This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
The Lexicon–Syntax Interface in Second Language Aquisition
Language Acquisition & Language Disorders Volumes in this series provide a forum for research contributing to theories of language acquistion (first and second, child and adult), language learnability, language attrition and language disorders.
Series Editors Harald Clahsen
Lydia White
University of Essex
McGill University
Editorial Board Melissa F. Bowerman
Luigi Rizzi
Max Planck Institut für Psycholinguistik, Nijmegen
University of Siena
Katherine Demuth
Bonnie D. Schwartz
Brown University
University of Hawaii at Manao
Wolfgang U. Dressler
Antonella Sorace
Universität Wien
University of Edinburgh
Nina Hyams
Karin Stromswold
University of California at Los Angeles
Rutgers University
Jürgen M. Meisel
Jürgen Weissenborn
Universität Hamburg
Universität Potsdam
William O’Grady
Frank Wijnen
University of Hawaii
Utrecht University
Mabel Rice University of Kansas
Volume 30 The Lexicon–Syntax Interface in Second Language Aquisition Edited by Roeland van Hout, Aafke Hulk, Folkert Kuiken and Richard Towell
The Lexicon–Syntax Interface in Second Language Aquisition Edited by
Roeland van Hout University of Nijmegen
Aafke Hulk University of Amsterdam
Folkert Kuiken University of Amsterdam
Richard Towell University of Salford
John Benjamins Publishing Company Amsterdam/Philadelphia
8
TM
The paper used in this publication meets the minimum requirements of American National Standard for Information Sciences – Permanence of Paper for Printed Library Materials, ansi z39.48-1984.
Library of Congress Cataloging-in-Publication Data The lexicon–syntax interface in second language aquisition / edited by Roeland van Hout, Aafke Hulk, Folkert Kuiken and Richard Towell. p. cm. (Language Acquisition and Language Disorders, issn 0925–0123 ; v. 30) Includes bibliographical references and index. 1. Second language acquisition. 2. Grammar, Comparative and general--Syntax. 3. Lexicology. I. Title: Lexicon-syntax interface in 2nd language acquisition. II. Hout, Roeland van. III. Series. P118.2.L49 2003 418-dc21 isbn 90 272 2499 4 (Eur.) / 1 58811 418 X (US) (Hb; alk. paper)
This volume contains a selection of papers presented at the NWCL/LOT Expert Seminar on ‘The interface between syntax and the lexicon in second language acquisition’, held in Amsterdam on March 30–31, 2001. The seminar was organized by the editors of this volume. We want to thank the participants of the seminar for reviewing the papers submitted for this volume. We are very grateful to the following institutions for financial support: the North West Centre for Linguistics (NWCL), the Landelijke Onderzoekschool Taalwetenschap (LOT; Netherlands Graduate School of Linguistics) and from the University of Amsterdam: the Amsterdam Center for Language and Communication (ACLC), the Chair of Linguistics of the Romance Languages and the Chair of Second Language Acquisition.
February 2003, The editors
Chapter 1
Introduction Second language acquisition research in search of an interface Richard Towell University of Salford
1.
Introduction
If it is to attain its eventual goal, second language acquisition research has to integrate the totality of second language acquisition processes. These must include the learning of the core syntax of a second language, the learning of the lexical items and determining the role of the cognitive mechanisms which are necessary for the use of linguistic forms in comprehension and production. It has been accepted for a long time that these three domains involve different kinds of learning: syntax is learnt through a process of implementing a particular set of universal structures (Chomsky 1986, White 1989); lexis is learnt by establishing a set of arbitrary associations which operate in a given society (Waxman 1996); comprehension and production are reliant on general cognitive procedures (Harley 2001). The learning of syntax is often characterised as a process of triggering (Sakas and J. D. Fodor 2001); the learning of lexis is characterised by the building up of associations (or connections) (Schreuder and Weltens 1993); comprehension and production are learnt by establishing and practising the required procedures (Pinker 1997). However, these three systems must come together in the creation of a whole linguistic capacity in the mind of an individual. The syntax will govern the structure of the grammar but the lexical items will govern how the structure is implemented. The linguistic knowledge which results from the interaction of these two systems can only develop and then find expression through the cognitive mechanisms associated with language comprehension and production. The researchers who attempt to provide accounts of the processes and outcomes of second language acquisition (SLA) are generally all too aware,
2
Richard Towell
therefore, that they have set themselves an ambitious, interdisciplinary task. Ideally, as a group, they wish to account for all aspects of second language acquisition from the phonetic to the intercultural (see Mitchell and Myles 1998). In particular, they set themselves the task of explaining those factors which have long been recognised as specific to the acquisition of a second or foreign language as opposed to the mother tongue. These are usually identified as transfer or cross-linguistic influence, evidence of a specifiable route of acquisition regardless of first language background, variability (also known as optionality) in the language of individual learners, and incompleteness or fossilisation in the final state of the majority of acquirers (Towell and Hawkins 1994). Clearly no one researcher could ever hope to deal with all aspects. At different times, the nature of the activities which are being described has led to the involvement of scholars from disciplines ranging from acoustics to anthropology. The central disciplines involved have, however, always been linguistics and psychology. For what now seems a brief period in the 1950s, linguistics and psychology came together to provide what was then thought of as a complete description of what language was and how it was learned: a powerful combination of structuralist linguistics and behaviourist psychology (Gass and Selinker 2001). Unfortunately, this period is only talked about in today’s classes on second language acquisition in order to show how misguided both of these initiatives were, forgetting rather that, despite the radical shifts of views which have followed, these efforts laid the foundations of the disciplines within which we all situate our research. It is probably true to say that linguistics and psychology began to follow different routes after the devastating criticisms of Skinner’s (1957) Verbal Behaviour put forward by Chomsky (1959), although many psychologists still continued to attempt to interpret transformational theory in psychological terms. These attempts foundered as the derivational theory of complexity (the belief that the more complex transformations were, the longer they would take to process) was denied by linguists. Linguists pointed out that, although terms like ‘least effort’ and ‘economy’ were essential to their endeavours, they were not defined with regard to processing effort but in relation to linguistic simplicity: “Chomsky’s economy principles are unambiguously matters of competence, in that they pertain to representations and derivations internal to the language faculty and exclude relations beyond the interfaces” (Smith 1999:114). Maintaining this position, linguists have gone on to develop their discipline significantly but within the boundaries which they have seen as necessary. They have therefore done so with little reference to any insights from psychology. Mainstream
In search of an interface
generative linguists have focused on syntax. That mainstream focus provides the background for the articles by Hawkins and Liszka, Corver, Van de Craats and Duffield in this volume. However, during the same period, the study of language within psychology also made rapid strides in exploring many aspects of psycholinguistics (see Harley 2001). One of these has concentrated on the lexicon, as is demonstrated in the chapters by Dijkstra and Williams in this volume; others have explored issues of how language may be stored in the brain and have made use of imaging techniques to enable us to begin to relate our theoretical analyses to physical realities (Perani et al. 1996, Dewaele 2002). These are represented in this volume by the chapters by Sabourin and Haverkort and by Green. There has been the occasional fruitful interchange in some areas of the discipline from time to time but there has been no real examination of how the two disciplines have evolved with regard to SLA and whether there are more global reasons for looking towards collaboration. There are now signs that both groups of researchers are coming to an understanding that their particular view of the world may not suffice to account for the overall process and that each will have to understand more about what the other knows. Whilst this book does not pretend to complete that task, it will seek to present current examples of the way linguists think about second language acquisition and of the way researchers working within a more psychological frame of reference think about the same subject in such a way as to show how there is a degree of complementarity in the work being done, even if, at the highest levels of argument, we are unlikely to see a swift return to the unity of view of the 1950s (cf. Smith 1999: 174). The belief that the time is right to seek such complementarity is encouraged by two fairly recent developments. Within linguistics, the advent of the minimalist theory and its consequences for SLA has caused researchers to look again at the relative roles of syntax and lexis. Under the minimalist view which, as Corver demonstrates in chapter three, applies as much to interlanguages as to any other natural languages, syntax is thought to be universal. It is “constituted of invariant principles with options restricted to functional elements and general properties of the lexicon” (Chomsky 1995: 170). The invariant nature of the syntax is possible because the functional elements are now seen as part of the lexicon: “It is clear that the lexicon contains substantive elements (nouns, verbs…) … And it is reasonably clear that it contains some functional categories.” (Chomsky 1995: 240). More recently, SLA specialists (Hawkins 2001: 345, Herschensohn 2000:80) have stated rather more firmly that, under minimalism,
3
4
Richard Towell
functional categories should be seen as part of the lexicon. We will argue below that this change of emphasis — Smith (1999) argues cogently that minimalism is a natural evolution of the generativist enterprise rather than a revolution — may well modify the way in which linguists have to think about development in SLA and indeed about the other central features of SLA research outlined above. Within psychology, there has been a welcome renewal of interest in how second languages are acquired. Since the mid 1980s, we have seen attempts to account for second language learning drawing on a rich vein of inductive research using computer modelling (Rumelhart and McClelland 1986). More recently, a variety of non-intrusive ways of providing physical evidence of brain processes have become available (Perani 1999). Both linguists and psychologists have become interested in how knowledge is stored in the mind and how it is retrieved from storage. Arguments have been put forward based on a distinction between declarative and procedural memory systems (Towell and Hawkins 1994, Ullman 2001) some of which suggest radical differences in the way first and second languages are acquired and stored in the mind. Most recently, arguments have been put forward to suggest that usage-based analyses can account for all linguistic units: “Psycholinguistic and cognitive linguistic theories of language acquisition hold that all linguistic units are abstracted from language use. In these usage based perspectives, the acquisition of grammar is the piecemeal learning of many thousands of constructions and the frequencybiased abstraction of regularities within them” (Ellis 2002: 144). All of this adds to the view that a full account of second language acquisition will require complementary input from both disciplines. One of the main keys must lie in how we see the concept of development and the psychological mechanisms which underlie development. The efforts of syntacticians focus on describing the syntactic structure which lies behind the interlanguage of the learner. This has always made it difficult for them to account for development (see Gregg 1996): the placing of functional categories within the lexicon makes this difficulty more acute. The invariant syntactic knowledge which learners have is a template present in the mind of the learner which can be modified by the information inserted within it. There cannot be a driving force for development in the syntax. It follows therefore that that driving force really comes from the lexis. However, up to now the learning of lexis has been thought of mainly in terms of one or other forms of associationist learning theory, with connectionism being the most powerful. Theorists pursuing this model have tended to argue that connectionist learning can account for the totality of language learning, including the learning of syntax
In search of an interface
(see quotation from Ellis (2002) above). But syntacticians cannot accept that the sophisticated structures which they observe and which provide no visible clues on the surface structure of the language can be learnt in this empirical fashion. They claim instead that innate knowledge (mediated or not by the L1) must be ‘guiding’ the learning (Hawkins 2001). It is not clear that this argument can be resolved by theoretical debates between ‘nativists’ and ‘non-nativists’. We need to examine in detail the evidence of how learners acquire a second language. This evidence must come from a variety of sources using a variety of techniques. It will take us into questions of what it is that learners acquire, how they acquire it and how that process modifies their linguistic capacity in both knowledge and use. Hawkins, Corver and Van de Craats in this volume present clear empirical accounts of how specific features of syntactic and lexical knowledge play a fundamental role in second language development. Their accounts cannot, however, tell us everything we need to know about the mental processes involved. Duffield asks fundamental questions about the nature of the competence which is acquired. Dijkstra provides an account of the way in which bilinguals store and access their knowledge. Williams examines the way in which linguistic knowledge may be built up on the basis of distributional evidence. Sabourin and Haverkort and then Green look at the way in which the knowledge may be stored and used. In this way we can see that a full account of the acquisition of a second language involves the three systems outlined at the beginning of this chapter. We will argue that none of the current arguments will suffice alone to account for the total process and that it is essential to attempt to integrate the sources of knowledge available to us (see Jackendoff (2002) for a similarly motivated position). In this chapter we shall seek first to outline the research context from the generativist point of view, initially in general, and then specifically with regard to the contribution of minimalism. We will then examine in more detail the contribution of psychological research and seek to show how a degree of complementarity may well exist, if there is the will to look for it. In this way, we are seeking to provide a context for the more detailed studies which figure in the rest of the volume by means of which the value of their contribution can be seen against the background of the evolving discipline of SLA research.
5
6
Richard Towell
2. The Linguistics dimension: The generativist research paradigm In outlining the contribution of generativist research, we will first briefly review the nature of this research paradigm and the reasons why it has become so central to SLA research. 2.1 The generativist position Noam Chomsky (1986: 3) poses three basic questions which linguists need to be able to answer: 1. What constitutes knowledge of language? 2. How is language acquired? 3. How is knowledge of language put to use? He and his followers have in effect concentrated on the first question. It is essential to — but separable from — the understanding of the other two. Indeed, Chomsky argues that in order to obtain a proper answer to this question it is necessary to idealise the data to be examined away from issues of performance so that the researcher can gain insight into the abstract knowledge which the native speaker of a language possesses, i.e. that person’s linguistic competence. Furthermore, the generativist position adopts a modular view of the mind in which the child possesses an innate language faculty. This is conceived of as separate from those parts of the mind which are devoted to general cognitive skills associated with the processing of information (perception, comprehension, production) and memory. It is argued that, because linguistic structure is universal and is not signalled overtly on the surface of any of the languages of which it is a manifestation, it is not possible for a child to acquire knowledge of language (in the sense of syntactic competence) on the basis of exposure to surface cues alone. This is frequently referred to as the logical problem of language acquisition. Surface cues are necessary to provide an indication for the child as to which of the limited number of possible languages he or she is confronted with, but they could never be sufficient to provide knowledge of the kind of organisation which is present in the syntactic structure of language. As each child acquires this knowledge with no conscious effort, no explicit instruction, following a regular pattern of acquisition not reflected in the data to which the child is exposed and without making the mistakes which piecemeal learning would imply, generativists conclude that linguistic knowledge is a biological innate endowment for humankind. This
In search of an interface
endowment is what enables the child to know more than the surface of the language reveals: the surface forms act only as a trigger for the underlying knowledge which the child already possesses. It is important to highlight four significant aspects of this theoretical position as these will give rise to further comment below and will be dealt with in subsequent chapters. First, a generativist approach involves idealisation of the data to be examined. For the study of adult competence in the mother tongue the researcher can frequently consult his or her own linguistic knowledge as a representative sample of the idealised speech community. This is not possible in second language research. SLA research requires data gathering methods which can isolate linguistic competence from performance factors. Second, the primacy of syntax within the generative paradigm has led to a separation between syntax and semantics. This is not without its problems as more and more researchers are finding that semantic factors influence syntactic phenomena (Juffs 1998, 2001). Duffield in this volume is concerned with how competence can be successfully defined within this paradigm and proposes that it is necessary to conceive of competence at two levels, one of which is more related to the surface structures of language. Third, the conception of the acquisition of syntactic knowledge through a process of triggering has given rise to debate (Lightfoot 1993, Carroll 2000, Sakas and J. D. Fodor 2001). Hawkins, Corver and Van de Craats, through an examination of the acquisition of syntactic and lexical features, define more clearly the nature of the features which have to be learnt and discuss the role of the L1 in providing the initial knowledge. This might provide a more satisfactory conceptual basis at least with regard to the initial state. It does, however, leave open the question of how the learners use empirical evidence to move to subsequent states. Fourth, the generativist position for SLA has to adapt to the fact that a second language acquirer has already learnt one language. The issue of how learners transfer or access universal knowledge in cases where one language has already been learnt is one which may have to be looked at again within the minimalist paradigm. This is an issue for Hawkins and Liszka, for Corver, and for Van de Craats. 2.2 Generativist second language research The particular strength of this approach has been in providing syntactic analyses within a theoretical framework. This enabled SLA researchers to predict in a precise way what learners needed to acquire in order to develop their interlanguage system. During the 1990s, this was manifested mainly
7
8
Richard Towell
through analyses presented within the conceptual basis provided by parameter (re-)setting (Flynn 1987). In this framework, languages could be compared on the basis of a single underlying syntactic phenomenon which was independently theoretically motivated by Universal Grammar (UG). This underlying syntactic phenomenon would have several surface expressions not linked together by other linguistic theories. Clear statements could then be made about what learners needed to do in order to re-set their existing parameters to the setting needed for their second language. Clear predictions could also be made about what the learner language would look like if the re-setting took place and what it would look like if it did not. To give two examples: the pro-drop parameter contrasted the presence or absence in the learner’s interlanguage grammar of such diverse phenomena as null subjects, of expletive it and there, the permissibility of subject-verb inversion and the possibility of extracting a wh-subject across an overt complementiser (that-trace filter) in languages such as English compared with Spanish or Italian. The verb raising parameter linked differences in adverb placement, negation and the use of quantifiers in French and English. The hypothesis in both cases was that exposure to the second language would enable learners to re-set the parameter through triggering and that researchers would be able to measure the differences in the learners’ linguistic competence at different points in time. Thus, if a native speaker of English acquiring Spanish triggered the pro-drop parameter to the Spanish setting, that person would immediately acquire knowledge of all the elements linked in the parameter. If a native speaker of French learning English re-set the verb-raising parameter, issues to do with adverb placement, the position of negatives and of floated quantifiers would be solved at the same time. Whilst this research provided a very positive move forward in SLA work, the empirical evidence frequently did not bear out the view that parameter re-setting was the essential process of SLA learning of syntax. Many studies (Hawkins, Towell and Bazergui 1993, White 1991) showed what might be called partial parameter re-setting, in the sense that some of the elements identified were learnt together but others were not. Whilst learners did not produce ‘wild grammars’ i.e. grammars which fell outside the constraints of linguistic theory, they could not be seen simply to re-set a parameter. This called into question either the nature of the underlying linguistic definition of the parameter or the process involved. There was also some uncertainty about what might constitute a trigger: would learning any form which was a surface manifestation of one element of the parameter trigger the other forms or was one form particularly privileged to act as a trigger?
In search of an interface
2.2.1 The minimalist perspective There are several important differences between the principles and parameters (P and P) model of syntax and the minimalist version. The most important one is the claim that the syntax is invariant and that the morpholexical system is the source of all variation. This has important implications for second language researchers. Herschensohn (1999) argues convincingly that minimalism should be better able to account for all of the main features of second language acquisition as defined above. It will be better at dealing with the area where it has always been most successful viz transfer but it should also be in a position to give a better account of the route of learning, incompleteness and variability (optionality). The problem with the P and P model was that it was all or nothing: either the parameter had been re-set and all features fell into place or it had not and they did not. As pointed out above, investigations based on this theory tended to find that partial re-setting took place, but the theory itself could not account for ‘partiality’ given that the re-setting process was one of ‘switch-flipping’. The notion contained within minimalism that acquisition proceeds more through “the gradual building of L2 grammar through the control of morpholexical constructions” (Herschensohn 1999: 81) allows for learners to be aware of the need to apply certain features or categories in some circumstances but not others. Hawkins (2001) and Hawkins and Liszka in this volume would probably agree that the minimalist approach opens the door to a more satisfactory account of variability (optionality) and incompleteness, but they base their arguments more on the presence or absence of features in the functional categories of the L2 than on the build up of constructions. Hawkins and Liszka also set out a specific view on the relationship between the L1 and the L2. They show that Chinese learners learning English do not mark tense consistently. Having investigated and rejected a range of alternative proposals, Hawkins and Liszka argue that this is likely to be because Chinese learners are unable to establish that the functional category English T is specified for +/- past. They suggest, furthermore, that this feature is not available to the learners because it does not exist in their first language. They claim that “where parametrised syntactic features are not present in a speaker’s L1, they will not be accessible in later L2 acquisition”. Such a point of view, if substantiated, would argue against full transfer from the L1 and would substantiate the partial access hypothesis. In the presentation of the article, Hawkins and Liszka contrast their account with that of Lardiere who adopts a full access point of view. Lardiere’s explanation is that the evidence suggests a failure of mapping from one component within the
9
10
Richard Towell
language faculty to another. Hawkins and Liszka feel that the evidence from Chinese learners argues more in favour of a partial transfer perspective. If Hawkins and Liszka are correct in their analysis, there are considerable consequences for acquisition in a more general sense. If L2 learners really cannot provide a feature based analysis for this part of their interlanguage syntax, they will have to learn the required forms (and store them) in another way. Whilst the discussion provided by Hawkins and Liszka remains within the framework of the role of features within generative syntax, the discussion of the alternative strategies available to learners provides food for thought as to whether the learning of the non-integrated forms must then be carried out in a different way e.g. stored as declarative as opposed to procedural knowledge as discussed in the article by Green. The notion of partial transfer also contrasts with the articles by Corver and Van de Craats. They allow for the full transfer of both lexical and syntactic features from the L1 to the L2. Their argument is that the full transfer of the L1 features into the L2 provides the starting point for learners beginning to acquire an L2. This is combined with a conservative strategy i.e. one in which the learners maintain those features unless and until they perceive the need to modify them. The learners create a series of interlanguages, all of which remain within the constraints of UG and thus provide checkable and interpretable information for the internal and external interfaces. As they progressively modify the features in response to the differences which they perceive on the basis of the input they receive from speakers of their ‘target’ L2, their interlanguages move towards a point where their interlingual ‘perfect’ systems correspond more to the natural language used by the L2 speakers. The above comments should serve to show that the generativist perspective on second language acquisition has the power to create well-defined hypotheses about the nature of language and to turn these into clearly defined investigative strategies. The shift to minimalism means, however, that the learning of items in the lexicon is potentially more significant than it was previously. The evidence from the Van de Craats article in particular shows how learners progressively revise their featural specification of both functional and lexical categories. As these are both part of the lexicon, it is clear that, from the minimalist perspective, the way in which second language learners come to revise their lexical forms — especially those which are linked to functional categories — is now a more important issue. We have suggested above that it provides the driving force for acquisition. The process of coming to know that forms and features need to be revised, however, necessarily involves comparative
In search of an interface
perception of language forms. How the learners trigger their knowledge or how they perceive the differences is something which cannot easily be dealt with within generative linguistics as currently defined, because it is more related to performance than competence. As soon as we mention perception we need to return to the psychological dimension and look again at second language acquisition from that perspective.
3. The psychological dimension It is probably as well to recognise immediately that the issues which divided linguists and psychologists fifty years ago have not gone away. Mainstream linguists still work within a rationalist framework which contrasts with the psychologists’ emphasis on empiricism. Linguists tend to reason on a top-down basis, psychologists base their theories on bottom-up evidence. Linguists believe in an innate, biological endowment specific to language; psychologists believe that language learning is one manifestation of cognition amongst others. Linguists believe that language is a symbolic system; some psychologists, at least, believe that it can be accounted for without the use of symbols. Linguists tend to reject computer modelling; many psychologists rely on computer modelling. Despite these differences, the argument that is being developed in this chapter (and which is the justification for this book) is that these differences actually provide us with perspectives which are complementary rather than contradictory, if we choose to look for the areas where insights from one field can contribute to the other (see Hulstijn 2002 for a similarly motivated view). Indeed, the fact that each field has excluded the domains covered by the other discipline surely leads to a position where significant aspects of the total process of SLA as described in Section 1 cannot be dealt with except by reference to the other discipline. It is therefore important to look more closely at the methods and results of psycholinguistic research in order to establish where the complementarity may lie. In order to do so we will now situate the articles in the book which have adopted a psychological reference point within the development of the methodologies used in psycholinguistics. Psychologists interested in language have made use of a variety of methods, many of which are shared with other branches of cognitive psychology. The three most important of these are experimental investigations which rely on the measurement of reaction times against a theoretically predicted outcome; the
11
12
Richard Towell
use of physical measurements of brain activity; and computer simulations of speaker or learner behaviour. Psychologists apply these methods to a variety of L1 and L2 speakers and to patients whose language ability is impaired in some way e.g. aphasics. All of these are represented in papers in this volume. Dijkstra’s article provides an excellent illustration of two out of the three methodologies typically encountered in the psycholinguistic literature. The problem space he addresses is whether bilinguals in possession of words from different languages (under his definition this includes language learners as ‘unbalanced bilinguals’) access both sets automatically in response to a given stimulus or whether one or the other is primed by different contexts or presentation methods. His stimuli include words which are cognates (similar in both form and meaning in the two languages), homographs (similar in orthographic form but not in meaning) and homophones (similar in sound but not in orthographic form or meaning). In a series of experiments Dijkstra and his colleagues have shown that bilinguals cannot do otherwise than access the items in both of their languages when presented with an applicable stimulus of isolated words (they call this nonselective access). They have also shown that frequency of use is the main determinant: highly automatised L1 words are accessed more swiftly than L2 words. But they have also shown that both items remain activated for a relatively long time before language selection takes place. Dijkstra also reports on a (limited) number of studies which examine words in sentential contexts by the measurement of Event Related brain Potentials (ERPs) through EEGs. The results have suggested that there are significant differences in the way bilinguals process their second language. Those related to semantic aspects seem quantitative in nature whilst those related to the syntax have a qualitative dimension as well, showing differences amongst early and late second language learners. Such studies, if replicated, could have considerable impact on the critical age hypothesis which is essential to some of the generative hypotheses, such as the general blocking principle discussed by Hawkins and Liszka. The study by Williams in this volume makes use at least in part of the third methodology regularly exploited by psycholinguists: computer modelling. Psychologists from Winograd (1972) onwards have been keen to exploit the processing abilities of computers for the purposes of modelling human behaviour. The most recent manifestation is connectionism which has been extensively (and controversially) used to model language and language learning. Connectionism (Bechtel and Abrahamsen 1991) attempts to show that many apparently complex processes can, in fact, be accounted for in a relatively simple way as long as the processing involved can be very large in quantitative
In search of an interface
terms and can operate in parallel (it is also known as parallel distributed processing or PDP). The basic idea is to have (a large number of) processing units which feed into one another at several levels (some of which are ‘hidden’) in sequence. The units involved can be given levels of activation or ‘weighting’ prior to any simulation. These may be random or specified in relation to the outcome envisaged. This will vary according to whether the model is being used simply to replicate what is assumed to happen in certain forms of processing or whether it is intended to model a developmental process (such as language learning). In the former case, random levels will be initially assigned and it will be intended that during the simulation the model will ‘learn’ new levels of activation. It is hoped that these levels of activation will represent eventually some kind of reality. Put (over-)simply, the lowest level or input units are then given different levels of activation and these levels feed through to the higher levels. Where the interaction with higher levels is ‘facilitated’ because certain units at those levels already have positive activation (excitation), the signal transmitted will be strengthened. Where the level of input activation encounters negative activation at a higher level, it will be ‘inhibited’ and, in the longer term, weakened. Over very many trials, the simulation establishes a stable level of output activation which is in essence derived through parallel processing from the relative frequency of the input it has received. This is arrived at through differentiated interaction between signals at the intermediate levels on the basis of the processes of strengthening and weakening. In a simulated task learning context, models can be modified in such a way as to include information about the extent to which the model is performing as it should in relation to some target (back-propagation). This means that the model can be trained to move nearer over time to a specified target. Researchers can then see how the model responds to the information it has been given and if and how it can ‘progress’ towards the target. It should be noted that this model only has one kind of unit which is linked to the signal strength of the connections: there is no symbolic level within connectionist models. The researchers can inspect the intermediate levels to see what activation levels the model produced at different times, they will be aware of different stages of ‘learning’ and they will see the extent to which the desired outcome is obtained and the relation with the input given. There is no doubt that theorising alone cannot discover the effect of enormous quantities of parallel processing and the computer is very efficient at examining this effect. The sixty thousand dollar question, however, is whether such modelling really reflects anything which goes on in the human brain.
13
14
Richard Towell
Those who favour this kind of modelling argue that it parallels human learning, that it can generalise beyond the data set which it is given and that it can produce ‘knowledge’ which is equivalent to what other approaches consider to be ‘abstract’ knowledge. It then follows naturally that proponents of this work see no need for abstract ‘symbolic’ categories in the mind because, from their point of view, these are ‘only’ the outcomes of the parallel processing described above, not primitive units. Numerous articles in the mid 1980s emanating from the PDP group led by Rumelhart and McClelland (Rumelhart, McClelland and the PDP Research Group 1986) made very strong claims about the way in which their network could parallel human language learning. These drew a detailed response, notably from Pinker and Prince (1988), in which the claims were thrown into question. As Fodor and Pylyshyn put it: “… Pinker and Prince argue (in effect) that more must be going on in learning past tense morphology than merely estimating correlations since the statistical hypothesis provides neither a close fit to the ontogenic data nor a plausible account of the adult data on which the ontogenic processes converge. It seems to us that Pinker and Prince have, by quite a lot, the best of this argument.” (Fodor and Pylyshyn 1988: 68). Connectionists have, however, gone on to refine their models and to enable them to call on other sets of information through which they have renewed their claim to model language. These are the issues which are interestingly explored in the study by Williams. He is critically interested in whether the learning which can be demonstrated by computer modelling really models that of humans or not. To investigate this, he combines computer-based simulations with experimental investigations on humans. The context of learning is that of the specification of abstract noun classes on the basis of a gender type classification. We already know that there is considerable evidence available to show that second language learners fail to assign gender in the consistent way that native speakers do. There are two central issues. The first is whether computer based inductive processes can genuinely be said to have gone beyond exemplar based generalisations to the creation of abstract classes. The second is whether what computers do and what humans do is in fact similar. Computers may very well demonstrate a remarkable ability to generalise from distributional examples, but do humans do the same? Do they rely on many examples to induce classes or do they induce classes in other ways, such as by making use of other clues e.g. animacy? Williams’ first simulation of the training kind with feedback seemed to show that the computer-based learning could generate productive knowledge
In search of an interface
of noun classes which enabled the network to generalise beyond the trained exemplars. It could behave as if it had formed abstract representations. His second simulation which did not involve feedback in the same way did not learn as well. Humans who learnt to classify the data into noun classes at 66% or above seemed to be using conscious explicit strategies, which clearly were not available to the computers, and/or to be influenced by prior knowledge of gender languages, equally not true of the computers. The conclusions which may be drawn are not simple: it seems as if some aspects of human behaviour may be similar to the inductive generalising of computers but that humans make use of other devices as well. Once feedback is introduced computer learning is considerably more powerful, but then it is probably more powerful than the human mind. Green and Sabourin and Haverkort take us into yet other areas of psycholinguistic research. Their concern is with how linguistic knowledge may be represented and stored in the mind. They share many reference points and to some extent a methodology. They both discuss evidence based on aphasics, they both rely on physiological evidence from ERPs to confirm or deny what other sources have indicated. Their conclusions appear to be slightly different. The broader argument which this research addresses concerns whether or not L1 and L2 learners acquire and store language in the same way. Virtually all researchers acknowledge the differences in learning environment for many L2 learners: they are generally older (beyond the so-called critical age) with fully developed memory and cognitive systems; they are often literate; they are already in possession of a first language and they are often exposed to explanations about language in classrooms as well as to language forms in more or less ‘authentic’ contexts of use. A key question has always been: do these differences mean that they will learn in a different way? Those who argue that they do suggest that they rely more on ‘explicit’ learning and that this will be reflected in the way their knowledge is stored in the mind. A separation is often made between ‘declarative’ knowledge, sometimes glossed as ‘knowing that’, the kind of knowledge which can be consciously accessed and articulated, such as a rule of grammar, and ‘procedural’ knowledge, sometime glossed as ‘knowing how’ i.e. the kind of knowledge which underlies skill activity, such as riding a bicycle, which cannot be consciously accessed. The two forms of knowledge are said to be acquired in different ways and stored in two different memories: a declarative memory and a procedural memory each of which is accessed in a different way (Anderson 1983, 1993, 2000). Declarative knowledge is acquired explicitly, consciously and quickly but cannot be used swiftly as a basis for any skill based action.
15
16
Richard Towell
Procedural knowledge is acquired implicitly, unconsciously and slowly by dint of a lot of practice. It is available swiftly in response to an appropriate stimulus. The argument has been put that L1 acquisition may be implicit and procedural and L2 acquisition may be explicit and declarative, although most researchers who discuss this issue allow for some overlap and some movement between categories over time as learners become more proficient and make their knowledge more ‘automatic’. Researchers such as Paradis (1997), however, claim categorically that explicit knowledge cannot become implicit. Green’s article questions the necessity for a separation between declarative and procedural memory and queries the evidence from aphasics on which he believes it is based. Computer modelling has indicated that the data derived from the aphasic experiments does not depend on having two memories. He therefore argues that it is worth looking at physiological evidence to see whether it confirms the necessity of two memory systems and whether there is evidence to suggest that L2 learners store knowledge in this differentiated way. His interpretation of the available evidence suggests that there is no difference for proficient L2 learners and he suggests that if there is a difference in the early stages of learning it soon disappears. Only longitudinal studies which included physiological studies could settle this argument conclusively. Sabourin and Haverkort do not refer explicitly to the two kinds of knowledge or the two memory systems outlined above but they do argue in favour of a clearer separation between the representation of linguistic knowledge and the representation of the knowledge which lies behind linguistic processing ability. This is because they believe that the empirical evidence that they have gathered by comparing the results obtained when proficient L2 learners are required to undertake the same grammaticality judgement test ‘off-line’ and ‘on-line’ shows that the knowledge base is different. The results of the ‘off-line’ task show no difference between the advanced learners and native speakers but the ‘on-line’ results do. When learners do a grammaticality judgement test in a paper and pencil way, they score as highly as native speakers. But when their performance on the same task is measured through ERPs, it is revealed that they are not responding in the same way. For Sabourin and Haverkort, this suggests that they are not accessing the same knowledge even though the observed outcome of correct answers is the same. Green therefore argues that the underlying knowledge base for advanced proficient learners is likely to be the same but Sabourin and Haverkort argue that it is likely to be different. Both agree that more research is needed.
In search of an interface
4. In what ways can the linguistic and psychological perspective be seen to be complementary? In this final section of the chapter, an attempt is made to build on the account given so far and to draw out some of the central themes which are treated in the following chapters. At the level of principle, as pointed out briefly in Section 3 above, there seems to be nothing but contradiction in the present stance of the founding disciplines of psycho-linguistics. And yet at a more pragmatic level, the accounts of the research outlined above suggest that the two disciplines may need each other rather more than they are prepared to admit. We will briefly explore this issue by looking at how interlanguages are created and how they develop bearing in mind the evidence presented in the various articles. Let us start with some notion of the initial state of knowledge for second language learners. This has to be defined for us by the linguists. In their chapters they argue cogently that learners transfer the features of lexical and functional categories from the L1 in ways which make up an operational interlanguage. Those features which are not available via transfer may be available through direct access to UG. In those many cases where the languages differ, the features are not combined in the same ‘bundles’ as those which are used by L2 speakers. The task for the learners is then to modify the relationship between features and forms in such a way as to create over time combinations which correspond more to the ‘bundles’ used by speakers of the L2. If they can create those bundles in an appropriate manner, Universal Grammar will ensure that they are interpretable by other cognitive systems. There are arguments about the extent to which the knowledge will transfer, and about the nature of competence, but the line of the argument is clear. The next question for second language acquisition researchers is how and why the necessary modification takes place. The linguist’s answer is that the learner must in some way come to know (but not in a conscious way) that the bundles of features in the interlanguage are not adequate to the purpose of full communication with native speakers in the L2. Once that unconscious realisation has taken place, the learner must then have a way of revising the existing feature bundles to make them correspond more to the L2 bundles. This is said by the linguists to happen implicitly and without conscious feedback. As was noted above, the term triggering is one which is frequently used in the literature but it is becoming more and more difficult to accept that what must be a complex process can adequately be summed up by that term. At this point we really have to turn to the psychologists to gain some insight into what may be happening.
17
18
Richard Towell
Dijkstra’s article makes it very clear that when possessors of two languages hear a given stimulus they activate all the relevant linguistic knowledge they have without separating it out into L1 and L2. All the activated forms have the possibility of giving rise to overt production: there is a selection process which determines which will actually be produced. The level of activation relates very much to frequency of use. This suggests that when second language learners acquire new forms they will compete with the existing forms. There will be differences in activation depending on the context when dealing with words in utterances (as opposed to the isolated words mostly studied) but the principle will hold that the ability to use second language forms accurately in fluent utterances will depend at least to some degree on the frequency of use experienced by the user. This raises interesting questions about how the activation of the interlingual forms for the purposes of communication in the interlanguage permits modification of the interlanguage system. If the effect of use is merely to strengthen the connections, as is implied by the non symbolic connectionist models explored by Williams and implied in the work of Dijkstra, then how can use give rise to modification of the forms? How also will they overcome the competition of existing forms? Sabourin and Haverkort suggest that whilst advanced L2 users may display the same knowledge in grammaticality judgement tests, the knowledge that lies behind their use of the language is not the same as that which lies behind the use in language production. Williams in his discussion of the comparative learning by humans and computers points out that humans appear to be influenced by factors other than the purely distributional. Is it possible that second language learners do indeed have some differential representation which allows them not to have to rely purely on the distributional analysis? Or is there another more symbolic version of implicit learning which could account for modification rather than strengthening? Green in his discussion of the storage of linguistic knowledge raises issues about the relative contribution of declarative and procedural knowledge. Within that area of reference, there are interesting questions about how learners deal with those lexical and functional categories which are not fully integrated within the syntax, as must be the case for interlingual systems which have not yet fully developed the syntactic system. We have seen that Hawkins and Liszka take the view that Chinese learners cannot integrate +/- past in the T category of their interlanguage syntax. They nonetheless produce correct past tense forms of regular verbs for at least some of the time. Where and how are these
In search of an interface
forms stored? If they are not generated by a function of the syntax, they must be stored as separate lexical items. This immediately opens the door to the notion that there must be a difference between the proportion of knowledge which is stored in the lexicon of the L1 and the L2 and in interlanguages: it would seem probable that in the early stages of learning at least second language learners must store a large proportion of the forms they have learnt as new lexical items and only work out later how they may be part of the syntax. Whilst there are no clear answers which immediately fall out of the existing state of knowledge, it should be evident that a combination of the insights of linguists and of psychologists will be required to answer these questions properly.
References Anderson J.R. 1983. The architecture of cognition. Cambridge, Mass.: Harvard University Press. Anderson J. R. 1993. Rules of the mind. New Jersey: Lawrence Erlbaum. Anderson J. R. 2000. Learning and memory. New York: John Wiley. Bechtel, W. and Abrahamsen, A. 1991. Connectionism and the mind. Oxford: Blackwell. Carroll, S. E. 2000. Input and evidence. Amsterdam: John Benjamins. Chomsky, N. 1959. “Review of Skinner 1957”. Language 35: 26–58. Chomsky, N. 1986. Knowledge of language. New York: Praeger. Chomsky, N. 1995. The minimalist program. Cambridge, Mass.: MIT Press. Dewaele, J. M. 2002 “Individual differences in L2 fluency: The effect of neurobiological correlates”. In Portraits of the L2 user, V. Cook (ed.), Clevedon: Multilingual Matters. Ellis, N. 2002. “Frequency effects in language processing: A review with implications for theories of implicit and explicit language acquisition”. Studies in Second Language Acquisition 24 (2): 143–189. Flynn, S. 1987. A parameter-setting model of L2 acquisition. Dordrecht: Reidel. Fodor, J. A. and Pylyshyn, Z. W. 1988. “Connectionism and cognitive architecture”. In Connections and symbols, Special Edition of Cognition, S. Pinker and J. Mehler (eds), 373. Cambridge, Mass.: MIT Press. Gass, S. and Selinker, L. 2001. Second language acquisition. New Jersey: Lawrence Erlbaum. Gregg, K. 1996. “The logical and developmental problems of second language acquisition”. In Handbook of second language acquisition, W. Ritchie and T. Bhatia (eds), San Diego: Academic Press. Harley, T. 2001. The psychology of language. Hove: Psychology Press. Hawkins, R. 2001. Second language syntax. Oxford: Blackwell. Hawkins, R., Towell, R. and Bazergui, N. 1993. “Universal Grammar and the acquisition of French verb movement by native speakers of English”. Second Language Research 9: 189–233. Herschensohn, J. 1999. The second time around minimalism and L2 acquisition. Amsterdam: John Benjamins.
19
20
Richard Towell
Hulstijn, J. 2002. “Towards a unified account of the representation, processing and acquisition of second language knowledge”. Second Language Research 18 (3): 193–224. Jackendoff, R. 2002. Foundations of language. Oxford: Oxford University Press. Juffs, A. 1998. “The Acquisition of semantics-syntax correspondences and verb frequencies in ESL materials”. Language Teaching Research 2: 93–123. Juffs, A. 2001. “Verb classes, event structure, and second language learners’ knowledge of semantics-syntax correspondences”. Studies in Second Language Acquisition 23: 305–313. Lightfoot, D. 1993. How to set parameters. Cambridge, Mass.: MIT Press. Mitchell, R. and Myles, F. 1998. Second language acquisition theories. London: Arnold. Paradis, M. 1997. “The cognitive neuropsychology of bilingualism”. In Tutorials in bilingualism. Psycholinguistic perspectives, A.M.B. de Groot and J.F. Kroll (eds), New Jersey: Lawrence Erlbaum. Perani, D. 1999. “The functional basis of memory: PET mapping of the memory systems in humans”. In Cognitive neuroscience of memory, L. G. Nilsson and H. J. Markovitsch (eds), 55–78. Seattle: Hogrefe and Huber. Perani, D., Dehaene, S., Grassi, F., Cohen, L. Cappa, S. F. and Dupoux, E. 1996. “Brain processing of native and foreign languages”. Neuroreport 7: 2439–2444. Pinker, S. 1997. How the mind works. Harmondsworth: Penguin. Pinker, S. and Prince, A. 1988. “On language and connectionism. Analysis of a parallel distributed processing model of language acquisition”. Cognition 28: 73–195. Rumelhart, D. and McClelland, J. 1986. “On learning the past tense of English verbs”. In Parallel distribued processing: Vol 1. Foundations, D. Rumelhart, J. McClelland and the PDP Research Group 1986. Cambridge, Mass.: MIT Press. Rumelhart, D., McClelland, J. and the PDP Research Group. 1986. Parallel distribued processing: Vol 1. Foundations. Cambridge, Mass.: MIT Press. Sakas, W. G. and Fodor, J. D. 2001. “The structural triggers learner”. In Language acquisition and learnability, S. Bertolo (ed.), 172–234. Cambridge: Cambridge University Press. Schreuder, R. and Weltens, B. (eds). 1993. The bilingual lexicon. Amsterdam: John Benjamins. Smith, N. 1999. Chomsky: Ideas and ideals. Cambridge: Cambridge University Press. Skinner, B. F. 1957. Verbal behaviour. New York: Appleton Century Crofts. Towell, R. and Hawkins, R. 1994. Approaches to second language acquisition. Clevedon: Multilingual Matters. Ullman, M. 2001. “The neural basis of lexicon and grammar in first and second language: The declarative/procedural model”. Bilingualism: Language and Cognition 4: 105–122. Waxman, S. 1996. “The development of an appreciation of specific linkages between linguistic and conceptual organisation”. In The acquisition of the lexicon, L. Gleitman and B. Landau (eds), Cambridge, Mass.: MIT Press. White, L. 1989. Universal Grammar and second language acquisition. Amsterdam: John Benjamins. White, L. 1991. “Adverb placement in second language acquisition: Some effects of negative evidence in the classroom”. Second Language Research 7: 133–61. Winograd, T. 1972. Understanding natural language. New York: Academic Press.
Chapter 2
Locating the source of defective past tense marking in advanced L2 English speakers* Roger Hawkins and Sarah Liszka University of Essex
1.
Introduction
It is well-known that advanced L2 speakers of English from certain L1 backgrounds show persistent optionality in marking thematic verbs for simple past tense in spontaneous oral production, as for example in The police caught the man and take him away. Speakers whose L1 is Chinese appear to be such a group. Bayley (1991, 1996) found the phenomenon sufficiently robust in Chinese speakers to undertake a variationist analysis of the factors which might be causing it. Wolfram and Hatfield (1984) had found similar optionality in the L2 English of Vietnamese speakers. More recently Lardiere (1998a, 1998b, 2000) has reported remarkably consistent optionality in simple past tense marking in a near-native speaker of English sampled with an eight-and-a-half year interval. Patty, a native speaker of Mandarin and Hokkien, marked simple past tense on thematic verbs around only one-third of the time in data collected from spontaneous speech. Native speakers, by contrast, do not apparently show endemic optionality of the same kind, although failure to inflect a verb for past tense is found sporadically in ‘slips of the tongue’ (Fromkin 1988).1 An important question for theories of SLA which assume that the mental grammars of individual L2 speakers are derived from Universal Grammar (UG) is why such optionality might exist in advanced/near-native speakers. Given that the morphophonology of forms in English provides clear positive evidence that past tense is marked, it is unexpected that advanced/near-native speakers should continue to have problems with it. Locating the source of the difficulty would be a small contribution to the broader goal of determining exactly how the language faculty is involved in the construction of grammatical knowledge by older L2 learners. Beck (1997) has argued that the kind of optionality in question is unlikely
22
Roger Hawkins and Sarah Liszka
to be the result of a deficit in the component of the language faculty which generates inflected phonological word forms: morphology. Beck compared the reaction times of 31 non-native speakers from a variety of L1 backgrounds with those of 32 natives on a task requiring the production of past-inflected verb forms. Speakers were presented with verb stems on a computer screen, and required to produce the simple past tense form orally, which activated a timing device. In previous studies with natives (e.g. Prasada, Pinker and Snyder 1990) it had been found that irregular verb forms show a frequency effect: the more frequent the stem of the verb in question (where frequency is defined as frequency of occurrence in corpora of English usage) the faster the reaction time to the past tense form. For regular verbs, however, frequency of the stem form had no effect on reaction time. Such findings have led to the claim that irregular past tense verb forms are stored associatively in memory (i.e. are ‘listed’) and hence show frequency effects as a function of strength of association between the stem and the listed form. By contrast, regular inflection is produced by rule, which applies in the same way to all regular stems, independently of frequency. Hence there is no reaction time effect. Beck’s findings with the non-native speakers were that they performed similarly to the natives: reaction times on low frequency irregular stems were slower than on high frequency stems (although not significantly so), and there was no difference in their performance on frequent and less frequent regular stems. This led Beck to suggest that it is not the morphological component which is involved in causing persistent optionality in past tense marking. Lardiere (1998a, 1998b, 2000) has suggested that the problem for Patty resides in the ‘mapping’ between fully specified syntactic phrase markers and surface morphophonology. Lardiere argues that other evidence from Patty’s spontaneous oral production suggests that she has a T(ense) category specified for finiteness. Firstly, her use of nominative case-marked pronouns is perfect; on standard assumptions the nominative case of subjects in English is the result of an agreement or checking relation between a T category specified [+finite] and the subject in the specifier of TP. Hence if Patty uses nominative pronouns perfectly she must have represented a finite feature on T. Secondly, there is no evidence for thematic verb raising (e.g. over negation) in Patty’s productions, suggesting that she has established that T has weak inflectional properties in English — another piece of evidence that Patty has a T specified for finiteness. Thirdly, there is evidence that Patty projects finite CPs, which on standard assumptions implies the presence of a T specified for finiteness. The ‘mapping’ difficulty is a problem accessing morphological forms which
Defective past tense marking in L2 English
have ‘layers’ of feature structure. Assuming a model of grammar where “an autonomous morphological component ‘reads’ the output of lexical and syntactic derivation, identifying those features which condition inflectional operations” (Lardiere 1998a: 20), Lardiere argues that the more inflectional features there are associated with a morphological form, the more likely a problem will arise for an L2 speaker. In the case of tense-marking, she assumes that the output of syntactic computations presents the morphological component with a terminal T node specified [+finite]. The morphological component must then determine whether it is [+past] or [−past], and if [+past] select suppletive forms in the case of irregular verbs, or invoke the regular rule for the affixation of -ed in the case of regular verbs. She proposes (speculatively) “that it is among the increasingly complex ‘outer’ layer mappings from morphology to PF that we are likely to find the greatest vulnerability to ‘fossilization’ and ‘critical period’ effects” (Lardiere 2000: 124). Additionally, if the phonological forms themselves involve complex phonology — for example, involving word final clusters like -kt, -skt and -mpst in the past tense forms of walked, asked and glimpsed — Lardiere argues that this may further affect successful mapping: “We can further imagine that an essentially morphophonological mapping procedure would be especially vulnerable to ‘derailment’ from a variety of postsyntactic or extra-syntactic factors, such as phonological transfer from the L1” (Lardiere 1998a: 21). Given that (Mandarin) Chinese is a language with basic (C)V(nasal) syllable structure and no syllable- or word-final consonant clusters (Hansen 2001), we can take Lardiere’s claim to be that while mapping of phonological forms onto terminal nodes which are the output of the syntax can cause problems generally for L2 speakers where layers of features are involved, in Patty’s case this is compounded by the fact that the L2 requires word-final consonant clusters where the L1 disallows them. The combined results of the studies by Beck and Lardiere point to the following possible conclusion: a deficit has occurred in one of the mechanisms of the morphological component — the mechanism referred to as the ‘vocabulary’ in models of distributed morphology (Halle and Marantz 1993, Embick and Noyer 2001) — which inserts vocabulary items (phonological forms) into terminal nodes where the feature specification of the vocabulary item and the feature specification of the terminal node are non-distinct. The representation of morphological forms themselves is not affected (as Beck’s study suggests), and the feature specification of categories manipulated by the syntax is not affected, if Lardiere’s analysis is correct. Moreover, the difference in phonotactic constraints between the L1 and L2 has a persistent influence, such that the mapping
23
24
Roger Hawkins and Sarah Liszka
problem is exacerbated. Thus this account of optionality in tense marking by L2 speakers is located at the interface between syntax and morphology. In this chapter we test this claim by comparing the spontaneous oral production of advanced L2 speakers of English from three different L1 backgrounds: Chinese, Japanese and German. If there is a general mapping problem for L2 speakers at the interface between syntax and morphology involving feature matching at the point of vocabulary insertion, we would expect this to appear in all three groups. Since Japanese is similar to Chinese in its phonotactic structure (disallowing word-final consonant clusters) but German is like English (cf. word-final clusters such as -ntst: getanzt ‘danced’), we would expect mapping problems to be more marked in the Chinese and Japanese informants.2 In fact we will argue that neither of the predictions is borne out by the data. The Chinese informants do mark simple past tense optionally in oral production, as previous studies had found, but the Japanese and German speakers are significantly less likely to do so. We will also show that our Chinese informants appear to know the morphological properties of English past tense verb inflection, as Beck’s study suggests that they should, and that they do not generally appear to have problems producing word-final consonant clusters. This will lead us to argue for a different locus for their deficit: at the interface between the syntactic component and the lexicon. In particular, we will claim that the Chinese speakers have difficulty assigning the formal (i.e. syntacticallyrelevant) feature [past], which determines the morphophonological forms of verbs in English, to the feature inventory of the category T(ense) in the lexicon, because this feature is not selected in Chinese, and is subject to a critical period.
2. Assumptions about the organisation of the language faculty We follow the spirit of recent work within the minimalist program (Chomsky 1998, 1999, 2001) and assume that the language faculty has a number of universally fixed and invariant computational procedures, and provides a universal inventory of phonological, semantic and syntactic (formal) features F from which lexical items can be assembled. One subset of the computational procedures is the syntax, which is capable of a small number of empirically and conceptually necessary operations: merge, agree and move. The syntax takes items presented to it from the lexicon and combines them into expressions. These expressions are interpreted by a semantic component (LF) (whose procedures are themselves universally invariant), and which makes the syntactic
Defective past tense marking in L2 English
expressions ‘legible’ to the conceptual-intentional modules of mind. The syntactic expressions are also interpreted by morphological/phonological procedures which are universally uniform, and which make syntactic expressions ‘legible’ to the sensori-motor systems for the production and understanding of speech. The universal inventory F of semantic, phonological and syntactic features is crucial in this model, because it is from this set that features are selected for the assembly of a lexicon whose items provide the input to the computational procedures. Individual languages select a subset of features from F and assemble them into lexical items. It is at this point — the selection of particular features for the assembly of lexical items — that languages vary. We focus here on the selection of syntactic (formal) features. Syntactic features are those which initiate syntactic operations, for example, wh-movement, case agreement, N-to-D movement, and so on. Some selections of syntactic features appear to be obligatory. For example, finite T appears necessarily to select the syntactic feature required to activate structural nominative case. Nominals also appear to obligatorily select case features which render them active for the purpose of case agreement. Languages are uniform in selecting these features. Parametric differences between languages arise when they make different choices of optional syntactic features. The distinction between obligatory and optional syntactic features is important for understanding tense and how it is realised morphologically in languages like English and Chinese. The view we adopt is that there is a syntactic (i.e. semantically uninterpretable) tense feature which, for the sake of exposition, we call [±past], which is available in the universal inventory F, but which is optional. English has selected it, but Chinese has not. In English, the presence of [±past] in finite T has a consequence for the morphology of the verb. [±past], being a syntactic (formal) feature, is not semantically interpretable at LF and so must be eliminated from a syntactic expression before such interpretation takes place. This elimination is effected through a checking (or matching) of the features of T with the morphological features of inflected verb forms like was, had, walked, ran. There are various complications that arise in this checking/matching operation depending on whether the verb is a light verb (be, have, do) or a thematic verb (walk, run). We will not expand on these here (see Lasnik 1999, Embick and Noyer 2001). But the basic claim is that finite English T has syntactic [±past] features which have morphological consequences for V. By contrast, we claim that Chinese does not have syntactic [±past] features on T, although it does have a syntactic [±finite] feature (Li 1990: 18).
25
26
Roger Hawkins and Sarah Liszka
As a consequence, bare Vs in Chinese can be interpreted either as past or nonpast, depending on context: (1) Zhangsan kan dianying. Zhangsan see movie ‘Zhangsan is seeing OR saw a movie.’
We take this up in more detail in Section 4.
3. The study To test the claim that the source of optionality in tense marking by L2 speakers lies at the interface between the syntactic and morphological components of the language faculty, we selected advanced L2 speakers of English from different L1 backgrounds, devised a test aimed at measuring informants’ knowledge of the morphological processes involved in simple past tense marking, and collected a sample of spontaneous production data including simple past tense verb forms from the same informants. 3.1 Informants Advanced L2 speakers of English were selected for this study on the basis of their performance on an independent measure of general proficiency. This consisted of two components: (a) the written multiple-choice grammar test component of the Oxford Placement Test (Allan 1992) which has 100 items covering a range of the core morphosyntactic properties of English; (b) Nation’s (1990) ‘vocabulary levels test’, which was designed as a language teacher’s aid for giving help with vocabulary learning, but provides a rough notional measure of the size of a speaker’s vocabulary up to the 10,000-word level. Using this combined grammar/vocabulary test, we selected informants whose L1 was Chinese, Japanese or German and whose mean scores broadly matched at the upper end (over 80% correct). This produced an experimental set of informants of two Chinese, five Japanese and five German speakers.3 Details of the proficiency scores are given in Table 1. 3.2 Test of knowledge of morphology If Beck’s (1997) findings are generalisable, we expect to find that the informants in our study can productively manipulate morphological processes involved in past
Defective past tense marking in L2 English
Table 1.Proficiency test scores: experimental informants L1 Chinese (n = 2) Japanese (n = 5) German (n = 5)
Mean proficiency score (%)
Range (%)
86.6 85.7 90.7
83.4–89.7 83.4–87.0 85.8–96.0
tense marking in English. To test this we designed a task which required informants to inflect both real and invented (nonce) verb stems for simple past tense. Our reasoning was that if speakers know the verb morphology associated with past tense, it should make no difference whether real or nonce forms are involved. 3.2.1 Design The task was adapted from one used by Prasada and Pinker (1993) with native speakers (‘experiment 3’ in their study). Prasada and Pinker were interested in the extent to which natives would produce regular and irregular past tense forms when presented with nonce verbs, and in particular whether novel irregulars displaying the ‘prototypical’ phonological shape of partially productive past tense irregulars (like string, sling, fling, cling Æ strung, slung, flung, clung) would elicit novel irregular past tense forms, such as spling Æ splung. To elicit such responses, informants were presented with nonce forms and an invented definition for each — six at the top of each page of a test questionnaire — followed by six sentences, each with a blank, where one of the nonce forms belonged (Prasada and Pinker 1993: 24). In our test we were interested simply in whether informants would inflect verbs appropriately for past tense in clear past tense contexts, and whether their knowledge was generative in the sense of allowing them to inflect verbs they had never encountered before correctly. They were therefore presented with six verbs at the top of each page of a test questionnaire with definitions (as in the Prasada and Pinker study), but half were real and half invented (18 of them in fact taken from the set of prototypical regular and irregular nonce verbs used by Prasada and Pinker). Below each set of six verbs with their definitions were six sentences where informants had to decide which verb to insert, and what its form should be. These contexts required either a simple past or a present perfect form. Thus informants could not simply produce a past tense form by rote, but had to make a definite choice of tense appropriate to the context. A partial illustration of the form of the test is given in (2):
27
28
Roger Hawkins and Sarah Liszka
(2) SPLING:
If you spling, you blow something out of your mouth (like smoke rings or air bubbles). CUT: If you cut something, you make it shorter, or divide it, or break its surface. a. The ground staff haven’t marked out the tennis court or put up the net yet, but the head gardener claims that they __________ the grass in readiness. b. As he rose slowly, the diver __________ bubbles in short bursts.
