Empty Categories in Sentence Processing

Empty Categories in Sentence Processing Linguistik Aktuell/Linguistics Today Linguistik Aktuell/Linguistics Today (...

59 downloads 1783 Views 2MB Size Report

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!

Report copyright / DMCA form

DOWNLOAD PDF

AUTHOR ""

TITLE "Empty Categories in Sentence Processing"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Empty Categories in Sentence Processing

Linguistik Aktuell/Linguistics Today Linguistik Aktuell/Linguistics Today (LA) provides a platform for original monograph studies into synchronic and diachronic linguistics. Studies in LA confront empirical and theoretical problems as these are currently discussed in syntax, semantics, morphology, phonology, and systematic pragmatics with the aim to establish robust empirical generalizations within a universalistic perspective.

Series Editor Werner Abraham University of California at Berkeley University of Vienna Rijksuniversiteit Groningen

Advisory Editorial Board Guglielmo Cinque (University of Venice) Günther Grewendorf (J.W. Goethe-University, Frankfurt) Liliane Haegeman (University of Lille, France) Hubert Haider (University of Salzburg) Christer Platzack (University of Lund) Ian Roberts (University of Stuttgart) Ken Saﬁr (Rutgers University, New Brunswick NJ) Lisa deMena Travis (McGill University) Sten Vikner (University of Stuttgart) C. Jan-Wouter Zwart (University of Groningen)

Volume 43 Empty Categories in Sentence Processing by Sam Featherston

Empty Categories in Sentence Processing

Sam Featherston Eberhard-Karls-Universität Tübingen

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 Sam Featherston Empty Categories in Sentence Processing / Sam Featherston. p. cm. (Linguistik Aktuell/Linguistics Today, issn 0166–0829 ; v. 43) Includes bibliographical references and index. 1. Grammar, Comparative and general--Syntax. 2. Grammar, Comparative and general--Sentences. 3. Principles and parameters (Linguistics) 4. Head-driven phrase structure grammar. I. Title II. Linguistik aktuell ; Bd. 43. P295.F4 2001 415--dc21 2001035023 isbn 90 272 2764 0 (Eur.) / 1 58811 069 9 (US) (Hb; alk. paper) © 2001 – John Benjamins B.V. No part of this book may be reproduced in any form, by print, photoprint, microﬁlm, or any other means, without written permission from the publisher. John Benjamins Publishing Co. · P.O. Box 36224 · 1020 me Amsterdam · The Netherlands John Benjamins North America · P.O. Box 27519 · Philadelphia pa 19118-0519 · usa

For Véronique

AUTHOR ""

TITLE "Table of contents"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Table of contents

Forewordxi Abbreviationsxv Chapter 1 Introduction1 Chapter 2 Empty categories in PPT and HPSG9 2.1 Introduction9 2.2 PPT and ECs10 2.2.1 The position of ECs in the theory10 2.2.2 The empty categories in PPT14 2.2.3 Summary19 2.3 HPSG and ECs19 2.3.1 The position of ECs in the theory20 2.3.2 Passive, raising and equi constructions without ECs21 2.3.3 Traces in HPSG22 2.3.4 HPSG without traces25 2.3.5 Summary27 2.4 The structure of German clauses27 2.4.1 German Clause Structure in PPT28 2.4.2 German clause structure in HPSG45 2.5 Summary48 Chapter 3 Parsers and gap processing49 3.1 Introduction49 3.2 Parameters of variation amongst parsers51 3.2.1 Parallel vs. serial51

viii Table of contents

3.2.2 Modular vs. non-modular53 3.3 Processing gaps58 3.3.1 Mechanisms for locating gaps58 3.3.2 Information available to the parser for locating gaps64 3.3.3 Identifying antecedents76 3.3.4 Summary78 3.4 Conclusions79 Chapter 4 Antecedent Reactivation and Trace81 4.1 The basic data of antecedent reactivation81 4.2 The Trace Reactivation Account85 4.3 Traceless accounts of the antecedent reactivation data87 4.3.1 Methodological objections88 4.3.2 The Direct Association Hypothesis88 4.3.3 The Semantic Processing Account99 4.3.4 The Depth of Processing Account103 4.3.5 How can we disentangle these accounts empirically?108 4.4 Experimental materials and design109 4.4.1 An excursus on object scrambling111 4.4.2 Methodology116 4.5 Experiment 1: Simple SVO121 4.5.1 Results122 4.5.2 Discussion124 4.6 Experiment 2: Particle verb constructions126 4.6.1 An excursus on particle verbs127 4.6.2 Results129 4.6.3 Discussion130 4.7 Experiment 3: VP in SpecCP133 4.7.1 Results and discussion134 4.8 Generalized discussion135 4.9 Conclusions136 Chapter 5 NP-trace and PRO: Local ECs141 5.1 Introduction141 5.1.1 The early CMLP evidence142 5.1.2 The evidence from probe recognition experiments145

Table of contents

5.2 The Trace Reactivation Account applied to local ECs147 5.3 Alternative accounts of the data149 5.3.1 Methodological criticisms of the visual probe recognition data150 5.3.2 The Direct Association Hypothesis151 5.3.3 The Semantic Processing Account151 5.3.4 The Depth of Processing Account156 5.3.5 Summary158 5.4 Experiment 4: Local ECs and end-of-sentence probes158 5.4.1 Procedure161 5.4.2 Results163 5.4.3 Discussion167 5.4.4 Summary168 5.5 Trace reactivation data and syntactic theories169 5.5.1 The empirical adequacy of HPSG and PPT169 5.5.2 An excursus on covert categories in HPSG170 5.5.3 The empirical adequacy of HPSG and PPT: Revised174 5.5.4 An excursus on diﬀerences between raising and equi in HPSG175 5.5.5 A new approach176 5.6 Experiment 5: Event Related Potentials at local EC sites177 5.6.1 ERP data and sentence processing178 5.6.2 Materials180 5.6.3 Procedure181 5.6.4 Results182 5.6.5 Discussion185 5.7 Conclusions189 Chapter 6 Evidence from sentence matching on wh-trace dependencies193 6.1 Accounts of the sentence matching data194 6.1.1 Constraint violations195 6.1.2 Correctability and coherence198 6.1.3 A critique of correctability202 6.1.4 Locality and operator-variable binding204 6.1.5 Summary206 6.2 Experiment 6: Matching multiple wh-questions207 6.2.1 Materials209

ix

x

Table of contents

6.2.2 6.2.3 6.2.4 6.2.5

Procedure212 Results213 Discussion215 Conclusions219

Chapter 7 Conclusions221 Experiments 1–3222 Experiment 4226 Experiment 5227 Experiment 6230 Appendix 1: Materials for CMLP Experiments 1–3233 Experiment 1: Gap ﬁnal233 Experiment 2: Particle ﬁnal234 Experiment 3: VP in SpecCP235 Appendix 2: Materials for probe recognition Experiment 4239 The Test Item Generator240 Appendix 3: Materials for ERP Experiment 5249 Appendix 4: Materials for sentence matching Experiment 6257 References261 Index275

AUTHOR ""

TITLE "Foreword"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Foreword

The work reported in this book came about as a result of the realization that the issue of the role of syntactic gaps in processing was unresolved. It is surprising that this should be the case, since there are few ﬁelds of study which seem to allow experimental approaches to produce answers to syntactic questions, and these few are generally investigated with great zeal. Gap processing shows the potential to be such a ﬁeld, and it too had been very popular in the late eighties and the ﬁrst half of the nineties. The early studies from both self-paced reading and cross-modal lexical priming (ﬁrst and foremost Nicol & Swinney 1989) had shown clear eﬀects at gap positions and this had been widely accepted as strong evidence that traces, or something similar, played an active part in human sentence processing. This, if conﬁrmed, would be a fascinating discovery, as it would reveal a correspondence between the functioning of the human parser and a construct of generative grammar far closer than is normally assumed. The excitement was deﬂated with the publication of Pickering & Barry (1991) and their demonstration that the data could be interpreted otherwise, as activation of a complement of the verb at the verb position. Subsequent empirical work such as Nicol (1993) tended to strengthen the impression that the excitement about trace activation had been misplaced, since other accounts, not making use of traces, were available. Roughly this was the situation when my attention was called to the phenomenon by Harald Clahsen at Essex. While alternative explanations had been put forward, the extensive work necessary to decide between them had not been undertaken, for two main reasons. First, it was diﬃcult to see how this might be achieved using English materials, since objects are normally adjacent to verbs in English, and this was one of the confounds in the data which needed to be resolved. Second, the excitement that psycholinguistics was ﬁnding hard answers to questions of syntactic theory had been given a douse of cold water, and disillusionment set in: few doubted that the theoretically less interesting answer of Pickering & Barry would prove to be correct. Harald Clahsen had noticed that, while English data could not distinguish between the accounts,

xii

Foreword

German materials could, and suggested I should have a look. Thus began a fascinating journey into the various facets of the phenomenon, which, perhaps not surprisingly, became more complex as time went on. The original intention had been to perform one experiment in order to settle the issue of what the previous cross-modal priming experiments had been measuring. Instead of one experiment, three were necessary in order to produce an answer which could not be accounted for by any of the competing theories which do not assume additional processing at gap positions. There remained, however, the unresolved question of what the data on the same topic from the other methodologies was showing. Surprising results require unanimity in the data to support them: since the cross-modal priming data showed one distribution of eﬀects and the probe recognition data (e.g. McElree & Bever 1989) showed another, the overall position still had to be regarded as doubtful. This led the other experiments reported here, using probe recognition, sentence matching and event-related potentials, as well as selfpaced reading. The result of this is perhaps as full an exploration of the topic as could be imagined. I am occasionally asked why I chose to discuss two diﬀerent frameworks’ analyses of the experimental materials instead of assuming just one as is more usual. The answer is twofold: ﬁrst, one of the aims of the study was to test the competing analyses of the two grammars. I regard it as deeply unsatisfactory that two generative grammar models can produce such diﬀerent analyses of the same data. This strikes me as demonstrating that neither model has a suﬃciently close relationship with the linguistic data it claims to represent. In order to rectify this situation, I look to psycholinguistic research and corpus-based work to provide a means of adjudicating between them, and this book is in part an attempt to do just that. The second reason for discussing the predictions of both Principles and Parameters Theory and Head-Driven Phrase Structure Grammar is that each of them has its descriptive strengths, but if we wish to gain the maximum insight into a syntactic structure it is often best to consider what each of them has to say about it. Various other researchers have contributed suggestions, criticisms and revisions to this work, above all Harald Clahsen, to whom I owe many of the ideas for experimentation here. Particular thanks also to Thomas F. Muente and Matthias Gross of the Hannover Medical School. Thanks also to Janet Fodor, Andrew Radford, Claudia Felser, Meike Hadler, Kerstin Maut and Sonja Eisenbeiß, as well as the Psycholinguistics Research Group at Essex for their comments and contributions, both to the design and construction of this work.

Foreword xiii

The research programme that this book reports was funded by the UK Economic and Social Research Council award number R00429534104. Closer to home, I must say thank-you to my parents for their care and support, but most of all to Véronique, to whom this book is dedicated.

AUTHOR ""

TITLE "Abbreviations"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Abbreviations

All abbreviations except those ubiquitous in the linguistic literature (e.g. VP) are additionally signalled within the text where they ﬁrst occur and sometimes subsequently where appropriate. AgrIOP Indirect object agreement projection AgrOP Object agreement projection CMLP Cross-modal lexical priming CP Complementizer phrase DAH Direct Association Hypothesis DO Direct Object DOP Depth of Processing Account EC Empty category ERP Event-related brain potentials GPSG Generalised Phrase Structure Grammar HPSG Head-driven Phrase Structure Grammar IO Indirect Object IP Inﬂexion phrase

NP p.c. PP PPT RC RT S SPA TP TRA UG V2 VP

Nominal phrase Personal communication Prepositional phrase Principles and Parameters Theory Relative clause Reaction time Saturated clausal projection Semantic Processing Account Tense phrase Trace Reactivation Account Universal Grammar Verb second Verb phrase/unsaturated clausal projection

AUTHOR ""

TITLE "Introduction"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 1

Introduction

The aim of this study is to examine certain assumptions about the way that syntactic structures containing extracted arguments are handled by the human sentence processing mechanism and its embedded grammar. In particular we focus on empty categories (= ECs), phonetically null place-holders for locally absent constituents. We give examples of sentences containing one type of empty category, trace, in example (1) (see Chapter 2 for more detail). In such sentences, it is argued that the displaced element (who, peanuts, thus far and no further), while overtly in sentence-initial position, also has a grammatically fully speciﬁed but silent copy in its canonical position following the verb. (1) a. Who did the boatman ferry [trace] across the river? b. Peanuts the eminent chef despises [trace] c. Thus far and no further will I go [trace]

While the overt occurrence of the displaced constituent is at the beginning of the sentence, it is thought that the trace enters into local grammatical relationships, such as subcategorisation by the verb. It is further assumed that there is a syntactic dependency between the displaced element and the trace, ensuring that they have identical features. However, while supporters of some grammatical frameworks make great use of ECs in their analyses of structures where categories are displaced or wholly absent, others view them as a theoretically undermotivated construct and prefer accounts which do not require them. Our research aim is therefore to ﬁnd empirical evidence which might conﬁrm or falsify the existence of ECs in the human sentence processing mechanism. To do this we contrast the predictions for processing of extraction structures of two models of grammar: the Chomskian Principles and Parameters Theory (= PPT) (Chomsky 1981, 1986, 1995) and the lexicalist Head-Driven Phrase Structure Grammar (=HPSG) (Pollard & Sag 1987, 1994). These are two of the most current theories of grammar in use in linguistics today, but also represent rather diﬀerent perspectives on the use of ECs. PPT makes use of ECs not only at the canonical location of displaced or absent constituents, but also

2

Empty Categories in Sentence Processing

in simple sentences, since it appeals to movement as a major explanatory mechanism in quite wide ﬁelds of the grammar. The assumption of ECs oﬀers the advantage that it permits other constraints on structure to hold exceptionlessly: for example, unsaturated projections, where more arguments are subcategorised for than are actually present, can thus be excluded by simple rule, which makes for a simple and elegant theoretical structure. HPSG takes another view, preferring not to make use of phonologically null constituents, but locating the additional structure necessary to dispense with ECs at the subcategorizing head. Essentially this extra structure consists of a distinction between local and non-local subcategorisation, with this latter applying to locally absent arguments. This arrangement allows the framework the considerable theoretical advantage of assuming only those constituents which are overtly represented in the input. The use of ECs is thus an interesting point of contention between syntactic theories, which means that we can test sets of competing predictions. It is the more signiﬁcant because of the importance that ECs occupy within PPT, for it is a central design feature of this approach that the descriptive facts should not be stipulated in construction-speciﬁc rules but rather derived from the interaction of abstract universal principles, which themselves are constraints from outside the language faculty. These principles must necessarily apply without exception, for they are part of the larger mental and physiological system in which the language faculty is embedded. In such a system, ECs provide a mechanism permitting the characterisation of these constraints as absolute and exceptionless. Within PPT, particularly within its most recent instantiation the Minimalist Program (e.g. Chomsky 1995), ECs are thus not an optional feature but rather an essential load-bearing member in the architecture of the grammar. PPT without ECs is therefore not readily imaginable, and a signiﬁcant portion of the credibility of PPT hangs on their psychological reality. This contrasts strongly with the position of many syntactians working in HPSG, who tend to view them as theoretically undesirable and, while they cannot yet rule out ECs absolutely, are keen to develop analyses which eliminate the need for them. The weighty implications for grammatical theories of ﬁrm answers to questions about the psychological reality of ECs thus make the issue challenging and worthwhile. In this study we investigate ECs with the methodology of psycholinguistic experimentation. In order to do this we must make an assumption: namely that there is some close correspondence between the functioning of the human sentence processor and the mental grammars suggested in the literature. We

Introduction

need to assume that the sentence processor is attempting to assign to its input the structures which grammatical theories would specify for them (see Chapter 3 for further detail). Given this parser-grammar transparency (cf. Berwick & Weinberg 1984 for discussion of stronger claims), we can generate hypotheses from grammatical theories about the behaviour of subjects presented with sentences containing displaced constituents, and so test the hypothesis that ECs are used in sentence processing. While transparency is an assumption that we cannot strictly test, it is a necessary presupposition within work in the ﬁeld, because even the most abstract work in syntactic theory must, at some level, base itself on the output of language processing. In fact we have no other way of knowing anything about the mental representation of human language except via the mediation of the language processor. It follows that every syntactic theory therefore is and can only be a model of the structure of the language as implemented by the processor. Transparency therefore is an assumption originally made by syntacticians when they make observations about the grammar from the language data, since this, even if it is introspective data, is nevertheless only accessible via the language processor. There is thus ample precedent for this assumption. Let us note here that we restrict our discussion of ECs in this work to experimental evidence and do not discuss the theoretical evidence or data from other areas of study such as phonology (see e.g. Sag & Fodor 1995). It is necessary for us to constrain our discussion to a certain manageable ﬁeld of theory and phenomena, and since we make no appeal to evidence from outside the immediate bounds of syntactic theory and sentence comprehension, we shall not review the argumentation. The structure of the book is as follows: in Chapter 2 we outline the syntactic background to our work and the use of ECs in PPT and HPSG, and in Chapter 3 we review previous psycholinguistic work on extraction dependencies. Chapters 4, 5 and 6 report our own experimentation, while Chapter 7 attempts to draw a balance of the evidence. In the remainder of this chapter we sketch in somewhat more detail the issues that we address and the experiments that we have undertaken, their design, aims and methodology. The major question we address in this study is whether there is reliable empirical evidence from behavioural measures which supports the assumption of ECs. A considerable amount of work has been done on this issue and a large body of data accumulated. However the application of psycholinguistic methodologies to syntax is a fast developing and continually changing ﬁeld, and as more is learned about the processing of sentences, past interpretations can

3

4

Empty Categories in Sentence Processing

rapidly be overtaken by newer insights. Because of this and because of certain confounds in the experimental structures, there is as yet no conclusive answer to the question whether there are or are not empirical reﬂexes of ECs. There are currently four psycholinguistic accounts purporting to be explanations of the data set bearing upon the reality of ECs in the representation of sentence structures, each having its own empirical base and set of predictions for processing (see Chapter 4 for details). One of these, the Trace Reactivation Account (e.g. Swinney et al. 1989; Nicol & Swinney 1989), attributes a range of results showing eﬀects in structures with gaps to reactivation of antecedents by ECs; the others are all traceless accounts, in that they attribute the same ﬁndings to other causes, and make no reference to ECs. The Direct Association Hypothesis (Pickering & Barry 1991) holds that experimental eﬀects at gap positions are due to processing of dependents at their subcategorizer and that this has been confused with gap position eﬀects because subcategorizers and gaps are normally adjacent. The Semantic Processing Account (Fodor 1989, 1993, 1995; Sag & Fodor 1995) suggests that the gap-related eﬀects may be due to semantic not syntactic processing, and that they can therefore oﬀer no support to a syntactic construct such as ECs. The Depth of Processing Account (Fodor 1995, p.c.) essentially attributes greater activation to greater processing complexity. Since an extracted argument must undergo more or deeper processing than an in-situ argument, it attains a higher level of mental activation, which causes the experimental eﬀects otherwise attributed to reactivation by trace. While the predictions of these four accounts are largely indistinguishable in English, the language on which most of this research has focused, we have succeeded in teasing them apart by using materials in another language, German, which diﬀers from English in some key characteristics. There are three key features of German which permit this. Firstly, German has a head-ﬁnal verb phrase, which means that verbs follow their direct objects. This makes it easier to distinguish between eﬀects at verbs and eﬀects at direct object gap positions. Next, German is a verb second language: ﬁnite verbs often appear in second position in the clause, but this causes no such change in order of arguments. This permits us to test direct object gaps which are remote from the overt position of their subcategorizing verb. Thirdly, German allows short scrambling, word order variations within the clause. This permits us to test extraction dependencies which do not cross the verb, which removes another confounding factor from the data. Our experiments use these features of German to distinguish between the competing psycholinguistic accounts of the antecedent reactivation data.

Introduction

Experiments 1 to 3 are reported in Chapter 4. They use the cross-modal lexical priming methodology and are an attempt to establish whether or not there is experimental evidence for trace, testing cases for which diﬀerent predictions are made by the diﬀerent psycholinguistic accounts. We achieved this by testing at the position from which a direct object is scrambled in a verb second structure. For example, in (2) from Experiment 1 the processing of the foot of the dependency of ihre Münze “her coin” and its extraction site can be examined remote from the subcategorizing verb, as this is in second position in the clause. (2) Die Frau gab ihre Münzei dem Mädchen ti the woman gave her coin to.the girl

Experiments 2 and 3 diﬀer only slightly and allow us to control for other factors which have previously obscured the diﬀerences between the accounts. The three experiments together provide empirical data which discriminates between the four diﬀerent accounts. In Chapter 6 we report Experiments 4 and 5. These investigate another subset of ECs, referred to in PPT as NP-trace and PRO. Previous evidence for these constructs is much weaker and more questionable than that for wh-trace. We report two experiments on this issue; again the aim is to diﬀerentiate between alternative interpretations of the existing data set. The ﬁrst uses an end-of-sentence probe recognition technique and compares raising, passive, and equi structures against three control conditions to assess whether these ECs, as covert anaphors, produce the same speeded response times as overt anaphoric elements. This design too allows us to test contrasting predictions of the diﬀerent accounts of previous ﬁndings. The second experiment looks at the issue from a new perspective using measurements of event-related brain potentials. The conditions we contrast are raising, equi, and transitive control structures, and the measurements are taken at the point where, on PPT analyses, one structure type has an NP-trace, the next a PRO, and the last no covert category. The aim is to determine whether the structural diﬀerence between raising and equi implicit in PPT but absent from traceless accounts is reﬂected in subjects’ brain potentials. Chapter 6 is our ﬁnal experimental chapter. This addresses a rather diﬀerent aspect of ECs, namely, the processing of their binding. The aim is to test the claim that trace binding, and only trace binding, has a particular characteristic. If it has, then it follows that the concept of trace is supported. We employ the sentence matching task, which has been argued to be sensitive to a particular

5

6

Empty Categories in Sentence Processing

feature of the structure of the language processor. Speciﬁcally it appears to distinguish between trace binding and other grammatical processes, a distinction made in modular theories of grammar and processing but not in integrative ones. We examine the processing of a particular island constraint, namely the one which excludes multiple wh-extraction in multiple wh-questions. In (3a), for example, we see that a multiple wh-question with one extracted wh-element who is grammatical; (3b) shows that the second wh-element what may not undergo wh-extraction even within the embedded clause; (3c) demonstrates that this must be attributed to some island constraint, since the local string …did you say what + clause is legitimate. (3) a. Who did you say bought what? b. *Who did you say what bought? c. Did you say what Jack bought?

The key issue is whether the parser distinguishes between local ungrammaticality (4a) and non-local ungrammaticality such as illegal wh-trace binding (4b). (4) a. *What does you say Jack bought? b. *What did you say Jack bought a parsnip?

Previous results using sentence matching have shown ungrammaticality eﬀects for local ungrammaticality but none for non-local ungrammaticality. This would suggest that the processor does indeed make such a distinction, which would lend support to modular parsing models which permit such features as movement, but here too there is more than one account of the experimental ﬁndings so far, and we address a test case. Conclusions can be drawn from our work at two levels: ﬁrst it has signiﬁcance in the debate between diﬀerent models of generative grammar, since diﬀerent frameworks make diﬀerent assumptions about ECs and their processing. While we highlight the implications just for PPT and for HPSG in the text here, our results are equally relevant to other frameworks and should contribute towards resolving some of the diﬀerences. At a higher level of generality our ﬁndings bear upon the validity of the linguistic as opposed to the psychological approach to the representation of language (e.g. Connectionism (Rumelhart & McClelland 1986), linguistic relativism (MacWhinney & Bates 1987)). If observationally and explanatory adequate models of syntactic processors can be developed on the basis of generative grammars, then this must be seen as a conﬁrmation of the linguistic approach to the characterisation of language as the product of a language faculty, a separate module in the mind with its own

Introduction

speciﬁc structure and qualities. This is in contrast to approaches which treat linguistic behaviour merely as one aspect of wider human behaviour, and assert that the speciﬁc characteristics of language are derivable from wider nonmodule-speciﬁc processes. On this level therefore, our work can provide evidence supporting generative grammars as realistic models of the human language faculty.

7

AUTHOR ""

TITLE "Empty categories in PPT and HPSG"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 2

Empty categories in PPT and HPSG

2.1 Introduction This chapter contains the syntactic background to our experimental studies on the processing of ECs. We begin with a brief sketch of how the aims and explanatory priorities of PPT and HPSG cause them to hold diﬀering analyses of the same empirical data and reach very diﬀerent conclusions about ECs. We also outline the variety and internal constituency of the ECs in each of the syntactic frameworks. The assumptions about ECs diﬀer markedly between the two models and form one of the most hotly debated areas of syntactic controversy. Since contrasting predictions about processing can be drawn from these assumptions about ECs, we can test the comparative empirical adequacy of the competing frameworks by comparing these predictions with the experimental data. In the second part of this chapter we specify certain syntactic assumptions we shall make in our analysis of the structures in our experimental materials. We do not attempt a comprehensive introduction to the two grammatical frameworks we discuss here (see Borsley 1996 for HPSG, Radford 1997 for PPT, and Borsley 1991 for a comparative approach), but rather focus only on those parts of the theories which concern ECs. Within Principles and Parameters Theory we subsume Government and Binding Theory (e.g. Chomsky 1981, 1986a) and the Minimalist Program (e.g. Chomsky 1995). For HPSG we refer chieﬂy to Pollard & Sag (1987) and (1994). Our view of the models is necessarily broad because the empirical studies we shall discuss span more than twenty years, during which time certain aspects of the frameworks have developed. Since research in processing tends to utilize the most consensual of syntactic analyses, it is rarely necessary to specify in detail which version of a grammar is being referred to in a particular study. This is a productive imprecision, because the diﬀerences between earlier and later versions are rarely crucial to the discussion; but we compare earlier and later versions of our two competing grammatical frameworks where this is relevant. In the ﬁnal section we outline which variants of each theory we shall adopt

10

Empty Categories in Sentence Processing

as they apply to the structure of German clauses. This is particularly necessary in our discussion of PPT, for there are a number of assumptions which are frequently adopted within current syntactic discussion but which we shall not take up here. In doing this we are characterising the consensual model assumed by workers in processing (e.g. for German Bader & Lasser 1994; Crocker 1994) rather than making speciﬁc stipulations for our own study; in fact these alternative analyses (with one exception) do not change the predictions in our experiments.

2.2 PPT and ECs 2.2.1 The position of ECs in the theory The position that ECs play in PPT is an important one. Trace theory was developed in the 70s as a mechanism to circumvent some problems encountered in the transformational syntax of the time (e.g. Fiengo 1977; Chomsky 1973, 1981; see also references in Chomsky 1981: 144, fn. 76). Essentially traces served to act as markers for transformations, so that all transformations were nondestructive. If a moved element leaves a trace behind it, the previous structure is still visible to the computational system; this removes the need for the syntax to specify in what order transformations apply. This is important because it permits PPT to be a model of comprehension as much as of production: nonstructure-preserving transformations cannot necessarily be run in reverse. ECs are an essential component of PPT for other reasons too. PPT has always aspired to be a grammatical theory which is not only descriptively but also explanatory adequate (e.g. Chomsky 1965:24f ); that is, it is able to produce a descriptively adequate grammar of every human language using only mechanisms and constructs which are psychologically plausible and realistically learnable. To achieve this Chomsky concludes (e.g. Chomsky 1965: 25ﬀ, 1981: 6ﬀ, 1986a:51ﬀ) that humans are born with a Universal Grammar (= UG) which speciﬁes quite narrowly what forms the syntax can take. The learning of syntax must thus be reduced to the smallest burden possible consistent with the attested variations in structure between human languages. It is this problem of acquisition which is the main motor behind the theoretical direction that Chomskian grammatical theory has taken. Now this minimal syntax is achieved by the use of principles and parameters. Principles are universal design features of human language, while parameters are options; they contain variables which

Empty categories in PPT and HPSG

are ﬁxed on the basis of input. In line with the aim of explanatory adequacy, the rule systems are abstract and over-arching; it is the interaction of the general rules which produces what appear to be construction-speciﬁc eﬀects. An important point for us to note is that these principles apply blindly and without exception. A locally absent argument should therefore cause the derivation to crash since the local grammatical constraints are not satisﬁed. This result is avoided by the assumption that extracted arguments leave behind traces as placeholders and that inﬁnitival clauses have PRO subjects. These ECs permit the principles to hold universally and exceptionlessly and thus be credible candidates for inclusion within UG. The assumption of ECs is therefore necessary for PPT to allow it to aspire to explanatory adequacy. We illustrate this with some examples here. The Projection Principle (Chomsky 1981: 29, cf. Chomsky 1995: 189) requires that lexical requirements be met throughout the derivation: traces permit this to hold even when constituents are locally absent. For instance, (5a) is ungrammatical because the verb beat requires an object. What is more, this object must be immediately following and adjacent to it, as (5b) and (5c) show. (5d) however is grammatical even though the overt local constituents Hilda regularly beats at tennis are identical to those in (5a): the assumption that the direct object has left a trace in the canonical object position removes the inconsistency. (5) a. b. c. d.

*Hilda regularly beats at tennis *Hilda regularly beats at tennis Dennis *Hilda regularly Dennis beats at tennis Dennis, whoi Hilda regularly beats ti at tennis, …

The Theta Criterion (Chomsky 1981:34,101,170; 1986b:86ﬀ; 1986a:13f; 1995: 30ﬀ) demands bi-uniqueness in theta role assignment: ECs permit this to be fulﬁlled.1 In (5d) we understand the wh-element to bear the PATIENT role which beat has to assign under sisterhood. Trace theory allows a natural account of this: the trace is in the appropriate position to receive the theta role. Binding theory (Chomsky 1995: 92, 1981:183) too can be simpliﬁed in an account which makes use of ECs: in a structure such as (6a) there is no overt clause-mate

1.Theta Criterion: Each argument A appears in a chain containing a unique visible theta position P, and each theta position P is visible in a chain containing a unique argument A (Chomsky 1986a: 97).

11

12

Empty Categories in Sentence Processing

antecedent for the reﬂexive.2 The assumption of a PRO subject in inﬁnitival complement clauses explains this in a principled manner, that is, without positing a diﬀerent binding domain for such cases as (6b). (6) a. Harriet tried to improve herself/*himself b. Harrieti tried PROij to improve herselfj/*himself

Similar considerations apply within Case Theory (1981: 49ﬀ; Chomsky 1995: 111ﬀ,308f ). Objective Case is normally only checked by complements of verbs and prepositions: whom in (7) is in neither of these positions. Again we can provide a ready account of this if we assume a chain as in (7b) with a trace at its foot. Objective Case can be checked at the trace position and shared with its antecedent. (7) a. Whom did Alﬁe invite to the dinner? b. Whomi did Alﬁe invite ti to the dinner?

