lexically licensed dative Cases surface because faithfulness takes precedence over markedness. The nominative and accusative candidates in (b) and (c) are both eliminated by FaithLex, because the lexical Case-licensing feature of the verb is not checked. That leaves the dative candidate in (a) as the only remaining candidate and thus the winner of the competition. The fact that (a) violates the lowerranked markedness constraint, *Dative, is irrelevant, because the game is over before Tableau T16.1 Icelandic intransitives (verb carries a feature-licensing dative Case) Input: NP V-[‡dat]

FaithLex

f a. NP-dat V-[‡dat]

*Dative

*Accusative

*Nominative

*

b. NP-nom V-[‡dat]

*!

c. NP-acc V-[‡dat]

*!

* *

Case Patterns

517

that lower-ranked constraint has a chance to in¯uence the outcome (as indicated by the shading). If the verb happens to be one that does not lexically license dative Case, the result is quite di¨erent (see tableau T16.2). Then the decision is made entirely by the markedness constraints, and nominative wins because it survives after the more marked dative and accusative candidates are eliminated. (If there is an (a) candidate, it has a structurally licensed dative.) FaithLex plays no role because there is no lexical licensing feature on the verb. Let us turn now to Japanese, where the ranking of FaithLex and *Dative is reversed (see tableau T16.3). The higher-ranked *Dative eliminates the dative candidate in (a) before the lower-ranked FaithLex has a chance to apply. FaithLex has no e¨ect on the outcome because both remaining candidates violate it, since they do not check the lexical Case feature of the verb. The accusative candidate in (c) is eliminated by *Accusative, leaving the nominative candidate in (b) as the winner. The same result occurs with verbs that do not carry a dative-licensing feature. The dative candidate (which could only have a structurally licensed dative) is eliminated ®rst. FaithLex has no e¨ect, but this time it is because none of the candidates violate Tableau T16.2 Icelandic intransitives (verb has no inherent Case-licensing feature) Input: NP V

FaithLex

*Dative

a. NP-dat V

*Accusative

*Nominative

*!

f b. NP-nom V

*

c. NP-acc V

*!

Tableau T16.3 Japanese intransitives Input: NP V-[‡dat] a. NP-dat V-[‡dat] f b. NP-nom V-[‡dat] c. NP-acc V-[‡dat]

*Dative

FaithLex

*Accusative

*Nominative

*! *

* *

*!

518

Ellen Woolford

it. Finally, as above, *Accusative makes the ®nal decision, leaving the nominative candidate the winner. Because this account makes the correct prediction for Japanese intransitives, even if there are verbs that lexically license datives, it is not necessary to assume that there is an accidental gap in the inventory of verb classes in Japanese such that Japanese simply happens to lack any intransitive verbs that have the ability to license dative Case. 16.3.3 Burzio's Generalization We now have an account for (a portion of ) Burzio's Generalization. That is, we now have an account for why, regardless of the ranking of constraints, unaccusative verbs never appear with a structural accusative argument. Although the argument of an unaccusative verb is potentially licensed for either nominative or accusative Case, nominative wins because of markedness.15 (See tableau T16.4.) We have now seen how the ranking of these simple faithfulness and markedness constraints produces the two types of intransitive patterns in dative subject languages; let us turn to transitive constructions. 16.3.4 Transitive Clauses with Dative Subjects Despite their di¨erence in intransitive constructions, Icelandic and Japanese share a common set of Case patterns in transitive clauses; see (10). (10) a. Japanese Case Patterns intrans: nominative *Dative transitive: nominative-accusative nominative-dative dative-nominative *Dative-Accusative

b. Icelandic Case Patterns intrans: nominative dative transitive: nominative-accusative nominative-dative dative-nominative *Dative-Accusative

We need to answer two questions about these patterns. First, why does Japanese allow datives in transitives but not intransitives? Second, what produces the depenTableau T16.4 Unaccusatives Input: V NP a. V NP-acc f b. V NP-nom

*Accusative

*Nominative

*! *

Case Patterns

519

dency e¨ect observed in both languages, wherein the Case of the object depends on the Case of the subject? We will see that the Icelandic pattern is just what we expect given the constraints discussed for intransitives. If a verb carries a lexically speci®ed licensing feature for the dative, that dative will surface due to the highly ranked faithfulness constraint, FaithLex. Markedness then takes over to pick the least marked Case for those arguments that do not get a lexically licensed Case: whenever there is a choice of nominative or accusative on an object, markedness always picks the nominative. The interesting problem is how to get the identical transitive pattern in Japanese, where the markedness constraint *Dative is ranked above the faithfulness constraint, FaithLex. Why does dative ever get to surface in Japanese? Intuitively, there are two possible answers to this question. One is that nominative and dative ®ght over the single argument in an intransitive, but both can win in transitives because there are two arguments. That intuition is the basis of all previous approaches to this problem, including the one in the version of this chapter presented at the Hopkins Optimality Workshop/Maryland Mayfest 1997. The second intuitive answer is that datives surface in Japanese transitives for the same reason that they surface in Icelandic transitives: faithfulness, despite the lack of faithfulness in intransitives. We will see that while both types of approaches can account for the Case patterns listed above, they make di¨erent predictions for what can happen crosslinguistically, and the second intuition appears correct. Let us now examine the details of the analysis of transitive constructions in Icelandic and Japanese. 16.3.4.1 Transitive Clauses in Icelandic We will begin with Icelandic transitive verbs that carry a lexical feature licensing a dative subject. In the lexical entry of such a verb, the experiencer role is marked to take dative Case; see (11). (11) Barninu batnaDi veikin. child-dat recovered-from disease-nom The child recovered from the disease. (Yip, Maling, and Jackendo¨ 1987:223) (12) verb hExperiencer, Themei | [‡Dative] The candidate set will include both dative and nominative subjects, and both nominative and accusative objects, since all these Cases can be licensed. (If we add candidates with a structurally licensed dative on the object, that would add candidates with a dative-dative pattern, as well as a nominative-dative pattern. However, these would be eliminated by high-ranking constraints, as shown below.)

520

Ellen Woolford

Tableau T16.5 Icelandic (input verb licenses dative subject) Input: V[‡dat sub] NP NP

FaithLex

f a. NP-dat NP-nom

*Accusative

*

b. NP-dat NP-acc c. NP-nom NP-acc

*Dative

* *!

*Nominative *

*! *

*

(13) a. dative-nominative b. dative-accusative c. nominative-accusative Now, let us submit these candidates to the two ranked constraints we used above for Icelandic intransitives. We know from the behavior of intransitive constructions that FaithLex is ranked above *Dative in Icelandic, because datives survive in intransitives. *Accusative is universally ranked below *Dative, because structural Cases in general are less marked than inherent Cases. In tableau T16.5 FaithLex rules out any candidate that does not check the dativelicensing feature of the verb against a dative feature on the subject of this verb. That eliminates the nominative-accusative pattern in (c). The remaining candidates (a) and (b) tie on the next constraint, *Dative, because they both have a dative subject. What makes the ®nal decision in this situation is *Accusative, which eliminates (b), leaving the dative-nominative pattern in (a) as the winner. (The verb is omitted from the candidates to save space, but it is important to remember that the verb in each output candidate carries a lexical Case-licensing feature that needs to be checked if FaithLex is to be obeyed.16) The intuitive explanation for why the dative-nominative pattern wins here is clear. In Icelandic, lexically licensed datives always surface. The decision between any remaining Case patterns in made by markedness. That is the source of the dependency e¨ect. When the subject is dative, and both nominative and accusative are available to the object, the less marked nominative wins.17 If a verb lexically licenses dative Case on its object instead of its subject, the situation is essentially the same, so no tableau will be shown here. For verbs that do not carry any lexically speci®ed inherent Case-licensing feature, candidate (c) with the nominative-accusative pattern will not be eliminated by FaithLex, in contrast to what happens in tableau T16.5. If candidates (a) and (b) are present, they would have to have structurally licensed datives, but those candidates would be eliminated by *Dative, leaving the nominative-accusative pattern in (c) the winner. Let us now turn to transitive clauses in Japanese.

Case Patterns

521

16.3.4.2 Transitive Clauses in Japanese In contrast to Icelandic, where the same two constraints we needed for intransitives automatically produce the right result for transitives, Japanese initially seems to present a serious problem. Above, we ranked *Dative above FaithLex in Japanese to prevent dative subjects from surfacing in intransitive clauses. But if we try to analyze transitive clauses with just these two constraints, ranked in the same order, we would incorrectly predict that *Dative would also eliminate datives in transitive clauses, before the lower-ranked FaithLex has a chance to preserve them. In fact, that ranking of those two constraints produces the pattern we see in English, where datives never get a chance to surface, even if there are verbs that can lexically license them. (See tableau T16.6.) To produce the Japanese transitive pattern, some higher constraint must eliminate the nominative-accusative pattern in (c) before *Dative can eliminate the dativenominative pattern that actually surfaces. In an earlier version of this chapter, it was proposed that the higher constraint in Japanese is *Accusative, which would eliminate the (b) and (c) candidates, leaving (a) as the winner. While that gets the right result here, it con¯icts with Grimshaw's (chap. 8, this volume) suggestion that *Dative is universally ranked above *Accusative (and that, more generally, all inherent Cases are universally more marked than structural cases). Moreover, that approach encounters some empirical problems that it shares with previous OT approaches, which we will discuss below. The alternative proposed here is that faithfulness preserves the dative in Japanese transitives, just as it does in Icelandic. Although inherent Cases are universally more marked than structural Cases, a pattern with an inherent Case (e.g., dativenominative) can beat out a pattern with only structural Cases (e.g., a nominativeaccusative pattern) if a higher-ranked faithfulness constraint favors the inherent Case. The additional faithfulness constraint we need here is a contextually dependent version of the FaithLex constraint. FaithLextrans requires checking of the lexical Case-licensing feature of transitive verbs (de®ned here as verbs with more than one argument). In Japanese, FaithLextrans is ranked above *Dative, so that Japanese transitive constructions act like Icelandic transitive constructions in preserving lexically licensed Cases. Tableau T16.6 English transitives Input: NP NP V[‡dat subj]