There were 120 contexts in all, 60 of which were expected to elicit simple past tense verb forms, and of these 30 involved nonce verb stems (15 prototypical regular types, like blark, and 15 prototypical irregular types, like spling). A control group of native speakers (n = 5) also took the morphology test. 3.2.2 Results Table 2 displays the frequencies of inflected and uninflected real and nonce verbs in simple past tense contexts. A χ2 test comparing the total frequency of inflected and uninflected forms produced by the non-native groups shows no significant difference between
Table 2.Frequencies of real and nonce verbs inflected for simple past L1 Chi (n = 2)
Jap (n = 5)
Ger (n = 5)
Eng (n = 5)
Verb type
Inflected Score (%)
Score (%)
Score (%)
Score (%)
Real reg
Yes No
23 3
88.5 11.5
60 2
96.8 3.2
64 1
98.5 1.5
74 0
.100 .0
Real irreg
Yes No
29 0
.100 .0
64 2
97.0 3.0
65 0
.100 .0
73 0
.100 .0
Nonce reg Yes No
25 2
92.6 7.4
58 2
96.7 3.3
64 0
.100 .0
72 2
97.3 2.7
Nonce irreg Yes No
18 4
81.8 18.2
54 9
85.7 14.3
53 7
88.3 11.7
71 0
.100 .0
Total
95 9
91.3 8.7
236 15
94.0 6.0
246 8
96.9 3.1
290 2
99.3 0.7
Yes No
NB: Cases where an informant selected a verb form other than simple past (e.g. past progressive or perfect) are excluded from the table. Hence scores do not necessarily add up to the expected maximum.
Defective past tense marking in L2 English
them (χ2 = 4.94, df = 2, p < .05). A comparison between the non-natives as a group and the native controls shows that there is a significant difference (χ2 = 12.64 (with Yates’ correction factor), df = 1, p < .05). This difference appears to be located primarily in the extent to which the non-natives fail to inflect irregular nonce forms for simple past (where the native controls do inflect in 100% of cases). This is particularly interesting because there are no significant differences between non-natives and natives in inflecting regular nonce forms. The result seems to suggest that speakers know that certain irregular nonce forms are irregular, hence do not attach a regular inflection to them, but do not know what the inflected form should be, producing an uninflected form. Broadly, and assuming that in some sense the non-native speakers are distinguishing nonce regulars from irregulars, frequency of past tense marking in the responses of these advanced non-natives is very similar to those of the natives. This is consistent with Beck’s findings, and suggests that the morphological component is operating similarly in these speakers to the way it operates in natives. 3.3 Inflected simple past tense in spontaneous production Spontaneous oral data were collected from two tasks: the retelling of a short extract from a Charlie Chaplin film (Modern Times), and the recounting of a happy or exciting experience each informant had had. The data were recorded and transcribed, and only verbs in unambiguously simple past tense contexts (i.e. those where a native could use no other form) were counted. For the purposes of this study, only thematic verbs were scored (e.g. walked, but not modals, copula/auxiliary was or auxiliary had). Verbs in contexts which were phonologically ambiguous were also discounted (i.e. regular past tense verbs followed by homophonic stops as in walked to work, or interdental fricatives, as in chased them). All unambiguous forms thus counted were rechecked against the original recordings to ensure accuracy. The frequencies of inflected and uninflected forms across the three groups of non-native speakers are presented in Table 3. χ2 tests show that there is a significant difference between groups both on frequency of inflection with regular verbs (χ2 =30.49, df=2, p<.01) and with irregular verbs (χ2 =8.13, df=2, p<.05). The difference appears to be located entirely in the Chinese speakers’ performance. This is in contrast to the performance of these speakers on the morphology test, where there was no significant difference between the three non-native groups, either on real verbs or nonce forms.
29
30
Roger Hawkins and Sarah Liszka
Table 3.Frequencies of inflected/uninflected verbs in simple past tense contexts: spontaneous oral production L1 Chinese (n = 2)
Japanese (n = 5)
German (n = 5)
Score
(%)
Score
(%)
Score
(%)
Inflected 25 Uninflect. 15
62.5 37.5
137 12
91.9 8.1
52 2
96.3 3.7
40
.100
149
.100
54
.100
Irregular Inflected 64 Ininflect. 12
84.2 15.8
252 18
93.3 6.7
79 4
95.2 4.8
Total
.100
270
.100
83
.100
Verb type Regular Total
76
These results are problematic for the view that L2 speakers generally have difficulty mapping phonological forms (‘vocabulary items’) with ‘layers’ of morphological features onto terminal nodes generated by the syntax. If this were the case, we would expect all three groups to perform similarly on the spontaneous production task, but the Chinese speakers are performing significantly differently from the Japanese and German speakers.4 Is the difference the result of L1 phonology interfering with and depressing the performance of the Chinese informants? This is unlikely. Recall that the prediction was that if L1 phonology had such an effect, we would expect the Chinese and Japanese to experience similar problems because the relevant property (absence of word-final consonant clusters) is present in both L1s. However, to pursue this possibility further, we considered performance of the non-native speakers on consonant clusters elsewhere in their spontaneous production, specifically in monomorphemic words like most, kind. If word-final clusters are problematic in production, some evidence for this should surface in these forms. Table 4 compares informants’ retention of -t/-d in inflecting regular past tense verbs with their retention of -t/-d in monomorphemes. Frequencies are small, but suggestive. Although Chinese speakers do drop final -t/-d in monomorphemes, they do not do so as frequently as in the simple past. This result matches a similar finding in Bayley (1996), who in a group of 20 L1 Chinese speakers of intermediate and advanced proficiency in L2 English found a 65% retention of final -t/-d in monomorphemes versus a 44% retention in simple past tense verb forms. This is in marked contrast with the pattern
Defective past tense marking in L2 English
Table 4.Absence of word-final -t/-d in monomorphemes and regular simple past tense forms compared L1 Chinese
Japanese
German
Word type
-t/-d
Score
(%)
Score
(%)
Score
(%)
Monomorphemes
Present Absent
9 2
82 18
27 1
96 4
48 0
100 0
Simple past (Regular)
Present Absent
25 15
63 37
137 12
92 8
52 2
96 4
of -t/-d deletion found in many studies of native speakers (Labov 1989), as already observed. Natives are more likely to drop -t/-d when a morphological boundary is not involved than when it is. Another possibility is that spontaneous oral production introduces performance pressures which make it difficult for L2 speakers to access inflected past tense verb forms in real-time (Prévost and White 2000: 129). If this were the case, performance on the morphology test would be a better reflection of the informants’ competence, because it lessens such pressures, while performance in spontaneous oral use of English underrepresents informants’ competence. This also looks implausible in the context of the three-group comparison. If performance pressures were involved, we would expect them to surface in all three groups, not just in the Chinese speakers. However, to pursue the possibility further, we looked at performance of the informants in using regular inflected participles in cases like were scared of, be sliced, is released, which are identical to the simple past tense forms. Table 5 brings together the frequencies of inflected/uninflected regular participles, monomorphemes and regular simple past tense forms. Again, although small frequencies are involved, if performance pressures were responsible for producing optionality in simple past tense marking, we might expect to see some evidence in the production of participles, where layers of morphological features also seem to be involved: if a verb form is [−finite] then it is either [+durative] (walking) or [−durative]. If it is [−durative] it is either [+past] ((have) walked) or [−past] ((to) walk). In summary, the results from spontaneous oral production are the following: although the non-native groups were matched for general high proficiency in L2 English, the Chinese informants were significantly less likely to inflect
31
32
Roger Hawkins and Sarah Liszka
Table 5.Absence of word-final -t/-d in regular participles, monomorphemes and regular simple past tense forms compared L1 Chinese
Japanese
German
Word type
-t/-d
Score
(%)
Score
(%)
Score
(%)
Participles
Present Absent
10 0
100 0
23 0
100 0
55 0
100 0
Monomorphemes
Present Absent
9 2
82 18
27 1
96 4
48 0
100 0
Simple past (Regular)
Present Absent
25 15
63 37
137 12
92 8
52 2
96 4
both regular and irregular thematic verbs for past tense than the Japanese or German speakers. The Chinese speakers retained word-final -t/-d with monomorphemes more often than with regular past tense verb forms (suggesting the problem is not caused by word-final consonant clusters). The Chinese speakers were perfect in inflecting past participles, in contrast to simple past tense verb forms, suggesting that ‘performance pressures’ are unlikely as a source of omission of inflections with past tense verbs.
4. Discussion We are interested in locating the source of optionality in the marking of thematic verbs for simple past tense in oral production by certain groups of high proficiency L2 speakers of English, for example speakers of L1 Chinese. To address this problem comparative data were collected from L1 speakers of Chinese, Japanese and German matched for general proficiency as ‘advanced’ speakers of English. An important caveat in discussing the results is that the number of informants was small, therefore the generalisability of any conclusions drawn must be treated with caution. The data were elicited from a test of knowledge of past tense morphology, and tasks allowing free oral production (an anecdote and the retelling of a film). Results of the morphology test suggest that all three groups, including the Chinese speakers, know the morphological properties of past tense marking in the context of real English verbs, and are productively aware of the regular versus irregular distinction even with verbs
Defective past tense marking in L2 English
they have never encountered before (nonce forms). This is consistent with earlier findings of Beck (1997), who showed that the reaction times of nonnative speakers on a verb inflection task were parallel to those of native speakers on the same task. This led her to conclude that the source of optionality of past tense marking is not located in the operation of the morphological component, a claim with which we concur. The comparative results from the free oral production of our three experimental groups show that the Chinese speakers are significantly different from the other two groups, producing uninflected regular verbs in unambiguously past tense contexts in over one third of cases. This was prima facie evidence against the idea that L2 speakers in general have difficulty mapping phonological forms (‘vocabulary items’) onto syntactic terminal nodes where that mapping implicates layers of morphological structure. If this were a general problem for L2 speakers, and on the assumption that morphological properties do not transfer from the L1 to the L2, we expected all three groups to show evidence of it. By looking at the Chinese speakers’ performance on word-final consonant clusters in monomorphemic words, and their performance in inflecting past participles, we also established that optionality in producing final -t/-d was lesser in these contexts. This suggested that the source of the problem with simple past tense was unlikely to be phonological in nature or the result of performance pressure because similar optionality would be expected across these contexts too. Having eliminated the morphological component, and the mapping of morphophonological forms onto syntactic representations as likely sources of observed optionality in past tense marking in the informants studied, an alternative explanation is needed. In this final section we explore the consequences of assuming that optionality results from a failure of L2 learners to include a syntactic feature for tense, that we are calling [±past], among the features which make up the lexical item T. Where this is the case, T enters syntactic derivations without the feature which, for native speakers of English, eventually forces the insertion of vocabulary items inflected for past tense into terminal nodes. Given such a hypothesis, we need to account for why optionality is characteristic of the Chinese speakers in our sample, but not the Japanese or German speakers; why the Chinese speakers nevertheless have greater than chance success in marking verbs for past tense in past tense contexts; and why the Chinese speakers might be more successful in marking irregular past forms and regular participles than in marking regular past forms. Central to this account is an understanding of the relation between how tense is interpreted by
33
34
Roger Hawkins and Sarah Liszka
the semantic component, and the presence of a syntactic (formal) tense feature in T. We therefore examine this relationship more closely. Languages appear to differ in whether they ‘grammaticalise’ tense through special morphophonological forms or not. Chinese is standardly assumed to lack such morphophonological forms, although it does have verbal aspect markers like -le, -guo and -zhe (Li and Thompson 1981, Li 1990, Packard 2000). For example, as already noted in Section 2, the verb kan ‘see’ can be freely interpreted as present or past depending on context: (3) a.
Zhangsan kan dianying. Zhangsan see movie ‘Zhangsan is seeing a movie.’ b. Zhangsan zuotian kan dianying. Zhangsan yesterday see movie ‘Zhangsan saw a movie yesterday.’
By contrast, Japanese and German do appear to grammaticalise tense, like English. Japanese has the forms -ta (past) and -ru (non-past) which appear to be tense auxiliaries (Okuwaki 2000). Thematic verbs in German inflect for past versus non-past tense like English, with regular and irregular variants (e.g. regular: kaufen ‘buy’, ich kaufte ‘I bought’; irregular: singen ‘sing’, ich sang ‘I sang’).5 Why might some languages have grammatical exponents of a ‘past/non-past’ distinction while others, like Chinese, simply lack such exponents? An interesting perspective on this question can be found in the work of Chierchia (1998). Chierchia suggests that the presence of a syntactic property in a grammar which is associated with a particular semantic operation has the effect of inhibiting the free application of that operation. In his own study he suggests that the presence of articles in a language blocks the free application of the semantic operations which give nominals generic, definite or indefinite meanings. So in English, count Ns can only be interpreted as definite when the is present, but in Russian, which lacks articles, bare Ns can be freely interpreted as generic, definite or indefinite ‘depending, presumably, on the context’ (Chierchia 1998: 361). This idea has been extended by Takeda (1999: 103) into a Generalised Blocking Principle (GBP): if a language has a certain functional category in its lexicon, the free application of the semantic operation that has the same function as that syntactic category is blocked in that language. Taking ‘functional category’ here to mean ‘feature of a functional category’, what the GBP proposes in terms of tense is that the semantic operation which interprets a T-V configuration as past or non-past can apply freely except where
Defective past tense marking in L2 English
a language assigns a syntactic [±past] feature to T. This then requires overt morphological expression and blocks the free operation of tense interpretation. So while finite bare Vs in Chinese can potentially be interpreted as past, present or future, depending on the context, finite bare Vs in English can only be interpreted as non-past, because past tense interpretation is associated with the specific features which give rise to forms like walked, ran. Consider now how this idea might be implemented in a grammar which incorporates some version of distributed morphology (Halle and Marantz 1993, Embick and Noyer 2001), the kind of model assumed in the work of Lardiere (2000) and Prévost and White (2000). The syntactic component generates expressions from bundles of syntactic and semantic features using the operations merge, move and agree (Chomsky 1998, 1999, 2001). Expressions consist of strings of terminal nodes which are presented to the morphological component for the insertion of vocabulary items (phonological forms). Insertion is an automatic process which takes place where the features of a vocabulary item are non-distinct from the features of a terminal node. For example, the terminal string T-V will present the morphological component with features like [+finite, −past] or [+finite, +past]. Vocabulary items with the relevant features will then compete for insertion into this position. Two important properties of insertion are firstly that a vocabulary item does not necessarily have to have all of the features of the terminal node to be inserted; it is sufficient that the features of the item are non-distinct from those of the terminal node (i.e. they may be a subset of those features). Secondly, that where vocabulary items are in competition for insertion into a terminal node, the most highly specified item compatible with the features of the terminal node is the one which wins the competition for insertion (Lumsden 1992: 480). For example, suppose that variants of walk have the following feature specifications: (4) walks [V, +finite, 3p, +sing] walked [V, +finite, +past] walk [V ]
Given such specifications, only walked can be inserted into a terminal node with the features [+finite, +past], even though walk is unspecified for those features and is in principle available for insertion. The reason is that although both forms have feature specifications which are non-distinct from the terminal node, walked is the more highly specified item. With such an account of the interaction between syntax and morphology, it is easy to see how the proposal that adult L2 speakers have difficulty with the
35
36
Roger Hawkins and Sarah Liszka
mapping of morphophonological forms onto terminal nodes might be made explicit: the procedure for inserting more specified vocabulary items over less specified items is not operating categorically in L2 speakers; less specified forms are being inserted where they should not be. The account we wish to explore here, however, is one where the syntactic feature [±past] is absent from T in the terminal string which is the output of the syntactic computations. The claim we will make is the following: where parametrised syntactic features are not present in a speaker’s L1, they will not be accessible in later L2 acquisition. This is equivalent to saying that syntactic features which invoke the Generalised Blocking Principle that have not been activated in early life do not operate in adult SLA. This immediately distinguishes Japanese and German speakers on the one hand, from Chinese speakers on the other. Japanese and German both have morphosyntactic exponents of ‘past tense’, hence the underlying syntactic features which invoke the GBP. In principle this should allow them to determine that past tense verb forms in English are syntactically motivated. By contrast, for L1 speakers of Chinese, T does not have the syntactic feature [±past]. Our proposal entails that when Chinese speakers learn English they are unable to establish that English T is specified for [±past]. This would mean that in their English grammars the terminal string T-V will include [±finite] but not [±past]. The distinction between vocabulary items like walks, walked, walk would then have no syntactic motivation for them.6 This is one kind of answer to the first question: why is optionality in past tense marking characteristic of the Chinese speakers in our sample, but not the Japanese and German speakers? It is also clear, however, that the Chinese informants studied here have acquired vocabulary items with past tense forms, and use them (at least superficially) in a highly target-like way in the morphology test, and at above chance level in spontaneous production. What might explain this if our claim is correct? One possibility is that Chinese speakers analyse participles and irregular past tense verb forms differently from regular past tense forms: they have a different morphological status in their grammars. Past participles are arguably different from simple past tense forms even in native grammars, in that they are aspectual in nature (realising ‘perfectivity’), and result from a verb-internal word formation process, which does not involve the T-V configuration. Since our assumption is that L2 speakers do not have difficulty with morphological operations per se, it would not be surprising if they did not have difficulty with participle forms. In other words, we expect Chinese speakers to have difficulty
Defective past tense marking in L2 English
with simple past tense forms because they involve a syntactic feature missing from T, but not participle forms because they do not involve T. In the case of irregular past tense forms, one possibility is that the Chinese speakers have acquired them as items independent from the equivalent bare V forms; i.e. ran is not the past tense exponent of run as it is in native grammars, but an independently acquired word form, in the same way that, say, amble and saunter are independent word forms for native speakers. While this is speculative, there is some limited evidence consistent with it in the transcripts of the spontaneous production data. There are some cases of ‘doubly inflected’ verb forms such as in (5a). The same speaker who produced (5a) also used ran in clearly non-past environments ((5b) and (5c)): (5) a. The girl ranned not far away. b. You should ran away together. c. She could not ran any more.
If ran were an independent word form it could be inflected for past tense and be used in non-past contexts. Observe also that (5b) and (5c) involve the use of ran as a non-finite verb. Given the model of vocabulary insertion assumed, ran could not be an inflected variant of run specified for the feature [+past]. If it were, its feature specification would clash with the feature specification of the non-finite terminal nodes into which it is inserted in (5b) and (5c). Insertion should simply be impossible. If past participles and irregular forms like ran have a different morphological status in the L2 English grammars of Chinese speakers from regular past tense verb forms like walked, this would be one kind of answer to the question: why are Chinese speakers more successful in marking irregular past tense verbs and regular participles than regular past tense verbs? They are not treating them as past tense verb forms at all; rather they have independent lexical item status for Chinese speakers. However, given this account, we have little idea currently of what these forms might ‘mean’ for Chinese speakers. This leaves the need to give an account of optionality in the marking of regular thematic verbs for past tense. We have no real answer to give in this case. Firstly, we have to assume that there is some reason why Chinese speakers do not treat regular verbs as independent vocabulary items as in the case of irregulars. One possibility is that it is the very regularity of the inflection and the frequency of such forms in the input which forces a morphological analysis of them as rule-based variants of the bare V. The problem then is to explain why some forms, but not others, are inflected in spontaneous production. To give a
37
38
Roger Hawkins and Sarah Liszka
flavour of the problem, consider the following short extract from the transcript of one of the Chinese informants in our sample: “When I saw the film ‘Lonely and Hungry’ and it reminded me of the old time when life was very hard. Some people they were very hungry and they have no work to do. They really don’t want to steal. But they had no other choice and when they become so hungry and they really want to just get the food and just want to eat the food so they stole or they just do something bad. But it’s understandable. It was not their mistake … And I watch it maybe 20 years ago. But it … I didn’t remember clearly about what it talk about. I just laugh a lot. But now when we see the film, I think. It gave me much imagination.”
What might be prompting this speaker’s use of the bare verb forms want, watch, talk and laugh in a context where past tense reference is often clearly intended? We have little idea yet of what the answer might be, but some possibilities appear unlikely. One explanation for selective verb inflection that has been advanced in studies of the early stages of L2 acquisition is the ‘Aspect Hypothesis’. This maintains that English past markers in early grammars are associated preferentially with verbs/predicates which have ‘telicity’ as part of their meaning (‘achievements’ and ‘accomplishments’ in the terminology of Vendler (1967)). See Bardovi-Harlig (1999) for a review of work on this topic. Does the same pattern obtain in the high proficiency Chinese speakers investigated here? That is, are they treating the regular past tense form as a marker of inherent verbal/predicate aspect? A breakdown of inflected regular forms by verb type is given in Table 6. Table 6.Inflected past tense regular verbs by aspectual type: Chinese speakers
Inflected tokens %
Statives
Activities
0/4 0%
10/14 71%
Accomplishments Achievements 1/2 50%
14/20 70%
The results are not conclusive. While statives are not inflected at all, which would be consistent with the Aspect Hypothesis, activities (which are atelic) are inflected to the same degree as achievements (which are telic). Interestingly, Lardiere (2002) has also analysed Patty’s past tense verb forms in terms of whether there is a correlation with the telicity of the verb/predicate, and found that 40% of all telic verbs (112/277, covering both regulars and irregulars) were marked for past tense, while 35% of all atelic verbs (130/371) were past-marked.
Defective past tense marking in L2 English
This is a non-significant difference (using χ2). Thus, although the Aspect Hypothesis might be an explanation for the distribution of English past tense verbs in low proficiency L2 speakers, it looks unlikely as an explanation for optionality in the marking of past tense on regular verbs by high proficiency L2 speakers. Another unlikely possibility considered (and rejected) by Lardiere in the case of Patty is the ‘Discourse Hypothesis’ (Bardovi-Harlig 1995). This holds that L2 speakers of English may initially use past forms of verbs to mark foreground events in narratives, but not mark background events. ‘Foreground’ events are defined as clauses that move time forward in a narrative and which, if interchanged, would change the sequence of events in the narrative; ‘background’ events are often ‘out of sequence’, provide additional information about the foreground events, set the scene, evaluate or explain (Bardovi-Harlig 1995: 265–267). Lardiere calculates that 32% of Patty’s past-marked verb forms (13/41) describe foreground events, while 30% (11/38) describe background events.7 Again, while the Discourse Hypothesis might explain the distribution of forms in low proficiency L2 speakers, it looks unlikely to be the source of optionality in high proficiency L2 speakers. The only hypothesis we are able to advance at present (but for which there is scant evidence in our current sample) is the following: linguistic theory appears to need to allow operations which apply to strings post-syntactically, that is in the morphological component or following vocabulary insertion. For example, Chomsky (1999) proposes that English has an output condition which bars surface adjacency between V and direct objects where the V is an unaccusative or a passive. For example, although (6a) is the expected output expression generated by the syntax, it is less natural than (6b) or (6c). Chomsky claims that a post-syntactic output constraint forces the object to move either leftwards or rightwards: (6) a. There was placed a large bowl on the table. b. There was a large bowl placed ___ on the table. c. There was placed ___ on the table a large bowl.
The surface ordering of clitic clusters in French also appears to be the effect of a post-syntactic output condition (Perlmutter 1971). If two third person clitics cooccur, an accusative form precedes a dative, but if a first or second person and a third person clitic co-occur, a dative form precedes an accusative (as in (7)): (7) a.
Elle le lui donne. she it-acc him-dat gives ‘She is giving it to him.’
39
40
Roger Hawkins and Sarah Liszka
b. Elle me le donne. she me-dat it-acc gives ‘She is giving it to me.’
This suggests that the language faculty allows some monitoring of surface strings. Perhaps in the normal case for Chinese speakers the morphology inserts bare verb vocabulary items into T-V strings. But because output checking is a possibility available in the grammar, Chinese speakers monitor the ambient discourse for ‘pastness’ and insert V-ed forms when they are able to detect it. So, as in French where ordering of object clitics is determined by the person of the forms in question (first and second person must precede third person), for the Chinese speakers selection of a V-ed form is determined by ‘pastness’. The difference between the two cases is that whereas [person] is a feature present in the specification of the French pronoun vocabulary items, ‘pastness’ has to be determined on the basis of context. Because context is involved, the monitoring process is unstable and gives rise to the kind of apparently random use of bare and inflected verb forms illustrated in the sample of informant speech given above. This is consistent with two observations about the marking of regular verbs for past tense by Chinese speakers. First, individual speakers of high proficiency can differ markedly in the extent to which they are successful in inflecting verbs. Our informants are apparently more successful in spontaneous oral production than Patty, for example. Secondly, different modalities allow a greater degree of success in past tense marking by the same individual. So our informants were more successful on the morphology test than in oral production, and Lardiere (2002) reports that Patty’s written output (in the form of e-mail messages) shows higher proportions of past tense marking than her spoken output.
5. Conclusion: The interface between syntax and the lexicon in adult SLA In a comparison of the performance of three groups of highly proficient L2 speakers of English (with Chinese, Japanese and German as L1s), it was found that the marking of past tense was optional in the spoken English of the Chinese speakers, but not in the case of the Japanese and German speakers. With the caveat that caution is required in generalising from small numbers of informants, we argued that the Chinese speakers’ knowledge of English morphological processes is intact (as Beck (1997) had already argued), and that phonological
Defective past tense marking in L2 English
properties (i.e. -t/-d deletion) did not appear to be a factor for our speakers. Furthermore, we found that past participle and irregular past tense forms were inflected more consistently than regulars. Although previous studies have argued that L2 speakers can fully acquire the syntactic features of lexical items like T (Lardiere 1998a, 1998b, 2000, Prévost and White 2000), we explored the alternative possibility that Chinese speakers cannot establish [±past] on T in English precisely because this feature is absent in their L1. By contrast [±past] is present on T both in Japanese and German. If Chinese speakers do not have access to such a feature in the construction of L2 knowledge, but Japanese and German speakers do, this would explain the observed difference between them in inflecting thematic verbs for past tense in spontaneous production. We linked this ‘deficit’ in the Chinese speakers’ grammars to the idea that optional syntactic features, when selected, ‘block’ the free application of semantic operations (Chierchia 1998, Takeda 1999). Our account assumes that these optional properties are subject to a critical period (an idea which originates in the work of Tsimpli and Roussou (1991) and Smith and Tsimpli (1995)). We then explored, in highly speculative mode, reasons why Chinese speakers might be successful at all in marking thematic verbs for past tense. This line of enquiry raises an interesting possibility for the interface between the syntactic component and the lexicon in adult SLA. Within the spirit of ‘minimalist’ enquiry, it might be proposed that all of the resources and computational procedures of the language faculty — LF, syntax, morphology — are intact and operative in adult SLA. However, optional (parametrised) syntactic features which play a major role in tying morphosyntactic structure to semantic operations in the L1 acquisition of specific languages are unavailable if not activated in early life. In parsing L2 data, older L2 speakers can use all of the resources except for these features. Phonological forms (‘vocabulary items’) which are selected by such features in native grammars are not selected by those features in the grammars of L2 learners beyond the critical period. Highly proficient speakers must nevertheless find some motivation for them. We rejected some of the conceivable ways in which Chinese speakers might motivate past tense verb forms in English (the ‘Aspect Hypothesis’ and the ‘Discourse Hypothesis’), and we suggested that they might be operating in terms of an ‘output condition’ which monitors the T-V string in relation to the ‘pastness’ of the discourse.
41
42
Roger Hawkins and Sarah Liszka
Notes * Parts of this work have been presented to audiences at McGill (2000) and GASLA (2000). We are grateful for the comments of those present in both cases. We have also benefited greatly from the comments of Donna Lardiere on an earlier draft discussion of the ideas outlined here. She will not agree with our conclusions, but her views have helped us sharpen our thinking on a number of issues. 1. The phenomenon of -t/-d deletion from word-final consonant clusters attested widely in informal varieties of native English (Labov 1989), is strongly disfavoured in the context of the regular simple past tense (Bayley 1996: 109). We return to this below, since it appears that the pattern is the reverse in non-native speakers: -t/-d is more likely to be missing from the regular past tense forms of verbs than in other contexts. 2. The assumption underlying the predictions made in this paragraph is that the morphological features of vocabulary items do not transfer from the L1 to the L2. Donna Lardiere (p.c. and 2000: 116–117) points out that the assumption is not self-evident. If the “L1 exhibits the same (or higher) degree of [morphological] complexity [L2 learners] may know what to look for”. Whether L2 speakers do or do not transfer morphological properties from their L1 is obviously an empirical question. If it turns out that they do, the observations reported in this article will need to be reinterpreted. Given that evidence bearing on the empirical question is currently lacking (although see Jarvis and Odlin (2000) for some relevant discussion), for the purposes of this article it will be assumed that transfer of verbal morphological properties from L1 to L2 is not a factor. 3. The small size of the Chinese group is a reflection of the difficulty we had in locating L1 Chinese informants who can achieve a score of 80% or above on the proficiency test (rather than a difficulty in finding L1 Chinese speakers of L2 English per se). Since, however, for the purposes of the present investigation, the frequency of past tense marking in non-nativespeaker English is the focus of interest, the variation in group size is relatively unimportant. In further work it would be desirable to have larger samples of advanced English speakers of the L1s in question. 4. Unless, of course, L2 speakers whose L1s are morphologically complex ‘know what to look for’. See footnote 2. 5. However, it should be noted that the present perfect — ich habe gekauft — predominates in everyday spoken German in what would be simple past contexts in English (Comrie 1976). 6. A question that might be raised here is whether this shouldn’t predict random usage by the Chinese speakers of all three forms (Donna Lardiere p.c.). This would be the case only if L2 speakers assumed that morphological variants of a lexical item can occur in free variation. However, we argue subsequently that Chinese speakers attempt to establish representations which distinguish morphologically distinct forms. The problem for them is that they have to do so without the benefit of the syntactic feature [±past]. 7. In practice the criteria seem to allow for considerable variation between raters in deciding which events are in the foreground and which in the background. We had difficulty applying them to our own sample, and do not report the results here.
Defective past tense marking in L2 English
References Allan, D. 1992. Oxford placement test. Oxford: Oxford University Press. Bardovi-Harlig, K. 1995. “A narrative perspective on the development of the tense/aspect system in second language acquisition”. Studies in Second Language Acquisition 17: 263–291. Bardovi-Harlig, K. 1999. “From morpheme studies to temporal semantics: tense-aspect research in SLA”. Studies in Second Language Acquisition 21: 341–382. Bayley, R. 1991. Variation theory and second language learning: Linguistic and social constraints on interlanguage tense marking. Unpublished doctoral dissertation, Stanford University. Bayley, R. 1996. “Competing constraints on variation in the speech of adult Chinese learners of English”. In Second language acquisition and linguistic variation, R. Bayley and D. Preston (eds), 97–120. Amsterdam: John Benjamins. Beck, M-L. 1997. “Regular verbs, past tense and frequency: tracking down a potential source of NS/NNS competence differences”. Second Language Research 13: 93–115. Chierchia, G. 1998. “Reference to kinds across languages”. Natural Language Semantics 6: 339–405. Chomsky, N. 1998. “Minimalist inquiries: the framework”. MIT Working Papers in Linguistics 15: 1–56. Chomsky, N. 1999. “Derivation by phase”. Ms. MIT. Chomsky, N. 2001. “Beyond explanatory adequacy”. Ms. MIT. Comrie, B. 1976. Aspect. Cambridge: Cambridge University Press. Embick, D. and Noyer, R. 2001. “Movement operations after syntax”. Linguistic Inquiry 32: 555–595. Fromkin, V. 1988. “The grammatical aspects of speech errors”. In Linguistics: The Cambridge survey, Vol. II, F. Newmeyer (ed.), 117–138. Cambridge: Cambridge University Press. Halle, M. and Marantz, A. 1993. “Distributed morphology and the pieces of inflection”. In The view from building 20: Essays in linguistics in honor of Sylvain Bromberger, K. Hale and S. J. Keyser (eds), 111–176. Cambridge, MA: MIT Press. Hansen, J. 2001. “Linguistic constraints on the acquisition of English syllable codas by native speakers of Mandarin Chinese”. Applied Linguistics 22: 338–365. Jarvis, S. and Odlin, T. 2000: “Morphological type, spatial reference, and language transfer”. Studies in Second Language Acquisition 22: 535–556. Labov, W. 1989. “The child as linguistic historian”. Language Variation and Change 1: 85–98. Lardiere, D. 1998a. “Case and tense in the ‘fossilized’ steady state”. Second Language Research 14: 1–26. Lardiere, D. 1998b. “Dissociating syntax from morphology in a divergent L2 end-state grammar”. Second Language Research 14: 359–375. Lardiere, D. 2000. “Mapping features and forms in second language acquisition”. In Second language acquisition and linguistic theory, J. Archibald (ed.), 102–129. Malden, MA: Blackwell. Lardiere, D. 2002. “Second language knowledge of [±past] vs. [±finite]”. Paper presented at GASLA 6, University of Ottawa. Lasnik, H. 1999. Minimalist analysis. Malden, MA: Blackwell. Li, A. Y-H. 1990. Order and constituency in Mandarin Chinese. Dordrecht: Kluwer.
43
44
Roger Hawkins and Sarah Liszka
Li, C. N. and Thompson, S. 1981. Mandarin Chinese: A functional reference grammar. Berkeley: University of California Press. Lumsden, J. 1992. “Underspecification in grammatical and natural gender”. Linguistic Inquiry 23: 469–486. Nation, I. P. S. 1990. Teaching and learning vocabulary. Boston, MA: Heinle and Heinle. Okuwaki, N. 2000. “Japanese -ta as an auxiliary verb”. Ms. University of Essex. Packard, J. 2000. The morphology of Chinese: A linguistic and cognitive approach. Cambridge: Cambridge University Press. Perlmutter, D. 1971. Deep and surface structure constraints in syntax. New York: Holt, Rinehart and Winston. Prasada, S. and Pinker, S. 1993. “Generalization of regular and irregular morphological patterns”. Language and Cognitive Processes 8: 1–56. Prasada, S., Pinker, S. and Snyder, W. 1990. “Some evidence that irregular forms are retrieved from memory but regular forms are rule-generated”. Poster paper, 31st Annual Meeting of the Psychonomic Society, New Orleans. Prévost, P. and White, L. 2000. “Missing surface inflection or impairment in second language acquisition? Evidence from tense and agreement”. Second Language Research 16: 103–133. Smith, N. and Tsimpli, I-M. 1995. The mind of a savant: Language learning and modularity. Oxford: Blackwell. Takeda, K. 1999. Multiple headed structures. Unpublished doctoral dissertation, University of California, Irvine. Tsimpli, I-M. and Roussou, A. 1991. “Parameter resetting in L2?” University College Working Papers in Linguistics 3: 149–169. Vendler, Z. 1967. “Verbs and times”. In Linguistics and Philosophy, Z. Vendler (ed.), 97–121. Ithaca, NY: Cornell University Press. Wolfram, W. and Hatfield, D. 1984. “Tense marking in second language learning: patterns of spoken and written English in a Vietnamese community”. ERIC document ED 25 960. Washington, DC: Centre for Applied Linguistics.
Chapter 3
Perfect projections* Norbert Corver Utrecht University
1.
An interface perspective on L2-knowledge
A central question in current generative research is the question of how perfect a system language is (cf. Chomsky’s minimalist research program: Chomsky 1995, 2000a, 2000b). Perfection is defined here from an interface perspective: the grammatical information provided by the linguistic expressions that are generated by the language L must be legible to the external performance systems within which the language faculty is embedded. In view of the traditional assumption that language is a relation of sound and meaning, i.e. a mental phonetic representation and a mental meaning representation, there are two points of access from external systems. There is a sensory-motor system that is looking at P(honetic) F(orm) and reads off information provided by the PF-representation. And there is some language use or conceptual system that reads off the meaning information provided by the L(ogical) F(orm)-representation. If the linguistic expression generated by the language system is legible both on the sound side and on the meaning side, the expression is said to converge at both interface levels. If there is some element or property which cannot be interpreted at the interface, the expression crashes. The sentence John met Mary Bill crashes, for example, because one of the noun phrases (say Bill) does not receive an interpretation at the LF (i.e. meaning) interface. The noun phrases John and Mary are interpreted as arguments of the two-place verbal predicate met. The noun phrase Bill cannot receive an argumental interpretation, nor any other meaning interpretation, and therefore turns the sentence into an LF-representation that crashes at the meaning interface. Taking this interface perspective on human language, one could say that language is an optimally designed, perfect system. It makes linguistic information available to the external systems in a form which is accessible to them. Knowledge of language, then, can be defined in interface terms: a person has
46
Norbert Corver
knowledge of language L if he is able to form linguistic expressions (i.e. soundmeaning pairs) that are fully-interpretable at the interface levels. This requirement that the linguistic expressions (PF-LF pairs) generated by the computational system be legible to the external systems plausibly holds for any state of L1-knowledge (both interlanguage knowledge and final state grammatical knowledge).1 In all stages of language acquisition, the language system (i.e the grammar) must interact with the external systems. If it does not, the linguistic knowledge is not usable at all and consequently there would not be any output products. Thus, from this interface perspective we could say that the L1-products in any stage of language acquisition are perfect linguistic objects (PF-LF pairs) in the sense that they fully consist of interface-interpretable properties. What does the interface perspective on linguistic knowledge contribute to our view on second language knowledge (and second language acquisition)? In fact, the conclusion seems inescapable that L2-expressions are also perfect grammatical objects. If they were not, the L2-objects generated by the computational system would not be legible and usable at all by the external systems which interact with the ‘L2-grammar’. There simply would not be any output (i.e. utterances). The conclusion must be that L2-products are interpretable both on the meaning side and the sound side. And just like it does for the various L1-knowledge states, this conclusion should also hold for the various knowledge states of the L2-grammar (initial state, interlanguage states and target state).2 Of course, the conclusion that L2-representations are natural language objects, in the sense of being objects that fall within the bounds of Universal Grammar, is not new. Over the last two decades, various researchers interested in investigating the linguistic competence of L2-learners have argued and tried to show that the (mental) L2-representations generated by the interlanguage grammar are constrained by principles of UG (and consequently can be analysed in the same way as other (e.g. L1 linguistic data; cf. Bley-Vroman 1990, Schachter 1989, White 1988, 2000). The importance of the interface perspective on linguistic expressions is that the conclusion of UG-consistency seems inescapable: if interlanguage representations did not obey the bare output conditions (i.e. the interface legibility requirements), these representations would be illegible and inaccessible to the external cognitive systems with which the interlanguage grammar interacts. The linguistic expressions generated by the interlanguage grammar would simply not contain the right ‘instructions’ to the performance systems, and consequently, they would not be able to put it to use. And as a consequence of that, there would not be any output.
Perfect projections
Thus, the L2-system, just like the L1-system, is a perfect system in the sense of being a system that is optimally designed to meet external conditions (bare output conditions) imposed by other cognitive systems that the language faculty interacts with. From a different perspective, though, linguistic expressions produced by L2-learners seem to be highly imperfect. They very often deviate, to a greater or lesser extent, from the linguistic expressions produced by adult mother tongue learners of the language that is acquired. Some L2-expression E with meaning representation LFx very often differs from the equivalent L1-expression generated by a mother tongue speaker of the language. Consider, for example, the ‘imperfect’ L2-expressions in (1a)–(4a), which are produced by Turkish L2-learners of Dutch.3,4 The b-examples represent the corresponding target expressions. (1) a.
Ik komt huis en Slenol wegt huis Stokhasselt. I come home and Slenol away-3sg house Stokhasselt ‘I go home and Slenol went to his house in Stokhasselt.’ (L2-expression) b. Ik kom thuis en Slenol gaat (weg) naar z’n huis in Stokhasselt. I come home and Slenol goes (away) to his home in Stokhasselt. (target expression)
(2) a.
Altijd uh alles woonte Klirsehir van Turkije. always uh all/everything live(d) Klirsehir of Turkey ‘Everyone still lives in Klirsehir in Turkey.’ (L2-expression) b. Nog altijd woont iedereen in Klirsehir in Turkije. still always lives everyone in Klirsehir in Turkey ‘Everyone still lives in Klirsehir in Turkey.’ (target expression)
(3) a.
Ik gaan school. I go-inf school ‘I go to school.’ (L2-expression) b. Ik ga naar school. I go-1sg to school ‘I go to school.’ (target expression)
(4) a.
En dan andere jongens komt. and then other boys-pl come-sg ‘And then, the other boys come.’ (L2-expression) b. En dan komen de andere jongens. and then come-pl the other boys-pl ‘And then, the other boys come.’ (target expression)
47
48
Norbert Corver
The L2-expressions in the a-examples in (1)–(4) are (superficially) imperfect in the sense that they deviate from the expressions generated by the grammar of mother tongue speakers of Dutch. In (1a), imperfection relates to the lexical item wegt, a verbal form which does not exist in (target) Dutch. In (2a), imperfection concerns the use of the lexical item alles, a quantificational expression that does not express quantification over persons in the target language. In (3a), imperfection relates to the linguistic expression of ‘location’: there seems to be no prepositional element available, which carries the locational (i.e. path) interpretation of school. In (4a), finally, we appear to have an imperfect agreement relation between the (plural) subject-noun phrase and the finite verb. Even though these L2-expressions may be imperfect from the perspective of the target language, I hope to show in this paper that they are perfect from the perspective of the interface conditions.5 L2-projections are perfect projections in the sense that, at the interface level, they consist of features (associated with lexical items) that are interpretable for the interface conditions. One could say: L2-projections are perfect in being externally-interpretable. On the basis of the types of L2-expressions illustrated in (1)–(4), I hope to show in this article that target (im)perfection (i.e. (non-)correspondence with the target pattern) should be distinguished from interface (im)perfection. L2-products of interlanguage grammars are typically ‘target-imperfect’ but ‘interface-perfect’. Importantly, those target-imperfect but interface-perfect interlanguage expressions can be of two types. First of all, there are interlanguage expressions that are legible at LF because the equivalent L1 expression is legible at LF. An example of such an expression resulting from transfer (or conservation; see Van de Craats, Corver and Van Hout 2000, Van de Craats this volume) is given in (5). Secondly, there are target-imperfect interlanguage expressions that are legible at LF and whose imperfection results from other mechanisms, e.g. the (non-target) merger of some root element and an inflectional suffix. An example of such an expression is the element wegt in (1). (5) examen van tolk exam of interpreter ‘the interpreter at the exam’
Let me briefly dwell on the example in (5), which from a superficial perspective looks like a perfect target Dutch pattern (‘surface perfection’). In possessive structures of native speakers of Dutch, the element van is an adpositional (i.e. prepositional) marker that is interpreted as the spell-out of the abstract genitive case feature associated with the possessor DP in postnominal position (like in:
Perfect projections
dat boek van Jan, ‘that book of Jan’). A Dutch-based analysis of the sequence examen van tolk is highly unlikely, however; under such an analysis, in which tolk is the complement to the noun examen, the entire noun phrase needs to be interpreted as ‘the examen taken by the interpreter’. This is not the reading it has, which is ‘the interpreter at/of the exam’; examen acts as the ‘possessor’. Given this semantic interpretation, Van de Craats, Corver and Van Hout reach the conclusion that the syntactic structure associated with the linear sequence in (5) is a structure transferred (i.e. conserved) from the first language (i.e. Turkish). This amounts to an analysis according to which van is an inflectional suffix attached to the possessed noun. It is the equivalent of the inflectional element -nin in Turkish expressions like Ayse-nin araba-si (Ayse-gen car-3sg, ‘Ayse’s car’). The Turkish syntactic structure which underlies the Dutch surface sequence in (5) is then the one in (6a). (6b) represents the syntactic structure of the Turkish sequence Ayse-nin araba-si. (6) a. [DP [AgrP [examen-van]i [Agr¢ [NP ti tolk] Agr]] D] b. [DP [AgrP [Ayse-nin]i [Agr¢ [NP ti araba] si]] D]
In this article, I will consider examples of interface-legible interlanguage expressions that are ‘target-imperfect’. One may wonder what interface-illegible interlanguage expressions look like. Being illegible at the interface such illegitimate patterns should never surface in the derivational output. They are ruled out by the bare output conditions at the interface. Potential examples of such illegible interlanguage patterns can be made up, of course. In Van de Craats, Corver and Van Hout (2000), for example, it is noted that even though Turkish L2-learners produce a great variety of interlanguage possessive patterns (see (7) for some examples), certain imaginable patterns are not attested in the L2-data; for example, the patterns in (8). The absence of these patterns in the L2 derivational output hints at the characterization of these patterns as ‘interfaceillegible’ interlanguage expressions. The external systems somehow cannot ‘read’ the linguistic instructions provided by these structures and, as a consequence of that, these patterns are never present in the L2-output. (7) a.
examen van tolk exam of interpreter ‘the interpreter of/at the exam’ (attested) b. auto z’n lamp car its light ‘the car’s light’ (attested)
49
50
Norbert Corver
(8) a.
tolk examen van interpreter exam of ‘the interpreter of/at the exam’ (unattested) b. z’n lamp auto his light car ‘the car’s light’ (unattested)
Through discussion of some illustrative L2-expressions, I hope to show in this chapter that target (im)perfection should be distinguished from interface (im)perfection. L2-products of interlanguage grammars are typically ‘targetimperfect’ but ‘interface-perfect’. As I will show, interface perfection applies both at the level of words and at the level of phrasal categories (i.e. lexical projections). L2-words that may be imperfect from the point of view of the target language are (interface-)perfect at the lexicon-syntax interface (see Section 2). And L2-phrases that are target-imperfect turn out to be fully legible (hence perfect) at the LF-interface (see Sections 3, 4 and 5).
2. Perfect L2-words at the lexicon-syntax interface Chomsky (1995) systematically refers to PF and LF as interfaces of syntax with respectively a perception/articulation system and an interpretation/use system — both being mental faculties. The syntactic structure generated by the computational system (say, merge and move/attract) is assigned a PF-representation and an LF-representation. Besides the PF- and LF-interface of syntax, a third interface of the syntax can be identified, viz. the syntactic representation built up from the lexicon. This is explicitly stated in Chomsky (1991: 46): …that there are three ‘fundamental’ levels of representation: D-structure, PF, and LF. Each constitutes an ‘interface’ of the syntax (broadly constructed) with other systems: D-structure is a projection of the lexicon, via the mechanisms of X-bar theory; PF is associated with articulation and perception, and LF with semantic interpretation.
Although a separate level of D-structure is no longer adopted in minimalist theorizing, the general idea that syntactic structure is built up from the lexicon still is a core assumption of generative linguistics. This is also clear from the following statement by Chomksy (1995: 225):
Perfect projections
Another natural condition is that outputs consist of nothing beyond properties of items of the lexicon (lexical features) — in other words, that the interface levels consist of nothing more than arrangements of lexical features.
For the lexicon-syntax interface, this implies that each lexical item (i.e. a constellation of lexical features) must be legible to the computational system (merge and attract/move) that accesses these objects (i.e. lexical expressions) and builds more complex expressions (i.e. syntactic expressions) from those lexical expressions. A question which then arises is: What makes a lexical item legible to the computational system (i.e. the ‘rules’ of grammar: e.g. merge, move/attract, agree)? To answer this question, let us first address the question of what a lexical item (LI) is. In line with De Saussure’s conception of words, a LI is typically defined as a sound-meaning pair (i.e. a PF-LF pair). In a sense, a LI is a structured object with a sound representation (the phonological matrix; sound properties) and a meaning representation (semantic properties). Presumably it is not the phonological and purely semantic properties which make a lexical item legible to the computational system. If a LI were just a sound-meaning pair one could wonder what the syntax (i.e. the recursive procedure) should do with it. That is, what would make such a sound-meaning pair legible to the computational system? Phonetic and purely semantic features do not seem to be accessed by the recursive syntactic procedures. In short, these PF-LF-pairs would remain illegible to the computational system. So, what makes these sound-meaning pairs legible to the computational rules, which combine these pairs into more complex sound-meaning constructs? The answer is: formal (i.e. syntactic) features. Suppose a formal feature must be ‘added’ (merged) to the sound-meaning pair (i.e. the lexical item) for the LI to be legible to the computational processes that generate larger structures. In other words, merger of a categorial feature with the sound-meaning pair turns the lexical item into an object that is visible at the lexicon-syntax interface (cf. Marantz 1997, Chomsky 2000b). Thus, what you have at the clausal level (syntax as a mediating representation between sound and meaning) is also what you have at the word level: (9) meaning
phonology
syntactic-formal feature
51
52
Norbert Corver
Surface imperfections in the L2 derivational output can now be due to miscategorisation (i.e. ‘mis-’ from the perspective of the target-language): an incorrect categorial feature is associated with some sound-meaning pair. In (10)–(13), some examples of miscategorisation by L2-learners are given: (10) Hier komt weg. Ik beetje momentes. here comes away I a-bit moment-infl ‘Here he goes away. I wait a bit.’ (11) Ik komt huis en Slenol wegt huis Stokhasselt. I come home and Slenol away-3sg house Stokhasselt ‘I go home and Slenol went to his house in Stokhasselt.’ (12) A: In de buit ligt die. in the outside lie those ‘Outside lay these.’ I: Hm? ·lack of understandingÒ A: In de buiten. Buiten ook heeft de steen. in the outside outside also has the stone ‘There are also stones (at the) outside.’ (13) a.
Ja verzeker betalen he. yes insure pay discourse-prt ‘Yes, my insurance will pay.’ b. Uh ik ongeluk beur maar ik heb nu geen verzeker. uh I accident happen but I have now no insure ‘I had an accident but I have no insurance now.’
In (10) and (11), moment and weg are treated as roots (i.e. sound-meaning pairs) that receive a verbal character after merger of a verbal categorial feature. Schematically (order irrelevant): v
(14)
v
moment/weg
After attachment of the verbal categorial feature to the root, the lexical item displays verbal behavior: it carries, for example, the verbal inflection -t (present tense, third person singular). From the perspective of the (grammar of the) L2-learner, there is nothing odd about these lexical items: they each represent a root (a sound-meaning pair), which carries a categorial feature that turns it into a verbal form that is legible at the lexicon-syntax interface, in the sense that
Perfect projections
the lexical item (carrying a categorial feature) is accessible to the computational (i.e. morphosyntactic) rules. In short, these lexical items represent perfect objects in the L2-learner’s grammar. Another interesting example of a target-imperfect but (lexicon–syntax) interface-perfect object is the verbal form bint in the following examples: (15) a.
Ja komt politieauto en hij bint in auto toe. yes comes police-car and he inside-3sg in car prt ‘Yes, there comes a police car and he goes into (enters) the car.’ b. I: Hij wat? he what? A: Hij bint auto. he inside-3sg car ‘He goes into the car.’
The use of bint in these examples seems to be a combination of L1-transfer (i.e. conservation) and creative use of the L2. Turkish has a verb binmek, which means ‘to get in’, and Dutch has a preposition/particle binnen, which means ‘inside’.6 The lexical items buit(en) (cf. (12)) and verzeker (cf. (13)) are also perfect objects in the L2-learner’s grammar. As can be concluded from their co-occurrence with determiner-like elements (de, geen), these items are nominal. This nominal behaviour is represented by the nominal categorial feature that is attached to the root of the lexical item i.e. the ‘bare’ sound-meaning pair. Schematically: n
(16)
n
buit(en)/verzeker
The verbal analysis of items like momentes, wegt and bint and the nominal analysis of buit(en) and verzeker are imperfect from the perspective (of the grammar) of the mother tongue speaker of Dutch. For him, binnen (meaning: ‘inside’) and buiten (meaning: ‘outside’) are both prepositions; moment (meaning: ‘moment’) is a noun and verzeker (meaning: ‘to insure’) is a verbal form. In short, the ‘wrong’ categorial feature is associated with the root in these L2-expressions. From the perspective of the L2-learner’s interlanguage grammar, however, these non-target-words are perfect projections: assignment of a categorial value makes these items legible at the lexicon-syntax interface and accessible to the computational rules that take these objects as their input. The noun buit, for instance, is merged with the determiner de.
53
54
Norbert Corver
Summarizing, ‘miscategorisation’ (i.e, ‘mis’ from the perspective of the target language) yields what could be called ‘target imperfect’ lexical items. Because of this miscategorisation, these L2-lexical items are often hard to understand for mother tongue speakers of that language. From the interface perspective, however, there is no reason to believe that these lexical items are illegible for the system (read: (morpho)syntax) which interacts with the lexicon. After assignment of a categorial feature, it is input to the combinatorial rules of the grammar.
3. Perfect quantificational expressions Thus far, we have seen that mis-categorisations on the part of the L2-learner lead to the formation of L2-expressions that are perfect from an interface perspective but may be deemed ‘imperfect’ when compared with the target language. These target imperfections do not only appear in the domain of content words (also called: lexical categories), but, not unexpectedly, also in the domain of function words (also called: functional categories). An interesting illustration of interface legibility of target imperfections comes from the domain of quantified noun phrases. As shown by the following examples, quantifying expressions may display different forms depending on their function and position within the syntactic structure. Consider, for example, the following variants of the universal quantifier al in present-day Dutch:7 (17) a.
Jan heeft alle mensen herkend. Jan has all people recognized b. Jan heeft alles herkend. Jan has everything recognized c. Jan heeft allen herkend. Jan has all recognized ‘Jan has recognized all of them/everyone.’
In (17a), we have the quantifying determiner alle, which is often treated as a fusion of the pre-determiner al (cf. note 7) and the definite article de: al+de Æ alle (cf. Paardekooper 1974, Verkuyl 1981; but see Zwarts 1992). In (17b), al is followed by the sequence (e)s, which presumably used to be a genitive case suffix in older variants of Dutch, but is no longer recognized as such anymore; i.e. alles seems to have developed into a non-composite form which carries a neuter meaning: ‘everything’. As shown by (17b), alles, as opposed to the
Perfect projections
quantificational form alle, occupies an argument position in the clause. The quantificational element allen, finally, represents a plural form. It always refers to human beings, and as such arguably carries the formal property [+human]. Just like alles, the lexical item allen in (17c) occupies an argument position within the clause.8 Importantly, all these quantified expressions obey the universal constraint that there is a restriction on the quantificational element (say al). This restricted reading on the quantifier can be represented as follows: (18) a. for all xi [xi: people] b. for all xi [xi: things] c. for all xi [xi: people]
(cf. (17a)) (cf. (17b)) (cf. (17c))
By allowing this restricted quantificational reading (i.e. the set of individuals (objects, persons) is specified over which the quantifier ranges) and binding a variable at LF, the nominal expressions in (17) are interpretable at the LF-interface. Quantification, being a core property of natural language, is also found in the L2-derivational output. Also with quantificational expressions we find patterns which are imperfect from the perspective of the target language but perfect from an interface perspective on natural language expressions. Consider, for example, the forms in (19) and (20), which are all produced by one and the same Turkish learner of Dutch (viz. Abdullah): (19) a.