Endocentricity expresses the X¢-theory notion that every projection has a head, but this would seem to be violated by a structure like (8a): the IP has no local head because of I to C movement. The assumption of a trace with the equivalent features to have in I resolves this problem (8b). (8) a. Have [IP they [I] no shame?] b. Havei [IP they [I ti] no shame?]

The development of the Minimalist Program (Chomsky 1995, esp p. 1–11, p. 167–172, p. 219–225) in PPT has brought two more reasons for it to require ECs as an integral part of the theory. One of these lies in the treatment of word order variation between languages: in the Minimalist Program this is brought about by morphological diﬀerences between languages. A strong feature is one which must be checked overtly; a weak one can be checked covertly after Spellout, the point at which phonological features are divided from the others. For example, the contrast between French main verbs, which overtly occupy the I position, and their English equivalents, whose overt position is V, is the 2.Binding: α binds β if α c-commands β and α, β are coindexed (Chomsky 1995: 93). Binding Principles: A. An anaphor must be bound in a local domain. B. A pronominal must be free in a local domain. C. An r-expression must be free. (Chomsky 1995: 96)

Empty categories in PPT and HPSG

diﬀerence between a strong and weak V-feature in I.3 Verb raising in French is to check a strong feature and is thus before Spellout and visible to PF. The equivalent features on the English I are weak and can be checked after Spellout, and are thus invisible to PF. In such a system, where movement operations account not only for extraction dependencies but also word order variation between languages, the use of ECs, in particular traces, is unavoidable. The other reason for the increased importance of trace in the Minimalist Program is the multiplication of movements for checking purposes. All ﬁnite verbs and all case-bearing NPs will move at some stage in a derivation in order to check their features.4 This necessitates some means for them to engage in their local grammatical relationships at their base-generated positions. This necessitates the assumption that all moved elements are members of chains, essentially conduits for the passing of grammatical features, which enable constituents to engage in local grammatical relationships at more than one place simultaneously. All non-head chain members are traces. Within the Minimalist Program, therefore, almost every argument will exist not only overtly but also as a trace, with a chain linking them. To summarize: ECs play an important part within the structure of PPT. They permit transformations to be structure-preserving, which frees the theory of the requirement to order them. They also allow principles to hold without the many exceptions which would be necessary if a purely surface analysis of sentence structure were adopted. There are clear advantages in explanatory adequacy if constraints on structures can be argued to be absolute and to apply throughout the derivation. Traces thus permit the development of a theory structure which relies upon the interaction of highly abstract principles to produce the observed eﬀects, with constructions such as passive being mere artifacts. Such a theory based on abstract over-arching principles could not be implemented if these principles could not be argued to be universal and hold without exception. It is plain therefore that ECs are not merely an optional addition to PPT, but an important structural member. Their role in the Minimalist Program is if anything more central still. With the drive to make constraints on representations ever more necessary and thus exceptionless, and

3.We shall use IP here to stand for a functional projection between VP and CP where further speciﬁcation is unnecessary. 4.We shall refer to nominal projections as NPs, since the internal constituency of nominal projections is not at issue here, noting that Chomsky does the same (e.g. Chomsky 1995). This also has the advantage of being a theory-neutral term.

13

14

Empty Categories in Sentence Processing

with forced movement for the checking of covert morphological features being the central explanatory mechanism, ECs become an essential part of the framework. 2.2.2 The empty categories in PPT In this work we restrict ourselves to the empty categories pro, PRO, NP-trace and wh-trace, which have the status of maximal projections and which are obligatory in the structures that license them. We illustrate these in (9). (9) a. b. c. d.

Jess seemed (to us) NP-trace to like cream best of all. What did James want wh-trace for Christmas? Josh hoped PRO to run 100m in less than 10 seconds. pro nataka wali na kuku. (Swahili) EC want.pres rice and chicken

We shall have nothing to say about other empty elements which are only optionally phonetically null or which represent only heads, since the theory relating to them has been much less well developed. Empty categories occupy a very central role within PPT precisely because they have no overt form (though see Sections 3.3.2.4 and 4.2). This has implications: their nature is very immediately determined by the grammar and much less by external factors; they can be thought of as the minimal element required for convergence, which makes them an interesting keyhole onto the requirements of the grammar. It might even be argued that it is implausible that their existence and characteristics should be acquired by the child exposed to the language, and that they are likely to reﬂect deeper principles of UG, universal and biologically determined (e.g. Chomsky 1981: 55ﬀ). Chomsky 1981 contains a principled justiﬁcation of the characteristics and distribution of ECs. Chomsky 1982 (p. 78f, see also Chomsky 1995: 41) adds to this the category pro (“small pro”) allowing the generalisation that the types of nominals, overt and covert, vary by two parameters [anaphor] and [pronominal]. In fact only seven of the eight cells in Table 1 are ﬁlled. The overt nominals are anaphors, pronouns and R-expressions; the covert categories are PRO, NP-trace, pro and wh-trace. The features [Anaphor] and [Pronominal] are most obviously deﬁned by their binding behaviour: anaphors being bound, and pronouns free, in their binding domain. R-expressions, which are neither anaphoric nor pronominal, are subject to neither of these binding conditions. No overt nominal category can be [+Anaphor] [+Pronominal] as this implies

Empty categories in PPT and HPSG

Table 1.The feature speciﬁcations of ECs (Chomsky 1982, 1995: 41) Features

Overt category

Covert category

[+Anaphor, +Pronominal] [+Anaphor, −Pronominal] [−Anaphor, +Pronominal] [−Anaphor, −Pronominal]

– reﬂexive/reciprocal pronoun R-expressiona

PRO NP-trace pro wh-trace

a

“Referring expression”: one which deﬁnes its own reference.

a contradiction, nevertheless, this was argued to be the case for the EC big PRO; since it has no governing category it can trivially be both free and bound in every binding domain that it has, i.e. none. These diﬀerent ECs vary considerably in their characteristics. PRO and pro are both assigned a theta role and have their Case features checked in situ. They do not therefore form parts of chains and are thus referentially more independent.5 Small pro resembles a pronoun which lacks phonetic form, since it has the same theta, Case, government and binding properties. Big PRO by contrast is very diﬀerent from overt pronouns and has something of a last resort ﬂavour. It occupies only positions which are ungoverned (Chomsky 1981: 56) or, on later analyses, which check null Case (Chomsky 1995:118ﬀ), i.e. SpecTP where T is [−tense]. Its binding properties are complex: it seems to exist in two forms, controlled PRO and arbitrary PRO (for discussion see Chomsky 1981). Control seems to be a special case of binding, with particular qualities (Chomsky 1995: 92ﬀ). Arbitrary PRO may simply be a control by a covert controller (Chomsky 1995: 36).6 The diﬀerences between ECs are summarized in Table 2 below. wh-trace and NP-trace are positions at the foot of chains, and are best thought of as sub-parts of these dependencies. In order to simplify the representational devices used within the Minimalist Program, Chomsky analyses all categories as chains, though most of these will be single membered ones. This uniﬁes the treatment of many grammatical relationships as it allows one to say that every

5.Except in so much as any category is a (single membered) chain. See Chomsky (1981: 333, 1995: 177) for deﬁnitions and further discussion of the notion of chain. We shall reserve the term chain for multi-membered chains here. 6.This presumes that for every arbitrary PRO there is a null element controlling its reference in a c-commanding position. (i) thus is covertly (ii). (i) PROarb to fool Her Majesty’s Customs and Excise is impossible. (ii) (For anyonei) PROi to fool Her Majesty’s Customs and Excise is impossible.

15

16

Empty Categories in Sentence Processing

Table 2.Features of empty categories in PPT pro

wh-tracea

NP-trace

Categorial status NP

NP

XP

NP

Chain member

no

no

È-chain

A-chain

Case

none/nullb

yes

yes

no

θ-role

yes

yes

at foot

at foot

Governmentc

no

yes, by AGR

proper

proper

Binding Theory variable

B

C

A

Antecedent

discourse determined

operator

NP in A-position

EC

PRO

Control or arbitrary

a

Such features as Case and theta-marking naturally apply only to nominal wh-trace, not adjectival or prepositional wh-trace; the reader should thus assume an “if appropriate” suﬃxed to each cell entry. b The Case status of PRO has changed under the Minimalist Program. c Government no longer forms part of PPT within the Minimalist Program, nevertheless we report it as part of the classical PPT account.

chain has exactly one theta role, one Case position and one argument. If a chain has more than one member, the lower chain links are coindexed traces, each trace binding the next in the appropriate conﬁguration. Table 3 summarizes EC dependency types. Table 3.Types of dependencies heading empty categories EC

wh-trace

NP-trace

Chain type

È-chain

A-chain

Dependency type

operator-variable

antecedent-anaphor

Head position

È-position, often SpecCP Caseless É-position

A-position Case-bearing θ-position

Foot position

A-position Case-bearing θ-position

A-position Caseless θ-position

Distribution

wh-questions, relative clauses, Raising, passive, SpecVP to SpecIP topicalisation subject raising

Empty categories in PPT and HPSG

A wh-trace is akin to a logical variable bound by an operator (10). (10) a.

Who/what does Harriet love? For which x, x a noun, Harriet loves x? b. How does William look? For which y, y an adjective, William looks y?

wh-traces are standardly analysed as being involved in a range of structures: wh-questions, relative clauses (RCs) and topicalization. In each of these Move α extracts an XP from a position where it can legitimately check its features and receive a theta role, and substitutes it in an È-position, usually SpecCP. In order to justify this movement we must assume some sort of strong operator feature in the C projection of a wh-question which lures the XP to move, for in a corresponding non-operator construction the XP can legitimately remain in its base-generated position e.g. (11b). (11) a. Which unemployed will Gordon cut beneﬁts for? b. Gordon will cut beneﬁts for the undeserving unemployed.

NP-trace on the other hand occurs only in positions where Case features cannot be checked, such as SpecVP and the complement positions of intransitive verbs. The movement is thus more clearly motivated by Case.7 The landing site is constrained to be a position with no associated theta role, however, as NP-trace is assigned a theta role at its base position. The diﬀerences between wh-trace and NP-trace dependencies are often characterized as being those of È-chains versus A-chains, though this landing site related distinction is no longer tenable under the assumptions of the Minimalist Program, and has been recast as L-related vs. not L-related (Chomsky 1995: 64).8 Binding and referential diﬀerences between A-chains and È-chains (such as reconstruction, binding of anaphors) have been much discussed (see especially Chomsky 1981: Ch. 4, 1986a: Section 11, 1995: 200ﬀ;

7.Here too we should note some exceptions: clausal complements can undergo passivisation (i) and raising (ii), though slightly more controversially (see Koster 1978 for discussion). (i) That Helmut was corrupt was not generally realised until much later. (ii) That Jacques was incompetent seemed to occur to none of the delegates. 8.Essentially the È/A distinction breaks down because of the redistribution of argument positions within Larsonian structures (Larson 1988, 1990; see also Hale & Keyser 1993), and because of the VP-internal subject hypothesis.

17

18

Empty Categories in Sentence Processing

Webelhuth 1989 and Mahajan 1990, 1994), but the diﬀerences noted have been contested (e.g. Ueyama 1994; Müller & Sternefeld 1994; Lee & Santorini 1994). As these characteristics are controversial and not addressed in our study we shall not discuss the details here. There are two other major inﬂuences on the distribution of ECs: they are naturally restricted by lexical requirements but also constrained by the Empty Category Principle, which holds that trace must be properly governed (Chomsky 1981; Lasnik & Saito 1984; Chomsky 1986a).9 The eﬀect of this on NP-trace is to ensure that NP-movement is only local from speciﬁer to c-commanding speciﬁer position; the eﬀect on wh-trace results in cyclicity (Subjacency in GB terms). The processing eﬀect of the ECP is to license traces only in predictable positions in sentences, though the inventory of these positions is subject to some cross-linguistic parametric variation. Let us note that Chomsky (Chomsky 1995: 181) suggests that this might be replaced by the concept of minimal domain. The internal feature constituency of ECs tends to be discussed only as far as writers need to specify for the context in which they are using them. Chomsky (1982: Ch. 2: “General Properties of Empty Categories”) mentions only person, number, gender and Case features, and implies that the semantic content is deﬁned only by binding (Chomsky 1982: 31). Bouchard (1984: “On the Content of Empty Categories”) speciﬁes the grammatical features person, number and gender, and adds a referential index which determines the EC’s reference. Lasnik & Uriagereka (1988) suggest that intermediate traces have no binding features, as they are invisible at LF. Chomsky 1995 (p. 202ﬀ, p. 251ﬀ) outlines the copy theory of movement. This would entail that all the features, grammatical, semantic and phonetic, of a displaced element were copied into the derived position, and those at the second position deleted only by a PF process. All features must be present at both head and foot during the computation however, because this analysis achieves reconstruction eﬀects by using selective interpretation at LF instead. Pied-piped elements would be interpreted in situ and only the features directly involved in the checking relation interpreted in the raised position. Now, this requires that traces contain semantic features and implies that they contain phonetic features, though these too would presumably be only selectively implemented. This conception of the content of ECs will be

9.Empty Category Principle: a non-pronominal empty category must be properly governed. Proper government: α properly governs β iﬀ α θ-governs or antecedent-governs β (Chomsky 1986a: 17).

Empty categories in PPT and HPSG

of importance to us in our discussion of experimental evidence for their psychological reality in Chapter 3 below. 2.2.3 Summary There are four obligatory ECs assumed in PPT: pro, which is identical to an overt pronoun except that it is silent; PRO, which is the subject of inﬁnitival clauses; NP-trace, which resembles an anaphor in being locally bound and which is the product of movement for Case reasons; and wh-trace, which is the foot of an operator-variable dependency. These form an integral part of the PPT framework of interlocking constraints and could not readily be removed or replaced by an alternative mechanism without major revision to other parts of the theory. This makes PPT vulnerable to the claim that these constructs are unmotivated and not psychologically real; however, this also makes the assumption of ECs a strong claim. If empirical evidence were to appear conﬁrming that ECs play a role in processing, PPT would be much strengthened by the success.

2.3 HPSG and ECs Like PPT, HPSG exists in several variants. We shall follow the same consensual approach here as we did with PPT, again tending towards the analysis in the standard presentations (Pollard & Sag 1987; Pollard & Sag 1994).10 Several factors make it easier to select a standard analysis in HPSG than in PPT. While the literature is extensive, it is still much smaller than that of PPT, and workers in the framework lay more weight on descriptive adequacy. In addition the most recent standard work (Pollard & Sag 1994) has more of the character of a summation of previous work than a bold speculative step like the Minimalist Program. Perhaps most importantly, we shall discuss in some depth comments from Sag & Fodor (1993, 1995) about the processing predictions of HPSG. It is convenient to adopt the same version of HPSG as they do to avoid recasting their argument. We comment on the implications of alternatives in HPSG where appropriate.

10.We should note here that we adopt the version of Pollard & Sag 1994 Chapters 1–8, not Chapter 9, and so assume a SUBCAT list rather than a SUBJ and a COMPS feature. Nothing hangs on this decision.

19

20

Empty Categories in Sentence Processing

2.3.1 The position of ECs in the theory The stated aim of HPSG is to provide a model of human language as a tool for investigations on a scientiﬁc basis (Pollard & Sag 1994: 6f ). It is therefore concerned to be descriptively adequate and places weight on being true to the overt data. It is thus much less interested in producing solutions at higher levels of abstraction than PPT. For example, although both frameworks espouse blind and across-the-board principles, HPSG is more ready to admit exceptions where particular construction types demand it. A clear example of this is to be seen in the frameworks’ attitudes to inﬁnitival clauses. PPT excludes the idea that these could lack a syntactic subject and posits an EC; HPSG prefers to admit the possibility that unsaturated projections might exist rather than assume a constituent which is not represented in the phonetic form. We have seen in Section 2.1 above that the aspiration to be explanatory requires PPT to have its principles apply exceptionlessly, and that this is one of the chief motivations for the assumption of ECs. Since HPSG can countenance exceptions this reason for ECs does not apply to HPSG. HPSG aims to be explanatory in its own way however, recognizing that a more economical account of any phenomenon is a superior one, other things being equal. HPSG is thus mono-stratal and makes use of a single mechanism: local feature structures. Since it makes no use of movement constructions, the idea of a trace lacks much of its motivation. Its nearest equivalent to movement, structure-sharing, is a relationship of balanced token identity: there is no hierarchy or direction of information ﬂow involved. It is therefore not modespeciﬁc: it can operate equally well as a model of the linguistic knowledge involved in comprehension or production, without the assumption of movement traces, which we saw were necessary for PPT to make the same claim. The last major reason for ECs in PPT was movement to functional projections for checking purposes, especially common in the Minimalist Program. HPSG has no equivalent to functional projections and locates this sort of feature at the lexical head itself. This is indeed central to HPSG because it derives much of its descriptive power from its complex syntactic categories, which can in principle have values for any linguistic feature necessary for the framework to account for the data. Since lexical categories do not move to functional projections to check their features, very few of the movement operations in PPT have any equivalent at all. Thus even in a version of HPSG which assumes some movement traces, these are only a small subset of those

Empty categories in PPT and HPSG

which would appear in the equivalent structure in an Minimalist Program version of PPT. It must be clear therefore that the mechanisms, theoretical structure and explanatory priorities of HPSG do not force us to posit ECs in the way that those of PPT do. Equally they do not rule ECs out. We shall see below that HPSG has increasingly made use of this freedom not to assume ECs. 2.3.2 Passive, raising and equi constructions without ECs There is no analogue in HPSG to PPT’s NP-trace in raising and passive structures. Raising is the reﬂex of a speciﬁcation in the lexical entry of raising verbs: it is handled by means of a structure-share between an argument of the matrix verb (the “controller”), and the (unexpressed) subject which is subcategorised for by the propositional complement (Pollard & Sag 1994: Ch. 3.5). Passive is similarly lexically treated (Pollard & Sag 1987:215f, 1994:118ﬀ). Essentially it is a valency alternation which promotes an object to subject and demotes the subject to an optional adjunct. While PPT claims that the two structures are the product of the same process of movement for Case-checking reasons, HPSG accounts for the two quite separately. Raising, being driven by the lexical entry of the raising verb, is permitted to occur only in verbs which do not assign a CONTENT value to one of their arguments (the Raising Principle).11 Passivisation, by contrast, depends on the availability of a “primary object” of the verb which can become the subject. It must be clear that NP-trace is redundant on these accounts. HPSG distinguishes between equi and raising, but utilizes essentially the same mechanism in each. A subject equi verb subcategorizes for a subject and an unsaturated clausal complement whose unexpressed subject is co-indexed with the matrix subject (Pollard & Sag 1994: Ch. 3.5). The same relationship obtains between the object of the matrix verb and the unexpressed embedded subject in the case of object equi verbs. Again there is no EC: “We posit no phonetically unrealized PRO as a constituent in syntactic representations” (Pollard & Sag 1994:123). In both raising and equi structures it is the representation of the unexpressed subject on the SUBCAT list of the VP complement which performs similar functions to PPT’s ECs. We discuss these analyses in the light of experimental ﬁndings in Section 5.6 below.

11.Raising Principle: Let E be a lexical entry whose SUBCAT list L contains an element X not speciﬁed as an expletive. Then X is lexically assigned no semantic role in the content of E if and only if L also contains a (nonsubject) Y[SUBCAT ·XÒ] (Pollard & Sag 1994: 140).

21

22

Empty Categories in Sentence Processing

2.3.3 Traces in HPSG Some ECs remain however, especially in earlier versions of HPSG. One major diﬀerence from PPT’s ECs is that they are not subject to Linear Precedence rules, since these apply to phonetic features only. They are thus conventionally thought of as having no overt position in the phonetic string (though see Section 5.6.2 for an alternative view). In many other ways the two frameworks make very similar assumptions: HPSG’s hierarchy of nominal sorts contains three nominal sorts which are deﬁned by their binding behaviour: anaphors, personal pronouns, and non-pronouns (Pollard & Sag 1994: 252). Table 4.Nominal sorts in HPSG (adapted from Pollard & Sag 1994: 253) PPT’s noun features

+ana −pron

−ana +pron

−ana −pron

HPSG nom–obj sort

anaphor

personal pronoun

non-pronoun

Overt examples

herself

her

Maria

Empty examples

subject of complement null subject of ﬁnite of equi verb, trace clause, trace

trace

The sort anaphor is subject to Binding Principle A and consists of two overt subsorts reﬂexive and reciprocal, but it also has covert exemplars. These are the covert subjects of non-ﬁnite complement clauses (like for example PRO in PPT), which appear only on the SUBCAT list of the non-ﬁnite verb in raising and equi constructions. We shall discuss the processing reﬂex of these constituents in detail in Chapter 5 below. The next sort is personal pronoun and is subject to Principle B of the binding theory. Overt pronouns are examples of this. The last is sort nonpronoun, subject to Binding Principle C, and consists of referring expressions. Both of these have covert forms. Null subjects in languages like Spanish and Italian will be null forms of personal pronoun (Pollard & Sag 1994: 252). An EC is required here on this version of HPSG in order to prevent the direct object becoming the ﬁrst item on the SUBCAT list since this is the deﬁnition of a subject. In the absence of a subject EC a direct object would be the least oblique complement and would behave like a subject: rules of linear precedence would require it to precede the verb and agreement rules would make it agree with the verb for person and number (but see discussion of HPSG without traces below).

Empty categories in PPT and HPSG

HPSG’s trace, the rough equivalent of PPT’s wh-trace, can be any of these three nominal sorts. To see why this is the case we shall have to outline the approach taken to unbounded dependency constructions in HPSG, but, anticipating this, we can say that the categorial status of the trace depends upon that of its antecedent. Because the relationship between ﬁller and gap is one of structure-sharing, then if the antecedent is of sort personal-pronoun, then the trace will be this too; if it is of sort non-pronoun, then this again will be the CONTENT value of the trace. Trace is assumed within the model of HPSG presented in Pollard & Sag 1994 Chapters 1–8, though not in Chapter 9 or Sag & Fodor (1995) and much subsequent work. We shall brieﬂy outline the approach taken to unbounded dependencies in HPSG in order to set this discussion in context. Pollard & Sag 1994 note that they are not aware of any evidence that transformations correspond to anything empirical and therefore assume a monostratal grammar. Unbounded dependencies therefore consist of structuresharing between the locally subcategorised trace and the remote ﬁller, using the NONLOCAL feature.12 In fact there is a multiple membered structure-share: in order to see this we need to look at the internal content of a trace. The lexical entry of a trace contains very little information, but it contains a structure-share between the LOCAL value and one of the NONLOCAL features. These are QUE for wh-question words, REL for relative pronouns and SLASH for others. In (12) we illustrate this with a SLASH value. (12)

PHON 〈 〉 LOCAL

1

SYNSEM NONLOCAL word

synsem

SLASH { 1 } QUE {} REL {}

nonlocal

This structure-share is thus internal to the trace. This dependency is grounded by an external structure-share, which is between these two values and (the LOCAL subset of) the corresponding item on the SUBCAT list of the subcategorising head.

12.In fact we shall limit ourselves to strong unbounded dependencies here (i.e. those with an overt ﬁller) since weak unbounded dependencies have a rather diﬀerent structure (Pollard & Sag 1994: 166).

23

24

Empty Categories in Sentence Processing

In this way the trace gains a LOCAL value which is that speciﬁed by its role in the local structure. This information is all at the foot of the dependency, however, and it must be connected to the head to produce the unbounded dependency. The ﬁller at the top of the dependency therefore participates in this structure-share too, providing more detailed information still. Since structures in HPSG are always strictly local, this information is shared on the cascade model, with the NONLOCAL values being successively shared between mothers and daughters, all the way from the trace to the ﬁller. This is eﬀected by the Nonlocal Feature Principle (13). (13) Nonlocal Feature Principle: The value of each NONLOCAL feature on a phrasal sign is the union of the values on the daughters. (Pollard & Sag 1994: 162)

NONLOCAL features are thus passed up the tree in a series of local structureshares until they reach the ﬁller, which will be in a head-ﬁller-struc. They can then unify with the ﬁller and thus the structure-share between the ﬁller and trace is made (14).13 S

(14)

S [SLASH { 1 }]

NP [LOCAL 1 ] Crumble

NP I

VP [SLASH { 1 }] V [SUBCAT 〈NP, 2 NP〉] adore

NP [SYNSEM 2 ] [LOCAL 1 ] [SLASH { 1 }] (trace)

In such a structure the trace plays the same role as traces in PPT in satisfying local subcategorisation, but it is also the foot of the SLASH feature cascade

13.We suppress discussion of INHERITED and TO-BIND NONLOCAL features for brevity here. Essentially this distinction serves to prevent a NONLOCAL feature which has uniﬁed with a ﬁller being passed any further.

Empty categories in PPT and HPSG

which ensures that its LOCAL features unify with those of the ﬁller. Traces in PPT have no equivalent function, since the dependency between them and their ﬁller is essentially one of binding. HPSG traces are not bound because they are token-identical with their ﬁllers under the structure-sharing relationship. 2.3.4 HPSG without traces In Pollard & Sag 1994 Chapter 9 and Sag & Fodor (1995), building on previous work such as Kaplan & Zaenen (1989), this account of unbounded dependencies is modiﬁed to exclude traces. To see how this works it is useful to schematise the three dependencies between four locations which interact to produce unbounded dependencies (see Figures 1 and 2). Filler

LOCAL 1

Subcategoriser

SUBCAT 1

Trace

LOCAL 1 NONLOCAL 1

Figure 1.Unbounded dependencies with traces

There are three dependencies connecting the ﬁller, the subcategoriser and the LOCAL and NONLOCAL features within the trace, each one representing a declarative constraint. In order for the structure to be grammatical, the LOCAL features on the trace must indeed be subsumed by the SYNSEM constituent speciﬁed on the SUBCAT list of the subcategoriser. In addition, the NONLOCAL features of the trace must be identical to the LOCAL values of the ﬁller, and lastly the LOCAL and NONLOCAL features of the trace must be identical. This pattern of dependencies guarantees that these constraints are fulﬁlled. The same eﬀect is achieved by the traceless version, however. Filler

LOCAL 1

Subcategoriser

SUBCAT 1 NONLOCAL 1

Figure 2.Unbounded dependencies without traces

In this structure, the ﬁller is directly associated with a NONLOCAL feature on the subcategoriser itself, instead of on the trace. The (LOCAL subset of the) SUBCAT value is structure-shared with the NONLOCAL value. This arrangement

25

26

Empty Categories in Sentence Processing

produces similar eﬀects to that above but with one fewer dependency and without making use of trace. The other method by which more recent versions of HPSG have enabled ECs to be eliminated is by distinguishing between subjects and other complements (Pollard & Sag 1994: Ch. 9; Borsley 1996 and references there). Rather than coding all arguments by a single feature, the SUBCAT list, and deﬁning a subject as the least oblique complement, the newer approach splits them into SUBJ and COMPS features. The advantage of this is that it avoids direct objects becoming least oblique complements and thus behaving like subjects, if the subject position is empty. Thus, under the older system, the verb send in She sent a cake to Katie would have three items on its SUBCAT list: subject, object and indirect object, whereas the Spanish equivalent, with a null subject, would have two overt items on the list: object, indirect object (15b), but the subject would have to appear as an EC to prevent the direct object behaving like a subject. This could involve the direct object preceding the verb (15c). (15) a.

English: She sends a cake to Katie SUBCAT ·NPsubj, NPobj, PPobj2Ò b. Spanish: Manda una torta a Katie SUBCAT ·ECsubj, NPobj, PPobj2Ò c. Spanish: *Una torta manda a Katie *SUBCAT ·NPobj, PPobj2Ò

In the new arrangement, however, the SUBJ feature can be empty without aﬀecting the status of the direct object. (16) a.

English: She sends a cake to Katie SUBJ ·NPsubjÒ COMPS ·NPobj, PPobj2Ò b. Spanish: Manda una torta a Katie SUBJ ·Ò COMPS ·NPobj, PPobj2Ò

This method of eliminating ECs is discussed in Borsley (1996: 93ﬀ, 180ﬀ). It seems clear that the SUBCAT feature (sometimes known as ARG-S) must be retained for binding purposes and some representation of the null subject cannot easily be removed from it. Nevertheless, the sign for a trace can be dispensed with.

Empty categories in PPT and HPSG

2.3.5 Summary Earlier versions of HPSG made use of trace in unbounded dependencies, though no variants make use of trace in passive and raising where PPT has NP-trace. Traces are now thought of as theoretically undesirable, though some work has still to be done in order to reformulate structures which make use of them (e.g. the empty complementizers in Pollard & Sag 1994: 213ﬀ). There are some points in common between the two frameworks, for example, both assume that ECs can exist which correspond to the diﬀerent sorts of nominal objects in the theory. However, there are also clear contrasts between the two frameworks. While PPT is making increasing use of ECs in its drive to derive all of syntactic theory from economy conditions, the number of ECs in HPSG is reducing and may reach zero. Both frameworks agree that ECs are not psychologically plausible candidates for learning, but their solutions to this problem diﬀer. PPT concludes that ECs must therefore be innate, while HPSG prefers to eliminate them. These diﬀerent assumptions make a fertile ground for the development of testable contrasting predictions for processing. Since there are a number of methodologies which have been argued to be sensitive to the presence of ECs, we can seek empirical evidence to test the hypothesis that ECs are part of the representation assigned by the parser to an input stimulus containing an extraction. In the experiments we report below, we make use of certain aspects of German sentence structure in order to resolve some confounds which have prevented clear answers being obtained from experiments conducted on English. In order to conduct such experiments, however, we must ﬁrst specify what structure we assume for the German clause. This is the content of the ﬁnal part of this chapter.