*Dative

a. NP-dat NP-nom

*!

b. NP-dat NP-acc

*!

f c. NP-nom NP-acc

FaithLex

*

522

Ellen Woolford

(14) Japanese ranking: FaithLextrans g *Dative g FaithLex Let us see how this works in tableau T16.7. The nominative-accusative candidate in (c) is eliminated by FaithLextrans because that candidate has an unchecked lexical Case-licensing feature on the transitive verb. The remaining candidates tie on *Dative, so that constraint has no e¨ect. The decision is made by *Accusative, which eliminates the dative-accusative pattern, leaving the dative-nominative pattern as the winner. When the input verb does not carry any lexical Case-licensing feature that needs to be checked, FaithLex will not eliminate any candidates. If the (a) and (b) candidates are present in the candidate set, with a structurally licensed dative, these would both be eliminated by *Dative. That would leave candidate (c), the nominative-accusative candidate, as the winner. 16.3.4.3 Theory Comparison The Icelandic and Japanese facts considered so far have been handled by three di¨erent OT approaches. Two of these, Nakamura 1977, 1999 and the approach in the earlier version of the present chapter given at the Hopkins Optimality Workshop/Maryland Mayfest 1997, make the prediction stated in (15). (15) Prediction of previous approaches If a language requires a nominative in intransitives, that language will also require a nominative in transitives. Nakamura's (1997) approach makes this prediction because it ranks the rule assigning nominative Case ®rst, to assign nominative in all intransitives. But that means that nominative assignment must also be ®rst in transitives as well, predicting that all transitives will have a nominative. This same prediction is made by the approach in an earlier version of this chapter, because if faithfulness is ranked low in intransitives, so that markedness constraints select a nominative, then faithfulness must also be ranked low in transitives, and markedness constraints should again select a nominative for some argument. In contrast, the contextual faithfulness approach advocated here makes a di¨erent prediction. Regardless of what happens in intransitives, a highly ranked contextually restricted variant of the faithfulness constraint will preserve inherent Cases in transitives, whether or not a nominative is present. (16) Prediction of this approach If a language allows inherent Cases in transitives, their realization will not depend on the presence of a nominative. One way to test the predictions of these approaches is to ®nd a situation where the verb is transitive, but yet there is only one argument that takes Case. Such a situation

c. NP-nom NP-acc

b. NP-dat NP-acc

f a. NP-dat NP-nom

Input: NP NP V[‡dat subj]

Tableau T16.7 Japanese transitives

*!

FaithLextrans

*

*

*Dative

*

FaithLex

*

*!

*Accusative *

*Nominative

Case Patterns 523

524

Ellen Woolford

occurs when a transitive verb takes a clausal object. Assuming that the clausal object does not get Case, the subject is predicted to be nominative under the previous approaches, because it is the only argument getting Case ( just as in an intransitive). In contrast, the contextual faithfulness approach predicts that the subject will be dative (if the verb lexically licenses a dative), because the construction is transitive (there are two arguments). An example of such a construction in Japanese is given in (27) (from Mariko Sugahara, personal communication), and we see that a dative subject can occur with a clausal object, as expected under the contextual faithfulness approach. (17) Mariko-ni(wa) [watashi-ga nihongo-o hanas-anai-to] omo-eru. Mariko-dat(topic) I-nom Japanese-acc speak-neg-comp think-able Mariko thinks that I do not speak Japanese. (or, It seems to Mariko that I do not speak Japanese.) Under the two previous approaches, one would have to claim that the clausal object must have nominative Case. Stronger evidence that the contextual faithfulness approach makes the correct prediction crosslinguistically will be presented in section 16.4. There we will see an ergative language that requires a nominative in all intransitive clauses, but nonetheless allows transitive clauses with an ergative subject and dative object (and no nominative). 16.3.5 Typology Let us now calculate the di¨erent types of languages that are predicted, based on the possible rankings of these violable constraints. Although we have used ®ve violable constraints thus far, several of these are ranked in universally ®xed orders, severely restricting the possible rankings; see (18). (18)

Universal rankings *Dative g *Accusative g *Nominative FaithLextrans g FaithLex

We will focus here on the possible rankings of *Dative with respect to the two faithfulness constraints.18 When both faithfulness constraints are above *Dative, we get Icelandic, where all datives surface. When the faithfulness constraints are both below *Dative, we get English, where no datives surface. (Note that under this approach, it is not necessary to stipulate that the dative Case is absent from the Case inventory of English, or that English verbs lack the ability to license dative subjects or objects. Any dative licensed by an English verb would never be allowed to surface, given this ranking.) Finally, when *Dative is ranked between the two faithfulness constraints, we get Japanese, where datives only surface in transitives.

Case Patterns

(19)

525

Typology (Icelandic pattern) g *Dative FaithLextrans g FaithLex g FaithLextrans g FaithLex (English pattern) *Dative g FaithLex (Japanese pattern) FaithLextrans g *Dative

Note that there is no way to rank these constraints that would produce a language that allows dative subjects only in intransitive clauses, but not in transitive ones. That is a desirable property because that logically possible type of language apparently never occurs.19 To conclude this section, let us take a brief look at ditransitives. 16.3.6 Ditransitives Let us examine the predictions of the above rankings in Icelandic and Japanese for the Case patterns in ditransitive constructions, focusing on how to account for the very regular appearance of datives in ditransitives. Ditransitive constructions are much more likely to contain a dative than transitives are, and the question is whether we need to add anything to this approach to account for this fact. One can think of several reasons for an increase in datives in ditransitives. First, ditransitives are more likely to contain a goal argument than transitives are, simply because there are more arguments present. With three arguments, there has to be something besides just an agent and a theme, and goal is a very likely choice for the third role. Since dative Case is associated with the goal role, that provides more opportunity for lexically licensed datives. A second, similar reason that ditransitives are more likely to contain a dative than transitives is because there is an additional argument that needs Case. In languages that prohibit two accusatives, there has to be a third Case in ditransitives besides nominative and accusative and dative is the usual choice. The question is what is the source of this dative? Is it always lexically licensed? To put this discussion on a more concrete footing, let us examine the predictions of the above constraint rankings for Icelandic and Japanese with respect to ditransitives. For Icelandic, where FaithLex is ranked above *Dative, any lexically licensed dative will surface in ditransitives. The same is true for Japanese, where FaithLextrans is ranked above *Dative. Let us examine tableau T16.8 to see this in Japanese. If the input verb carries a feature licensing dative on its goal argument, FaithLextrans will eliminate any candidate without a dative goal. The interesting question is whether there are any Japanese or Icelandic ditransitive verbs that do not license dative Case on the goal. The standard view has been that ditransitive verbs in these languages all lexically license dative Case on goals, perhaps due to the operation of a regular lexical rule (Yip, Maling, and Jackendo¨ 1987). That may be true, but let us at least ask what would happen if there are some ditransitive verbs in these languages that do not lexically license datives.

526

Ellen Woolford

Tableau T16.8 Japanese ditransitive (where the verb lexically licenses a dative goal) Input: agent goal theme V[‡dat goal]

FaithLextrans

f a. agent-nom goal-dat theme-acc b. agent-nom goal-acc theme-acc

*Dative

*Accusative

*

*

*!

**

Tableau T16.9 Ditransitive (with no lexically licensed cases) Input: agent goal theme V

FaithLextrans

a. agent-nom goal-dat theme-acc

*Dative

*Accusative

*!

*

f b. agent-nom goal-acc theme-acc

**

Tableau T16.10 Japanese ditransitive (without a lexically licensed dative goal) Input: agent goal theme V

OCP-case

f a. agent-nom goal-dat theme-acc b. agent-nom goal-acc theme-acc

FaithLextrans

*Dative *Accusative *

*!

* **

With no other constraints operating (and assuming that all languages can license two accusatives), one might expect that a ditransitive verb that licenses no inherent Cases would end up with a Case pattern like English ditransitives, with a nominativeaccusative-accusative pattern. That pattern is what emerges when markedness alone selects the Cases in a ditransitive, because dative is more marked than accusative. (See tableau T16.9.) However, in Japanese, double-accusative constructions are generally ruled out by an OCP-like constraint, known in the literature as the Double-o constraint (Shibatani 1973).20 (20) OCP

acc

prohibits two (adjacent) accusatives

By ranking this OCP constraint above *Dative in Japanese, the (b) candidate is eliminated before *Dative can eliminate the (a) candidate. (See tableau T16.10.)