Wat doet alles ik weet niet. what does all I know not ‘I don’t know what everyone does.’ b. Altijd uh alles woonte Klirsehir van Turkije. always uh all lived Klirsehir of Turkey ‘Everyone still lives in Klirsehir in Turkey.’ c. En dan vandaag hier komen alles. and then today here come all ‘Today, everyone comes here.’
(20) a.
Alles mensen # toerist ja. all people (were) tourists yes b. Ken je in Nederlands alles stad? know you in Netherlands every city c. Alles ja uh kinderen niet Turks spreken. all yes uh children not Turkish speak ‘All children don’t speak any Turkish.’
55
56
Norbert Corver
d. Wij maakte alles maar ik weet niet alles naam # Nederlands. we make everything but I know not everything name # Dutch ‘We make everything but I don’t know the name of everything in Dutch.’
In (19) and (20), the form alles is used instead of the target form allen and alle, respectively. The L2-learner has identified the universal quantificational meaning of alles, but he has not discovered yet that alles has a non-human interpretation and that it cannot occur as a quantificational determiner. Or to put it differently, the L2-learner has identified the quantifier feature associated with alles, but other lexical features, like its categorical feature (e.g. alle being a quantificational determiner and alles being a noun-like expression), do not seem to have been identified yet. Importantly, the quantificational expressions in (19) and (20) are LF-interpretable: they receive a restrictive reading (i.e. a set of objects is defined over which they range) and bind a variable at LF. Thus, an L2-expression like (19c) has the following LF-structure: (21) For all xi, [xi: people], xi come here today.
In short, the pattern alles (N) represents an LF-interpretable structure from an interface perspective. LF-legibility also holds for the lexical item alleen, as it is produced in the following examples by the same Turkish learner of Dutch: (22) a.
Ja nu alleen mag. yes now ‘alleen’ may ‘Yes now, everything is permitted.’ b. Ik wil alleen leren. I want ‘alleen’ learn ‘I want to learn everything.’
(23) a.
Maar niet alleen. but not ‘alleen’ ‘But not everyone (is unpleasant).’ b. Maar ik zeg niet alleen slechte mensen. but I say not ‘alleen’ bad people ‘But I don’t say that everyone is bad’
The item alleen in (22) and (23) receives a clearly (universal) quantificational interpretation. In (22), it receives the interpretation ‘everything’; in (23), it is interpreted as ‘everyone’. Also in these examples, then, the L2-learner has identified the universal quantificational element al.
Perfect projections
Interestingly, this quantificational reading is not the one which is associated with the lexical item alleen in the target language. In Dutch, alleen means ‘alone’, a reading which arguably derives from the two elements that compose this expression: the quantifier al and the numeral een: alleen actually means: ‘one is all’.9 As regards its distributional behaviour, alleen occurs as a floating element which enters into a predicative relationship with a noun phrase in the sentence. In (24), for example, alleen is predicated over the subject ik. (24) Ik ben toen alleen naar de bakker gegaan. I am then alone to the baker went ‘Then I went to the bakery shop alone.’
Contrary to the L2-expression alleen in (22) and (23), the target language item alleen can never occur in an argument position: (25) *Ik kende alleen. I knew alone
Although the L2-lexical item alleen, which has a universal quantificational meaning, does not occur in the target language, we should not conclude from this that it is an illegitimate object. At the LF-interface, this L2 quantificational expression receives a restrictive reading (‘for all xi [xi: persons/things]’) and binds a variable at LF. (26) For all xi, [xi: thing], I want to learn xi
As a matter of fact, the L2 quantificational expression alleen is just as perfect from an interface perspective as a target quantificational expression like iedereen (‘everyone’) and menigeen (‘many a one’). These forms are sometimes analysed as composite quantificational expressions, that consist of a quantifying element (ieder, menig) and an indefinite pronominal part (een), which, in the target language, refers to humans (just like English one in One shouldn’t do that). The L2 learner who produces the forms in (22) and (23) has possibly identified the composite character of the quantificational expression alleen: it consists of the universal quantifier al and the indefinite pronominal element een. As opposed to a target expressions like iedereen, however, the L2 expression alleen is not restricted to quantification over humans. This suggests that the L2-learner has not discovered yet that the indefinite pronominal een is restricted to a human interpretation. To summarize: forms like alles and alleen, as produced in (19)–(20) and (22)–(23) are legitimate expressions at the LF-interface. Even though their
57
58
Norbert Corver
distribution and interpretation (e.g. nonhuman versus human) may differ from that of the target lexical items alles and alleen, the conclusion seems inescapable that these quantificational expressions are fully legible at the LF-interface: they receive a restrictive reading and bind a variable at LF.
4. The interpretability of apparently P-less structures Target imperfection is also found with L2-structures expressing ‘prepositional’ features like ‘location’ and ‘path’. Consider, for example, the expressions in (27) and (28), that are produced by a Turkish learner of Dutch (see Schenning (1998) for extensive discussion). (27) a.
Hij ook woon Kirslehir. he also lives Kirslehir ‘He also lives in Kirslehir.’ b. Hij werkt Ankara. he works Ankara ‘He works in Ankara.’ c. Ik nooit geweest Istanbul. I never been Istanbul ‘I have never been in Istanbul.’
(28) a.
Ik gaan school. I go school ‘I go to school.’ b. Kom maar mijn huis. come just my house ‘Come to my house.’ c. Ja wij moet altijd moskee gaan een dag vijf keer moskee gaan. yes we must always mosque go a day five time mosque go ‘Yes, we must go to the mosque five times every day.’
As shown by the following target Dutch equivalents of (27a) and (28a), respectively, Dutch requires the presence of a prepositional element in those syntactic constructs that express the abstract property of location. (29) Hij woont ook in Kirslehir. he lives also in Kirslehir (30) Ik ga naar school. I go to school
Perfect projections
In (27), the elements Kirslehir, Ankara and Istanbul indicate a static location, i.e. ‘place’. In (28), the elements school, huis and moskee have a ‘path’ interpretation. It is obvious that abstract meaning properties like ‘place’ or ‘path’ are not directly related to these nouns. These nominal elements don’t have an inherent locative meaning, as is clear from sentences in which they fulfil a non-locative role, as in (31): (31) a.
Ik ken Ankara goed. I know Ankara well b. Ik zie mijn huis. I see my house
It is more likely that the locative properties ‘place’ (static location) and ‘path’ (dynamic location) are associated with the category P(reposition). In the (target) Dutch examples in (29)–(30), this prepositional element expressing the abstract meaning property ‘space’ or ‘path’ is phonetically realized; in the L2-expressions in (27)–(28) it is not. The ‘non-visibility’ (i.e. phonetic absence) of the prepositional element does not imply that the category P (and its projection) is absent in expressions like Kirslehir and school. In fact, the preposition may very well be empty: (32)
PP P·spaceÒ Ø
(33)
(cf. (27a)) NP Kirslehir
PP P·pathÒ Ø
(cf. (28a)) NP school
In Emonds (1985, 2000), a principle is proposed that permits a closed class category to be empty under the condition that it is realized on its phrasal sister. According to this principle, which Emonds calls the Invisible Category Principle (ICP), ‘empty P’ structures can be utilized in a language if features like path or space are realized on the NP-sister by means of a case marking. Thus, although the interpretable P-feature (i.e. path/space) itself is associated with the P-head of the prepositional structure, it can be alternatively realized on the NP-sister. These alternative spell-outs of the locative/path feature are pure spell-outs of features and appear late in the derivation (i.e. spell out at PF).
59
60
Norbert Corver
As noted in Kornfilt (1996), locative meanings are often expressed by means of case suffixes in a language like Turkish: (34) a.
Kitap masa-da. book table-loc ‘The book is on the table.’ b. Hasan Ankara-ya git-ti. Hasan Ankara-dat go-past ‘Hasan went to Ankara.’ c. Hasan Ankara-dan gel-di. Hasan Ankara-abl come-past ‘Hasan came from Ankara.’
Following Emonds’ ICP, these structures could be interpreted as empty P-structures, which have the prepositional feature (space/path) realized on the NP-sister as a case-suffix. For example (order of P and NP-sister irrelevant): PP
(35)
P·spaceÒ Ø
NP masa-DA (-DA as alternative realization of P-feature)
Under the assumption that L2-learners take a conservative approach towards the expression of location and path denoting expressions (i.e. PPs), it is expected that they initially do not realize the prepositional head. Just like in Turkish, they try to realize the prepositional feature by means of a case marking on the NP-sister of P, a strategy which is not available in Dutch.10 The following L2-variants of prepositional structures are interesting in this context: (36) a.
I: Waar woont hij? I: where lives he O: van Tilburg. O: of (= in) Tilburg b. I: Wanneer ziet ze die jongen dan? I: when sees she that boy then O: van Trabzon. O: of (= in) Trabzon c. En dan beetje wandelen van ·name of streetÒ. and then bit walk of ·name of streetÒ ‘And then I walk a bit in ·name of streetÒ.’
Perfect projections
In these examples, the prepositional element van appears in a prepositional structure which denotes a location. Van itself does not seem to carry any locative ‘meaning’. As a matter of fact, this meaningless preposition van also shows up in non-locative prepositional contexts (cf. Schenning 1998): (37) a.
Ik niet trouwen van Yvette. I not marry of Yvette ‘I don’t marry Yvette.’ b. Ik zegt: waarom jij niet praten van mij? I say why you not speak of (= with) me c. En dan moet ik vertellen van hun. and then must I tell of them ‘And then I must tell it to them.’
This distribution of the meaning-less element van is suggestive for an analysis in which it is a case-suffix which alternatively realizes the locative prepositional features and other types of prepositional features. Schematically, the van-variants would then have the structures in (38): PP
(38)
P·spaceÒ Ø
NP VAN-Tilburg
(VAN as a case-aYx)
In conclusion, both the prepositional structures in (32)–(33) and the prepositional structure in (38) are LF-interpretable objects. The empty preposition (and its projection) carries the interpretable property ‘space’ or ‘path’. The target imperfection is simply a surface phenomenon that relates to the phonetic spell out of the prepositional position: the L2-learner initially leaves the prepositional head empty and tries to realize the prepositional feature by means of a case-marking (i.e. alternative feature realization). Given the lack of clear case markings in Dutch, this alternative realization remains empty initially. At a certain stage in the acquisition process, the prepositional feature gets alternatively realized on the NP-complement by the semantically empty ‘preposition’ van.
5. Agreement and asymmetric spell out In traditional grammars, one often finds the observation that agreement is an asymmetric relation. A verb, for example, is said to agree with its subject-DP in
61
62
Norbert Corver
person and number; it is not the subject-DP that is dependent on the verb for agreement. And an adjective is said to agree in number and gender with the noun it modifies; it is not the noun that is valued for certain phi-features under agreement with the modifying adjective. In recent generative studies (cf. e.g. Chomsky 1995), this asymmetry of the agreement relationship is captured in terms of the notion ‘interpretable (formal) feature’. An interpretable (formal) feature is a feature that has a semantic contribution at the LF-interface (i.e. it is interpretable at LF). A non-interpretable (formal) feature has no interpretation at LF (or for that matter: PF). Structural case for nouns and phi-features for categories that agree with nouns are core examples of uninterpretable formal properties. The asymmetry in the agreement relationship, as observed in traditional grammatical studies, has been reinterpreted in terms of the notions [+interpretable] versus [−interpretable]. It is the element carrying the [+interpretable] feature that agrees with the element carrying the [−interpretable] feature. This agreement relationship involves feature matching and elimination of the uninterpretable feature that is associated with the matching constituent. In (39a), for example, it is the verb zagen that enters into an agreement relationship with the plural subject-noun phrase de mannen. Plurality (‘more than one’) and singularity (‘one’) is an interpretable property of nouns. In a way, the plurality feature on the verb zagen is redundant; it does not contribute any semantics and, as such, can be characterized as uninterpretable. The plurality marking on the verb is just a formal marker of the agreement relationship; semantically, it does not contribute anything to the linguistic expression. Therefore, in a certain intuitive sense, it is easier to imagine that a plural interpretation gets associated with the ill-formed expression (39b) than with the ill-formed expression in (39c). In (39b), plurality is morphologically specified on the noun, i.e. the category that carries the number feature as a [+interpretable] property, and the verb is not morphologically specified for plurality. Even though there is a morphological mis-match, there is a tendency to assign this sentence a plural interpretation: that is, the interpretation ‘the men-pl saw-pl me’ is much more likely than the interpretation ‘the man-sg saw-sg me’. In other words, it is the plurality marking on the noun that most strongly determines the semantic interpretation of this expression which does not satisfy agreement at the level of morphological expression. Consider next (39c), where we have the reverse situation: the subject noun phrase does not bear any marking of plurality; it is the verb zagen which is plural morphologically. In spite of the plural marking on the verb, it is intuitively more difficult to get a plural reading of the noun. As a matter of fact, it is the singularity of the noun
Perfect projections
which seems to be dominant again for the interpretation. In short, given the fact that in subject-verb agreement relations, it is the noun that determines plural or singular interpretation, it is expected that morphological marking of singularity or plurality is more likely to be realized on the noun than on the verb. (39) a.
De mannen zagen mij. the men-pl saw-pl me b. *De mannen zag mij. the men-pl saw-sg me c. *De man zagen mij. the man-sg saw-pl me
This asymmetry in the morphological realization of the number feature is reflected in the L2 derivational output: patterns are typically found in which the agreement properties of the noun are correctly realized overtly, but not those of the verb. In other words, morphological spell-out of the agreement property generally applies to the element carrying the [+interpretable] property and not to the element carrying the [−interpretable] property. Consider, for example, the following L2-utterances, which are produced by a Turkish learner of Dutch. (40) a.
En dan andere jongens komt. and then other boys-pl come-sg b. Maar twee meisen drie jongens weet ut wel maar heel but two girls-pl three boys-pl know-sg it prt but very klein beetje. little bit ‘But two girls and three boys knew it, but only a little bit.’ c. Twee jongens woont Oisterwijk. two boys-pl live-sg Oisterwijk ‘Two boys live in Oisterwijk.’ d. Die jongens komt Turkije. those boys-pl come-sg Turkey ‘Those boys come to Turkey.’
In these examples, the phi-feature number carries the value ‘plural’. This number feature is correctly spelled out on the noun of the subject-noun phrase that enters into an agreement relation with the verb. The verb carries what looks like a singular inflection. Arguably, this verbal form is unanalysed morphologically or, alternatively, the marking -t is underspecified for number. What is important is that, even though the L2-expressions in (40) may be regarded as imperfect from the perspective of the target language, they are
63
64
Norbert Corver
perfect from the perspective of the LF-interface: i.e. they are expressions that are fully legible semantically. Let me close off this section with another illustration of an LF-interpretable structure in which the agreement properties are morphologically spelled out asymmetrically. The relevant example comes from noun-phrase internal agreement between a numeral and a noun. As noted among others in Emonds (1985), plurality is a property of numerals: i.e. numbers above one are inherently marked for the property [+plural]. This formal property is interpretable, since plurality plays a role in the semantic interpretation of a linguistic expression. As illustrated by the following L2-expressions, plurality is not always marked on the ‘agreeing’ noun: (41) a.
Vijf minuut he uh en dan klaar # alles dit. five minute prt prt and then ready # all this ‘All this is ready in five minutes.’ b. I: Hoe lang zit je al weer op school # twee weken? I: how long sit you already at school # two weeks A: Vijf daag. A: five day c. I: Heb je nu vakantie van school? I: have you now holiday from school A: Ja twee week. A: Yes two week d. Vader, moeder en drie broer. father mother and three brother
In (41a), the noun minuut does not carry plural morphology. The target pattern would be: twee minuten. Also in this case, then, the L2-learner chooses the strategy of not morphologically expressing plurality if this semantic feature is already specified in the projected structure. In other words, redundant marking of the plurality feature on the noun is avoided. Again, it is important to stress that the numeral+noun-patterns in (41) are fully legible at the LF-interface. From a target language perspective, however, these structures are imperfect: in Dutch, plurality is (redundantly) marked on the noun, when it combines with a numeral that is inherently specified for plurality (i.e. ‘more than one’). In this respect, Dutch differs, by the way, from Turkish. As noted in Kornfilt (1996: 225), there are syntactic contexts in Turkish where, despite plural semantics of the noun phrase, the head noun cannot be marked for plurality. When the noun is preceded by a numeral or certain quantifiers, the plural suffix cannot occur:
Perfect projections
(42) a.
bes¸ çocuk(*-lar) five child(*ren) ‘five children’ b. birçok çocuk(*-lar) many child(*ren) ‘many children’
In view of the non-redundant marking in (42), it is likely that the L2-learner who has produced the numeral+noun-patterns in (41) has adopted a conservative strategy: the Turkish rule of not morphologically marking plurality on the noun when it is preceded by a numeral (or certain quantifiers) is also at the basis of the numeral+noun sequences in his second language, i.e. Dutch (see again Van de Craats (this volume) for further discussion of the notion of conservation).
6. Conclusion From the perspective of the target language, L2-expressions often seem highly imperfect. At the surface, these L2-expressions (e.g. Dutch L2-products of Turkish learners) seem to differ greatly from those produced by mother tongue speakers. From a different perspective, though, there does not seem to be much wrong with those L2-expressions: they are perfect expressions, in the sense that they meet conditions imposed by other cognitive systems that the language faculty interacts with (external requirements). That is, any L2 (interlanguage) grammar provides (grammatical) information that is ‘legible’ to the cognitive systems with which it interacts. I have tried to illustrate the interface-legibility of L2-expressions by means of four types of phenomena: (a) categorial labeling of words, (b) quantificational expressions, (c) the expression of location in prepositional structures, and (d) the morphological expression of certain agreement patterns. As regards the categorial labeling of words, it was noted that certain L2-words (e.g. an L2-verb like wegt (‘goes’)) that do not exist in the target-language, are perfect lexical constructs from a (lexicon-syntax) interface perspective: it is a sound-meaning pair that through assignment of a categorial value (i.e. V) becomes accessible to the computational system (merge, move, morphological rules) of the interlanguage grammar. I further argued that certain L2 quantificational expressions (e.g. alles mensen) that are imperfect from a target language perspective are fully legitimate from the perspective of LF-legibility: they are ‘normal’ quantificational expressions in the sense of
65
66
Norbert Corver
allowing a restricted quantificational reading and binding a variable at LF. I further argued that LF-legibility also holds for what, at the surface, looks like a bare nominal carrying a locative meaning (e.g. Ankara, meaning ‘in/to Ankara’). At a more abstract level, these L2-expressions are prepositional structures, which have the locative meaning associated with an empty P. These ‘empty-P’-structures are fully interpretable expressions at the LF-interface. Finally, it was observed that in L2-agreement patterns (of beginning learners) it is typically the [+interpretable] element that gets morphologically marked. Absence (or underspecification) of morphological marking of the [−interpretable] feature may yield a pattern which is imperfect from the perspective of the target language. From the interface perspective, however, the non-redundantly marked pattern is perfect: an agreement property like number is typically spelled out on those items for which singularity or plurality is an inherent semantic property. In this approach, I have only ‘glanced over’ a variety of L2-expressions from the interface perspective; a perspective which is characteristic of the minimalist thesis. The major purpose was to show that by taking this perspective, the conclusion seems inescapable that L2-expressions are perfect grammatical objects, where perfection amounts to legibility of its information to the systems with which it interacts.
Notes * I thank the participants at the workshop for their comments on the talk. I would also like to thank an anonymous reviewer for helpful comments and suggestions. 1. In a recent interview with Adriana Belletti and Luigi Rizzi, Chomsky states the following (cf. Chomsky, Belletti and Rizzi 1999 (rev. 2000:17)): “Every language meets minimalist standards. Now, that means that not only the language faculty, but every state that it can attain yields an infinite number of interpretable expressions. That essentially amounts to saying that there are no dead ends in language acquisition.” He further states: “The minimalist thesis would say that all states have to satisfy the condition of infinite legibility at the interface.” 2. A reviewer raises the following question: What can a generative-minimalist theory account for with respect to interlanguage expressions that other theories (e.g. a GB-based approach using presumed UG-notions like government) cannot account for? Although such a comparison of approaches may be useful in certain respects, it is not always easy to evaluate the benefits of one specific analysis of interlanguage data over another one, especially if the analytic tools are different. Important, though, is the different perspective that the minimalist approach towards language design provides: linguistic properties are not so much considered from an intra-grammatical perspective, but rather from an interface perspective (lexicon-
Perfect projections
syntax, syntax-semantics, syntax-phonology). This raises different sorts of questions about the linguistic objects one examines (e.g. What makes an (L2) representation (il)legible at the interface?) and arguably provides different sorts of accounts of the grammatical properties displayed by these representations. 3. The data are drawn from the European Science Foundation (ESF) Program in Second Language Acquisition by Adult Immigrants (for design, elicitation techniques, and topics, see Perdue 1993). This project was set up as a longitudinal and cross-linguistic multiple case study. Most of the data discussed in this article are from the Turkish informants Abdullah and Osman. Since the issue of language development is not central in this paper, I have left out information about the stages in which the expressions were uttered by these informants. 4. In each of the L2-expressions (1a)–(4a), there is more than one ‘target imperfection’. For the sake of discussion, I will pick out one type of imperfection for each of the examples. 5. I won’t consider in this paper the issue of (L2) perfection at the PF-interface. 6. I would like to thank the reviewer for discussion of this example. 7. Another pattern featuring the quantificational element al is: al de boeken (all the books). In this pattern, the quantificational element occurs in a pre-determiner position. 8. The quantificational form allen also shows up as a floating element, like in: Zij zijn gisteren allen gekomen (they are yesterday all come; ‘They all came yesterday.’). 9. Alleen can also mean ‘only’ in present-day Dutch. This (homophonous) adverbial element is not quantificational and displays a cross-categorial distribution, just like its English equivalent. 10. See Van de Craats (2000) and Van de Craats, Corver and Van Hout (2000) for a discussion of conservation of L1-grammatical features in L2-expressions. See also Van de Craats’ contribution in this volume.
References Bley-Vroman, R. 1990. The logical problem of foreign language learning. Linguistic Analysis 20: 3–49. Chomsky, N. 1991. “Some notes on economy of derivation and representation”. In Principles and parameters in comparative grammar, R. Freidin (ed.), 417–454. Cambridge, MA: MIT Press. Chomsky, N. 1995. The minimalist program. Cambridge MA: MIT Press. Chomsky, N. 2000a. New horizons in the study of language and mind. Cambridge, UK: Cambridge University Press. Chomsky, N. 2000b. “Minimalist inquiries (MI)”. In Step by step: Essays in minimalist syntax in honor of Howard Lasnik, Martin et al. (eds). Cambridge MA: MIT Press. Chomsky, N., Belletti, A. and Rizzi, L. 1999 (rev. 2000). An interview on minimalism. University of Siena. Craats, I. van de 2000. Conservation in the acquisition of possessive constructions. Doctoral Dissertation, Tilburg University.
67
68
Norbert Corver
Craats, I. van de, Corver, N. and Hout, R. van 2000. “Conservation of grammatical knowledge: on the acquisition of possessive noun phrases by Turkish and Moroccan learners of Dutch”. Linguistics 38 (2): 221–314. Emonds, J. 1985. A unified theory of syntactic categories. Dordrecht: Foris. Emonds, J. 2000. Lexicon and grammar: The English syntacticon. Berlin/New York: Mouton de Gruyter. Kornfilt, J. 1996. Turkish. New York: Routledge. Marantz, A. 1997. “No escape from syntax: Don’t try morphological analysis in the privacy of your own lexicon”. University of Pennsylvania Working Papers in Linguistics 4 (2): 201–225. Paardekooper, P. C. 1974. Beknopte ABN-syntaxis. Den Bosch: Malmberg. Perdue, C. (ed.) 1993. Adult language acquisition: Cross-linguistic perspectives, vol. I: Field methods. Cambridge, UK: Cambridge University Press. Schachter, J. 1989. Testing a proposed universal. In Linguistic perspectives on second language acquisition, S. Gass and J. Schachter (eds), 73–88. Cambridge: Cambridge University Press. Schenning, S. 1998. Learning to talk about space. The acquisition of Dutch as a second language by Moroccan and Turkish adults. Doctoral Dissertation, Tilburg University. Verkuyl, H. 1981. “Numerals and quantifiers in X-Bar syntax and their semantic interpretation”. In Formal methods in the study of language, J. Groenendijk, T. Janssen and M. Stokhof (eds), 567–599, Amsterdam: Mathematic Centre. White, L. 1988. “Island effects in second language acquisition”. In Linguistic theory in second language acquisition, S. Flynn and W. O’Neill (eds): 144–172. Dordrecht: Reidel. White, L. 2000. “Second language acquisition: From initial state to final state”. In Second language acquisition and linguistic theory, J. Archibald (ed.): 130–155. Oxford: Blackwell. Zwarts, J. 1992. X¢-Syntax — X¢-Semantics. On the interpretation of functional and lexical heads. Doctoral Dissertation. Research Institute for Language and Speech — OTS. Utrecht University.
Chapter 4
L1 features in the L2 output Ineke van de Craats University of Nijmegen
1.
Introduction
In current generative syntax (e.g. Chomsky 1995), the role of the lexicon has become more prominent in the generation of syntactic expressions. Under this view, the lexicon is not limited to vocabulary, but also contains important grammatical information. Lexical knowledge consists of grammatical properties as defined by language-particular knowledge of functional categories (parameter settings), language-specific knowledge of lexical items and their features (the vocabulary) and morphological knowledge. By means of lexical items selected from the lexicon, the computational system of human language can build phrases and sentences. A syntactic object — a clause or a phrase — is considered to be the structural projection of a series of linguistic properties associated with a lexical item. Those formal features (e.g. singular, accusative, human, +V, −N) are stored in the vocabulary, together with the semantics of that specific lexical item. Lexical items, like nouns and verbs, are base-generated as a unit, including case morphology and inflectional morphology like person, number, tense, under lexical heads. Functional heads, on the other hand, do not dominate inflectional morphology, they dominate bundles of abstract features. These features have to be eliminated or erased in the course of the derivation, which is done by feature checking. This feature checking is a matching of the features (e.g. case morphology is checked by its case assigner) and is done by adjoining the inflected N or V to the relevant functional head. So, morphology which is associated with a verb or a noun has to be checked by the abstract features dominated by a functional head (e.g. Agr or T for verbs, and Agr and D for nouns). What features are dominated by a functional head and whether these features are strong or weak is lexical knowledge which is necessary for the generation of a syntactic object, but is not part of the vocabulary. So, in the minimalist approach, a parameter is related to a feature of a functional head
70
Ineke van de Craats
that attracts an identical feature of a lexical item at some point in the derivation, and so, is essentially linked to the lexicon. If, in recent linguistic theorizing, the formal features of a lexical item and the specification of functional heads, can be seen as the seeds for building a syntactic structure they must play the same role in the acquisition of a new language, assuming that we are dealing with a natural language. Applying generative theory to second language acquisition is not new of course. Before the work of White (1982, 1985), Flynn (1986) and many others, Adjémian (1976) was the first to adopt a Chomskyan approach to interlanguage development. He considered grammatical interlanguage systems to be natural languages but different from L1 grammars only in their ‘permeability’ to aspects of the L1 system. In this chapter, the current view of generative syntax will be applied to the analysis of naturalistic second language data. Although L2 expressions are the syntactic products of the interaction between the computational system and a changing lexicon, the focus will be on how the grammatical knowledge of an L2 learner is encoded in the lexicon, not on derivation and syntactic representations. The question is more: what is the nature of this grammatical knowledge at the L2-initial state and how does it change? Through examples produced by L2 learners and by outlining the longitudinal development of some lexical items, it will be shown how features of a lexical item may change in the course of the acquisition process, giving rise to new syntactic structures. For a detailed discussion of the syntactic representations of nominal possessive constructions in which van (Subsection 5.1) and the realisation of the personal pronoun (Subsection 5.2) are involved, the reader is referred to Van de Craats, Corver and Van Hout (2000), for representations of clausal possessive constructions in which heeft is involved (Section 6) to Van de Craats, Corver and Van Hout (2002).
2. The L2-initial state, data and informants The central claim is that, initially, the grammatical system of an L2 learner is not only ‘permeable’ to the learner’s L1 system (Adjémian 1976) but even based on the L1 system. It is assumed that L2 learners exhibit conservative behaviour and take the fully fledged grammar of their L1 as the starting point of the L2 acquisition process (in case they have command of only their L1). This amounts to both the Full Transfer/Full Access Hypothesis (Schwartz and Sprouse 1996) and the Conservation Hypothesis (Van de Craats 2000, Van de Craats, Corver and Van Hout 2000). The latter, however, explicitly states that the learners’
L1 features in the L2 output
output cannot show all L1 properties because of a strongly limited L2 vocabulary. With the developing vocabulary, (more) L1 properties related to free and bound functional morphemes become manifest gradually, as we will see for genitive markers (Subsection 5.1) and copular forms (Subsection 6), which, initially, are not found in learners’ data but appear gradually. The following aspects of lexical knowledge may be conserved at the L2-initial state, when all learning starts: – – –
parameter settings (e.g. strength values); knowledge of morphology and morphological realization rules (e.g. realization of case); knowledge of lexical items: formal features (e.g. categorial features) and semantic-conceptual values (the meaning).
Because of acquisition, restructuring will apply at all levels of lexical knowledge: from parameter values to semantic-conceptual values. Initially, L2 learners rely on the old system of the L1. On the basis of primary linguistic input, they will depart from their conserved L1 parameter setting and L1 ‘vocabulary’ lexical knowledge. This implies that a parameter together with its possible values will remain available through UG. In the next sections, the changing grammatical knowledge at the basis of L2 expressions will be shown through the spontaneous production data of eight adults (18–24 years old) learning Dutch as a second language. These data were collected within the framework of the European Science Foundation (ESF) Program on Second Language Acquisition by Adult Immigrants (see Perdue 1993). The ESF project was set up as a longitudinal and cross-linguistic multiple case study; we only use the Dutch data here. The eight informants were followed for two and a half years. The period of investigation was divided into three cycles of nine sessions, one session a month. In the examples in the next sections, we refer to the learner and to the cycles and recording session (e.g. I.7 = first cyle, session 7) in which the utterance was produced. At the time of the first session, the informants had been living in the Netherlands for seven to twelve months. They had a very low level of proficiency in Dutch, were monolingual, and had a limited level of education. Several elicitation tasks were repeated in each cycle, such as interviews, role-playing, and film-retelling tasks. Some other examples used were produced by child L2 learners between six and nine years old. They come from another corpus of longitudinal and cross-linguistic data collected by Vermeer (1986). These informants were also from a Turkish and Moroccan Arabic background: 16 children from both language groups.
71
72
Ineke van de Craats
They were followed over 2.5 years from the time they entered primary school. At the time of the first recording the children’s age ranged from 6;4 to 7;9 years.
3. The nature of a lexical item As hinted at in the previous section, we distinguish lexical knowledge as defined by UG from language-specific knowledge of lexical items and their lexical entries. To avoid confusion, we refer to the latter type as the vocabulary. The first question that arises here is what knowledge learners have exactly of a lexical item. Lexical items are combinations of sound and meaning properties which can be read, or interpreted, by other cognitive systems (cf. Chomsky 1994, 1995). Phonetic features make up the phonetic representation and semantic features make up the semantic representation. A sound-meaning pairing is encoded by a phonological matrix. Each coding of a lexical item also contains a set of formal features: intrinsic features and optional features. The former are unpredictable, idiosyncratic grammatical properties of lexical items (e.g. the categorial feature [+N,−V] and the person feature [3 person]); the latter include grammatical features that are predictable from other properties of the lexical entry (e.g. the features number and (abstract) case, which might be derived from the categorial feature definition [+N,−V]).1 Table 1 gives an example of the concept ‘bicycle’ in three different languages. Only the phonological matrix differs. In what way might we conceive of conservation of vocabulary knowledge? Obviously, conservation does not apply at the level of the phonological matrix.2 One might conceive of early lexical acquisition as a process in which L2 learners try to match a meaning representation associated with some lexical item of their L1 vocabulary with a phonological matrix of the target language. Table 1.Lexical items of the concept ‘bicycle’ compared for three languages
– – –
phonological matrix semantics formal features intrinsic intrinsic intrinsic optional optional
The consequence of this conservation model at the level of the derivational output might be that, in an interlanguage, apparently empty constituents may exist. They are filled by L1 feature bundles of semantic-conceptual and formal features, but lacking a phonological matrix. The phonological representation is simply absent. The task of the L2 learner will be to fill in the empty slot of the phonological matrix. Schematically, the acquisition of a lexical item may be represented as in Table 2. What learners do, in fact, is add a new phonological matrix to the already existing bundle of semantic and formal features. They match, in Table 2 for instance, the L2 phonological matrix /fiets/ with the semantic and formal features belonging to the Turkish phonological matrix /bisiklet/. This combination of an L2 phonological matrix, L1 semantics and an L1 feature bundle essentially is a new lexical item in the learners’ L2 vocabulary. Evidence for this way of learning lexical items is (i) L2 lexical items showing imperfect matching or mismatching and (ii) empty phonological matrices in developmental sequences which are filled up later by elements based on the L1 syntax as will be shown by the developmental sequence of the lexical items van (Subsection 5.1) and heeft (Section 6). In the next section, some examples of mismatches will be presented. They are related to the semantic features, the formal features and the argument structure of lexical (not functional) elements. In Sections 5 and 6, we focus on mismatches of functional elements, both in the nominal and the clausal domain.
4. Mismatches in lexical development As long as there are no differences between the formal feature bundles of two lexical items in the source and target languages, learners can carry out this
73
74
Ineke van de Craats
matching operation without making any errors. For most content words, such a perfect match is possible as far as formal features are involved. We do not expect a different set of formal features in two languages because entities are typically nouns ([+N,−V]), actions are typically verbs ([−N,+V]), and qualities are typically adjectives, each of those with its own categorial values (N, V, A) and the formal features typically related to those categories. As for the semanticconceptual aspects, differences between L1 and L2 are to be expected, however. The first example of imperfect matching involves the domain of semantics. A meaning representation of a lexical item may have different aspects. The Turkish verb içmek, for instance, differs minimally from the Dutch verb drinken (‘to drink’). The formal features are the same, but the basic meaning of the verb /içmek/ is ‘to put something in something else’. This general meaning has several more specific meaning aspects, viz. ‘to drink’ and ‘to smoke’. Consider the L2 expression in (1). (1) Als ik Marlboro drinken. when I Marlboro drink ‘When I smoke a Marlboro.’
Turkish learner: Ergün: III-5
In (1), Ergün maps the L2 phonological matrix /drinken/ on to both L1 semantic aspects and grammatical properties of the verb içmek, which results in a mismatch.3 Mismatches, however, are not restricted to the domain of semantics. They may also involve formal features of lexical elements. Let us consider the examples in (2) and (3). Mahmut, a Turkish learner of Dutch, is retelling a scene from a silent movie in which Charlie Chaplin must pay the bill in a restaurant. Charlie refuses to do so because he wants to go to jail. The same episode was told twice, with an approximately ten months’ interval. In (2a) and (3a) the policeman orders Charlie to pay, in (2b) and (3b) Charlie answers that he has no money. Jij betalen geven. you to pay to give ‘You must pay.’ b. Ik niet betalen. I not pay ‘I have no money.’
(2) a.
(3) a.
Politie zegt: “Jij geld geven”. policeman says “you money give ‘The policeman says: “You must pay”.’
Turkish learner: Mahmut, II-9
Mahmut, III-9
L1 features in the L2 output
b. Ik heb niet geld. I have not money ‘I do not have money.’
The examples in (3) make clear what Mahmut meant to say in (2). We can assume that instead of betalen ‘to pay’ he intended to say geld ‘money’.4 From the perspective of the Conservation Hypothesis, the argument runs as follows. A Turkish learner expects to find the verb at the end of the sentence, as Turkish is basically a language with an SOV sentence structure. Hence, the Turkish learner in (2a) considers geven to be a verb. Since the L1 item ödemek (‘to pay’) has an internal argument, this L2 learner places this argument in the object position that normally precedes the verb, e.g. in (2a). Hence, betalen must be the argument, and a noun. In (2b) betalen might be meant as a verb, but (3b) makes that unlikely, the more so because, in cycle II, Mahmut is not yet able to produce possessive clauses in which hebben (‘to have’) occurs. As can be inferred from the sentences in (2) and (3), the categorial value of betalen in (2) is that of a noun: [+N,−V], and not that of a verb [−N,+V]. Only in (3), the phonological matrix /betalen/ is replaced by /geld/, as represented in Table 3. In this table, the subcategorization frame of ‘to pay’ has been integrated in order to underline the verbal character of the L2 item. Table 3.A learner variant of the lexical item /betalen/ compared to the relevant items in source and target languages; deviances from the target language are in italics L1 item (Turkish) – – –
A comparable mismatch due to misinterpretation of the L2 item terug (‘back’) by the L2 learner is presented in (4). (4) a.
Mijn vrouw ik thuis terug. my wife I home back ‘My wife came to my house.’
Mahmut, I-5
75
76
Ineke van de Craats
b. Ik meisje baby Ø. I girl baby ‘My girlfriend is expecting.’ Vijf maanden baby terug. five months baby back ‘In five months the baby will come.’
Mahmut, I-6
In (4a), Mahmut tells about his wedding. Before the wedding, he used to go to his girlfriend’s house, far away, but after the wedding his wife came to his house. In this early stage of acquisition, Mahmut had two options for expressing the possessive pronoun first person singular: the target variant mijn (‘my’) + possessee and the learner variant ik (‘I’) + possessee. He used them both in this sentence. The particle terug expresses the action of coming. In Dutch, the particle terug is the separable part of the compound verb terugkomen (‘to come back/to return’). In matrix clauses, the finite part of the verb appears in (more precisely: is moved to) the second position in the sentence, while the particle remains at the end. This is probably the cause of the misinterpretation by the Turkish learner, who expects the finite verb in end position, where he finds terug (instead of geliyor ‘comes’). In Dutch, as in English, it is even possible to leave out the past participle gekomen in a perfect tense and to say, as the result of the action terugkomen, ik ben terug (‘I am back’), which may be interpreted by this learner as: I have come. In (4b) the directional element is not so evident. Mahmut took part in a role playing task. He was asked to explain to a housing officer why he needed a house. The informant was given the information that his girlfriend was pregnant. The introducing sentence (ik meisje baby) cannot have another meaning than that his girlfriend is expecting. In Dutch, the internal argument (a baby) of the verb verwachten (‘to be expecting’) must be expressed overtly. Note that we are dealing here with two arguments that are strongly suggestive of the predicate verwachten (‘to expect’), so that it makes sense to assume a predicate with an empty phonological matrix, viz., verwachten. The second sentence of (4b) confused the housing officer. He understood that the baby was already born and that he or she would come back after a stay in Turkey or somewhere else. The cause of this misunderstanding is that Mahmut maps the L2 phonological matrix /terug/onto a L1 feature bundle linked to the verb /geliyor/. In that way, the particle terug can act as a verb and has the same argument structure as the verb ‘to come’.5 This is represented in Table 4. The examples above have shown that beginning L2 learners may have difficulty in discerning the grammatical properties of L2 content words, which
L1 features in the L2 output
Table 4.Lexical items of the concept ‘comes’; deviance from the target language is in italics L1 item (Turkish) – – –
are lexical elements with a relatively high salience in the environmental input. For the understanding of functional elements this must be still harder.
5. Functional elements in the nominal domain As functional elements such as determiners and affixes have little semantic load and are often unstressed, they are less salient to L2 learners than content words. So, L2 learners get less opportunity to perceive them in the L2 input and to comprehend them. This is even more so for the formal features. Therefore, it is not surprising that grammatical properties of L1 functional elements persist for a longer time than those of lexical elements. Consider for this purpose two mismatches of formal features at the level of the categorial value and see how these differences become manifest in the learners’ L2 expressions. The first example relates to the realisation of genitive case. The second example is a more indirect consequence of the possibility of dropping the subject (pro-drop) in Turkish and the impossiblity of doing so in Dutch. 5.1 The genitive case Speakers of Turkish are used to realizing case marking overtly by a rich morphological system of suffixes on the head noun. In Dutch, case marking is generally done covertly, except for pronouns and the marking of genitive case (cf. Corver in this volume). For a good understanding of the learners’ data, it is necessary to go into a bit more detail regarding the nominal possessive constructions in Dutch and in Turkish. A Turkish possessive construction like the one in (5) features agreement in person and number between the possessed
77
78
Ineke van de Craats
(pro)noun and the possessor. This agreement is manifested by the agreement suffix -sı. (5) Ays¸e-nin /o-nun araba-sı. Ays¸e-gen/(s)he-gen car-3sg ‘Ays¸e’s/her car.’
The possessor noun phrase (DP) carries a genitive case feature (-nin or -nun, choice determined by vowel harmony) and the possessed noun a genitive case agreement feature (-sı), as represented in Table 5. This possessive relationship is characterized by agreement, overtly realized both on possessor and possessee. In line with Chomsky (1986, 1995), it is assumed that the genitive case agreement feature of the possessee must check off the genitive case feature associated with the possessor DP. The required structural configuration is AgrP where the possessor DP and possessee N enter in Spec–Head configuration because of the strength properties of the Agr head (cf. Van de Craats et al. 2000 for details). Table 5.Formal feature complex of two lexical items and the functional head Agr in the Turkish possessive construction
/araba-sı/ ‘car’ [+N,−V] [3 person] [singular] [+genitive case assignment]
/Ø/ − [+N,−V,+D] ·strongÒ [+N,−V] ·weakÒ
Unlike Turkish, Dutch has several ways for expressing a possessive relationship. Two of them look, superficially, like Turkish (6a, 6b), the other (analytic) construction (7) does not. (6) a.
Ays¸e ’s auto Ays¸e ’s car ‘Ays¸e’s car’ b. Ays¸e d’r auto Ays¸e her car ‘Ays¸e’s car’
(Saxon genitive)
(Doubling possessive)
L1 features in the L2 output
(7) de auto van Ays¸e the car of Ays¸e ‘Ays¸e’s car’
(Analytic construction)
Although the possessive construction in (6a) is called the Saxon genitive construction, the -s should not be interpreted as an inflectional case suffix because, with a coordinated possessor, the possessive marker -s is phonologically attached to the rightmost noun, as in (8). If it were a normal case suffix realized on the head noun, it would be expressed on both nominal heads of a coordinated construction (cf. Corver 1990). We assume that the Saxon genitive is a clitic in the Agr head.6 (8) [Ays¸e en Jan]’s kritiek op elkaar. ‘Ays¸e and John’s criticism of each other.’
The major characteristic of the doubling possessive construction (6b) is the presence of a possessive clitic which doubles the possessor noun phrase and agrees in phi features with the possessor. It is assumed that the possessive clitic (d’r for feminine, z’n for masculine) heads the functional phrase AgrP (Miller 1991) and also that the doubled possessor DP originates within the lexical projection NP and raises overtly to the specifier position of Agr. In analytic constructions, as in (7), the dummy preposition van (‘of’) can be considered to be the morphological realization of the inherent genitive case (cf. Chomsky 1986). This implies that such genitive case will only be assigned by N to a noun phrase that receives a thematic role from it. In line with Chomsky’s (1995:285) reinterpretation of inherent case, this genitive case is a + interpretable feature of DP that need not, but could be, checked in a Spec–Head configuration. This has the consequence that the genitive bearing DP (i.e., Ays¸e) can remain within its base position since its genitive case feature need not be checked. The differences between the three types of possessive constructions vary in their modes of genitive case licensing and in what element is in the Agr head. Comparison of Tables 5 and 6 shows that the important differences between Turkish and Dutch lie in the mechanism of case licensing and the lexical material projected in Agr head, both abstract features and overt functional elements. But what about the realisation of genitive case? That is another point of difference and more transparent to learners than the properties discussed above. In Table 7, the lexical items associated with the concept ‘possessor of’ or, to put it in other words, the realisation of genitive case, are compared for Turkish, Dutch and two learner variants.
79
80
Ineke van de Craats
Table 6.Formal feature complex of two lexical items and the functional head Agr for three possessive constructions in Dutch Possessor
/auto/ ‘car’ [+N,−V] [3 person] [singular] [+gen. case assignment]
Agr Saxon genitive
Doubling possessive
Analytic construct.
/-s/ – [+N,−V,+D] ·strongÒ
/clitic/ Agr not (e.g. z’n) projected [+N,−V,+D] ·strongÒ
[+N,−V] ·weakÒ
[+N,−V] ·weakÒ
Table 7.Lexical items of the concept ‘possessor of ’; deviances from the target language are in italics L1 item (Turkish)
Learner variant 1
Learner variant 2
L2 item (Dutch)
/-(n)In/ ‘possessor’ [affix gen]
/Ø/ ‘possessor’ [affix gen]
/van/ ‘possessor’ [affix gen]
[N−]
[N−]
/van/ ‘possessor’ preposition [−N,−V] [DP−]
– – –
phonological matrix semantics categorial value
–
subcategorization frame [N−]
L2 learners who map the phonological matrix of the L2, van (‘of’), onto the grammatical properties of their L1, produce such nominal phrases as in (9) and (10). (9) pronominal possessor a. [die van] auto [that of car ‘his car’ b. [onze van] broer [our of brother ‘our brother’ (10) full noun possessor a. [examen van] tolk [exam of interpreter ‘the interpreter at the exam’
Ergün, III-4
child learner (number T25)
Ergün, III-5
L1 features in the L2 output
b. [die jongen van] zijn vader [that boy of his father ‘that boy’s father’ c. [de auto van] de lichten [the car of the lights ‘the car lights’
Osman, III-2
child learner (number T41)
In the examples in (9) and (10), the preposition van (‘of ’) should not be considered an element of an analytic construction, as in (7), but the genitive case marker of the preceding possessor noun phrase as in Turkish (cf. example (5). Notice that both child learners and adult learners produce these constructions based on the L1 bundle of formal features as presented in Table 7. It is a complicating factor that L2 learners do not show this genitive case marker from the earliest stage of L2 acquisition. L2 learners, in general, produce only a few functional elements in the beginning of the acquisition process and it is questionable whether they perceive functional elements in the L2 input at all. Nevertheless, they build syntactic constructions like the L2 expressions in (11), which are not simple two word utterances but can be extensive phrases. tante dochter7 Osman, I-5 aunt daughter ‘my aunt’s daughter (= cousin)’ b. tante zoon auto Mahmut, I-5 aunt son car ‘my aunt’s son’s car’ c. mijn vrouw oma andere man dochter Mahmut, II-8 my wife grandmother other man daughter ‘my wife’s grandmother’s second husband’s daughter’
(11) a.
The learner variants in (11) are almost incomprehensible to native speakers of Dutch, but make sense to Turkish speakers because, under the view of the Conservation Hypothesis, they base those expressions on their L1 grammar, more particularly on the fact that in Turkish, all possessor nouns are overtly marked by a genitive case and that the head noun (the possessee) is marked by a person and number marker that refers to the preceding possessor, as in (5). Under this view, an empty (Ø) phonological matrix associated with an L1 based feature bundle can be assumed for the learner variants in (11). (Compare also Corver’s contribution in this volume about the interpretability of apparently P-less structures.) This analysis is corroborated by the fact that the empty phonological matrix is filled in at a later developmental stage by the case marker
81
82
Ineke van de Craats
van (‘of’), and by the fact that only Turkish learners exhibit possessive constructions overtly marked for genitive case. Moroccan learners do not. One may object that the productions exemplified in (11) proceed directly from UG. But why would Turkish learners have access to UG and Moroccan learners not? In addition to possessive noun phrases in which the possessor precedes van, as in (5) and (6), L2 learners with a Turkish language background produce possessive noun phrases where van precedes the possessor, as in (12) and (13). (12) pronominal possessor a. [van hem] moeder Ergün, II-9 [of him mother ‘his mother’ (target: zijn moeder) b. [van ons] die fabriek Ergün, III-4 [of us that factory ‘our factory’ (target: onze fabriek) c. [van ons] buurman heeft goed gemaakt child learner (number T41) [of our neighbour has good made ‘our neighbour has repaired it’ (target: onze buurman) (13) full noun possessor a. [van Ergün] auto [of Ergün auto ‘Ergün’s car’ (target: Ergün z’n/s auto) b. [van schoenen] die touwtje8 [of shoes that rope ‘the shoelaces’ (target: de veters van de schoenen; de schoenveters)
Abdullah, II-7
Osman III-6
The examples in (12) and (13) exemplify a new stage of lexical development, viz., the stage in which van is no longer a genitive case suffix, but an adposition preceding the possessor, like van in analytic constructions as in (7). A second alteration vis-a-vis the former stage is that van precedes a full DP instead of being a suffix on a noun. In addition to these examples, there are some rare learner variants of the Saxon genitive and the doubling possessive constructions in which L1 and L2 elements are mixed. These are given in (14). In (14a), we see a target doubling
L1 features in the L2 output
possessive construction and in (14b) a Saxon genitive construction in which the genitive case is overtly realized together with a functional element in Agr head, i.e., ’s and zijn (instead of the target form z’n). (14) a.
die [jongen van] zijn vader that [boy of his father ‘the boy’s father’ (target: de jongen z’n vader) b. [van Ömers] huis [of Ömer’s house ‘Ömer’s house’ (target: Ömer’s huis)
Osman, III-2
child learner (number T29)
Likewise, the complete development of any lexical item can be represented from the L2-initial state, viz. the final state of the L1 lexicon, to the state where the lexical item of the target language is attained. The formal features required for building target syntactic objects, viz. those dominated by the functional head Agr and those of possessor and possessee, were already presented in Table 6. Table 8.Developmental stages of the L2 lexical item /van/ in constructions with full DP possessors; changes with regard to the previous stage are in italics
L1 agreement pattern – phonological matrix – semantics – categorial value – subcategorization frame – example L2 construct pattern – phonological matrix – semantics – categorial value – subcategorization frame – example L2 analytic pattern – phonological matrix – semantics – categorial value – –
subcategorization frame example
Stage 1
Stage 2
Stage 3
Target
/Ø/ ‘possessor’ [affix gen] [N−] Ays¸e-Ø auto
/van/ ‘possessor’ [affix gen] [N−] [Ays¸e-van] auto
/van/ d.n.a. ‘possessor’ [adposition] [−DP pos’sor] [van Ays¸e] auto
/van/ ‘possessor’ [affix gen] [N−] Ays¸e-van d’r auto
/van/ d.n.a. ‘possessor’ [adposition] [−DP pos’sor] van Ays¸e’s auto Ays¸e d’r auto Ays¸e’s auto
/van/ ‘possessor’ [preposition] [−N,−V] [DP pos’see−] auto [van Ays¸e]
/van/ ‘possessor’ [preposition] [−N,−V] [DP pos’see−] auto [van Ays¸e]
/van/ ‘possessor’ [preposition] [−N,−V] [DP pos’see−] auto [van Ays¸e]
83
84
Ineke van de Craats
Table 9.Distribution of variants of the lexical possessive item /van/ produced by four Turkish learners in constructions with full DP possessors Mahmut Cycles
In this way, lexical item learning in all aspects can be made visible. The vocabulary-internal development of the L2 lexical item /van/ can be represented as in Table 8 for constructions with full DP possessors. Three developmental stages are distinguished, each of them is characterized by one or more changes compared to the previous stage. The order of the stages is determined by the first emergence of the changed value. In stage 1, the possessive L2 expressions are based on an L1 pattern and the genitive case is not overtly realized. Stage 2 is characterized by the new (analytic) L2 pattern in which the genitive case is realized by van, both as a preposition and an affix. Likely, the new construction has impact on the agreement pattern from the L1 because learners interpret van correctly as the morphological realization of genitive case, but they mis-assess its categorial value. In stage 3, van is used as a preposition and adposition as well. In Table 9, the distribution of these lexical stages over the four Turkish informants is given. The time course represents three successive cycles, each representing ten months of data collection. The informants are arranged from the slowest learner on the left of the table to the faster learners on the right. From the combination of the data in the Tables 8 and 9, it can be inferred that Mahmut just entered stage 2, that Ergün attained this stage somewhat earlier, and that Abdullah and Osman have reached stage 3 at the beginning of the data collection. Osman is the only informant who produced a few doubling possessive constructions. Table 9, also shows a considerable overlap between the stages. In cycle III, Osman and Abdullah have abandoned stage 1, Mahmut and Ergün have not yet done so.
L1 features in the L2 output
5.2 The personal pronoun in an L1 pro-drop system Turkish is a pro-drop language (cf. Kornfilt 1997). This implies that the personal pronoun that functions as the subject is dropped if the subject has already been introduced in or is known from the discourse. To be more precise, the subject is assumed to be a lexically empty pronominal (pro) with the same person and number features as its lexical equivalent. A full pronoun is used either for the reintroduction of a person or thing, or for emphasizing the subject, e.g. in (15b). In the latter case, the subject would be stressed in Dutch. (15) a.
(pro) geli-yorum come-pres.1sg b. ben geli-yorum I (stressed) come-pres.1sg ‘I come’
(non-focused) (focused)non-
The same holds for the pronoun in a possessive noun phrase like his car. The presence of a lexical pronoun yields a reading in which the possessor is emphasized, e.g., in (16b), his car as opposed to your car. In neutral, non-focused readings, the pronoun is not lexically expressed but contains number and person features, as in (16a). (16) a.
So, in Turkish two different phonological matrices, /ben/ and /Ø/, can be linked to the same formal feature bundle, as illustrated in the columns 2 and 3 of Table 10. Notice here that empty phonological matrices are not restricted to learner variants as in (11) and in the Tables 7 and 8. Dutch does not have the possibility of pro-drop and therefore can only use a lexically filled pronominal ik (‘I’). According to recent analyses (see, among others, Corver and Delfitto 1999) Dutch pronouns should be considered transitive determiners (D). It is assumed that their complement is a pro which stands for the phrasal projection NP and not just for the head noun: [DP D [NP pro]]. The name prodeterminer seems to be more appropriate for these determiners than pronoun. Turkish pronouns, however, seem to be pronominals in the true sense of this word; i.e. they seem to be of the lexical category type N rather than of the functional
85
86
Ineke van de Craats
Table 10.Lexical items of the concept ‘I’ in Turkish and Dutch L1 item (Turkish) – – –
categorial type D (see Kornfilt 1997: 300). Support for this comes, for example, from their behaviour with respect to various morphological rules; like common nouns, pronominals in Turkish function as stems to which case and plural morphology can be attached. The two types of pronominals in Turkish and Dutch are compared in Table 10. The differences between the two pronominal systems consist in (i) the categorial value of the pronoun, and (ii) the use of overt pronouns as a means of focalising. Given the Conservation Hypothesis, Turkish L2 learners match a phonological matrix of the L2 with an L1 feature bundle. The question arises which combination they will make. As can be inferred from the L2 data, Turkish learners link the non-focused item (an empty pro), which is more frequently used in Turkish than the focused one and can be considered the default form, to the Dutch full pronoun. That is not surprising because, in initial stages, adult learners turn out to perceive full pronouns before clitic pronominals (cf. Broeder 1991), although the clitic pronouns occur more frequently in the spoken L2 input. However, the question remains for L2 learners how to express the focused pronoun in their L2 Dutch. Assuming that Dutch pronominals are of the categorial type N as well, their projections may contain a slot for demonstrative determiners. So these L2 learners may attach a demonstrative die or deze (‘that’) to the pronoun in emphatic contexts, as in (17a), in which Mahmut explains why his daughter must learn Turkish, and in (17b), in which Ergün talks about his brother. (17) a.