2.4 The structure of German clauses It is particularly necessary for us to be explicit about the structure of the German clause in PPT, for there are a number of assumptions which are commonly made in the syntactic literature but which we do not intend to adopt. Let us note that we are not making stipulations for our own study; we are simply characterising the assumptions generally made by workers in processing. These alternative analyses do not radically change the predictions in our experiments, but they typically have a somewhat abstract character and their assumption within processing studies would be at some cost in syntaxprocessing transparency (see Chapter 3 below for discussion). An example of

27

28

Empty Categories in Sentence Processing

this is the constraint widely adopted in PPT that all structures should be binary branching (Kayne 1984; Chomsky 1995: 131), which is not generally adopted in work on linguistic performance. 2.4.1 German Clause Structure in PPT Since there are a range of possible clause structures within PPT we must be explicit about the precise analysis we are adopting. In each case our assumptions will be those which are consensual in the processing literature. In particular we must clarify what structure we assume for the German clause and the positions that direct objects occupy within it. Since we shall be testing for eﬀects at presumed trace locations, it is necessary for us to specify where we assume traces and why. We shall ﬁrst present the basic clause structure we assume, and then discuss some possible alternatives and extensions. For the sake of clarity we shall deal with each possible variant of clause structure separately, and exemplify it with English examples where the structural point is parallel to that in English. This permits us to avoid translating example sentences and commenting on other irrelevant diﬀerences. 2.4.1.1 The CP > IP > VP Model Following Chomsky (1986a), Fanselow (1988), Stechow & Sternefeld (1988) and Grewendorf (1988, 1995) we posit a conservative CP > IP > VP basic clause structure as in (17), with head-ﬁnal IP and VP. CP

(17) Spec

C′ C

IP Spec

I′ VP

Spec

I V′

XP

V

Empty categories in PPT and HPSG

This basic format allows us to account for the diﬀerent overt positions of ﬁnite verbs. (18) shows the diﬀerent possibilities for main clauses. (18) Main clauses a. Man spielt im Sommer Krickett (declarative) one plays in summer cricket a¢. Im Sommer spielt man Krickett (declarative) in summer plays one cricket a≤. Krickett spielt man im Sommer (declarative) cricket plays one in summer b. Man kann im Sommer Krickett spielen (declarative) one can in summer cricket play b¢. Man hat im Sommer Krickett gespielt (declarative) one has in summer cricket played c. Wann spielt man im Sommer Krickett? (wh-question) when plays one in summer cricket? d. Spielt man im Sommer Krickett? (yes/no-question) plays one in summer cricket?

In (18a) we see the ﬁnite verb in second position (V2). The verb is generated under V, raises to I to check its tense and agreement features, and raises again from the head I in the head-ﬁnal IP to the C position in the head-initial CP. The reason for this movement is not well understood, but it occurs widely in the Germanic languages, even in English in negative inversion structures. It is this movement to C of the ﬁnite verb together with the topicalisation of an XP which make up the pattern familiar from Germanic languages known as Verb Second (V2). The initial XP may be almost any constituent, but is usually either the subject or an element prominent in discourse (cf. (18a), (18a¢) and (18a≤)). If there is only one verb and it is the subject which topicalises, the resulting overt word order is SVO. (18b) and (18b¢) show that the inclusion of a modal or auxiliary removes this eﬀect since it is only the ﬁnite verbal projection that moves to C, while subsequent verbs remain sentence-ﬁnal. In wh-questions (18c) the wh-element occupies the SpecCP position, but in yes/no-questions (18d) this topicalised element is obligatorily absent; perhaps blocked by a null interrogative operator.

29

30

Empty Categories in Sentence Processing

(19) Embedded clauses a. Richie sagt, man spielt im Sommer Krickett (declarative) R. says one plays in summer cricket a¢. Richie sagt, im Sommer spielt man Krickett (declarative) R. says in summer plays one cricket a≤. Richie sagt, Krickett spielt man im Sommer (declarative) R. says cricket plays one in summer b. Richie sagt, daß man im Sommer Krickett spielt (declarative) R. says that one in summer cricket plays c. Richie fragt, wann man im Sommer Krickett spielt (wh-question) R. asks when one in summer cricket plays d. Richie fragt, ob man im Sommer Krickett spielt (yes/no-question) R. asks if one in summer cricket plays

A declarative embedded clause can have the same V2 form as a declarative main clause after certain verbs (see (19a), (19a¢), (19a≤)). However, it is more frequently introduced by a complementizer which occupies the head C, preventing the ﬁnite verb from raising beyond I (19b). In embedded interrogatives, the ﬁnite verb does not raise; in yes/no-interrogatives the head C is occupied by an interrogative complementizer (19d), and there is perhaps a null complementizer in wh-interrogatives (19c). While there is some freedom to vary the order of the arguments and adjuncts, we may discern two structural tendencies. If the subject has not moved up into SpecCP, then it will be the ﬁrst element in IP. It seems natural to attribute this to its location in SpecIP, where it is in a Spec–Head conﬁguration with the inﬂexional features on I. Next, the direct object is the last nominal argument of the verb. The canonical order of complements of the verb thus resembles a mirror image of their ordering in English. (20) English: verb > direct object > indirect object German: indirect object > direct object > verb

This seems to support the supposition that the German VP has complement> head order, but with speciﬁers preceding heads. In the sections that follow we discuss some variations on the basic clause structure we have detailed above, mainly additional articulations, and note the reasons for our choices between them. 2.4.1.2 Alternative layouts of functional projections We shall not discuss AgrSP, TP and the Split INFL Hypothesis (Pollock 1989; Belletti 1990) but throughout this work refer only to IP, as the precise internal

Empty categories in PPT and HPSG

constituency of these projections is unimportant here. For similar reasons and in order to standardise terminology across frameworks, we use NP, not DP, as an abbreviation for nominal projections. We discuss AgrOP under the section on object raising below. Travis (1984, 1991) argues that all Germanic languages should be analysed as having the same clause structure with a head-initial IP (21). CP

(21) Spec

C′ C

IP Spec

I′ I

VP Spec

V′ XP

V

When elements other than the subject are focussed, then they raise to SpecCP and the verb moves to C, while the subject remains in SpecIP, as we have assumed above. But, she argues, when no element is focussed then nothing moves to CP. In an SVO clause the subject is in SpecIP and the verb is in I. In this case no CP is projected, and the main clause is an IP, as in English. This diﬀerence to our structures is motivated by her assumption that the diﬀerences in word order between Germanic languages should be derivable without parameterisation of precedence relations, so that the underlying structure of an unmarked clause should be the same in English and in German. She also justiﬁes these main clause IP structures by pointing out that there are focus diﬀerences between initial subjects and other elements. For example, nominative pronouns can readily occur in initial position, but accusatives and dative pronouns may only do so if they are heavily stressed. She gives the examples we show in (22).

31

32

Empty Categories in Sentence Processing

(22) a. *Es hat er gegessen it has he eaten ?? b. Ihn hat er gesehen. him has he seen a¢. Er hat es gegessen he has it eaten b¢. Er hat ihn gesehen He has him seen

The examples (22a,b), which are only possible with very strong contrastive stress, contrast strongly with (22a¢,b¢), which are very natural. Travis suggests that this contrast is due to the structural diﬀerence between a CP in (22a,b) and an IP in (22a¢,b¢). However, there is a factor which Travis omits to comment on here, namely that there are two sets of pronouns in German: the i-pronouns (as in Travis’s examples above) and the d-pronouns, the diﬀerence being that the second set are used in more focused contexts. So the reason that Travis’s example sentences sound so awkward is that the inappropriate choice of pronoun has been made. If we replace them with the d-pronouns, the sentences sound wholly natural, whether the accusative (23a,b) or the nominative forms (23a¢,b¢) begin the sentence. (23) a.

Das hat er gegessen it has he eaten b. Den hat er gesehen. him has he seen a¢. Der hat es gegessen he has it eaten b¢. Der hat ihn gesehen he has him seen

The strong contrast can therefore be seen to be caused by factors other than those that Travis assumes. The nominative i-pronouns may be used in initial position simply because SpecIP is the default position for subjects if there is no other focused element. So if an accusative or dative appears sentence-initially, then it is necessarily focused, but the nominative may or may not be focused. This accounts for the diﬀerences she argues from. This naturally does not exclude her view that unfocused subjects are in IP, but undermines evidence that she uses to support the view. There is however other evidence which makes it unlikely that German SVO structures are IPs: the impossibility of SVO after a complementizer. Recall that we argued that a

Empty categories in PPT and HPSG

complementizer forces the ﬁnite verb to remain in sentence-ﬁnal position because it occupies the head C position. Travis’s account predicts that unfocused SVO structures should be possible after a complementizer or indeed a wh-item. (24) Head-ﬁnal IP account: a. Angus sagt, [CP Darren [C wirft] [IP tsubj den Ball [I tverb]]] Angus says Darren throws the ball b. Angus sagt, [CP [C daß] [IP Darren den Ball [I wirft]]] Angus says that Darren the ball throws c. *Angus sagt, [CP Darren [C daß wirft] [IP den Ball [I tverb]]] Angus says Darren that throws the ball d. *Angus fragt, [CP wann Darren [C wirft] [IP den Ball [I tverb]]] Angus asks when Darren throws the ball (25) Travis’s account: a. Angus sagt, [IP Darren [I wirft] den Ball tverb]] b. Angus sagt, [CP [C daß] [IP Darren [I] den Ball wirft]] c. *Angus sagt, [CP [C daß] [IP Darren [I wirft] den Ball tverb]] d. *Angus fragt, [CP wann [C] [IP Darren [I wirft] den Ball tverb]]

The examples (24a) and (25a) are both possible, our version embedding a CP and Travis’s an IP. Our analysis (24b) shows how the presence of a complementizer prevents I to C movement, while Travis’s (25b) fails to motivate the verb staying in the VP, since the head I position is unoccupied. Our (24c,d) show why V2 is impossible in embedded clauses with a complementizer or wh-item, because these occupy the landing sites. Travis’s (25c,d) shows no reason to exclude these sentence forms. On these grounds we reject the suggestion that German has a head-initial IP (see also arguments advanced by Kathol (1990) and Schwartz & Vikner (1989)). Kathol (1990) suggests that all German clauses are IPs, and that only complementizers appear in CP, arguing that the verbs and complementizers do not form a natural class, and so cannot alternate in the head C position. He claims that coordination facts refute the CP analysis of clauses: for example he says that (26) and (27) are marginal and that this excludes the possibility that both clauses with complementizers and V2 clauses could be CPs. (26) ?Otto glaubt, daß zwei mal zwei vier sei und alle übrige Mathematik sei Otto thinks that two times two four is and all other mathematics is Zeitverschwendung. wasted.time

33

34

Empty Categories in Sentence Processing

(27) ?Otto glaubt, zwei mal zwei sei vier und daß alle übrige Mathematik Otto thinks two times two is four and that all other mathematics Zeitverschwendung sei. wasted.time is

He also claims that complementizers and ﬁnite verbs do not form a natural class, and are therefore unlikely alternative occupants of the C position. He suggests (following Travis 1984) that the ungrammaticality of sentences like (25c) above (repeated here as (28)) is due to the complementizer “identifying” the I position through “head feature transmission”, making its overt appearance unnecessary and illegitimate. (28) *Angus sagt, [CP [C daß] [IP Darren [I wirft] den Ball tverb]] Angus says that Darren throws the ball

He further argues that the grammaticality of coordinated sentences such as in (29a,b) provide independent evidence of the identiﬁcation of I by C. (29) a.

Wenn du nach Hause kommst und der Gerichtsvollzieher vor der if you to home come and the bailiﬀ at the Tür steht… door stands b. Wenn du nach Hause kommst und der Gerichtsvollzieher steht if you to home come and the bailiﬀ stands vor der Tür… at the door

Example (29a) is a clear example of two coordinated clauses forming a condition, and both conjuncts have the same verb-ﬁnal word order. In (29b), by contrast, the second conjunct has V2 word order, but it is nevertheless grammatical. Kathol suggests that this is because the complementizer only identiﬁes the ﬁrst I position. A very similar analysis but with the additional assumption of underlying VO order in line with Kayne’s (1994) Linear Correspondence Axiom (see Section 2.4.1.4 below) is presented by Zwart (1997). We shall not adopt these analyses here, noting that (26) and (27), although very marked, are not wholly ungrammatical, and their awkwardness is probably stylistic. They do not therefore exclude the possibility that both clauses are CPs. In addition, Kathol’s objection that the categories which can occupy C not being a natural class is weakened by a comparison with English, where exactly the same elements occupy C under almost identical circumstances, and where there is far stronger

Empty categories in PPT and HPSG

evidence that raised verbs move into head C. Furthermore, we doubt the empirical basis of the identiﬁcation of I by C, as this account seems to lack conﬁrmatory evidence. Lastly, we suggest that example (29b) which Kathol holds to be coordination of unlike pairs is not strict coordination but probably some looser paratactic connection: und has other functions as well as just coordination (e.g. der und singen können!, und so kam es, daß…). This use of und may be ordering. This becomes perhaps clearer if we make (29b) into two main clauses: Ich komme nach Hause und der Gerichtsvollzieher steht vor der Tür “I come home and the bailiﬀ is standing at the door”. The und gives no feeling of coordination here but merely ordering. We would argue that it is this use of und that we see in (29b). Haider (1993) argues that there is no evidence of an independent IP and VP in German and that the two categories can be subsumed into one. This gives the structure as in (30). FP(CP?)

(30) spec

F´ F

IP/VP spec

I´/V´ XP

I/V

We admit that it is diﬃcult to advance any ﬁrm proof of an IP separate from the VP using German data (but see Featherston 2000 for some recent experimental work), but note that there is for the same reasons no evidence against it either. The evidence both for and against can only ever be circumstantial since the head positions are necessarily adjacent in German. However, other languages oﬀer plentiful evidence of the non-identity of these two categories, and we shall assume this for German too, as we are unaware of any reason for thinking that German falls radically outside the mould of UG. Let us note that such a structure would make no diﬀerence to the predictions in our experimentation, but it would endanger our ability to draw conclusions about the structure of German from work done on other languages.

35

36

Empty Categories in Sentence Processing

2.4.1.3 Binary branching It is frequently assumed within PPT that a necessary constraint on phrase structure rules is to permit only binary branching nodes, not ternary or greater branching. Kayne (1984) argues in favour of this on the basis that it provides for “unambiguous paths”, while other researchers (e.g. Haegemann 1994) have shown that it can provide a signiﬁcant contribution towards solving the logical problem of language acquisition, since it reduces the number of possible structures for a given string. Most syntactic work within the framework of PPT tends to assume some version of binary branching. This constraint involves some complications in the analysis of co-ordinate structures (Goodall 1987) and more importantly for us here double object constructions. We shall brieﬂy review the implications here. Kayne (1984) suggests that the complement of a double object verb is in fact a single object consisting of the union of the two intuitive complements (31). (31) give [Mary books]

Haegemann (1994) presents a structure with the direct object as a sister of V¢ and the daughter of VP: (32) [VP [V¢ [V give] Mary] books]

Larson (1988, 1990; see also Jackendoﬀ 1990) suggests structures with additional verbal positions for third arguments. Consider the partial derivation (33) showing constituents in their base generated locations. vp

(33) NP John

v´ VP

v e

NP books

V´ V

PP

gave

to Bill

Empty categories in PPT and HPSG

The verb raises to the light verb position in the VP shell, thus giving the observed order of S > V > DO > IO, John gave books to Bill.14 The assumption of a Larsonian shell makes no empirical diﬀerence to the location of the traces relative to the overt position of the verb, but it does alter the relationship of certain arguments with the verb. We shall therefore not adopt the suggestion of multiple verbal projections in double object constructions, noting that Chomsky (1986a:2f, 1995:61f ) has not always committed himself to exclusively binary branching structures either.15 Again we adopt the consensus position in processing studies. We therefore assume that verbs are able to subcategorize and theta-mark more than one complement. There are two possible mechanisms for this: one is to assume a ternary branching structure such as (34). V´

(34) V

NP

PP

give

books

to Mary

This allows the complements to be base generated in their canonical positions relative to the verb and permits them to be its direct syntactic dependents; however, it violates binary branching. Alternatively one could envisage a hierarchical arrangement of the direct and indirect objects, as in (35) (e.g. Chomsky 1981: 171). VP(?)

(35) V´(?)

PP

V

NP

give

books

to Mary

14.Here and elsewhere S = subject, V = verb, DO = direct object, IO = indirect object. 15.In later discussion (1995: 179ﬀ, 305ﬀ, 329ﬀ, 356ﬀ) Chomsky does seem to adopt this analysis. However, even here it constitutes an assumption to be made, rather than a necessary fact: “Assume binary branching” (Chomsky 1995: 177).

37

38

Empty Categories in Sentence Processing

The weak points in this analysis are the weakening of the generalisation that theta roles are assigned under most local conﬁgurations (Chomsky 1981: 94), generally sisterhood, and the lack of any clear categorial status for the mother and sister nodes of the indirect object PP. Nevertheless, such structures are commonly assumed in work with a descriptive rather that theoretical emphasis (see e.g. Stechow & Sternefeld 1988: 433ﬀ for German). For our purposes, either of these two structures will be appropriate, there being no diﬀerence in their empirical predictions for our study. 2.4.1.4 Universal word orders Kayne (1994) argues against the standard assumption that head complement order is set by a head parameter, which has two values: head initial, like English, and head ﬁnal, like Japanese. He posits that base generated order is imposed by the Linear Correspondence Axiom which states that constituent order universally reﬂects the structural hierarchy. The vehicle he posits for this is Asymmetric C-command, which essentially requires that c-command imposes a unique order on any pair of categories, thus excluding mutual c-command. Kayne’s speciﬁc formulation of this derives a universal SVO order for base generation, all other orders being derived by movement. While Chomsky assumes this in the most recent versions of the Minimalist Program (1995: Ch. 4) and the implications for languages like German have been discussed (e.g. Zwart 1997), it is not generally adopted in work in processing of head-ﬁnal languages, as it represents a considerable abstraction from the observable data. Assuming the movement from post-verb to pre-verb position to be motivated by Case checking (see discussion of object raising below), this would produce a derivation like (36) for a simple transitive verb in a language like German generally assumed to have head ﬁnal VP.16

16.We apply here the assumptions of Chomsky 1995 Chapter 4.10 in the variant discussed on p. 358 where object raising takes place before the merger of the subject.

Empty categories in PPT and HPSG

(36)

vp subj

v´ obj

v´ VP

v v

v

tv

tobj

This structure yields SOV order at one stage of the derivation: assumptions about the distribution of strong and weak features can readily permit this ordering to apply at Spellout and thus enter the phonological component. Kayne’s Linear Correspondence Axiom has its equal and opposite in Haider’s (1992, 1993) Basic Branching Conjecture, which holds that basegenerated projections are universally left-branching.17 Haider argues that the prediction of a VP head direction parameter is that arguments should appear in mirror image order in head-initial and head-ﬁnal VPs, but that this prediction is not fulﬁlled. He suggests that the basic word order of IOs, DOs and PPs is the same in English and German: IO > DO > PP (37). (37) Geoﬀrey gave Peter(IO) a loan(DO) for his house(PP) …weil Geoﬀrey Peter(IO) ein Darlehen(DO) für sein Haus(PP) gab …since Geoﬀrey Peter a loan for his house gave

The Basic Branching Conjecture holds that all languages are underlyingly headﬁnal and project as (38). (38) [XP3 [XP2 [XP1 V]]]

Since the direction of projection and the direction of government are the same in a language like German, they have this basic structure. Languages such as English govern to the right but must also project to the left. As a result the verb is forced to project extended head-V positions and move up to them, in order to govern additional argument positions. This gives the structure (39), which

17.Basic Branching Conjecture: Base generated projections are left-branching (Haider 1992, 1993, my translation).

39

40

Empty Categories in Sentence Processing

Haider compares to Larsonian structures with light verb positions to which more embedded verbs can move. (39) [Vi [XP3 [ti [XP2 [ti XP1]]]]]

We shall adopt neither Kayne’s nor Haider’s analysis here, but rely on the setting of the head parameter as the mechanism for producing languages with head > complement and complement > head orders, even though it is frequently assumed parameters apply only to functional categories. In doing so we are as usual adopting the consensus position, for this is a standard assumption in processing studies of languages with head-ﬁnal constituents. In our experiments we shall be testing for reactivation at the location of argument traces. Since the status of intermediate traces is controversial it is most productive for us to be able to restrict our attention to traces occurring in base generated positions. The assumptions of a head ﬁnal VP in German allows us to exclude intermediate traces, and be maximally explicit about the positions at which PPT does and does not predict trace eﬀects to occur. 2.4.1.5 Object raising Chomsky (Chomsky 1995:120f,146ﬀ) argues that it is unsatisfactory that Case theory makes use of two diﬀerent structural relations: head–complement for accusative, and head–speciﬁer for nominative. Pollock (1989) had previously showed evidence for the split INFL hypothesis: separate tense and agreement phrases replacing INFL. Chomsky reduces all Case checking conﬁgurations to Spec–Head by positing an agreement phrase for objects too (AgrOP). Under this proposal object phrases raise to SpecAgrO and verbs raise to adjoin to the AgrO head, at which point the two are in a Spec–Head relation and can check their Case and perhaps agreement features (depending on language-speciﬁc morphology). The evidence for this concerns structures such as (40) (see Kayne 1989). (40) a.

Paul a repeint les tables Paul has repainted the tables b. Paul les a repeintes Paul them has repainted c. Combien de tables Paul a repeint(es) how.many part tables Paul has repainted (adapted from Kayne 1989)

In (40a) there is no evidence of object raising and the verb does not agree with the object. In (40b) the object clitic has moved perhaps to SpecAgrO and so the

Empty categories in PPT and HPSG

verb agrees for number and gender. (40c) shows that the overt agreement is optional if the object has undergone operator movement. Kayne’s suggestion is that the verb agrees only if the object raised to the SpecAgrO position before undergoing operator movement. The concept of object raising thus has some foundation in empirical data. Chomsky hypothesizes that this process occurs in languages such as English too, thus providing a uniﬁed clause structure across languages at the same time as homogenising the Case and agreement checking conﬁgurations. Given that we shall be testing for experimental eﬀects at direct object positions, it is necessary for us to be clear about where we assume direct objects to be generated. However, while object raising produces at least a temporary inversion of object and verb order in head-initial languages such as English and French, in a head-ﬁnal VP this does not occur. Object raising produces no overt change in the position of the direct object. IP

(41) subj

I´ AgrOP

obj

verb AgrOP´

VP tsubj

tv V´

tobj

tv

In (41) we illustrate the results of object raising on the assumptions about clause structure of Chomsky 1995: Ch. 3, with a AgrOP between the VP and the IP.18 The object has raised to SpecAgrOP from its base-generated position immediately preceding the verb. The verb has moved to AgrO and subsequently to I.

18.Notice that we simplify the features adjoined to I here and below. On Chomsky’s (1995: Ch. 3) assumptions these should be the head features of V adjoined to the head features of AgrO adjoined to the head of I (perhaps T and AgrS).

41

42

Empty Categories in Sentence Processing

The subject has raised from SpecVP to SpecIP. The salient fact to note here is that the overt order S > O > V is unchanged, and so this model of object raising provides no obstruction to the working hypothesis we adopt in our experimentation that the direct object position is immediately preceding the verb. vp

(42) subj

v´ obj

v´ verb23

VP tobj

tv

In (42) we apply the assumptions of Chomsky 1995 Chapter 4.10: 358. The AgrOP projection is replaced by additional speciﬁer positions on the upper verbal projection; the object-verb checking relation is attained when the object raises to the (lower) Specvp position, and the V adjoins the upper verbal position v. The surface order is unaﬀected and the object is located immediately preceding the verb. Again the overt SVO order is not compromised. However, we shall not assume object raising here for several reasons: ﬁrst, there is no overt morphological reﬂex of object raising in German; that is, verbs never agree with their objects. Next, many syntacticians prefer to posit diﬀerent Case checking conﬁgurations for nominative and accusative as part of the determination of the diﬀerences between them (e.g. Bobaljik 1995; Boškovic´ 1995). It is thus taken to be part of the deﬁnition of structural accusative Case that it is checked in a head-complement conﬁguration. This naturally removes the requirement for the object to raise to SpecAgrO. Third, in our study we shall consider double object constructions, and the interaction of object raising with these has been insuﬃciently worked out for us to use it as a descriptive basis for the analysis of our experimental structures.19 The fourth and last reason why we shall not assume object raising is that it predicts a multiplication of the

19.For example Chomsky (1995: 285 and 386, fn. 55) suggests that NPs with inherent case do not need to check their φ-features because they are interpretable. Radford (1997) on the other hand shows how the assumption of an AgrIOP can account for a number of word order phenomena with three-place particle verbs.

Empty categories in PPT and HPSG

number of traces in the proximity of the verb, and this in turn would make the ﬁnding of trace reactivation at any one position unlikely. In order to establish an eﬀect at any one testpoint in a sentence, it is necessary to show that it diﬀers from another control testpoint. If we assume multiple argument movements even in a simple clause, no suitable control position will be available, because they will be contaminated by a background eﬀect of multiple trace position eﬀects. To illustrate this, let us consider a clause structure in which a direct object has been extracted from its canonical preverbal position and moved to SpecCP. CP

(43) obj

C´ C

IP subj

I´ AgrOP

tobj

verb AgrOP´

VP X

tv V´

tobj

tv

While there is a trace at the canonical object position, sister of V, daughter of V¢, even this minor articulation of the clause structure means that the identiﬁcation of an earlier control position becomes considerably more diﬃcult. This is particularly the case because identity of the position X is disputed. On the VP-internal subject hypothesis (e.g. Sportiche 1988), the position X would be the subject trace. However, on a Larsonian (1988) analysis of the position of prepositional arguments, X would be the PP position, a position it would not leave. It must be clear that additional NP or PP arguments or indeed additional functional projections in the clause would complicate the picture further. Since the position of the intermediate trace is not easily locatable and we anyway

43

44

Empty Categories in Sentence Processing

doubt that it will have any processing reﬂex, we shall assume that objects do not raise in German. 2.4.1.6 Predicate Linking Rothstein (1983) puts forward her Predicate Linking Rule which essentially suggests that every predicate must have a subject.20 This has been inﬂuential: Chomsky (1986b: 116, 1986a: 92 fn5) attributes the Extended Projection Principle to this. In addition, and more importantly for our purposes, this implies multiple NP-traces in raising and passive constructions with modals or auxiliaries. For example in (44) we might be led to assume as many as seven intermediate trace positions. (44) Morgani seemed ti to ti be ti thought ti to ti have ti been ti discredited ti

For the reasons we discussed above, such multiplication of ECs renders investigation into their processing virtually impossible. In our own work we avoid structures which would be vulnerable to such complications, but we shall also forbear to comment on the implications of this theory for other experiments on raising and passive sentences. 2.4.1.7 Summary It will be clear from the discussion above that clear predictions for the processing of ECs can only be made on the basis of syntactic analyses which are close to the overt data. There are a number of steps that we can take and are generally taken to avoid abstract analyses clouding the issues. We adopt a CP > IP > VP clause structure with a head-ﬁnal IP and VP. We shall not assume object raising, binary branching, predicate linking nor Kayne’s nor Haider’s universal word orders. Next, we avoid any possible structural ambiguity by excluding sentences with PP arguments or pre-verbal negatives, either of which might cause uncertainty as to the location of the direct object trace. Thirdly, we shall restrict our discussion to traces at the foot of dependencies, a move for which there is both theoretical and empirical justiﬁcation. Chomsky (1995: 300ﬀ) argues that [−interpretable] features will be deleted if they are recoverable. Since the intermediate traces in a derivation are usually recoverable, it is likely that many if not all intermediate traces are inaccessible to the computational

20.Predicate Linking Rule: Every non-theta-marked XP must be linked at S-structure to an argument which it immediately c-commands and which immediately c-commands it (Rothstein 1985: 11).

Empty categories in PPT and HPSG

system beyond a certain point in the derivation. It would be a reasonable assumption that such traces are less psycholinguistically active than traces at the foot of dependencies since it is exactly these traces which correspond to the processing notion of a “gap” (see Ch. 2 below for discussion). Thus there are good reasons to restrict investigations to these ECs alone. 2.4.2 German clause structure in HPSG While it is possible to identify a consensus on the structure of German clauses in PPT, in HPSG the balance of the alternatives is more even. This would be a signiﬁcant problem for our study were it not that the basic diﬀerence between PPT and HPSG that we wish to address with our experiments is the existence of ECs, in particular object trace. This makes it imperative that we specify exactly what clause structure we assume in PPT, because the location of the traces that we shall test for depends upon it. The HPSG analysis that we shall be comparing does not assume traces for non-local constituents, and so it is not necessary for us to be quite so explicit about our assumptions. Nevertheless, we shall brieﬂy review the competing proposals here. Much of the detail of the structure of a German clause is identical to that of an English clause, and therefore consists of a sign of sort phrase with a head daughter and complement daughters. Thus far the analysis is uncontroversial. The diﬃculties appear when one wishes to account for the diﬀerence between (45a) and (45b). (45) a.

Kaut Svenja öfter Gummibären? chews Svenja often jelly.babies b. …daß Svenja öfter Gummibären kaut. …that Svenja often jelly.babies chews

In (45a) the verb is initial, in (45b) ﬁnal. Since Linear Precedence in HPSG depends on the order of sisters on a SUBCAT list, these contrasting orders require some explanation. Put informally the problem is as follows: how can a constituent occupy two so diﬀerent positions in what otherwise looks to be an identical clause? Notice that the preposing of an XP in a declarative form of (45a) to make a V2 clause is not at issue, as this is standardly considered to be topicalization. The problem relates thus only to the location of the verb, which is considered to be in either initial (45a) or ﬁnal (45b) position. The justiﬁcation of orders of phonetic features is the content of Linear Precedence (LP) rules, which specify in what order sisters appear. This is

45

46

Empty Categories in Sentence Processing

governed by the principle of Exhaustive Constant Partial Ordering, which basically states that any given set of sisters should be ordered the same, no matter what their mother or daughters. The orders are deﬁned by the concept of Obliqueness (i.e. order of arguments on SUBCAT lists) plus a number of other factors such as lexicality. Our problem here is caused by Exhaustive Constant Partial Ordering. This permits only one ordering of a given set of constituents, but our clause above seems to show two. There are a number of possible solutions to this problem (see also discussion in Borsley 1996: 220ﬀ). Pollard (1991) assumes that German has a ﬂat VP and generates the diﬀering verb positions with a linear precedence rule. This requires the initial and ﬁnal verbs to be distinguished by a feature. As soon as they are distinguishable, then a linear precedence rule can identify them separately and produce the desired word order variations. Pollard (1990) uses this to account for German V2. He utilizes the [INV] feature which was originally introduced to account for auxiliary inversion in questions in English. When [+INV], main clause ﬁnite verbs precede all their complements (except any topicalized ones). S

(46) V[+INV]

NP

NP

Kaut chews

Svenja Svenja

Gummibären? jelly babies

Hinrichs & Nakazawa (1994) use the similar approach and a [FLIP] feature to deal with the auxiliary ﬂip in (47). (47) Niemand glaubt, daß Harald das hätte machen können nobody believes that Harald that had do can “No one believes that Harald could have done that”

In this sentence the hätte must be [+FLIP] to produce the desired order. There are disadvantages to the account however. Netter (1992) argues that it requires a commitment to an entirely ﬂat clausal phrase structure for German, which would imply that English and German clauses should behave very diﬀerently in subcategorisation, for example. German verbs would combine with all of their complements simultaneously, whether they be arguments, adjuncts or particles, whereas in fact scopal asymmetries seem to occur which disconﬁrm this. Netter also points out that a ﬂat structure is incompatible with

Empty categories in PPT and HPSG

the Head-Adjunct phrase structure schema, in which the semantic head is the adjunct while the syntactic head is the modiﬁed phrase. This method of enabling the inheritance of semantic features is only compatible with the attachment of a single adjunct or complement at a time. However, this is the nearest to a standard answer to the problem, as Pollard & Sag argue in favour of it in Pollard & Sag 1994 (pp. 38–42). A suggestion by Jacobson (1987) applied to clause structure by Borsley (1989) is to use an additional SLASH mechanism for non-local arguments and apply it to the fronting of verbs. This approach essentially takes the movement analysis of PPT but gives it an HPSG implementation. The advantage of such a system is that it preserves the structure of the remainder of the clause and makes no demands on the structure of the rest of the sentence, permitting a hierarchical or ﬂat S, as other factors demand. The disadvantage is that it requires some representation of the verb at its canonical position to serve as the repository of its subcategorisation, essentially, a trace. While this is a possible analysis, it does not make use of the strengths of HPSG, and sits uneasily with the general rejection of movement in the framework. Netter (1992) argues that other data such as scopal asymmetries make it necessary for the German VP to be analysed as hierarchical. He hypothesizes that there is an empty functional head in ﬁnal position in V2 clauses, and that the initial ﬁnite verb is itself not a head, but a complement of it. Since the functional head must subcategorize for the same complements as the initial verb, but in addition for the verb itself, this involves specifying the SUBCAT list of the functional head as being a structure-share of the SUBCAT list of the initial verb with the verb itself appended to it. Netter introduces a functional constraint snoc to perform this. This approach suﬀers some of the disadvantages of the previous one, in that it requires the assumption of a trace-like category, but additionally necessitates an operation on SUBCAT lists which has no precedent. None of the accounts of the variable verb positions in German clauses that we have considered can be seen as wholly satisfactory, since each has its drawbacks. For explicitness we shall assume that variable verb position is driven by an [INV] feature, while the initial clausal element is topicalised using a [SLASH] feature. This gives the following structures for embedded and unembedded clauses.