Case Patterns

527

While this is the right result for Japanese ditransitives, the question is how the dative in the winning candidate is licensed. Unless there is a lexical rule that gives all ditransitive verbs the ability to lexically license dative Case on their goals, a ditransitive verb that does not lexically license a dative must have a structurally licensed dative on its goal. 16.3.7 Conclusions In this section, we have seen how valency e¨ects, dependency e¨ects, and inventory e¨ects involving the dative Case can be handled if standard Case-licensing principles are supplemented with a small set of markedness and faithfulness constraints. The dependency e¨ect is the result of markedness. When the subject is dative, both nominative and accusative Case are potentially licensed on the object; but because nominative is less marked, a nominative object is always selected over an accusative one. The valency e¨ect involves both markedness and faithfulness. Nominative is favored over dative when *Dative is ranked above FaithLex. The situation in which datives are allowed in transitives but not intransitives is due to the e¨ect of a higherranked, contextually restricted faithfulness constraint that holds in transitive contexts. The inventory e¨ect also involves both markedness and faithfulness. Above we saw that whether or not datives surface in a language may depend on the relative ranking of *Dative and the faithfulness constraints that preserve lexically licensed datives, or on the relative ranking of the OCP constraint prohibiting double accusatives. In the next section, we turn to ergative languages. We will see that this same OT approach correctly predicts the basic types of languages that occur with respect to the distribution of the ergative Case. 16.4

Types of Ergative Languages

The two basic types of ergative languages are the classic type (limiting ergative Case to transitives) and the active-stative type (where ergatives also occur in intransitives, on the more agentive subjects). An example of the classic type is Niuean, where intransitives can only have nominative subjects, as in (21), but transitives may have ergative subjects, at least when the subject is an agent as in (22) (Seiter 1980).21 (21) Ko e tohitohi a au (he) mogo-nei. pres write nom I on time-this `I am writing at the moment.' (Seiter 1980:30) (22) To lagomatai he ekekafo a ia. Fut help erg doctor nom him `The doctor will help him.' (Seiter 1980:29)

528

Ellen Woolford

An example of the active type of ergative language is Basque (Levin 1989). An intransitive construction with an ergative subject is shown in (23). (23) Gizona-k kurritu du. man-erg ran 3sg-aux `The man ran.' (Levin 1989:57) In Basque, only agentive subjects take ergative Case; unaccusative subjects take nominative Case (Levin 1989). (24) Ni etorri naiz. I-nom arrive 1sg-aux `I arrived.' (Levin 1989:43) Let us summarize the Case patterns that we wish to focus on. (25) Classic ergative type intrans: nominative *Ergative trans: ergative-nominative *Ergative-Accusative

Active-stative ergative type nominative ergative ergative-nominative *Ergative-Accusative

These patterns are very similar to what we saw above with datives. The classic ergative patterns look a lot like Japanese, with ergative substituted for dative, while the active-stative patterns look like the Icelandic patterns, again with ergative substituted for dative. As we saw with the dative languages above, the transitive patterns are the same in both types of ergative language, despite the di¨erence in intransitives. These languages show the same dependency e¨ect we saw with dative subjects: constructions with an inherently Cased subject have a nominative object.22 The above account of the two types of dative subject languages crucially rests on the fact that dative is an inherent Case. Under the view that ergative is also an inherent Case, we can capture the parallel dependency and valency e¨ects in ergative constructions. The constraints we need are FaithLex and *Ergative. If FaithLex is ranked higher than *Ergative, the ergative must surface, because FaithLex requires that any inherent Case-licensing feature on the input verb be checked. That ordering produces active-stative languages where the ergative is preserved even in intransitive clauses. The reverse ranking produces classic ergative languages (and languages where ergative never surfaces) where nominative is selected over ergative in intransitives. (26) Constraint ranking for the two types of ergative languages FaithLextrans g FaithLex active-stative: g *Ergative classic ergative: FaithLextrans g *Ergative g FaithLex nonergative: *Ergative g FaithLextrans g FaithLex

Case Patterns

529

Tableau T16.11 Active-stative Input: NP V[‡erg subj]

FaithLex

f a. NP-erg b. NP-nom

*Ergative *

*!

Tableau T16.12 Classic ergative Input: NP V[‡erg subj] a. NP-erg f b. NP-nom

*Ergative

FaithLex

*! *

Let us now see how these rankings produce these Case patterns. 16.4.1 Intransitive Constructions Active-stative languages preserve all lexically licensed ergatives because of the highly ranked faithfulness constraints. (See tableau T16.11.) In contrast, classic ergative languagesÐas in tableau T16.12Ðrank FaithLex lower, allowing markedness to make the decision in intransitives. 16.4.2 Transitive Constructions In the active-stative type of ergative language, where FaithLex is ranked above *Ergative, FaithLex requires ergatives to surface in transitives as well as intransitives. (See tableau T16.13.) Nominative is a better choice than accusative for the object, due to markedness (the dependency e¨ect). The classic type of ergative language works much like Japanese; see tableau T16.14. Although FaithLex is ranked below *Ergative, the e¨ect of *Ergative is nulli®ed in transitive constructions by the e¨ect of the higher-ranked FaithLextrans . Thus ergative subjects surface in transitive constructions, even though they cannot in intransitives. What happens in ergative languages when the subject is not lexically licensed for ergative Case? The prediction is that candidate (c) below, with the nominativeaccusative pattern, will be the winner because it will not be eliminated by the faithfulness constraint, and the (a) and (b) patterns will be eliminated by *Ergative. That is exactly what happens in Hindi with verbs such as `bring' that do not lexically license an ergative; see (27).

530

Ellen Woolford

Tableau T16.13 Active-stative Input: V[‡erg subj] NP NP

FaithLex

f a. NP-erg NP-nom

*Accusative

*

b. NP-erg NP-acc c. NP-nom NP-acc

*Ergative

* *!

*Nominative *

*! *

*

*Accusative

*Nominative

Tableau T16.14 Classic ergative Input: V[‡erg subj] NP NP

FaithLextrans

f a. NP-erg NP-nom

*

b. NP-erg NP-acc c. NP-nom NP-acc

*Ergative

* *!

* *! *

*

(27) LaRkiyaÄaÄ sabzii laayõÄ Äõ. girls-nom vegetable-acc brought-FEM-PL `The girls brought vegetables.' Comrie (1984:858) In section 16.5, we will discuss more complex ergative systems that manifest other Case patterns. However, before we do that, let us consider languages with both datives and ergatives. 16.4.3 Datives in Ergative Languages We have seen that languages can be classi®ed into three types depending on the patterning of datives (no datives, datives only in transitives, datives in intransitives and transitives), and into three parallel types depending on the patterning of ergatives. What does this approach predict with respect to types of languages with both ergatives and datives? The answer to that question depends partially on whether or not there is any universal markedness relation between ergative and dativeÐthat is, are *Ergative and *Dative in any universally ®xed ranking? While one might assume that datives are less marked than ergatives, based on the fact that familiar languages such as German, Icelandic, and Japanese have datives but not ergatives, there are languages where *Dative is ranked above *Ergative.

Case Patterns

531

Basque is such a language. Basque has both ergative and dative Case; we saw ergative examples above, and a dative example is given in (28). (28) Ni-ri zure oinetako-a-k-j gustatzen zaizkit. I-dat your shoes-det-nom like aux `I like your shoes.' (Austin and Lopez 1995:12) However, the distribution of these Cases di¨ers; ergatives can occur in intransitives, but datives are limited to transitives (Austin and Lopez 1995). That suggests that *Ergative is ranked above FaithLex (so ergatives surface even in intransitives), while *Dative is ranked above FaithLex but below FaithLextrans (so that datives surface only in transitives). See (29). (29) Basque ranking: FaithLextrans g *Dative g FaithLex g *Ergative It appears, therefore, that the ranking of *Ergative and *Dative is not universally ®xed, and we thus expect the distribution of the dative Case to be independent of that of the ergative Case. Thus the ``type'' of a language with respect to datives should be independent of the ``type'' of that language with respect to ergatives.23 Languages with both ergatives and datives have a special theoretical interest, however, because they allow us to test the predictions of di¨erent kinds of OT approaches to the valency problem. 16.4.4 The Ergative-Dative Problem The type of language that will allow us to test these predictions prohibits both ergatives and datives in intransitive constructions, but allow both Cases in transitive clauses. Such a language is Uradhi (Crowley 1983).24 In Uradhi, intransitive subjects are always nominative (Crowley 1983:334), but there are transitive constructions with an ergative-dative pattern. (30) qapa-al ulu ipi-eu in" t"u-B}Oa. blue dove-erg clitic water-dat look for-pres `The blue dove is looking for water.' (Crowley 1983:341) The existence of ergative-dative transitive patterns in languages that require a nominative in intransitives is problematic for any approach that ranks rules that associate Cases to arguments. The reason is as follows. Such approaches account for the obligatory presence of nominative in intransitives by ranking the rule that assigns nominative highest, so that nominative will be assigned before the rule that assigns ergative or dative has a chance to apply. (31) Hypothetical ranking of case association rules for Uradhi assign nominative g assign dative, assign ergative g assign accusative

532

Ellen Woolford

That approach does a ®ne job for intransitives, which will always get nominative, and it also does a ®ne job for any transitive construction with one inherent Case plus a nominative, as in the dative-nominative example in (28). However, that approach makes an incorrect prediction for transitive examples like (30) with an ergative-dative pattern. In those constructions, the nominative assigning/association rule should also apply ®rst, so that either the subject or the object should be nominative. The existence of ergative-dative clauses in languages like Uradhi is equally problematic for the approach taken in the oral version of this chapter. Under that approach, a language such as Uradhi would have a constraint ranking as in (32). (32) *Accusative g *Dative, *Ergative g *Nominative, FaithLex That grammar would select nominative Case for all intransitives, because the faithfulness constraint is ranked too low to allow dative or ergative to win. But with FaithLex ranked so low, decisions in transitives would also be made by markedness and an ergative-dative pattern could never beat a pattern containing a nominative. In contrast, the Uradhi pattern is not problematic, and is in fact expected, under the contextual faithfulness approach advocated here. The Uradhi pattern will result whenever both *Dative and *Ergative are ranked between the two faithfulness constraints, FaithLextrans and FaithLex. (33) Uradhi ranking: FaithLextrans g *Dative, *Ergative g FaithLex Ranking *Dative and *Ergative above FaithLex prevents either of these inherent Cases from surfacing in intransitive constructions, but ranking *Dative and *Ergative below FaithLextrans means that both inherent Cases will surface in transitives when the verb lexically licenses them. Thus languages like Uradhi provide strong evidence for a contextual faithfulness approach to the valency problem, over any approach that tries to get valency e¨ects merely from the ranking of Case association/assignment rules or markedness constraints. 16.4.5 Conclusions In this section, we have seen that the types of dative subject languages discussed in section 16.3 are paralleled by the types of ergative constructions that occur and the same approach accounts for both. This initial success suggests that this OT approach is a promising way to deal with the problem of the typology of Case patterns. Nevertheless, the few simple constraints introduced so far cannot account for anything like the full range of Case patterns that occur in the languages of the world. Although a complete account of the typology of Case patterns is far beyond the scope of this chapter, the last section of the chapter will discuss possible ways of dealing with some other sorts of dependency e¨ects involving Case.