[Die ik] Hollands dochter Hollands. [that ik Dutch daughter Dutch ‘When I speak Dutch, my daughter speaks Dutch.’
Mahmut, I-4
L1 features in the L2 output
[Die mijn] mama [die ik] Turks praten mijn dochter [that my mother [that I Turkish speak my daughter Hollands praten. Dutch speak ‘When my mother and I speak Turkish, my daughter speaks Dutch.’ b. [die van mijn] broer Ergün, I-3 [that of my brother ‘my (emphatic) brother’
Table 11 represents schematically how L1 knowledge is linked to L2 knowledge in the learners’ lexicon. Table 11.Learner variants of the lexical item /ik/ compared to the target language; deviances from the target language are in italics Learner variant 1 – – –
6. Functional elements in the clausal domain An interesting example of how small differences in feature bundles play a role in L2 acquisition is the acquisition of the possessive verb hebben (‘to have’). A number of studies (e.g. Benveniste 1966, Freeze 1992, Moro 1997) have pointed out that possessive have-constructions derive from an underlying locative construction, and differ only in some respects from possessive constructions without have. Under this view, the form heeft (‘has’) is considered a form of to be in which a locative preposition has been incorporated. In that way the Dutch form heeft is the spell out of the features Tense (present) (T) + Agreement (3sg) (Agr) + Locative (LOC). Basically, the possessor is considered to be the complement of the locative preposition (P); this PP is moved to the subject position (in SpecAgrsP) after the incorporation of the locative prepostion. The derivation is as follows in (18).9
87
88
Ineke van de Craats
…dat Paul een motor heeft.10 …that Paul a motorcycle has ‘… that Paul has a motorcycle.’ [CP dat [AgrP Pauli [[TP ti [SC een motor [PP Pk ti]] [T+Pk]j]] Agr+ [T+P]j (= heeft)]] b. Paul heeft een motor. Paul has a motorcycle ‘Paul has a motorcycle.’
(18) a.
Some languages also show a non-incorporated variant like French (ce livre est à moi ‘this book is to me’). Moroccan Arabic, however, does not have a variant in which the locative preposition is incorporated in the be-copula, viz. no equivalent of the verb to have. Moroccan Arabic uses only a locative sentence with the PP at+clitic (i.e. Eend-u at+him), both for a present tense sentence (19a) and for a past tense sentence (19b) (cf. Harrell 1970). Abder, Eend-u dar kbira. (present tense) Abder at-him house-f big-f ‘Abder has a big house’ [CP Abder, [CP [AgrP [Agr¢ [Eend-u]i [SC dar kbira [PP ti]]]]]] b. Abder, kanet Eend-u dar kbira. (past tense) Abder cop.past.3sg.f at-him house-f big-f ‘Abder had a big house.’ [CP Abder, [CP [AgrP [Agr¢ kanetk [TP [Eend-u]i [T¢ tk [SC dar kbira ti]]]]]]]
(19) a.
Notice that, in (19a), a copular form is lacking in the present tense and the preposition Eend-u occupies the position of the verb (in Agr), whereas in (19b), the position of the verb in Agr head is taken by the past tense copula kanet and Eend-u is in SpecTP. This may have interesting consequences for L2 learners who are in search of equivalents of their L1 grammar in the environmental input of the L2. Let us make first a comparison of the concept ‘has’ in the source and the target languages (see Table 12). Just as was the case for the genitive case marker in Section 5.1, L2 learners start with an empty phonological matrix linked to the L1 feature bundle, as can be seen in the following L2 expressions produced by adult Moroccan learners. In (20a), Fatima is asked about her work in Morocco: she had a small shop with some knitting machines. In (20b), Mohamed describes his situation in an interview with a housing official.
L1 features in the L2 output
Table 12.Lexical items of the concept ‘have’ compared for source and target languages L1 item (Moroccan Arabic)
L2 item (Dutch)
/Eend- / ‘with/at’ [−N,−V]
/heeft/ ‘has’ [Agr +T +P] 3sg +PRES +LOC [DP−]
– – –
phonological matrix semantics categorial value
–
subcategorization frame [−DP ·cliticÒ]
(20) a.
pronominal possessor Ik klein winkel. (= ik Ø klein winkel) Fatima, I-3 I small shop ‘I had a small shop.’ b. nominal possessor Mijn vrouw ook klein huis. (= mijn vrouw Ø ook klein huis) my wife also small house ‘My wife has also a small house.’ Mohamed, I-6
Slow learners often show a stage in which they produce a preposition between the possessor and the possessee, e.g. in (21). The preposition appears in the same position where it would appear in the L1, between the strong pronoun (not the clitic) or the full noun phrase in the dislocated position at the beginning of the sentence (cf. the position of Abder in (19a, b) and second noun phrase. In (21a), Fatima talks again about her work and, in (21b), about her relatives on a photograph: three of them are half sisters and brothers. (21) a.
pronominal possessor Ik met klein winkel. I with small shop ‘I had a small shop.’ [CP ik [CP [AgrP metj [SC klein winkel [PP tj]]]]] b. nominal possessor Fatiha Mustafa Khilifiye met andere moeder. Fatiha Mustafa Khilifiye with other mother ‘Fatiha, Mustafa and Khilifiye have another mother.’
Fatima, I-3
Fatima, II-4
In more advanced stages of L2 development, learners happen to produce sentences in which the locative character of the possessive clause becomes manifest, e.g. in (22a) and (22b), in which the entire PP has been moved to
89
90
Ineke van de Craats
SpecAgrP, or (22c) in which the locative preposition is overtly spelled out and incorporated in heeft as well. Met kind een jaar.11 with child one year ‘The child is one year old.’ b. Bij hem kief. at him hashish ‘He has hashish.’ c. Met Soumiya heeft veel pijn. with Soumiya has much pain ‘Soumiya suffers very much.’
(22) a.
Fatima, II-7
HassanK, III-2
HassanM, III-5
Except for Fatima, such locative constructions instead of the verb have are rare in the data produced by the Moroccan informants. When their have-constructions are considered more closely, however, we become more sceptical about the character of the have-forms. To start with Fatima, she uses the form heeft (3sg) not only for all person roles, but she alternatively uses met and heeft with the same meaning for more than 13 months. Another informant produces instances of heeft in which it can be a synonym for met (23a). The sentence in (23b) is even a direct copy of Moroccan Arabic in which heeft is followed by a clitic as if it were a locative preposition. Die heeft geel haar mag binnen. that has yellow hair may inside ‘The boy with blond hair is allowed to go inside.’ b. Die meisje, heef-ze een oom. that girl has-she an uncle ‘That girl has an uncle.’
(23) a.
HassanK, II-2
HassanK, II-2
The situation becomes crucial when this informant wants to express a possessive clause in the past tense. Recall from (19b) that, in Moroccan Arabic, the past tense of a have-clause is expressed by a be-copula in the past tense followed by the locative preposition and see what HassanK produced in (24). Die was heeft (-pro) 30 jaar. HassanK, II-3 he was-cop.past.3sg has [=Eend+3sg] 30 years ‘He was 30 years.’ b. Die meisje was nooit heeft (-pro) verkering. II-4 that girl was-cop.past.3sg never has [=Eend+3sg] relationship ‘That girl was never in a relationship.’
(24) a.
L1 features in the L2 output
c.
Dan was heeft (-pro) een huis. then was-cop.past.3sg has [=Eend+3sg] a house ‘Then he had a house.’
HassanK, II-9
From the examples above we infer that HassanK and Fatima base their syntactic have-structures on the feature bundle of the Moroccan Arabic preposition Eend (‘at, with’). In this developmental stage, they still match the L1 feature bundle with an L2 phonological matrix. Heeft is in fact a prepositional heeft. The other two informants do not produce past tense forms of the verb hebben; that is already sufficient reason to be sceptical about the real identity of their have-forms. The question can be raised how and when we can be sure that heeft is a real copula and no longer a prepositional form. It is not easy to decide when we are dealing with a copular verb, but it can safely be claimed that it is no longer a preposition when past tense forms of hebben are expressed in one verbal form (had ‘had’). HassanK succeeds in producing several past tense forms and Mohamed does only once (see Table 14). A second indication is that the verb hebben has extended from a possessive copula to an auxiliary of the perfect tense, e.g. hij heeft gezien (‘he has seen’). In this light, it is relevant to signal that all Moroccan learners produce have-copulas six months or more before they start using the auxiliary have (see Van de Craats 2000). Finally, the development of the lexical item /heeft/ is represented in Table 13, from the L2-initial state (i.e. the empty phonological matrix with the L1 feature bundle) via various developmental stages to the state in which the item has attained the complete feature constellation it has in the target language. Tables 13 focuses on the properties of the lexical item heeft, which is the spell out of the features [agreement], [tense] and [−N,−V]. This already suggests a process of adjunction of several functional heads. In Table 13, we abstract away from the formal features dominated by T and Agr triggering the movement of P(P) (see Van de Craats et al. 2002: 156 for the derivation). Table 14 shows how the stages are distributed over the data of the four Moroccan learners. Slow learners exhibit more tokens of the developmental stages than fast learners. The order of the stages is determined again by the first emergence of a changed value. For the slow learners, a considerable overlap of the stages can be observed. The overlap is smaller for learners like Mohamed and HassanM. At the end of the data collection, stage 3 is attained by all four learners; only two of them can form past tense forms according to target language standards, however.
91
92
Ineke van de Craats
Table 13.Stages in the development of the L2 lexical item /heeft/; differences with regard to the previous stage are in italics
– – – – –
phon.matrix semantics categorial value subcategorization frame example
Stage 1
Stage 2
Stage 3
Target state
/Ø/ ‘with, at’ [−N,−V] [−DP clitic] Abder Ø boek
/met, bij/ ‘with, at’ [−N,−V] [−DP clitic] Abder met boek
/heeft/ ‘with, at’ [−N,−V] [−DP clitic] Abder was heeft boek
/heeft/ ‘has’ [+Agr,+T,+P] [DP−] Abder heeft boek
Table 14.Distribution of variants of the lexical possessive item /hebben/ produced by four Moroccan learners over developmental stages Cycles
7. Conclusions From the examples discussed above, we can infer that, in general, L2 learners are more aware of the fact that words differ from language to language, i.e., that phonological matrices differ, than that they are aware of the fact that other lexical properties differ. Adult L2 learners are inclined to map or to attach a new phonological matrix from the L2 to a bundle of semantic and formal features from the L1. Only if those features bundles differ, can we get an insight in how L2 learners proceed in their learning of lexical items: they start by assuming an L1 feature constellation and they gradually change the features of the bundle one by one, as we saw in Tables 8 and 13 for the acquisition of van and heeft. In this way, the L2 learners’ output becomes more and more target-like; not only the surface form (= the phonological matrix) alters, but also the features. In this light, it can be explained why functional elements are more difficult to acquire than lexical elements (i.e. content words). It is not only the case that
L1 features in the L2 output
they are less salient in the environmental input, but they also differ more in the structure of their feature bundle. Therefore, it takes more time to discover each of the composing features. The formal features, however, are of crucial importance for the development of the L2 syntax: they are the input for the computational system and, hence, decisive for the result of derivation: the syntactic objects. They are the interface itself, so to say.12 Until a Moroccan learner of Dutch, for instance, has not discovered each of the formal features of heeft (‘has’), he cannot attain the formation of the verb with subject-verb agreement, he cannot form the past tense of that verb, nor the perfect aspect by means of the auxiliary hebben. This shows the direct link between lexicon and syntax. Why is there progress? The syntactic objects (i.e. the product of the computational system and the selection of lexical elements, including formal features) in interlanguages seem to provide L2 learners with a reason to change their L2 output, because it is not sufficiently understood by native speakers, incomplete (recall, for instance, the empty phonological matrices in the output of beginners in example (11)), etc. Therefore, L2 learners are forced to constantly reanalyse their L2 output and to change the underlying formal features and/or the parametric values. This process of restructuring seems to occur in a continuous interplay between syntax and formal features as the seeds of syntax and as a part of the lexicon as well.
Notes 1. In Chomsky (1994) it is argued that only the idiosyncratic formal features are part of the lexical item. Optional formal features are added to the lexical item when it is selected from the lexicon. Here, we will abstract away from that distinction. 2. Except for the case in which L1 and L2 are so much similar that an L2 learner decides to use an L1 lexical item (with the L1 phonological matrix) directly in the L2 output. 3. It is certainly not the case that learners keep on using all meaning aspects of a lexical item in their L2: the more transparent and concrete a specific semantic aspect is, the earlier it is used in L2 expressions (cf. the 17 different aspects of breken (‘to break’) the acceptability of which Kellerman (1979) asked his Dutch subjects to judge). 4. A reviewer proposed that betalen geven would be a serial verb construction, or geven ‘to give’ a light verb; geven can function as a light verb in Dutch. Although this account is not quite impossible, the reviewer is arguing too much from a target language perspective. Moreover, there is no light verb geven in (2b), and the example in (3b) makes clear that it is not the verb betalen that is meant.
93
94
Ineke van de Craats
5. A reviewer advanced that Adjémian (1983) already found that L2 learners tend to transfer lexical patterns from their L1 to their L2, and assume that verbs take the same kinds of subject and object in their L2 as they do in their L1. Mapping an L2 phonological matrix onto L1 lexical properties goes even further. 6. Following insights from Longobardi (1996), we take the position that the Saxon genitive and the doubling construction are in essence hidden construct noun phrases, because they share certain properties with the Construct-State construction (as in Arabic and Hebrew). The Saxon genitive and the possessive doubling have the same distribution as a definite article, viz. they block the presence of a definite article in D0. 7. The English translation of these examples suggests that the word order is not uncommon, but possessors DPs in the Saxon genitive are restricted to proper names, and to human beings in the doubling possessive construction. Moreover, recursion is odd in those constructions. A native speaker of Dutch would only use the analytic construction for recursive patterns due to processing limitations. (i) de dochter van mijn tante (ii) de auto van de zoon van mijn tante (iii) de dochter van de tweede man van de oma van mijn vrouw 8. Note that in Dutch, the order Possessor-Possessee is restricted to human possessors; the Saxon Genitive (e.g. Jans fiets ‘John’s bike’) is even restricted to proper names and nouns equivalent to proper names, e.g. tante ‘aunt’. This points all the more to the L1 grammar in example (11b). 9. See also Van de Craats (2000) and Van de Craats, Corver and Van Hout (2002) for details on this construction in Dutch and Moroccan Arabic. 10. In this example, a subordinated clause is given first because it shows the basic position of the verb in Dutch (SOV language). 11. In Moroccan Arabic, one’s age is also expressed by a have-construction. 12. Although the lexicon plays an important role in this view on adult L2 acquisition, it is not identical to the Lexical Learning Hypothesis (e.g. Clahsen, Eisenbeiss and Penke 1996) nor to other recent views on L1 acquisition (e.g. Radford 2000, Roeper 1996) in which the acquisition of (L1) syntax is seen as gradual structure building.
References Adjémian, C. 1976. “On the nature of interlanguage systems”. Language Learning 16: 297–320. Adjémian, C. 1983. “The transferability of lexical properties”. In Language transfer in language learning. S. Gass & L. Selinker (eds), 250–268. Rowley, MA: Newbury House. Benveniste, E. 1966. “‘Etre’ et ‘avoir’ dans leur fonctions linguistiques”. Bulletin de la Société de la linguistique de Paris 55 (1): 113–134. Reprinted in Problèmes de linguistique générale. Paris: Gallimard. Broeder. P. 1991. Talking about people. A multiple case study on adult second language acquisition. Amsterdam/Lisse: Swets & Zeitlinger. Chomsky, N. 1986. Barriers. Cambridge, Massachusetts: MIT Press.
L1 features in the L2 output
Chomsky, N. 1994. “Bare phrase structure”. In MIT occasional papers in linguistics. Cambridge, Massachusetts: MIT Press. Chomsky, N. 1995. The minimalist program. Cambridge, Massachusetts: MIT Press. Clahsen, H., Eissenbeiss, S. and Penke, M. 1996. “Lexical learning in early syntactic development”. In Generative perspectives on language acquisition, H. Clahsen (ed.), 129–159. Amsterdam/Philadelphia: John Benjamins. Corver, N. 1990. The syntax of left branch extractions. Doctoral dissertation. Tilburg University. Corver, N. and Delfitto, D. 1999. “On the nature of pronoun movement”. In Clitics in the Language of Europe, H. van Riemsdijk (ed.), 799–861. Berlin: Mouton & de Gruyter. Flynn, S. 1986. A parameter-setting model of second language acquisition. Reidel, Dordrecht. Freeze, R. 1992. “Existentials and other locatives”. Language 68: 553–595. Harrell, R. 1970. A short reference grammar of Moroccan Arabic. Washingon, D. C.: Georgetown University Press. Kellerman, E. 1979. “Transfer and non-transfer: Where we are now”. Studies in Second Language Acquisition 2: 37–57. Kornfilt, J. 1997. Turkish. London, New York: Routledge. Longobardi, G. 1996. The syntax of N raising: A minimalist theory. Manuscript, Utrecht University. Miller, P. 1991. Clitics and constituents in phrase structure grammar. Doctoral dissertation, University of Utrecht. Moro, A. 1997. Predicative noun phrases and the theory of clause structure. Cambridge: Cambridge University Press. Perdue, C. 1993. Adult language acquisition: Cross-linguistic perspectives. Volumes I and II. Cambridge: Cambridge University Press. Radford, A. 2000. “Children in search of perfection: towards a minimalist model of acquisition”. In Essex research reports in linguistics, 34. Roeper, T. 1996. “The role of merger theory and formal features in acquisition”. In Generative perspectives on language acquisition, H. Clahsen (ed.), 415–449. Amsterdam/ Philadelphia: John Benjamins. Schwartz, B. D. and Sprouse, R. 1996. “L2 Cognitive states and the full transfer/full access model”. Second Language Research 12: 40–72. Van de Craats, I. 2000. Conservation in the acquisition of possessive constructions. A study of second language acquisition by Turkish and Moroccan learners of Dutch. Doctoral dissertation, Tilburg University. Van de Craats, I., Corver, N. and Van Hout, R. 2000. “Conservation of grammatical knowledge: on the acquisition of possessive noun phrases by Turkish and Moroccan learners of Dutch”. Linguistics 38 (2): 221–314. Van de Craats, I., Corver, N. and Van Hout, R. 2002. “The acquisition of possessive haveclauses by Turkish and Moroccan learners of Dutch”. Bilingualism: Language and Cognition 5 (2): 147–174. Vermeer, A. 1986. Tempo en struktuur van tweede-taalverwerving bij Turkse en Marokkaanse kinderen. Doctoral dissertation, Tilburg University. White, L. 1982. Grammatical theory and language acquisition. Dordrecht: Foris. White, L. 1985. “The pro-drop parameter in second language acquisition”. Language Learning 35: 47–62.
95
Chapter 5
Measures of competent gradience* Nigel Duffield McGill University & Max Planck Institute for Psycholinguistics, Nijmegen
1.
Introduction: Reconsidering the competence–performance distinction
An assumption of much current generative research is that grammaticality is a strictly categorical property. By extension, underlying competence, of which grammaticality is a reflex, is also assumed to be categorical in nature.1 In most analyses, this assumption about grammaticality is explicitly formalized in the convention of placing asterisks in front of ‘ungrammatical’ sentences: starred sentences are categorically ‘bad’, unstarred sentences are unequivocally ‘good’.2 A second, equally standard assumption in generative research is that acceptability judgments are not categorical, at least if one ignores such uninteresting deviances as arbitrary word scrambles, nonce-word insertions, and so forth. Everybody knows that acceptability judgments on most theoretically relevant sentences are only relative, that the particular level of acceptability of a given sentence is a complex function of a variety of factors, including extragrammatical ones. It is notable that L2 researchers have been more explicit in recognizing this fact than pure theorists (see especially Birdsong 1989, Ellis 1991, Hedgcock 1993, Martohardjono 1998; cf. also Coppieters 1987, Mandell 1999, but the view is more generally held: see Bever 1970, Greenbaum 1977, Schütze 1996, for discussion and further references). In spite of the tension between these assumptions, most theoretical linguists use continuous performance data to draw inferences about (what is assumed to be) categorical grammaticality, and by extension, categorical underlying competence.3 Having once ‘idealized to competence’, theoretical linguists are free to ignore those aspects of performance that they deem irrelevant to grammaticality: this includes the property of gradience that is a principal concern of this paper.
98
Nigel Duffield
Most of the foregoing discussion is familiar to anyone who has considered the competence issue, so it might be wondered why I am bringing it up again. There are three reasons, two of which apply to linguistic research in general, one which is more specifically concerned with the relationship between nativespeaker (NS) and non-native-speaker (NNS) data in second language acquisition research. The first reason for critically re-examining the competence–performance distinction is to save it from attack from an increasingly coherent body of research that takes seriously the complexity of grammatical systems whilst at the same time questioning the empirical validity of that distinction.4 One articulate statement of this new position is presented in the ‘emergentist’ work of Allen and Seidenberg (1999). Allen and Seidenberg’s attack on the competence–performance distinction as it relates to language acquisition is motivated by what they claim to be the obscure basis of grammaticality judgments: The mapping between competence grammar and performance is at best complex, as we have noted: it is also largely unknown. A problem arises because the primary data on which the standard approach relies — grammaticality judgments — are themselves performance data (Bever 1970). The methodology of the standard approach holds that properties of the hypothesized language faculty can be identified on the basis of experts’ intuitive judgments of the well-formedness of utterances. However, the relationship between grammaticality judgments and the structure of the grammar is no more transparent than that between other aspects of competence and performance. (Allen and Seidenberg 1999: 117–118)
While one might take issue with the reference to ‘experts’ intuitive judgments’, the rest of this paragraph is unexceptionable. Allen and Seidenberg further point out that misgivings about the basis of grammaticality judgments are to be found within the generativist camp. The authors cite a telling paragraph in Schütze (1996: 20): It is conceivable that competence in this sense of a statically represented knowledge does not exist. It could be that a given string is generated or its status computed when necessary, and that the demands of the particular situation determine how the computation is carried out, e.g., by some sort of comparison to prototypical sentence structure stored in memory. Since such a scenario would demand a major re-thinking of the goals of the field of linguistics, I will not deal with it further.
Measures of competent gradience
Allen and Seidenberg claim that this sidesteps an important issue. Their response is to try to derive the effects of grammaticality from an input-driven system lacking any autonomous syntactic representations: that is, without any reference to grammatical competence as theoretical linguists would construe it. For various reasons, including both the results of their own simulation and the results of other experiments to be discussed below, I believe that this alternative approach to grammaticality is mistaken, and that something like the competence–performance distinction is empirically more correct. At the same time, their work demonstrates the urgent need for a clearer articulation of the nature of the distinction, and for a more explicit statement of the content of grammaticality judgments. While I will not be able to offer that statement here, I hope to sketch out the direction in which it might be found. A second reason for reconsidering the competence–performance distinction is that I agree with those who claim that the line between competence and performance is presently drawn in the wrong place: that the current idealization to competence presents too narrow a view of scope of competence. This is a view held not only by connectionists, but also by some generativists, most notably Culicover (1998, 2000), though cf. Fodor (2001). The specific notion that I would like to include within a revised version of competence is that of syntactic gradience. As the title of this article implies, I will argue that part of competence consists, not in correctly making categorical judgments about grammaticality, but in correctly making gradient ones. To cite one brief example, to be a competent ‘knower’ of English is to know, not merely that the sentences in (1) and (2) below are all less than perfect, but also that they are rather precisely ordered in acceptability, each example being slightly less acceptable than the one that immediately precedes it. These examples, cited in Kluender (1992), are originally due to Chung and McCloskey (1983) (where the symbol ‘>’ indicates ‘more acceptable than’): (1) a.
This is the paper that we really need to find someone who understands. > b. This is the paper that we really need to find a linguist who understands. > c. This is the paper that we really need to find the linguist who understands. > d. This is the paper that we really need to find his advisor who understands. > e. This is the paper that we really need to find John, who understands.
99
100 Nigel Duffield
(2) a.
This is a paper that you really need to find someone that you can intimidate with. > b. Which paper do you really need to find someone that you can intimidate with. > c. How many papers do you really need to find someone that you can intimidate with. > d. What do you really need to find someone that you can intimidate with.
Kluender’s concern in presenting such examples is to draw out the ‘hidden’ factors determining the acceptability of these sentences: all of the sentences in (1), and those in (2), involve very similar syntactic structures, yet their grammaticality status varies from (almost) completely acceptable to strongly unacceptable. In this case, the degrees of acceptability are a function, not of syntactic structure, but of a semantic factor, what Kluender terms ‘referential specificity’. One conclusion to be drawn from such examples is that we are failing to account for speaker’s intuitions by attributing these grammaticality effects to purely syntactic conditions. This is Kluender’s contention. I return to the ‘hidden factors’ issue in Section 2.4. below, in discussing the factors determining another putatively syntactic effect — the parallelism constraint on VP-ellipsis. At this point, however, my concern is with the fact that this is an instance of ‘competent gradience’: competent knowers of the language judge these sentences as increasingly less acceptable, and they also converge on the particular ordering involved. This becomes especially relevant to SLA when we consider a hypothetical advanced L2 learner’s response to such sentences.5 On the categorical view of competence, an L2 learner who systematically rejected all of these sentences as unacceptable (*) should in fact be ‘more correct’ relative to the hypothesized grammar (I-language) than the native speaker who is inclined to accept most or all of them (most or all of the time). Although unusual, such instances of L2 learners ‘outperforming’ native-speakers are not unknown; the existence of this phenomenon is the third reason to re-examine the competence–performance distinction. In fact, I will suggest that the hypothetical L2 learner’s performance does reveal an underlying categorical competence. Intuitively, however, a learner who judges as unacceptable what nativespeakers judge as less acceptable has somehow failed to achieve native-speaker competence, if we define competence as whatever implicit knowledge determines successful language use. By one measure, then, the L2 learner whose judgments fail to match those of the native-speaker will be judged more competent than the native speaker,
Measures of competent gradience
where competence is defined in terms of convergence between a discrete pattern of behaviour and underlying categorical knowledge. Another measure — and common sense — dictates that the same L2 learner exhibiting the same behaviour should be judged less competent than the native-speaker — where competence is defined in terms of convergence on native-speaker’s judgments. Although most generative SLA researchers would claim that they are probing the first type of competence, I know of very little work that does not determine this competence — implicitly or explicitly — in terms of the second type.6,7 There have been two responses to this apparent paradox. As noted above, the anti-generativist response is simply to reject the idea of competence as statically represented, autonomous knowledge. The standard generativist response is to continue to pretend that gradience is not part of syntactic competence. Neither alternative seems satisfactory.8 A middle way is to claim that both impulses are correct, because there are two types of competence. This is the proposal I will try to motivate here. For the purposes of exposition, I’ll term these hypothesized competences, underlying versus surface competence, respectively, with no preference intended. By hypothesis, underlying competence (UC) is categorical, and consists of formal (phonological and syntactic) principles, autonomous from the lexicon. It is plausible to think of UC as innate. Surface competence (SC), by contrast, is intimately determined by the interaction of contextual and specific lexical properties with the formal principles delivered by UC; as a consequence, SC generates gradient effects. SC is largely language-specific learned knowledge. In principle, grammaticality judgments can be a reflection in performance of either type of competence; generally, however — again, by hypothesis — explicit grammatical judgment tasks will tap surface competence, whereas implicit tasks will tap underlying competence.9 Now, it may happen that, in a particular task, a given set of judgments will be ‘weakly consistent’ with both types of competence, and also that L1 and L2 learners’ judgments will converge. This state-of-affairs is schematised in Figure 1 below. Gradient judgments will be weakly consistent with UC just in case (i) all sentences judged less than perfectly acceptable violate some formal constraint, (ii) none of the sentences judged to be perfect violate this constraint, and (iii) it is reasonable to attribute the relative degrees of acceptability of the unacceptable set to some, non-formal (preferably extra-grammatical) factor. This I take to be the standard generativist line.
101
102 Nigel Duffield
Common set of judgments (shared by NSs & NNs)
Surface competence
Underlying competence
Figure 1.Full convergence (SC and UC generate the same set of grammatical sentences, NS and NNS converge on this set).
However, several legitimate alternatives exist. These are cases where either the performance judgments on the data presented are a reflex of only one type of competence (UC or SC), or the judgments of two subject groups (L1 versus L2) reflect different competences. These logical alternatives are schematised in Figures 2–5 below. Common set of judgments (shared by NSs & NNs)
Surface competence
Underlying competence
Figure 2.Convergence on SC only.
Common set of judgments (shared by NSs & NNs)
Surface competence
Figure 3.Convergence on UC only.
Underlying competence
Measures of competent gradience 103
Native Speaker Judgments
Non-Native Speaker Judgments
Surface competence
Underlying competence
Figure 4.Parallel, disjoint convergence (Type 1: NS converge on SC, NNS on UC). Non-Native Speaker Judgments
Surface competence
Native Speaker Judgments
Underlying competence
Figure 5.Parallel, disjoint convergence (Type 2: NNS converge on SC, NS on UC).
Notice that all of these scenarios assume convergence between NS and NNS competences; that is to say, I am ignoring those cases where L2 learners have internalised a surface or underlying competence that is different from that of native-speakers; cf. Sorace (1993, 1996). What this model is intended to account for are cases in which the L2 learner has attained the target grammar, but where his/her judgment patterns nevertheless diverge systematically from those of the native-speaker. Clearly, hypothesizing two different types of competence complicates any theory of the relationship between competence and performance. Parsimony requires that this complication be shown to be empirically necessary. The purpose of the rest of this article, therefore, is to provide such motivation. The organization is as follows. The following section provides some experimental evidence motivating two types of competence. Next, I consider two types of gradient effect, both of which I will argue are properties of surface, rather than underlying, competence. Then, I discuss two instances of principled mismatch between NS and NNS judgments, motivating both Figures 4 and 5 above. Finally, I examine a special case of Figure 2 above, where NS and NNS apparently converge on the same SC, but for different reasons.
104 Nigel Duffield
2. The dual competence approach 2.1 Motivating two types of competence This section draws attention to two pieces of recent work that suggest a distinction between two types of competence. In each case, complementary methodologies have been used to investigate speakers’ judgments of particular syntactic phenomena, which have been argued (in the theoretical literature) to exhibit a categorical distinction. In each case, these methodologies have yielded divergent, but nevertheless coherent, results. This divergence, I would claim, reflects the fact that two different types of competence are being tapped. 2.1.1 McKoon and MacFarland (2000) McKoon and MacFarland’s (2000) study investigates the theoretical argument for a discrete representational contrast between two classes of verbs: externallyversus internally-caused change of state verbs. As these labels suggest, externally-caused change-of-state (ECS) verbs are those whose result state may be brought about by some external agent/cause, whereas internally-caused change of state (ICS) verbs describe result states whose cause is internal to the theme NP undergoing change-of-state. An example of the former is the verb redden, an example of the latter, bloom. The theoretical literature on such predicates analyses the difference in terms of a discrete lexical contrast: the lexical argument-structure of ECS verbs is said to involve an additional abstract predicate (CAUSE), which is systematically absent from the argument-structure of ICSs. In the specific analysis of Levin and Rappaport Hovav (1995), this entails a structural difference in the lexicosyntactic representation of the two verb-types: as illustrated in (3) below, ECS verbs project more structure than ICS verbs. (3) a. ECS: ((α) CAUSE (BECOME (x ·STATEÒ))) b. ICS: (BECOME (x ·STATEÒ))
One piece of empirical evidence supporting this structural contrast comes from the alleged (in)ability of these intransitive verbs to form transitive counterparts: it has been claimed that ECSs, but not ICSs, permit transitive alternants. So, for example, there is claimed to be a categorical contrast in the acceptability of (4a) versus (4b): (4) a. The sunshine reddened Amy’s cheeks. b. *The sunshine bloomed the tulips.
Measures of competent gradience 105
McKoon and MacFarland tested the validity of this empirical claim through a number of different studies. First, they examined a corpus of approximately 180 million words of written and spoken English. The results of that investigation, summarized in Table 1, below demonstrate that ICSs considered as a class are in fact just as likely to transitivise as ECSs. Table 1.Probability of transitive use for ECS and ICS verbs (adapted from McKoon and MacFarland 2000: Table 2) SentenceType
Low prob. Transitive
Higher prob. Transitive
External Cause
Internal Cause
verb
prob. ‘yes’
verb
prob. ‘yes’
atrophy awake crumble abate
.03 .05 .05 .10
bloom deteriorate germinate rot
.00 .01 .06 .08
mean
.06
mean
.06
redden dissipate fray fossilize
.24 .41 .52 .60
blister ferment corrode erode
.22 .54 .63 .67
mean
.48
mean
.45
All of these predicates vary individually in transitivity — some ECSs such as abate show a low probability of transitivity, others such as fossilize are more likely to transitivise: the same is true of ICSs. In terms of usage, then, the transitivity constraint does not distinguish the two classes of predicate; rather, transitivity is a lexically gradient effect. Thus, the corpus study results apparently invalidate a key piece of empirical evidence for a representational contrast between the two verb classes. The proposed theoretical distinction becomes even more questionable in light of a follow-up study, in which subjects were asked, in an offline grammatical judgment task, to rate the acceptability of various transitivised ECS and ICS verbs. The results of this task confirmed the patterns already observed in the spontaneous production data, namely, that the acceptability of the transitive alternant of a given verb was a function of individual lexical differences, and of selectional restrictions, rather than putative verb-class. Statistically, there is no main effect of verb class, nor an interaction between verb-class and transitivity,
106 Nigel Duffield
in the acceptability ratings. Overall, the acceptability rate for both classes of verbs was extremely high: even those verbs judged least acceptable had a mean acceptance rate of over 80%. The results of the first two studies speak against the theoretical claim that these two types of verbs are represented differently. On the contrary, subjects’ close convergence on a scale of relative acceptability (for the transitive forms of individual predicates) suggests that a fine-grained, lexically determined, and inherently gradient, type of competence informs native-speakers’ performance. Had McKoon and MacFarland stopped at this point, their work would provide support for the usage-based, probabilistic models of language performance favoured by connectionists and others (cf. especially MacDonald et al. (1994), Barlow and Kemmer (2000)). A final experiment, however, suggested a different conclusion. In this latter, implicit experiment, McKoon and MacFarland measured the response latencies involved in reading ECSs versus ICSs that had previously been matched in terms of length and offline acceptability. In direct contrast to the previous results, this implicit measure revealed a reliable contrast in reading time between ECS and ICS verbs; ECS verbs — whether presented in intransitive or transitive frames — took significantly longer to read than their matched ICS counterparts. The main results are reproduced in Table 2 below (from McKoon and MacFarland: Tables 7 and 8: 848, 851). That is to say, the results of these last experiments support the idea of distinct representations for these verbs on the basis of the hypothesized verb classes. McKoon and MacFarland argue that the results of the final experiment speaks against probabilistic models of lexical representation, and confirm the psycholinguistic reality of the theoretical model. Both conclusions are too strong, I think. On the one hand, whether or not there is a categorical distinction in the representation of the two classes of verbs, it is still necessary to account for subjects’ ability to converge on a scale of gradient judgments for individual predicates. As discussed earlier, a speaker whose judgments of the acceptability of such predicates was the inverse of all other subjects (in Experiment 2) could reasonably be judged ‘less competent’ than one whose relative judgments were in accordance with other native speakers (at least by the second definition of competence), even if their reading times in Experiment 3 were comparable. Moreover, the fact that subjects’ performance in an implicit task yields a statistically discrete result does not prove that the competence underlying this behaviour is itself categorical, and certainly is no more than consistent with the
Measures of competent gradience 107
Table 2.Mean reading times for ECS and ICS verbs (from McKoon and MacFarland 2000: Tables 7 & 8) Sentence Type
External Cause
Internal Cause
Intransitive frames
JTime (ms)
prob. ‘yes’
JTime (ms)
prob. ‘yes’
All test sentences Low prob. Transitive Higher prob. Transitive
1551 1561 1538
.91 .92 .90
1400 1392 1413
.96 .96 .96
Intransitive frames
External Cause
Internal Cause
JTime (ms)
prob. ‘yes’
JTime (ms)
prob. ‘yes’
2220 2230 2210
.86 .81 .93
2069 2131 1963
.96 .96 .96
All test sentences Low prob. transitive Higher prob. transitive
theoretical analysis. To infer the reality of a particular contrast in underlying competence from an apparently isomorphic contrast in processing is to assume a very direct interpretation of the derivational theory of complexity that most psycholinguists would view with some caution. Having said that, the results of the third experiment do provide support for the idea that some correlate of the categorical contrast described by the theoretical analysis has psychological reality, and that this correlate does seem to be categorically expressed. I suggest that an adequate model is one that accommodates both sets of results. Rather than attempting to reconcile these within a single type of competence, the proposal is to assume a dual-competence model, as represented in Figures 1–5 above. In this model, the results of the first two experiments can be represented as in Figure 2 (tapping SC), those of the final experiment as in Figure 3 (tapping UC).10 2.1.2 Duffield and White (1999), Duffield et al. (2002) The dual competence model receives additional support from recent SLA work using complementary methodologies to investigate L2 learners’ knowledge of pronominal clitic placement in Spanish and French: Duffield and White (1999), Duffield et al. (2002). In this section, my concern is with the divergent results of the French native-speaker control group in one section of experiments. Here, unusually, an offline grammaticality judgment task revealed a grammaticality contrast to which the implicit (sentence-matching) task was apparently insensitive.
108 Nigel Duffield
(5) a.
Je veux le voir. I want 3sg to.see ‘I want to see him.’ b. *Je le veux voir. c. Je le fais chanter. I 3sg make to.sing ‘I make him sing.’ d. *Je fais le chanter.
The phenomenon of interest, illustrated in (5) above, is the contrast in pronominal clitic placement between so-called restructuring verbs (such as pouvoir, vouloir) and the causative verb faire. The distinction is of theoretical interest in that the two structures are assumed to involve distinct syntactic representations. Briefly, many theoretical analyses assume that restructuring verbs — as the name suggests — restructure the syntax of the lower clause, creating a monoclausal structure. By contrast, sentences involving causatives are assumed to remain fundamentally bi-clausal: at an abstract level of representation, the pronominal clitic is still syntactically associated with the lower verb; see Duffield et al. (2002), for more detailed discussion. Notice that this analysis, derived on the basis of cross-linguistic comparison with other Romance varieties, is precisely the opposite of that suggested by a naïve inspection of the French facts. In French, the pronominal clitic stays close to the verb of which it is an argument in restructuring contexts, and is displaced from it in causative contexts: hence, one might expect that restructuring would be required for the interpretation of clitics with causatives, rather than with verbs like vouloir. The theoretical analysis, however, predicts a distinction between the acceptability of two types of ‘ungrammatical’ sentence, namely, between (5b) and (5d) above: whereas (5b) is predicted to be ungrammatical at all levels of representation, (5d), in which the clitic is attached to the verb with which it is thematically associated, is predicted to be ungrammatical only at surface level. This predicts that if a task could be found that taps ‘underlying’, rather than surface, grammaticality, then sentences such as (5d) should pattern with other grammatical, as opposed to ungrammatical, sentences. In Duffield et al. (2002), we employed just such a task to elicit implicit grammaticality judgements. This is the Sentence-Matching (SM) paradigm — introduced by Freedman and Forster (1985), later developed for SLA research by Bley-Vroman and Masterson (1989); see also Eubank (1993), Eubank and Grace (1988). In this task, subjects are asked to determine whether or not two
Measures of competent gradience 109
visually-presented sentences are identical in form. Previous research has shown repeatedly that, in general, subjects take significantly less time to decide that matching pairs of grammatical sentences are identical than to match corresponding ungrammatical pairs of sentences. Hence, statistically discrete differences in reaction times constitute an implicit measure of grammaticality. This measure has proven useful in comparing native-speakers’ implicit grammaticality judgments with those of L2 learners, since one is — theoretically, at least — able to avoid a potential pitfall of explicit grammaticality judgment tasks, in which L2 learners may have explicitly learned a particular rule, and be able to apply it in a grammatical judgment task, but nevertheless have a radically different interlanguage competence from that of native-speakers. Contrary to some other researchers, I see little value in using the SM paradigm if the only purpose is to confirm results achievable using more traditional methodologies. Arguably, the cases of interest are those where traditional grammaticality judgment tasks and SM yield divergent results. Indeed, it seems unlikely that sentence-matching would have received much attention at all had it not been for the fact that it ‘fails’ in some interesting contexts. These contexts were the focus — arguably, the raison d’être — of Freedman and Forster’s original paper from 1985. Freedman and Forster showed that subjects who otherwise reliably distinguished in response latency between identical pairs of grammatical versus ungrammatical sentences, appeared to treat a particular subset of ungrammatical sentences — namely, specified subject condition violations — as though they were in fact grammatical. That is to say, there was no significant difference in response latency in this condition. For example, subjects treated pairs of sentences rated offline as ungrammatical, such as (6a), on a par with grammatical sentences as in (6b), and distinct from matching pairs of ungrammatical sentences, such as (6c), which elicited significantly longer response latencies. (6) a. *Who did you see Rick’s picture of? Æ implicitly treated as ‘grammatical’ b. Who did you see a picture of? c. *What you did of a picture see?
Freedman and Forster interpreted this systematic absence of an effect in terms of the level of syntactic representation tapped by the SM task: given the theoretical framework they assumed, they argued that the sentence was grammatical at s-structure, but ungrammatical at at the level of logical form. Clahsen, Hong and Sonnenstuhl-Henning (1995) reinterpret these findings
110
Nigel Duffield
in terms of operations applying at different levels of structure: SM, they argue, inspected an underlying syntactic level (LF), but was insensitive to operatorvariable binding relations holding at that level of representation. This reinterpretation captured both the Freedman and Forster results as well as their own results on verb-position in finite root versus embedded clauses in German. These latter results showed that speakers treat as grammatical main clauses in which the verb (incorrectly) occupies the final position, as in (7a), but treat as ungrammatical embedded clauses in which the verb is incorrectly raised to second position, as shown in (7c): (7) a. *Hans den Hund gesehen hat. Æ treated as ‘grammatical’ in SM. Hans the dog seen has ‘Hans has seen the dog.’ b. Hans hat den Hund gesehen. c. *daß Hans hat den Hund gesehen. Æ treated as ‘ungrammatical’ in SM. that Hans has the dog seen ‘that Hans has seen the dog.’ d. daß Hans den Hund gesehen hat.
In our own experiments on clitic placement (see Duffield et al. 2002), we determined that in SM native-speakers treat ‘ungrammatical causatives’ differently from ‘ungrammatical restructuring’ sentences. As predicted by the theoretical analysis, French native speakers treat cases such as (5d) on a par with other grammatical sentences: there was no significant difference in response latency between (5c) pairs and (5d) pairs; by contrast, it took subjects significantly longer to match (5b) pairs than their grammatical counterparts (5a). Crucially, in all other conditions involving incorrect clitic placement, French nativespeakers were significantly slower to match sentences compared to the corresponding correct placements of these clitics. Thus, it was not the case that SM overall was insensitive to constraints on clitic placement; quite the contrary. Instead, SM was selectively insensitive to surface violations of grammaticality: sentences that were ‘underlyingly grammatical’ were accepted as grammatical, even though the surface string was ungrammatical. In offline grammatical judgment tasks, on the other hand, French nativespeakers consistently treat surface ungrammatical sentences equally: offline, (5d) is considered no more grammatical than (5b). Thus, there is once again a divergence between the results obtained by implicit versus explicit methodologies. In contrast to the McKoon and MacFarland experiments discussed in the previous section, here it is the selective absence of a specific result from the
Measures of competent gradience
implicit experiment that is significant. Whichever the direction, though, both sets of experiments require a dual competence approach to underlying linguistic knowledge, if we wish to accommodate and model both online and offline results. Both sets of experiments — here again, I am considering only the native-speakers’ results in our experiment — exemplify Figures 2 and 3 above, with the implicit experiment tapping UC, and the results of the explicit grammatical judgment task reflecting surface competence. 2.2 Types of gradience Having established the basic framework of a dual competence model to accommodate both categorical and gradient effects, I would now like to consider this latter notion more closely, in order to distinguish different types of gradience. This is the issue that bears most directly on the more general concern of this volume, namely, on the nature of the lexicon-syntax interface. Just as I have suggested that there are two types of competence, it is also necessary to draw a distinction between two types of gradience. Again, for want of better terms, I will refer to these subtypes as lexical and syntactic (constructional) gradience, respectively. Since this latter distinction is somewhat more intuitive than the UC/SC contrast, a couple of illustrative examples should suffice. 2.2.1 Lexical gradience Lexical gradience refers to cases where the acceptability of a given sentence varies as a function of the properties of particular lexical items. Such properties may be semantic or idiosyncratic. An example of a semantic property might be intentionality: verbs that select +intentional subjects may be more acceptable in a given sentential context than those that do not. Idiosyncratic properties, by contrast, distinguish lexical items from near neighbours: for example, highly frequent nouns may be more acceptable than (near) synonyms of lower frequency; some items may be more appropriate than others in a given register. In both cases, the acceptability of the carrier sentence is determined by properties of the specific lexical entries. McKoon and MacFarland’s work just discussed exemplifies this type of gradience: their corpus study showed that some intransitive verbs transitivise much more readily than others, independently of the verbal class to which they belong. For instance, whereas atrophy (ECS) and deteriorate (ICS) have an extremely low probability of occurring in transitive frames, fray (ECS) and ferment (ICS) are much more likely to be transitivised. This difference is partly
111
112
Nigel Duffield
Change of Location Change of State Continuation of a pre-existing state Existence of State Uncontrolled process Controlled process (motional) Controlled process (non-motional)
selects be (least variation)
selects have (least variation)
Figure 6.Auxiliary Selection Hierarchy (adapted from Sorace 2000).
a function of inherent semantic factors, and partly one of frequency.11 Another article in the same issue of Language, by Sorace (2000), provides a different example of lexical gradience. Sorace is concerned with variation in auxiliary selection (have versus be) in constructions in the perfect in Germanic and Romance (especially Italian). The standard theoretical assumption is that, for any particular variety, the auxiliary associated with a particular predicate is rigidly lexically-determined: a given predicate either categorically selects the be auxiliary or the have auxiliary (the latter being the default value). Sorace’s article suggests that this categorical view is incorrect. Her paper shows that — both cross-linguistically and within a given variety — auxiliary selection is a gradient rather than categorical property. According to their semantic properties, verbs occupy a position on a continuum (or ‘hierarchy’, to use Sorace’s term) between be and have selection (see Figure 6). At either end of this semantically-defined continuum, there is little variation in which auxiliary is selected, so that for a non-motional controlled process, such as chat (It. chiaccherare) have is the only acceptable auxiliary, whereas for a pure change of location predicate such as come (It. venire), only be is possible. This is illustrated by the examples in (8a versus 8b) below. By contrast, predicates whose inherent semantic properties place them in the middle of this continuum show much more flexibility as to which auxiliary is selected. This flexibility is reflected both in terms of linguistic variation with respect to selection — as illustrated by examples in (9) below — and in terms of ‘coercability’ within a given variety: as the examples in (10) illustrate, verbs in the middle of the continuum can be pragmatically coerced into preferentially selecting either one or the other auxiliary.12 (8) a.
Maria è /*ha venuta alla festa. Maria is/has come to-the party ‘Maria came to the party.’ [1a]
Measures of competent gradience
b. I colleghi hanno/*sono chiaccherato tutto il pomeriggio. the colleagues have /are chatted whole the afternoon ‘The colleagues chatted the whole afternoon.’ [33a] Gli atleti svedesi hanno corso/ ?sono corsi alle Olimpiadi. the athletes Swedish have run /are run at-the Olympics ‘The Swedish athletes ran at the Olympic Games.’ [37] b. De temperatuur is /heeft 3 uur lang gestegen, maar is toen the temperature is/has 3 hours risen but is then weer gezakt. again dropped ‘The temperature rose for three hours but then dropped again.’ [11]
(9) a.
Il pilota ha / ?è atterato sulla pista di emergenza. the pilot has/is landed on-the runway of emergency ‘The pilot landed on the emergency runway.’ [44a] b. L’aereo ?ha /è atterato sulla pista di emergenza. the plane has/is landed on-the runway of emergency ‘The plane landed on the emergency runway.’ [44b]
(10) a.
If only the poles of the continuum are considered, this contrast gives the appearance of being categorical; however, once one considers the middle range, it becomes clear that auxiliary selection is a lexically gradient phenomenon.13 2.2.2 Syntactic/constructional gradience In other constructions, however, gradient effects are observed that are independent of the particular lexical items involved. That is to say, certain sentences are regularly judged to be less than perfectly acceptable without being deemed wholly unacceptable. In the theoretical literature, such sentences are typically designated as ‘marginal’, a status denoted by one or two question-marks (?/??).14 However, as noted above, since standard theoretical models have no way of representing such judgments, marginal sentences are usually ‘re-classified’ as grammatical or ungrammatical ad hoc, depending on the analysis that is being pursued. Such reclassification immediately obscures an essential feature of most acceptability judgments, namely, their syntactic gradience. One example of this type of gradience has already been mentioned, viz., the influence of referential specificity in determining the relative strength of syntactic island effects: Kluender (1992), see also Kluender and Kutas (1993). A different example of syntactic gradience is provided by my work with Ayumi Matsuo on VP-ellipsis constructions in English. Ellipsis constructions,
113
114
Nigel Duffield
and the constraints pertaining to them, have provided core data for generative analyses for several decades, their importance first brought to general attention in Sag’s dissertation (1976). The aspect of ellipsis constructions relevant to the present discussion is a constraint on structural parallelism, which — the theoretical literature claims — requires the VP of the antecedent clause to be syntactically parallel to that of the understood ellipsis. This structural parallelism constraint is used to explain the contrast between (a) versus (b) examples in (11) and (12) below: examples (11a) and (12a) show VP-ellipsis with ‘parallel’ active/verbal antecedent clauses; those in (11b) and (12b) illustrate two types of ‘non-parallel’ antecedent, passive VPs and nominal antecedents, respectively. The examples in (11/12c) and (11/12d) are intended to show that this parallelism constraint fails to apply — or, at least, does not apply so strongly — if the ellipsis (VPE) is replaced with the semantically equivalent VP-anaphora (VPA) clause. (11) a.
Someone had to take out the garbage. – But Barney refused to. b. The garbage had to be taken out. – ?/??But Barney refused to. c. Someone had to take out the garbage. – But Barney refused to do it. d. The garbage had to be taken out. – But Barney refused to do it.
(VPE)
(VPA)
(12) a.
It always annoyed Sally if anyone mentioned her sister’s name. – Tom did, out of spite. (VPE) b. The mention of her sister’s name always annoyed Sally. – ??/*Tom did, out of spite. c. It always annoyed Sally if anyone mentioned her sister’s name. – Tom did it, out of spite. (VPA) d. The mention of her sister’s name always annoyed Sally. – ?Tom did it, out of spite.
The structural parallelism effect is interesting for at least two reasons. First, for native-speakers, the parallelism constraint has generally gradient, rather than categorical effects. That is to say, native-speakers typically disprefer, but do not necessarily exclude, violations of structural parallelism with VPE (the (b) examples above). This has been demonstrated experimentally in Tanenhaus and Carlson (1990), as well as in our own work (Duffield and Matsuo 2001, 2002). The availability of ‘non-parallel ellipsis’, in contrast to some other kinds of
Measures of competent gradience
‘ungrammatical’ sentence, has also been documented in corpora of spontaneous speech, as reported in Hardt (1993). The following examples, taken from Hardt (1993), attest to the productivity of violations of the parallelism constraint. (13) a.
This information could have been released by Gorbachov, but he chose not to. (Daniel Schorr, National Public Radio broadcast 10/17/92) [Hardt (131)] b. A lot of this material can be presented in a fairly informal and accessible fashion, and often I do. (Chomsky 1982, cited in Dalrymple et al. (1991)) [Hardt (134)] c. [Many Chicago-area cab-drivers] … sense a drop in visitors to the city. Those who do, they say, are not taking cabs. (Chicago Tribune 2/6/92) [cf. Hardt ex. 118]
Hence, it seems fair to claim that such sentences have a different status from those that native-speakers quite generally reject as unacceptable. A second point to observe about non-parallel ellipsis is that constructiontype seems to be a factor in determining acceptability. Once again, experimental evidence just cited confirms the intuition that non-parallel ellipsis where the antecedent is a derived nominal (12b) is significantly less acceptable than nonparallel ellipsis where the antecedent is a passive VP (11b) (though it still remains significantly more acceptable than some other kinds of ungrammatical sentence). Standard theoretical analyses of ellipsis have no way to represent either the gradient effects of the parallelism constraint overall, or the differential effects of construction type. Hence, in the theoretical literature, the relative acceptability judgments just described get ‘recoded’ categorically, with non-parallel ellipsis being considered uniformly ungrammatical (*), irrespective of the particular type of antecedent, and non-parallel anaphora ((11d)/(12d)) deemed perfectly acceptable, native-speakers’ intuitions notwithstanding. This contrast is represented schematically in Table 3. It should be stressed that this type of gradience is orthogonal to lexical gradience: for each condition tested in our experiments, different verbs and different auxiliaries were deemed more or less acceptable in ellipsis contexts; crucially, though, all of the verbs were accepted some of the time in nonparallel contexts. The main statistical findings for native speakers were as follows. First, in both experiments, there was a reliable main effect of parallelism: in particular, VP-ellipsis following non-parallel antecedents was significantly less acceptable than following parallel antecedents, irrespective of construction type. Second,
115
116
Nigel Duffield
Table 3.Designated vs. actual acceptability judgments for parallel vs. non-parallel antecedents (VPE and VPA completions). RH column shows acceptance rates as percentages for trials in Duffield & Matsuo (2001), (2002), respectively. Antecedent-ellipsis type
Designated grammaticality judgment
Actual acceptability judgments
Active-VPE Passive-VPE
– *
– ?
90 52
88 48
Verbal-VPE Nominal-VPE
– *
– ??
89 39
93 57*
Active-VPA Passive-VPA
– –
– ?
96 91
87 84
Verbal-VPA Nominal-VPA
– –
– ?