47

48

Empty Categories in Sentence Processing

(48)

...weil because

S [–INV]

NP

NP

P

V [–INV]

Paul Paul

das Licht the light

an on

macht puts

“… because Paul puts the light on” S

(49) NP

Paul Paul

S [INV+] [SLASH NP] V [INV+] [SLASH NP]

NP

P

macht puts

das Licht the light

an on

“Paul puts the light on”

2.5 Summary In this chapter we have seen that the two syntactic frameworks hold very diﬀerent views of ECs, in part inﬂuenced by their internal theory structure and explanatory priorities. Aspiring to be an account of language acquisition and to derive all constraints from outside the language module, PPT must insist on its principles applying blindly and exceptionlessly. HPSG lays greater weight on descriptive adequacy and evades where possible the assumption of categories which lack overt form. Since these issues have not been settled deﬁnitively by the process of conventional syntactic debate, it is tempting to apply experimental methods to them. A number of researchers have preceded us in this venture. We review the results of their work in Chapter 3 below.

AUTHOR ""

TITLE "Parsers and gap processing"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 3

Parsers and gap processing

3.1 Introduction We can deﬁne the human sentence processing mechanism, or the parser or (sentence) processor as we shall generally refer to it here, as the linguistic knowledge (often assumed to be a grammar) applied to sentences or sentence fragments plus whatever extra faculty, knowledge or functions are necessary to apply it and achieve comprehension. Since on most accounts we have two independent variables (the application function and the grammar) and only one dependent variable (the performance data), and, in addition, the boundary between the grammar and the application function is ill-deﬁned, it can be diﬃcult to establish clearly whether a result reﬂects on the application function or the grammar itself. For this reason researchers usually take the grammar to be part of the parser and phrase their conclusions in terms of this. Questions which need to be answered in a full account of the structure of the parser would include details of: –

–

the external relations of the parser, i.e. the boundaries of competence with non-linguistic processing, and whether these boundaries are permeable or impermeable to information from outside. the internal relations of the parser, i.e. the division of labour of the various subparts, if these exist.

Information we might expect about the operation of the parser would include: – –

the input and output representations what computational processes it uses

Since the parser is a “black box” phenomenon in the same way as the grammar, and we are only able to study the parser’s characteristics indirectly, we are only able to address a limited number of secondary questions about the parser, the answers to which can be used to make deductions about the primary questions above. Research has so far managed to address such issues as:

50

Empty Categories in Sentence Processing

– – – – – – –

the time course of processing what information sources are available to the parser how the processor deals with sentential ambiguity what determines processing complexity to what extent the parser seems to compute representations compatible with existing grammatical models the ways in which syntactic and semantic processing interact how the processor assigns syntactic structure to phonologically null input

Before we consider the evidence that has been gathered about the parser we shall brieﬂy point out those issues which we shall not discuss at any length, but rather assume (somewhat sketchy) answers to. One basic premise of this work is that a theory of grammar must be a model of human language competence, that parsing is one example of the application of this competence, and that by examining empirical data of sentence processing it is possible to make conclusions about the grammar. This is usually known as the Strong Competence Hypothesis (Bresnan 1978; Bresnan & Kaplan 1982; see Berwick & Weinberg 1984 for the similar Type Transparency). This transparency, as we shall call it, is not held by all researchers in the ﬁeld of sentence processing, for some take the view that there is no representation like a grammar in the mind (e.g. McClelland 1987). We shall assume that the parser is universal much as UG is (e.g. Representational Hypothesis of Frazier 1985; see e.g. Bates et al. 1982; MacWhinney & Bates 1989 for the alternative view). On this basis, one can meaningfully compare processing results from diﬀerent languages in the same way that one can compare syntactic phenomena. This does not exclude the possibility that there are subparts of the parsing mechanism that are made more use of in some languages than in others, but suggests that these will be diﬀerences of use rather than availability, perhaps akin to such inter-language variation as word order and the extent of morphological marking. Another assumption that we shall make is that the input to the sentence processor is at the word level, and the output at the sentence level. Thus we shall not discuss lexical access or processing, nor the additional interpretation that is available to input strings consisting of more than one sentence. Plainly these limits are somewhat arbitrary, and evidence could be adduced to suggest that information sources beyond these boundaries play a role at the sentence level, but we limit ourselves to the processing of sentences here.

Parsers and gap processing

3.2 Parameters of variation amongst parsers There are two parameters along which parsing models standardly vary: serial to parallel, and modular to interactive. The ﬁrst refers to the parsers ability to compute more than one structure simultaneously for an ambiguous input, and the second refers to its architecture as deﬁned by information ﬂow. Frequently both are used together, as parsing models tend to be either serial and modular, or parallel and interactive (see e.g. Mitchell 1994). 3.2.1 Parallel vs. serial The most common perspective on the structure of the parser asks how it deals with ambiguity. This issue is much discussed in the literature and we shall do no more than summarise it here (see MacDonald 1994; Gorrell 1995: Ch. 3; Crocker 1996: Ch. 2 for broad brush discussions). The ways that a parser might behave when confronted with input of indeterminate structure and interpretation are often classiﬁed into three types: serial processing, delayed processing and parallel processing. Serial parsers are those which make instant phrase structure decisions as the input arrives. They are generally modular parsers as this mode of operation ﬁts well with such an architecture. For example, Frazier’s model (1985, 1987) has an automatic and autonomous structure building module that makes quick phrase structure decisions and uses syntactically-deﬁned simplicity-based strategies to make initial attachment decisions when other data is insuﬃcient. At the point of local ambiguity, this initial module makes an immediate decision and passes only this chosen analysis on to the next modules for checking. This is argued to make for processing eﬃciency, as all analyses except one are rejected before any computational eﬀort has been expended on calculating their exact syntactic status and reference, which would take place in later modules. A disadvantage is that wrong decisions will be relatively frequent and may cause signiﬁcant extra eﬀort to reanalyse. This is the garden path eﬀect which is even noticeable to the conscious reader in particularly severe cases. More recent instantiations of serial parsers have tended to compromise somewhat and have features which fudge the clear predictions of the serial parser. For example, Crocker (1996) allows the parser to vary its commitment to its analysis, depending on the goodness of the ﬁt into the context, and also suggests that diﬀerent depths of reanalyses will aﬀect diﬀerent numbers of modules. Both of these weaken the strong predictions of serial processing and allow

51

52

Empty Categories in Sentence Processing

compatibility with a wider range of data. The approach of Delay Models may be summarised as “wait and see”. For example Berwick & Weinberg (1984, 1985) describe a grammar implementation parser which operates deterministically. When the input is ambiguous such a model puts oﬀ making decisions until disambiguation arrives. Thus, in (50) below, on analyzing as far as believes and receiving his sister in the input, the parser does no further computation, but rather looks at the next token in the input, since the continuation of the sentence cannot be uniquely determined (51). (50) [IP[NPJohn] [Vbelieves]] … [?his sister] (51) John believes his sister. John believes his sister is smart. John believes his sister being arrested is a disgrace.

More recently this approach has been developed by Weinberg (1992) and Gorrell (1992) and dubbed Minimal Commitment. In these models parsers draw only those conclusions from the input that they can be sure of and postpone uncertain decisions. This architecture predicts few garden paths but greatly increased processing load during long ambiguities. The parallel parser computes alternative analyses simultaneously, dropping variants as disambiguation is met. A pure form of parallel parsing would predict not only no garden paths at all, but also enormous processing load in highly frequent constructions such as: (52) “The girls landed …”

which could have any of the following structures: (53) The girls (who have been) landed (with extra work)… The girl is landed (with extra work)… The girls landed (on runway two) … The girl has landed (on runway two) … The girls’ landed (uncle gave them estates) … The girl’s landed (uncle gave her estates) … The girl has landed (uncles in every county) … The girl is landed (she has extensive estates) …

However, Kennedy, Murray, Jennings & Reid (1989) found no sign of processing complexity in sentences such as (54) even with long ambiguous strings. (54) The workers considered the last oﬀer from the management was an insult

Parsers and gap processing

In addition, the processing of head-ﬁnal languages such as Japanese would be almost impossible because of the multiple ambiguities which can easily occur, especially with premodifying relative clauses (RCs) (see papers in Mazuka & Nagai 1995). To meet these challenges, supporters of parallel parsing tend to assume that possible structures computed in parallel are additionally ranked either by their frequency and plausibility (Kurtzman 1985) or else structural simplicity (Gorrell 1989). Gorrell (1987 [reported in 1995], 1989, 1995) argues that lower ranked structures are more diﬃcult to access and may eventually be dropped if they fall beneath a threshold. In constraint-based models (e.g. Trueswell & Tanenhaus 1994; MacDonald, Pearlmutter & Seidenberg 1994) there is constant ambiguity, and constant ambiguity resolution from the full range of information sources. Each new input word has a ranked list of its possible grammatical features and semantic collocations, the choice among which is made on the basis of the frequency and preference information associated with the lexical entry. This version of parallelism is consistent with the data but diﬃcult to distinguish in terms of empirical predictions from a serial model, as the processing complexity argued to reﬂect the promotion of a low ranked structure is very similar to the reanalysis cost predicted by a serial model. 3.2.2 Modular vs. non-modular There are questions about modularity at two levels. At the higher level: Is the language faculty a cognitive module?21 At the lower level: Is the sentence processor made up of self-contained sub-routines? Jerry Fodor (1983) refers to these as External and Internal Modularity. There are two major reasons why this question has occupied a pre-eminent position in the research dialogue about parsing mechanisms. The ﬁrst is that modularity is particularly important to linguists: a major reason for the existence of Linguistics as a separate discipline is that mental operations on language are argued to operate diﬀerently to other functions of the human mind. The modularity hypothesis is an attempt to capture and deﬁne this dissociation. The second reason for the attention paid to modularity is that it has

21.Jerry Fodor (1983) summarizes cognitive modules as “domain-speciﬁc computational system characterized by informational encapsulation, high speed, restricted access, neural speciﬁcity, and the rest” (p. 100). “The rest” are mandatory operation (p. 52), ﬁxed neural architecture (p. 98), speciﬁc breakdown patterns (p. 99), and characteristic pace and sequencing (p. 100). See e.g. Marslen-Wilson & Tyler (1987) for counter-arguments.

53

54

Empty Categories in Sentence Processing

implications for grammatical models. Chomskian grammars (Chomsky 1981, 1986a, 1995) distinguish between modules of the grammar (Case theory, binding theory, theta-theory) and even between conditions on transformations and conditions on representations (Chomsky 1995: 27). A view of the parsing process in which there are distinct processes or groups of processes which can be identiﬁed as applied asynchronously is readily accounted for in such a grammar. However, most other grammar models do not make such distinctions. For example the rule system in HPSG is a rather homogenous set of constraints on structure, which does not easily lend itself to an instantiation in a modular parser, nor provide any ready way of accounting for data suggesting that sentence processing proceeds in stages. Below we present some models of parsers exhibiting decreasing degrees of modularity. Forster (1979) provides the classic example of an autonomous sentence processor. It consists of three subsystems: the Lexical Processor, the Syntactic Processor, and the Message Processor; and one data store: the Lexicon. Each subsystem has access to the General Problem Solver, the generalized and ﬂexible cognitive system. The input to each system consists only of the output of the previous system, and so the outputs are invariably ordered, furthermore the functioning of each module is automatic and beyond conscious control (cf. Jerry Fodor’s (1983) “mandatory”). Message processor

Syntactic processor

General cognitive system

Lexical processor

Figure 3.Forster’s (1979) modular autonomous parsing model

Forster argues that his model is a strong hypothesis which should be tested to destruction before any other is adopted which would oﬀer fewer insights into

Parsers and gap processing

the language faculty. However it is widely accepted that this model is too strong, and it is rarely argued for even by those supporting a modular position; nevertheless, it provides a useful marker of the extreme end of the modularity continuum. Tanenhaus, Carlson & Seidenberg (1985) sketch a sentence processor which can be seen as a modiﬁed version of Forster’s in that it posits similar modules deﬁned at the traditional linguistic levels, but the general cognitive system has access to the input as well as the output representations of each module, which allows it to ﬁlter at each stage. This is necessary because the automatic operation of each component produces multiple parallel outputs, many of which need to be pruned. This pruning accounts for garden paths and is justiﬁed by the limitations of size of the memory buﬀer of the general cognitive system. semantic component syntactic component lexical component

general cognitive system

phonological component

Figure 4.Tanenhaus et al.’s (1985) weakly autonomous parsing model

As models posit more but smaller subsystems, the distinction between modular and non-modular becomes diﬃcult to maintain. Crain & Steedman (1985) propose a “weakly interactive” parser. This is a hybrid model designed to account for the conﬂicting data suggesting that plausibility and context contribute to the outcomes of syntactic operations, but still to respect the existence of syntax as a formal system. The syntactic processor proposes multiple parallel analyses at the word level and the semantic processor working virtually simultaneously disposes of inappropriate ones. The syntactic and semantic strands can be seen as sorts of evidence used in interpretation, rather than separate operations. Nevertheless, the syntax has

55

56

Empty Categories in Sentence Processing

OUPUT Syntax proposes

Semantics disposes

INPUT Figure 5.Crain & Steedman’s (1985) weakly interactive parser

independent existence, because it is the syntactic function which has prior access to the input signal and provides possible analyses for the semantic function to choose from. This diﬀerentiates it from a strong interactive version in which the semantics would have prior access and dictate to the syntax which analyses were plausible and thus could be computed (e.g. Marslen-Wilson 1975; Tyler & Marslen-Wilson 1977; Marslen-Wilson & Tyler 1980, 1987). At the extreme interactive end of the continuum we have models whose structure derives from the implementation. The architecture of Parallel Distributed Processing involves all sources of information being available at once. This network architecture has more frequently been used to model lexical storage and access (e.g. Rumelhart & McClelland 1986; see e.g. Pinker & Prince 1988 for a critique), but there are also models of parsing (e.g. Elman & McClelland 1984). Stevenson (1994) presents a model which attempts to implement the principles of PPT as probabilistic constraints on structure. The activation of nodes is determined by atomic features on the lexical entry as well as frequency, lexical preference and recency. It is plain that the sentence processor is merely an epiphenomenon, just one function of the mind on this account. While there are still assaults upon the conception of language as a cognitive module from some psychological perspectives (e.g. Connectionism, see also papers in Garﬁeld 1987), the evidence in favour of external modularity for the computational system (though not the lexicon) does seem fairly robust. Whether the sentence processor has an internal modular architecture is far from being decided. One can distinguish two common approaches to the question in recent work. The ﬁrst holds that a parsing model which is a direct implementation of a modular theory of grammar, and is descriptively adequate, can provide evidence for modularity. This is at the heart of the recent spate of principle-based parsers (Pritchett 1988, 1992a; Gorrell 1989, 1993, 1995; Crocker 1994, 1996). Our Experiments 1–5 examine the hypothesis that ECs are used in on-line processing. Empirical evidence suggesting that a non-obvious syntactic construct has psychological reality must provide strong support for grammars which

Parsers and gap processing

Syntactic information

Lexical information

World knowledge General cognitive system

Semantic rules

Pragmatic rules

Phonological information Phonetic information

Figure 6.A fully interactive parsing model

predict ECs and the model of linguistic functioning that such theories embody. The second approach is to look for processes within sentence comprehension which demonstrate the qualities that cognitive modules are argued to have. The modular characteristics which are most frequently tested for are informational encapsulation and restricted access (Jerry Fodor 1983). The argument is made that certain sorts of information relevant to the structure or interpretation of the input data are not, or not immediately, made use of in particular processing tasks. Since it is not clear that the parser could systematically chronologically separate two kinds of knowledge which are organisationally integrated in the grammatical base, this sort of data can support modularity (e.g. Frazier & Rayner 1982; Rayner, Carlson & Frazier 1983; Ferreira & Clifton 1986; Frazier 1987b, 1993; Flores D’Arcais 1990; but see e.g. Crain & Steedman 1985; Marslen-Wilson & Tyler 1987; Altmann & Steedman 1988; Osterhout & Swinney 1989; MacDonald, Pearlmutter & Seidenberg 1994 for studies suggesting exactly the opposite). Our Experiment 6 examines the evidence that the binding of certain ECs has diﬀerent parsing characteristics from most other processing. This would support the view that there are separate structurebuilding and structure-checking functions in the human sentence processing system.

57

58

Empty Categories in Sentence Processing

3.3 Processing gaps The major question addressed here is how the parser decides where to assume that there is a “gap” in a sentence, a position at which syntax might posit an EC. Its prominence arises partly from the fact that it is a form of ambiguity which can wrong-foot the parser and thus allows a peephole onto its functioning. But there is another factor, namely that it is a syntactic issue which divides grammatical models and thus allows conclusions to be drawn about the nature of the embedded grammar. The questions commonly posed concern how the parser ﬁnds gaps and how it ﬁlls gaps: – – – – – – – –

Does the parser use any particular strategy in positing gaps? Does it make errors and in which direction? What is the time course of gap location? What linguistic information does it have access to? Does it also have access to non-linguistic information? Is there any asynchrony in the access to these two sources? What guides the parser in choosing an antecedent for a gap? Does it make antecedent assignment errors?

We ﬁrst discuss gap ﬁnding strategies proposed, then look at the information available to the parser, and lastly consider work done on ﬁlling gaps. 3.3.1 Mechanisms for locating gaps Janet Fodor (1978) contrasted two tactics that the parser might adopt in trying to decide where to assume there to be a gap: the Last Resort and First Resort strategies. The ﬁrst would consist of waiting until there was no other possibility but a gap before assuming there to be one; the second would be assuming a gap whenever possible. There is intuitive evidence for and against both of these (Fodor 1978 and references therein). She rejected both strategies in favour of a Lexical Expectation Model, which holds that the parser has access to a lexically marked ranked list of the complement possibilities of every verb, and uses it to decide whether to assume a gap. This was supported by experimental evidence of Clifton, Frazier & Connine (1984), who used a 350ms visual word-by-word presentation and a timed grammaticality judgement task.22 The variable

22.In a timed grammaticality judgement task, sentences are presented to subjects who

Parsers and gap processing

between the sentence types is transitivity, with (a) being strongly preferred transitive, (b) preferred transitive, and (c) preferred intransitive.23 These three sentence types were presented both in transitive and intransitive uses.24 (55) Transitive use Tommy’s girlfriend was impressed with… a. what Tommy built _ at the racetrack b. what Tommy stole _ at the racetrack c. what Tommy drove _ at the racetrack

(1002ms) (1054ms) (1251ms)

(56) Intransitive use Nobody told the visitors… a. what they should pass (_) quietly by _ b. what they should hurry quietly by _ c. what they should tiptoe quietly by _

(1251ms) (1223ms) (1157ms)

Both grammaticality judgements and reaction times (RTs) co-varied with extent to which the use of the verb corresponded to its transitivity preference. The mechanism suggested is that subjects assumed verbs would be used in accordance with their preferences, and were garden-pathed when this was incorrect. This seems to support a lexical expectation model (see also e.g. Holmes et al. 1989; Trueswell, Tanenhaus & Kello 1993; Shapiro, Nagel & Levine 1993). Clifton & Frazier (1989) accept that the Lexical Expectation Model can account for many results, but argue that it cannot deal with head-ﬁnal languages like Dutch, since it would predict no eﬀects before the verb is reached, but in fact eﬀects can be shown (see Frazier 1987b, 1993). Frazier tested Dutch sentences with ﬁnal verbs in RCs like (57):

press a button to indicate whether or not the sentence was grammatical. Both the proportion of correct judgements and the time taken can give information about the complexity of the sentence. This experiment used visual word-by-word presentation at 350ms intervals to exclude subjects’ reading speed as a variable. 23.For my language feeling drove is a very bad example of an intransitive preference verb, but the authors give no other example for this sentence set. 24.Note the symbols we shall use here when the nature of the gap is relevant to the experimental design: _ (_) *_ _?

= real gap = doubtful gap = illegal gap = implausible gap

59

60

Empty Categories in Sentence Processing

(57) Het meisje dat (_) de man (_) kuste the girl that the man kissed “The girl that the man kissed” OR “…that kissed the man”

Since neither word order nor case marking disambiguates such sentences, in theory both subject and object readings are available for the extracted element. However, there is a strong preference for the interpretation where the absent element is the subject. Clifton & Frazier show how the Active Filler Strategy can account for all the data.25 Essentially it says that when the parser is holding an unassigned ﬁller of an unbounded dependency it prefers to posit a gap than an overt constituent. This model neatly combines the advantages the First Resort Strategy and the Lexical Expectation Model, by making the behaviour it predicts restricted only to occasions when a ﬁller is in fact available for assignation and also permitting, though not relying on, argument structure eﬀects. This model presupposes a modular parser architecture, since locates all of the decision making process in the syntactic component. They further argue that eﬀects such as those in Clifton et al. (1984) (see above) only demonstrate that lexical preference and pragmatic information is used before the end of the sentence, but not on-line. This interpretation of the data is quite plausible, because earlier studies had used relatively oﬀ-line tasks such as timed end-of-sentence grammaticality judgements, and even a serial parser need not exclude some degree of feedback to the phrase structure module over a distance of several words. There is strong evidence to support the view that the parser posits gaps very quickly and sometimes erroneously, in line with the predictions of the Active Filler Strategy. Tanenhaus, Stowe & Carlson (1985, Experiment 2) used selfpaced reading to test sentences with transitive and intransitive preference verbs (ask and race), and both “doubtful” (i.e. “possible but unconﬁrmed” Fodor 1978) and real gap positions. The ﬁllers were plausible at their real late gap positions, but either plausible or implausible at the doubtful gaps; in this example, one can ask a witness but not a church, but one can ask about anything. We can detect whether subjects assume a gap at the early position, for if they do, we shall see the eﬀect of the implausibility of (58b) relative to (58a). This approach is known as the embedded anomaly technique.

25.Active Filler Hypothesis: When a ﬁller of category XP has been identiﬁed in a nonargument position, such as COMP, rank the option of assigning its corresponding gap in the sentence over the option of identifying a lexical phrase of category XP (Clifton & Frazier 1989).

Parsers and gap processing

(58) a.

Transitive preference — plausible The DA found out which witness the reporter asked (_) anxiously about _ b. Transitive preference — implausible The DA found out which church the reporter asked (_?) anxiously about _ c. Intransitive preference — plausible The sheriﬀ wasn’t sure which horse the cowboy raced desperately past _ d. Intransitive preference — implausible The sheriﬀ wasn’t sure which rock the cowboy raced desperately past _

The results show slower reading times at the late gap position in transitive preference conditions (58a,b) than in the intransitive preference conditions (58c,d). This would suggest that subjects needed to reanalyse the gap dependency as they had previously located its foot at the early gap position after the transitive preference verbs. There was also a slowing at the early gap position in the implausible condition (58b) relative to the plausible condition (58a). This may be seen as the implausibility of the dependency linking church and ask slowing its processing. These results demonstrate that the plausibility of the antecedent did not prevent the subjects from positing a gap after the verb, but the preferred argument structure did. This evidence is consistent with a syntaxonly view of gap location. The “ﬁlled gap eﬀect” provides particularly strong evidence for early commitment to a gap. It is a subconscious but experimentally measurable garden path caused by the parser positing a gap but rapidly having to reanalyse. The was ﬁrst noted in an experiment by Crain & Fodor (1985a), who tested sentence pairs in which the doubtful gap position was overtly ﬁlled by an argument, but still found a slowing in reading speed consistent with the forced reanalysis of the relative clause dependency (59).26 (59) a. Who could the little girl have forced (_) us to sing those songs for _ ? b. The little girl could have forced us to sing those songs for Cheryl.

There is no local diﬀerence at forced between sentences (59a) and (59b); but the parser has an unassigned ﬁller in (59a) but not in (59b). There is a slowing of

26.We simplify the materials slightly.

61

62

Empty Categories in Sentence Processing

reading speed at us in (59a), which demonstrates that the parser had already posited a gap after forced and must retract it. As Fodor (1989) notes, this testiﬁes to extreme rapidity in the parser since the disambiguator is the very next word. This high speed has important implications for the class of parsers that are credible. It does seem thinkable that there is a grammatical module which produces an initial structure hypothesis so quickly that it can be gardenpathed during the course of the verb, even when this is only one syllable long as here. This initial hypothesis would be only milliseconds later checked or further processed by other modules working with other sorts of information, such as world knowledge. Behaviour like the ﬁlled gap eﬀect is predicted by a modular parser but only with diﬃculty incorporated into an interactive account. More recently adjustments have been suggested to the Active Filler Strategy in order to deal with other data. De Vincenzi (1991, 1996) has looked at Italian and shown that null subject languages require the reformulation of the Active Filler Strategy as her own Minimal Chain Principle.27 This essentially says that the parser will posit the shortest and fewest chains consistent with the input. A preference for short chains would produce Active Filler Strategy eﬀects. She uses self-paced reading with following comprehension questions to show that readers prefer associating wh-ﬁllers with subjects rather than objects in sentences where the gap location is ambiguous. For example, in (60) Giovanni can be the object or an extraposed subject, leaving in each case the other position for the gap. (60) Chi (_) ha chiamoto (_) Giovanni? who has called Giovanni “Who has called Giovanni?” OR “Who has Giovanni called?”

Another development from the Active Filler Strategy is the Crocker’s (1994) Active Trace Strategy. This diﬀers from the Active Filler Strategy in that it is unconstrained by the surface order of constituents, because, he argues, surface ordering is a feature of items present in the PF representation, and trace is not part of the PF representation. This means that trace can be posited in a fully predictive manner. He refers to this as making use of the “second dimension” of the structure of an input sentence, i.e. the hierarchical phrase structure; the ﬁrst dimension being merely chronological. Not everyone assumes that gap location is driven by a mechanism speciﬁc to the task. Proponents of sentence processors which account for the gap

27.Minimal Chain Principle: Avoid postulating unnecessary chain members at surface structure, but do not delay required chain members (De Vincenzi 1991).

Parsers and gap processing

location data without a separate stipulated function claim that theirs are more parsimonious models. Gibson, Hickok & Schütze (1994) show how a ranked parallel parser (see also Gibson & Hickok 1993) can use certain grammatical information to locate gaps without special stipulation. They utilise “beam theory”, an explicit economy metric from the ﬁeld of artiﬁcial intelligence. This model assumes a cost in processing load units incurred by the violation of the grammatical constraints of theta assignment and subcategorization by a head. Similar to this but making more direct use of a speciﬁc grammar is Pritchett (1992), a principle-based model. Such a parser is an attempt to show how the application of all but only grammatical constraints simultaneously can account for all processing data including gap location. A very diﬀerent view of the processes involved in identifying gaps is associated with the lexicalist tradition. This can be exempliﬁed by the lexical model of MacDonald, Pearlmutter & Seidenberg (1994), which argues that lexical and syntactic ambiguity are essentially the same thing, since string ambiguity originates in the ambiguity of a lexical item, and shows how a wide variety of extra-syntactic information, but primarily frequency and context, can aﬀect the syntactic analysis. This sentence processing model is not an informationally encapsulated system but rather falls into the group of interactive parsing models whose impulse is the satisfaction of constraints from any and all information sources. On this account, the parser will posit a gap whenever the balance of evidence from the many information sources suggest it. MacDonald et al. account for previous data showing that gap location seems to be just syntactically driven by arguing that verb complementation preferences, frequency and context have not been suﬃciently controlled for in the materials tested. Since even De Vincenzi’s Minimal Chain Principle and Crocker’s Active Trace Strategy are motivated by wider processing principles, it can be seen that there is general trend towards accounting for gap location as part of the wider parsing process and not simply by positing a strategy speciﬁc to this problem (see also Pickering 1994). There are theoretical advantages to dealing with this process as just one function of a generalised processing system and without stipulation. The basic choice is between the accounts of principle-based parsers and those of constraint-based parsers. In spite of much discussion and research these two types of models are still competing. Our experiments below address this debate.

63

64

Empty Categories in Sentence Processing

3.3.2 Information available to the parser for locating gaps As a ﬁrst step towards adjudicating between gap location models we can ask what information is immediately available to inﬂuence the parser’s decision whether to posit a gap. This question has generated a considerable quantity of research and discussion. The motivation for this attention is clear: if it could be shown that certain types of grammatical information were not made use of in the initial structure hypothesis, this would strongly support a modular embedded grammar. We shall consider two sorts of grammatical information here, subcategorization and extraction islands. 3.3.2.1 Subcategorization information It is fairly uncontroversial that verb transitivity inﬂuences gap location, as the eﬀect has been frequently replicated (see discussion of Tanenhaus, Stowe & Carlson (1985, Experiment 2) above). Only one study, Mitchell (1987), has suggested that transitivity can be ignored. It is an attachment study, but the question it addresses, whether the parser has on-line access to transitivity, can be generalised to gap location. Mitchell used self-paced reading of word strings to test whether readers would even consider attaching objects to intransitive verbs. After the child had sneezed the doctor/prescribed a course of injections b. After the child had visited the doctor/prescribed a course of injections

(61) a.