Case Patterns

16.5

533

Other Sorts of Dependency E¨ects

We have discussed dependency e¨ects in which the Case of the object depends on the Case of the subject. Other dependency e¨ects include Cases that depend on the aspect of the verb and Cases that depend on the semantic features of the argument (e.g., person, animacy, speci®city, and so on). 16.5.1 Aspect Split in Hindi: Contextually Restricted Faithfulness Hindi is a well-known example of a language with an aspectually based ``split'' in its Case patterns, such that ergative Case is restricted to the perfective aspect (e.g., Comrie 1984; Kachru 1987; Mahajan 1990). In the examples below, the verb `eat' occurs with a nominative-accusative pattern in the imperfective construction in (34a), but an ergative-nominative pattern in the perfective construction in (34b). In both constructions, agreement is with the nominative argument. (34) a. Raam roTii khaataa thaa. Ram(masc)-nom bread(fem)-acc eat(imp-masc) be(past-masc) `Ram (habitually) ate bread.' (Mahajan 1990:72) b. Raam-ne roTii khaayii thii. Ram-erg bread-nom eat(perf-fem) be(past-fem) `Ram had eaten bread.' (Mahajan 1990:73) It is proposed here that this is another instance of a contextual restriction on the distribution of ergatives and we should treat it like we treated the contextual restriction of ergatives to transitives in the last section. Speci®cally, the proposal is that there is a variant of the FaithLex constraint that preserves ergatives that operates only in the context of perfective aspect. When that contextually restricted FaithLex constraint is ranked higher than *Ergative, the result is that lexically licensed ergatives will surface in perfective clauses. But outside that context, markedness takes over, and the subject is nominative. (35) Hindi ranking: FaithLexperf g *Ergative g FaithLex, *Accusative This approach makes the correct prediction that not all agentive or transitive verbs will take an ergative subject in the perfective; instead, only the verbs that are lexically marked to license an ergative subject will do so. The Hindi verb for `brought' is an example of an agentive, transitive verb that takes a nominative subject even in the perfective; see (36). (36) LaRkiyaÄaÄ sabzii laayõÄ Äõ. girls-nom vegetable-acc brought-FEM-PL `The girls brought vegetables.' Comrie (1984:858)

534

Ellen Woolford

Any approach that ignores the lexicon and simply assigns ergative Case to all agentive or transitive subjects in the perfective aspect will make the wrong prediction for such examples. 16.5.2 Subject-Person Splits: the Morphological Realization of Case A language is said to manifest a subject ``split'' if it marks subjects di¨erently (in terms of Case or agreement) depending on the features of that subject. The standard example of a subject split is from Dyirbal, where transitive subjects take an ergative Case morpheme only if they are third person; ®rst- and second-person subjects are never marked with (overt) ergative Case (Dixon 1979:87). The proper treatment of such splits depends on whether they represent an alternation between di¨erent abstract Cases (nominative versus ergative) or whether they involve one abstract Case that may or may not be morphologically realized. Comrie (1991) argues that all transitive subjects in Dyirbal have ergative Case that is simply not morphologically realized on ®rst- and second-person pronouns. There are good arguments that this is the correct approach for Dyirbal and for other languages with this kind of subject-person split. Dixon (1972) shows that ®rst- and second-person transitive subjects behave like ergative subjects in a variety of syntactic constructions in Dyirbal. In languages with agreement, such as Marathi, the agreement pattern indicates that subject-person splits do not involve a change in abstract Case (in contrast to aspect splits, which do). We see this contrast in the examples below, where the nominative subject in the imperfective construction in (37a) controls the agreement, whereas the unmarked ®rst-person subject in the perfective construction in (37b) does not, suggesting that it is still ergative. ambe khate. (37) a. Sudha Sudha(fem) mangoes(masc) eat(fem,3sg) `Sudha eats mangoes.' (Comrie 1984:861) b. Mi kame keli. I jobs did-(neuter,3pl) `I did the jobs.' (Comrie 1984:861) We can eliminate the possibility that the agreement system simply has a di¨erent pattern in the two aspects, because transitive verbs that do not license an ergative subject show the agreement of a nominative subject. (38) Mi khup goSTi bollo. I many things said(masc,1sg) `I said many things.' (Comrie 1984:861) The conclusion is that these subject-person splits do not involve an alternation between di¨erent Cases, but only whether or not the ergative abstract Case is morphologically

Case Patterns

535

realized.25 The present chapter is limited to dealing with situations involving choices among di¨erent abstract Cases, and nothing will be said about how to deal with morphological realization problems. However, the reader is referred to Aissen (chap. 3, this volume), who treats the morphological realization problem with constraint conjunction. 16.5.3 Marked Objects and Three-Way Systems It is well known (e.g. Comrie 1981; Croft 1990; Dryer 1986) that in many languages, objects with certain features are Case marked di¨erently than objects without those features (e.g., speci®c and nonspeci®c objects in Turkish (EncË 1991) and Inuit (Bittner 1994), and pronominal versus nonpronominal objects in Dyirbal (Dixon 1972)). How should we treat such object Case ``splits''? Aissen (chap. 3, this volume) translates the functional approach of Silverstein (1976) into OT using the technique from Prince and Smolensky 1993 of aligning prominence scales. Under Silverstein's approach, subjects prototypically have features high on the various animacy/topicality hierarchies (e.g., ®rst person, human, speci®c, and so on), while objects prototypically have low features (third person, inanimate, nonspeci®c, and so forth). When a subject or object has unexpected features, it needs to be marked or ¯agged as atypical, so as to alert the listener not to make a mistake about its grammatical relation. This ¯ag can be a morphologically overt Case, preposition, or agreement. One problem with Silverstein's approach is that it treats subject splits and object splits alike, as two ends of the same continuum. But if that is so, we expect to ®nd an equal diversity of types of subject and object splits in the world's languages; however, that prediction is not borne out. Instead, there are very few kinds of subject splits, in contrast to an enormous diversity of object splits.26 For example, Comrie (1981:123) notes that while de®niteness is frequently the basis of object splits, there is an embarrassing absence of clear attestations of the predicted marked inde®nite subject. Object splits often involve disjoint sets of features and combinations of features, as in Palauan where human objects and singular, speci®c objects are marked (Woolford 1995), but subject splits seem never to involve more than one feature. In addition, object splits can be ``marked'' with a range of devices, including prepositions, several di¨erent Cases, and agreement, while subject splits only involve Case. Croft (1990) argues that agreement does not show the inverted pattern for subjects that Silverstein's approach would predict. My conclusion is that (virtually) all subject splits (and some object splits) involve an alternation between realizing or not realizing one abstract Case, whereas the majority of object splits involve an alternation between two distinct abstract Cases. My approach to the latter type of object split (Woolford 1995, 1996, 1999) is parallel in some sense to the treatment of coda constraints in phonology. The basic intuitive idea is that syntax ``wants'' an empty VP (or at least one empty of structurally Cased

536

Ellen Woolford

arguments), just as phonology ``wants'' an empty coda in syllables. Structurally Cased VP-internal arguments are worse if they have marked features, just as codas are worse if they have marked features. The strategies that languages use to avoid having a VP-internal structurally Cased argument are of two types: one can make the argument not VP-internal by moving it out of VP (which may consequently alter its Case), or one can simply alter the Case of the argument by adding a preposition or an inherent Case. The reason that subjects are so much less often a¨ected by such constraints is that subjects are usually already outside of the VP for independent reasons, and thus cannot violate these constraints. For reasons of space, it is not possible to describe this approach to object splits and contrast it with that of Aissen (chap. 3, this volume), so let us instead focus here on what these two approaches have in common and how they allow us to explain why nominative is not always the best choice for objects in ergative languages. Simplifying somewhat, both my approach and Aissen's postulate a family of markedness constraints that prohibit ``bare'' objects with marked/high features such as [‡human], [‡speci®c], [‡topic], or [‡pronominal]. Normally, ergative languages have nominative objects, as we have seen in the above sections, but nominative objects are ``bare'' objects. If the relevant bare object constraint is ranked high enough, it will eliminate candidates with a VP-internal nominative object from the candidate set, leaving a candidate with a more marked object Case as the winner. 16.6