97 74
88 76
* The relatively high acceptance rate for VPE following nominal antecedents is due to the different balance of finite and non-finite ellipsis sentences in the latter experiment. See 2.4 below for further discussion.
two clear interactions were observed: between ellipsis type and parallelism (VPE following non-parallel antecedents is reliably less acceptable than VPA in the same context), and between construction type and parallelism (VPE following passive antecedents is significantly more acceptable than following nominal antecedents, though still less acceptable than following active antecedents). Third, VPA also shows a reliable parallelism effect with nominal antecedents, albeit a smaller one compared to the VPE effect. Before discussing the second language learners, let us consider how we might model the native-speaker results, first, given the traditional competence–performance model, and then within the dual competence model proposed here. As far as I can determine, it is simply impossible to model these gradient effects in the traditional framework without arbitrarily recoding them as categorical effects. One could, for example, model the main effect of parallelism by reclassifying the circa 50% acceptance rate for VPE in passive contexts as equivalent to categorically unacceptable sentences, say, those with less than 5% acceptance ratings by native speakers. One could also gloss over the statistically reliable difference between construction types: since the principles and parameters approach allows no construction-specific rules, it cannot allow construction-specific effects to bear on grammaticality. Finally, one could dismiss the
Measures of competent gradience
small, but significant, effect of parallelism in VPA contexts. Viewed charitably, this way of treating gradient effects obscures subtle and empirically valid distinctions, while reconciling them with an explanatory model; a less charitable interpretation would regard this is as fixing the data. However it is viewed, something important is lost. I suggest that these gradient effects are an essential feature of linguistic competence, not something to be factored out. By contrast, the dual competence model allows us to model and to interpret these gradient results without abandoning the idea that some aspects of syntactic competence are indeed categorical and autonomous of lexical and constructional knowledge. As was the case for the McKoon and MacFarland results discussed above, I suggest that the results of this experiment be interpreted in terms of Figures 2 and 3 respectively, in which the structural parallelism constraint, which shows its effects across constructions, is represented in UC, whereas specific lexical and constructional information, including frequency information, is represented in SC. The interaction between these two types of competence gives rise to the various types of gradience observed. The dual competence model not only permits modelling of lexical and syntactic gradience; as discussed above, it also provides a potentially explanatory model of principled divergences between native-speakers’ and second language learners’ behaviour, and a way to resolve the apparent paradox of L2 learners outperforming native-speaker ‘controls’. In the remaining sections of this paper, I will consider some cases of what I will term parallel disjoint convergence. 2.3 Parallel disjoint convergence As noted above, the dual competence model allows non-native speakers’ results to differ from those of native-speakers on any given acceptability judgment task in two principled ways: either non-native speakers results can reflect UC while native speakers’ reflect SC, or vice versa, these options being schematised in Figures 4 and 5 above, respectively. In the following sections, I will outline one instance of each type of disjoint convergence. 2.3.1 Type 1 disjoint convergence Type 1 disjoint convergence (Figure 4 above) refers to instances discussed at the outset of this paper, in which L2 learners seem to outperform their nativespeaker counterparts; that is, instances where L2 learners’ behaviour actually
117
118
Nigel Duffield
comes closer to the categorical behaviour predicted by standard theoretical models than does that of native-speakers. This mismatch is illustrated in a recent cross-linguistic study of investigating L2 learners’ knowledge of English derivational morphology (Duffield, Sabourin and Curtin 1998). This study was a partial replication for SLA of a set of experiments with native-speakers reported in Marslen-Wilson et al. (1994). The Marslen-Wilson et al. experiments examined morphological relatedness between derivationally related words, as evidenced by priming effects. Among the interesting findings of the original study, the two most relevant were as follows. First, it was determined that words related by simple phonetic overlap did not yield priming effects (tin does not prime tinsel, nor asp, asparagus); in other words, only words that could be decomposed into a shared stem plus a legitimate affix (for the derived form) are considered by native-speakers to be related. Second, it was determined that being formally morphologically related was a necessary, but not a sufficient condition for lexical relatedness; in addition, the two forms had to be semantically related. Thus, govern was shown to prime government (and vice versa), but depart fails to prime department, since the latter pair do not share any meaning. From a theoretical point of view, the second finding is somewhat unexpected, since most theoretical morphological models assume formal rules to be autonomous of specific lexical-semantic information. Psycholinguistically, however, for native-speakers, the results show clearly that relatedness is encoded in particular lexical entries (which is the only possible locus of specific semantic information). Thus, there is a mismatch between native-speakers’ psycholinguistic representations and what the theoretical models predict. In our replication of the Marslen-Wilson et al. experiments with L2 learners, we predicted — given our theoretical assumptions — that L2 learners might diverge in a principled way from native-speakers. Specifically, while we expected that both native-speakers and L2 learners should fail to show priming for purely phonetically related (non-morphologically-related) pairs, we hypothesized that intermediate learners might initially over-generalise the formal rule, showing priming in depart-department cases, in contrast to native speakers. This was precisely what we found: whereas our native-speaker group and our advanced L2 group replicated the Marslen-Wilson et al. findings, the intermediate L2 group (native speakers of Japanese) showed priming effects for morphologically-related pairs irrespective of semantic relatedness. By demonstrating categorical, autonomous behaviour, the intermediate group better approximated the theoretical ideal than either the advanced group or the native
Measures of competent gradience
speakers: in this sense, these intermediate learners were more competent — rather, closer to underlying competence — than the others. On the other hand, they were clearly less competent than the advanced learners in converging on the judgments, and by extension, on the overall competence, of native-speakers, since in this case target competence is lexically-constrained, gradient behaviour. Although this contrast is not exactly comparable to the other phenomena discussed in this paper, since it is purely lexical, rather than syntactic, the same logic applies: some grammatical phenomena are categorical and autonomous properties, others show lexically-specific, gradient effects; both need to be accommodated. 2.3.2 Type 2 disjoint convergence The converse behaviour — where second language learners’ results reflect surface competence when native-speakers’ results show the influence of UC — can be seen in the sentence-matching experiments on clitic placement discussed earlier. Recall that the claim was that native-speakers’ failure to show a grammaticality effect (in the online task) in French causative constructions — in clear contrast to the grammaticality effect they exhibited for restructuring verbs — was due to the ‘underlying grammaticality’ of the clitic placement in sentences such as (5d) above. If this is the correct explanation of the observed asymmetry, then the predictions for the second language learners on this task are somewhat paradoxical, raising the possibility that L2 learners’ implicit acceptability judgments might fail to match those of native speakers by outperforming them with respect to the presumed theoretical target. In our experiment, this is precisely what happened: both English and Spanish speaking L2 learners of French showed a grammaticality effect for both restructuring and causative contexts in the online task, in contrast to the French native speakers. Once again, the standard model provides no satisfactory account of these results: either one is forced to exclude the causative condition altogether on the grounds that it ‘did not work’ for the native-speaker controls, or one accepts (paradoxically) that the L2 learners have achieved ‘native-speaker competence’ in this condition, as measured by approximation to the theoretical target, although their implicit judgments are wholly distinct from those of the nativespeakers. Once more, something significant is lost. By contrast, this pattern of results can be accommodated directly by the dual competence model, as schematised in Figure 5 above. The necessary assumption would be that whereas native-speakers analyse syntactic structures
119
120 Nigel Duffield
in terms of a general computational system, (at least some) L2 learners’ analyses are at the level of the surface properties of specific constructions. This assumption, though controversial, is in line with a respectable body of L2 research (see especially Clahsen and Muysken 1989, Bley-Vroman 1989, 1990). Clearly, considerably more research is necessary to demonstrate this version of the Fundamental Difference Hypothesis (the idea that SLA is constrained by fundamentally different principles and mechanisms than those that guide first language acquisition). The point here is that the present model is able to treat such divergences between native and non-native speakers in a principled fashion, without totally excluding L2 learners’ access to UG (UC in present terms). 2.4 Factoring out gradient effects: L1 versus L2 differences Before concluding, I wish to draw attention to another aspect of competence that is revealed when one studies gradient effects, but which remains obscured in the traditional paradigm. By focussing on gradience, it is possible to determine which of several logically independent variables contribute(s) to a particular acceptability judgment and — just as importantly — to determine the relative strength of these variables. One potential outcome of this type of factor analysis in SLA studies is that native speakers and second language learners converge on the same overall result for quite different reasons: in other words, their common acceptability judgments are determined by distinct ‘constraint rankings’.15 In work reported in Duffield and Matsuo (2002, in preparation), we carried out two follow-up experiments on the VP-Ellipsis study reported above. The previous experiments — in line with most other psycholinguistic work in this area — had assumed that the parallelism effect was entirely due to the syntactic properties of the antecedent clause. The follow-up experiments re-examined this assumption, the goal being to explain the parallelism effect in VP-ellipsis constructions, by teasing apart the other linguistic factors that may contribute to that effect, and (again) to compare native speakers and L2 learners’ sensitivity to such factors. For these latter experiments, we considered two properties in addition to syntactic parallelism, namely (conceptual and syntactic) recoverability, and finiteness. Recoverability refers to the idea that the parallelism effect may be partly due to the relative salience in the discourse representation of the material to be reconstructed: that is, non-parallel antecedents might be dispreferred not for structural reasons, but for interpretive ones. To test this, we manipulated
Measures of competent gradience
the antecedent clause in the active-passive experiment (see above), such that the linguistic information necessary to (re)-construct and interpret the ellipsis clause was more or less recoverable from the antecedent. Specifically, we hypothesized that (conceptual) recoverability of non-parallel passive antecedents would be enhanced by the presence of a by-phrase, as for example in (14a) versus (14b) below.16 (14) a.
Mary was busy, so the package was set by Tom. – ?He had promised that he would. (with by-phrase) b. When we got back, our driveway had been cleared of snow. – ??A neighbour told us that Tom had. (no by-phrase)
The other property manipulated was the finiteness of the ellipsis clause. Standard theoretical accounts do not distinguish between non-finite ellipsis — involving to — as in (11) above, on the one hand, and finite ellipsis — involving do, or some other auxiliary verb — as in (12) above, on the other. That is to say, the parallelism effect is generally claimed to constrain finite and non-finite ellipsis equally. Intuitively, however, the parallelism effect is considerably weaker with non-finite ellipsis. These experiments tested that intuition experimentally. Detailed results and discussion of these experiments are reported in Duffield and Matsuo (in preparation). Here, it suffices to report the main findings, which were as follows (see also Duffield and Matsuo 2002). First, contrary to standard theoretical assumptions, our experiments show that the parallelism effect in ellipsis is not uniquely due to the structural properties of the antecedent clause: other lexical and conceptual factors interact to determine the strength of the effect. Of these factors, finiteness is crucial: non-finite ellipsis displays significantly weaker parallelism effects in non-parallel contexts than finite ellipsis. Second, conceptual recoverability does have an effect on the acceptability of non-parallel antecedents, but only — for native speakers, at least — in interaction with finiteness: recoverability weakens the parallelism effect only in non-finite contexts. The comparison between native-speakers and L2 learners’ performance was also revealing. Overall, L2 learners’ performance parallels that of nativespeakers: both groups exhibit a significantly lower acceptance of ellipsis in nonparallel contexts; and, for both groups, this effect is gradient, rather than categorical (just as in the previous experiments). This clearly indicates that gradient effects can be successfully acquired in SLA. On the other hand, the constraint ranking underlying native speakers’ and L2 learners’ common overall results appear to be quite different. Whereas
121
122 Nigel Duffield
finiteness shows a robust main effect for both groups, for L2 learners conceptual recoverability has an ameliorating effect even in finite clauses (which was not the case for native speakers). This suggests that this type of conceptual information plays a larger role in determining L2 learners’ judgments than it does for native speakers, who rely more on purely formal information. Whatever the final interpretation of these results should be, it should be clear that such findings are only attainable in principle if attention is paid to the details of gradient effects: a categorical model can neither describe nor accommodate them.
3. Conclusion The purpose of this paper has been to draw attention to various types of gradient effects, both lexical and syntactic. As suggested in the title, I have argued that these effects form an essential part of our implicit grammatical competence. A revised model of competence was proposed that accommodates these effects, but which maintains the generativist assumption that certain core aspects of grammatical knowledge are still categorical, and autonomous of the lexicon. The proposed model was also shown to offer an explanation for apparent paradoxes that arise in SLA whenever second language learners outperform native-speakers: by distinguishing two types of implicit knowledge, it is possible to offer a principled account for certain systematic mismatches between nativespeakers and second language learners. Finally, this model may ultimately allow us to bridge the gap between those who argue for strong continuity in SLA and those who advocate fundamental differences. There is considerable empirical evidence for both positions; it could be that both are correct.
Notes * I am grateful to two anonymous reviewers for constructive comments and suggestions. I would also like to thank those people who commented on previous drafts of this paper, including David Birdsong, Jonathan Bobaljik, Tom Roeper and Lydia White. I am especially grateful to Jonathan Bobaljik for clarifying many misunderstandings on my part, and for offering a persuasive defense of the standard approach. Unfortunately, due to time constraints, I have not been able to integrate all suggested revisions into this paper. No-one other than myself is responsible for remaining errors and inconsistencies.
Measures of competent gradience
1. These assumptions were not always held: as noted by Levelt et al. (1977), theoretical researchers in the sixties and early seventies developed theories on degrees of ungrammaticality (Levelt et al. cite Chomsky 1964, Katz 1964, Ziff 1964, and Lakoff 1971). 2. I postpone any discussion of the status of marginal (?) sentences, since most theoretical analyses actually end up designating such sentences either as grammatical or ungrammatical, usually the former; their relative deviance — or amelioration — relative to the other sentences of their designated type being attributed to peripheral, extra-syntactic, factors. For further discussion, see Schütze (1996: especially pp. 41ff). 3. In this sense, there has been comparatively little progress from a notion of grammaticality defined in terms of extensional (infinite) sets: in this model, particular sentences are either generated by the grammar, or they are not. Yet most generative linguists would claim, following Chomsky (1986), that I-language, rather than E-language, is the proper object of study (see also Hoekstra 1990). 4. Of course, denying the competence–performance distinction is nothing new; in the past, though, its detractors often misunderstood or vastly underestimated the intricacy and complexity of grammatical knowledge. 5. The paradox arises most clearly in L2 research simply because L2 researchers are, of necessity, more acutely aware than are theoretical linguists of the metalinguistic nature of linguistic judgments, and of the methodological and analytic problems of data collection and comparison. 6. The comparison involved here may be direct, where L2 subjects’ judgments are compared with the judgments of a ‘control group’ of native-speakers, or indirect, where L2 judgments are compared with a pre-established set of judgments (perhaps taken from a theoretical paper), which native-speakers presumably would agree on; either way, L2 learners are judged competent if their judgments ‘match’ those of native-speakers in some statistically reliable way. 7. It might be objected that such a conflict only arises where native-speakers’ judgments are gradient. However, the contention here is that almost all apparently categorical judgments are in fact gradient (when properly analyzed); hence, there is a real problem here. 8. This is a possible move, provided that there is something interesting left to a judgment once the gradient properties have been factored out of the equation. Often, though, it seems as if there is nothing left, no interesting residue that UG could explain. This again echoes Culicover (1998: 48): Chomsky has argued consistently that this perspective about linguistic theory [including the notion of UG as an ‘idealized characterization of linguistic competence’: NGD] is rational and scientific, virtually indisputable. In fact, it cannot reasonably be disputed given the presumptions that: (i) a language faculty exists that contains specific syntactic knowledge; (ii) what is left after stripping away the dynamical aspects of language is something that really exists, in some sense, in the mind/brain… (emphasis mine: NGD). 9. This latter assumption may of course be incorrect. The guiding intuition here is that the introspection involved in explicit tasks is inevitably mediated (in adult native speakers) by lexical knowledge, a product of surface competence. My speculation is that direct introspection of the computational system is impossible.
123
124 Nigel Duffield
10. Since McKoon and MacFarland did not test L2 learners, this interpretation is intended to apply to native-speakers only. 11. Here, I make simplifying assumption that these are independent factors. Obviously, this is not always the case: if, for example, inherent semantic constraints restrict or reduce the occurrence of a verb in a transitive frame (see immediately below), this will affect the token frequency for that item, which in turn may further inhibit its acceptability. 12. Numbers in square brackets designate Sorace’s original example numbers. Example 9b [11] above is originally due to van Hout (1993: 7). 13. Similar remarks would seem to apply to other ‘unaccusative effects’: for example, there-insertion in English (see Levin & Rappaport Hovav 1995). 14. Crucially, the marginal status of these sentences emerges from a uni-modal pattern of acceptances: all speakers accept these sentences sometimes, and reject them on other occasions; see Avrutin and Wexler (1992), for a relevant discussion of uni-modal vs. bimodal patterns of acceptance, and their proper interpretation. 15. In response to a reviewer’s query, the expression ‘constraint ranking’ is not intended here to imply a treatment in terms of Optimality Theory necessarily. It is not obvious that standard OT models capture gradient effects any better than mainstream generative models, since ‘violable constraints’ do not yield gradient judgments (in most models, at any rate). Rather, the term is intended to refer to differences in the relative weighting of various lexical and syntactic factors that determine the judgment. How these should relate to a particular theoretical description is an independent question. 16. In the case of non-parallel nominal antecedents, we manipulated syntactic, rather than conceptual, recoverability. Here, we contrasted zero-derived versus non-zero-derived alternations (e.g., visit versus discussion), since it has been argued that the former (zeroderived nominals) are more easily reconstructable as verb-phrases in VPE contexts. No effect was found for this type of syntactic recoverability (see Duffield and Matsuo in preparation).
References Allen, J. and Seidenberg, M. S. 1999. “The emergence of grammaticality in connectionist networks”. In The emergence of language, B. Macwhinney (ed.), 115–152. Mahwah, NJ: Erlbaum. Avrutin, S. and Wexler, K. 1992. “Development of principle B in Russian: Co-indexation at LF and coreference”. Language Acquisition 2 (4): 259–306. Barlow, M. and Kemmer, S. 2000. Usage-based models of language. Stanford: Center for the Study of Language and Information. Bever, T. G. 1970. “The cognitive basis for linguistic structures”. In The development of language, J. R. Hays (ed.), 279–362. New York: John Wiley & Sons. Birdsong, D. 1989. Metalinguistic performance and interlinguistic competence. New York: Springer-Verlag.
Measures of competent gradience
Bley-Vroman, R. 1989. “The logical problem of second language learning”. In Linguistic perspectives on second language acquisition, S. Gass and J. Schachter (eds). Cambridge: Cambridge University Press. Bley-Vroman, R. 1990. “The logical problem of foreign language learning”. Linguistic Analysis 20: 3–49. Bley-Vroman, R. and Masterson, D. 1989. “Reaction time as a supplement to grammaticality judgements in the investigation of second language competence”. University of Hawai’i Working Papers in ESL 8 (2): 207–237. Chomsky, N. 1964. “Degrees of grammaticalness”. In The structure of language: Readings in the philosophy of language, J.A. Fodor and J.J. Katz (eds). Englewood Cliffs: Prentice Hall. Chomsky, N. 1986. Knowledge of language: Its nature, origin and use. New York: Praeger. Chung, S. and Mccloskey, J. 1983. “On the interpretation of certain island facts in GPSG”. Linguistic Inquiry 14: 704–713. Clahsen, H. and Muysken, P. 1989. “The UG paradox in L2 acquisition”. Second Language Research 5: 1–29. Clahsen, H., Hong, U. and Sonnenstuhl-Henning, I. 1995. “Grammatical constraints in syntactic processing: Sentence-matching experiments in German”. The Linguistic Review. Coppieters, R. 1987. “Competence differences between native and near-native speakers”. Language 63: 544–573. Culicover, P. 1998. “The minimalist impulse”. In The limits of syntax, P. W. Culicover and L. McNally (eds), 44–77. New York: Academic Press. Culicover, P. 2000. “Minimalist architectures (Review of Jackendoff 1997)”. Journal of Linguistics 35: 137–150. Dalrymple, M., Shieber, S. and Pereira, F. 1991. “Ellipsis and higher-order unification”. Linguistics and Philosophy 14 (4): 399–452. Duffield, N. and Matsuo, A. 2001. “A comparative study of ellipsis and anaphora in L2 acquisition”. In Proceedings of the 25th Boston University conference on language development, A. H.-J Do, L. Domínguez and A. Johansen (eds), 238–249. Somerville, MA: Cascadilla Press. Duffield, N. and Matsuo, A. 2002. “Finiteness and parallelism: Assessing the generality of knowledge about English ellipsis in SLA”. In Proceedings of the 26th Boston University conference on language development, B. Skarabela, S. Fish, S. and A. H.-J. Do (eds), 197–207. Somerville, MA: Cascadilla Press. Duffield, N. and Matsuo, A. in preparation. “Acquiring competent gradience: Factoring out the parallelism effect in VP-ellipsis”. ms. McGill University/University of Ottawa. Duffield, N., Sabourin, L. and Curtin, S. 1998. “UG constraints on derivational morphology in SLA”. McGill Working Papers in Linguistics: Proceedings of GASLA 1997 13 (1, 2). Duffield, N. and White, L. 1999. “Assessing L2 knowledge of Spanish clitic placement: Converging methodologies”. Second Language Research 15 (2): 133–160. Duffield, N., White, L., Bruhn De Garavito, J., Montrul, S. and Prévost, P. 2002. “Clitic placement in L2 French: Evidence from sentence matching”. Journal of Linguistics 38 (3): 1–37. Ellis, R. 1991. “Grammaticality judgments and second language acquisition”. Studies in Second Language Acquisition 132: 161–186. Eubank, L. 1993. “Sentence matching and processing in L2 development”. Second Language Research 9: 253–280.
125
126 Nigel Duffield
Eubank, L. and Grace, S. 1988. “V-to-I and inflection in non-native grammars”. In Morphology and its interface in L2 knowledge, M.-L. Beck (ed.), 69–88. Amsterdam: John Benjamins. Fodor, J. D. 2001. “Parameters and the periphery: Reflections on syntactic nuts”. Journal of Linguistics 37 (2): 367–392. Freedman, S. E. and Forster, K. I. 1985. “The psychological status of overgenerated sentences”. Cognition 19: 101–131. Greenbaum, S. 1977. Acceptability in language. The Hague: Mouton. Hardt, D. 1993. Verb phrase ellipsis: Form, meaning and processing. Computer and Information Science, University of Pennsylvania: Ph.D. dissertation. Hedgcock, J. 1993. “Well-formed vs. ill-formed strings in L2 metalingual tasks: Specifying features of grammaticality judgments”. Second Language Research 91: 1–21. Hoekstra, T. 1990. “Markedness and growth”. In Logical issues in language acquisition, I. Roca (ed.), 63–83. Dordrecht: Foris. Katz, J. J. 1964. “Semi-Sentences”. In The structure of language: Readings in the philosophy of language, J. A. Fodor and J. J. Katz (eds), 400–416. Englewood Cliffs: Prentice Hall. Kluender, R. 1992. “Deriving island constraints from principles of predication”. In Island constraints: Theory, acquisition and processing, H. Goodluck and M. Rochemont (eds), 195–222. Dordrecht & Boston: Kluwer. Kluender, R. and Kutas, M. 1993. “Subjacency as a processing phenomenon”. Language and cognitive processes 8 (4): 573–640. Lakoff, G. 1971. “Presuppositions and wellformedness”. In Semantics, D. D. Steinberg, and L. A. Jakobovitz (eds). London: Cambridge University Press. Levelt, W. J. M., Van Gent, J. A. W. M., Haans, A. F. J. and Meijers, A. J. 1977. “Grammaticality, paraphrase, imagery”. In Acceptability in language, S. Greenbaum (ed.), 87–101. The Hague: Mouton. Levin, B. and Rappaport Hovav, M. 1995. Unaccusativity: At the syntax-lexical semantics interface. Vol. 26. Linguistic Inquiry Monograph Series. Cambridge, Mass.: MIT Press. Macdonald, M.-E. C., Pearlmutter, N. J. and Seidenberg, M. A. 1994. “Syntactic ambiguity resolution as lexical ambiguity resolution”. In Perspectives on sentence processing, C. Clifton Jr., L. Frazier and K. Rayner (eds), 123–154. Hillsdale, NJ: Erlbaum. Mandell, P. B. 1999. “On the reliability of grammaticality judgment tests in second language acquisition research”. Second Language Acquisition 15 (1): 73–99. Marslen-Wilson, W., Tyler, L. K., Waksler, R. and Older, L. 1994. “Morphology and meaning in the English mental lexicon”. Psychological Review 101 (1): 3–33. Martohardjono, G. 1998. “Measuring competence in L2 acquisition: Commentary on part I”. In The generative study of second language acquisition, S. Flynn, G. Martohardjono and W. O’Neil (eds), 151–157. Mahwah, NJ: Lawrence Erlbaum Associates. McKoon, G. and MacFarland, T. 2000. “Externally and internally cause change of state verbs”. Language 76 (4): 833–858. Sag, I. 1976. Deletion and logical form. MIT: Doctoral dissertation. Schütze, C. 1996. The empirical base of linguistics. Chicago: University of Chicago Press. Sorace, A. 1993. “Incomplete and divergent representations of unaccusativity in non-native grammars of Italian”. Second Language Research 9: 22–48.
Measures of competent gradience 127
Sorace, A. 1996. “The use of acceptability judgments in second language acquisition Research”. In Handbook of language acquisition, T. Bhatia and W. Ritchie (eds). New York: Academic Press. Sorace, A. 2000. “Gradients in auxiliary selection with intransitive verbs”. Language 76 (4): 859–890. Tanenhaus, M. and Carlson, G. N. 1990. “Comprehension of deep and surface verbphrase anaphors”. Language and Cognitive Processes 5 (4): 257–280. Van Hout, A. 1993. “On unaccusativity: The relation between argument and aspect.” Paper presented at the Arbeitsgruppe Strukturelle Grammatik, MPG, Berlin. Ziff, P. 1964. “On understanding utterances”. In The structure of language: Readings in the philosophy of language, J. A. Fodor and J. J. Katz (eds). Englewood Cliffs: New Jersey.
Chapter 6
Lexical storage and retrieval in bilinguals* Ton Dijkstra NICI/University of Nijmegen
1.
Introduction
Language users, monolinguals and bilinguals alike, usually communicate in sentences. Because sentences consist of words, a complete understanding of how language users process sentences includes an understanding of how they recognize their constituent words. Although language users appear to recognize words embedded in sentences of their mother tongue almost effortlessly, the underlying word recognition process must surely be very complex. First, word identification must depend on the characteristics of the lexical item itself, for instance, on how often it has been encountered in the past (e.g. does it have a high or low frequency of usage?) and on whether it is ambiguous with respect to its syntactic category (e.g. is dance used as a noun or verb?) or semantics (e.g. does bank refer to the river side or the institution?). In addition, a word’s recognition process could be affected by the syntactic and semantic aspects of the preceding sentence context, which may be more or less constraining or predictive. For bilinguals reading in their second language (L2), the recognition of words in sentences must be even more complex, because several additional factors comes into play. At the lexical level, for instance, it is likely that the subjective frequency of the L2 words is considerably lower than that of their L1 words (due to the participants’ lower proficiency in L2), making L2 words harder to recognize. Furthermore, it may not always be clear in advance which language a presented word belongs to, because a word form may be ambiguous across languages (e.g. the word room occurs both in English and in Dutch, but in Dutch means ‘cream’) or because there may be code switches (language alternations) in the sentence (e.g. dit is een voorbeeld van bottom-up processing, meaning ‘this is an example of bottom-up processing’). In addition, the syntactic and semantic aspects of the preceding sentence are not the only possible constraints that may affect target word recognition: there is an additional factor,
130 Ton Dijkstra
namely the language of the preceding words, which might provide an independent source of constraint on bilingual word recognition. Because the recognition of words in sentence context is so complex, it is no wonder that most studies in this area during the last decades have focussed on the bilingual’s recognition of isolated words. This process already requires a distinction of different types of structural representations for words (for instance, orthographic, semantic, and phonological; a storage issue), and is inextricably bound up with how words are retrieved (a processing issue), for which purposes they are retrieved (issues with respect to task demands and instruction), and in which non-linguistic context their retrieval takes place (e.g. is the word positioned within a stimulus list containing words from the same or a different language?). The major part of this chapter is specifically concerned with the issue of isolated word recognition, and it addresses the following questions: 1. Structure: which representations are activated during bilingual word recognition? 2. Process: what is the time-course of activation for words of different languages? 3. Contextual constraints: how do non-linguistic experimental factors, such as participant expectancies and instruction, affect lexical selection in bilinguals? In the short second part of this chapter, we will discuss a few recent studies that have examined effects of linguistic context (syntactic, semantic, and lexical) on the recognition of words in sentence context. At present, only a handful of reaction time (RT) studies has been done, investigating quite diverse linguistic questions within the perspective of different theoretical frameworks. Given this sorry state of affairs, we will focus on the more coherent studies that have investigated the bilinguals’ brain activity during the processing of syntactic and semantically incorrect sentences in terms of Event-Related Potentials (ERPs). We will argue that while semantic processing may be quantitatively different between monolinguals and bilinguals, syntactic parsing may be both quantitatively and qualitatively different, in complex ways that depend, for instance, on the L2 proficiency of the bilinguals involved.
2. How bilinguals recognize words presented in isolation Before we examine recent studies on the recognition of isolated words by bilinguals, we will first consider three important general aspects of their
Lexical storage and retrieval in bilinguals
organization. First, what kind of bilinguals participated in these studies? Second, what type of stimulus materials did they involve? Third, what were the empirical paradigms the studies used for investigation? Perhaps unexpected by some readers, the term ‘bilinguals’ in many psycholinguistic studies does not specifically refer to those language users who use the words of two different languages at the same rate and with the same ease (socalled ‘balanced bilinguals’). Instead, the bilinguals participating in these studies usually are persons who use their two languages in daily life but to a different extent, implying that they are more proficient in one language (usually their native language or L1) than the other (their second language or L2; they are ‘unbalanced bilinguals’). In many of the studies we will be talking about, participants are university students, who are quite proficient in English (generally having eight or more years of experience) but have a different language (often Dutch) as their strongest, native language. These bilinguals, who have acquired their L2 relatively late (at puberty or later), will be referred to as ‘late bilinguals’, while earlier L2 acquisition makes bilinguals ‘early bilinguals’. The stimulus materials that are often used in studies on isolated words are referred to as ‘interlingual homographs’ and ‘cognates’. Interlingual homographs are words that have identical orthographic representations across languages but different semantics (such as angel, meaning ‘heavenly messenger’ in English, but ‘sting’ in Dutch), while cognates are words that overlap across languages in both their orthographic form and their meaning (e.g. film). To address how words of different languages are stored and retrieved in bilinguals, RTs to interlingual homographs or cognates are compared to matched onelanguage control items in different experimental paradigms. Any latency differences that arise between the two item types are assumed to be a consequence of the special bilingual status of interlingual homographs and cognates. Among the many experimental paradigms used in bilingual RT studies are variants of lexical decision, language decision, progressive demasking, (language) go/no-go, word naming, and word association. In a language-specific lexical decision task, participants press one button if they encounter a word in the target language, and another button if they see a nonsense letter string or non-word. For instance, in the English lexical decision task, participants press a ‘yes’ button if a presented letter string was English, and a ‘no’ button if it is not. In a generalized lexical decision task, participants press the ‘yes’ button for any word they encounter, irrespective of its language. A presented word could, for instance, be Dutch or English. In contrast, in the language decision task nonsense letter strings (non-words) do not occur. Only words are presented,
131
132
Ton Dijkstra
belonging to one of two languages. One button must be pressed when a word belongs to one language (e.g. English), and another if it belongs to the other language (e.g. Dutch). In the go/no-go task, only words from two languages are presented as well. However, participants react only when they identify a word from the target language, for instance, English (English go/no-go), but they do nothing if a word of the non-target language (for instance, Dutch) is presented. In progressive demasking, participants identify a word that is presented in an alternating sequence with a pattern mask, for instance a checkerboard or a row of hash marks. Across alternations, the pattern mask decreases in duration, while stimulus presentation increases, until the target word is recognized. In word naming, the participants must read aloud words in the target language, for instance, in their mother tongue (e.g. Dutch) or in their second language (e.g. English). Finally, in word association, participants respond to a presented word by producing the first word that comes to their mind in the target language. 2.1 Structure: Which representations are activated during bilingual word recognition? Research has shown that in the initial stages of monolingual word recognition, an input letter string leads to the activation of multiple word candidates in the mental lexicon that closely match the input (see Grainger and Dijkstra 1996). For instance, when the letter string word is presented to English monolinguals, the stored representations for words like word, cord, ward, wood, and work will initially become active (such candidates are called neighbours). When the word identification process proceeds, inappropriate candidates will gradually be reduced in activation and no longer be considered as a possible input word. Finally, only the lexical representation corresponding to the presented word remains active and becomes recognized. For bilinguals, the interesting question arises if word candidates from different languages are activated if they overlap sufficiently with the input letter string. For instance, are both the English and the Dutch readings of the interlingual homograph angel activated in parallel if Dutch-English bilinguals read the word in an English book? According to the ‘language-selective access’ view on bilingual word recognition, only lexical candidates of the task-relevant language (in this example English) are activated (Gerard and Scarborough 1989, Macnamara and Kushnir 1971). Thus, when the word word is presented, Dutch word candidates that are similar to word, like bord and wond would not become activated. In contrast, according to the ‘language non-selective access’ view, word candidates from
Lexical storage and retrieval in bilinguals
both languages (here English and Dutch) become active (Altenberg and Cairns 1983, Grainger and Beauvillain 1987, Van Heuven, Dijkstra and Grainger 1998). As we shall see later, the majority of recent studies support the language nonselective access hypothesis. However, formulating the lexical access views in this general way ignores at least two important points. First, the issue of (non)selective access should be differentiated with respect to different types of lexical representations: e.g. orthographic, phonological, and semantic representations. Second, the answer to the question might depend on whether one is processing in one’s native language (L1) or in a second language (L2). Dijkstra, Grainger and Van Heuven (1999) found evidence of crosslinguistic competition between words of different languages that are similar in form and/or meaning. They investigated whether Dutch-English bilinguals recognized interlingual homographs faster or slower than matched onelanguage control items in English lexical decision and progressive demasking tasks. The English stimulus words varied in their degree of orthographic (O), phonological (P), and semantic (S) overlap with Dutch words. Examples of items in their six test conditions are sport (overlap in S, O, and P codes), wild (SO), wheel (SP), pink (OP), angel (O), and pace (P). The first two conditions (SOP and SO conditions) consist of what are usually called ‘cognates’, while the last three conditions contain ‘interlingual homographs’ (OP and O conditions) or ‘interlingual homophones’ (P condition). Participants were faster to make a lexical decision to the target words with cross-linguistic overlap than to exclusively English control words if the overlap was orthographic and/or semantic in nature (e.g. in the SO and O conditions). In contrast, cross-linguistic phonological overlap produced inhibitory effects. Responses to test items of the P condition, for example, were slower than to matched purely English words. To show that the observed result pattern did not arise because the test and control items were not well matched in some unknown aspects, the lexical decision experiment was replicated with American-English monolinguals. For these participants, no RT or error differences were found between test items and their matched controls. Furthermore, to show that the results were not restricted to lexical decision, the experimental materials were also included in a progressive demasking task. The result pattern obtained with this paradigm was strikingly similar to that in lexical decision, indicating that it was not the task that induced the facilitation and inhibition effects for homographs relative to controls. In sum, a presented word form in L2 appears to initially activate orthographic, phonological, and semantic lexical representations in both L2 and L1.
133
134 Ton Dijkstra
The opposite effects of orthographic and phonological overlap may help to explain observed differences in the result patterns of other available studies, because the stimulus materials in these studies may have varied in terms of the degree of cross-linguistic phonological overlap and therefore in the relative amount of inhibition (e.g., Dijkstra, Van Jaarsveld and Ten Brinke 1998, Font 2001, Gerard and Scarborough 1989, Von Studnitz and Green 2002). In a follow-up experiment (Van Heuven and Dijkstra 1999), English pseudo-homophones were added to the stimulus list. Pseudo-homophones are nonsense words for which the pronunciation sounds like a real word, like brane and bloo. The reasoning behind this manipulation was that the presence of such items would discourage the use of phonology, and would therefore lead to a reduction of the earlier found phonological inhibition effect (see Davelaar, Coltheart, Besner and Jonasson 1978, for similar arguments in a monolingual study). A reduction of the phonological inhibition effect was indeed found in several conditions, but the effect did not disappear completely in conditions where cross-linguistic overlap occurred in several codes (e.g. sop). This finding suggests that phonology may be re-activated by interactions between codes. For instance, via its semantic and orthographic overlap an interlingual homograph like film might reactivate its phonology in both languages (see Gottlob, Goldinger, Stone and Van Orden 1999, for such ‘resonance effects’ in a comparable monolingual study; and Sebastián-Gallés and Kroll in press, for an overview of the role of phonology in bilingual lexical processing). Recent studies by Van Hell and Dijkstra (2002) and Font (2001) indicate that language non-selective access also occurs for cross-linguistically ambiguous target words of the native language (L1) and even when targets are not completely identical in form across languages. In the study by Van Hell and Dijkstra (2002), trilinguals with Dutch as their native language, English as their second language, and French as their third language performed a word association task or a lexical decision task in their L1 (Dutch). Stimulus words were (mostly) non-identical cognates such as tomaat or non-cognates. Shorter association and lexical decision times were observed for Dutch-English cognates than for noncognates. For trilinguals with a more equal (high) proficiency in French and English, faster responses in lexical decision were found for both Dutch-English and Dutch-French cognates. In other words, even when their orthographic and phonological overlap across languages is incomplete, cognates may be recognized faster than non-cognates. For French-Spanish bilinguals, Font (2001) has found that in lexical decision cognates differing in one letter between languages (called ‘neighbour cognates’
Lexical storage and retrieval in bilinguals
by her) are still facilitated but significantly less so than identical cognates. Furthermore, she has shown that the amount of facilitation that is observed depends on the position of the deviating letter in the word. Neighbour cognates with the different letter at the end of the word (e.g. French texte – Spanish texto) are facilitated more than neighbour cognates with the different letter inside (e.g. French usuel – Spanish usual). In fact, facilitatory effects for the latter type of cognate disappeared and effects tended towards inhibition when such cognates were of low frequency in both languages. Similar patterns of results were found in both L1 and L2 processing. These results make it likely that the size of RT effects observed for cognates and interlingual homographs depends on their degree of cross-linguistic overlap (also cf. Cristoffanini, Kirsner and Milech 1986). Note that it follows logically that across language pairs that do not share orthography at all (e.g. Chinese and English), no ‘orthographically similar’ word candidates can be activated, while effects of phonological similarity might still occur (depending on, for instance, the way tonal information affects the establishment of the set of lexical candidates). 2.2 Process: What is the time-course of activation for words from different languages? The issue of language (non)selective access can also be examined from a processing point of view by considering the time-course of lexical activation and selection in bilingual word recognition. As a first question, we may consider the rate of code activation in L1 and L2: how fast do orthographic, phonological, and semantic codes from the two languages become active? From the monolingual domain, we know that high frequency words are generally recognized faster than low frequency words, and, because the words of L2 must have a lower subjective frequency than those of L1 (simply because the former have been encountered less often), it seems likely that L2 codes become available slower than L1 codes. A comparison of the study by Dijkstra et al. (1999), discussed in the previous section, with a study by Lemhöfer and Dijkstra (submitted) provides information that supports this viewpoint. Dijkstra et al. (1999) showed that in a lexical decision task where L2 (English) was the target language of the bilingual participants, cross-linguistic effects arose for L1-L2 (Dutch-English) homographs with respect to all three codes. Because English was the target language in this task, task execution implied the verification of the English language membership of possible word
135
136 Ton Dijkstra
candidates, even when Dutch codes would be available faster than English ones. In other words, Dutch codes had time to establish themselves and exert effects on later available English targets that were necessary for responding. Lemhöfer and Dijkstra (submitted) presented the same stimulus materials to Dutch-English bilinguals in a generalized lexical decision task. In this task, participants responded with ‘yes’ to both English and Dutch words, but with ‘no’ to non-words. In contrast to English lexical decision, participants in this task can use both Dutch and English lexical representations as a reliable basis for response. Thus, in this task cross-linguistic effects will arise only to the extent that L1 and L2 codes can affect each other before the fastest codes (usually Dutch ones, we assume) are retrieved and responded to. The results of this study were quite clear: no facilitation effects arose for interlingual homographs, while cognates were facilitated relative to control words. The pattern of results for homographs indicates that responses were based upon the fastest available code, usually the Dutch orthographic code, while cross-linguistic overlap with respect to semantics in the case of cognates apparently can be used to speed up the response. In sum, even though L1 and L2 codes become active in parallel, L2 codes are often activated more slowly than L1 codes, probably because of differences in subjective frequency between languages. As a consequence, the development of cross-linguistic effects depends on the target language in the experiment (L1 or L2) and on other temporal characteristics of the task involved. There is a different way of approaching the issue of the time-course of lexical selection in bilinguals. Rather than asking how fast lexical representations from different languages become active, we might wonder how long they remain active. Even if there is an initial activation of various codes from different languages, lexical selection might be relatively fast or slow afterwards. This issue has been investigated in experimental studies by varying the frequency ratio of the two readings of interlingual homographs (e.g. angel is relatively more frequent in English than in Dutch). Dijkstra, Timmermans and Schriefers (2000) examined how long the two readings of an interlingual homograph compete for selection and whether language information provided by the item can be used to facilitate the selection of one of these readings. In three experiments, each with a different instruction, bilingual participants processed the same set of homographs embedded in identical mixed-language lists. Homographs of three types were used: high-frequent in English and low-frequent in Dutch; low-frequent in English and high-frequent in Dutch; and low-frequent in both languages. In the
Lexical storage and retrieval in bilinguals
first experiment (involving language decision), one button was pressed when an English word was presented and another button for a Dutch word. In the second and third experiments participants reacted only when they identified either an English word (English go/no-go) or a Dutch word (Dutch go/no-go), but they did not respond if a word of the non-target language (Dutch or English, respectively) was presented. In all three tasks, clear inhibition effects arose for homographs relative to one-language controls. Even in the Dutch go/no-go task for Dutch-English bilinguals performing in their native language, participants were unable to completely exclude effects from the non-target language on homograph identification. More important for the present discussion, however, is the finding that target-language homographs were often ‘overlooked’, especially if the frequency of their other-language competitor was high. The relative frequency of the two readings of the interlingual homograph was found to affect both RTs and error rates. In the Dutch go/no-go task, participants did not respond to low-frequency items belonging to their native language in about 25 percent of the cases! Inspection of cumulative distributions showed that if they did not respond after about 1500–1600 ms, they did not respond anymore within the time window of two seconds. The observed flattening of the cumulative distribution towards an asymptotic value suggests that recognition of the homograph reading from the non-target language in some way ‘prohibits’ the subsequent recognition of the target language reading (e.g. after recognition, all other lexical candidates may be suppressed). Thus, selection of one of the readings of the interlexical homographs takes place rather late during processing. The results suggest that until that time both readings of a presented homograph are involved in a (frequency-dependent) ‘race to recognition’ that is won by the fastest candidate. It is clear that the system must at some time arrive at a selection of one lexical item only, but apparently the role played by the language of that item in aiding selection is only minor. In fact, determination of the language of the item may depend on lexical selection having taken place. In addition, it does not seem possible to discard the homograph reading from the non-target language and to focus on the target reading only on the basis of the instruction that just the target language needs to be responded to. Similar results were found when the target language was English (L2) and when it was Dutch (L1), even though fewer target words were overlooked in the second case. Again, this finding points to activation from both readings of the
137
138
Ton Dijkstra
interlingual homographs irrespective of whether the target language is the native language or a second language. This study allows two important additional conclusions. First, there appear to be serious limitations on the degree of control that participants can exert on the relative activation of their two languages. Second, the selection of the target word appears to be based on item characteristics (such as word frequency) and not on the language membership of the item. Language membership appears to be available relatively late (maybe only after item identification) and therefore cannot help to speed up lexical selection. 2.3 Contextual constraints: How do non-linguistic experimental factors affect lexical selection in bilinguals? In the previous sections, we have argued that under the experimental circumstances of the presented studies involving isolated words, the word recognition system functions in a language non-selective way. However, that the system can function in a non-selective way does not imply that it does so irrespective of the experimental circumstances. In the following two sections, we will consider to which extent the observed language non-selectivity may be modulated by context. We will make a distinction between two types of contextual factors: non-linguistic or experimental and linguistic. Non-linguistic or experimental context aspects are concerned with participant expectations based on, for instance, (the explicitness of the) instruction and task demands. Linguistic context aspects have to do with lexical, syntactic, semantic, and language information, such as provided by a sentence context. We note that for lists of individual items, stimulus list composition and in particular language intermixing could have both linguistic (lexical) and non-linguistic effects. Language intermixing refers to whether an experiment contains exclusively items belonging to one language (blocked presentation) or items from two languages (mixed presentation). The effects of language intermixing and task instruction on bilingual word recognition were the focus of three DutchEnglish lexical decision experiments by Dijkstra, Van Jaarsveld and Ten Brinke (1998). In Experiment 1, Dutch bilingual participants performed an English lexical decision task including Dutch-English homographs, cognates, and purely English control words. The mean RTs to interlingual homographs were unaffected by the frequency of the Dutch reading and did not differ from those to monolingual controls. In contrast, cognates were recognized faster than controls. In Experiment 2, Dutch participants again performed an English
Lexical storage and retrieval in bilinguals 139
lexical decision task including interlingual homographs, but, apart from nonwords, Dutch words were also incorporated, requiring a ‘no’-reaction. Strong inhibition effects were now obtained for interlingual homographs relative to English control words. The size of the inhibition effect depended on the relative frequency difference of the two readings of the homograph. It was largest when the Dutch reading of the homographs had a high frequency relative to the English reading. In Experiment 3, Dijkstra et al. (1998) used the same stimulus materials but changed the task demands. Participants now performed a general lexical decision task, responding ‘yes’ if a word of either language was presented (rather than saying ‘no’ to Dutch words). In this experiment, frequencydependent facilitation effects were found for the interlingual homographs (relative to English control words). The authors argued that the null-results for interlingual homographs in the first experiment did not constitute conclusive evidence that bilingual word recognition involves a language selective access process, because in that case the different stimulus list composition of Experiment 2 should not have affected the results. Instead, the results of Experiment 2 were considered as evidence supporting the language non-selective access view, and this view was tested and supported again under the different task demands of Experiment 3. This last experiment further showed that task demands may affect the direction of the observed effects: changing the task from language specific lexical decision in Experiment 2 towards generalized lexical decision in Experiment 3 turned the inhibition effects of Experiment 2 into facilitation effects in Experiment 3. The null-effects in Experiment 1 make one wonder if the Dutch readings of the interlingual homographs were activated at all in this English lexical decision task. Recent reanalysis of the data suggests they were. A regression analysis showed that (despite the over-all null results) homograph responses became slower as the frequency of their Dutch reading increased, while they became faster with increasingly high English frequency readings. Furthermore, De Moor (1998) demonstrated that the L1 semantics of the interlingual homographs was apparently activated as well. De Moor first replicated the null-result for homographs relative to controls. Then, on the trial after the homograph appeared, she presented the English translation of its Dutch reading. For instance, brand was followed by fire, which is the English translation of the Dutch word brand. A small but reliable translation priming effect of 11 ms was found. In a replication of this experiment with different stimulus materials, Van Heste (1999) observed a reliable 35 ms difference between translation and control trials. The Dutch reading of the homograph on the previous trial had apparently been activated
140 Ton Dijkstra
even though this did not affect its RT (cf. De Bruijn, Dijkstra, Chwilla and Schriefers 2001). Finally, Dijkstra et al. (1999) performed an analogous experiment that was reviewed in Section 2, also involving an English lexical decision task with interlingual homographs and controls. In this study, significant facilitation effects were found for homographs having cross-linguistic overlap in orthography but not in phonology (stage), and no effects for items with overlap in both (step). The items in this study were comparable to those in Dijkstra et al. (1998), making it likely that the null-effects in the earlier study were due to mixing the two types of items. (Indirect support for this reasoning comes from a Spanish lexical decision study involving French-Spanish bilinguals by Font (2001: 115), who found facilitatory effects for French-Spanish homographs that had little phonological similarity across languages.) Several other accounts have been proposed for the null-results in Experiment 1 and the inhibitory effects in Experiment 2 from Dijkstra et al. (1998). These accounts have either referred to differences in the relative activation of words from the two languages in Experiments 1 and 2 (Dijkstra et al. 1998, Grosjean 2001), or to differences in participants’ decision strategies (De Groot et al. 2000, Dijkstra et al. 2000). Let us take a closer look at the various proposals. Dijkstra et al. (1998) assumed that the degree of activation of Dutch (the non-target language) was higher in Experiment 2 than in Experiment 1, because Dutch words were only included in Experiment 2. As a consequence, the English readings of the interlingual homographs suffered from more competition by the Dutch reading in Experiment 2. As an underlying mechanism, this view assumes that lexical activation effects can last across trials and can affect relative language activation. In sum, the different results in the two experiments are assumed to be a consequence of different bottom-up activation processes due to the composition of the stimulus list. This is basically an explanation in terms of lexical context effects. Similarly, Grosjean (2001) interpreted the results in terms of the participants’ ‘language mode’, referring to the relative state of activation of the bilingual’s languages and language processing mechanisms. The mode is ‘monolingual’ if only one language is relatively active and ‘bilingual’ if both languages are active (though one language may be more active than the other). In Experiment 1, the participants only read English words and non-words (although some words were homographs and cognates) and they were instructed to decide whether the items were English words or not. This would have positioned them towards the monolingual end of the mode continuum, but
Lexical storage and retrieval in bilinguals
they did not reach this position totally as they knew they were being tested as bilinguals. Thus, although their Dutch was partly active (which would explain the cognate effect) it was not sufficiently active to create a homograph effect. In sum, Grosjean (2001) proposed that both the participants’ expectancies with respect to the English lexical decision task and the degree of language intermixing (encountering mostly English words) affected the bilinguals’ performance. This explanation implies that both non-linguistic and linguistic context aspects affected relative language activation. De Groot et al. (2000) replicated the null-results observed in Experiment 1 by Dijkstra et al. (1998) using different stimulus materials and different DutchEnglish bilinguals. They proposed that the participants were instructed to perform a ‘language specific’ English lexical decision task, but on some trials may instead have treated the task as a ‘language neutral’ lexical decision task. The adoption of a ‘language specific’ processing mode would induce slower responses to homographs than to matched controls due to lexical competition between the activated target and non-target readings of the interlingual homographs (just as in Experiment 2 by Dijkstra et al. 1998). In contrast, in a ‘language neutral’ processing mode the response to a homograph would be based on the availability of any reading, irrespective of language, and homographs could then be responded to faster than controls (as in the generalized lexical decision task of Experiment 3 by Dijkstra et al. 1998). In sum, a mixture of the two processing modes adopted by the participants led to a mixture of facilitation and inhibition effects for homographs relative to controls, yielding an overall null-result. (Note that this account would predict larger standard deviations for the homographs in the condition where the Dutch reading of the homograph has a high frequency than in the conditions where it has a low frequency, because in the former type of condition, the Dutch reading would be available much faster than the English reading, while that would not be the case for the last type of words.) Dijkstra, De Bruijn, Schriefers and Ten Brinke (2000) pointed out that the participants in the three studies that reported null-results were apparently not told in advance that some of the presented letter strings would be words in both Dutch and English. Participants might sometimes have adopted a ‘language neutral’ processing mode because they were in an uncertain situation. To disentangle the effects of instruction and language intermixing, Dijkstra et al. designed an experiment that combined features of Experiments 1 and 2 by Dijkstra et al. (1998). Participants were explicitly instructed that they would encounter Dutch words requiring a ‘no’ response and were provided with
141
142 Ton Dijkstra
examples in the practice set. However, exclusively Dutch words were presented only in the second part of the experiment. No significant RT differences were found between the interlingual homographs and matched English control items in the first part of the experiment. In contrast, strong inhibitory effects for homographs relative to control words appeared in the second part. Examination of the transition from Part 1 to Part 2 showed that, as soon as Dutch items started to come in, the RTs to homographs were considerably slowed down compared to control words. These results converge completely with those of Experiments 1 and 2 by Dijkstra et al. (1998) discussed above. They suggest that language intermixing rather than instruction-based expectancies drives the bilingual participants’ performance. Instead of interpreting the pattern of results in Part 1 and Part 2 of the experiment as evidence for differences in relative language activation (depending on the local absence or presence of Dutch items in the stimulus list), Dijkstra et al. propose that participants used different decision criteria in the two parts of the experiment, depending on the types of lexical items they encountered. Ignoring the details of the proposed underlying mechanisms, we can draw a number of general conclusions on the basis of these and other studies (see Dijkstra and Van Heuven (2002), for an elaborated model of bilingual word recognition based on this evidence). First, word candidates from both target and non-target languages are activated in parallel in a ‘bottom-up’ way (via the signal), even though their rates of activation may differ depending on subjective frequency. Second, stimulus list composition and task demands are important determinants of the response patterns. Third, task demands, instruction details, and other ‘top-down’ information sources do not ‘override’ the activated bottom-up information; instead, the activated representations in the two lexicons are used for responding in accordance with the requirements of the task at hand. In all, the conclusion is that for isolated words presented in stimulus lists, bilingual word recognition is based on the input signal and is basically automatic. Non-linguistic context effects (due to the composition of the stimulus list, the specific instruction, or the task to be performed) appear to affect the decision criteria that are used to accept one lexical candidate or another during the lexical selection process, but not to affect the relative degree of activated word candidates from one or the other language.
Lexical storage and retrieval in bilinguals 143
3. How bilinguals recognize words presented in sentences In everyday life, the contextual influences on word recognition are not provided by previous words in an unrelated word list or by the demands of the experimental task that must be performed, but by the syntactic, semantic, lexical, and language aspects of the sentence context that precedes a particular word that is to be recognized. In the following section, we briefly consider such linguistic context effects on word recognition. We will first examine an RT study on the general effects of sentence context on bilingual word recognition, showing that there may be complex interactions between different aspects of the sentence context and word identification. Next, we will zoom in on syntactic and semantic aspects of sentence processing as reflected in studies that measure the bilinguals’ brain activity using Event-Related Potentials. Altarriba, Kroll, Sholl and Rayner (1996) examined semantic and lexical form effects of a preceding sentence context on bilingual word recognition in two experiments. In the first experiment, they monitored the eye movements of Spanish-English bilinguals who were reading English (L2) sentences that contained either an English (L2) or a Spanish (L1) target word (Experiment 1). Sentences provided either high or low semantic constraints on the target words. An example sentence of the high constraint and Spanish target condition is He wanted to deposit all his dinero at the credit union, where dinero is Spanish for ‘money’. The experiment led to an interaction between the frequency of the target word and degree of sentence constraint for Spanish target words with respect to the first fixation duration, but not for English target words. Thus, when the Spanish target words were of high frequency and appeared in highly constrained sentences, the participants apparently experienced interference. This result suggests that sentence constraint influences not only the generation of semantic feature restrictions for upcoming words, but also that of lexical features. The high-frequency Spanish word matched the generated set of semantic features, but not the expected lexical features when the word appeared in the alternate language (Altarriba et al. 1996: 483). The same pattern of results was found in a second experiment, where the sentences were presented word by word using the rapid serial visual presentation (RSVP) technique and participants named the capitalized target word in each sentence. The findings of this study indicate that linguistic sentence context interacts with target word recognition, suggesting that linguistic context functions in a different way than non-linguistic context. Furthermore, it is interesting to note that the data pattern showed an interaction of word frequency (a lexical
144 Ton Dijkstra
information source) and the sentence constraint, and not of language membership and the sentence constraint. This suggests that (just like for isolated words) lexical characteristics are more important than language characteristics in the determination of word recognition in sentences. Only a limited number of studies have investigated syntactic effects of sentence context on word recognition in some detail (for a full review, see Kroll and Dussias in press). Here we will briefly describe a few recent studies that used Event-Related brain Potentials (ERPs) to compare syntactic and semantic aspects of sentence processing in bilinguals (Weber-Fox and Neville 1996, Hahne 2001, Hahne and Friederici 2001). Like the authors of these studies, we will argue that there are processing differences between monolinguals and bilinguals with respect to semantic aspects that appear to be especially quantitative in nature, while the differences with respect to syntax appear to be qualitative as well. In the study by Weber-Fox and Neville (1996), five groups of ChineseEnglish bilinguals performed an acceptability judgment task for sentences in their L2, English, while their EEG was recorded. These groups of participants had learned English at different ages. Apart from normal control sentences, they read semantically anomalous sentences in English (e.g. The scientist criticized Max’s event of the theorem), sentences that contained violations of English phrase structure rules (e.g. The scientist criticized Max’s of proof the theorem), sentences that contained specificity constraint violations (e.g. What did the scientist criticize Max’s proof of?), and sentences that contained subjacency constraint violations (e.g. What was a proof of criticized by the scientist?). In terms of their brain activity, early L2 learners (those who had learned English before age 11) responded to the semantic anomalies in a very similar way as monolingual language users. The other bilingual groups differed from the monolinguals only quantitatively: an N400 effect, a marker associated with the processing of semantic anomalies, was present in their EEGs, but it was delayed in time relative to that in monolinguals. In contrast, several qualitative differences between proficiency groups were found with respect to the syntactic processing of phrase structure violations. First, none of the bilingual groups displayed a so-called early left anterior negativity (N125) in the EEG that was present in monolinguals. The N125 is an early effect in the EEG that may reflect automatized first-pass parsing processes. Second, a second left lateralized negativity (N300–500) was found in all groups, which was left lateralized (found in the left hemisphere of the brain) in monolinguals and early bilinguals, but more bilaterally distributed in late bilinguals.