The ﬁrst part of (61a) was slower than the ﬁrst part of (61b), the second part of (61a) was faster than the second part of (61b). Subjects were presumably garden-pathed at doctor in (61a) and at prescribed in (61b). This suggests that people puzzled over the abnormal use of sneeze as a transitive but, having established that sneeze can’t have a transitive reading, were then quick to assign doctor to the next verb. They were happy to take visit as a transitive, but took time to reanalyse on meeting the verb. Fodor (1989) tentatively concludes on the basis of this that the parser has access to subcategorization information, but not so fast as to prevent it making initial hypotheses about structure on the basis of category information alone. But the presentation of the materials divided into sections which usually closely resemble, but at the crucial points do not coincide with, the constituent clauses makes this is a very biasing task, and we think it likely that the garden path at doctor represents subjects rechecking at sneezed that they haven’t misread. The absence of ﬁlled gap eﬀects with

Parsers and gap processing

intransitive verbs in other experiments makes us fairly conﬁdent that subcategorization information is used immediately, in spite of this result. Tanenhaus, Boland, Garnsey & Carlson (1989, Experiment 3) used word by word self-paced reading with a nonsense button to test whether only transitivity or also more complex argument structure is taken into account by the parser when deciding whether to posit a gap.28 (62) Transitive verbs a. Which customer did the secretary call _ on the oﬃce phone? b. Which article did the secretary call _? on the oﬃce phone? Verbs allowing clausal complements c. Which child did your brother remind (_) to watch the show? d. Which movie did your brother remind (_?) to watch the show?

The question runs: given that we expect an implausibility eﬀect at call in (62b) relative to (62a), will we ﬁnd the same at remind in (62d) relative to (62c), or will the parser look further downstream, knowing that remind can take a clausal complement, and that the gap for movie may be in it, as in Which movie did your brother remind … the child to watch _? The results showed the expected embedded anomaly eﬀect for the call type verbs, with slowed RTs and lower acceptability after the verb when the ﬁller is implausible. In contrast, for remind type verbs there were eﬀects only after it was clear that there must have been a gap. One might dub this an “unﬁlled gap eﬀect”. This, the authors claim, shows that relatively detailed information about verb argument structure, not just transitivity, can be accessed instantly and consulted when the parser decides whether to posit a gap. 3.3.2.2 Island constraints The evidence on the on-line application of island constraints is complex and murky. Frazier (1985) looked at sentences containing RCs with real and decoy gaps in a timed whole sentence presentation grammaticality judgement task.29

28.Subjects read the sentence in self-paced reading format by pushing one button, but use another button to indicate that the sentence “no longer makes sense”. Boland et al. (1990) calls this a “sensibility judgement task”; the COD gives “sensibleness” with this meaning. The term “nonsense button” seems brief and clear. 29.In this procedure subjects are presented with the sentence as a whole and the time they take to read and judge it is measured. This can show evidence of garden paths since these should slow reading and reduce grammaticality ratings.

65

66

Empty Categories in Sentence Processing

It is a standard descriptive generalization that wh-extraction is not possible from RCs; they are islands. The control items demonstrate that there is no underlying diﬀerence in grammaticality judgement time between the sentences. The verb win is optionally transitive; the verb excel is not. If the island status of the RC prevents subjects positing a gap after win, then (63a) should receive the same grammaticality judgement scores as (63b), with the same time delay. (63) Islands reaction time a. What did the girl who won (*_) receive _ ? 1245ms b. What did the girl who excelled receive _ ? 1200ms Controls c. The girl won all the time. d. The girl excelled all the time.

1170ms 1171ms

judgement 68% 78% 89% 85%

Interestingly, the results showed tentative evidence that subjects were positing gaps at the early doubtful gap position in (63a). The grammaticality judgements were 10% less and the time taken 45ms more in the optionally transitive condition than in the intransitive condition, suggesting that dependencies were being posited at the early gap position, and reanalysed at the late gap position. This result would imply that not all grammatical knowledge is utilized in locating gaps; the island in (63a) seems not to prevent dependencies being made there. This result is still controversial however; not all work has found the same results. Clifton & Frazier (1989, Exp. 2) attempted to replicate this data with very similar materials but found only weakly signiﬁcant eﬀects for both judgement time and error rate. Stowe (1986) used self-paced reading to test sentences in which there was a potential gap site in a clausal subject island. (64) Non-island condition The teacher asked … a. if the team laughed about Greg’s older brother fumbling the ball. b. what the team laughed about (_) Greg’s older brother fumbling _. Island condition c. if [the silly story about Greg’s brother] was supposed to mean anything. d. what [the silly story about (*_) Greg’s brother] was supposed to mean _.

The results show Greg is read more slowly in (64b) than in (64a), which is a classic ﬁlled gap eﬀect, but there is no similar contrast between (64c) and (64d), which Stowe takes to mean that island constraint information is used in structure building. However, even the positive (64a,b) results in this study were

Parsers and gap processing

only weakly reliable and Stowe herself suggests that they are equivocal (1986: 238). It is possible that other materials would achieve diﬀerent results. There is, for instance, no intuitive garden path in (64d), unlike in Frazier’s materials above.30 Bourdages (1992) tested sentences with embedded anomalies in RC islands in French. He used self-paced reading to observe potential ﬁlled gap eﬀects in sentences which contain verbs such as jouer which permit two avec PPs, but which have a strong preference for only one to be realised. The ﬁller is a wh-PP with avec and he therefore predicts a ﬁlled gap eﬀect when the second avec is met. However, if the parser applies island constraints on-line, there will be no such eﬀect when the verb jouait is in an RC island. (65a) is the potential island constraint violation and (65b) the non-island control.31 (65) a.

Island condition Avec qui le voisin a-t-il dit à la petite ﬁlle qui with who the neighbour has.he said to the little girl who jouait *_ avec son amie que sa mère est partie _ vers was.playing *_ with her friend that her mother is left _ about trois heures? three o’clock “With whom did the neighbour say to the little girl who was playing *_ with her friend that her mother had left _ around three o’clock?”

b. Non-island condition Avec qui le voisin a-t-il dit que la petite ﬁlle jouait _ with who the neighbour has.he said that the little girl was.playing _ avec son amie qui est partie vers trois heures? with her friend who is left about three o’clock “With whom did the neighbour say that the little girl was playing _with her friend who left at three o’clock?”

30.This is perhaps due to lexical factors. About is easy to use intransitively. More strongly transitive prepositions might well give diﬀerent results. Perhaps: (i) The hungry guest asked what time the lunch at (*_) The Red Lion was. (ii) Gus asked who the Christmas card to (*_) Aunt Mabel had shocked. 31.It should be noted that these sentences are standardly regarded as ungrammatical, except on a rather unnatural reading where the ﬁller is associated with the matrix clause, which is possible because the ﬁllers are adjuncts. It is methodologically undesirable to present too many ungrammatical sentences in a processing task as it may encourage subjects to develop strategies.

67

68

Empty Categories in Sentence Processing

c.

End-of-sentence eﬀect control À qui le voisin a-t-il dit (_) à la petite ﬁlle que son to who the neighbour has.he said (_) to the little girl that her amie avait parlé _ avant de partir? friend had spoken _ before of leaving “To whom did the neighbour say to the little girl that her friend had spoken before leaving?”

The reading times showed a ﬁlled gap eﬀect in the control sentence but not in the island sentence. However, Bourdages hypothesizes that gaps are only posited when they would complete a possible sentence. The (65c) sentence type provides evidence of this as the doubtful gap after the matrix verb would not form a complete sentence, and no ﬁlled gap eﬀect is found. Bourdages argues that there is a confound between the eﬀect of island constraints and an end-ofsentence eﬀect in gap location which his experiment has not unravelled. Pickering et al. (1994, Exp. 1) used both eye-tracking and self-paced reading to test sentences with subject relatives, so that the sentence was interpretable at the gap, if island constraints are ignored. Sentence types (66a) and (66b) are legal extraction and non-extraction structures, and permit the identiﬁcation of the ﬁlled gap eﬀect at the early gap. Sentence types (66c) and (66d) are parallel but the extraction in (66c) is in an RC island. If the contrast (66c):(66d) is parallel to (66a):(66b), then this would oﬀer evidence that subjects posited a dependency at the early gap position in spite of the constraint violation. (66) a.

Non-island extraction I realise what the artist painted (_) the large mural with _ today b. Non-island control I realise that the artist painted the large mural with skill c. Island extraction I realise what the artist who painted (*_) the large mural ate _ today d. Island control I realise that the artist who painted the large mural ate cakes

The results showed slowed reading times from both eye-tracking and self-paced reading at the critical painted in the extraction conditions (66a,c), suggesting that their subjects were making dependencies at this position. Since the results from the island condition (66c) resemble those from the legal extraction condition (66a), this would support the view that island constraint are not used in structure building. However, the predictions of ﬁlled gap eﬀects at the NP following the verb were not fulﬁlled in either methodology. While this is

Parsers and gap processing

puzzling, it does not greatly weaken the ﬁnding of dependency formation in the island condition, since the results were very similar in the two extraction structures and from the two methodologies.32 Pickering et al.’s Experiment 2 looked at extractions from complex subjects and Experiment 3 tested coordinate structures, however they found no reliable eﬀects in either. They hypothesise that gaps may be posited in islands, but the processing of the resulting dependency necessary to create a ﬁlled gap eﬀect may be only undertaken at the proposition boundary.33 While there is fairly robust evidence suggesting that under some circumstances certain island constraints may be ignored in on-line processing, there are also contrary results and a number of unanswered questions. We may summarise that this question has no deﬁnitive answer as yet. This is a pity, as the issue has potentially revealing theoretical implications for other areas of work. If it were shown that the parser ignored islands in initial structure building, this would be a very interesting correspondence with the constructions which seem to be invisible to the sentence matching task (see Chapter 6 below and the references there). 3.3.2.3 Pragmatic information Other studies have addressed the question whether pragmatic knowledge is used in gap ﬁlling. A number of experimental results have been argued to support this. For example, Clifton et al. (1984) found that pragmatic information could override verb preferences.

32.Pickering et al. wonder whether this may have been due to their use of the moving window version of self-paced reading, in which subjects can see the length of the sentence before they read it. This seems plausible, as it could be motivated by the end of sentence eﬀect found elsewhere. 33.An interesting footnote by Janet Fodor (1989) notes that some the materials in these island experiments could have legitimate parasitic gaps in the islands of the form: i.

This is the article that nobody who received _ understood _

While this is true, it seems to us unlikely that this could have played a role because of the extreme diﬃculty, even for a linguistically informed reader, of getting such a parasitic-gapin-island reading with an embedded verb which is not obligatorily transitive. We invite the reader to test their own intuitions: This is the prize that nobody who won # appreciated # This is the car that the woman who drove # crashed # Which purse did the thief who stole # leave ﬁnger prints on # ? This kind of violin only the musicians who play # appreciate the value of #

69

70

Empty Categories in Sentence Processing

(67) Tommy’s girlfriend was impressed with… a. what Tommy drove _ at the racetrack b. which car Tommy drove _ at the racetrack

So (67a) caused a small garden path because subjects assumed it would be intransitive, but (67b) didn’t as the ﬁller is a likely complement of the verb. However, such evidence does not show that world knowledge is used immediately, for the RTs were from grammaticality judgements at the end of the sentence. A more on-line task is required. Tanenhaus, Boland, Garnsey & Carlson (1989, Exp. 3) used self-paced reading with nonsense button to look for ﬁlled gap eﬀects, to test whether plausible ﬁllers are assigned to remind type verbs, given that the parser could always posit that the implausible ﬁller would ﬁnd its gap in the complement clause (see discussion of Tanenhaus et al. 89 Experiment 2 above). (68b) is plausible at the early gap position, (68a) is implausible, and (68c) is the control condition with no extraction. (68) a. Which movie did Mark remind (_?) them to watch _ this evening? b. Which child did Mark remind (_) them to watch _ this evening? c. Sam wondered whether Mark reminded them to watch the movie.

Tanenhaus et al. say the results show that there is a ﬁlled gap eﬀect if the ﬁller is plausible (68b), but not if it isn’t (68a), and conclude that plausibility plays a role in gap positing if the gap is not obligatory, but not if it is. Such a conclusion would force us to assume that it is always available, but ignored if the gap is obligatory, which seems theoretically unattractive. However the results are not very clear and even reveal some evidence of experimental error. There is a small diﬀerence between the plausibles and implausibles from to in the grammaticality judgement data (see Figure 7), but Tanenhaus et al. do not suggest that it is signiﬁcant.

Figure 7.Ungrammaticality judgements in % from Tanenhaus et al. (1989) exp. 3

Parsers and gap processing

Figure 8.Reading times per word in milliseconds from Tanenhaus et al. (1989 exp. 3)

The only diﬀerence between the plausibles and the implausibles in the RTs (see Figure 8) is at to, and this looks more like a speeding up for the implausibles than a slowing down for the plausibles, because the implausible to is suddenly faster even than the control sentence with no dependency. Recall that if no gap is posited, the implausibles should correspond to the control condition, but this strong, dramatic even, negative spike in the implausibles has no equivalent in the control condition. Since the implausible condition does not show the same behaviour as the control condition, this experiment cannot being seen as showing the application of plausibility, as Tanenhaus et al. assume.34 There are other reasons to doubt that plausibility plays a role in gap positing. We follow Hickok, Canseco-Gonzalez, Zurif & Grimshaw (1992) in considering that the self-paced reading with nonsense button methodology of the Tanenhaus school is an inappropriate instrument with which to examine

34.In fact this negative spike looks like some unrelated eﬀect or even experimental error of some kind, although this has not been commented on in the literature. In the control condition, the RTs vary only marginally (see graph in Figure 8 above). The two wh-ﬁller sentence types are very close to each other everywhere except at the disambiguating to. Here the plausibles jump about 100ms, which can be reasonably be assumed to be either the eﬀect of reanalysis or a gap being ﬁlled. The implausibles plunge about 100ms but then return to shadowing the plausibles closely by the next word. This contrasts strongly with the control condition, and is clearly not just the absence of reanalysis. As no model of gap ﬁlling predicts anything like this, and it is not reﬂected in the grammaticality judgements, it seems barely conceivable that this represents the behavioural reﬂex of subjects positing a dependency. The only other place in the literature where such an odd spike occurs is Tanenhaus et al. (1993), using the same methodology (see ﬁg. 10, p. 74 and table 5, p. 97).

71

72

Empty Categories in Sentence Processing

semantic rocessing. Crucially, the subjects are told to look out for implausibility in order to use the nonsense button. Eﬀectively therefore the experiment is circular with regard to plausibility: an experiment in which subjects are instructed to look out for plausibility can logically never provide evidence that listeners always look out for plausibility. Other empirical evidence throws doubt on the use of context too. Hickok et al. replicated Tanenhaus et al.’s fourth experiment using the cross-modal lexical priming methodology (CMLP).35 In (69a) the ﬁller is plausible as the object of remind, in (69b) it isn’t. At each position marked # an associate of the ﬁller and a control were presented. (69) a. Which doctor did John #1 remind #2 Mary to see b. Which movie did John #1 remind #2 Mary to see

The results showed the same reactivation at the object position in the implausible case as at the plausible case. So we have clear results from CMLP that the gap ﬁlling in this case is structurally driven. Other studies too contradict the Tanenhaus school claim of immediate use of plausibility. Garnsey, Tanenhaus & Chapman (1989) in an event-related brain potentials (ERP) experiment (see Chapter 5 for details) showed immediate accessing of an implausible wh-antecedent just like a overt implausible object: (70) Testing at a gap position a. The businessman knew which customer the secretary called _ at home b. The businessman knew which article the secretary called _? at home Testing at an object position c. The businessman knew whether the secretary called the customer at home d. The businessman knew whether the secretary called the article at home

35.Cross-modal lexical priming involves the subject listening to an auditory stimulus, but making lexical decisions on visually presented targets. Standardly measurements are taken at two diﬀerent testpoints: the experimental testpoint (often a gap position) and a control testpoint; using two diﬀerent visual targets: a semantic associate of the antecedent and a matched unrelated control. A speeded lexical decision on the related target relative to the unrelated target and the control position is interpreted as evidence that the antecedent is relatively highly activated at the moment the visual target is presented. It is an interaction of Testpoint and Target type therefore which provides a positive result. In this way the methodology can provide evidence about the time course of processing at gap positions. See Swinney et al. (1979) for the original presentation of the methodology and our Chapter 4 for further details.

Parsers and gap processing

There was an N40036 at the embedded verb in sentence (70b) and at the object in (70d), because these are implausible. It was shorter but more pronounced in the overt object sentence type, perhaps because it is more diﬃcult to ﬁnd an alternative plausible analysis. This result is a challenge to the hypothesis that pragmatic information plays any role in initial gap location and is consistent with gap location being structurally driven. More recently Tanenhaus, Boland, Mauner & Carlson (1993) have argued that some sort of interpretation of the content of a sentence is carried out before the syntactic analysis is complete and phrase this in terms of thematic role assignment at the verb. They suggest that this is driven by a grammatical process, theta assignment, but it should be noted that the empirical predictions that this makes are not obviously distinguishable from those of Tanenhaus et al. (1989) arguing for the use of pragmatic information. It seems a pragmatic fact that one does not normally contribute alcoholic drinks to public library, not a grammatical one; in fact the NP public library is as capable of bearing a BENEFICIARY theta role as the NP campus party. They suggest that the processing system has immediate access to the possible arguments that a verb can take and the semantic restrictions upon them, and that this information is used in gap location and ﬁlling (p. 303). They do not discuss whether the results of their experiments conﬁrm this, since this point was not the aim of their experiment, but the predictions and results are clear. The main experiment contrasts sentences with plausible and implausible ﬁller assignments: (71) a.

Which campus party did John contribute some cheap liquor to # Friday night? b. Which public library did John contribute some cheap liquor to # Friday night?

If real-world information were used in gap location, we should expect in the plausible sentence a standard gap ﬁlling slowing of reading times at the to. However, if the ﬁller is implausible and the parser makes immediate use of this information, the predictions change radically. The parser will reject the possibility of a gap after to since the ﬁller is implausible, and so there should be no sign of processing complexity there. However there should be an “unﬁlled gap eﬀect” thereafter at Friday, as it becomes clear that no further gap is forthcom-

36.The N400 is a generally thought to be a response to lexical access, the more diﬃcult the access, the stronger the wave form. Such factors as implausibility, low frequency and unpredictability can trigger a stronger N400. See Section 5.7.1 for further detail.

73

74

Empty Categories in Sentence Processing

ing and the ﬁller is unassigned. The data shows neither of these predicted patterns (see Figures 9 and 10); the reading time for to in the implausibles is about 135ms slower than that for the plausibles. This experiment can give us no reason to assume the use of pragmatic or speciﬁcally thematic information in gap positing.

Figure 9.Grammaticality judgements from Tanenhaus et al. (1993)

Figure 10.Reading times in milliseconds from Tanenhaus et al. (1993)

There are of course other studies on the availability of contextual information to the ﬁrst-pass parse of an input string, of which gap location may be just a subpart. However, more recently these have tended to concentrate on the issue of new item attachment, looking at structures such as the restricted relative clause vs. main verb ambiguity, and PP attachment (e.g. MacDonald et al. 1994; Trueswell & Tanenhaus 1994; Gibson & Pearlmutter 1994). The reason is probably that the modular accounts of these structures are less observat-

Parsers and gap processing

ionally adequate than the generalization that gaps are posited as early as the grammar permits, and the evidence for contextual inﬂuence is stronger. We conclude that the balance of evidence is currently against the use of pragmatic information in gap location. 3.3.2.4 Acoustic information One published study has produced very interesting results showing that acoustic information can provide the parser with evidence about the location of gaps. Nagel, Shapiro & Nawy (1994) used CMLP to show that prosodic contours are used in on-line gap location. They tested for antecedents at doubtful early gap positions in two sorts of sentences: those with early gaps and those with late gaps but early doubtful gaps. Each sentence was read with the appropriate prosody, but the sentences were otherwise indistinguishable at the testposition #. (72) a.

Which doctor did the supervisor call _ # to get help for his young daughter? b. Which doctor did the supervisor call (_) # to get help for _ during the crisis?

The results show signiﬁcant priming of the antecedent at the real early gap in (72a) but none at the unreal early gap in (72b). Since the sentences are superﬁcially identical up to and beyond this point, it follows that the gap position must have been signalled by the prosodic form of the sentence. They follow this with a phonetic analysis of the acoustic diﬀerence between the two subcategorizing verbs and ﬁnd a lengthening and greater drop in pitch in the gap condition. This they characterise as the phonological representation of a gap, but this claim seems too strong. There are other diﬀerences between the two sentence types, such as the identiﬁcation of the subject of the inﬁnitival clause. The gap vs. non-gap contrast is just part of a larger structural contrast at the verb position, and it cannot be asserted with conﬁdence that it is precisely the presence of a gap which is signalled by the prosodic diﬀerences. Nevertheless this is a most interesting experiment as it demonstrates fairly conclusively that there is an element of prediction of gap position, and that neither a structural simplicity strategy such as the Active Filler Strategy nor lexical preference can account for all the data. 3.3.2.5 Summary No model of gap location we considered is wholly supported and so we must content ourselves at this stage with the identiﬁcation of which sorts of data the

75

76

Empty Categories in Sentence Processing

parser can make use of when faced with a doubtful gap. Some sorts of evidence are clearly used in locating gaps. Considerable detail is available to the parser on encountering a verb about its subcategorization preferences, and these are used at such great speed that conclusions can be drawn about the continuation of the sentence before the next word is processed, which can on occasions cause gaps to be posited erroneously. Acoustic evidence also plainly plays a role. There is evidence both for and against island information, and we hesitate to draw any strong conclusion here. Further studies would be welcome, as they might conﬁrm the data converging with that from sentence matching, which would be indicative of operator binding being a separate module (see Chapter 6). The immediate application of semantic and extralinguistic information is hotly disputed. Our own view is that the balance of evidence is consistent with gap location being a function of an informationally encapsulated initial grammar only processor, albeit with rapid but subsequent feedback from other mechanisms where other information types are present. This hypothesis makes testable predictions which are too routinely conﬁrmed for them to be wholly false; in spite of numerous attempts, no reliable falsifying evidence has been shown. And if true, this is a surprising fact about language performance which has rich implications. 3.3.3 Identifying antecedents Once the parser has decided that there is a gap in a particular position it still needs to ﬁnd an antecedent. Frazier, Clifton & Randall (1983) argued for a Most Recent Filler Strategy. It hypothesized that control information was not immediately available to the parser on encountering the embedded inﬁnitival verb’s empty subject position, and that the parser simply assumed that the most recent potential ﬁller would be the ﬁller. In sentences where this strategy caused the gap to be misﬁlled, the Most Recent Filler Strategy predicted a garden path eﬀect and consequent longer comprehension time. They produced evidence for this using a speeded comprehension task: subjects read sentences in a 300ms word-by-word presentation, and then made a sentence-ﬁnal “got it” or “missed it” decision. They composed four sentence types with the variables subject control vs. object control, and ambiguous vs. unambiguous:37

37.Actually there were a number of other variables, but we simplify here. We shorten these sentence types somewhat for convenience. “We” was actually “Everyone”, “child” was “little child”, “NP” was “those stupid French songs” and PP was “last Christmas”.

Parsers and gap processing

(73) a.

Subject control — unambiguous We liked the woman who the child started # to sing NP for # PP b. Object control — unambiguous We liked the woman who the child forced # to sing NP PP c. Subject control — ambiguous We liked the woman who the child begged # to sing NP for # PP d. Object control — ambiguous We liked the woman who the child begged # to sing NP PP

The results showed that the Recent Filler sentences were indeed comprehended faster and more frequently than the Distant Filler sentences. Critically, this eﬀect held true for the sentences containing verbs with unambiguous control properties, which shows that the Most Recent Filler Strategy is not just used in ambiguous cases. Frazier et al. argued that control information is accessed only later in the parsing procedure, and that this is an argument for modular grammars and against GPSG. This set oﬀ a furious methodological and theoretical debate (Crain & Fodor 1985a; Clifton & Frazier 1986; Fodor 1988; Ford & Dalrymple 1988; Nicol & Osterhout 1988; Osterhout & Nicol 1988 [both reported in Nicol & Swinney 1989]). Perhaps the most persuasive empirical counter-evidence is Boland, Tanenhaus & Garnsey (1990) who use self-paced reading and a nonsense button. The basic idea is that one can test whether it is the correct ﬁller or the recent ﬁller which has been interpreted as the empty subject of the inﬁnitive by looking for implausibility eﬀects (since snakes can’t hop, but girls can). The frog conditions provide baselines. (74) a. Which snake/frog did the girl force to hop over the rock b. The girl forced the snake/frog to hop over the rock

There are nonsense judgements and slowing down in reading speed from the embedded verb onwards in the implausible snake condition, showing that the empty subject is interpreted as being the correct antecedent. This plainly contradicts the Recent Filler Strategy since the which snake ﬁller is not the most recent ﬁller, and establishes ﬁrmly that control information is quickly available to the parser. Perhaps more noteworthy even than empirical falsiﬁcation is Fodor’s (1988) ﬁnal word against the Recent Filler strategy. She convincingly shows that the Most Recent Filler Strategy cannot in fact distinguish between grammars, and that its proponents have wrongly characterized object control constructions as Distant Filler sentences when in fact the trace of the real controller, the

77

78

Empty Categories in Sentence Processing

matrix object, is closer than the matrix subject.38 More recent evidence for the identiﬁcation of gap antecedents refers to the work which is customarily argued to establish their psychological reality. Since we shall review this work in detail below, we shall simply oﬀer a summary here. It is widely accepted that pronouns and anaphors have ﬁxed ﬁelds of reference which are syntactically determined (Chomsky 1981; see Shillcock 1982; Corbett & Chang 1983; Dell, McKoon & Ratcliﬀ 1983; Cloitre & Bever 1988 for experimental conﬁrmation). Work using CMLP has shown that gaps have the same characteristic (Nicol & Swinney 1989; Swinney, Ford, & Bresnan 1989). There is no evidence from these studies that any “strategies” are in operation. We can therefore be sure that in a RC structure such as (75a), the processor considers only the cherry tree and no other argument as a possible antecedent. Similarly, there is no probabilistic element in the identiﬁcation of the logical object of the passive embedded verb in (75b). (75) a.

Behind the shrubbery stood a cherry tree that uncle felt the hawthorn rather outclassed Gap. b. Uncle felt that the cherry tree standing behind the shrubbery was rather outclassed Gap by the hawthorn.

3.3.4 Summary To sum up, we have seen quite robust evidence that gaps share with overt anaphoric elements the feature that their antecedent identiﬁcation is structurally driven, and makes no use of processing strategies based upon essentially extra-linguistic criteria like economy or saliency, such as the Most Recent Filler Strategy. Subcategorization and control information are both immediately available, but the context has often been shown not to play a role at the earliest stage of processing. Such conclusions would be contested by proponents of fully integrated parsers, however, who would argue that the range of interrelated information available to the parser has not been fully controlled for in many studies (e.g. Trueswell & Tanenhaus 1994). They would seek to show that gap location, like other syntactic ambiguity resolution, is very sensitive to frequency, subcategorisation preference and semantic ﬁt. The issue is very ﬁnely balanced since the two schools of thought react to adverse evidence by tweaking their own models to account for it. We address this question further in Chapter 6 in our sentence matching experiment.

38.Ford & Dalrymple (1988) contemporaneously make the former of these points.

Parsers and gap processing

3.4 Conclusions In this chapter we have outlined some of the work which addressed the functioning and the structure of the sentence processor, particularly with reference to the processing of gaps. We reviewed the range of diﬀerent models of parser which have been advanced in the literature and the parameters of structure and operation on which they vary. We have brieﬂy mentioned some of the explanatory strengths and weaknesses of the options. In general experimental evidence has tended to rule out the extremes in any direction, whilst not providing unambiguous conﬁrmation for any given model. This may be seen as a very positive result in that it would conﬁrm that the parser does indeed possess many of the characteristics that are attributed to it, but while much has been learnt, there are many questions still outstanding. The processing of gap dependencies is our particular focus in this work, and here too we have seen a range of suggestions for the procedures involved. It seems theoretically preferable not to assume a dedicated mechanism here, since the none of the task-speciﬁc models which have been suggested is fully descriptively adequate, and gap location integrated merely as one function of a generalised processing operation is more parsimonious. This still leaves us with a choice between principle-based parsers which implement an existing syntactic model, and which argue for a major role for grammar, and constraint-based parsers which integrate all information sources in the task of on-line comprehension and can dispense with a grammar as conventionally understood by linguists. This is an unﬁnished debate and our current work will contribute to it. Our ﬁrst programme of experiments addresses the psychological reality of empty categories, a test case for principle-based parsers. Measurable processing reﬂexes of these phonologically null constructs would provide very solid evidence for a grammatical approach, since these models predict such eﬀects but the reality of silent constituents would be actually a very surprising fact. Our second experimental study uses the sentence matching task to test for eﬀects which are accounted for only by modular parsers, namely a processing dissociation between local and non-local grammatical phenomena. Integrative models cannot readily assimilate such data and certainly do not predict it. In Chapter 4 below we consider the experimental evidence for the existence of empty categories and report our own studies.

79

AUTHOR ""

TITLE "Antecedent Reactivation and Trace"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 4

Antecedent Reactivation and Trace

4.1 The basic data of antecedent reactivation In Chapter 3 we reviewed work addressing the processing of gaps in sentences. We shall now consider whether these positions play merely the passive role in the computation of a sentence representation that the name gap implies, or whether there is evidence that these positions are only phonologically empty but syntactically active. On the one view, gaps are locations where a subcategorized argument is absent and it is this quality of being missing which deﬁnes them. Alternatively, it has been suggested that the processor actually computes a silent equivalent of the antecedent at the gap location which plays its full part in the processing of the sentence. Since this is in line with the suggestions of Trace Theory (e.g. Fiengo 1977; Chomsky 1977; Wasow 1979), this possibility is often thought of as the assumption that gaps contain traces, constituents with measurable psychological and behavioural reﬂexes. In this section we shall concentrate on trace, particularly wh-trace, reserving the rather special status of NP-trace and PRO for separate discussion.39 The experimental evidence for syntactic categories which have grammatical but not phonetic features has its roots in work on the antecedents of pronouns and other overt anaphoric elements. It has been shown experimentally that the processing of these overt elements produces priming of their antecedents (see below). Since the class of appropriate antecedents is exactly deﬁnable, there can be little doubt that a speciﬁc syntactic process is involved. Essentially the argument for ECs from this data rests on the analysis of ECs as covert anaphoric elements, producing experimental priming of their antecedents in the same way

39.Notice that we do not wish to prejudge the issue of the existence or theoretical validity of these constructs by using their names, but there is no theoretically neutral way of referring to these empty categories and the distinctions between them that PPT makes, and so we shall use the vocabulary of ECs proper to PPT, but with the mental reservation that they may not exist. Similarly, when we mark the location of an empty category in an example sentence, this should be understood as indicating simply where it would be assumed by PPT.