Conclusion

Case patterns that are problematic for Case theory (e.g., Chomsky 1981, 1986, 1995) include dependency e¨ects, valency e¨ects, and inventory e¨ects. These can be dealt with if we add a set of violable constraints to Case theory, which come into play only when more than one Case is licensed on a particular argument. This will allow us to maintain the Case-licensing principles of Chomsky 1995 as universal, inviolable principles, formulated in a quite restrictive way that mentions only which head licenses which Case. All situations that previously seemed to require language-speci®c licensing principles, or language-speci®c features or parameter settings, or any other exceptions to or quali®cations of the universal Case-licensing principles can now be handled by the violable Case constraints. However, even the violable constraints are universal: what di¨ers between languages is their relative ranking. The violable Case constraints include faithfulness and markedness constraints. Faithfulness constraints require faithfulness to (checking of ) any inherent Caselicensing features carried by verbs. Markedness constraints select less marked structural Cases over more marked inherent Cases, and within the structural Cases, they select the less marked nominative over the more marked accusative. In addition,

Case Patterns

537

there are OCP-type constraints that prohibit two instances of the same Case or the same kind of Case in certain domains. This approach subsumes the behavior of unaccusatives that motivated Burzio's generalization under markedness: unaccusatives can be licensed for nominative or accusative Case, but markedness prefers the less marked nominative. Similarly, in constructions with a dative or ergative subject, the object surfaces as nominative rather than accusative because of markedness. With the removal of D-structure from the theory, we can no longer maintain the idea that inherent Cases are assigned at D-structure in order to explain why inherent Cases occur in positions normally licensed for structural Case. Instead, inherent Case surfaces when faithfulness constraints (favoring inherent Case) are ranked above markedness constraints favoring a structural Case. However, faithfulness is sometimes contextually restricted, producing valency e¨ects as well as ``split'' systems. The reason that ergative and dative do not surface in all languages is that these are rather marked Cases. The same markedness and faithfulness constraints that account for the dependency and valency e¨ects discussed above also account for the (apparent) Case inventory of each language. There is no need to claim that some languages lack the ability to license certain Cases. Notes I would like to thank Judith Aissen, Rakesh Bhatt, Stanley Dubinsky, Lyn Frazier, Lisa Green, Jane Grimshaw, Roger Higgins, Andre Isaak, Kyle Johnson, Chisato Kitagawa, John McCarthy, Peggy Speas, and Sten Vikner for interesting discussion of the issues in this chapter and valuable comments on earlier drafts and presentations of the chapter. 1. The idea that faithfulness can be restricted to certain contexts is proposed in McCarthy and Prince 1995 and developed in Beckman 1998. 2. The idea that the inventory of sounds in a language can be accounted for, rather than merely stipulated, comes from Prince and Smolensky 1993. 3. In the version of this chapter presented at the Hopkins Optimality Workshop/Maryland Mayfest 1997, I assumed that the input to syntax was something more like the old D-structure (without the structure) where inherent Cases were already assigned to the relevant arguments. 4. The fact that nominative Case can be licensed on an argument inside the VP means that overt NP Movement is no longer driven by Case checking; instead, movement to the external subject position is driven by a di¨erent principle, a version of the EPP (Extended Projection Principle) formulated in Chomsky 1995 as a need to check a D feature. 5. There are constructions that can optionally have what looks like a nominative Case morpheme on both subject and object in Korean, but SchuÈtze 1996 argues that one of these actually marks the topic. Multiple nominatives mark multiple topics in Japanese as well (Tateishi 1991). See also Dubinsky 1992 on other special Japanese constructions with a double nominative. I know of no language that marks normal agentive transitive clauses with a nominative on both subject and object, although languages that have a morphologically unmarked accusative are sometimes glossed in this way in the literature.

538

Ellen Woolford

6. Some of the claims that dative is structural are based on passives where dative changes to nominative. However, under the approach argued for here, lexically licensed Cases do not always get to surface, and thus the traditional ``Case change'' diagnostic is not reliable. 7. Nevertheless, we need to capture the correlations between thematic roles and particular inherent Cases. One idea is that some verbs carry truly quirky inherent Case features, but other verbs receive such inherent Case-licensing features as a result of general lexical rules (Yip, Maling, and Jackendo¨ 1987). Both types of verb will carry such a feature in the input to syntax. 8. The idea that there are constraints such as *Dative and *Accusative is independently proposed by Grimshaw (1997; chap. 8, this volume). 9. One might doubt that nominative is universally the least marked Case, given the fact that in English, accusative rather than nominative is selected in several contexts where there is no obvious head to license that Case. One is when a question such as ``Who wanted co¨ee?'' is answered ``Me.'' Another is in coordinate constructions where, in nonstandard dialects and in the speech of many children, accusative occurs even when the coordinate constituent is a subject: ``Me and Joe went ®shing.'' I assume that there is an as yet unknown higher constraint operating that selects accusative (pronouns) in such contexts, although I can only suggest some possible lines of inquiry as to what that constraint (or constraints) might be. One is suggested by the fact that coordinate constructions require what Cardinaletti and Stark (1999) call strong pronouns. Another, suggested to me by Sten Vikner, is that in the ®rst context the answer is elliptical, and that the full answer would license accusative Case: ``It was me.'' 10. The constraint *Nominative is not needed for any situation discussed here, or any other situation that I know of. If the theory does not require the least marked element in such a series to have a corresponding markedness constraint, *Nominative should be eliminated. 11. This does not exclude the possibility of a language with datives but no (surface) accusatives, because there are other constraints that favor datives (faithfulness) and discriminate against accusatives (see section 16.5.3). 12. The question of what contexts are associated with special faithfulness and why is an interesting one, in both syntax and phonology. There is a functionalist argument that there is a need for marked Cases in transitive clauses, to prevent ambiguity between subject and object (e.g., Silverstein 1976). That account does not extend to aspectual restrictions, but those have been argued to have a historical account (e.g., Anderson 1977). 13. This valency restriction in Japanese holds in matrix clauses but not in embedded clauses (Shibatani 1977:807; Dubinsky 1992), and in statements but not in questions (Kitagawa, personal communication). This suggests that there is also contextually restricted faithfulness in these contexts. 14. Nakamura (1997, 1999) independently proposes an OT solution to this problem of the two types of dative subject languages, embedded in Role and Reference grammar. For him, the Icelandic pattern results when the constraint ``Non-macrorole core arguments take DATIVE case'' is ranked above ``Core arguments take nominative Case.'' The opposite ranking produces the Japanese pattern. However, we will see in section 16.4 that any approach that captures this valency e¨ect by assigning nominative before dative encounters empirical problems. 15. Note that there is no need to invoke partitive Case to explain how a subject that remains in the VP gets Case, contra Belletti 1988, because a VP-internal argument can be licensed for nominative Case by feature raising (Chomsky 1995) and unaccusative subjects agree like any

Case Patterns

539

other nominative in Italian (Burzio 1986). The de®niteness e¨ect is argued by Diesing (1992) to be related to structural position of the object rather than to partitive Case. Chomsky (1995) rejects the partitive Case hypothesis as incompatible with the assumptions of the Minimalist Program. The idea that unaccusative objects inside the VP are potentially licensed for either nominative or accusative Case, but some principle selects nominative over accusative, has been proposed more or less independently in a wide range of work on Burzio's Generalization (see for example Brandner 1993, 1995; Laka 1993; Haider 1995; Burzio, forthcoming; Mahajan, forthcoming), although the deciding principle proposed in these works is not markedness. Woolford 2000 provides a summary of this literature. 16. A nominative-dative pattern with a structurally licensed dative would be eliminated by FaithLex because the subject has no dative feature to check against the verb's dative licensing feature. A dative-dative pattern would be eliminated by *Dative because of its extra violation. 17. While nominative is always selected over accusative when the relative markedness of these two Cases is making the decision, in OT there is always the possibility that other constraints, when high ranked, will make a di¨erent decision. A family of constraints that selects a more marked object when that object has certain features such as human or speci®c will be discussed brie¯y at the end of this chapter. There may also be a constraint penalizing the nonlocality of checking nominative Case on an argument inside VP, which may be responsible for the dativeaccusative pattern in Faroese (see Barnes 1986), a language closely related to Icelandic. See Woolford 2000 for a discussion of how this constraint should be formulated. 18. The relative ranking of the FaithLex constraints with respect to markedness constraints barring structural Cases makes no empirical di¨erence because FaithLex constraints have no e¨ect on structural Cases. 19. To strengthen this result, we need to rule out the possibility of a faithfulness constraint that is contextually restricted to intransitives. 20. The Double-o constraint in Japanese has exceptions, ``especially when the two constituents are non-adjacent'' (Dubinsky 1994:54). 21. Verbs such as like with an experiencer subject take a nominative subject. 22. There are ergative languages that manifest a three-way pattern, due to the e¨ect of additional markedness constraints discussed in section 16.5. 23. The claim in Woolford 1997 that the behavior of ergative and dative is parallel refers only to the fact that both Cases participate in dependency and valency e¨ects, in a way that structural Cases do not. 24. Another such language may be Tabassaran. There are no one-place verbs with either dative or ergative Case according to Kibrik (1985), who also gives an example of a transitive sentence with an ergative subject and a dative object (Kibrik 1985:281). 25. Additional evidence that there is no abstract Case change from ergative to nominative in languages that overtly mark ergatives only in third person comes from Kabardian (Colarusso 1992). Kabardian has a marked nominative Case. The nominative, ergative, and dative (the latter two look alike) all show the same pattern of overt morphological marking of the Case in the third person but not in ®rst and second person. However, it is clear from the fact that nominative is morphologically marked in Kabardian that this alternation is between the presence and absence of a Case morpheme, rather than re¯ecting a change in abstract Case.