Lexical storage and retrieval in bilinguals
Third, a P600 effect was present in the monolinguals and early bilinguals but not in the late learners. The P600 effect is considered to be the most important EEG marker associated with syntactic reanalysis and repair. In sum, late L2 learners consistently displayed large differences in ERPs patterns relative to monolinguals, suggesting that (especially late) syntactic processes are different in late L2 learners. Hahne (2001) came to similar conclusions on the basis of an auditory sentence processing study involving proficient late Russian-German bilinguals and German monolinguals. Her participants listened to German sentences that were either correct (e.g. Die Tür wurde geschlossen, ‘The door was being closed’), contained a semantically incorrect item (selection restriction violation: Die Ozean wurde geschlossen, ‘The ocean was being closed’), or a syntactically correct item (word category violation: Das Geschäft wurde am geschlossen, ‘The shop was being on closed’). As before, ERP differences in processing semantic incongruities between native and L2 speakers were only quantitative in nature, while there were qualitative differences with respect to syntactic processing between the two participant groups. This suggests that the second language learners did not process syntactic information in the way that native listeners did. Hahne and Friederici (2001) examined sentence comprehension in Japanese speakers who had learned German as a second language after puberty. These bilinguals listened to German sentences that were correct or contained semantic and/or syntactic violations. A variety of differences was found in the ERPs for the Japanese-German bilinguals and German monolinguals. Semantically incorrect sentences induced an ERP pattern similar for the two groups (an N400 effect), while correct sentences led to a different pattern (greater positivity) in L2 learners than in native listeners. The latter finding may reflect the greater difficulties the learners had with respect to syntactic integration. For sentences containing a phrase structure violation, L2 learners, in contrast to native listeners, did not show significant modulations of the syntax-related ERP components mentioned above (the early anterior negativity and the P600). Furthermore, sentences containing a pure semantic or combined syntactic/ semantic violation elicited effects not found in native listeners. These effects may reflect additional conceptual-semantic processing in late bilinguals. These ERP studies indicate that future RT studies examining sentence processing in bilinguals are likely to yield evidence for complex interactions between lexical and syntactic knowledge in L1 and L2. Of course, there is a vast number of research issues to be addressed. For a psycholinguist working on the bilingual lexicon, one interesting issue to explore is to which extent the language
145
146 Ton Dijkstra
non-selective access mechanism found at the lexical level also holds at the syntactic level. The assumption that the syntactic rules and syntactic categories of different languages are not incorporated in language-specific databases but in an integrated store leads to a variety of predictions. For instance, for a DutchEnglish bilingual processing differences could arise for a noun phrase like ‘the light of a distant star’ and a noun phrase like ‘the man sat in his room’, because the interlingual homograph star is an adjective in Dutch (meaning ‘rigid’) but a noun in English, while the homograph room is a noun in both languages (which means ‘cream’ in Dutch). Furthermore, language non-selective access of syntactic rules might lead to specific cross-linguistic priming effects. For instance, hearing a sentence like ‘the librarian handed the reader a book’ might prime the production of de vader gaf het meisje een appel (‘The father gave the girl an apple’) but not de vader gaf een appel aan het meisje (‘The father gave an apple to the girl’) (cf. Bock 1986). Another interesting issue is whether there is a separate effect of the language of the sentence context on the recognition of a target word. In other words, could a noun phrase or larger sentence context elicit some kind of language frame that affects the processing of later arriving words? In that case, processing a sentence with a code-switch like ‘I see a huis’ might be more difficult than processing a regular sentence like ‘I see a house’, simply because the words in the first sentence context do not all belong to one and the same language. Finally, it seems likely that we may be expecting some unexpected results in future RT studies on bilingual syntactic processing, due to the complex interactions between lexical, syntactic, and semantic factors. One possibility is that quantitative differences in working memory capacity for L2 syntactic processing may lead to qualitative processing consequences between L2 monolinguals and bilinguals (cf. Michael, Dijkstra and Kroll 2002). For instance, in a pilot study in our lab we found that although both Dutch and German readers tended to resolve local ambiguities in subject- and object-relative clauses in their L1 by using syntactic information only, Dutch-German readers in their L2 used semantic information as well (Caelen 1998, but also see Frenck-Mestre and Pynte 1997). A similar pattern was found in L1 readers under higher processing load conditions. All these exciting questions and many others are amenable to empirical research by means of existing research techniques. Unfortunately, the collection of empirical data addressing these questions has only just started and we have no answers to these questions yet.
Lexical storage and retrieval in bilinguals 147
4. General conclusions In this chapter, we have considered a number of questions about bilingual lexical processing and provided answers based on the presently available empirical evidence with respect to interlingual homographs and cognates. First, we have argued that during the recognition of isolated words by bilinguals, lexical candidates from several languages are activated in parallel. Such parallel activation does not only hold for orthographic representations, but also for phonological and semantic codes. Moreover, there is evidence that language non-selective access occurs even when bilinguals are processing words in their native language and are not aware that their second language knowledge is important. These findings indicate that the bilingual word identification system, just like the monolingual system, is to a large extent ‘automatic’ in nature, in the sense that lexical candidates from both languages are activated in a fast recognition process that in itself is largely unaffected by intentional and attentional factors. Second, we have examined the time-course of lexical activation with respect to L1 and L2 and found that L2 is slower to be activated than L1, depending on relative L1/L2 proficiency and therefore on (subjective) L1/L2 word frequency. For interlingual homographs, we have found that in spite of differences in L1/L2 activation rates, both readings of interlingual homographs remain active during lexical processing for a relatively long time. This finding has several important theoretical consequences. For instance, if the language membership of word candidates could be used quickly to suppress lexical candidates that are irrelevant in the experimental context, effects of the non-targeted reading of the homographs should quickly disappear. However, if language membership information becomes available late during processing, both readings of the homograph would remain active for quite long. The available empirical studies support the latter position, which is in correspondence with the automatic nature of bilingual word recognition. Third, we have demonstrated that both non-linguistic experimental and linguistic context factors may affect the result patterns that are observed in experiments. It appears that non-linguistic factors such as task demands and instruction affect the performance of bilingual participants at the level of task and decision processes as well as participant strategies. Linguistic factors such as lexical, syntactic, and semantic aspects of the sentence context appear to affect the word identification process more directly. Evidence from ERP studies indicates that syntactic processing in bilinguals may differ both quantitatively
148 Ton Dijkstra
and qualitatively from that in monolinguals, but RT studies are badly needed in order to specify from which underlying mechanisms the differences originate. To conclude, empirical studies on bilingual word recognition in the last decade have uncovered a number of fundamental characteristics of the bilingual word recognition system. They have answered some major questions that are unique to the bilingual domain, such as that about language selective or nonselective access, as well as more generally important questions, such as how language users handle lexical ambiguity and how task and stimulus context affect word recognition. The conclusions of these studies will have to be taken into account during the development of a more general model of bilingual processing. However, much more empirical evidence on the interaction between lexical, syntactic, and semantic processing is needed before we can even attempt to build such a model.
Note * The author thanks Folkert Kuiken and two anonymous reviewers for their comments on a previous version of this paper. The author also thanks Judy Kroll for her continuous support and the many discussions that shaped the ideas in this chapter.
References Altarriba, J., Kroll, J. F., Sholl, A. and Rayner, K. 1996. “The influence of lexical and conceptual constraints on reading mixed-language sentences: Evidence from eye fixations and naming times”. Memory & Cognition 24: 477–492. Altenberg, E. P. and Cairns, H. S. 1983. “The effects of phonotactic constraints on lexical processing in bilingual and monolingual studies”. Journal of Verbal Learning and Verbal Behavior 22: 174–188. Bock, J. K. 1986. “Syntactic persistence in language production”. Cognitive Psychology 18: 355–387. Caelen, M. 1998. Extending the study on the processing of relative clauses to bilingualism. Unpublished Master’s Thesis, University of Nijmegen. Cristoffanini, P., Kirsner, K. and Milech, D. 1986. “Bilingual lexical representation: The status of Spanish-English cognates”. Quarterly Journal of Experimental Psychology 38A: 367–393. Davelaar, E., Coltheart, M., Besner, D. and Jonasson, J. T. 1978. “Phonological recoding and lexical access”. Memory & Cognition 6: 391–402. De Bruijn, E., Dijkstra, A., Chwilla, D. and Schriefers, H. 2001. “Language context effects on interlingual homograph recognition: Evidence from event-related potentials and response times in semantic priming”. Bilingualism: Language and Cognition 4: 155–168.
Lexical storage and retrieval in bilinguals 149
De Groot, A. M. B., Delmaar, P. and Lupker, S. J. 2000. “The processing of interlexical homographs in a bilingual and a monolingual task: Support for nonselective access to bilingual memory”. Quarterly Journal of Experimental Psychology 53: 397–428. De Moor, W. 1998. Visuele woordherkenning bij tweetalige personen. [Visual word recognition in bilinguals.] Unpublished Master Thesis, University of Ghent. Dijkstra, A., De Bruijn, E., Schriefers, H. J. and Ten Brinke, S. 2000. “More on interlingual homograph recognition: Language intermixing versus explicitness of instruction”. Bilingualism: Language and Cognition 3: 69–78. Dijkstra, A., Grainger, J. and Van Heuven, W. J. B. 1999. “Recognition of cognates and interlingual homographs: The neglected role of phonology”. Journal of Memory and Language 41: 496–518. Dijkstra, A., Timmermans, M. and Schriefers, H. 2000. “Cross-language effects on bilingual homograph recognition”. Journal of Memory and Language 42: 445–464. Dijkstra, A. and Van Heuven, W. J. B. 2002. “The architecture of the bilingual word recognition system: From identification to decision”. Bilingualism: Language and Cognition 5: 175–197. Dijkstra, A., Van Jaarsveld, H. and Ten Brinke, S. 1998. “Interlingual homograph recognition: Effects of task demands and language intermixing”. Bilingualism: Language and Cognition 1: 51–66. Font, N. 2001. Rôle de la langue dans l’accès au lexique chez les bilingues: Influence de la proximité orthographique et sémantique interlangue sur la reconnaissance visuelle de mots. Unpublished Doctoral Thesis of the Université Paul Valery, Montpellier, France. Frenck-Mestre, C. and Pynte, J. 1997. “Syntactic ambiguity resolution while reading in second and native languages”. Quarterly Journal of Experimental Psychology 50: 119–148. Gerard, L. D. and Scarborough, D. L. 1989. “Language-specific lexical access of homographs by bilinguals”. Journal of Experimental Psychology: Learning, Memory and Cognition 15: 305–313. Gottlob, L. R., Goldinger, S. D., Stone, G. O. and Van Orden, G. C. 1999. “Reading homographs: Orthographic, phonologic, and semantic dynamics”. Journal of Experimental Psychology: Human Perception and Performance 25: 561–574. Grainger, J. and Beauvillain, C. 1987. “Language blocking and lexical access in bilinguals”. Quarterly Journal of Experimental Psychology 39A: 295–319. Grainger, J. and Dijkstra, A. 1996. “Visual word recognition”. In Computational Psycholinguistics: AI and connectionist models of human language processing, A. Dijkstra and K. De Smedt (eds), 139–165. London: Taylor and Francis. Grosjean, F. 2001. “The bilingual’s language modes”. In Language processing in the bilingual, J. L. Nicol and T. D. Langendoen (eds), 1–25. Oxford: Blackwell. Hahne, A. 2001. “What’s different in second-language processing? Evidence from eventrelated brain potentials”. Journal of Psycholinguistic Research 30: 251–266. Hahne, A. and Friederici, A. 2001. “Processing a second language: late learners’ comprehension mechanisms as revealed by event-related brain potentials”. Bilingualism: Language and Cognition 4: 123–141. Kroll, J. F. and Dussias, P. E. In press. “The comprehension of words and sentences in two languages”. Chapter to appear in Handbook of bilingualism, T. Bhatia and W. Ritchie (eds). Cambridge, MA: Blackwell Publishers.
150 Ton Dijkstra
Lemhöfer, K. and Dijkstra, A. Submitted. “Recognizing cognates and interlingual homographs: Time course and code similarity effects in generalized lexical decision”. Macnamara, J. and Kushnir, S. L. 1971. “Linguistic independence of bilinguals: The input switch”. Journal of Verbal Learning and Verbal Behavior 10: 480–487. Michael, E., Dijkstra, A. and Kroll, J. F. 2002. “Individual differences in the degree of language nonselectivity in fluent bilinguals”. Paper presented at the meeting of the International Linguistic Association, Toronto, Canada. Sebastián-Gallés, N. and Kroll, J. F. In press. “Phonology in bilingual language processing: Acquisition, perception, and production”. In Phonetics and phonology in language comprehension and production: Differences and similarities, N. Schiller and A. Meyer (eds). Berlin: Mouton de Gruyter. Van Hell, J. and Dijkstra, A. 2002. “Foreign language knowledge can influence native language performance in exclusively native contexts”. Psychonomic Bulletin and Review 9: 780–789. Van Heste, T. 1999. Visuele woordherkenning bij tweetaligen. [Visual word recognition in bilinguals.] Unpublished Master Thesis, University of Leuven. Van Heuven, W. J. B., Dijkstra, A. and Grainger, J. 1998. “Orthographic neighborhood effects in bilingual word recognition”. Journal of Memory and Language, 39: 458–483. Van Heuven, W. J. B. and Dijkstra, A. April 1999. The role of phonology in the recognition of interlingual homographs and cognates. Paper presented at the Second International Symposium on Bilingualism, Newcastle, UK. Von Studnitz, R. E. and Green, D. 2002. “Interlingual homograph interference in GermanEnglish bilinguals: Its modulation and locus of control”. Bilingualism: Language and Cognition 5: 1–23. Weber-Fox, C. M. and Neville, H. J. 1996. “Maturational constraints on functional specializations for language processing: ERP and behavioral evidence in bilingual speakers”. Journal of Cognitive Neuroscience 8: 231–256.
Chapter 7
Inducing abstract linguistic representations Human and connectionist learning of noun classes John N. Williams University of Cambridge
1.
Introduction
Noun class information is a crucial component of the interface between the lexicon and the grammar. In order to explain linguistic productivity it is necessary to assume that linguistic rules are defined not over specific words, but classes of word. This is not only true given the classical distinction between lexicon and grammar, but also in ‘emergentist’ views which see no clear separation between these two systems (Ellis 1998, Tomasello 2000). Even though the latter stress the lexical-specificity of many ‘grammatical rules’, it is still recognised that adult productivity can only be explained if words are grouped into classes, even if those classes do not map neatly onto traditional linguistic categories. The way in which words are grouped into grammatical classes is therefore an important issue in understanding language development, particularly in explaining the leap from lexical learning to grammar learning. Noun classes, such as grammatical gender, are fundamentally abstract, grammatical notions (Corbett 1991). However, attempts have been made to uncover subtle phonological and semantic cues that can be used to predict a word’s gender (Kelly 1992). For example, masculine nouns in German are more likely to be monosyllabic, and monosyllabic words that are masculine contain more consonants than those of other classes. In French, feminine nouns tend to end in closed stressed syllables (e.g. personne, tomate, viande), and masculine nouns tend to end in open stressed syllables (e.g. avion, bruit, chapeau, bain). There are also a number of characteristic derivational morphemes associated with each gender (e.g. -eur and -ment are masculine, and -tion, -euse, -ière are feminine). Sokolik and Smith (1992) trained a connectionist network to classify French nouns as either masculine or feminine. The network was presented with the orthographic, rather than phonological, forms of the words. They found
152
John N. Williams
that it could then indicate the gender for nouns that it had not received during training, although its performance was not perfect (ranging between 73% and 75%). This indicates that there are regularities in the form (in this case spelling) of French words which can to a certain extent predict gender category. Yet there are always words which fall stubbornly outside such generalisations. In the case of French, Carroll (2001) argues that in any case, the kinds of phonological cues that have been appealed to are more subtle than could reasonably be expected to be represented in the lexicon. This is not to say that phonological and semantic cues do not play a role in learning gender systems, or that they do not affect how easy it is to remember the gender of specific words. But ultimately gender classes impose an abstract categorisation on words which is independent of their phonological and semantic properties. Learning gender systems, then, requires the formation of abstract grammatical categories, and producing grammatically well-formed utterances involves applying agreement rules which make reference to those categories. There is a growing body of evidence which suggests that even quite advanced second language learners continue to make gender errors (Hawkins 2001, Holmes and De la Batie 1999). In contrast, such errors are relatively rare in first language acquisition (Caselli, Leonard, Volterra and Campagnoli 1993). There is also evidence for qualitative differences between first and second language acquisition and processing of gender. A number of studies have shown that second language learners are more sensitive to phonological agreement patterns that correlate with gender classes than either children or adults in their native language. For example, for the Italian il pettine (‘the comb’, masculine singular) a second language learner might produce *le pettine, using the article which is more often associated with the -e ending on feminine plural nouns (Holmes and De la Batie 1999). In contrast a child would be more likely to produce *il pettino, choosing an article that is correct for the noun’s gender and number, and providing the noun with the characteristic -o ending for masculine singulars. This demonstrates a grasp of the noun’s abstract gender as the controlling influence in determiner selection (Caselli et al. 1993). In reaction time tasks on adults, Taraban and Kempe (1999) showed that non-native speakers of Russian are more sensitive to phonological cues to gender than are natives. Finally, a study by Guillelmon and Grosjean (2001) showed that whereas native speakers of French and early bilinguals show certain gender congruency effects in reaction time tasks, such effects are absent in late bilinguals. These studies suggest that second language learners do not achieve native-like representation or processing of gender information.
Inducing abstract linguistic representations
In this chapter I shall explore the possibility that the reason why gender is a persistent problem for second language learners is precisely because the underlying abstract grammatical concepts are difficult to acquire through associative learning. I shall address this issue through behavioural studies of semi-artificial language learning in tandem with computational (connectionist) simulations. These simulations were used as a means of assessing the viability, and potential limitations, of a purely associative learning account of the behavioural data.
2. The issue of abstraction in human and connectionist learning Noun class induction provides a well-constrained domain in which to examine the broader issue of abstraction in both human and connectionist learning. In the case of adult implicit learning there has been a good deal of debate over whether the knowledge that is acquired in, say artificial grammar learning experiments can really be characterised as abstract (compare Johnstone and Shanks 1999, Knowlton and Squire 1996, Meulemans and Van der Linden 1997). Some degree of abstraction is suggested by the ability to transfer rule knowledge between stimulus sets (Knowlton and Squire 1996, Mathews et al. 1989). But this appears to be no more than knowledge of patterns of alternation or doubling of stimuli, for example the common abstract ABA structure which underlies the syllable sequences ga-ti-ga and wo-fe-wo (Marcus, Vijayan, Bandi Rao and Vishton 1999). Gómez and Gerken (2000) refer to this as “patternbased abstraction”. But language structure depends upon patterns that are defined over abstract categories, such as the common NVN structure underlying ‘Dogs eat pizza’ and ‘John loves books’. Gómez and Gerken, (2000) refer to this as “category-based abstraction”. Very little implicit learning research has examined this kind of abstraction, even though it is a prime area in which implicit learning of language structure can be evaluated. In connectionist networks rule-like behaviour, such as the ability to generalise to novel inputs, is an emergent property of the system, and there is no separation between rote memory for examples and the representation of underlying generalisations (consider, for example, the well-known models of past tense formation (Rumelhart and McClelland 1986), and reading (Seidenberg and McClelland 1989)). But it has been argued that the human generative capacity in linguistic domains can not be accounted for without the ‘classical’ distinction between knowledge of instances and knowledge of rules, or the
153
154
John N. Williams
traditional computational distinction between data and symbolic programs (Fodor and Pylyshyn 1988). According to this view, the problem with connectionist models is that they respond to novel inputs purely on the basis of their similarity to trained examples, and not by applying abstract rules (Berent, Marcus, Shimron and Gafos 2002, Marcus 1999, Marcus et al. 1999). Categorybased abstraction provides an ideal arena in which to explore this issue.
3. Previous research into human and connectionist learning of word classes In his work on sequence learning Elman (1990) showed that there is a sense in which a connectionist network can learn abstract noun classes. This network learned the sequential probabilities of words in simple sentences through a prediction task (attempting to predict the next word in a sentence on the basis of the preceding ones). When the internal states of the network were examined (see below for an illustration of how this is done) it was found that the activation patterns produced by words clustered into classes that reflected the distributional properties of the training sentences. The two largest clusters were for nouns versus verbs, and within these groups there were smaller sub-clusters corresponding to transitivity preference for verbs, and animacy for nouns. These clusters were based purely on a distributional analysis of the words in the input. For example, what made a noun ‘inanimate’ was nothing more than the fact that it only occurred before certain kinds of verb (e.g. move, break) and not others (e.g. smell, see). This work is widely cited as proof that networks can induce word classes by performing distributional analysis, and as support for a statistical approach to language learning (Redington and Chater 1998). Given the apparent power of distributional information to deliver noun class information it is perhaps surprising that there is only limited evidence from experimental studies that humans are able to exploit it in order to learn noun classes. Saffran (2001) examined incidental learning of a set of hierarchical phrase structure rules in which each phrase was associated with a distinct class of nonsense words. She argued that the results of the grammaticality judgment tests showed that the participants developed sensitivity to phrase structure and word class, and that this was based on a statistical analysis of the distribution of the words in the input. However, abstract representations of word class would permit test items containing word sequences that had never occurred in the input to be judged as grammatical (or more grammatical than similar
Inducing abstract linguistic representations
sequences which violated phrase structure). Because no such test was performed it is difficult to know whether abstract word classes had really been learned. More stringent tests of word class learning become possible when noun classes, such as gender systems, are considered. Brooks, Braine, Catalano and Brody (1993) used an artificial language in which there were two noun classes, and each class used different affixes to mark the location of the actor in relation to the object denoted by the noun. Neither the form nor meaning of the nouns provided any clue to their class. Adults were first taught the vocabulary, and then performed both comprehension and production tasks (e.g. acting out phrases, or describing pictures with feedback in the form of the correct answer). After training they were tested on knowledge of the trained items, and also on their ability to produce the correct response for noun-affix combinations that had not been presented during training. Whilst their performance on trained items was at around 75%, they were at chance on the generalisation items. Not one of the 16 subjects showed evidence of having learned the system. Similar results have been obtained in a number of other studies (Braine 1987, Braine et al. 1990, Frigo and McDonald 1998). Frigo and McDonald (1998: 237) argue that models of noun class learning that depend on pure distributional analysis (Anderson 1983, Maratsos and Chalkley 1980, Pinker 1984) are “too powerful”. The question is, then, does connectionism fall into this class of overly powerful learning mechanisms for learning noun classes? The experiments and simulations presented below further explored the circumstances under which arbitrary and non-arbitrary noun class systems can be learned by humans and connectionist networks. 4. Experiment 1 Williams and Lovatt (2003) tested whether humans can learn the arbitrary noun class system shown in Table 1. There were eight nouns divided into two arbitrary classes ‘masculine’ and ‘feminine’. Words in the ‘masculine’ class occurred with the determiners ig, i, ul, and tei. Words in the ‘feminine’ class occurred with the determiners ga, ge, ula, and tegge.1 The training items were the nonitalicised phrases shown in Table 1. The italicised items were withheld for testing generalisation. It would only be possible to know that ‘the ball’ should be translated as ig johombe by knowing that johombe belongs to the ‘masculine’ class. Neither its form, its -e ending, nor its meaning provide any clues. The participants first learned the nouns and determiners as isolated vocabulary items. They then received the determiner-noun combinations for
155
156 John N. Williams
Table 1.The items employed in Experiments 1 and 2. Items used for testing generalisation are in italics. definite singular (the)
each training item as part of an exercise in rote memorisation that cycled through phases of presentation and cued recall over sets of four items. Four phrases were presented with their English translations, for example: ‘the nose’ ga nawase, ‘the birds’ i migeni, ‘some balls’ tei johombi, ‘a kiss’ ula tisseke. The participants repeated each novel phrase immediately after they had seen and heard it. After the four phrases had been presented participants attempted to recall each phrase given the English translation and stem as cues, for example: ‘the birds’ _ migen_, ‘the nose’ _ nawas_, ‘a kiss’ _ tissek_, ‘some balls’ _ johomb_. They were provided with feedback after each recall attempt in the form of the correct answer. After receiving the 24 training items they performed a generalisation test on the withheld items in Table 1. The generalisation test was similar to the recall component of the training phase. The English translation of each phrase was presented (e.g. ‘the ball’), along with the form of the corresponding stem (johomb_), and the participants had to produce the appropriate determiner and appropriately inflected noun. No feedback was given. This sequence of memory and generalisation tasks was repeated five times. Across 21 participants the mean generalisation performance over the five cycles was 36%, 48%, 54%, 66%, and 67%. A repeated measures Anova showed that the improvement in performance was significant, F(4,80) = 13.11, p < 0.001. This shows that the participants learned something of the underlying noun class organisation. However, there were large individual differences in the level of learning. Two factors were found to independently predict performance on the final generalisation test. The first was the participants’ phonological short-term
Inducing abstract linguistic representations
memory, as measured prior to the experiment by their ability to recall lists of three nonsense words (the singular forms of the nouns in the target language) in the order of presentation. The correlation between this memory measure and performance on the final generalisation test was r = 0.528, p < 0.05. There was evidence that the relationship between phonological short-term memory and rule learning was mediated by memory for determiner-noun combinations received during training. Clearly memory ability is crucial to performing the kind of distributional analysis upon which learning of this kind of system depends. The second factor was a measure of the participants’ breadth and depth of knowledge of other gender languages. All of the participants’ L1s were non-gender languages (in fact all but one of them was a native speaker of English), but the more gender languages they knew as L2s, and the better they knew them, then the better their performance on the generalisation test (r = 0.520, p < 0.05). This suggests that the learning process was facilitated by linguistic knowledge. There are a number of possible reasons why our participants managed to learn an arbitrary noun class system whereas those in the previous studies did not. First, the systems used by Brooks et al. (1993) and Braine et al. (1990) involved agreement between spatial prepositions and nouns, and Frigo and McDonald (1998) used a system involving agreement between greetings and names. Participants may have had relatively little familiarity with similar systems in other languages that they knew. Second, it is possible that the size of the languages is important. Braine et al. (1990) used a 24-word vocabulary, Brooks et al. (1993) used 30 words, and Frigo and McDonald (1998) used 20 words, whereas Experiment 1 used only eight words. Clearly, keeping track of the collocates of 20 to 30 words is much harder than keeping track of the collocates of eight words. A third potentially important factor is that in the present case some of the determiners in each class had the same ending. The feminine class contained the pairs ga-ula and ge-tegge; the masculine class contained i-tei, and the remaining determiners ig and -ul were the only ones to end in consonants. This similarity structure may have facilitated the learning process. Experiment 1 demonstrates that an arbitrary noun class system is in principle learnable. The question now is whether a connectionist simulation of the same learning problem will be similarly successful. 4.1 Simulation 1 For this and all other simulations reported here, the simulation package Tlearn was used (Plunkett and Elman 1997). The aim in the first simulation was to
157
158
John N. Williams
train the network in a way which resembled as closely as possible the training task performed by the participants in Experiment 1. I decided to focus on the recall component of the training task. The network was taught to produce the correct determiner for each phrase in the training set shown in Table 1. The input consisted of representations of the noun stem, the inflection, the English determiner, and the number of the noun. For example, the input for the item tei johombi was ‘johomb’, ‘-i’, ‘some’, and ‘plural’. This is the information that is relevant to predicting the determiner, and which was explicitly provided to the participants in the recall component of the training task in Experiment 1.2 Following Elman (1990) one unique input node was used to represent each element of the input (for example one unit was used to represent johomb), yielding a total of 15 input nodes (eight stems, two inflections, three English determiners, singular, plural). The input nodes were connected to five hidden units, which were in turn connected to eight output units, one for each of the eight possible determiners. For each input pattern the network was taught to produce the correct determiner. For example given the input johomb, -i, ‘some’, and ‘plural’ it was taught to predict tei. This involved comparing the actual output from the network with the correct output, and making appropriate changes to the connection weights within the network according to the degree of error. In this sense the network was provided feedback in the same way as the participants in the experiment. The network was initially trained until the root mean square (RMS) error for the training items was 0.1 (this required an average of 2,479 cycles through the training set).3 An error of this magnitude indicated that for each input pattern the network was able to activate the correct determiner on the output layer to a value close to the target value of 1.0, and all other output units had values close to zero. Testing involved presenting the input patterns for the generalisation items in Table 1 (i.e. the network was presented with patterns that it had not received during training). For each input pattern, the activation level of the output units was recorded, compared to the correct answer and the degree of error calculated. The training and test procedure was repeated 20 times, and on each run the connection weights were given random starting values. Generalisation performance on each run was perfect in the sense that the activation on the node for the correct determiner was far greater than that of the others. Over 20 runs the mean RMS error was 0.118 (which is not much greater than that for trained items). That is, the network was able to correctly predict the determiner for input patterns that it had never encountered during training with an accuracy which was almost as high as for the trained items.4
Inducing abstract linguistic representations 159
In order to explore the nature of the network’s internal representations the word stems were presented alone to the input layer at test (i.e. all of the other elements of the input were given values of zero). The activation patterns over the five hidden units were recorded and submitted to a cluster analysis (for a similar procedure see Elman 1990). The logic of only presenting the word stems was that the aim was to ascertain the similarity structure of the hidden unit activations to the nouns in a way that was not contaminated by the activations produced in specific contexts of definiteness and number. Over six separate runs a similar result was obtained — the activation patterns clustered according to gender. That is, nouns within the same class produced activation patterns that were similar to each other and distinct from the patterns produced by the nouns in the other class. It should be clear that this network is not simply producing responses to the generalisation items on the basis of their similarity to trained items. For example, for the test item ig johombe the stimuli were johomb, -e, ‘the’, and ‘singular’. During training johomb and -e only occurred with ul. The elements the and ‘singular’ occurred with both ig and ga with equal frequency. Yet the network was able to produce a strong output on ig and much lower levels of activation on the remaining determiners. Simulation 1 therefore shows that a connectionist network can achieve linguistic productivity, and can behave as if it has formed abstract representations, even though there are no abstract representations as such within the network. There are various ways in which the power of Simulation 1 could be varied in order to account for the effects of individual differences in Experiment 1, or the failures to obtain learning of arbitrary noun classes in previous experiments. The effect of memory ability could be dealt with by changing the learning rate parameter (which determines the size of the weight changes in response to a given amount of error). Factors such as the similarity structure of the determiners, or the number of nouns in the training set, would be expected to influence learning rate as well. However, the influence of knowledge of other gender languages is more problematic and will be considered after the remaining experiments and simulations have been reported. Connectionist networks are commonly regarded as models of the associative mechanisms underlying implicit learning (Cleeremans and Jiménez 2002). However, when we debriefed our participants after Experiment 1 it was clear that the more successful amongst them had been employing intentional learning strategies, and that there was a good correspondence between their conscious understanding of the system and their performance in the final
160 John N. Williams
generalisation test. It therefore becomes important to test whether learning could be obtained under implicit conditions.
5. Experiment 2 This experiment employed a training task that was thought to be unlikely to induce an intentional learning strategy. Participants first performed the same phonological short-term memory test and vocabulary learning exercise as in Experiment 1. Determiner-noun combinations from the training set were then auditorily presented in a semi-random sequence, avoiding immediate repetitions of the same noun or determiner. For each item the participants had to perform the following tasks: (1) repeat the phrase aloud, (2) indicate whether it refers to a living or non-living thing by pressing one of two response keys, and (3) translate the phrase into English. For example, for the item ul johombe they would respond by saying ‘ul johombe’, pressing the non-living key, and saying ‘a ball’. The meanings of the words were altered so that half of the nouns in each class referred to living things and half to non-living things. The living/nonliving decision was included because this experiment was also a control for a subsequent version in which noun animacy predicted noun class membership (see Experiment 3 below). Here it serves as a means of increasing task demands so that participants would be less likely to attempt to engage explicit learning processes. The participants were told that the purpose of the experiment was to see how their decision and translation performance improved with practice and so they were encouraged to make their responses as quickly and as accurately as possible. Training extended over 15 cycles through the 24-item training set, giving a total of 360 training trials. This took between 60 and 75 minutes including rest breaks after every five cycles. The training phase was followed by the test phase. On each trial the English translation of a test phrase was visually presented (e.g. ‘the ball’) and the participants had to choose between a grammatical and ungrammatical translation in the target language, where the determiner for the ungrammatical item was always of the correct number and definiteness, but the incorrect gender (e.g. ig johombe versus ga johombe for ‘the ball’). First the eight generalisation items were presented (see Table 1) followed by 16 trained items. There were 18 participants who were selected on the basis of their good knowledge of gender languages so as to increase the potential for obtaining learning in this experiment. They all rated themselves as intermediate or better
Inducing abstract linguistic representations
in at least two gender languages (mean = 2.8, range = 2 to 6). Twelve of the participants spoke a gender L1. Using the same scale for assessing knowledge of gender languages as employed by Williams and Lovatt (2003) they scored 5.8, which is much higher than the mean of 2.6 for the participants in Experiment 1. Their phonological short-term memory was also somewhat superior, the mean score being 71% as opposed to 64%. None of the participants were aware of the noun class system either during training or test phases. The average percentage correct on the generalisation items was 56%, which was not significantly different from the chance level of 50%, t = 1.34, p > 0.1. On the other hand, performance on the trained items was 69%, which is significantly better than chance, t = 5.53, p < 0.001, and significantly better than performance on generalisation items, t = 2.58, p < 0.05. Thus, although the participants had quite good memory for trained items, there was no evidence of learning the underlying noun class distinction. This conclusion is emphasised by the fact that the ten participants who scored 75% or better on the trained items (mean = 80%) had a mean generalisation score of 50%. Nor were there any correlations between generalisation test performance and either phonological short-term memory or language background, and participants who spoke a gender L1 did no better than those that did not (generalisation test scores were 56% for both groups). Given the failure to obtain learning in this experiment one may conclude that Simulation 1 was in fact too powerful, and that the learning that occurred in Experiment 1 was a result of purely explicit processes which fall outside the scope of the model. However, there is an alternative possibility. We should also consider the relationship between the task performed by Simulation 1 and the tasks performed by the participants in Experiments 1 and 2. Simulation 1 was intended as a model of the recall component of the training task used in Experiment 1. But in Experiment 2 the participants’ task was very different. They did not have to generate any determiners at any point during training, but only had to perform animacy decisions and produce English translations. Simulation 1 could not be said to be a good model of this task. A second simulation was therefore conducted that made different assumptions about the learning task. 5.1 Simulation 2 Incidental learning is best regarded as a relatively passive process of recording correlations between attended features in each experience. Cleeremans and
161
162 John N. Williams
Jiménez (2002), following O’Reilly and Munakata (2000: 18), have referred to this as ‘model learning’, the goal of which is to “enable the cognitive system to develop useful, informative models of the world by capturing its correlational structure”. Connectionist models of model learning do not require feedback because the system merely attempts to represent the structure of the inputs it is provided. This is in contrast to ‘task learning’ which has the aim of “mastering specific input-output mappings (i.e. achieving specific goals) in the context of specific tasks through error-correcting learning procedures” (ibid. p. 18). Crucially for present purposes they assume that model learning operates continuously, regardless of the task. Simulation 1 instantiated task learning, and was successful because the underlying noun class distinction happened to be relevant to the task the network was required to perform. But in Experiment 2 the tasks that the participants were performing (animacy decisions and translation) exerted no pressure to learn the noun class distinction. The same would be true of simulations of those tasks. The only way in which the noun class distinction could be learned, therefore, would be through model learning, which requires a different kind of network from that used in Simulation 1. One way of instantiating model learning is to train a three-layer network to associate each input to itself. That is, the network learns to reproduce the input pattern on the output layer. These are called “autoassociation” networks (Plunkett and Elman 1997). Because there are fewer hidden than input/output units the network is forced to discover an economical means of representing the patterns so that they can be reproduced on the output. This gives the network the potential to extract generalisations. Autoassociation networks do not require feedback because the input itself provides the reference point against which the accuracy of the output can be judged. How does such a network fare on the arbitrary noun class induction problem? In Simulation 2 there were 31 input units representing the eight determiners, eight stems, two inflections, three English determiners, eight English nouns, and units for singular and plural. All of the relevant information in a training item such as ul johombe, ‘a ball’ was represented as a pattern over the input layer. The 31 output units represented the same information as the input units. The network had 20 hidden units.5 For each item in the training set the network was trained to reproduce the input pattern on the output layer. Training continued until output error ceased to decline (which was after about 2,500 cycles). In Experiment 2, learning was assessed by forcing participants to choose between two translations for a phrase, for example, between ga johombe and ig
Inducing abstract linguistic representations 163
johombe as translations of ‘the ball’. The model can be tested in the same way by presenting both grammatical and ungrammatical determiner-noun combinations and comparing the strength of the output on the determiner units. For a trained item, such as ul johombe, the strength of activation of the corresponding determiner in the output, in this case ul, was, as one would expect, very high (0.996 when averaged over eight training items on five separate runs, where the required activation level was 1.0). Ungrammatical items such as ula johombe produced much weaker activation of the corresponding output determiner node, in this case ula (0.214). Clearly the network had not simply learned to reproduce input patterns on the output layer. Rather, its ability to do so was affected by whether it had received those patterns during training. In human terms this would be the equivalent of a greater feeling of familiarity for ul johombe than ula johombe. But for generalisation items the output activation on determiners in both grammatical and ungrammatical items, e.g. ig johombe versus ga johombe, was very low and not significantly different (0.054 and 0.055 respectively). In other words, both items appeared equally unfamiliar to the network. Therefore, like the human participants in Experiment 2, the autoassociation network had good memory for trained items, but was unable to distinguish between grammatical and ungrammatical generalisation items. The contrast between Simulations 1 and 2 demonstrates that task learning enabled a connectionist network to become sensitive to an abstract noun class distinction whereas model learning did not. This is a rather surprising result when one considers that there is a sense in which the networks were performing rather similar tasks. In both cases they had to remember which determiners occurred with which configurations of noun, definiteness, and number in the training items. The difference was that in Simulation 1 the network’s resources were focused on predicting the determiner from the cues that it was provided, whereas in Simulation 2 the network was actually attempting to remember the unique combination of determiner, noun, definiteness, and number that occurred in each training item. This exercise in episodic memory for entire training episodes apparently did not exert sufficient pressure on the network to discover the underlying noun class distinction. The contrast between task learning and model learning is reminiscent of the procedural-declarative distinction in Anderson’s ACT framework (Anderson 1983). Productions are sets of rules which match their ‘IF’ conditions against the current contents of working memory, and if these are satisfied, they ‘THEN’ produce some action, or deposit some other kind of representation in working memory. Although stated in a symbolic formalism in ACT, a connectionist
164 John N. Williams
network can be conceived as a subsymbolic model of the entire set of productions which perform the transformation between one type of input to another type of output (Sun et al. 2001). Both procedural learning and connectionist learning of this type depends on error correction. In contrast, Simulation 2 could be identified with the ‘declarative’ memory component of the ACT framework. The idea that these two kinds of memory system might have differential power to extract generalisations from the environment is clearly relevant to attempts to construct a theory of second language acquisition in terms of their interaction (Towell and Hawkins 1994), or to identify them with different brain regions (Ullman 2001). Indeed, the idea that procedural learning is more powerful at extracting abstract linguistic rules would be consistent with the proposal that such a mechanism supports first language acquisition, whereas second language acquisition is supported by declarative learning (Ullman 2001). If Simulation 1 is accepted as a valid model of the learning process in Experiment 1 then there is another interesting consequence. The learning that was occurring in that experiment was characterised as ‘explicit’. Not only did the participants appear to have an intention to learn, but some of them also made comparisons between consciously recalled input items, and formed conscious hypotheses. Simulation 1 captures the intentional component of the learning process, since it too evaluated its outputs with respect to feedback for the purpose of learning in order to be able to generate determiners. But it obviously does not model the other components of what, in human terms, we regard as explicit learning. However, this does not necessarily detract from the relevance of the model. Shanks (1995) reviews a range of studies on human learning where there is a good fit between human behaviour and connectionist models. Yet in many of these experiments the participants were actively searching for rules. For example, in a medical diagnosis task (see Shanks 1995: 42) participants were presented with hypothetical patients with certain symptoms and were instructed to diagnose what illness each patient had. Each trial was accompanied by feedback in the form of the correct diagnosis. Performance was directly related to the degree of contingency between different cues (symptoms) and outcomes (diseases) in the training data. Shanks showed that the results could be adequately modelled by a simple connectionist network in which symptoms were presented as inputs, diagnoses as outputs, and the correct diagnosis was provided as feedback (Shanks 1995:120). Yet the participants in the experiment presumably had the experience of actively trying to work out the relationship between symptoms and diseases. Whilst it is presently unclear how the
Inducing abstract linguistic representations 165
conscious states of the learner influence the learning mechanism, it should not be assumed that the possibility of there being such interactions rules out a unified associative explanation (Cleeremans and Jiménez 2002).
6. Experiment 3 With the benefit of hindsight Experiment 2 was a poor test of implicit learning because the kind of associative learning mechanism supposed to underlie implicit and incidental learning would not be expected to learn the underlying generalisations. We6 therefore decided to run Experiment 2 again, but this time using a system that would be learnable even by the kind of autoassociation network used in Simulation 2. The language was essentially the same as that shown in Table 1 except that the meanings of the words were altered so that all of the words in class I referred to living things and all of the words in class II referred to inanimate objects. For simplicity, the living/non-living distinction will be referred to here in terms of an ‘animacy’ cue to noun class. It has been shown that under intentional learning conditions humans have no problem grasping semantically-based noun classes (Braine 1987, Carroll 1999). A version of Simulation 2 that included animacy information confirmed that the present system was also learnable by an autoassociation network. This is presumably because there are direct associations between the units that encode animacy and certain determiners. Note, therefore, that in this experiment we are no longer concerned with whether implicit learning of abstract noun classes is possible. Rather the issue is whether implicit learning of a noun class distinction can be obtained under conditions where the connectionist model predicts that there should be an effect. The tasks and procedure were exactly the same as in Experiment 2. For each phrase presented during training the participants had to repeat it, indicate whether it referred to a living or nonliving thing, and translate it into English. Note that this time the living/nonliving decision coincided with the noun class of the word. The same learning tests were used as in Experiment 2. There were 37 participants with varied language backgrounds. Only seven of the participants became aware of the noun class distinction and its relation to animacy during the training phase, and their performance was perfect, or near perfect, on the generalisation and trained items. None of the remaining 30 participants became aware of the system during the training phase and none of them claimed to have been consciously trying to work out
166 John N. Williams
the system during the generalisation test. Even at the end of the whole testing phase none of them realised the relevance of animacy. Nevertheless, performance on generalisation items was 61%, which was significantly above the chance level of 50%, t = 3.25, p < 0.01. They scored 71% correct on trained items, which was also significantly above chance, t = 6.09, p < 0.001. Therefore, Experiment 3 succeeded in demonstrating at least some degree of implicit learning of a system that was also learnable by an autoassociation network. However, there were large individual differences in generalisation test performance. Just as in Experiment 1 there were correlations with phonological short-term memory (r = 0.50, p < 0.01) and knowledge of gender languages (r = 0.586, p < 0.001), which in this case was quantified simply in terms of the number of gender languages in which the participants rated their proficiency as intermediate or better (mean = 1.8, range = 0 to 5). We also evaluated whether, amongst the 30 unaware participants, speakers of gender L1s did better than speakers of non-gender L1s. For the 13 speakers of gender L1s mean generalisation test performance was 71%, which is significantly above chance, t = 4.08, p < 0.01, whereas for the 17 speakers of non-gender L1s it was 54%, which is not significantly above chance, t = 0.96. The difference between these two groups was significant, t = 2.78, p < 0.01. The two groups did not differ significantly in terms of the number of L2s spoken to an intermediate level or better (3.54 and 3.12 respectively, t < 0.92), the number of gender languages known as an L2 (the means were 1.46 and 1.23 respectively), but they did differ slightly in terms of phonological short term memory (77% versus 68%, p = 0.08). Better matched groups resulted from removing the three participants with the lowest phonological short term memory scores from the sample (all scores were less than 50%, and all three participants were in the non-gender L1 group). The 13 speakers of gender L1s and remaining 14 speakers of non-gender L1s were well matched in terms of number of gender languages spoken as an L2 (1.46 and 1.43 respectively) and in terms of phonological short term memory scores (77% and 73%). Yet the generalisation scores were 71% and 55% (the difference being significant, t = 2.31, p < 0.05). Note that for the non-gender L1 group the mean for the trained test items was well above chance (67%, t = 3.74, p < 0.01).
7. Discussion In one sense it could be argued that there is a good alignment between the connectionist models and the human data in the present studies, provided
Inducing abstract linguistic representations 167
assumptions are made about which kind of network is appropriate to which task conditions. Where the model was able to generalise there was also evidence for generalisation amongst the participants in the experiment (Simulation 1 and Experiment 1, Simulation 2 supplemented by animacy information and Experiment 3). Where the model was not able to generalise there was no evidence for generalisation amongst the human participants (Simulation 2 and Experiment 2). The problem is, however, that the networks only seem to account for learning amongst those participants who already possessed knowledge of other gender languages. Yet none of the networks contained any prior knowledge. Seen in this light they provide a poor fit to the human data. In this final section I shall consider ways in which prior knowledge could have influenced human learning, and whether the data then become more amenable to a connectionist interpretation. I shall then consider the implications of the present results for second language acquisition. 7.1 The role of prior linguistic knowledge One way in which prior knowledge could facilitate learning is through its effect on the learners’ strategy. Recall that the success of Simulation 1 depended upon using number, definiteness, and an abstract representation of the nouns (represented as single nodes) to generate the determiners. But this presupposes a certain understanding of the nature of gender systems. Participants who did not have this understanding may simply have approached the task in the way that it was presented to them; that is, as a short-term memory exercise for determiner-noun combinations. In that case their learning processes would be more appropriately modelled by Simulation 2 than Simulation 1. Indeed, the contrast between Simulations 1 and 2, between task learning and model learning, could be seen as a computational account of a more general contrast between analytic and non-analytic, memory-based, learning strategies (Skehan 1998). In the present case the probability of adopting an appropriate analysis strategy could have also depended upon metalinguistic knowledge of other gender systems that was derived from second language learning experience. Obviously a learning strategy account can not apply to the kind of incidental and implicit learning occurring in Experiment 3. However, in this case learning failures could be accounted for simply by assuming that animacy was not perceived as being relevant to the determiners. In Williams (in preparation) I argue that implicit learning of form-meaning connections (such as between determiners and animacy information) is problematic because of the requirement
168 John N. Williams
that form and meaning are unitised at encoding; learners must actually perceive them as being relevant to each other. Merely paying attention to the relevant elements does not appear to be sufficient, at least not under the task conditions of Experiment 3. In terms of the model learning mechanism instantiated by Simulation 2 this means that even though animacy information was attended, it did not enter the same memory trace as information about the determiner, noun identity, definiteness, and number. The problem is, therefore, to explain why participants who spoke a gender L1 defied this principle and were able to unconsciously associate the determiners with animacy information. There is no obvious connectionist answer to this problem. Is the classical linguistic approach any more promising? Linguistic (Carroll 1989, Hawkins 2001) and psychological (Levelt, Roelofs ans Meyer 1999, Vigliocco, Antonini and Garrett 1997) analyses of gender representation and processing in the L1 assume abstract gender features that are attached to nouns in the lexicon. How gender features are acquired is not often considered. However, Carroll (2001) proposes an induction procedure which is triggered by the presence of alternating determiner forms in the input (e.g. two words for ‘some’). The first occurrence of one of the determiners, for example in tei johombi (‘some monkeys’) has no effect. But when another phrase involving a word for ‘some’ is encountered, for example tegge nawasi (‘some vases’), the learner seeks to rationalise the contrast by marking the noun with a [+gender] feature. In this way, one of the determiners becomes an assigner of the gender feature whilst the other remains the default. Remembering which of the alternating pair of determiners assigns the gender feature is likely to be problematic, however. In the (admittedly artificial) case that [+gender] also corresponds to some other active feature of the noun, such as [+inanimate], one can imagine that this problem would be alleviated. To account for the influence of gender L1s in Experiment 3 it would have to be assumed that this kind of induction mechanism can only operate in L2 if it was used in the L1. This is perhaps not too implausible if one considers that each time a speaker of a gender language encounters a novel noun the same process of using the accompanying determiner to assign gender to it must operate. On the other hand, it is another matter to assume that, when confronted with a new language, learners are able to assign new gender features on the basis of newly observed alternations between determiners. It is also relevant to consider that at present there is no evidence that speakers of gender L1s have any less problem with gender in an L2 than do speakers of non-gender L1s (Bruhn and White 2000). Thus, although the gender L1 advantage found in Experiment 3
Inducing abstract linguistic representations 169
is intriguing, there is no obvious way of accounting for it at the present time from either connectionist or classical perspectives. 7.2 Implications for second language acquisition When considering second language acquisition, particularly under naturalistic un-instructed conditions, it is relevant to consider the power of incidental learning mechanisms; that is, learning that takes place as a natural consequence of processing the relevant stimuli for purposes other than discovering the underlying regularities. This means that we should consider implicit learning conditions like those in Experiments 2 and 3 and learning mechanisms of the type exemplified by Simulation 2 as being the most relevant. Granted this assumption, then the prospects for associative learning of abstract noun classes would appear to be bleak. However, one limitation of Experiment 2 and Simulation 2 is that they employed a completely arbitrary noun class system. As mentioned earlier, it has been argued that in many natural languages at least a proportion of the members of the same noun class share phonological and semantic properties. Could the presence of these cues facilitate learning? In fact a number of experimental studies have shown that partial phonological and semantic cues do indeed facilitate noun class induction (Braine 1987, Brooks et al. 1993, Frigo and McDonald 1998). However, these studies have only demonstrated an effect of partial cues under intentional learning conditions similar to those in Experiment 1. There have been no demonstrations of their effect upon implicit learning. Indeed, my own preliminary investigations of learning such systems using networks of the type used in Simulation 2 have failed to generalise to unmarked words (whereas a network such as Simulation 1 would clearly have no problem). Even under the intentional learning conditions of the earlier experiments there was very little evidence of generalisation to items that did not carry the appropriate cues. The adults in the study of Brooks et al. (1993) showed barely a significant effect using a one-tailed test (which assumes that the direction of the difference is predicted), and for the children in their second experiment there was no evidence of generalisation at all. Given that seven out of the 16 adults had explicit knowledge of the word classes, whereas only one of the children did, then it seems likely that these participants were responsible for the slightly above-chance performance of the group as a whole. Generalisation to unmarked nouns would therefore seem to be unlikely under implicit conditions.
170 John N. Williams
Only in one of three experiments of Frigo and McDonald (1998) was performance on unmarked generalisation items significantly above chance, and this was when word class was indicated by a characteristic initial and final syllable (e.g. wanersumglot, wanolovglot, wanalglot versus kaisalmrish, kaisilvrish, kaisalbrish). Braine (1987) also obtained good generalisation to unmarked words, but half of the nouns in one class referred to males and the other half to females. Thus, generalisation appears to be limited to cases where the cues are more salient than in natural languages. Somewhat counter-intuitively, where the above studies did find evidence of generalisation to unmarked items was when entirely novel nouns were introduced in the final test phase. The equivalent test in the context of the language used here would involve telling participants that ul vark means ‘a dog’ and asking them to produce the translation of ‘the dog’ (the correct answer being ig vark). Such a test only requires knowledge of the associations between the determiners. Therefore, it does appear that partial phonological cues can facilitate acquisition of inter-determiner associations (or rather, their equivalent in the languages that were used). Determiners in the same class presumably become associated by virtue of their frequent association to the same phonological cue. Generalisation is then achieved by a process of inference from another determiner-noun combination that is provided at test or recalled from memory. As argued by Frigo and McDonald (1998), poor performance on generalisation tests involving nouns that occurred in training could be because of problems recalling an example of a determiner that occurred with that noun. But the native speaker of a gender language is assumed to generate an appropriate determiner directly on the basis of an abstract specification of the noun’s gender in the lexicon, not by inference. It is far from clear that the participants in these experiments acquired knowledge of noun classes in that sense. The results from these studies do not, therefore, offer much prospect of incidental learning of noun classes. This is of course consistent with the claim that gender is a persistent problem for second language learners. Assuming an underlying model learning mechanism such as that in Simulation 2, learning would be predicted to be limited to rote storage of determiner-noun combinations, and associations between determiners and partial phonological and semantic cues. This would explain L2 learners’ sensitivity to phonological cues in gender processing tasks (Guillelmon and Grosjean 2001, Holmes and De la Batie 1999, Taraban and Kempe 1999). Unmarked nouns would have to be dealt with through rote storage, putting a strain on phonological memory (Williams and Lovatt 2003). The lack of a true underlying noun class organisation
Inducing abstract linguistic representations
would make storage of determiner-noun combinations particularly prone to error, but if at least one instance of a determiner-noun pair can be retrieved, other appropriate determiners could be inferred using knowledge of interdeterminer associations. Thus, second language learners can acquire a semblance of competence but the failure to organise the underlying representations in terms of abstract noun classes will cause persistent problems. I have argued that this reflects a weakness in the type of associative learning mechanism that is assumed to underlie incidental learning.
Notes 1. This language was derived from Italian. The determiners were derived from the Italian il, i, un, dei, la, le, una, and delle by systematically substituting consonants (lÆg, dÆt, nÆl). The nouns correspond to Italian nouns which end in -e in the singular and -i in the plural regardless of gender, e.g. cliente (masculine), stazione (feminine). Note that none of the participants in Experiments 1 and 2 (reported below) had any knowledge of Italian, and only two participants in Experiment 3 knew Italian at an intermediate level or better as an L2. 2. The only difference was that in the experiment they also had to produce the inflection, whereas in the simulation the inflection was provided on the input. However, in the experiment the participants learned the correct plural inflections in the preliminary vocabulary learning phase, and not in the training phase of the main experiment. In any case the inflection provides no clue as to the correct determiner over and above the presence or absence of the plurality of the noun. 3. A Root Mean Square error of 0.1 means that over all of the input patterns presented on a particular cycle the average difference between the actual output and the required output on each node was 0.1 units of activation. The point at which the correct output node was simply the most active occurred well before an RMS error of 0.1 was achieved. 4. The Luce ratio was also used as a measure of network performance — the activation level of the correct output node divided by the sum of the activation over all output nodes. Perfect output would be indicated by a Luce ratio of 1.0. In this simulation the mean Luce ratio over 20 runs was 0.87. 5. The number of hidden units was set to about two thirds of the number of input/output units so as to force the inputs through a reduced representational space, exerting pressure on the network to extract generalisations. Other simulations were performed with either 10 or 40 hidden units but the generalisation performance was similar to that reported here. 6. This experiment was run in collaboration with Helen East.