82

Empty Categories in Sentence Processing

that their overt equivalents do. So the evidence for the psychological reality of ECs rests upon the parallelism of their behaviour and that of their overt equivalents. Let us review this data. It has been known since the early eighties that overt anaphors cause the processor to reaccess their antecedents in ways that are experimentally measurable (e.g. Shillcock 1982; Dell, McKoon & Ratcliﬀ 1983; Corbett & Chang 1983).40 Nicol (1988, reported in Nicol & Swinney 1989 and Fodor 1989) used CMLP to show that Anaphors and pronouns access only the syntactically predicted antecedents. (76) The boxer told the skier that the doctor for the team… a. would blame himself # for the injury. b. would blame him for # the injury.

She tested for doctor and skier at the positions marked # after the anaphoric elements and found priming of only doctor after himself in (76a) and both boxer and skier after him in (76b). Since the standard generalization is that an Anaphor such as himself requires a clause-mate antecedent, and a pronoun such as him must not have an antecedent within its clause (Principles A and B of Binding Theory; Chomsky 1981; also Pollard & Sag 1994), we can see that exactly the syntactically appropriate antecedents are reaccessed. Nicol also showed that only antecedents with the correct number and gender are activated (77). (77) a.

The boxer told the actress that the doctor for the team would blame her # for the injury. b. The boxers told the skier that the doctor for the team would blame them # for the injury.

Thus in sentence (77a) her primes only actress, not boxer, and in (77b) them primes only boxers not skier. There is also evidence that pragmatics does not aﬀect pronoun coreference. Corbett & Chang (1983) used a post-sentence probe task to examine whether context would limit the antecedents accessed by a pronoun.

40.The word “anaphor” is a technical term within PPT (Chomsky 1981: 330ﬀ), but is also a useful word in its broader sense, something like “constituent coreferential with another preceding element”. In order to prevent the former blocking the latter, we shall use a capital A (i.e. “Anaphor”) when the technical sense is intended.

Antecedent Reactivation and Trace

(78) Ellen aimed a pistol at Harriet, … a. but Ellen did not pull the trigger b. but she did not pull the trigger

ELLEN HARRIET ELLEN HARRIET

fast slow fast fast

When the subject of the second sentence was speciﬁed as Ellen, then the response to the probe ELLEN was faster than to HARRIET. When the subject was a pronoun that could refer to either Ellen or Harriet, the response times were similar to the two probe words. This shows that the pronoun in (78b) reactivated both possible antecedents, even though one of them was a much more natural continuation of the sentence content than the other. Plausibility appears to play no role here. We may summarise the results of these experiments with the statement that the antecedent identiﬁcation for overt anaphoric elements seems to have the characteristics of a structurally-driven process, and that all but only grammatical antecedents are reactivated. Towards the end of the eighties, researchers started to ask whether this would apply to covert anaphors too. Two classic studies showed that it did. Swinney, Ford, & Bresnan (1989) looked at the antecedents reactivated at gaps in relative clauses. They used the CMLP methodology with both lexical decision and naming tasks, and tested at three points marked #1–3 (79). (79) The policeman saw the boyi that the crowd at the party #1 accused ti #2 of the #3 crime

The core result is that boy is responded to faster than its control at testpoints #2 and #3, both of which followed the gap position. There was no eﬀect at the control testpoint #1; this excludes continued activation as an explanation of the eﬀects found at #2 and #3. Neither crowd nor policeman, which were also tested, showed the same eﬀect. This is ﬁrm evidence of syntactically appropriate antecedents only being reactivated, parallel to the ﬁndings on overt anaphors. Following this, Swinney, Ford & Bresnan (1989) tested at earlier points in RCs to show that there is a speciﬁc syntactic character to any dependency even in complex cases, and that they are processed according to grammatical predictions. They presented targets at three testpoints (#1–3) in a RC which itself contained a complement clause, producing a structure with doubtful and real gaps (80). (80) The boxeri visited the doctorj that the swimmerk at the competition had #1 advised (tj) #2 himi #3 to see tj about the injury

83

84

Empty Categories in Sentence Processing

The results showed: #1 no signiﬁcant priming #2 signiﬁcant priming of doctor only #3 signiﬁcant priming of boxer and doctor Testpoint 1 shows that there is no residual priming, and so all that follows is reactivation. At testpoint 2 there is priming of doctor; there is in fact no real gap here, but the result shows reactivation of the correct antecedent of the doubtful gap (see Chapter 3). The third point shows (presumably) residual activation of doctor, but also new activation of boxer, the only possible antecedent for this EC in the sentence. This quick switch of antecedent priming makes for very convincing evidence that these anaphoric relations too are structurally driven, since only a syntax-based antecedent identiﬁcation mechanism will predict or even account for this. No parsing strategy or probabilistic competitive activation account will do so. This conclusion is conﬁrmed by other work designed to test relevance of plausibility here. Swinney (1991, reported in Hickok et al. 1992) used CMLP to look at strongly biasing contexts such as: (81) The police captain said that the copi from his precinct that the soup in the bowl had eaten ti # was going to give a talk on public policy

Swinney obtained reactivation of cop after eaten in spite of the obvious impossibility of soup eating a cop. Again the eﬀects are structurally determined in deﬁance of plausibility. We can summarise that gaps seem to behave as real syntactic constituents in that they are subject to the same syntactic constraints on their interpretation as overt anaphoric elements. They cause reactivation of all and only grammatically legal antecedents, and their reference is unaﬀected by salience or plausibility. This data set (and the frequent replications of the eﬀect since) constitutes the basic phenomenon that any account of the experimental results must address. The Trace Reactivation Account (= TRA) suggests that this data is evidence that traces, the grammatical construct argued to be the foot of movement chains in PPT (e.g. Chomsky 1995), are present in gap locations and have all the features of overt lexical items except phonetic features. This account further hypothesizes that traces are posited and used in on-line sentence processing. We review this in the section below and competing accounts in the subsequent section.

Antecedent Reactivation and Trace

4.2 The Trace Reactivation Account The TRA is based upon the process of gap ﬁlling, which we shall outline here (following Wanner & Maratsos 1978; Swinney et al. 1989; Nicol & Swinney 1989). When the parser encounters a displaced element, it retains it in some separate short-term store, perhaps a dedicated coreference processor, and retrieves it on encountering a gap later on. The parser reconstitutes the grammatical and semantic features of the antecedent as a silent syntactic element and sets up the dependency between the ﬁller and gap (a “chain” in more recent PPT, e.g. Chomsky 1993). Only then does the foot of this dependency, the trace, enter into a relationship with its subcategorizer in the same way as an overt element would. Thus in (82a–c) below the pencil is the antecedent and is recognized as such by the parser.41 (82) a. The pencili Maria threw ti into the bin b. Aunt Daphne sent the pencili that Maria threw ti into the bin c. Which pencili did Maria throw ti into the bin?

At the gap position it reaccesses the stored ﬁller, constructs it as a trace, and this occupies the same local syntactic role as its overt counterpart in (82a). The question how exactly what happens at the gap site causes the reactivation which our experiments are able to detect is little discussed in the literature, and tends to be assumed rather than argued for. The answer naturally depends on the view taken of the direction of dependency formation. The parsing literature tends to look at the process as the reactivation of a ﬁller, with the information being transferred from left to right; HPSG represents the dependency as a balanced relationship of identity; while work in PPT tends to see the process as leftward movement of features. These contrasting views do not make for a very transparent relationship between processing and syntactic perspectives on such dependency formation. One possibility is that the processing of the dependency involved in the construction of an EC may cause the reactivation we measure. On meeting (or expecting to meet) a gap, the processor would take the decision that this should be the position with which the ﬁller is associated, re-access the ﬁller, and integrate it into the structure of the sentence so far. It is imaginable that this

41.In the relative clause the antecedent is not actually the occurrence of the pencil in the matrix clause, but the precise mechanism is poorly understood. We follow standard practice in processing literature in abstracting from this technical detail.

85

86

Empty Categories in Sentence Processing

extra processing at the gap position is what we detect. It is however also possible that what we can measure is merely the same eﬀect of activation as is caused by an overt constituent. As we have seen, the parsing of any constituent in the input causes measurable activation of that lexical item, and it may be that the antecedent activation at a gap position is just the same, but triggered by a covert rather than an overt category. One of the few aspects of this process open to empirical enquiry is the question which features are and are not reactivated. Since we have seen that only antecedents with the correct number and gender features are reactivated, we can assume that the syntactic features of the antecedent are accessed. This is in line with the content that PPT argues traces must have, namely formal features. Whether a trace also carries the semantic features is a question the literature on PPT neglects, but here we can appeal to experimental data. The data taken as trace reactivation comes standardly from the activation of semantic associates (see discussion of methodologies above), and so we can be sure that these are reactivated too. One point which is generally uncontroversial is that trace has no phonetic features (e.g. Chomsky 1981; Pollard & Sag 1994). A ﬁnding which has excited little comment in the literature but which bears on this issue is Tanenhaus, Carlson & Seidenberg (1985) (see also review of Nagel et al. 1994 in Chapter 3 above). Their experiments one and two used auditory input with visual targets presented for recognition at the sentence end. The unusual factor in this study was that the targets were matched by rhyme. They ﬁrst tested with overt antecedents and found activation when the target rhymed with an (orthographically distinguished) lexical item four words (but not seven words) earlier in the sentence. (83a) is a control condition, (83b) shows the experimental eﬀect. (83) Since Jane forgot to put in… a. all the milk, the cake was dry. b. all the cream, the cake was dry.

THEME slow THEME fast

The next experiment was essentially the same thing but tested for activation by ﬁllers at gap positions. (84) The man was surprised … a. at which winei the judges awarded the ﬁrst prize to ti. b. at which beeri the judges awarded the ﬁrst prize to ti.

FEAR slow FEAR fast

This data seems to demonstrate that phonetic features too are reactivated at the gap position. This result is not obviously compatible with the TRA, since traces

Antecedent Reactivation and Trace

are standardly said to have no phonetic features. While this does not rule out the TRA, it does place some limits on what accounts can be given of what happens at this position. For example, it would seem inconsistent with this data to suggest that the eﬀects found at a gap site simply represent the existence of a constituent which has semantic and syntactic features, but no phonetic features. We may summarise that the precise mechanism of trace reactivation is as yet unclear, but that the data demonstrates that syntactic, semantic and phonetic features are all reaccessed, a ﬁnding which is not completely in line with standard views of the constituency of a trace (but see the copy theory of movement in Chapter 2 above). The TRA has implications for both processing and grammatical theories. Various researchers have made claims about processors which involve the parser positing and making referential use of traces (e.g. Pritchett 1992a,b; Crocker 1994, 1996). If it were shown that presumed trace eﬀects were merely an epiphenomenon, this would tend to focus attention on other processing accounts which make no use of traces, and whose proponents view ECs as an unmotivated construct. On the other hand, if something quite so functionally unmotivated in parsing as traces could be shown to play an integral role in sentence comprehension, principle-based parsers would inevitably gain credibility. This controversy is even more topical within syntactic theory, however, as the treatment of extraction dependencies is a central diﬀerence between grammars. We have seen above the role that traces play in Theta Theory, Binding Theory and Case Theory. Since recent versions of HPSG have nothing syntactic at these positions, but achieve the same eﬀects with features on the subcategorizer (i.e. SYNSEM|NONLOCAL), there is signiﬁcant pressure on PPT to show some evidence in their favour. For either account to be disproved could have far-reaching consequences. Syntactic models tend to have closely interwoven sets of assumptions which interact to produce accounts of the data; it is not a simple matter to alter a construct which has been projected on theory-internal grounds, on the contrary, such tinkering is likely to produce a house of cards eﬀect. This means that if one of the theories were to be shown to be in error on this issue it would throw into serious doubt the fundamental assumptions upon which the framework rests. This makes this research question an interesting one.

87

88

Empty Categories in Sentence Processing

4.3 Traceless accounts of the antecedent reactivation data 4.3.1 Methodological objections McKoon & Ratcliﬀ (1994) cast doubt on the validity of CMLP results suggesting that antecedents are reactivated at trace positions. They argued that the related and unrelated targets which subjects are asked to identify at test positions might be responded to at diﬀerent rates because the related targets were a better semantic, pragmatic and syntactic ﬁt. For example, in the sentence below, PAPER would be identiﬁed more quickly than THINK at the gap position because it is a better ﬁt as an object of the verb. (85) When writing the exam, Jimmy used the new peni that his mother-inlaw had purchased ei # for his birthday. Related target: PAPER Unrelated target: THINK

In addition, they reported a visual presentation experiment of their own, in which they had demonstrated that “good ﬁt” words were responded to more quickly than “bad ﬁt” words. This ﬁnding is not controversial, since this eﬀect has been noted before and even made use of as an experimental technique (e.g. Boland 1993; Corbett & Chang 1983), however CMLP has generally been thought to be relatively free of such eﬀects. The challenge was quickly taken up by Nicol et al. (1994) and Love & Swinney (1996). The latter group argued that a strength of the cross-modal element was precisely that it excluded artifacts of integration, and conducted two CMLP experiments to demonstrate this. The variable of “goodness of ﬁt” was meticulously excluded in the ﬁrst and deliberately built in to the second. Neither showed any signiﬁcant eﬀects of the goodness-of-ﬁt that McKoon & Ratcliﬀ argued was the relevant factor. On the other hand, both experiments showed reactivation of antecedents in line with syntactic constraints. These results supported the consensus view that the CMLP task, which has been the methodology of choice in the trace reactivation debate, is generally free of integration eﬀects. 4.3.2 The Direct Association Hypothesis This is the single most important challenger to the TRA. Recall that the mechanism of antecedent reactivation in the TRA is a dependency between the ﬁller and the gap. The ﬁller is held in store until the gap position and then reaccessed,

Antecedent Reactivation and Trace

from when on the reconstituted element enters grammatical relations in the same way as an overt constituent. There are thus two steps for the parser to take in order for it to identify the ﬁller as the logical object of the verb. In (86), it ﬁrst constructs a dependency between the ﬁller whom and the trace t but only thereafter adjudges it to be the direct object of admire and processes it accordingly (e.g. theta assignment, Case checking).42 (86) Whomi does Cheri admirej tij?

The Direct Association Hypothesis (DAH) of Pickering & Barry (1991) uses only one link to achieve the same eﬀect. This account hypothesizes a direct connection between moved item and its subcategorizer (87). (87) Whomi does Cheri admirei?

The ﬁller is not separately stored and retrieved in a discrete coreference process as in the TRA, but undergoes the same holding operation as any other element which is not adjacent to its subcategorizer. In (88) the two occurrences of Robin are equidistant from their subcategorizing verbs: the TRA supposes diﬀerent mechanisms for its storage between them, the DAH the same. (88) It was Robini that you said Peter thought Gordon irritatedi. Robini (though Peter wouldn’t believe it) irritatedi Gordon.

But the most important novelty of the DAH for our purposes is that the ﬁller is directly associated with its subcategorizer when the parser encounters it. The parser has no need to wait for an overt gap position, nor does it make use of trace. This account reﬂects the way that some lexicalist grammatical models allow for the subcategorization of non-local arguments (e.g. the NONLOCAL features in HPSG). According to Pickering & Barry, the antecedent reactivation data thought to be located at the gap position is in fact simply the association of the dependent with its subcategoriser. Since most experimentation uses direct object gaps and these immediately follow their subcategorizers in English, the erroneous attribution of the eﬀect to the gap is readily comprehensible. There is much corroborating evidence that might be advanced in support of this account. Firstly, it helps to solve the puzzle of why UG should contain real but silent constituents, with the ambiguities and obvious complications

42.Note that we shall use superscripts to indicate direct association (=subcategorization). Note also that in our discussion we shall sometimes use the word verb when subcategorizer would be more exact. We consider the brevity and euphony gained worth the exactness lost.

89

90

Empty Categories in Sentence Processing

that they add to the parsing process. A ﬁller-gap dependency, for example, would be considerably simpler if there were a resumptive pronoun at its foot. A resumptive pronoun is after all equivalent to a trace with its phonetic features restored, so why is it that structures such as (89), which is, while ungrammatical, immediately comprehensible to listeners, is not made more use of in languages with wh-movement? (89) What did Clare say that she liked (*it)?

In a system without traces the foot of the dependency is the overt subcategorizer. Since subcategorizers are overt, a large part of the problem dissolves, and the lack of resumptive pronouns is explained — with no trace, the task of inserting a resumptive pronoun becomes far more syntactically complex and its rarity is explained. The DAH is also compatible with evidence suggesting that subcategorization frames play a major role in the parser’s gap location decision process (see discussion of verb preference in Chapter 3). Consider also the ambiguous question (90a), which has the two possible answers (90b) and (90c). (90) a. Whoi does the boss like (ei) to work (ei) hard? b. She likes the sales force to work hard c. She likes to work the sales force hard

The lack of any overt evidence for the gap position leaves the decision to be made entirely on the basis of the verb preferences. It would appear that the interpretation of like as an ECM verb is preferred to that of work as a transitive. While this sort of data is not incompatible with the use of traces as the foot of dependencies, it is more natural if the foot is at the verb, since it is features of the verb which appear to be bearing the load of dependency disambiguation. Self-paced reading data on displaced objects provides further evidence. When the parser encounters a verb and has an unassigned ﬁller, reading time is slowed (e.g. Crain & Fodor 1985a; Stowe 1986; Tanenhaus et al. 1989). This typically lasts for two words (Verb and Verb+1) and so it appears that the slowed reading times surround the trace position. Slowing at the verb position has been seen as the parser checking the verb’s argument structure to see whether it should posit a trace, and the delay on the next word is the integration of the trace itself. But it may well be that the eﬀect at the verb is the association itself, and the eﬀect at the next word is simply a continuation of this. There is plenty of data suggesting that eﬀects in self-paced reading show some delay; for example Pickering et al. (1994) found eﬀects in self-paced reading to be delayed

Antecedent Reactivation and Trace

relative to eye-tracking data on the same materials. The self-paced reading data is thus fully consistent with the DAH. The ﬁlled gap eﬀect (see Chapter 3 above) is consistent with the DAH too. Whenever the parser encounters a locally possible subcategorizer for a ﬂoating ﬁller, it attaches it (cf. Clifton & Frazier 1989). There is thus a perceptible garden path eﬀect at Donna in (91): (91) Whoi did the detective interrogate (ei) Donna about ei?

If the parser waited for a gap before positing a dependency, it would have to wait until the direct object position after interrogate, the position of Donna. But the ﬁlled gap eﬀect, which is the processing cost of undoing the incorrectly posited dependency, demonstrates that the parser must have made some degree of commitment to the association at the verb position, before Donna is reached. In order to account for this, the TRA must make reference to predictive gap location or consultation with the verb’s preferences to establish whether a gap will be possible. The DAH needs no special additional assumptions. The ﬁlled gap eﬀect is thus more consistent with the DAH than the TRA. The DAH also provides a natural account of an embarrassment for trace accounts, namely, adjuncts (cf. Hukari & Levine 1994). Adjunct ﬁllers will never have overt gaps, but there is no intuitive or, to our knowledge, empirical evidence that adjunct ﬁllers are processed at all diﬀerently. This makes any major contribution of gaps and traces to processing unlikely. As yet no one has shown reactivation of an adjunct trace, and indeed their existence must be regarded as controversial. The DAH allows adjunct ﬁllers to be attached to their head verb just like arguments, but the TRA must again rely on information from the subcategorizer, and posit a gap only on this basis. The DAH can thus oﬀer a natural account of a wide range of phenomena associated with dependencies without the use of traces or even the assumption of gaps. We have seen that, in doing this, it is often rather more explanatory than the TRA. In the following section we shall brieﬂy review Pickering & Barry’s justiﬁcation of the DAH and point out how their dismissal of the need for traces can be seen as speciﬁc to the framework they adopt, Flexible Categorial Grammar. 4.3.2.1 Parsing models and the DAH Pickering & Barry (1991) provide arguments against traces using two sorts of sentences: those with an extraction position remote from its subcategorizing verb, and those with multiple embeddings (see Gibson & Hickok 1993 and Gorrell 1993 for counter-arguments, Pickering 1993 for the reply). We shall

91

92

Empty Categories in Sentence Processing

consider just one of the conﬁgurations which Pickering & Barry suggest provide evidence against traces, in order to give the core of the discussion. Consider (92) and (93) below. Pickering & Barry argue that (92) is awkward and that the assumption of traces wrongly predicts that (93) should be equally so. (92) We gave every student capable of answering every single tricky question on the details of the new and extremely complicated theory about the causes of political instability in small nations with a history of military rulers a prize. (93) [iWhat prize] did you give every student capable of answering every single tricky question on the details of the new and extremely complicated theory about the causes of political instability in small nations with a history of military rulers ei?

Pickering & Barry suggest that the awkwardness of (92) is caused by the need to associate the argument the prize with gave after our faculties have been overloaded by such a long indirect object. The lack of such an eﬀect in (93) suggests that traces are not real syntactic constituents; if they were, they would trigger the same awkwardness. Gibson & Hickok (1993) and Gorrell (1993) both show that this distance between ﬁller and foot is an inadequate predictor of complexity. In fact it seems rather more likely that it is the ﬂagrant violation of the ordering constraint that heavier constituents should follow lighter ones which causes the eﬀects (Behaghel 1909). In fact the whole discussion can be seen as a debate between radically different parsers: Pickering & Barry’s bottom-up incremental parser and the predictive models of the others. This becomes clearer when we note that Pickering & Barry’s evidence argues as strongly against standard constituent structures as against traces. This is entirely in line with their assumptions, since they are arguing from a position of a ﬂexible categorial grammar which embodies a very particular view of constituent boundaries. For an example of why this is argued for, consider the fairly standard constituent boundaries in (94) below. (94) a.

I saw [the [cat [that [bit [the [rat [that [nibbled [the [cheese]]]]]]]]]]]] b. What did you [give [every [student [capable [of [answering […]]]]]] Gap]?

Pickering (1993) suggests that (94a) is evidence that the parser is not bound by traditional right-branching constituency boundaries; if it were, he argues, it could not parse the complement of the verb saw as an NP until it had reached

Antecedent Reactivation and Trace

cheese, because the NP would not be complete. Since the interpretation of the cat … as the object plainly occurs earlier, the bottom-up parser cannot be bound by these constituent boundaries. But notice now that (94b) essentially reveals the same problem: Pickering & Barry assume that the gap cannot be posited until the preceding constituent has been entirely processed. Again, this seems unlikely. But this argument applies to the gap only because we locate the gap at the closure of the VP, and (94a) shows that it is this VP closure that Pickering & Barry’s bottom-up parser is unable to predict, rather than the gap itself. Given this state of aﬀairs, there are two ways to go; one route is to maintain a bottom-up parser but eliminate traces and traditional right-branching constituency boundaries. This is Pickering & Barry’s preferred route. They posit no role for traces and adopt an dependency categorial grammar which assigns leftbranching constituent status to input strings using NP and S as both additive and subtractive terms, based on Montague grammar (see e.g. Ades & Steedman 1982; Steedman 1987).43 The elimination of traces and right-branching structures are not readily dissociated, as the evidence is essentially the same for both. The other route is that taken by Gibson & Hickok (1993) and Gorrell (1993) in their replies to Pickering & Barry. This is to permit some degree of top-down predictive processing. There is a problem with this, however. This requires the parser to structure-build a trace before its position in the linear precedence of sentence constituents, which could be held to oﬀend against the “no crossing branches” limitation which is standardly assumed for syntactic structures. Both replies to Pickering & Barry address this diﬃculty. Gibson & Hickok describe a parser which allows traces to be posited as soon as they are licensed by a subcategorizer, reserving for the trace position the role of structure checking. They circumvent the no crossing branches problem by suggesting that it only applies to items which contain lexical material. The authors admit that their parser is empirically indistinguishable from the DAH on the available data.

43.Thus a string such as John saw is S/NP, because it is a constituent which would be a sentence (S) if it were followed by a NP. In the same way, a transitive verb is (NP\S)/NP, because it is a constituent which would be an S if it were preceded by an NP and this new constituent were followed by an NP. Notice that this incremental analysis necessarily tends to favour a left-branching structure for almost any string. For example, the string Jack saw Jill is assigned structure in the stages NP, S/NP, S: (i) [[[Jack NP] [saw (NP\S)/NP] S/NP] [Jill NP] S] It also has the interesting quality of making it seem more natural to assign structure subscripts to the right and not the left bracket.

93

94

Empty Categories in Sentence Processing

Gorrell (1993) also outlines a parser with predictive capabilities. He employs the technique of delayed commitment. The parser posits a trace dominated by its subcategorizer but otherwise in an unspeciﬁed position relative to other constituents. Only when more certain information is available does the parser assert greater detail and carry out further grammatical processing. This will produce an incremental interpretive eﬀect in line with Pickering & Barry’s data and at the same time evades the problem of structure-building elements out of sequence. An alternative solution to this problem is Crocker’s (1994, 1996) parser with the Active Trace Strategy (see also Section 3.3.1 above). This permits gap location without constraint by the surface order of constituents, because, he argues, this applies only to items present in the PF representation, which excludes trace. Trace can thus be posited in an up-and-over manner using Grimshaw’s concept of extended projections. He thus cuts out the need for the subcategorizer, justifying this with data such as that of Frazier (1987b), who tested verb-ﬁnal Dutch sentences, and showed that there are parsing preferences before the end of the sentence. The data shows that the parser prefers to place the gap early in the case of ambiguity, even before the verb has been reached, which supports Crocker’s model. To summarize: even though their own evidence did not sustain the strong claims they based on it, Pickering & Barry started a debate which clariﬁed that it was untenable to assume a bottom-up parser using traces, but that a parser with some predictive capacity may use traces. Work in combinatory models argues that traces are unnecessary to the processor, and thus a more economical parsing model can be adopted without them. Principle-based parsers show that descriptively adequate models can make use of traces; this success necessarily supports the psychological reality of the grammar models such parsers implement. Since the question has not been answered simply by examining the adequacy of parsing models with and without traces, it will be instructive to consider the experimental evidence. 4.3.2.2 Experimental evidence for the DAH The ambiguity of this situation has given rise to experimental attempts to resolve it. Let us ﬁrst note the experimental predictions of the DAH: if traces are not used in processing, there will be no activation of antecedents at a gap position, but there will be at the subcategorizer, as it is here that the association is made. That this distinction was not noted previously is because most experiments wishing to examine trace reactivation examine direct object traces, as their position is readily identiﬁable. But at this position the diﬀerent predictions

Antecedent Reactivation and Trace

of the DAH and the TRA are indistinguishable, as direct objects are canonically positioned adjacent to the verb. Nicol (1993) is an attempt to dissolve this confound. She used CMLP to test for reactivation of the PP wh-item at two testpoints: one adjacent to the verb #1 and one at the presumed trace position non-adjacent to the verb #2 (95). (95) To which butcheri did the woman who had just inherited a large sum of money give #1 the very expensive gift ei #2 the other day?

She found signiﬁcant reactivation of the ﬁller at both testpoints. This was an unexpected result, and in the absence of a control testpoint rather left the issue unresolved. Nicol puts the case that this supports a version of the DAH, but this is rather surprising conclusion, as it seems easier for the TRA to accommodate eﬀects at the verb than it is for the DAH to accommodate eﬀects at the gap. It is not inconsistent with the TRA that some predictive gap location should be based on information which becomes available at the verb. Since the verb occurs during the dependency, a parser with any degree of incrementality can readily accommodate some processing there. For the DAH, the dependency ends at the verb site, which makes it less easy to account for eﬀects at the gap position beyond this. Nevertheless, one might argue that there is some structure checking at this position, simply to conﬁrm that there is no overt argument present which would necessitate reanalysis. Nicol accepts that there are confounding factors. It is for example possible that the activation at the verb was simply an erroneous ﬁrst resort eﬀect, and that the verb, which subcategorizes for a PP, immediately licensed a gap position adjacent to it for the ﬁller. Examples such as (96) demonstrate this to be a possible structure. (96) To which charitiesi did the woman who had inherited a large sum of money give ei?

The second alternative explanation is that the eﬀect at the trace position could be due to clausal processing at the point where the verb’s argument structure has been saturated. Nicol does not mention this, but this explanation could also be applied to the reactivation at the verb. See Section 4.3.4 below on depth of processing eﬀects for discussion of this. In an attempt to control for some of these uncertainties, Nicol next carried out an experiment to test directly for activation of the subject at the verb. The idea behind this was that this would establish whether the preceding arguments of a verb are reactivated when it is encountered. If it happened with a subject, it could be assumed to take place with a ﬁller too.

95

96

Empty Categories in Sentence Processing

(97) [The actress [RCwho had caused such a sensation among the critics]] is # a failure with the general public.

Signiﬁcant reactivation was indeed found immediately after the copula, showing evidence that all arguments are activated at the verb. There are however here too possible alternative analyses. Firstly, it is quite probable that the RC, which was included to create distance between the subject and the verb, itself causes priming of its head. It would be no great surprise to ﬁnd additional activation of the head noun phrase during the RC and perhaps particularly at its oﬀset, when it is integrated into the head. While the distance between the minimal NP subject the actress and the verb is large, the distance between the maximal NP subject the actress who had caused such a sensation among the critics and the verb is nil. Since it would be very natural to posit some summary processing of the complex NP at its oﬀset, it seems not unlikely that this played a role. Additionally, there is both syntactic evidence (e.g. Stowell 1981; Pollard & Sag 1994) and experimental evidence (Samar & Berent 1991) to suggest that be is a raising verb. This would imply that the subject is base generated at Nicol’s test position, which raises weighty questions about the cause of the eﬀects that Nicol reports. To summarise, Nicol’s experiments seem to rule out the strong versions of both the DAH and the TRA, which would each deny any eﬀect at the opponent’s preferred processing location, while not obviously favouring the weaker form of either. Another study performed at the same time oﬀers quite detailed evidence on the subject from a diﬀerent methodology. Tanenhaus et al. (1993) were trying to show evidence that some interpretation of the content of a sentence is carried out before the syntactic analysis is complete. Their claim was that the implausibility of a ﬁller is detected before the trace position, but our interest in their data is as evidence of the time course of ﬁller processing, in order to compare the DAH and the TRA. They used accumulative self-paced reading with a nonsense button to test sentences such as (98):44 (98) a.

Which campus partyi did John contribute some cheap liquor to ti Friday night? b. Which public libraryi did John contribute some cheap liquor to ti Friday night?

44.In accumulative self-paced reading that portion of the sentence which has already been read does not disappear, but is added to with each button press.