540

Ellen Woolford

26. This observation is also noted in Isaak 2000. Isaak proposes an interesting modi®cation of Silverstein's approach wherein certain features associated with object splits are privative, so that there could not be subject splits based on the negative value of features such as human. References Anderson, Stephen. 1977. On Mechanisms by Which Languages Become Ergative. In Charles Li (ed.), Mechanisms of Syntactic Change, 317±363. Austin: University of Texas Press. Austin, Jennifer, and Luis Lopez. 1995. Nominative, Absolutive, and Dative Languages. In NELS 25, 1±15. GLSA, University of Massachusetts, Amherst. Barnes, Michael. 1986. Subject, Nominative, and Oblique Case in Faroese. Scripta Islandica 37: 13±46. Beckman, Jill. 1998. Positional Faithfulness. Doctoral dissertation, University of Massachusetts, Amherst. Belletti, Adriana. 1988. The Case of Unaccusatives. Linguistic Inquiry 19: 1±34. Bittner, Maria. 1994. Case, Scope, and Binding. Dordrecht: Kluwer. Brandner, Ellen. 1993. The Projection of Categories and the Nature of Agreement. In Gisbert Fanselow (ed.), The Parameterization of Universal Grammar. Amsterdam: Benjamins. Brandner, Ellen. 1995. Speci®ers and Functional Projection. Doctoral dissertation, University of Stuttgart. Broekhuis, Hans, and Leonie Cornips. 1994. Undative Constructions. Linguistics 32: 173±189. Burzio, Luigi. 1986. Italian Syntax. Dordrecht: Reidel. Burzio, Luigi. Forthcoming. Anatomy of a Generalization. To appear in Eric Reuland (ed.), Argument and Case: Explaining Burzio's Generalization. Amsterdam: Benjamins. Cardinaletti, Anna, and Michael Starke. 1999. The Typology of Structural De®ciency: On the Three Grammatical Classes. In H. van Riemsdijk (ed.), Clitics in the Languages of Europe. Berlin: Mouton. Chomsky, Noam. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, Noam. 1986. Knowledge of Language. New York: Praeger. Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press. Colarusso, John. 1992. A Grammar of the Kabardian Language. Calgary: University of Calgary Press. Comrie, Bernard. 1981. Language Universals and Linguistic Typology. Oxford: Blackwell. Comrie, Bernard. 1984. Re¯ections on Verb Agreement in Hindi and Related Languages. Linguistics 22: 857±864. Comrie, Bernard. 1991. Form and Function in Identifying Cases. In F. Plank (ed.), Paradigms, 41±55. Berlin: Mouton de Gruyter. Croft, William. 1990. Typology and Universals. Cambridge, England: Cambridge University Press. Crowley, Terry. 1983. Uradhi. In R. M. W. Dixon and B. Blake (eds.), Handbook of Australian Languages, vol. 3, 306±531. Amsterdam: Benjamins.

Case Patterns

541

Czepluch, Hartmut. 1988. Case Patterns in German: Some Implications for the Theory of Abstract Case. In McGill Working Papers in Linguistics: Special Issue on Comparative Germanic Syntax, 79±122. Department of Linguistics, McGill University, MontreÂal. Diesing, Molly. 1992. Inde®nites. Cambridge, Mass.: MIT Press. Dixon, R. M. W. 1972. The Dyirbal Language of Northern Queensland. Cambridge, England: Cambridge University Press. Dixon, R. M. W. 1979. Ergativity. Language 55: 59±138. Dryer, Matthew. 1986. Primary Objects, Secondary Objects, and Antidative. Language 62: 808±845. Dubinsky, Stanley. 1992. Case Assignment to VP-Adjoined Positions: Nominative Objects in Japanese. Linguistics 30: 873±910. Dubinsky, Stanley. 1994. Predicate Union and the Syntax of Japanese Causatives. Journal of Linguistics 30: 43±79. EncË, MuÈrvet. 1991. The Semantics of Speci®city. Linguistic Inquiry 22: 1±26. Grimshaw, Jane. 1997. The Best Clitic: Constraint Con¯ict in Morphosyntax. In Liliane Haegeman (ed.), Elements of Grammar, 169±196. Dordrecht: Kluwer. Haider, Hubert. 1995. The License to Check: Structural Case ‡ Economy Yields Burzio's Generalization. In Hubert Haider (ed.), Studies on Phrase Structure and Economy. Arbeitspapiere des Sonderforschungsbereichs 340: Sprachtheoretische Grundlagen fuÈr die Computerlinguistik, Bericht Nr. 70. UniversitaÈt Stuttgart. Harada, S. I. 1973. Counter Equi-NP Deletion. In Annual Bulletin 7: 113±147. Research Laboratory of Logopedics and Phoniatrics, University of Tokyo. Isaak, AndreÂ. 2000. Split Case-Marking and Prominence Relations. Doctoral dissertation, University of Massachusetts, Amherst. Kachru, Yamuna. 1987. Ergativity, Subjecthood, and Topicality in Hindi-Urdu. Lingua 71: 223±238. Keer, Edward. 1999. Geminates, the OCP, and Faithfulness. Doctoral dissertation, Rutgers University, New Brunswick, N.J. Kibrik, A. E. 1985. Toward a Typology of Ergativity. In A. Woodbury and J. Nichols (eds.), Grammar Inside and Outside the Clause, 268±323. Cambridge, England: Cambridge University Press. Kuroda, S.-Y. 1992. Japanese Syntax and Semantics: Collected Papers. Dordrecht: Kluwer. Laka, Itziar. 1993. Unergatives That Assign Ergative. In Jonathan D. Bobaljik and Colin Phillips (eds.), MIT Working Papers in Linguistics 18: Papers on Case and Agreement I, 149± 172. MITWPL, Department of Linguistics and Philosophy, MIT, Cambridge, Mass. Legendre, GeÂraldine, W. Raymond, and P. Smolensky. 1993. An Optimality-Theoretic Typology of Case and Grammatical Voice Systems. In Joshua S. Guenter, Barbara A. Kaiser, and Chenyl C. Zoll (eds.), Proceedings of the Nineteenth Annual Meeting of the Berkeley Linguistics Society, 464±478. Berkeley Linguistics Society, University of California, Berkeley. Levin, Beth. 1989. The Basque Verbal Inventory and Con®gurationality. In L. MaraÂcz and P. Muysken (eds.), Con®gurationality: The Typology of Asymmetries, 39±62. Dordrecht: Foris.

542

Ellen Woolford

Levin, Beth, and Jane Simpson. 1981. Quirky Case and Lexical Representations of Icelandic Verbs. In Roberta A. Hendrick, Carrie S. Masek, and Mary Frances Miller (eds.), CLS 17, 185±196. Chicago Lingvistic Society, University of Chicago. Mahajan, Anoop. 1990. The A/A-Bar Distinction and Movement Theory. Doctoral dissertation, MIT, Cambridge, Mass. Mahajan, Anoop. Forthcoming. Oblique Subjects and Burzio's Generalization. In Eric Reuland (ed.), Argument and Case: Explaining Burzio's Generalization. Amsterdam: Benjamins. McCarthy, John, and Alan Prince. 1995. Faithfulness and Reduplicative Identity. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 655±700. GLSA, University of Massachusetts, Amherst. Mohanan, K. P. 1982. Grammatical Relations and Clause Structure in Malayalam. In Joan Bresnan (ed.), The Mental Representation of Grammatical Relations, 504±589. Cambridge, Mass.: MIT Press. Nakamura, Wataru. 1997. A Constraint-Based Typology of Case Systems. Doctoral dissertation, SUNY, Bu¨alo, N.Y. Nakamura, Wataru. 1999. Functional Optimality Theory. In Michael Darnell et al. (eds.), Functionalism and Formalism, vo1. 2. Amsterdam: Benjamins. Prince, Alan. 1997. Paninian Relations. Paper presented at the University of Massachusetts, Amherst. Prince, Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint Interaction in Generative Grammar. Technical Report TR-2, Center for Cognitive Science, Rutgers University, New Brunswick, N.J., and Technical Report CU-CS-696-93, Department of Computer Science, University of Colorado, Boulder. Available at http://www.ruccs.rutgers.edu/ publicationsreports.html. SchuÈtze, Carson. 1996. Korean ``Case Stacking'' Isn't: Unifying Noncase Uses of Case Particles. In Kiyomi Kusumoro (ed.), Proceedings of the North East Linguistic Society 26. GLSA, University of Massachusetts, Amherst. Seiter, William. 1980. Studies in Niuean Syntax. New York: Garland. Shibatani, Masayoshi. 1973. Semantics of Japanese Causativization. Foundations of Language 9: 327±373. Shibatani, Masayoshi. 1977. Grammatical Relations and Surface Cases. Language 53: 789± 807. Silverstein, Michael. 1976. Hierarchy of Features and Ergativity. In R. M. Dixon (ed.), Grammatical Categories in Australian Languages, 112±171. Atlantic Highlands, N.J.: Humanities Press. Tateishi, Koichi. 1991. The Syntax of ``Subjects.'' Doctoral dissertation, University of Massachusetts, Amherst. Tsunoda, Tasaku. 1981. Split Case-Marking Patterns in Verb-Types and Tense/Aspect/Mood. Linguistics 19: 389±438. Webelhuth, Gert 1995. X-Bar Theory and Case Theory. In Gert Webelhuth (ed.), Government and Binding Theory and the Minimalist Program, 15±95. Oxford: Blackwell.