171
172 John N. Williams
References Anderson, J. R. 1983. The architecture of cognition. Cambridge MA: Harvard University Press. Berent, I., Marcus, G. F., Shimron, J. and Gafos, A. I. 2002. “The scope of linguistic generalizations: Evidence from Hebrew word formation”. Cognition 83: 113–139. Braine, M. D. S. 1987. “What is learned in acquiring word classes: A step towards an acquisition theory”. In Mechanisms of language acquisition, B. MacWhinney (ed.), 65–87. Hillsdale, NJ: Lawrence Erlbaum. Braine, M. D. S., Brody, R. E., Brooks, P. D., Sudhalter, V., Ross, J. E., Catalano, L. and Fisch, S. M. 1990. “Exploring language acquisition in children with a miniature artificial language: Effects of item and pattern frequency, arbitrary subclasses, and correction”. Journal of Memory and Langage 29: 591–610. Brooks, P J., Braine, M.D S., Catalano, L. and Brody, R. 1993. “Acquisition of gender-like noun classes in an artificial language: The contribution of phonological markers to learning”. Journal of Memory and Language 32: 76–95. Bruhn, J. and White, L. 2000. “L2 acquisition of Spanish DPs: the status of grammatical features”. In Proceedings of the 24th annual Boston University conference on language development. Vol. 1, S. C. Howell, S. A. Fish and T. Keith-Lucas (eds), 164–175. Somerville, Mass.: Cascadilla Press. Carroll, S. 1989. “Second language acquisition and the computational paradigm”. Language Learning 39: 535–594. Carroll, S. E. 1999. “Input and SLA: Adults’ sensitivity to different sorts of cues to French gender”. Language Learning 49: 37–92. Carroll, S. E. 2001. Input and evidence: The raw material of second language acquisition. Amsterdam: John Benjamins. Caselli, M. C., Leonard, L. B., Volterra, V. and Campagnoli, M. G. 1993. “Toward mastery of Italian morphology: A cross-sectional study”. Journal of Child Language 20: 377–393. Cleeremans, A. and Jiménez, L. 2002. “Implicit learning and consciousness: A graded, dynamic perspective”. In Implicit learning and consciousness, R. M. French and A. Cleeremans (eds), 1–40. Hove: Psychology Press. Corbett, G. 1991. Gender. Cambridge: Cambridge University Press. Ellis, N. C. 1998. “Emergentism, connectionism and language learning”. Language Learning 48: 631–664. Elman, J. L. 1990. “Finding structure in time”. Cognitive Science 14: 179–211. Fodor, J. A. and Pylyshyn, Z. W. 1988. “Connectionism and cognitive architecture: A critical analysis”. Cognition 28: 3–71. Frigo, L. and McDonald, J. L. 1998. “Properties of phonological markers that affect the acquisition of gender-like subclasses”. Journal of Memory and Language 39: 218–245. Gómez, R. L. and Gerken, L. 2000. “Infant artificial language learning and language acquisition”. Trends in Cognitive Sciences 4: 178–186. Guillelmon, D. and Grosjean, F. 2001. “The gender marking effect in spoken word recognition: The case of bilinguals”. Memory and Cognition 29: 503–511. Hawkins, R. 2001. Second language syntax: A generative introduction. Oxford: Blackwell.
Inducing abstract linguistic representations
Holmes, V. M. and De la Batie, B. D. 1999. “Assignment of grammatical gender by native speakers and foreign language learners”. Applied Psycholinguistics 20: 479–506. Johnstone, T. and Shanks, D. R. 1999. “Two mechanisms in implicit artificial grammar learning? Comment on Meulemans and Van der Linden 1997”. Journal of Experimental Psychology: Learning, Memory, and Cognition 25: 524–531. Kelly, M. H. 1992. Using sound to solve syntactic problems: The role of phonology in grammatical category assignments. {sychological Review 99: 349–364. Knowlton, B. J. and Squire, L. R. 1996. “Artificial grammar learning depends on implicit acquisition of both abstract and exemplar-specific information”. Journal of Experimental Psychology: Learning, Memory, and Cognition 22: 169–181. Levelt, W J. M., Roelofs, A. and Meyer, A. S. 1999. “A theory of lexical access in speech production”. Behavioural and Brain Sciences 22: 1–75. Maratsos, M. P. and Chalkley, M. A. 1980. “The internal language of children’s syntax: The ontogenesis and representation of syntactic categories”. In Children’s Language Vol. 2, K. Nelson (ed.), 127–214. New York: Gardner Press. Marcus, G. F. 1999. “Language acquisition in the absence of explicit negative evidence: Can simple recurrent networks obviate the need for domain-specific learning devices?” Cognition 73: 293–296. Marcus, G. F., Vijayan, S., Bandi Rao, S. and Vishton, P. M. 1999. “Rule learning in 7-monthold infants”. Science 283: 77–80. Mathews, R.C., Buss, R.R., Stanley, W.B., Blanchard-Fields, F., Cho, J.-R. and Druhan, B. 1989. “The role of implicit and explicit processes in learning from examples: A synergistic effect”. Journal of Experimental Psychology: Learning, Memory, and Cognition 15: 1083–1100. Meulemans, T. and Van der Linden, M. 1997. “Associative chunk strength in artificial grammar learning”. Journal of Experimental Psychology: Learning, Memory, and Cognition 23: 1007–1028. O’Reilly, R. C. and Munakata, Y. 2000. Computational explorations in cognitive neuroscience: Understanding the mind by simulating the brain. Cambridge: MA: MIT Press. Pinker, S. 1984. Language learnability and language development. Cambridge, Mass.: Harvard University Press. Plunkett, K. and Elman, J. L. 1997. Exercises in rethinking innateness: A handbook for connectionist simulations. Cambridge, MA: MIT Press. Redington, M. and Chater, N. 1998. “Connectionist and statistical approaches to language acquisition: A distributional perspective”. Language and Cognitive Processes 13: 129–191. Rumelhart, D. E. and McClelland, J. L. 1986. “On learning the past tense of English verbs”. In Parallel distributed processing: Explorations in the microstructure of cognition Vol. 2, J. L. McClelland and D. E. Rumelhart (eds), Cambridge, MA: MIT Press. Saffran, J. R. 2001. “The use of predictive dependencies in language learning”. Journal of Memory and Language 44: 493–515. Seidenberg, M. S. and McClelland, J. L. 1989. “A distributed, developmental model of word recognition and naming”. Psychological Review 96: 523–569. Shanks, D. R. 1995. The psychology of associative learning. Cambridge: Cambridge University Press. Skehan, P. 1998. A cognitive approach to language learning. Oxford: Oxford University Press.
173
174 John N. Williams
Sokolok, M. E. and Smith, M. E. 1992. Assignment of gender to French nouns in primary and secondary language: A connectionist model. Second Language Research 8: 39–58. Sun, R., Merrill, E. and Peterson, T. 2001. “From implicit skills to explicit knowledge: a bottom-up model of skill learning”. Cognitive Science 25: 203–244. Taraban, R. and Kempe, V. 1999. “Gender processing in native and nonnative Russian speakers”. Applied Psycholinguistics 20: 119–148. Tomasello, M. 2000. “The item-based nature of children’s early syntactic development”. Trends in Cognitive Sciences 4: 156–163. Towell, R., and Hawkins, R. 1994. Approaches to second language acquisition. Clevedon: Multilingual Matters. Ullman, M. T. 2001. “The neural basis of lexicon and grammar in first and second language: The declarative/procedural model”. Bilingualism: Language and Cognition 4: 105–122. Vigliocco, G., Antonini, T. and Garrett, M. F. 1997. “Grammatical gender is on the tip of Italian tongues”. Psychological Science 84: 314–317. Williams, J. N. In preparation. “Implicit learning of form-meaning connections”. Williams, J. N. and Lovatt, P. 2003. “Phonological memory and rule learning”. Language Learning 53: 67–121.
Chapter 8
Neural substrates of representation and processing of a second language* Laura Sabourin and Marco Haverkort University of British Columbia / University of Nijmegen & Boston University
1.
Introduction
Most research in second language acquisition — as well as in first language acquisition — does not make a careful enough distinction between the different levels at which language behaviour and changes in the language of the learner are described. The description is usually cast in terms of a representation of grammatical knowledge available to the learner, and changes in language behaviour are viewed as the result of qualitative changes in that knowledge, for instance the addition of a rule or the resetting of a parameter. In this paper, we want to argue that it is important to distinguish between the representation of grammatical knowledge, the language processor, and general cognitive strategies in adult second language acquisition. In order to be able to distinguish between grammar and processor in second language acquisition, we will compare results obtained with different methods. In particular, we will use an off-line grammaticality judgment task to tap grammatical knowledge of second language learners, and on-line EEG measurements to investigate to what extent the processing strategies of second language learners are qualitatively similar to those used by native speakers. Specifically, we will investigate the use of grammatical gender to see if L2 processing is strictly linguistic in nature or depends to some degree on more general cognitive strategies. In the next sections we will first look at evidence from the field of aphasia that a distinction between knowledge and grammar on one hand and the processor on the other should be made. We will then look at two L2 experiments to see if this distinction also holds for L2.
176 Laura Sabourin and Marco Haverkort
2. Grammar versus processor: Evidence from aphasia There is quite extensive evidence from another domain of linguistic inquiry — the study of aphasia — that supports the idea that the representation of grammatical knowledge on the one hand and language processing on the other are dissociated entities. According to this view, aphasics have the knowledge of their language available, but cannot process language on-line, due to working memory or other processing limitations (cf. Kolk 1995, 1998). There are a number of observations that support the idea that aphasics still have access to the correct grammatical representations, but that their access is too slow for adequate on-line processing. First, although there is a spontaneous recovery process post-onset for virtually all patients, there is no indication that the representations must be re-acquired or relearned. Second, the fact that patients exhibit task-dependent variation supports aphasics having access to the grammatical representations. They perform at chance level with certain constructions — such as object relative clauses, passives, and object clefts (Caplan and Hildebrandt 1988, Grodzinsky 1990) — in a sentence-picture matching task, where they have to select the picture that correctly depicts the sentence they were just given. These same patients can perform much better (close to ceiling level), in a grammaticality judgment task. Linebarger et al. (1983) and Grodzinky and Finkel (1998) however, found that their patients had problems with the grammaticality judgment task, especially when the sentences to be judged involved antecedent-trace dependencies. The former task is more complex and involves more processing than the latter: a syntactic structure needs to be established, onto which a semantic representation is then mapped; subsequently, the pictures need to be analysed, resulting in a conceptual structure, and finally these conceptual structures have to be compared with the semantic representation of the sentence in order to find the best match. In a grammaticality judgment task, on the other hand, only the first step needs to be taken: a syntactic representation needs to be established, and if the structure under construction fails before it is finished, the sentence is marked as ungrammatical (Chomsky 1995). A full semantic structure does not need to be computed, as in the sentence-picture matching task. Thus, if patients have a processing problem, it is to be expected that they will perform better on the latter task, which is simpler computationally and requires less storage capacity. The fact that they perform much better on grammaticality judgments clearly indicates that the grammatical knowledge must be available at some level. Third, it has been shown in several studies (Burkhardt et al. 2001, Haarmann
Neural substrates of representation and processing of a second language 177
1993, Kolk 2002, among others) that aphasics exhibit syntactic and semantic priming effects. In a syntactic priming task, unimpaired subjects are quicker in a lexical decision task if the target is a word that syntactically fits into the sequence of words heard or read up to the point of presentation of the target. It reflects the fact that language users have clear expectations about what syntactic category is to come next. For aphasics, these effects also show up, but only with stimulus onset asynchronies (SOA) that are larger than the optimal SOA for unimpaired subjects: whereas the optimal SOA for unimpaired subjects is 300 ms (with longer SOAs the effect gradually disappears), the optimal SOA for the aphasics is much larger. Haarman (1993) presents data from a syntactic priming study. He compared sentences such as those in (1) and found a priming effect for an unimpaired control group of about 65 ms on the last word: if that word fit the syntactic context, the unimpaired control group made the lexical decision 65 ms quicker than if it did not. The agrammatic patients showed the same priming effect (a quicker response to words that fit the syntactic context), but only when the SOA was increased from 300 (normals) to 1100 ms. (1) a.
Wij zijn getest/*gewandeld. ‘We are tested/walked.’ b. Wij kunnen praten/*neus. ‘We can talk/nose.’ c. op de tafel/*rood ‘on the table/red’
The fact that the aphasics showed a syntactic priming effect can only be explained under the assumption that they have the relevant knowledge (regarding phrase structure and subcategorization) at their disposal and hence have syntactic expectations as to what word class the next word will be; otherwise, no effect should be found. The fact that the optimal SOA is a little over three times as large for the aphasic population as for the control group, however, indicates that the aphasics cannot make use of the relevant knowledge quick enough online; as soon as they are given more time, the exact same effect shows up as for the unimpaired population. However, for the unimpaired control group, the priming effect disappeared when items were presented at longer SOAs. A similar priming effect has been shown to exist in filler-gap dependencies, using semantic priming. Burkhardt et al. (2001), using sentences with moved wh-phrases and DPs, presented semantically related or unrelated words (Examples 2 and 3 below) at the trace or 600 ms after the trace in object position (indicated by ti).
178 Laura Sabourin and Marco Haverkort
(2) The kid loved the cheese whichi the brand new microwave melted ti yesterday afternoon while the entire family was watching tv. (3) The butteri in the small white dish melted ti after the boy turned on the brand new microwave.
In this experiment as well, the priming effect can only be observed in a small window, which for normals is immediately at the object position; if the semantic prime is presented with a delay, the priming effect is gradually lost in the unimpaired population, an indication that it is indeed the reactivation of the semantic content of the moved phrase at the trace position that causes the effect. Here, again, the aphasics exhibit a priming effect, but only when the semantically (un)related word is presented with a delay of 600 ms, indicating that the patients can construct the trace associated with the moved wh-phrase or DP. Thus, the representation of the relevant syntactic knowledge must be available to them; otherwise no priming effect would be expected. These observations all point in the direction that a processing-based account of aphasic behaviour is on the right track: the knowledge base seems to be available, and can be used by the patients under particular conditions. However, the task cannot be too complicated or involve too many sub-tasks, and the patients need to be given sufficient time to do the task. At the behavioural level, young children and second language learners behave similarly to the aphasics in a number of respects (subjects are omitted, verbs are not inflected for tense and agreement but occur in the infinitival form in the corresponding syntactic position instead, and functional categories — conjunctions, determiners, pronouns, auxiliaries, copula verbs and prepositions, for instance — are omitted), which suggests that their behaviour should be explained along similar lines, viz. in terms of processing limitations, in line with Ockham’s razor (see also Avrutin, Haverkort and Van Hout 2001 and the different papers in that volume). We hypothesize that in other populations, particularly second language learners, however, these limitations are of a different nature: not so much timing restrictions, as in the aphasic population, but the use of qualitatively different processing strategies (see below).
3. Second language processing As indicated above, the aim of this paper is to investigate the role of the representation of grammatical knowledge, language processing and general cognitive strategies in adult second language learners. It is possible that successful
Neural substrates of representation and processing of a second language 179
second language learners have native-like knowledge, just like the aphasics (suggesting that access to Universal Grammar (UG) for a second language is possible). However, they may actually process this knowledge in a non-nativelike manner. Their non-native processing, though, may not be due to timing limitations as in aphasic populations (a quantitative difference) but may be a qualitative difference. We will now look at studies that investigate whether advanced second language learners of Dutch, even if they exhibit the same knowledge as native speakers in an off-line grammaticality judgment task, exhibit the same neurophysiological responses to grammatical violations. This would indicate that, even though their knowledge is comparable to that of native speakers, their online processing differs qualitatively. Comparison of data obtained using the traditional grammaticality judgment technique with those obtained by tapping directly into electrophysiological activity in the brain associated with a specific processing phenomenon allows us to study knowledge and processing separately. Grammatical gender is specifically interesting in this respect, because it involves both lexical and syntactic aspects; hence storage, computation, and their interaction can be studied simultaneously.
4. Knowledge versus processing: Two experiments 4.1 Grammatical gender Grammatical gender or noun classification systems are found in many of the world’s languages. Dutch is a gender language with two gender classes, marked by the definite articles de (common gender) and het (neuter gender). Originally, the language employed a three gender system with masculine, feminine and neuter categories, but the former two were conflated into one common gender. The earlier three-way system is similar to the system presently used in German, a language that is closely related to Dutch. The following experiments investigated how second language speakers deal with local grammatical gender agreement within the noun phrase. There are two different types of agreement that fall into this category. One type is the agreement between definite determiner and noun. Common gender nouns (such as tafel, ‘table’) take the common definite determiner de and neuter gender nouns (such as kind, ‘child’) take, in the singular, the neuter definite determiner het. In the plural, the determiner de is used for both common and
180 Laura Sabourin and Marco Haverkort
neuter gender nouns. The indefinite determiner is the same for both genders, i.e. een, and the agreement is only evident on the adjective (adjective-noun agreement). For indefinite common gender nouns, the suffix -e is added to the adjective, while for indefinite neuter nouns the adjective remains uninflected, as shown in the following examples: (4) a.
Een klein kind. a small-Ø child-neut b. Een klein-e tafel. a small-agr table-com
4.1.1 Experiment 1: Grammatical knowledge This first experiment was designed to determine the level of knowledge second language speakers can achieve concerning the Dutch gender agreement system. Only advanced participants with German as their native language were tested in the second language group. The task here, as for Experiment 2, was to judge the grammaticality of sentences. There were 2 types of sentences in the experimental items: the first sentence type contained either the correct or incorrect definite determiner, while the second type contained either correct or incorrect adjectival agreement.
Participants In total 59 participants were tested on this task: 34 native speakers of Dutch formed the control group, while there were 25 second language learners with German as their native language. As we were interested in studying advanced second language learners, participants were required to have a high level of proficiency. Therefor attain such participants had to have been using Dutch for at least three years. A proficiency score was also obtained from each second language participant; a score of 90% or more correct was required. This proficiency score was determined by testing participants on their knowledge of number and finiteness agreement.1 Information about the participants is summarized in Table 1. Materials and methodology The grammaticality judgment test contained 80 sentences of interest, each of which contained the critical determiner-adjective-noun sequence. Half of these items belonged to the determiner-noun agreement condition and the other half to the adjective-noun agreement condition. For the first condition, the critical
Neural substrates of representation and processing of a second language
Table 1.Participant information. The number of participants included for each language group along with information as to the average duration and range of exposure the German participants had to the Dutch language. Native language
Exposure to Dutch
Accuracy on proficiency test
Dutch (n = 34)
N/A
Range: 90–100% Average: 98%
German (n = 25)
Range: 2–49 yrs* Average: 11.6 yrs
Range: 92–100% Average: 97%
* The one German subject who had less then 3 years of exposure to Dutch, started teaching himself Dutch while still living in Germany, but those years were not counted, as the amount of Dutch used before moving to the Netherlands could not be determined.
nouns in the sentences were preceded either by the correct definite determiner or by the incorrect definite determiner (Example 5). In the second condition there were sentences containing indefinite NPs in which the critical noun was preceded either by the correctly or incorrectly inflected form of the adjective (Example 6). The test also included 200 filler sentences with different types of violations: 80 sentences that were used in the proficiency measure, 80 sentences looking at the use of the relative pronouns and 40 sentences looking at the form of the predicative adjective. Full details can be found in Sabourin (2003). (5) Het/*De kleine kind probeerde voor het eerst te lopen. the-neut/*com small child-neut tried for the first to walk. ‘The small child tried to walk for the first time.’ (DET-N agreement) (6) Hij loopt op een gekke/*gek manier. (A-N agreement) he walks in a funny-com/*neut way.
The critical nouns used in this experiment were controlled for frequency. Half of the items were of high frequency while the other half were of a middle frequency. The middle frequency items were still of a fairly high frequency to ensure that second language participants would know them. The frequency of each item was determined through the CELEX database (Burnage 1990). The log frequency of each high frequency item was between 1.96 and 2.98 (average 2.28); for each middle frequency item, it was between 1.11 and 1.49 (average 1.31). Items were also broken down into gender class: half of the nouns were common gender nouns (de) and the other half were neuter gender ones (het). Each subject received a grammaticality judgment questionnaire. The participants were asked to first go through the test making a yes/no decision as to the grammaticality of each sentence. They were required to complete this
181
182 Laura Sabourin and Marco Haverkort
task in 30 minutes. After judging the grammaticality of each sentence, they were asked to go back to the beginning and correct every sentence they had marked as ungrammatical. This was done to ensure that subjects were rejecting a sentence for the right reasons and not, for instance, due to the fact that they felt that an incorrect preposition or incorrect word order had been used. In scoring the grammaticality judgments only sentences that were both correctly judged as grammatical or ungrammatical and that contained a relevant correction were considered as correct answers. For example, in the ungrammatical version of the sentence in (5), repeated below as (7), the participant correctly might have said the sentence was ungrammatical, but, in the correction of the sentence only changed the position of the prepositional phrase voor het eerst. If this was the case, the answer was scored as incorrect. (7) *De kleine kind probeerde voor het eerst te lopen. the-com small child-neut tried for the first to walk ‘The small child tried to walk for the first time.’
Similarly, if the sentence was supposed to be marked as grammatical but the subject rated the sentence as ungrammatical, making a correction that was unrelated to the condition being tested, the answer was scored as correct. For example, if the above sentence had been grammatical (with het kind instead of de kind), but the subject still rated it as ungrammatical due to the position of the prepositional phrase, the answer would have been considered as correct, since correct judgment was given with respect to gender.
Results The results for this experiment will be analysed using a four-way Anova (analysis of variance). Only responses to the ungrammatical items will be analysed as scores on the grammatical items were near perfect for both groups. The within-subjects effects were definiteness (definite and indefinite), frequency (high and middle), and gender (common and neuter). The between-subjects effect was native language (Dutch and German). Only one significant interaction was found in this analysis; definiteness significantly interacted with L1 (F(1,57) = 15.8, p < .001). This effect can be seen in Figure 1. The main effects of definiteness (F(1,57) = 32.84, p < .001) and L1 (F(1,57) = 25.37, p < .001) were also significant. What is most important to note here is that while there was a difference between the native speakers and the L2 learners, this is only clearly the case when indefinite NPs are being used. The L2 learners perform significantly less
Neural substrates of representation and processing of a second language
100 90 80
def indef
70 60 50 Dutch
German
Figure 1.Scores (in percent) comparing the Dutch and German scores on the definite and indefinite NPs.
worse than the native speakers on the definite NP items compared to the indefinite NP items. To summarize, the German group shows that, for the definite NPs (the determiner-noun agreement condition), their knowledge is similar to that of the native speakers. However, for indefinite NPs (the adjective-noun agreement condition), the German group performs very poorly. One way to interpret these results is by noting that determiner-noun agreement is similar to simply assigning gender to nouns and can, therefore, be done on the basis of lexical information rather than via a syntactic process. On the other hand, adjectivenoun agreement requires the participants to take the lexical knowledge of which gender an item is and apply this information in order to correctly inflect the adjective. In the next experiment we look at the on-line processing of the same sentences. The question now is how the L2 group processes these different kinds of data on the on-line version of the task. 4.1.2 Experiment 2: Processing As was seen in the first experiment, the German group showed that when they are given an overt determiner they can judge the grammaticality quite accurately. If only this off-line measure had been used, the conclusion might be that second language speakers acquire a native-like competence of their second language. But, as was argued above, there are some reasons to think that the representation of knowledge on the one hand and language processing on the other can be dissociated. For aphasics, evidence was presented, indicating that they still have knowledge of the language but that they cannot process the language on-line in a quick enough manner. It is therefore also possible that
183
184 Laura Sabourin and Marco Haverkort
second language learners acquire the knowledge of the second language grammar but are not able to process this language in the same manner as native speakers, though for different reasons than the aphasics. The fact that accuracy decreases when syntactic agreement must be processed is suggestive that this may be the case. There are numerous techniques that can be used to test on-line language processing. Some experiments make use of reaction time (RT) measurements such as lexical decision and self-paced reading. Unfortunately, although these techniques can tell us a lot about how language processing is organized in terms of its general architecture, they may not be fine-grained enough to determine whether similar or different processing mechanisms are being used. One on-line technique that provides detailed information on the qualitative aspects of processing is electroencephalography (EEG). The neuroimaging technique of Event-Related Potentials (ERPs) is able to measure the electrophysiological activity in the brain that is thought to directly reflect neural activity. ERPs are negative or positive changes in the voltage of the ongoing brain activity that can be elicited by sensory input or a cognitive task. The technique provides information on the latency, amplitude, polarity, and distribution over the scalp of the EEG-signal. It has been found in previous language studies that signals elicited by, for instance, grammatical and ungrammatical sentences can be discriminated (Kutas 1993, Rugg and Coles 1995). One well-understood ERP-correlate of syntactic language processing is the P600 or Syntactic Positive Shift (SPS) which is associated with processing of morpho-syntactic anomalies and complexity (Osterhout and Holcomb 1992, Hagoort, Brown and Groothusen 1993). The P600 is a positive deflection in the EEG-signal that starts approximately 500 ms after the presentation of the word that renders a sentence ungrammatical; this positivity continues for about 400 ms and reaches its maximum amplitude at around 600 ms after the presentation of the word that renders the sentence ungrammatical. This component is most prominent in centro-parietal regions of the scalp. There has recently been some research on the neural correlates of grammatical gender processing in both Dutch and German. These studies have only looked at native speakers and they have looked only at determiner-noun agreement. For Dutch, Hagoort and Brown (1999) showed that grammatical gender incongruencies result in an increase in the amplitude of the P600 component as compared to sentences with congruent determiner-noun agreement. A P600 component was also found in German for gender violations (Gunter, Friederici and Schriefers 2000). Thus, grammatical gender violations
Neural substrates of representation and processing of a second language
in both Dutch and German result in an increase in amplitude of the P600 component. The question then is: will Germans also show this P600 in their L2 processing?
Participants In total 39 participants were tested on the ERP version of the above experiment. There were 23 native speakers of Dutch and 14 second language speakers with German as their native language. The L2 participants have lived in the Netherlands between two and 32 years with an average of 9.8 years. None of the participants took part in Experiment 1. Materials and methodology The critical stimulus sentences used in this experiment were the same sentences used in Experiment 1. While the task used in the ERP version also contained a grammaticality judgment, there were a few important differences in how the on-line task was run compared with the off-line task from Experiment 1. During the ERP measurement, participants were seated in a dimly lit soundproof room facing a computer monitor. Sentences were presented word by word in the middle of the screen (the word was on the screen for 250 ms, followed by a blank screen for 250 ms before the next word appeared). Each sentence was preceded by an asterisk (to let participants know that a new sentence was about to start). After each sentence, a delay screen was displayed, followed by a screen requesting subjects to give a grammaticality judgment by pushing one of two buttons. After each sentence participants were given two seconds in which they were allowed to blink.2 The experiment started with a practice session to allow participants to get used to the presentation style of the sentences and to practice not blinking during the sentence trials. The actual experiment lasted approximately one hour. The native speakers were given three breaks while the L2 participants were given a total of seven breaks; the length of the pause was chosen by the participant, so the total testing time varied, depending on the length of the pauses that were taken. EEG recording The EEG activity was recorded by means of tin electrodes mounted in an elastic cap (Electro-Cap International) from 12 electrode sites, based on the international 10–20 system. The 12 electrodes analyzed were: F7, Fz, F8, T3, Cz, T4, T5, Pz, T6, O1, Oz and O2. The ‘F’ represents frontal electrodes, ‘T’ represents temporal electrodes, ‘C’ represents central electrodes, ‘P’ represents parietal
185
186 Laura Sabourin and Marco Haverkort
electrodes and ‘O’ represents occipital electrodes. Odd numbers represent electrodes on the left half of the scalp, even numbers represent electrodes on the right half, and the ‘z’ represents electrodes along the midline. All electrodes were referenced to linked mastoids. Both horizontal and vertical electro-oculograms (EOGs) were measured for both eyes. Electrode impedances were kept below 5 kΩ. EEG and EOG signals were sampled at 1000 Hz, amplified and digitally filtered with a cut-off frequency of 30 Hz; effective sample frequency was 100 Hz.
Results First the behavioural results (accuracy on the grammaticality judgment) will be presented followed by separate analyses of the on-line ERP data for the native speakers and the German group. For the L2 group only the sentences which they correctly judged as grammatical will be analyzed. The average accuracy scores for the Dutch and German groups are presented in Table 2. Table 2.Accuracy scores (in %) ont the ERP version of the grammaticality judgment task NP Definites
Dutch (n = 23) German (n = 14)
NP Indefinites
grammatical
ungrammatical
grammatical
ungrammatical
95% 88%
90% 80%
93% 93%
92% 68%
Average ERPs were computed at the above electrode sites for each participant in all conditions. The averaging was done for an interval starting at the onset of the critical noun and continuing for 1500 ms post-onset. All averages were aligned with a 200 ms pre-stimulus baseline (200 ms before onset of the critical noun is set to zero for both conditions to correct for pre-existing differences in the EEG). For analysis purposes, averaged ERPs of each 1500 ms epoch were divided into 50 intervals of 30 ms. This method allows one to see the onset and duration of effects more clearly. In each of these 50 intervals, mean amplitudes were statistically analysed with a Manova. Effects will be reported only if three or more successive intervals reach significance at the .05 level for native speakers and at the 0.1 level for the German group. Three successive significant intervals are more likely to reflect a real and reliable effect despite the use of multiple comparisons. A .1 level will be allowed for the L2 speakers in order to avoid false negative results of the P600 as less of their data can be analysed (only the sentences for which they made a correct judgment) and it is expected that
Neural substrates of representation and processing of a second language 187
their data will also be more variable. Both of these reasons will likely make it more difficult to find significant differences in the wave patterns so a less strict level of significance will be taken, but the reader must be aware then, that unexpected significant results should be looked at carefully. Each language group will be analysed separately and then discussed comparatively. For each 30 ms interval a three-way Manova was carried out looking at the effects of grammaticality (two levels), front to back scalp distribution (four levels: frontal, temporal, parietal and occipital), and left to right scalp distribution (three levels: left, midline and right). The four levels of the front to back scalp distribution are: F7, Fz and F8 as the first level, T3, Cz and T4 as the second level, T5, Pz, and T6 as the third level and O1, Oz and O2 as the fourth level. The three levels of the left to right scalp distribution are: F7, T3, T5 and O1 as the left hemisphere electrodes; Fz, Cz, Pz and Oz as the midline electrodes and; F8, T4, T6 and O2 as the right hemisphere electrodes.
Definite NPs Upon a visual inspection of the ERP patterns for the definite NPs for the native speakers a clear P600 pattern is seen; the ungrammatical sentences are more positive than the grammatical sentences over the more posterior electrodes. This can be seen by looking at Figure 2. Looking at the waves statistically the presence of a P600 is supported. Within each of the 30 intervals from 570 to 1500 ms the effect of grammaticality is significant at the .05 cut-off. Within this same time frame there was also a significant interaction with the front to back factor. From 570 to 900 ms, there is a largely distributed P600 component which is significant over the following electrodes: Fz, C3, Cz, C4, T5, P3, Pz, P4, T6, O1, Oz and O2. From 900 ms to the end the posterior positivity is maintained, while a frontal negativity starts which is only significant to the .05 level at electrodes F3 and F8. Upon visually inspecting the ERP patterns for the German participants (see Figure 3) we can also see a P600 component. Statistically we see that indeed a P600 effect is present. However, it is much more restricted and starts later than the one found for native speakers. Between 840 and 990 ms there is a significant positivity for the ungrammatical sentences at electrodes C3, Cz, C4, T5, P3, Pz, P4 and T6. This can be seen in Figure 4 where the difference waves at electrode Pz are compared for the Dutch and German groups. Difference waves represent the wave found for the ungrammatical sentence minus the wave found for the grammatical sentence; thus a positivity in the difference wave reflects a positivity in the ungrammatical sentences.
188 Laura Sabourin and Marco Haverkort
Figure 2.ERP wave patterns for the grammatical (darker line) versus ungrammatical sentences for the native speakers in the NP definite condition. The y-axis represents a voltage of ±5 microvolts with positive plotted up.
In Figure 4 we can see that the P600 effect starts later for the German group and that its maximal amplitude occurs later as well. Another difference between the ERP waves for the German and Dutch groups can be seen. The German group does not show the late frontal negativity that is seen in the native speaker ERP.
Indefinite NPs A visual inspection of the native speaker data (see Figure 5) shows that a P600 effect is also present for these sentences. Statistically, the P600 effect is significant from 600 to 710 ms at electrodes C3, Cz and C4 and between 600 and 1500 ms at electrodes T5, P3, Pz, P4, T6, O1, Oz and O2. Electrodes Fz and F8 show a late frontal negativity. This can be seen in Figure 5. Visual inspection of the German data reveals no obvious effects (see Figure 6). It is important to note that only sentences that were correctly judged
Neural substrates of representation and processing of a second language 189
Figure 3.ERP wave patterns for the grammatical (darker line) versus ungrammatical sentences for the German group in the NP definite condition. The y-axis represents a voltage of ±5 microvolts with positive plotted up.
in the grammaticality judgment portion of this task are included in the analyses below. Looking at the statistics there are three time frames where the effect of grammaticality is significant: 300 to 420 ms, 780 to 960 ms and 1080 to 1200 ms. Within none of these time frames does the effect of grammaticality significantly interact with either the front to back factor or the left to right factor. Looking at each electrode separately also does not result in any significant differences between grammatical and ungrammatical waves. However, upon visual inspection, it appears that in the first time frame there is a frontal positivity, followed by a left posterior negativity and a frontal negativity for the ungrammatical sentences. The final significant time frame seems to be due to a largely distributed positivity. A comparison of the difference waves for the native and German speakers at electrode Pz can be seen in Figure 7.
190 Laura Sabourin and Marco Haverkort
Figure 4.The difference waves for the ungrammatical minus the grammatical condition for the NP definite sentences. The P600 component as seen at electrode Pz for the native speakers (the darker line) and the German speakers. The y-axis represents a voltage of ±8 microvolts with positive plotted up.
In the case of the definite NP where gender agreement between the overt determiner and noun can be seen as equivalent to gender assignment to nouns, which is similar in Dutch and German, the Germans are able to perform well off-line and their on-line processing looks very similar to that of native speakers, although the P600 component found for the ungrammatical sentences occurs later and is more restricted. However, in the case of Dutch indefinite NPs, where agreement can only occur at a more purely syntactic level, since more than just knowing whether an item is common or neuter is required, the German speakers have quite a bit of difficulty in the off-line judgment, and in looking at processing of only the items they correctly judged in the on-line task they do not show a P600 component.
Neural substrates of representation and processing of a second language
Figure 5.ERP wave patterns for the grammatical (darker line) versus ungrammatical sentences for the native speakers in the NP indefinite condition. The y-axis represents a voltage of ±5 microvolts with positive plotted up.
5. General conclusions The main goal of this paper was to show that in the study of language behaviour in general and in the field of second language acquisition in particular the representation of grammatical knowledge and the processing in which this knowledge is employed need to be carefully distinguished. We have shown that for the phenomenon of grammatical gender and for one specific group of second language learners this is indeed an important distinction, because this group shows a difference in processing at least for the indefinite NPs. While quantitative differences in language processing can be seen in aphasic populations, when compared with unimpaired language users, we have shown that there is actually a qualitative difference between native speakers and second language learners in the processing of language. The results presented in this
191
192 Laura Sabourin and Marco Haverkort
Figure 6.ERP wave patterns for the grammatical (darker line) versus ungrammatical sentences for the German group in the NP indefinite condition. The y-axis represents a voltage of ±5 microvolts with positive plotted up.
paper suggest that the German participants may be making use of a translation strategy to learn Dutch gender assignment knowledge and that they may be using their L1 processing strategies to process their L2 for cases where the grammars are very similar. Further support for this is seen in the ERP processing patterns for sentences involving subject-verb agreement and finiteness, both phenomena for which the Germans show native-like knowledge in the off-line task. For finiteness structures, for which the German group can translate their L1 strategies, they show native-like processing but for subject-verb agreement, which exists in German but is different in this language, the processing is different from that observed in native speakers (see Sabourin 2003). Another important thing to note is that the very clear late frontal negativity which was seen in the ERPs of the native speakers was not found for the German group even for the NP definite condition for which a P600 effect can be seen.3 These
Neural substrates of representation and processing of a second language 193
Figure 7.The difference waves for the ungrammatical minus the grammatical condition for the NP indefinite sentences. The P600 component as seen at electrode Pz for the native speakers (the darker line) and the German speakers. The y-axis represents a voltage of ±8 microvolts with positive plotted up.
findings suggest that linguistic processing (as reflected by the P600) can only occur in the L2 when the processing strategy from the L1 can be used relatively directly in L2 processing. Processing of grammatical structures that are not similar in the L1 and L2 may be learned and handled by more general cognitive strategies. Ullman (2001) discussed the declarative/procedural memory distinction in terms of L1 processing, lexical knowledge being declarative and syntactic knowledge being procedural in nature. For L2 speakers, Ullman claims that both lexical and syntactic knowledge usually rely on declarative memory, although Ullman suggests that factors such as age of exposure and practice may influence the ability to use procedural memory in L2 learners. Using this terminology the results presented here suggest that in L2 processing of grammatical knowledge, only in the case where the L1 and the L2 are similar can procedural memory be used by advanced adult L2 learners although probably with a quantitative difference compared to native speakers, viz. a temporal delay.
194 Laura Sabourin and Marco Haverkort
Further research should shed light on whether this non-native-like processing is an across-the-board second language effect or whether differences may be found depending on the particular phenomena studied and particular language groups involved.
Notes * We would like to thank Laurie Stowe, John Hoeks, Liz Temple and two anonymous reviewers for their comments on an earlier draft of this paper. The research of the first author was funded by the School of Behavioral and Cognitive Neurosciences (BCN) of the University of Groningen; the research of the second author was funded by a grant from the Royal Netherlands Academy of Sciences (KNAW). 1. For more details on the proficiency test see Sabourin (2001, 2003). 2. Participants were asked to try their best to not blink during presentation of the sentences as eye movements and blinks greatly distort the EEG signal. 3. This suggests that whatever the L2 speakers are doing, it is not exactly the same as the native speakers.
References Avrutin, S., Haverkort, M. and Van Hout, A. 2001. “Introduction: Language acquisition and language breakdown”. Brain and Language 77: 269–273. Burkhardt, P., Piñango, M. and Wong, K. 2001. The role of the anterior left hemisphere in real-time sentence comprehension: Evidence from split intransitivity. Ms. Yale University. Burnage, G. 1990. A guide for users. Nijmegen: CELEX Centre for Lexical Information. Caplan, D. and Hildebrandt, N. 1988. Disorders of syntactic comprehension. Cambridge: MIT Press. Chomsky, N. 1995. The minimalist program. Cambridge: MIT Press. Grodzinsky, Y. 1990. Theoretical perspectives on language deficits. Cambridge: MIT Press. Grodzinsky, Y. and Finkel, L. 1998. “The neurology of empty categories: Aphasics’ failure to detect ungrammaticality”. Journal of Cognitive Neuroscience 10 (2): 281–292. Gunter, T. C., Friederici, A. D. and Schriefers, H. 2000. “Syntactic gender and semantic expectancy: ERPs reveal early autonomy and late interaction”. Journal of Cognitive Neuroscience 12 (4): 556–568. Haarmann, H. 1993. Agrammatic aphasia as a timing deficit. Doctoral dissertation, University of Nijmegen. Hagoort, P. and Brown, C. 1999. “Gender electrified: ERP evidence on the syntactic nature of gender processing”. Journal of Psycholinguistic Research: Special Issue on “Processing of Grammatical Gender” 28 (6): 715–728.
Neural substrates of representation and processing of a second language 195
Hagoort, P., Brown, C. and Groothusen, J. 1993. “The syntactic positive shift (SPS) as an ERP-measure of syntactic processing”. Language and Cognitive Processes 8: 439–483. Kolk, H. 1995. “A time-based approach to agrammatic production”. Brain and Language 50: 282–303. Kolk, H. 1998. “Disorders of syntax in aphasia: Linguistic-descriptive and processing approaches”. In Handbook of neurolinguistics, B. Stemmer and H. Whitaker (eds), 250–260. San Diego: Academic Press. Kolk, H. 2002. Language production in agrammatic aphasics: an experimental study. Paper presented at the University of Nijmegen Linguistics Colloquium. Kutas, M. 1993. “In the company of other words: Electrophysiological evidence for singleword and sentence context effects”. Language and Cognitive Processes 8 (4): 533–632. Linebarger, M., Schwarz, M. and Saffran, E. 1983. “Sensitivity to grammatical structure in socalled agrammatic aphasics”. Cognition 13: 361–392. Osterhout, L. and P. J. Holcomb. 1992. “Event-related brain potentials elicited by syntactic anomaly”. Journal of Memory and Language 31: 785–806. Rugg, M. D. and Coles, M. G. H. 1995. “The ERP and cognitive psychology: Conceptual issues”. In Electrophysiology of the mind: Event-related brain potentials and cognition, M. D. Rugg and M. G. H. Coles (eds), 27–39. Oxford: Oxford University Press. Sabourin, L. 2001. “L1 effects on the processing of grammatical gender in L2”. In Eurosla Yearbook, Volume 1, S. Foster-Cohen and A. Nizegorodcew (eds), 159–169. Amsterdam: John Benjamins. Sabourin, L. 2003. Grammatical gender agreement in L2 processing. Doctoral dissertation, University of Groningen. Ullman, M. T. 2001. “The neural basis of lexicon and grammar in first and second language: The declarative/procedural model”. Bilingualism: Language and Cognition 4 (1): 105–122.
Chapter 9
Neural basis of lexicon and grammar in L2 acquisition The convergence hypothesis* David W. Green University College London
1.
Introduction
In acquiring a second language (L2) individuals must grasp its grammar and its vocabulary but differences in the context of acquisition from a first acquired language (L1) may mean that different learning mechanisms are involved. Such differences, in turn, carry implications for the neural representation of L1 and L2. Understanding the neural basis of the representation of L1 and L2 can therefore contribute to a deeper understanding of the interface of syntax and lexicon in L2 acquisition. The basic orienting question is this: is a person’s lexical and grammatical knowledge represented differently if it is learned as an L2 as opposed to an L1? In this chapter, I contrast this proposal, termed the differential representation hypothesis, with an alternative, termed the convergence hypothesis. This hypothesis states that as proficiency in L2 increases, non-native speakers represent, and process, the language in the same way as native speakers of that language. In addressing these hypotheses, I consider both their computational basis, and relevant neuropsychological and neuroimaging data. The chapter is structured as follows: I first consider the differential representation hypothesis. On one version an L2 is represented in a person’s righthemisphere rather than in their left-hemisphere. I reject this possibility. Next I consider a version in which in L2, in contrast to L1, grammatical and lexical information is represented in a common memory system. This version of the hypothesis relies on a distinction between two memory systems that has been justified primarily by studies on amnesic patients. However, computational modelling shows that a single memory system is sufficient to generate the data
198 David W. Green
of these amnesic patients. Such a result prompts consideration of the computational basis for the alternative, convergence hypothesis. By themselves computational arguments are not decisive and so we consider neuropsychological and neuroimaging data in an effort to adjudicate between the two hypotheses empirically. A final section considers the kind of studies needed to further our understanding of the issue.
2. A distinct hemispheric representation of L2? The human brain is a product of evolution and it makes computational sense for an evolved system to have redundant and duplicate mechanisms for performing tasks (Edelman 1989). A strong version of the differential representation hypothesis is therefore possible in which L2 is represented in a completely distinct neuroanatomical substrate from L1 (Scoresby-Jackson 1867). Language functions in monolingual, right-handed individuals are typically represented in a distributed left-hemisphere network. 91% of right-handed participants showed left-hemisphere dominance for language in a study in which they were injected with a barbiturate of sodium amytal either into the right, or into the left, carotid artery (Loring, Meador, Lee et al. 1990). Sodium amytal causes an anaesthesia for 1 to 2 minutes in the cerebral hemisphere on the same side as the injection. If language is lateralised to that hemisphere the person will be unable to speak. Less invasively, in a large functional imaging study, 94% of right-handed participants showed left-hemisphere dominance for language (Springer, Binder, Hammeke et al. 1999). These data are consistent with the notion that the left-hemisphere contains circuitry specialised for language processing. Nonetheless, in principle, L2 might be represented in homologous areas of the right-hemisphere (Albert and Obler 1978). However, Rapport, Tan and Whitaker (1983) in a study of right-handed polyglot aphasics prior to surgery found no evidence of the disruption of picture naming following intracarotid injection of sodium amytal into the right-hemisphere. In contrast, naming was massively disrupted following injection into the lefthemisphere. Further, in a study of 88 reported cases of right-handed bilingual aphasics, Fabbro (1999: 210–211) found that only 8% presented with a lesion to the right-hemisphere. Taking into account reporting biases, he concluded that the incidence of aphasia in bilinguals with right-hemisphere lesions is not in fact higher than that shown by monolingual aphasics. These observations suggest that both L1 and L2 are represented in a common substrate in the left-hemisphere
Neural basis of lexicon and grammar in L2 acquisition 199
though perhaps with different microanatomical representations (Paradis 2001). The critical question then concerns the neural representation of lexical and grammatical knowledge for L2 within this hemisphere. I consider a more subtle version of the differential representation hypothesis in the next section.
3. The specific representation of lexicon and grammar Researchers have taken different views on the extent to which the lexicon and grammar of a language are subserved by distinct neural mechanisms that are language-specific. According to one view, words are processed in one dedicated, posterior system and grammar is processed in another dedicated anterior system (e.g. Chomsky 1995, Pinker 1994). An alternative view is that these two components of the language system are indeed mediated by distinct neural mechanisms but that these mechanisms are not in fact specific to language. Ullman (2001a) proposed that the lexicon is stored in a neural system that subserves declarative memory in general. By contrast, grammar is represented in a procedural memory system that is implicated in the learning of motor and cognitive skills in general. The declarative memory system is held to be involved in the learning of facts and events and to be particularly important in the learning of arbitrarily related information from different sources such as the associations between the sounds of words and their meanings. Information in this system is available for ‘explicit’ (i.e. conscious) recollection. Whilst initial learning may depend on the medial temporal structures (e.g. the hippocampus), neocortical regions subsequently become the principal site of representation (e.g. temporo-parietal region). In contrast, the use of grammar (including syntax, morphology, and phonology) is achieved by a system that underlies the performance of motor skills in general — a procedural system mediated by structures in the frontal cortex and basal ganglia and the inferior parietal region (Squire, Knowlton and Musen 1993, Squire 1994). The nondeclarative or procedural system is held to influence behaviour ‘implicitly’, i.e. in the absence of conscious recollection. Hence the contrast between declarative and procedural (or nondeclarative) memory systems is sometimes referred to, in summary terms, as a contrast between explicit and implicit memory systems (see, for example, Paradis, 1994). But this identification cannot be taken too far. The acquisition of vocabulary for instance is not simply a matter of declarative memory. Gupta and Dell (1999) argue that the learning of vocabulary involves the explicit learning of the
200 David W. Green
relationship between a phonological representation and meaning and the implicit learning of the mapping of input phonemes onto an articulatory chain (see also Ellis 1995 and Segalowitz and Segalowitz 1993). Consistent with this view, Paradis (1997: 333–334) argues that the acquisition of vocabulary is a partially explicit process. Likewise Lebrun (2002: 304) argues that common words and phrases are not only stored neocortically but as verbomotor subcortical patterns. But the point to note here is that this contrast applies equally to L1 and L2 vocabulary learning. The distinction between declarative and nondeclarative memory systems has been exploited as a means to contrast the neural representation of L1 and L2. Consider one possible difference between the acquisition of L1 and L2. The representation of a language acquired in an oral, conversational setting (e.g. Quebecois; Friulian) may differ from one acquired in the formal setting of a school. In particular, there may be a difference in the representation of grammar and morphology (morphosyntax). Individuals who acquire L1 in a conversational setting achieve proficiency in morphosyntax implicitly. In contrast the grammatical rules in the school setting are part of an explicit declarative knowledge. Maturational constraints may also affect the acquisition of morphosyntax more than the acquisition of vocabulary (Paradis 1994: 398). In consequence, Paradis (1994) has argued that L1 and L2 may load differently on these two memory systems. L1, especially its morphosyntax, but also its lexicon to an extent, may load more on the implicit, procedural memory system whereas L2 may load more on an explicit, declarative memory system. Ullman (2001b) has made a related proposal based on the notion that linguistic abilities are sensitive to the age of exposure to the language (Lenneberg 1967). It is generally considered that attainment in L2 is constrained by the age at which learning begins. For instance, there is a negative correlation between the age at which learning begins and eventual performance (Johnson and Newport 1989). But not all language capacities are affected equally: the use of grammar is more adversely affected than the use of lexical items. As a result, in L2 acquisition, there is a specific shift in processing of grammatical computation from the procedural memory system to the declarative memory system (Ullman 2001b: 108; Note 2: 110 contrasts his proposal with that of Paradis). There is no shift for lexical processes. These are held to depend on the declarative memory system for both L1 and for L2. I take this version of the differential representation hypothesis and consider it in a little more detail. In Ullman’s (2001b) view, vocabulary in both L1 and in L2 is represented in a declarative memory system — in the form perhaps of an associative
Neural basis of lexicon and grammar in L2 acquisition 201
network linking meanings and sounds. By contrast, whereas grammatical processing in an L1 (e.g. forming the past-tense of a regular English verb such as walk by adding -ed to the stem) relies on a procedural system, grammatical processing in an L2 (such as English), is achieved declaratively. The basic notion is that linguistic forms that are compositionally computed in L1 are memorized in L2 as if they were words or idioms. Given that the associative lexical memory can generalize patterns, such a system can still be productive. Certain rules may also be learned, though these will differ in type from any implicitly learned rules of L1. Ullman acknowledges that age of exposure to L2 is not the only factor affecting the dependence on declarative memory: “even older learners may show a degree of dependence on procedural memory if they have had a large amount of practice — that is, a fairly substantial amount of use of the language” (Ullman, 2001b: 110). But the clear implication is that even proficient speakers of L2 will differ from native speakers of that language in relying much more on declarative memory for grammatical computations. There are two elements to Ullman’s (2001b) proposal. First, it is motivated by the claim that linguistic abilities are sensitive to the age of exposure. Second, it appeals to two distinct types of memory that have been inferred from research on amnesic patients. The following section considers each element.
4. Some grounds for doubt Ullman (2001b) motivated the shift towards a declarative representation of grammatical knowledge in L2 by appealing to data (Johnson and Newport 1989) on the limits of L2 (English) attainment for native Korean, and Chinese; speakers who learned English post-puberty (after the age of 17 years). In the Johnson and Newport study, L2 learners were asked to judge whether or not auditorily presented sentences were grammatically correct or not. Roughly half of the sentences were grammatical and half were minimally different ungrammatical variants. A key finding was that age of acquisition was negatively correlated with performance before puberty, but there was no systematic relationship between age of acquisition and performance post-puberty. Further, few if any of the 46 participants in their study, achieved native-like levels of performance post-puberty. These results are consistent with a critical period view of language acquisition (Lenneberg 1976). In contrast to such data, Birdsong and Molis (2001) in a replication of Johnson and Newport, but using 61 native Spanish speakers, found that the age of acquisition did predict
202 David W. Green
attainment in L2 post puberty. They also found evidence of native-like attainment in late learners of L2. They argue, in line with Flege, Yeni-Komshian and Liu (1999), that practice is an important factor in determining the eventual level of attainment. The nature of the L1 and L2 pairing may also be relevant. Turning now to the key element of Ullman’s (2001b) proposal, namely the idea that emphasis is shifted to the declarative memory system in L2 learners and that there is little or no involvement of a procedural memory system in grammatical processing. This proposal presumes that there is good evidence for the existence of these two types of memory systems. Data from amnesic patients appear to provide compelling support. Amnesics have poor declarative memory but show normal performance on various tasks involving nondeclarative memory (Gabrieli 1998). A study by Knowlton and Squire (1993) is exemplary. They used dot patterns created by systematically distorting a prototype pattern. Amnesic patients were able to classify these patterns normally (a nondeclarative task) but were severely impaired in their ability to recognize whether or not a particular pattern had been presented previously (a declarative task). Knowlton and Squire interpreted these data as evidence that performance in the two tasks was mediated by two different memory system one of which (the declarative system) was impaired and the other of which was not. However, computational work by Nosofsky and Zaki (1998) challenged this interpretation by showing that differences in performance on these two tasks can be obtained within a single memory system. A slight reduction in the value of a sensitivity parameter in their computational model reduced classification performance marginally but exerted a marked effect on recognition performance. More pertinent to the present concern is work on the learning of artificial grammars. Knowlton and Squire (1994, 1996) contrasted normals and amnesics in their ability to learn an artificial grammar. In such studies individuals first memorise a set of strings generated by a (finite-state) grammar and are then informed about a set of rules generating the strings. In the classification task, they have to classify a new set of strings into those that are grammatical and those that are not. In the recognition task, they have to indicate whether or not a string of symbols was presented. In the studies by Knowlton and Squire classification performance in amnesic patients was normal but recognition performance was impaired. Knowlton and Squire interpreted this as evidence that the two tasks are mediated by different memory systems — only one of which, the declarative memory system, is impaired in amnesics. Other studies support a dissociation between recognition and repetition priming in amnesic patients (e.g. Hamann and Squire 1997: Experiment 1).