Antecedent Reactivation and Trace

If subjects process ﬁllers at the gap site in line with the TRA, they will respond to the implausibility in (98b) there. If they process the ﬁller at its subcategoriser, as predicted by the DAH, they will respond at the site of the to. We present the results in the table below for convenience. The ﬁgures in the ﬁrst line represent the increased nonsense judgements in the implausible condition over the plausible condition. Those in the second line represent the increased reading time required in the implausible condition over the plausible condition.45 Table 5.Results of Tanenhaus et al. (1993) Increments Nonsense % Reading time ms

contribute

some

cheap

liquor

to

Friday

night

0 5

3 5

4 75

−18 −50

14 130

5 145

1 50

The nonsense judgements show some evidence that subjects did not wait for subcategoriser or gap. The eﬀect at Friday is evidence for the TRA, while the much larger eﬀect at to is consistent with the DAH. The largest nonsense increment is at liquor, preceding both the subcategoriser and the gap, and is thus predicted by neither account. The reading time data shows just one thing unambiguously: evidence of gap-ﬁlling. The slowing of reading time to be expected when a ﬁller is integrated is clearly located at the subcategorizer to and gap site.46 This is in line with both the DAH and the TRA. In fact it is possible to argue within both accounts that they can encompass this data. Bottom-up trace processing predicts no eﬀect before the gap, but any even

45.Note that these are as accurate as can be obtained from the published form of this paper, which oﬀers the data in graph form and makes no attempt at statistical analysis. We discuss the data also in Chapter 3 above. 46.We do not see any straightforward explanation of the odd reading time variations at cheap, liquor, and to. We should expect the maximum implausibility eﬀect precisely at liquor, but in fact we see a sudden reversal; if we believe this data then subjects are judging it more plausible that liquor be contributed to a public library than to a campus party. The oddity is a 175ms spike on liquor in the plausible condition, which is completely unreﬂected in the implausible condition. It seems likely that this unmotivated spike is some sort of experimental artifact or error, because it is out of line with the rest of this data, is not found in other self-paced reading data, and consists of only one data point, but this is admittedly speculation (but see also footnote 34, p.71). If it were true, then the reading time data would parallel the nonsense judgements in showing evidence that the implausibility was detected before the gap or the subcategoriser.

97

98

Empty Categories in Sentence Processing

moderately predictive parser (e.g. Gibson & Hickok 1993; Gorrell 1993; Crocker 1994, 1996) would allow us to attribute the eﬀects at to and liquor to gap processing, since the gap position is not only unambiguous but even obligatory by this point. Since the sentence has only one grammatical outcome as soon as the some is encountered, it is possible to attribute both slowing and implausibility judgements to syntax-based predictive processing quite compatible with the use of traces. It is also possible to attribute the eﬀect at liquor to direct association, however. Pollard & Sag (1994: 255) suggest that contentless prepositions are transparent to the CONTENT features of their complement nouns, which can therefore appear on the verb’s SUBCAT list.47 If the NP complement of such a PP were then extracted, the verb and not the preposition would be the subcategorizer of the ﬁller (99). (99) Which partyi did John contributei? some cheap liquor toi? last week?

This would alter the predictions of the DAH in Tanenhaus et al.’s experimental sentences since the predictions of the DAH are linked to the position of the subcategorizer. The DAH would then predict the detection of implausibility before the gap site, as indeed happens, so this alternative account would permit the data to be argued to support the DAH on the basis of the nonsense judgements. On the other hand it would mean that the reading times unambiguously supported the TRA since these show eﬀects only around the PP and trace position. We may summarize that Tanenhaus et al.’s data is most obviously supportive of the DAH but can also readily be accounted for within a parser using traces and some small degree of grammar-based predictive interpretation. A further type of experimental study which can throw light on the conﬂicting claims of the DAH and the TRA is work concerning languages with headﬁnal VPs. Frazier has done work on ﬁller attachments in Dutch (Frazier 1987b; Frazier & Flores D’Arcais 1989; Frazier 1993; see also Pritchett 1991 on Japanese). The relevant sentences are verb ﬁnal and have two arguments which are not disambiguated by case marking. One of these is an extracted argument, and

47.Both of these relations are deﬁned by positions on SUBCAT lists, and in these cases the objects of the PPs need to appear not only on the SUBCAT list of the PP but also on that of the verb. (i) Gerry appealed [to [jDavid]] [ej to back the agreement] (ii) *Gerryj depends [on [jhim]]

Antecedent Reactivation and Trace

so we have a string locally ambiguous between NP > gap and gap > NP. The verb morphology disambiguates by identifying the number of the subject. (100) a.

Object gap Karl hielp de mijnwerkersi diei de boswachter ei vond SLOW Karl helped the mineworkers who the woodman found-sg “Karl helped the mineworkers who the woodman found” b. Subject gap Karl hielp de mijnwerkersi diei ei de boswachter vonden FAST Karl helped the mineworkers who the woodman found-pl “Karl helped the mineworkers who found the woodman”

The results show a reanalysis eﬀect at the verb in the dispreferred condition, and reveal a preference for the subject gap reading, as in (100b). This means that the attachment must take place before the subcategorizing verb. The exact nature of the preference is irrelevant here, it is only important to us that there is one. Plainly some commitment to the attachment is made without reference to the subcategorizer, against the requirements of the DAH (and indeed Gibson & Hickok 1993). This data cannot exclude direct association however for the same reasons that the Tanenhaus et al. (1993) study cannot exclude the use of trace; in fact it requires only that direct association be embedded in a parser with some capacity for predictive interpretation. Now this is not compatible with the original Pickering & Barry (1991) version of the DAH, but is accepted by Pickering (1993) and is certainly available to an incremental parsing model such as Pickering (1994). We have seen that previous experimentation, even when it has tried to distinguish between subcategorizer and trace based eﬀects has not yet been wholly successful, although the strong versions of each which permit dependency processing only at their chosen location can be shown not to hold. One of the major aims in our experimentation below is to test the competing claims of the DAH and the TRA and to produce clear evidence to distinguish between them. However, other accounts of the antecedent reactivation data have been put forward, and we review these here. 4.3.3 The Semantic Processing Account Fodor (1989, 1993, 1995) and Fodor & Sag (1994) have repeatedly suggested that psycholinguistic results purporting to reveal evidence of empty categories

99

100 Empty Categories in Sentence Processing

cannot in principle do this, as it is impossible for experimental ﬁndings to count as evidence for any aspect of syntactic theory unless it can be established that the data is a result of the syntactic representation of the input materials. Since our knowledge of the boundary between syntactic and semantic processing is incomplete and our understanding of the nature of the processing that our experimental techniques tap into partial, we cannot exclude the possibility that the eﬀects found are merely a result of the computation of a semantic representation of the sentence. We shall refer to this as the Semantic Processing Account (= SPA). The SPA sounds persuasive but is on reﬂection weaker than at ﬁrst glance. One major problem is that it is not obviously testable: since Fodor makes no reference to any speciﬁc theory or implementation of the SPA, it is diﬃcult to examine its validity. One piece of work for which it has been claimed that it unambiguously represents syntactic and not semantic processing is De Vincenzi (1996, Exps. 2 & 3). She tested biclausal wh-questions with and without complementizers in which the wh-element is ambiguous between a subject and object reading. She shows that the presence or absence of a complementizer aﬀects the preference for a subject reading of the wh-item, and accounts for this with an interaction of the ECP and her Minimal Chain Principle. Since the diﬀerence between the sentences with and without complementizers is purely a syntactic one with no semantic reﬂex, she argues that this meets Fodor & Sag’s criterion of clearly syntactic evidence for wh-trace, because it must be the length of the wh-chains which causes the preference. We shall not attempt to tease apart the validity of this claim in detail, noting only two possible weaknesses. First, one might turn De Vincenzi’s conclusions round and argue that her results conﬁrm precisely that the presence or absence of a complementizer occasions a semantic diﬀerence: the preference for a subject or object reading. If this is the case, then the evidence for wh-chains is no longer necessarily syntactic. Also, we cannot be sure that the diﬀerence at the trace position is syntactic in nature, since it is shown only by self-paced reading complexity data. There is plenty of evidence that additional semantic processing can slow reading times (e.g. Tanenhaus et al. 1989). De Vincenzi’s data may thus not be conclusive. Another weakness of the SPA is that it fails to account for the similarity in time course and amplitude of EC antecedent reactivation and activation by overt anaphors (compare Nicol 1988 and Swinney et al. 1989). Since pronouns and reﬂexives are unambiguously existing syntactic objects, it seems not unreasonable to assume that other constructs which trigger identical eﬀects are

Antecedent Reactivation and Trace

also syntactic objects, whether these experimental eﬀects themselves are the result of semantic or syntactic processing. Uncertainty about the nature of the evidence does not remove the striking parallels between overt anaphors and their suggested covert equivalents. This similarity is all the more persuasive because of the lack of resemblance between the EC reactivation data and eﬀects which are unambiguously semantic. Experiments on semantic processing show a very diﬀerent pattern of reactivation. For example, Dell, McKoon & Ratcliﬀ (1983) examined the priming of antecedents by hypernyms using 250ms incremental word by word presentation, which was interrupted for the presentation of the probe. Experimental texts such as (101) were four sentences long and contained ﬁve testpoints, the ﬁrst of which (#1) was a control testpoint, which was before the hypernym (here criminal). Both the real antecedent of the hypernym (burglar) and a clause-mate control (garage) were presented. The ﬁnal sentence existed in a form with the hypernym and one with an unconnected lexical item (cat). (101) A burglar surveyed the garage set back from the street. Several milk bottles were piled at the curb. The banker and her husband were on vacation. (The #1 criminal)/(A #1 cat) #2 slipped #3 away #4 from the #5 streetlamp. Antecedent: BURGLAR Clause-mate: GARAGE

Results showed that not only the antecedent but also the clause-mate item from the ﬁrst sentence is primed after the hypernym at #2; the antecedent remains so from #2 to #5, whereas the clause-mate priming dies away again quickly. There are two diﬀerences with the results from syntactic priming by overt anaphors and wh-trace shown by Nicol (1988) and Swinney et al. (1989): ﬁrstly here clause-mate lexical items are primed, whereas only grammatical antecedents were reactivated by the anaphors; second, the semantic activation here seems to continue for the whole clause, while the syntactic reactivation by anaphors rapidly degrades (the data from Swinney et al. 1989; Osterhout & Nicol 1988; and Osterhout & Swinney 1992 converges on a ﬁgure of about 500ms). Experimental data on semantic processing thus seems to produce eﬀects markedly diﬀerent from those that are caused by overt anaphors and trace. While this evidence is suggestive it cannot be conclusive, and there is plainly a need for further studies to address this question. However, as Fodor nowhere to our knowledge speciﬁes how a semantic account of reactivation

101

102 Empty Categories in Sentence Processing

might function, considerable care must be taken. It is generally assumed that some semantic processing of a sentence takes place at its end; an assumption which is backed up by data showing that reading times are longer for ﬁnal words (e.g. Tanenhaus et al. 1989; Pickering et al. 1994). Since there is no overt syntactic motivation for this slowing of reading towards the end of a sentence, this phenomenon is frequently attributed to semantic processing, perhaps the conversion of a syntactic representation into a semantic one. Although this is largely speculation, we might reasonably accept that eﬀects found at sentence ends could be semantic in origin. We shall discuss this prediction of the semantic processing in Chapter 5 below, but it is inapplicable here since our antecedent eﬀects are found mid-sentence. Since Fodor argues (1989, 1993, 1995) that certain uses of semantic features in HPSG are more compatible with the processing data, we feel justiﬁed in deriving some predictions from HPSG’s semantic feature structures. It seems reasonable to assume that a semantic account based on HPSG semantic features may predict reactivation of an argument wherever in a structure HPSG locates a semantic feature referring to it. In (102) we illustrate a (simpliﬁed) verbal sign in HPSG. We take the CONTENT features to be those which are semantic in the relevant sense here. (102)

PHON 〈give〉 HEAD CAT

verb

VFORM Wn

SUBCAT 〈NP[nom] 1 , NP[acc] 2 , NP[acc] 3 〉

SYNSEM LOCAL CONTENT

RELATION give GIVER 1 GIVEN 2 GIFT 3

Now consider a construction with a displaced argument such as (103).

Antecedent Reactivation and Trace 103

S

(103) NP [LOCAL 1 ] Crumble

S [SLASH { 1 }] 2 NP I

V [SUBCAT 〈 2 NP〉] [SLASH { 1 }] adore

The features of the NP crumble are located in sentence-initial position. Since crumble is topicalised, a subset of its features, the LOCAL values which includes the semantic features, are structure-shared, and might thus be argued to be represented elsewhere as well. They are passed from head to head down to the subcategoriser of crumble, namely adore. It is thus clear that the semantic features of a displaced element are available also at its subcategoriser. It follows that an HPSG-based model of the SPA can predict reactivation of an extracted category at the location of its subcategoriser, as well as sentence-ﬁnally. In our experimentation we shall test these predictions. 4.3.4 The Depth of Processing Account Janet Fodor (1995, p.c.) has recently put forward a very diﬀerent account of antecedent reactivation. This most naturally applies to data using an end-ofsentence testpoint, but can be sharpened to address experimental results showing activation at the trace location itself. She refers to work showing that arguments of less predictable structures are better recalled, perhaps because they are more deeply processed (Cairns & Blank 1976; Cairns, Cowart & Jablon 1981). For example, Cairns, Cowart & Jablon tested pairs of sentences such as (104) which diﬀered in the predictability of one word (here camera). (104) a.

Kathy wanted a snapshot of my baby, but she unfortunately forgot her camera today. b. Kathy ﬁnally arrived at my baby shower, but she unfortunately forgot her camera today.

In their ﬁrst experiment, subjects listened to the sentences over headphones and pushed a button when they had understood it. (104b) was read more slowly

104 Empty Categories in Sentence Processing

than (104a). The second was a phoneme monitoring experiment designed to test for processing load at the experimental item (here camera) Subjects identiﬁed the /t/ in today more slowly in (104b) than in (104a). These two results show that (104b) is more complex, and some of that complexity is located at the oﬀset of camera. In the third experiment subjects were visually presented with a probe at the end of the sentence. The item camera was recognized faster after (104b) than in (104a). The argument therefore runs: unpredictable items are less easy to process (Experiments 1 and 2), but the additional processing makes them better recalled (Experiment 3). Fodor refers to this as the depth of processing eﬀect (= DOP). Fodor (1995) suggests that a sentence with a displaced argument requires greater processing, because it is more complex, and thus its arguments will be better recalled. Thus far this story produces a very clever alternative account of sentence end probe recognition data purporting to show evidence of trace reactivation. Fodor herself refers to studies showing that subjects of passives are better recalled than subjects of actives. Thus in (105a), cheesemonger will be faster identiﬁed than in (105b). (105) a.

The cheesemonger transferred into ﬁrst class CHEESEMONGER slow b. The cheesemonger was transferred into ﬁrst class CHEESEMONGER fast

A supporter of the TRA would attribute this to a recency eﬀect caused by the presumed NP-trace after the passive participle; the DOP maintains that it is due to the diﬀerence in complexity between the two sentence types. However, the DOP can also account for diﬀerences in activation between moved and unmoved constituents from the same clause. The Cairns et al. data suggest that there is additional processing complexity at the displaced argument itself, relative to other constituents of same clause. This would explain why just this argument will be more activated than other arguments of the same clause. The DOP can thus explain why we might ﬁnd that camembert will be recognized faster than its controls in (106), where the TRA would make reference to reactivation by a wh-trace after the verb. (106) Which camembert did the cheesemonger send to the drayman? CAMEMBERT fast CHEESEMONGER slow DRAYMAN slow

Antecedent Reactivation and Trace 105

This localisation of the eﬀect in the displaced element itself would successfully account for eﬀects relative to controls from the same clause. Although this hypothesis has not been experimentally tested, a review of the literature reveals data which seems likely to be attributable to it. Consider for example Lewis, Shapiro, Afton & Tuller (1996), whose intention was to compare the eﬀects of plausibility and subcategorization in the verb-object relationship using two diﬀerent tasks, self-paced reading and CMLP. They tested sentences which were syntactic and plausible (107a), syntactic and implausible (107b), and nonsyntactic (107c). (107) a. Christa composed #1 the sentence #2 last night. b. Christa followed #1 the sentence #2 last night. c. Christa ached #1 the sentence #2 last night. Related probe: PRISON Control probe: FOREST

In their self-paced reading experiment the object was read more slowly in the nonsyntactic condition than in the other two cases. In their CMLP experiment there was more priming of the related probe at #2 in the nonsyntactic condition than in the other two conditions, which is probably caused by the longer reading time shown in the self-paced reading experiment. This pattern of results, for which the authors had no explanation, would seem to be a fairly clear-cut example of the DOP eﬀects in action, since it replicates the ﬁnding of a slowing of processing apparently causing an acceleration of probe recognition. This neatly accounts for data showing priming of displaced arguments from experiment types which are not location speciﬁc, such as sentence end probe tasks. In fact, if taken together with the DAH, the DOP can account for almost all the research ﬁndings. However, there are cases that it cannot cover. Consider, for example, an experimental sentence type such as (108), with CMLP testpoints marked #1 and #2. (108) What cheesei did the haberdasher send the shy #1 milliner ei #2 for Michaelmas?

Experiments on similar sentences would lead us to expect reactivation at #2, but no similar eﬀect at #1. This is because the eﬀect is location speciﬁc, unlike the maintained eﬀect that our formulation of the DOP so far predicts. Neither the DOP nor DAH accounts predicts this distribution of eﬀects, since the DOP requires eﬀects at #1 too, and the DAH requires #2 to be adjacent to a subcategorizer. However, Janet Fodor (p.c.) has recently made a suggestion how this

106 Empty Categories in Sentence Processing

approach can successfully include location speciﬁc reactivation. She suggests that there is additional processing of arguments at (potential) clause boundaries, perhaps as deﬁned by a saturated projection in HPSG or a Complete Functional Complex (Chomsky 1986b: 15). Although not experimentally conﬁrmed, such additional processing has sometimes been argued for. It has been suggested for dependency structures by Goodluck et al. (1991) and Bourdages (1992) that such processing occurs at a potential end of sentence (in order to explain why certain eﬀects seem to occur at object gaps but not subject gaps) (see also discussion of Experiment 1 in Nicol 1993 in 4.3.2.2). The predicted pattern of activation is illustrated in Figure 11.

Figure 11.Activation at gap positions in the Depth of Processing account

The graph illustrates schematically the activation levels of a displaced argument A and two other in situ arguments B and C for comparison, over the course of the processing of their clause. The background activation of the displaced argument A is higher than that of the others B and C, because it has been more deeply processed. At the clause boundary, the sentential processing causes reactivation of all arguments, but greater reactivation of A, for two reasons. First, because this extra activation can be thought of as being relative to the already existing state of activation. If the eﬀect is to double the activation level, the absolute increase will be greater for a extracted constituent. Second, the role that an extracted constituent plays in the sentential processing is a larger one, since it is accessed at two non-consecutive positions in the sentence. If the assumptions underlying the DOP hold, then the DOP can provide a traceless rationale for eﬀects at trace positions, since these will tend to be potential clause completion points. Consider (109a–c) below.

Antecedent Reactivation and Trace 107

(109) a.

What cheesei did the haberdasher send the shy #1 milliner ti #2 for Michaelmas? b. What cheesei did the haberdasher #1 eat ti #2 at Michaelmas? c. What cheesei did the haberdasher #1 speculate #2 that the shy milliner #3 ate ti #4 at Michaelmas?

Send in (109a) is a three-place predicate. There will be no potential end of sentence until its three arguments have been identiﬁed. The ﬁller may be one, but it still needs to ﬁnd a gap. The haberdasher and the milliner are the ﬁrst two ﬁrmly identiﬁed, and the parser can process the ﬁller as the third at the gap position. Since this is a potential clause boundary, the clause undergoes further processing, and since the DOP speciﬁes greater activation for displaced arguments, we would predict greater reactivation for the ﬁller than for the others at #2 but not at #1. This would exactly mimic presumed trace eﬀects. Sentence (109b) has a two-place predicate. We therefore predict greater eﬀects for the ﬁller than for the haberdasher at #2 but not at #1, since #2 is a potential clause boundary. Again, this is precisely what we would predict from antecedent reactivation. The verb in (109c) takes a clausal complement. The DOP predicts no diﬀerence at #1, #2 and #3, because none of these represent a possible clause closure. At #4, however, both verbs have their argument structures saturated and the DOP predicts greater reactivation of the ﬁller than of the other arguments. Once again the eﬀect will resemble antecedent reactivation at a trace site. An additional piece of empirical support for the DOP is its successful prediction of presumed ﬁrst resort gap location eﬀects (cf. Frazier 1987a; Clifton & Frazier 1989). Consider a sentence such as (110). (110) What cheesei did the cheesemonger send #1 the grocer ti #2 for May Day?

Since send takes three arguments, we would expect clausal boundary processing at #2; but this is also the gap location. Many empirical studies of gap location have however also shown eﬀects at #1. The traditional explanation of this has been to assume that gaps are posited on a ﬁrst resort basis, i.e. as early as structurally possible. The DOP oﬀers another account of this: it is also a potential clausal boundary. Since send can optionally appear with just two arguments, #1 is a possible termination of the sentence. Since the DOP actually predicts such eﬀects, unlike the TRA within which the ﬁrst resort proposal was formulated, this aﬀords strong support to the DOP. Thus the DOP is a very real contender for the best explanation of the experimental eﬀects found at gap positions, as it makes use of no intuitively implausible constructs and accounts for all current experimental results. It is

108 Empty Categories in Sentence Processing

not immediately easy to devise an experiment which would test its predictions against the TRA, because, as we have seen in (110a–c) above, complement gap sites which are not adjacent to verbs (to exclude the DAH) are usually also possible clause boundaries. 4.3.5 How can we disentangle these accounts empirically? In the light of the various competing accounts we have outlined above and the many demands on experimental materials which will be necessary to avoid confounds, is it possible to produce experimental evidence to distinguish between the TRA and other accounts and thus shed processing light on the contrasting treatments of gap dependencies in PPT and HPSG? The number of constraining factors is large. First we must use a CMLP methodology, since this has been shown to be free of integrative eﬀects, and produces location-speciﬁc evidence for the processing reality of a trace, and not just evidence of complexity. Next, we must test for activation of a trace which is non-adjacent to a subcategorizer, to control for the DAH and the SPA. In addition, the subcategorizer must not be a contentless (“case-checking”) preposition as there is some doubt whether these are actually real subcategorizers. Fourthly, the ﬁller gap dependency should not cross the subcategorizer, as this produces unclear results due to the possibility of trace prediction. Fifthly, the gap location must not be at a potential end of clause, so that we can control for the depth of processing account. The list is intimidating, but fortunately several of these conditions repeat themselves in the limits that they put on possible structures. In our experiments below, we succeed in constructing sentences which fulﬁll these conditions by making use of verb second clauses in which both ﬁller and gap follow the subcategorizer. Let us ﬁnally note that our experimental materials, which we shall present in the next section, are primarily analysed in the framework of PPT. In many cases the HPSG analyses will diﬀer from these, and we shall sometimes note this, but the PPT analysis is the crucial one. The reason has to do with the logical structure of our experiments to test for trace reactivation. Our primary objective is to establish whether empty categories do or do not play a role in processing, and to do this we must test the TRA. Our interest in the alternative accounts, the DAH, the SPA and the DOP, is only to ﬁnd out if they can render the TRA as an account of the attested data superﬂuous; they do so if they can account for all the trace reactivation data. It is not our primary aim to establish whether the predictions of these alternative accounts hold or not, only whether

Antecedent Reactivation and Trace 109

they prove to be more empirically adequate than the TRA, in that they predict the occurrence of trace reactivation eﬀects more accurately than the TRA. It must be clear, therefore, that testing for eﬀects at positions where the alternative accounts predict eﬀects, but the TRA does not, may conﬁrm aspects of the alternative accounts, but cannot answer our primary question. Testing at locations where the TRA predicts eﬀects, but the others do not, on the other hand, will allow us to draw conclusions about the processing of traces. Since PPT uses traces, this is the framework that our hypothesis must be phrased in.

4.4 Experimental materials and design While it is not possible in English to construct experimental materials which will permit us to distinguish the predictions of the various accounts of antecedent reactivation data, German allows this. German diﬀers from English in a number of ways which are helpful to us here. First, the German VP is headﬁnal. This assists us in distinguishing between reactivation at the subcategorizing verb and reactivation at a gap location. Any eﬀect which is part of the processing of a dependency on the verb requires that the verb be recognized ﬁrst. This would suggest that any reactivation of a dependent will be slightly downstream of the verb. If objects follow their verbs, as in English, then this will exactly coincide with the gap position. However, if objects precede verbs, as in German, then eﬀects of trace can appear before the verb has been processed. While already this diﬀerence gives rise to measurably diﬀerent predictions of the TRA and the DAH, other characteristics of German permit us to draw even clearer distinctions. German clauses conform to a constraint known as Verb Second (= V2). Although this is not wholly uncontroversial, it is generally assumed in PPT that this involves the ﬁnite verb raising from head V to a clause-initial functional head, perhaps C (cf. von Stechow & Sternefeld 1988:388ﬀ), perhaps via a lower clause-ﬁnal functional head, perhaps I (for a contrary view see Haider 1993 Chapters 3 and 4).

110

Empty Categories in Sentence Processing

CP

(111) spec

C´ C verb

IP subject

I´ VP

spec

tverb V´

object

tverb

This leaves the object position remote from the overt position of the verb. Now, if we extract the object from this position, we are left with a hypothesized trace which is not adjacent to its subcategorizing verb. This, as we noted in Section 3.5 above, is the conﬁguration required to distinguish between the TRA on the one hand and the DAH and the SPA on the other. The movement from object position that we chose for our experimental materials was short scrambling. We discuss the structural analyses of this phenomenon in the section below, but here we shall ﬁrst illustrate the structure of the experimental sentences in our materials. The word order changes introduced by V2 and short scrambling are illustrated informally in (112). (112) a.

Base generated verb ﬁnal Tante Anna dem Mädchen den Hund schickte >>>>> aunt Anna the girl the dog sent “Aunt Anna sent the dog to the girl” b. Verb second form: Tante Anna schickte dem Mädchen den Hund >>>>> aunt Anna sent the girl the dog c. Scrambled form: Tante Anna schickte den Hundi dem Mädchen ti aunt Anna sent the dog the girl

Antecedent Reactivation and Trace

Sentence (112a) shows the word order of the sentence with the constituents in their canonical order; note that this is a grammatical embedded clause. (112b) shows the eﬀect of V2. The verb has moved from ﬁnal position into second position, assumed to be C. (112c) shows the eﬀect of scrambling. In our experimental sentences, we introduce the direct object (= DO) ahead of the indirect object (= IO), an inversion known as short scrambling. On the assumption that this is a movement operation, this last step leaves behind a trace which is not adjacent to its subcategorising verb. This combination of V2 and object scrambling forms the structure which underlies our experimental materials in this chapter. 4.4.1 An excursus on object scrambling “Scrambling” is used as a cover term for a wide range of seemingly optional or stylistically motivated word order variations (see Introduction to Corver & Riemsdijk 1994; and Kiss 1994 for discussion). Here we shall restrict ourselves to a discussion of the derivation of DO > IO order, a marked order of complements in German. This is illustrated in (113), where (113b) is the marked word order, often attributed to scrambling. (113) a.

Tante Anna schickte dem Mädchen den Hund aunt Anna sent the.dat girl the.acc dog b. Tante Anna schickte den Hund dem Mädchen aunt Anna sent the.acc dog the.dat girl

The classic source for the descriptive generalisations about object scrambling is Lenerz (1977). He ﬁrst identiﬁes the order IO > DO as the “unmarked order” of NP complements (= AB in (114)–(117) below), using a criterion of acceptability in the widest range of circumstances but excluding such performance factors as absolute frequency (pp. 26ﬀ).48 He then isolates four factors which can inﬂuence the acceptability of the marked order (= BA in (114)–(117) below). We list and summarise these here (from Lenerz 1977: 63):

48.“Unmarked order”: If two constituents A and B can appear both in the order AB and in the order BA, and if BA is subject to deﬁnite testable conditions that AB is not subject to, then AB is the “unmarked order” and BA is the “marked order” (Lenerz 1977, my translation).

111

112

Empty Categories in Sentence Processing

(114) Theme–Rheme Condition Scrambled order BA is less acceptable if B is more rhematic than A. (115) Deﬁniteness Condition Scrambled order BA is less acceptable if B is not deﬁnite. (116) Law of Increasing Members There is a stylistic tendency for heavier constituents to follow lighter ones. (117) Clause-ﬁnal Verb Condition There is a stylistic tendency not to allow a light constituent to end a clause without a clause-ﬁnal verb.

Other factors might be added to this list; pronominal arguments must always precede full NPs, for example, and certain verbs seem to be lexically marked as preferring DO > IO as the unmarked order of their complements, but the basic premise of there being an underlying order, deviations from which must be justiﬁed by broadly pragmatic factors is widely accepted as a descriptive generalisation (but see e.g. Reis 1986 for an alternative view). Syntactically there is no consensus on the analysis of scrambled structures and the literature is large. Fortunately, for the purposes of our processing experiment, the exact analysis is unimportant as long as the supposition that scrambling is a movement construction holds. This is necessary because our experiment necessitates a movement trace at the canonical position of the object. We brieﬂy review analyses of scrambling below. Analyses of scrambling in PPT may be divided into those which assume that the DO > IO order is derived by movement and those which assume that it is base-generated (see Stechow & Sternefeld 1988; Grewendorf 1995 for review, papers in Corver & Van Riemsdijk 1994; Grewendorf & Sternefeld 1990 for discussion). The movement analysis of German short scrambling seems to have originated with Thiersch (1978, 1982). Thiersch accepts that IO > DO is the D-structure representation and argues that DO > IO is derived by the adjunction to VP of the DO. This would give a structure such as (118) for a clause with object scrambling.

Antecedent Reactivation and Trace

IP

(118) Tante Anna Aunt Anna

VP NP

VP

den Hundi the dog

NPDAT

dem Mädchen the girl

V´ ti

V schickt sends

Thiersch does not specify a particular motor for this movement process, but it is central to the account that scrambling consists of leftward movement, leaving a trace. Later movement accounts (e.g. Webelhuth 1989; Saito 1989, 1992; Mahajan 1990; Fanselow 1990; Haider 1993, 1994, 1997; Deprez 1994; Müller & Sternefeld 1994; Poole 1996; Vikner 1997; Kitahara 1997; Grewendorf & Sabel 1998) use a range of diﬀerent landing sites and motivations for movement, but can essentially be seen a variants of this model for the purposes of our experiment. These later contributions are often attempts to reformulate the movement model of scrambling in terms of the Minimalist Program, which is problematic because scrambling seems to be optional, a quality not permissible for movement within the Minimalist Program. As an example of this we may cite Kitahara (1997: 77ﬀ), who suggests that scrambling is driven by the checking requirements of a category bearing “a strong argument feature”. The contents of the VP would thus be merged as something like [VP dem Mädchen [V¢ den Hund [V schickt]]]. A functional category with a strong feature [+D] would then be merged into the derivation, triggering the movement of den Hund to its speciﬁer position. Clearly some explanation is required for the inability of the indirect object to check this strong feature, since it is plainly closer, and thus economy principles would demand that it, and not the more remote direct object, be attracted. Kitahara does not suggest such a mechanism, but one can readily imagine the strong feature on the functional category being more exactly speciﬁed, perhaps for the Case of the DP it will attract. This solution is implicit in the informal tree diagrams of the underlying clause structure of double object structures current in the literature which distinguish and AgrOP and an AgrIOP.