Case Patterns

543

Woolford, Ellen. 1993. Symmetric and Asymmetric Passives. Natural Language and Linguistic Theory 11: 679±728. Woolford, Ellen. 1995. Object Agreement in Palauan: Speci®city, Humanness, Economy, and Optimality. In J. Beckman, L. Walsh Dickey, and S. Urbanczyk (eds.), University of Massachusetts Occasional Papers 18: Papers in Optimality Theory, 655±700. GLSA, University of Massachusetts, Amherst. Woolford, Ellen. 1996. Conditions on Object Agreement in Ruwund. Ms., University of Massachusetts, Amherst. Woolford, Ellen. 1997. Four-Way Case Systems: Ergative, Nominative, Objective, and Accusative. Natural Language and Linguistic Theory 15: 181±227. Woolford, Ellen. 1999. Animacy Hierarchy E¨ects on Object Agreement. In Paul Kotey (ed.), New Dimensions in African Linguistics and Languages. Trenton, N.J.: Africa World Press. Woolford, Ellen. 2000. Deriving the E¨ects of Burzio's Generalization. Paper presented at the Case Theory Workshop, DGfS, Marburg, Germany, March. Yip, Moira. 1995. Identity Avoidance in Phonology and Morphology. Available as ROA-82 at Rutgers Optimality Archive: http://www.ruccs.rutgers.edu/roa.html. Yip, Moira, Joan Maling, and Ray Jackendo¨. 1987. Case in Tiers. Language 63: 217±250. Zaenen, Annie, and Joan Maling. 1984. Unaccusative, Passive, and Quirky Case. In Proceedings of the Third West Coast Conference on Formal Linguistics, 317±329. Stanford, Calif.: CSLI Publications. Zaenen, Annie, Joan Maling, and H. ThraÂinsson. 1985. Case and Grammatical Functions: The Icelandic Passive. Natural Language and Linguistic Theory 3: 441±483.

Index

A/A0 binding, 143, 147, 157, 159, 161, 163±164 movement, 144, 146, 147, 165n5, 181 property, 143, 147, 157, 163±164, 165n5 scrambling, 143, 147, 155±156, 159, 162, 164 A½xation derivational vs. in¯ectional, 49±53 phrasal, 242±243, 247±249, 252, 263 Aghem, 316, 334 Agreement, 397±398, 407, 412±413 Alignment, 205, 211, 213±215, 218±220, 222±229, 236, 237n11, 238n12, 242, 251 constraint alignment, 66±89 harmonic alignment of scales, 66±69 prosodic domain of, 246, 252±253, 270n1 syntactic domain of, 246, 252, 269 Alignment constraints, 7±9, 11, 18, 22n14, 105± 106, 109, 131, 176±180, 213±224, 242, 251±259, 261, 263±267, 269, 270n2, 272n14, 272n14, 273n15, n19, n20, 315, 317±318, 320±322, 325, 328±341 Allotaxy, 122±123, 131 Arabic, 173 Aranda, 87 Auxiliary verbs, insertion of, 438, 440, 442, 452, 461 Basque, 247±248, 271n3, 531 Berber, 173 Binding, 466±467, 471, 473, 478, 488, 492 mixed, 143, 146, 154 split, 143, 163, 164 Binding constraints BindingTheoryB, 401±413 MostLocalAntecedent, 470±479 PrBd/ProperBinding, 432, 436±439, 441, 444, 448, 454±458, 461 Breton, 242±246, 249±250, 253±264, 271n4, n6, n9, 272n15, 273n21 Bulgarian, 248, 251, 254, 264, 279, 282±287, 289± 290, 301±303 Burzio's Generalization, 510, 518, 537

Candidate set, 4, 19±20, 281, 284, 291±292, 429. See also Harmonic bounding; Inventories Case, 84±87, 176±179, 183±187, 189±191, 193± 198, 199n10, 200n13, 394±396, 401±402, 404, 407, 419, 421, 510±532, 525, 536±537 dependency e¨ect, 509, 510, 527, 533 ergative languages, 385, 527±532 nominative-accusative languages, 382, 511, 521± 522 split ergativity, 62, 82±88, 91fn2, 534±535, 537 Case constraints, 176, 401±406, 432, 513±533 Catalan, 182 Cebuano, 355, 373 Chamorro, 173, 469, 482±486, 488±492, 498±503 Chichewa, 115, 126±127, 132 Clitics auxiliary, 241, 243±250, 252, 254, 256±258, 261, 264, 267, 269, 270, 271n7, 272n11, 273n21, 274n21 combinations, 211±213, 225±228, 235, 237n9 doubling, 248 PF spell out, 242, 248±249 positioning of, 242±243, 253 pronominal, 114, 241, 245±246, 248±250, 254, 264±266, 271n10, 272n12, 273n20 relative order, 205, 207, 211±213, 216±217, 236 splitting, 266, 272n14 template, 205, 211±212, 214±216, 226 Complementizers absent, 103±105, 434 obligatory, 101±103, 106, 436±437, 454 optional, 97, 100±101, 104, 108±109, 33, 436± 437 Compounds, 33±36, 49±52 Synthetic, 33±36 N-V, 50±51 Con, 4±5 Constraints, general issues. See also Alignment; Faithfulness; Information structure; Markedness; and Operator binding for speci®c constraints conjoined (see Local conjunction)

546 Constraints, general issues (cont.) ¯oating (or partial ordering of ), 17±18 freely rankable, 12±13 gradient, 16, 227±228 subhierarchies, 12±13, 21, 61±90 tied, 17±18, 98, 105±106 Danish, 279, 282, 288, 290±291, 295±297, 299, 301±303, 306n13, 307n16, 308n22, 431±458 Discourse features of the input, 7±9, 148±154, 156, 158±164, 163, 165n8. See also Information structure Discourse prominence, 68, 72±73, 148±151, 356, 416 levels of, 360±363 Dispersion Theory, 355, 373±375. See also Inventories Do emphatic, 432, 447, 452 expletive, 5, 12, 257, 445±449, 456 Double object construction, 288±289, 291, 296± 297, 299±300, 302. See also Object shift Dutch, 174 Dyirbal, 83±88, 130, 534 Economy referential, 466±467, 472±475, 479±480, 485±503 structural, 8, 12, 16±17, 437±438 Economy constraints EcVP/Economy-of-VP's, 438±448, 452, 454±458, 461 HMC/HeadMovementConstraint, 261, 274n21, 445, 446±449, 455±458, 461 LxMv/NoLexicalMovement, 12, 430, 436±439, 441, 444±449, 454±458, 461 MLC/MinimalLinkCondition, 301±303, 306n7 MinProj/MinimalProjection, 8±9, 16, 23n18, n22 NoMorphology, 32 RefEcon/ReferentialEconomy, 473±478 Stay, 176±179, 183±187, 189±191, 193±198, 199n10, 200n14, 256, 281±282, 286, 289±294, 316±317, 320±322, 325, 327±341, 349n3, 433, 437, 444, 446, 448, 455, 458, 461 (see also *t) (*)Struc/Structure, 81, 123 *t (see also Stay), 16, 256±259, 261 *0/zero, 81±89 Emergence of the Unmarked, 9, 126±130, 175± 180, 263, 280 English, 4±6, 12, 14±15, 23n16, 97±99, 107±108, 282±290, 301±302, 306n8, 316, 331±332, 337, 410±414, 428±458 Expletives do (English), 5, 12, 257, 456 es (German), 1±2, 6±8, 11, 23n16 it (English), 4±5 ober (Breton), 257±259, 262±263, 273n19, n21

Index Eval, 4, 401, 415, 440 Evaluation. See Optimization Faithfulness, 4, 20±21, 21n5, 24n26, 97±101, 106± 107, 110, 117, 205, 208, 210, 220±222, 226, 230± 232, 234, 236, 238n16, 369±370, 470, 513±514, 516, 519, 522, 524, 526±527, 533, 536±537 contextual, 514, 524, 527 Faithfulness constraints Faith (or Max), 19, 101±110, 123, 208±210, 221±222, 230±235, 255±258, 273n20, 401±414, 470±503 FullInterpretation, 5±12, 17, 21n5, 22n9, n10 Focus, 143, 145±150, 152, 154, 156, 159, 161, 163± 164, 176±179, 186±187, 189±191, 193±198 lack of structural, 315, 319, 327±329, 332, 334, 336, 341 partial structural, 331±334, 341 structural, 315, 317, 319, 333±334, 336, 341 uniform leftward, 330, 334, 336, 339±341, 342n1 uniform rightward, 330, 334, 336, 338, 341, 342n1, n2 Fox, 71±72, 81 Franco-ProvencËal, 266 French, 18, 120, 122, 212±215, 428±429, 431±458 Gen, 4, 280, 399±401, 440 German, 1, 6±14, 18, 22n14, 23n18, 143, 151, 153±154, 156±157, 159, 163, 165n8, 241±242, 249±250, 273n15, n21, 274n23, 279, 295±299, 301±303, 305n8, 306n13, 308nn21, 23±35, 309n28 Government constraints, 102±105 Grammatical functions Subject, 172±174, 176±179, 181, 183±198, 199n8, n11 Object, 174, 176±189, 186±187, 189±191, 193± 194, 196±198, 199n7, n8, n10 Greek, 172±173, 181±184 Harmonic boundary, 326±327, 334, 340±341 Harmonic bounding (of candidates), 257, 315, 320±322, 324±326, 341, 349n3 Hierarchies Person/animacy, 61 Semantic roles, 61 Hindi, 530, 533 Icelandic, 466±467, 511, 514±522, 525, 528, 530, 282, 291±292, 296, 301±303, 307n17, 18, 436± 458 Iconicity, 121 Idioms, 32, 43±49 Ine¨ability, 2±3, 19±20, 97±98. See also Neutralization Information structure, 148, 150, 151, 157, 159, 165n8. See also Discourse features