Neural basis of lexicon and grammar in L2 acquisition 203
Repetition priming refers to the improvement in the identification, detection or production of a stimulus as result of having experienced it previously. It is considered to be mediated by nondeclarative memory because repetition priming occurs even when there is no conscious recollection of the prior experience of the stimulus (Gabrieli 1998). Priming is held to be mediated by neocortical structures that are spared in amnesic patients (e.g. McClelland, McNaughton and O’Reilly 1995). Hamann and Squire (1997: Experiment 1) presented amnesic patients and controls in a priming phase with a set of four-letter consonant strings for three seconds each. In a later study phase, these stimuli and others were presented for 170 ms each and participants had to identify them. Priming was operationalised as the difference in the identification of old and new strings. After two priming and study phases participants were tested for their recognition of stimuli presented in the study phases. The recognition test consisted of pairs of old and new stimuli and participants had to decide which string was old. Amnesics showed the same degree of priming as the normal controls but their recognition performance was at chance. But do these dissociations (between recognition and classification and between recognition and repetition priming) require us to postulate two distinct memory systems? On the basis of simulation data, Kinder and Shanks (2001) argue that they do not. They used a simple recurrent network (see Cleeremans 1993) to simulate performance in artificial grammar learning. In order to differentiate an amnesic network from a normal network they reduced the learning rate during acquisition and, in a separate simulation, reduced the number of hidden units prior to test. A change in either parameter was sufficient to induce a dissociation between classification and recognition. In a second set of simulations, they showed that a simple recurrent network could also simulate the dissociation between recognition and repetition priming. These simulation results show that the dissociations observed in the clinical population do not require a dual-memory system. Instead, such results are consistent with a single system or network and so weaken support for the procedural/declarative distinction at the heart of the differential representation hypothesis. Although these simulation results lead us to be wary of using performance differences as direct evidence of different cognitive and neural systems,1 computational results only provide an existence proof and do not establish that the brain does not in fact have distinct declarative and nondeclarative memory systems. Further, I know of no computational studies aimed at showing whether or not the kinds of specific dissociations referenced
204 David W. Green
by Ullman can emerge from single networks (see also Ullman 2001b, Note 1:10). At a minimum, such results encourage the search for an alternative formulation. In fact, computational considerations lead us to expect a rather different outcome for L2 acquisition. The next section considers a computational justification for the convergence hypothesis. 5. The convergence hypothesis and its computational basis According to the convergence hypothesis, any qualitative differences between native speakers of a language and L2 speakers of that language disappear as proficiency increases. Such a hypothesis is broadly in line with the idea that proficiency in language involves identifying, and using, the various cues to meaning — see, for example, the competition model (MacWhinney 1997). The convergence hypothesis is specifically concerned with the neural, and not simply the cognitive, representation of L2. Given the diversity of languages, in order to consider the computational grounds for the convergence hypothesis, we need to characterise languages at a certain level of abstractness. There are four linguistic means for communicating experience (see, for example, Tomasello 1995): individual symbols (lexical items); markers on symbols (grammatical morphology), ordering patterns of symbols (word order) and prosodic variations of speech (e.g. stress, intonation, timing). Languages differ in the weight they attach to these different linguistic means. In some languages, word order is basically free and information on ‘who did what to whom’ is conveyed by word endings or by prosody in tone languages. By contrast, in English, such information is conveyed by word order and this is relatively rigid. These different linguistic means or signals require different devices for their processing. The first step of the computational argument for the convergence hypothesis is that the neural representation of the various linguistic devices is similar across languages. The next paragraph spells out the basis for this argument. Such devices may be represented by specific networks with a distinct neural anatomical representation or they may be mediated by a specialised network with a distributed neural representation. Specialised networks can emerge from unique interactions amongst a set of regions each fulfilling a number of different functions (e.g. Mesulam 1990). Consider the development of a system using these devices to communicate meaning. First, different neural regions may compete to process input. Those regions, whether innately specified, or possessing some small processing advantage, will come to mediate processing of a given linguistic means. Neural regions active at the same time will connect
Neural basis of lexicon and grammar in L2 acquisition 205
together (Hebb 1949, Robertson and Murre 1999) giving rise to a specialised network. Second, once a network has come to process signals of a particular type it will resist processing other types of signals unless input to it is curtailed in which case it may process signals of another type given that plastic reorganization is possible. One line of support for this view is evidence of ‘crowdingout’: when language is displaced to the right-hemisphere as a result of neurological damage to the left-hemisphere, a person’s visuo-spatial skills (typically mediated by the right-hemisphere) are impaired (Teuber 1974, Strauss, Satz and Wada 1990). Third, given commonalities across brains in the initial sensitivities of different regions — call this the commonality assumption2 — there will be commonalities in the neural representation of the different devices for speakers of different languages. The second step of the computational argument for the convergence hypothesis is that the acquisition of an L2 arises in the context of an already specified, or partially-specified system, with a specific neural network mediating each device.3 It follows that L2 will receive convergent representation with L1. Further, given the commonality assumption (see above) the representation of L2 will converge with the representation of that language learned as an L1. The convergence hypothesis does not entail that a speaker of L2 will necessarily achieve native-like levels of performance (for example in achieving certain phonetic norms, Flege 1995) nor does it exclude the possibility that tasks such as mental arithmetic are carried out exclusively in L1. Clearly, also, the contexts of acquisition (e.g. a formal school setting versus an immersion setting) affect the initial registration of linguistic information. However, in contrast to the differential representation hypothesis, the convergence hypothesis is committed to the prediction that as proficiency in L2 increases, the same linguistic means involve the same neural networks as native speakers. The hypothesis would be refuted if there is no change in representation with proficiency and if a normal, proficient L2 speaker activated neural networks disjoint from those of a native speaker, especially when encoding and decoding syntactic information. The fact that explicit, declarative representations of grammatical information, play only an initial role in on-line processing, according to the convergence hypothesis, does not mean that they are unimportant. Explicit (metalinguistic) representation may well benefit the recovery of L2 over L1 following brain-damage (e.g. Lebrun 2002, Paradis 1994, 1997). But such a possibility, it seems to me, cannot be used to claim a continuing role for such representations in on-line processing, once the relevant procedures are in place. However, this possibility is open to test.
206 David W. Green
Of course, certain differences in processing profiles and neural activation are to be expected when L2 speakers are contrasted with monolingual speakers of that language. The acquisition of an L2 carries consequences. Alternative means for expressing communicative intentions can induce competition both in production (e.g. Bialystok 1992, Gollan and Kroll 2001, Green 1986, 1998, Hermans 2000) and in comprehension (e.g. Dijkstra, van Jaarsveld and ten Brinke 1998, de Groot, Delmaar and Lupker 2000) though the range of conditions under which this occurs is unknown. Depending on how the system is controlled there may be a difference in processing profiles despite convergence. But there will be a marker for such an effect: increased competition (and hence increased activation perhaps) in the areas associated both with lexical and with grammatical encoding will be associated with increased activation in the areas associated with language control. Such effects will be apparent both in L2 and in L1. The next section considers empirical data with a view to adjudicating between the differential representation and convergence hypotheses.
6. Empirical data: Can we adjudicate? Both neuropsychological and neuroimaging studies provide data that may help in adjudicating between the two hypotheses. Under the latter we include EventRelated Potential (ERP) data and haemodynamic methods (Positron Emission Tomography, PET and functional Magnetic Resonance Imaging, fMRI (Note 4 briefly describes these classes of method). We first consider evidence for the distinct representation of lexical and grammatical information in L1 and then consider what we can infer from neuropsychological and neuroimaging studies of bilingual speakers. 6.1 The representation of L1 Both neuropsychological and neuroimaging data suggest that there is a degree of specialisation within monolingual speakers for syntactic and semantic processes. For instance, Breedin and Saffran (1999) reported a patient, D. M., who was good at detecting grammatical violations despite a pervasive loss of semantic knowledge. ERP data from normal individuals also indicate that there are distinct mechanism mediating at least post-lexical syntactic and semantic processes (Hagoort, Brown and Osterhout 2000). For instance, N400 (found 400 ms after an event) is sensitive to violations of semantic expectancy whereas
Neural basis of lexicon and grammar in L2 acquisition 207
P600 (found 600 ms after an event) is sensitive to syntactic violations. ERP data cannot provide direct evidence of the neural sources of such effects but haemodynamic studies are informative. Studies on grammatical processing and encoding in native speakers (Hagoort, Brown and Osterhout 2000) suggest a common syntactic component subserved by the left frontal area (a dorsal part of Broca’s area and adjacent parts of the middle frontal gyrus) and studies on the semantic representation of words identify regions in the temporo-parietal region — the left extrasylvian temporal cortex and the left anterior inferior frontal cortex (Price 2000). Neuropsychological data (Donkers, Redfern and Knight 2000) and also neuroimaging data (e.g. Price, Moore, Humphreys and Wise 1997) suggest a specific area in the anterior temporal region as a site critical for the interpretation of sentences (i.e. syntactic-semantic integration). 6.2 The representation of L2 What empirical evidence is there that L2 is represented differently from L1 as proposed by the differential representation hypothesis? According to Ullman (2001b), L2 learned late will be sensitive to damage to neocortical temporal/ temporal-parietal regions for those linguistic forms that depend on grammatical processing in L1. A case reported by Ku, Lachmann and Nagler (1996) seems to support his position. A 16 year old native Chinese speaker who had been living in the United States for six years and who had received intensive training in English over this period suffered a circumscribed lesion to the left temporal lobe (as a result of herpes simplex encephalitis). For three weeks following the lesion he lost the ability to comprehend and to speak English. In contrast, naming in Mandarin was normal. However, in speaking Mandarin his syntax was simplified and so this case is not decisive support for the claim that grammatical information is represented differently in L2. The notion that L1 grammatical processing is mediated by a frontal-basal ganglia circuit predicts that damage to the basal ganglia will lead to a selective loss of L1. Fabbro and Paradis (1995) report the case of E. M. with such a lesion and, true to prediction, her spontaneous speech in her L1 (Venetan) was poor whereas her speech was better in her L2 (Italian) that she rarely used prior to the lesion. Ullman (2001b) considered the nature of her errors. There was a similar proportion of word finding difficulties in both languages but a tendency for poorer grammatical performance in L1 (e.g. the omission of grammatical function words in obligatory contexts). However, these effects are small and the overwhelming difference is her spontaneous use of L2 in preference to L1.
208 David W. Green
Green and Price (2001) argue that language control (e.g. the ability to select between one language and another) is also mediated by frontal-basal ganglia circuits. An impairment in this system will also give rise to problems in modulating the output from the lexical-semantic system. Leaving the neuropsychological data on one side, how convincing is the ERP and neuroimaging data for a distinct representation of L2? Individuals acquiring L2 can vary in terms of when they acquired L2, how they acquired L2 and how proficient they are in using it. Typically, proficiency is confounded with the age of acquisition. In terms of proficiency it is natural to expect that less proficient users of L2 will show quantitative differences on a range of measures (e.g. naming time, ERP effects and activation patterns). The critical issue is whether or not there are qualitative differences indicating that different neural mechanisms are involved. If there are, it is important to determine whether these necessarily imply different representations. ERP data point to both quantitative and qualitative differences in processing between L1 and L2. Kutas and Kluender (1991), for instance, found that the N400 component in response to a semantic anomaly was delayed and of lower amplitude in a bilingual’s less fluent language. Likewise, Webber-Fox and Neville (1996) found N400 present in all groups of Chinese-English second language learners though it was more delayed in those learning L2 after reaching the age of 11–13 years. More critically, in contrast, to monolinguals, there was a distinct pattern of response to phrase structure violations in bilinguals. Only individuals acquiring L2 before the age of four showed no difference from native learners of L2. Such data are compatible with the notion that there is a critical period for language learning and are consistent with the notion that different brain mechanisms mediate syntactic processing in late learners of L2. But they are not decisive as later exposure to English was associated with worse performance in identifying syntactically anomalous sentences. Such individuals may have been circumventing syntactic processing. Hahne and Friederici (2001) examined the effects of phrase structure violations and semantic anomaly in Japanese late learners of German. These individuals also showed substantial error rates in a grammaticality judgement task and so cannot be considered proficient. Hahne and Friederici (2001) confirmed a delayed N400 effect in response to semantic anomaly but also found a right anterior central negativity. Unlike native German speakers there was no early anterior negativity in response to a syntactic violation. They propose that late learners identify lexical content independently of morphological form (e.g. the past participle form of the verb) and construct a representation directly
Neural basis of lexicon and grammar in L2 acquisition 209
based on conceptual information. Hahne and Friederici speculated on the source of these effects based on other functional imaging data (Falk, Durwen, Müller et al. 1999, Opitz, Mecklinger and Friederici 2000). They proposed that late learners of L2 (at least their Japanese participants) supplement lexicalsemantic information by using the right prefrontal cortex to construct a semantic-conceptual representation of sentence content. Unfortunately, there is a dearth of functional imaging studies of L2 grammatical processing and encoding. On the production side more generally, Kim, Relkin, Lee and Hirsch (1997) used fMRI to study the representation of L1 and L2 while bilinguals covertly described what they had done the previous day. Half of their sample acquired their L2 in infancy and half after puberty. L1 and L2 were represented in spatially segregated parts of the left inferior frontal cortex (Broca’s area) in late learners but in overlapping parts of Broca’s area in early learners. Regions activated in Wernicke’s area (traditionally linked to language comprehension) overlapped for both groups. Kim et al. concluded that age of acquisition affected neural representation. However, there was no assessment of proficiency in L2 and so we cannot tell whether or not age of acquisition is critical. Late learners could have been less proficient in their L2. In fact, when L2 proficiency is high, Chee, Tan and Thiel (1999) found no difference within the left prefrontal cortex (including Broca’s area) when comparing word generation in early bilinguals (L2 acquired before the age of six) and late bilinguals (L2 acquired after the age of twelve) for Mandarin-English speakers in Singapore. The pattern of brain activation in response to Mandarin words was similar to that observed in response to English words, and did not vary as a function of age of acquisition. Klein, Milner, Zatorre et al. (1995) reached a similar conclusion: a common network of brain areas is engaged in L1 and in L2 in highly-proficient bilinguals despite late acquisition of L2. In terms of comprehension, Abutalebi, Cappa and Perani (2001) concluded that both languages are processed in a single and common left-sided network, comprising all the classical language areas when L2 is acquired early (before the age of five). In contrast for late bilinguals, the degree of language proficiency is the critical factor. Highly proficient late bilinguals activate similar left hemispheric areas for L1 and L2 (Perani, Paulesu, Sebastian-Galles et al. 1998) whereas less proficient subjects have different patterns of activation for their two languages (Perani, Dehaene, Grassi et al. 1996, Dehaene et al. 1997, Price, Green and Von Studnitz 1999). Critically, more extensive activations are associated with the less proficient language (e.g. greater temporal lobe dispersion) perhaps indicating that in comprehending stories individuals process grammatical forms differently.
210 David W. Green
At the macroanatomical level then current functional imaging data indicate that there is little difference in the representation of L1 and L2 for highlyproficient bilinguals. The implication is that age of acquisition is less critical than proficiency. However, as Vaid and Hull (2002) observed, we still need studies that directly compare individuals differing in L2 proficiency.
7. Ways forward The differential representation hypothesis and the convergence hypothesis concur that the initial representation of L2 may differ from that of native speakers of that language. The fundamental research requirement then is to conduct withinparticipant longitudinal studies to chart changes in behaviour on various tasks, e.g. picture naming (Kroll, Michael, Tokowicz and Dufour 2002) and to examine changes in ERP and neuroimaging profiles as proficiency in L2 changes. Further, such studies need to involve both syntactic and lexical tasks. To my knowledge there are currently relatively few such studies. As discussed above, in contrast to L1, grammatical knowledge of L2 may be represented explicitly and declaratively. According to the convergence hypothesis, ERP responses to syntactic anomalies should change with proficiency. Osterhout and McLaughlin (2000) studied responses to semantic and syntactic anomalies in native speakers of French and in novice learners. Semantic anomalies yielded N400 and syntactic anomalies yielded a P600 in native French speakers. French learners after four weeks of instruction showed an N400 in response to semantic anomalies. In contrast, there was either an N400 or no effect in response to syntactic anomalies. After just four months, however, syntactic anomalies yielded a P600 but no N400. These data suggest that if there are qualitative differences between native speakers and L2 learners, these can be rather short-lived. Any differences in responding to syntactic anomalies are presumably negatively correlated with the increasing grasp of syntax. Consistent with the convergence hypothesis, Weber-Fox and Neville (submitted, cited in Ullman 2001b) examined responses to open-class and closed-class words. Native speakers of English showed a left anterior negativity to closed class words (N280) and an N400 for open-class words. L2 speakers of English (with Chinese as their first language) showed the same open class N400 as native English speakers. Interestingly, the response to closed class words related to an independent test of their grammatical ability. The higher the score on the test, the earlier the anterior negativity for closed class words.
Neural basis of lexicon and grammar in L2 acquisition
It is important to extend longitudinal investigation to examine the neural correlates of parsing in more detail. A number of behavioural studies have examined the extent to which L2 learners of English show an influence of their L1 on resolving local syntactic ambiguities of various kinds (e.g. Frenck-Mestre and Pynte 1997, Juffs 1998). The behavioural picture is complex but not inconsistent, in my view, with the convergence hypothesis (see Kroll and Dussias in press for a recent review). In examining the extent to which L2 parsing profiles converge with those of native speakers it is possible that we will need to identify instances where an inappropriate parse leads to a high cost in recovering the intended interpretation. After all, if the intended interpretation can be recovered quickly, what computational constraint is there for the neural processing profile of an L2 learner to converge with that a native speaker? On the other hand, the processing cost of recovery may be a function of language background. As Juffs (1998: 135) proposes, languages (e.g. Japanese, Korean) with Subject Object Verb structure may lead speakers to become adept at recovery from garden-paths. Speakers of these languages may routinely make parsing decisions about theta-roles that must be revised on encountering the verb. Regardless of these kinds of possibilities, the convergence hypothesis predicts that for proficient speakers of L2, sentence interpretation will activate an area in the anterior temporal pole and this area, as in the case of L1 (Noppeney and Price submitted), will show evidence of syntactic priming, i.e. reduced activation in circumstances where the same syntactic structure is repeated either within- or between-languages. The achievement of proficiency also entails attending to the world in the manner of native speakers so that lexical and grammatical processes can be coordinated appropriately (Black and Chiat 2003, Levelt, Roelofs and Meyer 1999, Slobin 1996). It is this pattern of coordination that also needs to be considered. Neuroimaging allows us to consider how areas work together. Büchel, Friston and Frith (2000: 339), for instance, describe methods for examining effective connectivity (“the influence one neuronal system exerts on another”) using structural equation modelling of the patterns of activation in different regions of interest. We should expect convergence of these patterns of effective connectivity with those of native speakers of the language as proficiency increases. By way of illustration, consider differences between languages in the way they package together different aspects of a movement. English packages manner and motion together (hop, float) and, unlike some other languages (e.g. Spanish), has fewer verbs that express motion and direction together (e.g. rise, fall). In order to select the correct verb in Spanish an English speaker must
211
212 David W. Green
explicitly encode the direction of motion — i.e. make that property of the scene salient rather than the manner of motion. Proficiency should be associated with changes in the explicit representation of the properties of scenes. In terms of activation patterns, there should be changes in the activation and connection of cortical regions mediating those properties so that the correct words can be selected and expressed in a suitable syntactic frame.5
8. Conclusion This chapter has contrasted two main hypotheses about the representation and processing of lexicon and grammar in L2. The subtle form of the differential representation hypothesis proposes that declarative representations play a much more important role in the representation of grammar in L2 than in L1. This chapter has considered the computational and empirical basis for the differential representation hypothesis and argued, on both computational and neuroimaging grounds, for an alternative, convergence hypothesis. According to this hypothesis, as proficiency in L2 increases, the networks mediating L2 converge with those mediating language use in native speakers of that language. Current evidence marginally favours the convergence hypothesis. However, we lack appropriate longitudinal studies of L2 acquisition. Crucially, we have little or no information about the functional integration of different neural regions during second language use. Scope indeed for discovery!
Notes * I thank the editors of this volume for the opportunity to contribute to our understanding of the interface between syntax and the lexicon in L2 acquisition and for constructive comments on a previous draft of this chapter. 1. It would be useful to have converging evidence for the existence of distinct neural systems beyond that offered by amnesia. Fortunately, we do not need to rely exclusively on neuropsychological data. Neocortical activity is reduced for stimuli that have been processed before (Ungerleider, 1995) and this datum has been interpreted as evidence that these structures mediate priming. But amnesics can also be impaired on priming in certain conditions (Ostergaard, 1999). 2. The commonality assumption is compatible with anatomical variability. Brains differ in the location of quite major features (e.g. Rickard 2000). The commonality assumption refers to the sensitivity of neural regions, not to their precise anatomical location. Neuroimaging evidence for convergence must take such variability into account.
Neural basis of lexicon and grammar in L2 acquisition
3. At the lexical level, the impact of a prior representation is captured by cognitive models such as the revised hierarchical model (Kroll and Stewart 1994) and the distributed features model (e.g. De Groot 1993, Kroll and de Groot 1997). It might also be argued that prior representation of L1 induces a radically distinct representation of L2. Jiang and Forster (2001) propose, on the basis of experimental evidence, that lexical items in L2 (they tested native Chinese speakers with English as the L2) are represented in a non-lexical memory. This memory allows the meaning of translation equivalents to be retrieved indirectly via the L1 lexical item. However, it is unclear what it means for a non-lexical system to represent the syntactic properties of lexical items. Their findings, as they acknowledge, need to be replicated with proficient L2 speakers and with a different language pairs. 4. ERPs provide high resolution temporal evidence of the existence of different processes during language processing. These are derived by averaging signals from an eletroencephalogram over a series of trials that time-locked to the presentation of a particular type of stimulus. An ERP itself comprises various components (i.e. positive and negative voltage peaks). Where these are affected by some experimental manipulation they are termed ERP effects. ERP data are compatible with an infinite number of neural generators (the “inverse problem”) and so they need to be complemented by data from haemodynamic methods. Haemodynamic methods (Positron Emission Tomography, PET; or functional Magnetic Resonance Imaging, fMRI) rely on a close coupling between changes in the activation of a population of neurons and change in blood supply. A haemodynamic effect arises only when there is a change in the overall metabolic demand in a neuronal population. PET and fMRI track different signals. PET measures the decay of a short-lived isotope which accumulates in a neural region in proportion to the amount of blood flowing through that region. The most typical fMRI method indexes metabolic demand and hence relative neural activity by assessing the ratio of deoxy- to oxyhaemoglobin in the blood (see Rugg (2000) for a critical appraisal of these methods). It is worth noting that haemodynamic methods, along with other electrophysiological methods, allow us to identify regions that are sufficient for task performance but they do not allow us to identify regions that are necessary for task performance. Other data are needed to identify which regions are necessary. For instance, if task performance is impaired in a patient with a lesion at given site then this region, or the network of which it is part, is necessary for task performance. Likewise, “virtual” lesions induced by drugs or by transcranial magnetic stimulation, may help identify regions necessary for task performance. 5. Lower levels of proficiency in L2 might also be associated with more reliance on conceptual/pragmatic information (see Hahne and Friederici above). In native speakers, conceptual factors affect grammatical encoding (Vigliocco and Hartsuiker 2001) and it is reasonable to expect that such effects might be more marked in novice learners. A sentence completion task offers one behavioural measure. Vigliocco and Franck (2001) showed that there were more errors in generating a predicate when the sex of the referent was incongruent with the gender of the noun. Given that novice learners of French or Italian, for instance, know the syntactic gender of the noun, they might also show greater effects of incongruity.
213
214 David W. Green
References Abutalebi, J., Cappa, S. F. and Perani, D. 2001. “The bilingual brain as revealed by functional imaging”. Bilingualism: Language and Cognition 4: 179–190. Albert, M. L. and Obler, L. K. 1978. The bilingual brain: Neuropsychological and neurolinguistic aspects of bilingualism. New York: Academic Press. Bialystok, E. 1992. “Selective attention in cognitive processing: The bilingual edge”. In Cognitive processes in bilinguals, R. J. Harris (ed.), 501–513. Amsterdam: Elsevier Science Publishers B. V. Birdsong, D. and Molis, M. 2001. “On the evidence for maturational constraints in secondlanguage acquisition”. Journal of Memory and Language 44: 235–249. Black, M and Chiat, S. 2003. “Noun-verb dissociations: a multi-faceted phenomenon”. Journal of Neurolinguistics 16: 231–250. Breedin, S. D. and Saffran, E. M. 1999. “Sentence processing in the face of semantic loss: A case study”. Journal of Experimental Psychology: General 128: 547–562. Büchel, C., Frith, C. and Friston, K. 2000. “Functional integration: Methods for assessing interactions amongst neuronal systems”. In The neurocognition of language, C. M.Brown and P. Hagoort (eds), pp. 337–355. Oxford: Oxford University Press. Chee, M. W. L., Tan, E. W. L. and Thiel, T. 1999. “Mandarin and English single word processing studied with functional Magnetic Resonance Imaging”. Journal of Neuroscience 19: 3050–3056. Chomsky, N. 1995. The minimalist program. Cambridge, MA: MIT Press. Cleeremans, A. 1993. Mechanisms of implicit learning: Connectionist models of sequence processing. Cambridge, MA: MIT Press. Dehaene, S. D., Dupoux, E., Mehler, J., Cohen, L., Paulesu, E., Perani, D., van de Moortele, P. F., Lehéricy, S. and Le Bihan, D. 1997. “Anatomical variability in the cortical representation of first and second languages”. Neuroreport 8: 3809–3815. De Groot, A. M. B. 1993. “Word-type effects in bilingual processing tasks: Support for a mixed representational system”. In The bilingual lexicon, R. Schreuder and B. Weltens (eds), 27–51. Amsterdam: John Benjamins. De Groot, A.M.B., Delmaar, P. and Lupker, S.J. 2000. “The processing of interlexical homographs in translation recognition and lexical decision: Support for non-selective access to bilingual memory”. Quarterly Journal of Experimental Psychology 53A: 397–428. Dijkstra, A., van Jaarsveld, H. and ten Brinke, S. 1998. “Interlingual homograph recognition: Effects of task demands and language intermixing”. Bilingualism: Language and Cognition 1: 51–66. Dronkers. N. F., Redfern, B. B. and Knight, R. T. 2000. “The neural architecture of language disorders”. In The new cognitive neurosciences, M. S. Gazzaniga (ed.), 949–960. Cambridge, Mass: MIT. Edelman, G. M. 1989. The remembered present: A biological theory of consciousness. New York: Basic Books. Ellis, N. C. 1995. “The psychology of foreign language vocabulary acquisition: Implications for CALL”. Computer Assisted Language Learning 8: 103–128.
Neural basis of lexicon and grammar in L2 acquisition
Fabbro, F. 1999. The neurolinguistics of bilingualism: An introduction. Hove, Sussex: Psychology Press. Fabbro, F. and Paradis, M. 1995. “Differential impairments in four multilingual patients with subcortical lesions”. In Aspects of bilingual aphasia, M. Paradis (ed.), 139–176. Oxford, UK: Pergamon. Falk, A. R., Durwen, H. F., Müller, C., König, M., Müller, E. and Heuser, L. 1999. “Determination of eloquent cortical areas in Russian bilinguals performing a word generation task”. Neuroimage 6: S1002. Flege, J. E. 1995. “Second-language speech learning: Theory, findings and problems”. In Speech perception and linguistic experience: Issues in cross-language research, W. Strange (ed.), 233–277. Timonium, MD: York Press. Flege, J. E., Yeni-Komshian, G. H. and Liu, S. 1999. “Age constraints on second-language acquisition”. Journal of Memory and Language 41: 78–104. Frenck-Mestre, C. and Pynte, J. 1997. “Syntactic ambiguity resolution while reading in second and native languages”. Quarterly Journal of Experimental Psychology 50A: 119–148. Gabrieli, J. D. 1998. “Cognitive neuroscience of human memory”. Annual Review of Psychology 49: 87–115. Gollan, T. H. and Kroll, J. F. 2001 “Lexical access in bilinguals”. In A handbook of cognitive neuropsychology: What deficits reveal about the human mind, B. Rapp (ed.), 321–345. New York: Psychology Press. Green, D. W. 1986. “Control, activation and resource: A framework and a model for the control of speech in bilinguals”. Brain and Language 27: 210–223. Green, D. W. 1998. “Mental control of the bilingual lexico-semantic system”. Bilingualism: Language and Cognition 1: 67–81. Green, D. W. and Price, C. 2001. “Functional imaging in the study of recovery patterns in bilingual aphasics”. Bilingualism: Language and Cognition 4: 191–201. Gupta, P. and Dell, G. S. 1999. “The emergence of language from serial order and procedural memory”. In The emergence of language, B. MacWhinney (ed.), 447–481. Mahwah: Lawrence Erlbaum Associates. Hagoort, P., Brown, C. M. and Osterhout, L. 2000. “The neurocognition of syntactic processing”. In The neurocognition of language, C. M. Brown and P. Hagoort (eds), 273–316. Oxford: Oxford University Press. Hahne, A. and Friederici, A. D. 2001. “Processing a second language: Late learners’ comprehension mechanisms as revealed by event-related brain potentials”. Bilingualism: Language and Cognition 4: 123–141. Hamann, S. B. and Squire, L. R. 1997. “Intact priming for novel perceptual representations in amnesia”. Journal of Cognitive Neuroscience 9: 699–713. Hebb, D.O. 1949. The organization of behaviour: A neuropsychological theory. New York: Wiley. Hermans, D. 2000. Word production in a foreign language. Doctoral dissertation, Katholieke Universiteit Nijmegen. Jiang, N. and Korster, K. I. 2001. “Cross-language priming asymmetries in lexical decision and episodic recognition”. Journal of Memory and Language 44: 32–51. Johnson, J. S. and Newport, E. L. 1989. “Critical period effects in second language learning: The influence of maturational state on the acquisition of English as a second language”. Cognitive Psychology 21: 60–99.
215
216 David W. Green
Juffs, A. 1998. “Main verb versus reduced relative clause ambiguity resolution in L2 sentence processing”. Language Learning 48: 107–147. Kim, K. H. S., Relkin, N. R., Lee, K. M. and Hirsch, J. 1997. “Distinct cortical areas associated with native and second languages”. Nature 388: 171–174. Kinder, A. and Shanks, S. 2001. “Amnesia and the declarative/non-declarative distinction: A recurrent network model of classification, recognition and repetition priming”. Journal of Cognitive Neuroscience 13: 648–669. Klein, D., Milner, B., Zatorre, R., Meyer, E. and Evans, A. 1995. “The neural substrates underlying word generation: A bilingual functional-imaging study”. Proceedings of the National Academy of Sciences USA 92: 2899–2903. Knowlton, B. J. and Squire, L. R. 1993. “The learning of categories: Parallel brain systems for item memory and category knowledge”. Science 262: 1747–1749. Knowlton, B.J. and Squire, L.R. 1994. “The information acquired during artificial grammar learning”. Journal of Experimental Psychology: Learning, Memory, and Cognition 20: 79–91. Knowlton, B. J. and Squire, L. R. 1996. “Artificial grammar learning depends on implicit acquisition of both abstract and exemplar-specific information”. Journal of Experimental Psychology: Learning, Memory, and Cognition 22: 169–181. Kroll, J. F. and Dussias, P. E. (in press). “The comprehension of words and sentences in two languages”. In Handbook of bilingualism, T. Bhatia and W. Ritchie (eds), Cambridge, MA: Blackwell Publishers. Kroll, J. F. and de Groot, A. M. B. 1997. “Lexical and conceptual memory in the bilingual: Mapping form to meaning in two languages”. In Tutorials in bilingualism: Psycholinguistic perspectives, A. M. B. de Groot and J. F. Kroll (eds), 169–199. Mahwah, NJ: Lawrence Erlbaum Associates. Kroll, J. F. and Stewart, E. 1994. “Category interference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations”. Journal of Memory and Language 33: 149–174. Kroll, J. F., Michael, E., Tokowicz, N. and Dufour, R. (2002). “The development of lexical fluency in a second language”. Second Language Research 18: 137–171. Ku, A., Lachmann, E. A. and Nagler, W. 1996. Selective language aphasia from herpes simplex encephalitis. Pediatric Neurology 15: 169–171. Kutas, M. and Kluender, R. 1991. “What is who violating? A reconsideration of linguistic violations in light of event-related brain potentials”. In Cognitive Electrophysiology, H.-J. Heinze, T. F. Münte and G. R. Mangun (eds),183–210. Boston: Birkhäuser. Lebrun, Y. 2002. “Implicit competence and explicit knowledge”. In Advances in neurolinguistics of bilingualism, F. Fabbro (ed.), 299–313. Udine: Forum. Lenneberg, E. H. 1976. Biological foundations of language. New York: Wiley. Levelt, W. J. M., Roelofs, A. and Meyer, A. S. 1999. “A theory of lexical access in speech production”. Behavioral and Brain Sciences 22: 1–75 Loring, D. W., Meador, K. J., Lee, G. P., et al. 1990. “Cerebral language lateralization: evidence from intracorotid amobarbital testing”. Neuropsychologica 28: 831–838. McClelland, J.L., McNaughton, B.L. and O’Reilly, R.C. 1995. “Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory”. Psychological Review 102: 419–437.
Neural basis of lexicon and grammar in L2 acquisition 217
MacWhinney, B. 1997. “SLA and the competition model”. In Tutorials in bilingualism: Psycholinguistic perspectives, A. M. B. de Groot and J. F. Kroll (eds), 113–142. Mahwah, NJ: Lawrence Erlbaum Associates. Mesulam, M. M. 1990. Large scale neurocognitive networks and distributed processing for attention, language and memory. Annals of Neurology 28: 597–613. Noppeney, U. and Price, C. J. (submitted). “The neural basis of syntactic priming”. Ms. Wellcome Department of Imaging Neuroscience, UCL. Nosofsky, R. M. and Zaki, S. R. 1998. “Dissociations between categorization and recognition in amnesic and normal individuals: an exemplar-based interpretation”. Psychological Science 9: 247–255. Opitz, B., Mecklinger, A. and Friederici, A. D. 2000. “Functional asymmetry of human prefrontal cortex: Encoding and retrieval of verbally and non-verbally coded information”. Learning and Memory 7: 85–96. Ostergaard, A. L. 1999. “Priming deficits in amnesia: Now you see them now you don’t”. Journal of the International Neuropsychological Society 5: 175–190. Osterhout, L. and McLaughlin, J. 2000. “What brain activity can tell us about secondlanguage learning”. Paper presented at the 13th Annual CUNY conference on Human Sentence Processing, San Diego. Paradis, M. 1994. “Neurolinguistic aspects of implicit and explicit memory: Implications for bilingualism and second language acquisition”. In Implicit and explicit language learning, N. Ellis (ed.), 393–419. London: Academic Press. Paradis, M. 1997. “The cognitive neuropsychology of bilingualism”. In Tutorials in bilingualism: Psycholinguistic perspectives, A. M. B. de Groot and J. F. Kroll (eds), 331–354. Mahwah, NJ: Lawrence Erlbaum Associates. Paradis, M. 2001. “Bilingual and polyglot aphasia”. In Handbook of neuropsychology, 2nd edition, vol. 3 Language and aphasia, R. S. Berndt (ed.), 69–91. Amsterdam: Elsevier Science. Perani, D., Dehaene, S., Grassi, F., Cohen, L., Cappa, S. F., Dupoux, E., Fazio. F. and Mehler, J. 1996. “Brain processing of native and foreign languages”. NeuroReport 7: 2439–2444. Perani, D., Paulesu, E., Sebastian-Galles, N., Dupoux, E., Dehaene, S., Bettinardi, V., Cappa, S. F., Fazio, F. and Mehler, J. 1998. “The bilingual brain: Proficiency and age of acquisition of the second language”. Brain 121: 1841–1852. Pinker, S. 1994. The language instinct. New York: William Morrow. Price, C. J. 2000. “The anatomy of language: contributions from functional neuroimaging”. Journal of Anatomy 197: 335–359. Price, C. J., Green, D. and Von Studnitz, R. 1999. “A functional imaging study of translation and language switching”. Brain 122: 2221–2236. Price, C. J., Moore, C., Humphreys, G. and Wise, R. 1997. “Segregating semantic from phonological processes during reading”. Journal of Cognitive Neuroscience 9: 727–733. Rapport, R. L., Tan, C. T., and Whitaker, H. A. 1983. “Language function and dysfunction among Chinese and English speaking polyglots: Cortical stimulation, Wada testing, and clinical studies”. Brain and Language 18: 342–366. Rickard, T.C. 2000. “Methodological issues in functional magnetic resonance imaging studies of plasticity following brain injury”. In Cerebral reorganization of function after brain damage, H. S. Levin and J. G. Grafman (eds), 304- 317. Oxford: Oxford University Press.
218 David W. Green
Robertson, I. H.and Murre, J. M. 1999. “Rehabilitation of brain damage: Brain plasticity and principles of guided recovery”. Psychological Bulletin 125: 544–575. Rugg, M. R. 2000. “Functional neuroimaging in cognitive science”. In The neurocognition of language, C.Brown and P. Hagoort (eds), 15–36. Oxford: Oxford University Press. Scoresby-Jackson, R. E. 1867. “Case of aphasia with right hemiplegia”. Edinburgh Medical Journal 12: 696–706. Segalowitz, N. S. and Segalowitz, S. J. 1993. “Skilled performance, practice, and the differentiation of speed-up from automatization effects: Evidence from second language word recognition”. Applied Psycholinguistics 14: 369–385. Slobin, D. I. 1996. “From ‘thought and language’ to ‘thinking for speaking’”. In Rethinking linguistic relativity, J. J. Gumperz and S. C. Levinson (eds), 177–202. Cambridge: Cambridge University Press. Springer, J. A., Binder, J. R., Hammeke, T. A. et al. 1999. “Language dominance in neurologically normal and epilepsy subjects: A functional MRI study”. Brain 122: 2033–2046. Squire, L. R., Knowlton, B. and Musen, G. 1993. “The structure and organization of memory”. Annual Review of Psychology 44: 453–495. Squire, L. R. 1994. “Declarative and nondeclarative memory: Multiple brain systems supporting learning and memory”. In Memory systems, D. L. Schacter and E. Tulving (eds), 203–231. Cambridge, MA: MIT Press. Strauss, E., Satz, P. and Wada, J. 1990. “An examination of the crowding hypothesis in epileptic patients who have undergone the carotid amytal test”. Neuropsychologia 28, 1221–1227. Teuber, H. L. 1974. “Why two brains?” In The Neurosciences. Third study program, F. G. Worden (ed.), 71–74. Cambridge: MIT Press. Tomasello, M. 1995. “Language is not an instinct”. Cognitive Development 10: 131–156. Ullman, M. T. 2001a. “The declarative/procedural model of lexicon and grammar”. Journal of Psycholinguistic Research 30: 37–69. Ullman, M. T. 2001b. “The neural basis of lexicon and grammar in first and second language: the declarative/procedural model”. Bilingualism: Language and Cognition 4: 105–122. Ungerleider, L. G. 1995. “Functional brain imaging studies of cortical mechanisms for memory”. Science 270: 769–775. Vaid, J. and Hull 2002. “Re-envisioning the bilingual brain using functional neuroimaging: Methodological and interpretive issues”. In Advances in neurolinguistics of bilingualism, F. Fabbro (ed.), 315–355. Udine: Forum. Vigliocco, G. and Franck, J. 2001. “When sex affects syntax: Context effects in sentence production”. Journal of Memory and Language 45: 368–390. Vigliocco, G. and Hartsuiker, R. J. 2001. “The interplay of meaning, sound, and syntax in sentence production”. Psychological Bulletin (under review). Weber-Fox, C. and Neville, H. J. 1996. “Maturational constraints on functional specializations for language processing: ERP and behavioral evidence in bilingual speakers”. Journal of Cognitive Neuroscience 8: 231–256. Weber-Fox, C. and Neville, H. J. (submitted). “Sensitive periods differentiate processing subsystems for open and closed class words: An ERP study in bilinguals”. cited in Ullman, 2001b (above).
Chapter 10
The interface Concluding remarks Roeland van Hout, Aafke Hulk and Folkert Kuiken University of Nijmegen (Hout) / Utrecht University (Hulk) / Universiteit van Amsterdam (Kuiken)
1.
Interfaces in generative grammar
In its bare form, a grammar can be defined as a set of elements or symbols (the lexicon) and a set of rules (the syntax) that together produce output strings, the utterance of the language belonging to the grammar. The format of the lexicon is evident: it has no structure, and it does not need to have one. Chomsky (1965: 84) characterizes the lexicon as “simply an unordered list of all lexical formatives”. Although the assumed absence of order may be accepted as a heuristic device, the lexicon of human languages and of human speakers is not an unordered list, far from that, as is apparent from many recent lexical studies as well as from the chapters in this book. The form, role and structure of lexical items, words, lemmas, or whatever the lexical ‘formatives’ are called, and the relationships or connections between them constitute the pivotal domain of research in the chapters of this book. The nearest neighbour to which lexical items and their structural properties connect is syntax, the machinery by which utterances can be computed, on the basis of lexical input. The questions then are: How are lexicon and syntax linked precisely? What kind of information do they exchange? What is the interface between lexicon and syntax? Which interface levels need to be distinguished? There are several possible answers to this question depending on one’s theoretical perspective. Within the generative, modular paradigm, Chomsky (1995: 131) distinguishes two interface levels: the level of phonetic form (PF) is the interface with sensorimotor systems, the level of logical form (LF) is the interface with systems of conceptual structure and language use. The two performance systems involved are the articulatory-perceptual system and the conceptual-intentional system,
220 Roeland van Hout, Aafke Hulk and Folkert Kuiken
or to put it in a more straightforward way, they relate to sound and meaning. This generativist view is captured by Jackendoff (2002: 197) in the following diagram: lexicon syntax phonology
semantics
This diagram illustrates the central position of syntax, and, at the same time, it raises the question of the relationship between lexicon and syntax. In the chapter by Norbert Corver, this relationship was identified as the third interface level. Corver cited Chomsky (1991, 46): …that there are three ‘fundamental’ levels of representation: D-structure, PF and LF. Each constitutes an ‘interface’ of the syntax (broadly constructed) with other systems: D-structure is the projection of the lexicon, via the mechanisms of X-bar theory; PF is associated with articulation and perception, and LF with semantic interpretation.
D-structure is no longer a separate level in minimalist theory, as Corver points out, but the internal interface of syntax and lexicon is still to be distinguished. The lexical input needs to provide the information required to put grammar to work. This means that lexical items in generative grammar contain information related to syntax (the formal features), to phonology (the phonological matrix), and to semantics (the semantic features). The minimalist approach puts syntax in a central position. Other generative view points exist, for instance Jackendoff (2002) attaches equal generative capacity and autonomy to each of the three levels. From his perspective, phonological structures, syntactic structures and conceptual structures are part of the same processing architecture (Jackendoff 2002: 199), and that implies that interface issues between lexicon and syntax are basically not different from conceptual and phonological interface issues, as is illustrated in the following diagram: lexicon phonology
syntax
semantics
The interface 221
The contributions in this book concentrate on the lexicon-syntax dimension, without any direct claims about the relative importance of other interfaces. Interesting contributions on the relationships between semantics/conceptual structures and syntactic structures, for instance, can be found in Bowerman and Levinson (2001), in the context of (first) language acquisition. In the Jackendoff approach, the key position of the lexicon in relation to phonological, syntactic, and semantic structures is reflected in the format of lexical items. Three layers of information are distinguished for lexical items: phonological features, semantic features, and syntactic (or formal) features. The lexicon contains the fuel to put language to work, which applies to both content and function words. In his chapter Ton Dijkstra added the orthographic layer, a consequence of our literate society that tends to be overlooked by linguists.
2. Learning the syntax, learning the lexicon As discussed in Richard Towell’s chapter, L2 acquisition, in contrast to L1 acquisition, is marked by variability and incompleteness. Any theory on human language should be able to explain these two phenomena, but, in addition, the claim can be made that L2 acquisition can shed light on the properties of human language. The interaction between one language (for instance the L1) and another language (for instance the L2) in the heads and mouths of real speakers, can produce evidence for the core properties of lexical and syntactic structures (cf. Muysken 2000). Including the developmental track of acquisition and stagnation stresses the relevance of bilingual language processing as a primary topic in language research even further. What is the relationship between syntax and lexicon in L2 acquisition? A strong argument in favour of the high status of lexical information is that second language learners always have been aware of the kernel value of a lexicon. Learners prefer to walk around with dictionaries, not with grammars. It has been understood for some time that syntax and lexicon involve different kinds of learning: syntax is learnt through a process of implementing a particular set of universal structures; lexis is learnt by establishing a set of arbitrary associations which operate in a given society. The learning of syntax is often characterised as a process of triggering; the learning of lexis is characterised by the building up of associations (or connections). Yet these two systems must come together in the creation of a whole linguistic system in the mind of an individual. The syntax will govern the phrase structure of the grammar but the
222 Roeland van Hout, Aafke Hulk and Folkert Kuiken
lexical items will govern how the phrase structure is implemented, notably the argument structure and the feature composition of the lexical items is essential to the implementation of the syntax in the language production process. This book was designed to examine the relative contribution of these two dimensions in a clear fashion, through illustrations of exemplary research carried out within each paradigm and to examine how they can be made to inter-relate in a way which would enable us to explain better the overall process of SLA. An examination of the interface between syntax and the lexicon is both timely and important. Both groups of researchers are now coming to an understanding that their particular view of the world may not suffice to account for the overall process and that each will have to understand more about what the other knows. From the point of view of the researcher interested in syntax (generally coming from a background in linguistics), the shift of linguistic theory away from principles and parameters and into minimalism has resulted in a crucial increase of the significance of the lexicon. Within minimalism so much of the information essential for the working of the system has been assigned to the lexicon that it has become crucial for syntacticians to reflect more on how the lexicon works. From the point of view of the researcher interested in the lexicon (generally coming from a background in psychology), it is important to integrate the outcome of lexical learning within the overall acquisition process. Unless one adopts a purely connectionist position, it is clear that the use of the lexical items studied can only take place within the syntactic system. An understanding of the acquisition of the syntax is therefore essential to the understanding of the whole second language processing and acquisition. The introductory chapter by Richard Towell explicates the complementarity of linguistics and psychology in doing language research. Taking the other chapters in this book in consideration, he balances the linguistic and psychological dimensions of basic questions in second language acquisition research. Towell argues that the lexical part of the lexicon-syntax interface is the driving force of language acquisition and that we need to investigate the psychological mechanisms behind this development.
3. Some final considerations The chapters in this book demonstrate the many different perspectives required to study second language acquisition over its full range. A whole series of contrasts keeps returning: symbolic learning vs. connectionist learning, L1 vs.
The interface 223
L2 acquisition, procedural vs. declarative knowledge, structure vs. process, competence vs. performance, etc. One main conclusion is that, over and over again, we need to plea for a cocktail of treatments, for many sorts of data, and for many sorts of expertises to give an appropriate answer to the questions belonging to such contrastive pairs. In this cocktail, the following distinctions can be made. 3.1 Variety in methodology It will be clear that no single particular methodology can produce all the answers we need. We need the linguistic analysis of spontaneous and elicitated speech data (see the chapters by Hawkins and Liszka, Van de Craats, Corver), but also on-line and off-line grammatical judgment tasks (see the chapters by Duffield, and Sabourin and Haverkort). We need to carry out psycholinguistic experiments with reaction times (see the chapters by Duffield and Dijkstra), but also newer methodologies should be applied like eye-tracking, and neuroimaging techniques (see the chapters by Sabourin and Haverkort, and Green). Another promising methodology is the use of computer simulations (see the chapter by Williams). 3.2 Variety in learners and languages The discovery and testing of general or universal principles and parameters in language and language acquisition require the full range of learners: from real beginners (see Corver and Van de Craats) to (near)native speakers (see, e.g. Hawkins and Liszka), from unguided acquisition to classroom learning. At the same time we need to consider the full language typological range, both as a source and a target language in L2 acquisition (the chapters in this book cover a range larger than in many other books on L2 acquisition). As for the computer simulations, different learning algorithms should be probed, including algorithms where previous knowledge (L1) can be implemented to explore its effect on acquiring another language system (L2). Without studying real beginners, it seems not feasible to get an answer to the question of the role of cognitive vs. linguistic principles in speaking a new language (cf. Klein and Perdue 1997), and how, perhaps, cognitive principles are matched by linguistic structures. Cognitive principles can be influential in an indirect way via the syntax-semantic interface, but maybe they directly trigger specific syntactic mechanisms. That would contradict approaches based solely on lexical formal features as the main sources of information for syntactic structures.
224 Roeland van Hout, Aafke Hulk and Folkert Kuiken
3.3 Variety in linguistic domains The same linguistic domains must be investigated, in both L1 and L2 acquisition, including not only inflectional processes but derivational morphological devices as well, and including not only syntactic parameters related to the whole clause, but also parameters for subdomains (possessive DPs, for instance; see Van de Craats). Studying different linguistic domains can be helpful in determining which words are stored and which explicit rules are active in producing lexical items. Schreuder and Baayen (1997) show that even regular inflectional forms can be stored instead of being computed. Morphology is interesting here because it constitutes, due to its paradigmatic organisation, the link between syntactic structures and sets of lexical items. In L1 acquisition the acquisition of past tense forms has been an important domain for testing associative learning algorithms. (cf. Plunkett and Juola 1999). Past tense forms, from a generative point of view, are the topic of research of Hawkins and Liszka in this book. They observe that high proficiency adult L2 speakers of English do not always mark thematic verbs like walk, notice for past tense in obligatory past tense contexts. They find that the morphology cannot be the source of such optionality. In their view, optionality in adult L2 performance results from the interaction between a perfectly functioning syntactic component with an impaired lexicon (as far as tense features are concerned). Gender systems turn out to be another attractive domain in L2 acquisition because of their irregularities and their covert regularities: Williams suggests that problems learning gender in a second language may reflect the weakness of the kind of associative learning mechanism that underlies incidental learning. Sabourin and Haverkort argue that there is a qualitative difference between native speakers and second language learners in language processing. They suggest that the German participants use a translation strategy to learn Dutch gender assignment and that they use their L1 processing strategies to process their L2 in cases where the grammars are very similar. 3.4 Variety in contexts and tasks Both Hawkins and Liszka, and Haverkort and Sabourin make it clear that different tasks may return different results. A simple production task may show that a learner has control over a specific process, whereas spontaneous language production may show strong differences between native and L2 speakers.
The interface 225
Dijkstra’s research corroborates the naturalness of such task differences. The lexicon is a flexible store whose properties may differ with the task it is confronted with. Only different forms of processing can be held responsible for such task differences, which means that processing is an inherent part of linguistic structuring. The chapters by Dijkstra and Green show how intricate the organisation of the bilingual lexicon is. Differences in organisation may be related to the way differences between speakers and communities develop, dependent upon the type of bilingualism and, in addition, the linguistic distance between the language involved. 3.5 Variety in perspectives Towell’s introduction to this book is a plea for a stronger cooperation between linguistics and psychology and a plea for longitudinal studies. Generative linguistics alone will not do. Whilst generative linguists claim to describe the acquisition process, most of their efforts pertain to the classification of successive linguistic stages of learners’ interlanguages. Towell concludes that the driving force of language learning must come from lexis, as there is no driving force available in the computational system (CS), as it is defined in modern generative syntax. Syntactic knowledge is a template simply present in the mind of the learner that automatically operates given the lexical information inserted. This is illustrated by Corver who argues, both for word level categories and phrasal level categories, that L2-expressions are just as perfect as L1-expressions from an interface perspective, even though from the perspective of the target language they may seem highly imperfect. On the other hand, the learning of lexis has been thought of mainly in terms of some form of associative learning theory, with connectionism being the leading variant in language acquisition research. Some theorists have concluded that connectionist learning can account for the totality of the learning, including the learning of syntax. Generative syntacticians argue that the sophisticated structures which they observe and which are not visible at the surface structure of the language, cannot be learnt in an empirical fashion alone and therefore they claim that innate knowledge (mediated or not by the L1) must be ‘guiding’ language acquisition. It is clear that both groups have a strong case to make but that neither is able on their own to account for the total process. Syntax needs to be fed by lexical information, including formal features; this information has to be collected by the lexicon from output strings
226 Roeland van Hout, Aafke Hulk and Folkert Kuiken
generated by syntax. This is particularly clear in the contribution of Van de Craats who shows how syntax and lexicon interact in the data of Moroccan and Turkish adults learning Dutch outside the class-room. The starting point of their developmental process is assumed to be the fully-fledged grammar of the L1 that under the impact of the L2-environmental input changes the underlying grammar of the L1 towards a more target-like L2 output. The computation of syntactic information on the basis of output strings seems to require storage devices, associative linking and analogical strategies. Such computational efforts need time (see Van de Craats), and sometimes a lot of time before the proper information can become available in the production of spontaneous speech (see Hawkins and Liszka). We hope that the readers of this volume have come to appreciate the complexities involved in the issue of interface and will be encouraged to do more interdisciplinary research in order pave the way for a deeper understanding of the interface between syntax and the lexicon in second language acquisition.
References Bowerman, M. and Levinson. S. (eds). (2001). Language acquisition and conceptual development. Cambridge: Cambridge University Press. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge. Mass: MIT Press. Chomsky, N. (1991). Some notes on economy of derivation and interpretation. In Principles and parameters in comparative grammar, R. Freidin (ed.), 417–454. Cambridge MA: MIT Press. Chomsky, N. (1995). The minimalist program. Cambridge MA: MIT Press. Jackendoff, R. (2002). Foundations of language. Brain, meaning, grammar, evolution. Oxford: Oxford University Press. Klein, W. and Perdue, C. (1997). The basic variety. Or: Couldn’t natural languages be much simpler? Second Language Research 13, 301/347. Muysken, P. (2000). Bilingual speech. A typology of code-mixing. Cambridge: Cambridge University Press. Plunkett, K. and Juola, P. (1999). A connectionist model of English past tense and plural morphology. Cognitive Science 23, 436–490. Schreuder, R. and Baayen, H. (1997). How complex simplex words can be. Journal of Memory and Language 37, 118–139.
T task learning 13, 162, 163, 167 task-dependent variation 176 terminal node 23, 35 test of knowledge of morphology 26 time-course of lexical activation 135, 147 triggering 1, 7, 8, 17, 91, 221 V VP-ellipsis 100, 113-115, 120, 125 W word association 131, 132, 134 word naming 131, 132 working memory 146, 163, 176
In the series LANGUAGE ACQUISITION AND LANGUAGE DISORDERS (LALD) the following titles have been published thus far or are scheduled for publication: 1. WHITE, Lydia: Universal Grammar and Second Language Acquisition. 1989. 2. HUEBNER, Thom and Charles A. FERGUSON (eds): Cross Currents in Second Language Acquisition and Linguistic Theory. 1991. 3. EUBANK, Lynn (ed.): Point Counterpoint. Universal Grammar in the second language. 1991. 4. ECKMAN, Fred R. (ed.): Confluence. Linguistics, L2 acquisition and speech pathology. 1993. 5. GASS, Susan and Larry SELINKER (eds): Language Transfer in Language Learning. Revised edition. 1992. 6. THOMAS, Margaret: Knowledge of Reflexives in a Second Language. 1993. 7. MEISEL, Jürgen M. (ed.): Bilingual First Language Acquisition. French and German grammatical development. 1994. 8. HOEKSTRA, Teun and Bonnie SCHWARTZ (eds): Language Acquisition Studies in Generative Grammar. 1994. 9. ADONE, Dany: The Acquisition of Mauritian Creole. 1994. 10. LAKSHMANAN, Usha: Universal Grammar in Child Second Language Acquisition. Null subjects and morphological uniformity. 1994. 11. YIP, Virginia: Interlanguage and Learnability. From Chinese to English. 1995. 12. JUFFS, Alan: Learnability and the Lexicon. Theories and second language acquisition research. 1996. 13. ALLEN, Shanley: Aspects of Argument Structure Acquisition in Inuktitut. 1996. 14. CLAHSEN, Harald (ed.): Generative Perspectives on Language Acquisition. Empirical findings, theoretical considerations and crosslinguistic comparisons. 1996. 15. BRINKMANN, Ursula: The Locative Alternation in German. Its structure and acquisition. 1997. 16. HANNAHS, S.J. and Martha YOUNG-SCHOLTEN (eds): Focus on Phonological Acquisition. 1997. 17. ARCHIBALD, John: Second Language Phonology. 1998. 18. KLEIN, Elaine C. and Gita MARTOHARDJONO (eds): The Development of Second Language Grammars. A generative approach. 1999. 19. BECK, Maria-Luise (ed.): Morphology and its Interfaces in Second Language Knowledge. 1998. 20. KANNO, Kazue (ed.): The Acquisition of Japanese as a Second Language. 1999. 21. HERSCHENSOHN, Julia: The Second Time Around – Minimalism and L2 Acquisition. 2000. 22. SCHAEFFER, Jeanette C.: The Acquisition of Direct Object Scrambling and Clitic Placement. Syntax and pragmatics. 2000. 23. WEISSENBORN, Jürgen and Barbara HÖHLE (eds.): Approaches to Bootstrapping. Phonological, lexical, syntactic and neurophysiological aspects of early language acquisition. Volume 1. 2001. 24. WEISSENBORN, Jürgen and Barbara HÖHLE (eds.): Approaches to Bootstrapping. Phonological, lexical, syntactic and neurophysiological aspects of early language acquisition. Volume 2. 2001. 25. CARROLL, Susanne E.: Input and Evidence. The raw material of second language acquisition. 2001.
26. SLABAKOVA, Roumyana: Telicity in the Second Language. 2001. 27. SALABERRY, M. Rafael and Yasuhiro SHIRAI (eds.): The L2 Acquisition of Tense– Aspect Morphology. 2002. 28. SHIMRON, Joseph (ed.): Language Processing and Acquisition in Languages of Semitic, Root-Based, Morphology. 2003. 29. FERNÁNDEZ, Eva M.: Bilingual Sentence Processing. Relative clause attachment in English and Spanish. 2003. 30. HOUT, Roeland van, Aafke C.J. HULK, Folkert KUIKEN and Richard J. TOWELL (eds.): The Lexicon-syntax Interface in Second Language Acquisition. 2003.