Index Information structure constraints New, 151±154, 157±164, 166n17, 255 Prom, 151±164, 166n17 Inputs, 2, 20±21, 98±101, 118, 148, 356±358, 399± 400, 404, 413, 422n7, 422n12, 468±469. See also Faithfulness cross input consistency, 336 Inventories, 20±21, 205, 207±208, 230±231, 236, 375±378, 498. See also Dispersion Theory of clitic forms, 230, 236 of pronominal forms, 125 of voice systems, 355±387 Irish, 173 Italian, 4±5, 11±12, 19, 22n9, 23n16, 172, 183, 187±194, 198, 199n8, 212, 215±216, 316, 330, 348n2, 349n2iv Japanese, 120, 127±128, 514±530 Kera, 316, 327±328 Kichaga, 131 Korean, 397, 413±414, 422n4 Left-dislocation, 172±174, 176, 178, 181, 183±184, 186, 191 Lexicon, 6, 30±31, 36, 41, 44, 205±208, 211, 226, 230±231, 233, 235, 236, 250 Local conjunction, 23fn20, 80±89, 157 Logical Form (LF), 17, 19±20 Lumni, 76±77 Lushootseed, 75 Macedonian, 242±243, 245±249, 251±254, 263± 269, 271n4, n7, n10, 274n22 Malagasy, 174 Mandarin (Chinese), 396±397, 407, 410, 412±414 Marathi, 465, 470±471, 477±478, 534 Markedness, 12, 79, 97±99, 101, 106±107, 110, 113±125, 205, 207±208, 211, 221±222, 226±227, 230±236, 237n10, 358±359, 470±471, 510, 513, 516, 519, 520, 522, 524, 526, 536±537 Minimality relativized, 466±469, 477±481, 490, 493, 495± 496 rigid, 467, 470, 477±479 Movement case-driven, 282, 294, 301±302, 368n21 V-to-I, 427±429, 234, 442, 445, 461 wh-, 97±98, 107, 282±292, 302±304 Negative clauses, 442, 445±454 Neutralization, 101, 468±470, 474, 494±495 Norwegian, 97±99, 107±108 Object shift, 282, 288±294, 301±303, 307n17. See also Double object construction Ojibwe, 130

547 Operator binding constraints, 155±164, 166n17 Optimization bi-directional, 465±503 expressive, 466, 468, 472±476, 479±480, 483, 495±498, 500±503 global vs. local, 54 interpretive, 468±469, 472, 479±480, 493, 495± 499, 501, 503 Lexicon, 503 Optionality, 17±19, 97±98, 100±101, 109±110, 433, 436. See also Neutralization Particle verbs, 36±43 Phi features, 402, 410, 470, 473, 485, 488 Philippine languages, 371±372, 375±376, 380, 382, 387 Podoko, 316, 339 Point of view, 393, 419±421 Portuguese Brazilian, 120 European, 171±177, 175±177, 179, 181±190, 197±198, 199n7, n10 West African Pidgin, 122 Pronouns. See also Clitics alternating with clitics, 208, 227 bound, 114 de®ning properties of, 115 form-function constraints on, 119 forms of, 114 null object, 397, 407, 410, 413, 415±416, 422n3, n4 overt, 119 pro, PRO, 393±398, 401±421, 422n4 reduced, 119 typology of, 125, 134 weak, 114, 288±289, 294, 296 zero, 114 Pronoun fronting, 294±296 QR (Quanti®er Raising), 299±301 Reconstruction, 143±145, 147, 157, 163 Anti-, 144, 147, 157, 163 Relative clauses, 99, 102, 108 that-relatives, 103, 106 wh-relatives, 107±108 Richness of the Base, 24fn28, 117 Salish, 133 Sama, 387 Scottish Gaelic, 316, 336 Scrambling, neutral vs. focus, 146, 157, 159, 164 Second-Position F2/Finite-second, 243, 245±246, 249±250, 255, 258, 262±264, 267, 269, 273n15, n21 V2/Verb-second, 241±243, 457±459

548 Semantic roles. See also Thematic roles Agents, 360±362, 368, 371±372, 379, 381, 383, 386 Patients, 360±362, 368, 371±372, 376, 379, 381, 383, 386 Serbo-Croatian, 253, 264, 270n1 Spanish, 115, 172, 181, 183±187, 190, 195±198, 199n6, 200n13, n14, 217±222, 316, 338, 349n2, 410±414, 422n3 Spurious se, 205, 221±222, 237n5, 237n10 Squamish, 77±79 Strict domination, 3, 13±14, 18 Structural constraints Control, 393, 398, 401±402, 404±416, 418±421 F-Crit/F-Criterion, 292±294 MarkXP, 101±102 ObHd/ObligatoryHeads, 247, 257±258, 432, 437, 439, 441, 444, 446, 448, 455, 458, 461 Operators-in-Scope, 457±459 OpSpec/Operators-in-Spec, 45, 247, 258, 457 Par-Mov/ParallelMovement, 279±284, 286, 288±299, 301, 304n1, 305n3, 4, 307n15, 16, 308n27, 309nn28, 30±31 Pron-Crit/Pronoun-Criterion, 289±292, 296± 299, 306n13, 15, 307n16, 308n26 PrPr/ProjectionPrinciple, 433±434, 437, 439, 440±441, 444, 448, 454±458, 461 Pure-EP, 283±284, 286±292, 297, 305n7, 306n14 Subject, 5±12, 14, 21n2, 22n9, 23n18, 151±153, 156±164, 166n17, 316±317, 320, 322, 325, 327± 341 V-in-V/Verb-in-V-zero, 438±448, 452, 454±455, 458, 461 Wh-Crit/Wh-Criterion, 282±282, 286±292, 297, 308n26 Superiority e¨ects, 282±285 Swahili, 122 Swedish, 122 Thai, 397, 407, 409±410, 412±414, 416±417, 422n4, 422n18 Thematic roles (theta-roles), 113, 440±454. See also Semantic roles Topic, topicality, 115±116, 121, 147, 148, 149, 152, 156, 159, 161, 163, 176±178, 180, 183±187, 189± 191, 193±198, 200n14, 360, 416, 418, 121 Topicalization, 144, 146±147 Trentino, 184 Typology (general issues) factorial, 13, 15±16, 18, 315±341 via harmonic bounding, 320±321, 324±326, 341 monotonic enrichment, 321±322 Uradhi, 531±532 Voice, 355±387, 559, 561±569, 573±574, 577, 580, 584

Index active/passive, 63, 71±82, 356±359, 361, 363, 366, 369±372, 374, 376, 379±387 antipassive, 356±358, 361, 363, 366, 370, 372, 374, 376, 379±381, 383±387 direct/inverse, 62, 71±72, 88±89, 361±363 Warumungu, 131 Walpiri, 129, 272n11, 385 Weak crossover, 145 West Flemish, 122 Winnebago, 133 Word Order preservation of, 279±280, 282, 297, 301±304, 305n3 unmarked/basic, 171, 197, 315±316, 319, 321, 336, 341, 349n4

Language, Speech, and Communication Statistical Language Learning, Eugene Charniak, 1994 The Development of Speech Perception, edited by Judith Goodman and Howard C. Nusbaum, 1994 Construal, Lyn Frazier and Charles Clifton, Jr., 1995 The Generative Lexicon, James Pustejovsky, 1996 The Origins of Grammar: Evidence from Early Language Comprehension, Kathy Hirsh-Pasek and Roberta Michnick Golinko¨, 1996 Language and Space, edited by Paul Bloom, Mary A. Peterson, Lynn Nadel, and Merrill F. Garrett, 1996 Corpus Processing for Lexical Acquisition, edited by Branimir Boguraev and James Pustejovsky, 1996 Methods for Assessing Children's Syntax, edited by Dana McDaniel, Cecile McKee, and Helen Smith Cairns, 1996 The Balancing Act: Combining Symbolic and Statistical Approaches to Language, edited by Judith Klavans and Philip Resnik, 1996 The Discovery of Spoken Language, Peter W. Jusczyk, 1996 Lexical Competence, Diego Marconi, 1997 Finite-State Language Processing, edited by Emmanuel Roche and Yves Schabes, 1997 Children with Speci®c Language Impairment, Laurence B. Leonard, 1997 Type-Logical Semantics, Bob Carpenter, 1997 Statistical Methods for Speech Recognition, Frederick Jelinek, 1997 WordNet: An Electronic Lexical Database, Christiane Fellbaum, 1998 WordNet 1.6 CD-ROM, edited by Christiane Fellbaum, 1998 Investigations in Universal Grammar: A Guide to Experiments on the Acquisition of Syntax and Semantics, Stephen Crain and Rosalind Thornton, 1998 A Prosodic Model of Sign Language Phonology, Diane Brentari, 1998 Language Form and Language Function, Frederick J. Newmeyer, 1998 Semantics and Syntax in Lexical Functional Grammar: The Resource Logic Approach, edited by Mary Dalrymple, 1998 Understanding Language Understanding: Computational Models of Reading, edited by Ashwin Ram and Kenneth Moorman, 1999 The Syntactic Process, Mark Steedman, 1999 Toward a Cognitive Semantics, Volume 1: Concept Structuring Systems, Leonard Talmy, 1999 Toward a Cognitive Semantics, Volume 2: Typology and Process, Leonard Talmy, 1999 The Syntax of American Sign Language: Functional Categories and Hierarchical Structure, Carol Neidle, Judy Kegl, Dawn MacLaughlin, Benjamin Haban, and Robert G. Lee, 1999 Presumptive Meanings: The Theory of Generalized Conversational Implicature, Stephen C. Levinson, 1999 Converging Methods for Understanding Reading and Dyslexia, edited by Raymond M. Klein and Patricia A. McMullen, 1999 Optimality-Theoretic Syntax, edited by GeÂraldine Legendre, Jane Grimshaw, and Sten Vikner, 2001