113

114

Empty Categories in Sentence Processing

AgrSP

(119) AgrS

TP T

AgrIOP AgrIO

AgrOP AgrO

VP IO, DO, ...V

In a schematic clause structure such as (119), not only the subject and object agreement projections are distinguished but also the direct and indirect object agreement projections. For such an arrangement to work, these two must be permitted to select DP whose features they are to check. The alternative non-movement view of scrambling originated in discussion of non-conﬁgurational languages (Farmer 1980; Hale 1980, 1981). Scherpenisse (1986) formulated this account of short scrambling in German, on which the marked order DO > IO is base-generated. He motivates this by suggesting that the focusing of a constituent is a lexical process, and that this causes the incorporation of this constituent into the verb. This then would result in the following structure for our clause. IP

(120) NP

VP

Tante Anna Aunt Anna

V´ NP denHund the dog

V NPDAT

V

dem Mädchen the girl

schickt sends

Antecedent Reactivation and Trace

For our purposes, the important quality of this analysis is that it involves no movement and thus leaves no trace. More recent non-movement accounts of scrambling put forward a range of diﬀerent mechanisms which would permit the generation of a direct object non-adjacent to its subcategorising verb (e.g. Haider 1988; Bayer & Kornﬁlt 1991; Neeleman 1994; Kiss 1994; Tonoike 1994; Fanselow 1997). For example, Fanselow concludes that no derivational account of scrambling is possible within the Minimalist Program, because the Minimal Link Condition excludes the movement of any category over another of the same type. On Fanselow’s account, word-order variation is freely permitted because features of the subcategorizing head are checked against features of the DP-objects under c-command, not adjacency, and theta assignment is subsumed within Case checking. It should be noted that, while this account permits the DO > IO word order of short scrambling, it does not contribute towards accounting for its appearance in certain speciﬁc contexts, and wrongly predicts free word order. The data set called scrambling in PPT is known in HPSG as free word order phenomena. Analyses in HPSG tend to resemble PPT base-generation analyses in spirit, but vary in mechanism (see Borsley 1996 for review). Perhaps the consensual approach is that of Pollard & Sag (1994: 38–42). This may be summarised as assuming a ﬂat VP and then employing Linear Precedence rules which produce the requisite ordering of sisters. This however is problematic because of the principle of Exhaustive Constant Partial Ordering (Pollard & Sag 1987), which states that any given set of sisters should be ordered the same no matter what their mother or daughters, thus excluding “free” word order. There are at least three solutions which might be envisaged. First, one might simply introduce a new feature to motivate the word order change, similar to the use of [INV] by Pollard (1990), and [FLIP] for verbal complexes by Hinrichs & Nakazawa (1994). Uszkoreit (1986) posits much more complex Linear Precedence rules to account for the pragmatic inﬂuence on word order (basically Lenerz 1977 put in PSG terms). Reape (1994) suggests “Domain Union”. He assumes that linear precedence rules are purely binary, that is they relate only to the relative order of pairs of elements. He then suggests that these binary pairs may be interleaved. Thus if X > Y and A > B, then XABY, XAYB etc but *XBAY, *YXAB etc. A very diﬀerent approach to clausal structure is to assume a hierarchically organised VP and allow a head to combine with its arguments one at a time instead of all at once (Netter 1992; cf. Gunji 1987 for Japanese). This removes the Exhaustive Constant Partial Ordering problem and allows diﬀerent orders to be generated freely. To conclude we shall mention Jacobson (1987). This

115

116

Empty Categories in Sentence Processing

approach to variations in clausal word order is the most closely related to the PPT movement accounts as it involves a [DSL] (double slash) category, which is an implicit acceptance that the category is not in its canonical position. Jacobson applies this to extraposition rather than free word order, but in principle such an approach could be applied to produce variant clause-internal argument orders as well. We shall not pursue details of the competing scrambling analyses here, but note that our experimentation using scrambled structures can contribute some evidence which will help adjudicate between the movement and base-generation analyses of scrambling. If we ﬁnd evidence of trace reactivation, this will support movement models of scrambling. Should there be no evidence of traces at the canonical DO position, this would oﬀer no such support. It would not, however, constitute any positive evidence for or against any particular view of scrambling, since the failure to ﬁnd trace eﬀects might be due to some other characteristic of processing. 4.4.2 Methodology 4.4.2.1 CMLP Procedure The basic methodology was common to all three experiments we report in this chapter: minor diﬀerences will be pointed out as we proceed. We adapted the cross-modal lexical priming procedure (discussed for example in Swinney et al. 1979; Nicol & Pickering 1993) in which subjects listen to texts over headphones whilst simultaneously performing a recognition task on targets presented visually. Targets are presented at two points: at the trace position and at a control position; this provides assurance that eﬀects found are speciﬁc to the trace location. Two sorts of targets are presented, related to the antecedent and unrelated: this ensures that eﬀects are speciﬁc to the antecedent. This methodology is the preferred one for investigation of antecedent reactivation eﬀects. The cross-modal element excludes eﬀects based on pure short term memory within a single perceptual mode. It permits the presentation of the prime without interruption of the stimulus input, allowing an on-line measure to be made (cf. Hickok 1993). CMLP also taps more reliably into sentence processing than a task such as probe recognition which measures the response time of a consciously performed end-of-sentence task (e.g. Bever & McElree 1988). Furthermore, it allows the subject to parse the stimulus materials at normal speed, unlike techniques such as self-paced reading or ERP measurement, in which the subject processes the input at an unnaturally slow

Antecedent Reactivation and Trace

rate (cf. Pickering et al. 1994). The experiment was controlled by the NESU software package (Baumann et al. 1993). Subjects listened to 24 short texts over headphones whilst seated in front of a 17≤ monitor. Of these 12 were experimental texts and 12 ﬁller texts. Every 5–10 seconds a word in white Arial 36 point font appeared on the black screen, and the subjects pushed a green button if it was a real word, and a red one if it was a non-word (a lexical decision task). About half of these targets were words and half non-words. Non-words were real German words in which one letter or cluster was changed in order to make a phonotactically licit but non-existent string: (121) Real word Brunst >>> Saugen >>> Baum >>>

Derived non-word *Frunst *Saugel *Baim

The aim of the visual targets is to tap into the on-line processing of the auditory input that the subject is performing. To this end the experimental targets are positioned very precisely to coincide with the points in the input which the researcher wishes to investigate. If antecedents undergo some processing at gap positions, then this should be detectable as activation, since it should enable subjects to recognize identical visual targets more quickly. We illustrate the experimental sentences in (122). (122) Anna schickte den Stifti einer ihrer vielen begabten #1 Nichten ti #2 Anna sent the pencil one of.her many gifted nieces Identical target: STIFT “pencil” Unrelated target: BLITZ “lightning”

Den Stift, the direct object, has been scrambled into a position before the indirect object leaving a presumed trace at its canonical position. The indirect object was made 8–10 syllables long to remove the antecedent from short term memory. The gap position #2 was determined using Triple Magic, a digital sound editor. It was assumed to be at the location where the oﬀset of the preceding word and onset of the following word intersect. The control position #1 was always exactly 400ms earlier. In order that subjects should simultaneously pay attention to the aural texts, every ﬁve minutes the experiment interrupted and subjects answered questions about the content of the most recent texts by ticking boxes on a sheet. Accuracy on these questions was not analyzed. Before commencing the experiment, all

117

118

Empty Categories in Sentence Processing

subjects read an instruction sheet and were given a chance to ask questions. They then completed two practice sessions with further opportunities to ask questions after each, before moving on to the main part of the experiment. Each recorded experimental sentence was presented identically but with visual targets in four diﬀerent counterbalanced conditions: a quarter of the subjects saw each condition. For each experimental sentence, the ﬁrst group of subjects saw the antecedent target at the trace position; another at the control position 400ms earlier. The third group saw an unrelated target at the gap position, and the last saw it at the control position. Every subject saw equal numbers of targets in each condition. 4.4.2.2 Subjects Subjects were students who participated for payment either at the University of Essex or at the Heinrich-Heine Universität Düsseldorf. All were native speakers of German. They were tested individually in dedicated rooms without disturbance. The experiments were conducted entirely in German. There were 58 subjects in Experiment 1 (34 women, 24 men, mean age 28.3 years), 82 subjects in Experiment 2 (48 women, 34 men, mean age 27.7 years), and 30 subjects in Experiment 3 (17 women, 13 men, mean age 28.9). 4.4.2.3 Input stimulus Each experimental sentence was embedded in a short narrative text. These texts varied between 60 and 180 words in length and were semantically coherent in that they told a short anecdote. This was felt to be important for three reasons: ﬁrst, because it made it much easier for the subjects to pay full attention to their content. It is plain that they must be listening with full attention for them to notice that an element is out of canonical order and bother to reconstruct it. Second, it was necessary to motivate the scramble: German scrambled sentences without discourse justiﬁcation sound deviant. For subjects to reliably posit a trace of a moved argument at the right place, the structure of the sentence must seem at very least possible and preferably almost predictable. The third reason was to tie in the stress and intonation patterns of the text with the appropriate structure. It is very diﬃcult to read a scrambled sentence aloud with realistic prosodic pronunciation out of context, but scrambles sound extremely odd if read with ﬂat intonation and unmarked sentence stress. Some recent work on similar structures in Japanese (Nakano 1998), where scrambled sentences presented without contextual justiﬁcation produced no signiﬁcant results, may be taken as support for this move. Some ﬁller texts contained no scramble but

Antecedent Reactivation and Trace

frequently another marked structure or feature to help the scrambles to blend in. In order to ensure that the structure was indeed interpreted as an instance of scrambling, the exceptional verbs which seem to show DO > IO as the unmarked order were excluded, using the results of Mergel (1994). The texts were recorded by a female German native speaker from the dialectally neutral area of Hannover. The full set of input stimuli and an example context can be found in the appendix. 4.4.2.4 Targets The targets to be judged within the CMLP procedure (Swinney et al. 1979) are usually strong semantic associates of the preceding priming stimulus, following on the priming within semantic ﬁelds demonstrated by Meyer & Schvanenveldt (1971). The mental activation therefore occurs indirectly: with one anaphoric connection from antecedent to the foot of the dependency, and with a second from the reconstructed semantic content of the dependency to a semantic associate within the lexicon. We did not adopt this methodology but rather presented as targets the trace antecedents themselves, thus eliminating one of the two associations that need to be made to feed the priming eﬀect. There were two reasons for this. The ﬁrst was that there is no published directory of experimentally determined semantic associates for German and so we should have had to do this work ourselves, which would have been a study in its own right. The second was that we wished to maximize the priming eﬀect, in case the optional movement phenomenon should produce weaker eﬀects than the operator movement dependencies that are usually studied. Since it enables priming to occur at ﬁrst and not at second hand, the use of the antecedents themselves may well make the methodology more sensitive. The reason that semantic associates rather than the antecedents themselves are standardly used is in order to avoid subjects realizing that certain targets had just appeared in the input. If subjects started to anticipate the connection, they might start to attempt to remember antecedents or develop other conscious strategies. We avoided this problem by the alternative approach of using a larger proportion of ﬁller targets than is usual. This method was easy to implement because, as mentioned above, we wished to avoid presenting our scrambling sentences in a semantic void, and so had embedded them in a motivating narrative context. The proportion of ﬁller targets to experimental targets was more than 10:1, but since half of the experimental items were unrelated control targets, fewer than one in twenty targets had occurred in the immediately preceding text.

119

120 Empty Categories in Sentence Processing

The use of real antecedents as targets was also camouﬂaged by the inclusion of ﬁller targets which had a semantic association with the current auditory text but which had not themselves appeared in the text. The examples of semantic contexts and targets given below demonstrate the eﬀect. (123) “and so the wicked witch turned the prince into a frog” “the hushed court rose as the ancient judge entered” “the carpenter scratched the desk with the chisel” “to pay all that money to the taxman is laudable”

MAGIC JURY WOOD MONEY

A subject hearing the ﬁrst three will not be surprised by the fourth and will probably scarcely be aware that it diﬀers from the others in that the identical target appeared in the preceding text. This approach had the added advantage of encouraging subjects to pay full attention to the content of the texts, as it would assist them in making their lexical decisions. After they had ﬁnished the experiment, subjects were asked whether they had any idea what the aim of the experiment was. No subject mentioned that the targets in scrambled sentences had previously occurred in the text. The proportions of items, words and non-words amongst the targets are given in the table below (includes practice sessions). For each subject, half of the items were identical targets and half unrelated targets. Table 6.Numbers of items and ﬁllers in our Experiments 1–3 Real words

Exp. 1 Exp. 2 Exp. 3

Items

Fillers

Total

12 12 12

60 54 58

72 66 70

Non-words

Total

64 67 76

136 133 146

The experimental targets were carefully matched with their unrelated equivalents for length, frequency (lemma frequency from the CELEX lexical database (Baayen et al. 1996)), syllabic structure and etymological origin. The full set of targets can be found in Appendix 1. 4.4.2.5 Statistical treatment The data from each experiment was processed in a parallel manner. First the data of six subjects (out of 170) was excluded on the basis of three criteria: because of experimental error (three cases), because the subject had a mean RT

Antecedent Reactivation and Trace

of more than 1000 milliseconds (two cases), or because the subject had a standard deviation of RT of more than 250 milliseconds (one case). Next all RTs of erroneous lexical decisions were eliminated, and lastly all scores exceeding two and a half times the standard deviation by experimental condition (i.e. by testpoint and target type) were excluded from further analysis. In each experiment the items excluded made up less than 5% of the data. The RTs were converted into logarithms in order to reduce their skewedness for input into the anova calculations, though we shall use the unconverted ﬁgures within the text because they are more perspicuous.

4.5 Experiment 1: Simple SVO Our ﬁrst experiment used sentences with the single lexical ﬁnite verb raised to the V2 position and the direct object scrambled over the indirect object (124): (124) Sie übergab das Paari einem unbewaﬀneten Ver#1mittler ti #2 she handed.over the pair an unarmed mediator “She handed the pair over to an unarmed mediator” Identical target: PAAR “pair” Unrelated target: BROT “bread”

Since this leaves the DO gap in sentence-ﬁnal position, remote from any verbal material, we may test the contrasting predictions of the TRA on the one hand and the DAH and the SPA on the other hand. The test positions are therefore the oﬀset of the ﬁnal word in the sentence, which is where we would locate the trace of the scrambled DO, and a control position 400ms earlier #1. While the TRA predicts eﬀects at the trace location #2 and the DAH and SPA predict eﬀects at the subcategoriser, and crucially these two positions are not adjacent here, the data will allow us to distinguish between them. As well as having a gap position remote from the subcategoriser, there is one additional feature which also distinguishes our materials from those previously used. Recall that Nicol (1993) attempted to clarify whether antecedent eﬀects were really located at trace positions or in fact at subcategoriser positions as suggested by the DAH (see Section 4.3.2.2 above). She used CMLP to test for reactivation of the PP wh-item in sentences where this would be nonadjacent to the verb (125).

121

122 Empty Categories in Sentence Processing

(125) To which butcheri did the woman who had just inherited a large sum of money give #1 the very expensive gift ti #2 the other day?

She found signiﬁcant reactivation of the ﬁller at both testpoints. This left the question unresolved, because it allowed the data to be interpreted either way. One could either claim that the eﬀect at the trace was just a secondary eﬀect of structure checking and that the main eﬀect was at the verb, as predicted by the DAH, or equally but oppositely, one could claim that the main eﬀect was at the trace, as predicted by the TRA, and that the eﬀect at the verb was merely a First Resort eﬀect, given that To which butcher did the woman (…) give? is a legitimate sentence (see Section 3.3.1 above). Our experimental materials exclude such ambiguity. Consider Nicol’s example experimental sentence once more. The alternative interpretations appear because the ﬁller-gap dependency crosses the verb (126). (126) To which butcheri did the woman (…) give #1 the (…) gift ti #2

In our sentence type both the head and the foot of the dependency occur after the verb and the DO is in a standard argument position. The structure checking interpretation of the eﬀect at the gap position is thus not possible, since the DO is unambiguously a subcategorised dependent of the main verb. It is not in an È position and cannot have undergone cyclic movement and thus possibly belong to a diﬀerent clause. There is therefore no question of a provisional dependency being made at the verb and structure checking carried out at the trace position: the relationship between the verb and the argument is made. Similarly, there is no potential for a First Resort eﬀect account of ﬁller assignment. Our German materials using short scrambling thus oﬀer us the opportunity to clarify this question. Note that this experiment does not control for the DOP, as this alternative account had not been suggested to us at the time this experiment was undertaken. 4.5.1 Results Figure 12 shows the mean lexical decision times on the visual targets collapsed across subjects and items. We performed two analyses of variance on the data, ﬁrst by Subject (= F1) and then by Item (= F2). The main eﬀect for Target type was signiﬁcant on both analyses (F1 = 62.01, p < 0.01, F2 = 63.43, p < 0.01), as was the main eﬀect for Subject on the by subjects analysis (F = 3.89, p < 0.000, and for Item on the by items analysis (F = 3.04, p = 0.001). The main eﬀect for Testpoint was nonsignif-

Mean RT in ms

Antecedent Reactivation and Trace

740 720 700 680 660 640 620 600 580

724

707

622

Unrelated targets

612

Identical targets

Testpoint Control Trace

Figure 12.Experiment 1 reaction times in milliseconds by target type and testpoint

icant on both analyses (F1 = 0.42, p > 0.5, F2 = 0.42, p > 0.5). There was no signiﬁcant interaction of Testpoint × Target type (p > 0.7). The most marked feature of these results is the diﬀerence in RT between the target types. Since the identical target is a word that has occurred recently in the input, it is unsurprising that it is recognized more quickly than the unrelated target. The eﬀects for Subject and Item show us that variation between subjects was noticeable and that the experimental items were less homogeneous than one might wish. It is fairly clear that the testpoint makes little diﬀerence to the RTs, on the other hand. Table 7.Experiment 1: Mean reaction times and priming eﬀects in milliseconds Experiment 1 Control position (X) gap position (Y) Testpoint eﬀect (X–Y)

Unrelated targets (A)

Identical targets (B)

Target type eﬀect (A–B)

724 707 17

622 612 10

102* 95* 7*

* Eﬀects are signiﬁcant

Let us brieﬂy recall what we are looking for in this data. The data from the unrelated targets forms a baseline which will expose any irrelevant background eﬀects, such as a diﬀerence in processing complexity at the two test positions. In fact the diﬀerence is small, only 17ms. Similarly, we can look at the data just from the control position, to identify any baseline diﬀerence between the two target types. This is large at about 100ms, representing the maintained activation of the identical targets. The statistic which would reveal antecedent reactivation eﬀects is the interaction of testpoint and target type. In fact the data from the identical targets at the gap position is almost exactly what we should

123

124 Empty Categories in Sentence Processing

expect on the basis of the two baselines. We can therefore conclude that there is no sign of antecedent reactivation in this data. 4.5.2 Discussion Since the TRA predicted reactivation eﬀects at the presumed trace position, it is not supported by this result. The DAH and the SPA predicted no experimental eﬀects at this position, and are therefore supported as accounts since they are shown to be more empirically adequate. The data is similarly consistent with base-generation analyses of the short scrambling construction, since they predict the results as recorded. However, after this experiment was carried out we became aware of some alternative explanations. The ﬁrst relates to the syntactic analysis of this sentence type. Some syntacticians have suggested (e.g. Reis 1980) that the standard head-ﬁnal VP analysis of such clauses is wrong and that such German V2 clauses with a single, ﬁnite verb in second position are actually head-initial. There are two strands of the evidence for this. The ﬁrst is that there is no overt evidence of a sentence-ﬁnal verbal element in such sentences, and the second is that such sentences show eﬀects unlike other sentence types. Consider the following sentences (from Reis 1980): (127) Causal adverbial > manner adverbial a. daß Hans wegen des Tadels sorgfältig schreibt that Hans due.to the reprimand carefully writes b. Hans hat wegen des Tadels sorgfältig geschrieben Hans has due.to the reprimand carefully written c. Hans will wegen des Tadels sorgfältig schreiben Hans wants due.to the reprimand carefully to.write d. Hans schreibt wegen des Tadels sorgfältig Hans writes due.to the reprimand carefully (128) Manner adverbial > causal adverbial a. *daß Hans sorgfältig wegen des Tadels schreibt that Hans carefully due.to the reprimand writes b. *Hans hat sorgfältig wegen des Tadels geschrieben Hans has carefully due.to the reprimand written c. *Hans will sorgfältig wegen des Tadels schreiben Hans wants carefully due.to the reprimand write d. Hans schreibt sorgfältig wegen des Tadels Hans writes carefully due.to the reprimand

Antecedent Reactivation and Trace

The manner adverb compulsorily precedes the causal adverbial except in those sentences where there is no overt evidence of a sentence-ﬁnal verb position, where the reverse is not only permissible, but even perhaps preferred. This can be seen as evidence that the verb phrase has a reversed head direction in this sort of clause. While it is less economical to posit two diﬀerent clause structures for declarative clauses, it cannot be entirely rejected. One possible motivation for such a situation is that German is in the historical process of changing from a head-ﬁnal to a head-initial VP, as English has. In order for such long-term alterations to come about, there will plainly be times when rather theoretically messy situations will occur, and it is possible, though somewhat speculative, that the German VP is in such an intermediate situation. If this is correct, we would only expect trace-based reactivation from sentences with some clauseﬁnal verbal material. Another complicating factor becomes visible when one examines the processing characteristics of our experimental sentences. Since the basic structure of the clauses is Subj+V+DO+IO, the verbs chosen subcategorise for three arguments. Both the experimental testpoints appear after the beginning of the IO, the third and ﬁnal argument. Since this can be plainly identiﬁed as such as soon as its dative article is reached it is possible for the remainder of the sentence to be unambiguously predicted. The encountered argument must be the IO, and no further arguments are possible. The sentence intonation underlines this: the intonational contour signals clearly that the sentence is being ﬁnished. That being the case, it is possible that the parser starts the sentence wrap-up processing (see discussion of DOP above) already at this stage. This would involve predictive location of the trace position too. Since the continuation of the sentence becomes unambiguously predictable before the control testpoint, it is possible that this predictive sentential processing is contaminating this position by producing extra activation here. At our gap position, on the other hand, the reactivation would on this account be much reduced, as the trace location had been posited and processed in advance. This would predict the pattern of results we found. A third possible reason for the lack of antecedent reactivation in this experiment is the position of our second testpoint at a clause boundary. As a rule researchers do not use testpoints at clause boundaries precisely in order to avoid end of sentence noise obscuring results. There is plenty of psycholinguistic evidence that the processes involved in responding to a sentence ﬁnal word and an embedded word diﬀer (e.g. Carpenter & Just 1987). One diﬀerence which can be easily seen is that reading times are longer for ﬁnal words;

125

126 Empty Categories in Sentence Processing

the data from self-paced reading tasks conﬁrms that there is a general slowdown of reading speed at the end of a sentence (see Tanenhaus et al. 1989; Pickering et al. 1994 and our own self-paced reading experiment). Diﬀerences have also been shown to exist in the ERP response to ﬁnal words (Friedman, Simpson, Ritter & Rapin 1975; Osterhout 1994). It seems possible that this was a factor which prevented antecedent reactivation being observed. In the light of these doubts about the reliability of the data from our ﬁrst experiment, we revised our experimental materials to avoid the problems we identiﬁed.

4.6 Experiment 2: Particle verb constructions In our second experiment we used the same methodology but changed the verbs in the experimental materials. In this second experiment, the trace position is followed by a particle at the end of the sentence. (129) Nach zwei Tagen Streit sprach der Richter das Geschäft dem after two days dispute awarded the judge the shop the ziemlich überraschten Andreas # zu. somewhat surprised Andreas part ‘After two days of dispute the judge awarded the shop to the rather surprised A.’ Identical target: GESCHÄFT “shop” Unrelated target: WAGEN “car”

This has three implications for our experiment. Firstly, this is a structure which has an uncontroversial head-ﬁnal VP. This removes the doubt from the syntactic perspective whether antecedent eﬀects should be expected in the structure. Secondly, the argument structure of the main part of the verb alone, the sprach at the V2 position, does not permit three arguments; only in combination with the particle zu does this become a saturated projection. The continuation of the sentence is thus not unambiguously predictable when the IO is reached, since further lexical material must follow to ensure grammaticality of the sentence, and so the parser cannot commence sentential wrap-up processing on a predictive basis. The alternative processing account we identiﬁed above is thus controlled for. Thirdly, the gap testpoint is no longer at a clause boundary and therefore the possibility of noise from the processing of the subsequent clause is excluded. In addition, the location of this testpoint at a point which is not a clause boundary allows us to control for the predictions

Antecedent Reactivation and Trace 127

of the DOP, which suggests that presumed trace reactivation eﬀects are in fact sentential processing at potential clause boundaries. 4.6.1 An excursus on particle verbs Because of their importance to the interpretation of the results of this experiment we shall brieﬂy discuss the syntactic status of these sentence-ﬁnal elements. These particles (in this case zu) would originally have been prepositions or adverbs, but many verb–particle combinations have been lexicalised to some extent, the extent varying quite widely from case to case. As examples from opposite ends of the spectrum one might cite zurückbringen and beibringen. The ﬁrst is a close translation equivalent of English bring back and has similar qualities: both parts of the verb may combine with other items or indeed stand on their own with little change of meaning and no change in argument structure. The second is compositionally akin to English get over in the sense “convey a message”, but is the most common informal word for “teach”. While both bringen and bei exist as free forms or in other combinations, they then have completely diﬀerent meanings quite unrelated to “teach” and take a diﬀerent range of arguments: bringen is English “bring” or “take”, bei is French “chez” or Latin “apud”. The syntax of sentence ﬁnal particles is somewhat uncertain (see Stiebels & Wunderlich 1994; von Stechow 1991; also Schreuder 1990 and Frazier, Flores D’Arcais & Coolen 1993 for processing). The suspicion is that particle verbs are discontinuous elements, which theories would often wish to proscribe. Nevertheless, one can readily formulate a syntactic analysis of these items within PPT, because verb movement from ﬁnal to V2 position provides a ready account of their position. The word order in the embedded clause (130a) is equivalent to the hypothesized unmoved order. V2 causes the verb to raise from its ﬁnal position to the second position in the clause, head C on our assumptions. The particle however is stranded. (130) a.

[CP [C] der Lehrer den Kindern Latein beibringt] the teacher the children Latin teaches b. [CP [C bringt] der Lehrer den Kindern Latein bei tV] teaches the teacher the children Latin part

We shall follow Radford (1997: 437f ) here in assuming that the particle, a head P, is incorporated into the V2, giving a complex verbal projection V1.

128 Empty Categories in Sentence Processing

V1

(131) P

V2

bei

bringen

When the verb V2 is raised it excorporates out of the V1 leaving the particle stranded, but still part of a structure. V1

(132) P

V2

bei

t

This account of ﬁnal particles is not available to HPSG, however, as it makes no use of verb movement. The standard assumption seems to be that these particles are subcategorised dependents of their head verbs and appear on their SUBCAT lists (Sadler p.c.). The reason that they appear separately in V2 structures is simply that they are no more intimately connected with the head verb than any other constituent. We follow Pollard (1996) and Uszkoreit (1984) in assuming an inversion feature [INV] on German verbs.49 If this feature is positive, the verb precedes all its complements instead of following them, much in the same way as interrogative inversion is accounted for. Any particle is unaﬀected by the position of the verb, and simply remains following all other arguments in clauses both [+INV] and [−INV]. There is one diﬀerence between this and the PPT account which is due to the framework being non-derivational. This version does not allow the incorporation of the particle into the verb as in (130) above. It would of course be possible in the embedded sentence (130a) where the two are adjacent, but the fact that verb is generated directly in its V2 position in the matrix sentence prevents it incorporating. Since the verbs are identical except for the [INV] feature, we assume that it will not happen in the embedded clause either.

49.Netter (1992) proposes something essentially similar, but with a hierarchical VP structure.

Antecedent Reactivation and Trace 129

4.6.2 Results Let us recall that in Experiment 1, although the RT of the identical target was faster at the trace position than at the control position, so too was the baseline set by the RTs to the unrelated targets, and so there was no evidence of additional priming of the antecedent at the trace position. Here the situation is diﬀerent: the baseline of the unrelated targets is slower at the trace position than at the control position, and so even equal RTs for the identical targets at the control and gap test positions would testify to some priming of the antecedent at the trace position. In fact the identical targets elicited faster reaction times at the trace position than at the control position. The graph in Figure 13 illustrates this. 720 Mean RT in ms

700 680

692

707

660 640 620

626

600 580

597 Unrelated targets

Identical targets

Testpoint Control Gap

Figure 13.Experiment 2 reaction times by testpoint and target type

The unrelateds were 15ms slower at the gap position and the identicals 29ms faster. We may thus say that the identical targets were 44ms faster at the trace position than we would have expected on the basis of the unrelated targets. The two target types thus show very diﬀerent behaviour at the two test positions, and this diﬀerence proved to be statistically signiﬁcant. Again both by subjects and by items analyses were performed. As before there was a signiﬁcant main eﬀect for Target type (F1 =101.09, p<0.01, F2 =89.97, Table 8.Experiment 2: Mean reaction times and priming eﬀects in milliseconds

Control testpoint (X) Gap testpoint (Y) Position eﬀect (X–Y) * Figures are signiﬁcant (p < 0.05)

Unrelated targets (A)

Identical targets (B)

Target type eﬀect (A–B)

692 707 −15

626 597 29*

66* 110* 44*

130 Empty Categories in Sentence Processing

p < 0.01). There was no signiﬁcant eﬀect for Testpoint (F1 = 2.40, p = 0.12, F2 = 0.24, p = 0.63). The eﬀect for Subject was signiﬁcant (F1 = 5.92, p < 0.001), as was the eﬀect for Item (F2 = 5.68, p < 0.001). However, unlike in Experiment 1, the crucial interaction of Target type and Testpoint was signiﬁcant on both analyses (F1 = 4.45, p = 0.035, F2 = 5.22, p = 0.023), showing that the target type diﬀerence was not producing the same eﬀects at the two testpoints. We investigated this interaction further by performing one-way analyses of variance on each of the target types separately. While there were no signiﬁcant positional diﬀerences for the unrelated targets (707ms vs. 692ms, F = 0.46, p = 0.50), the RTs for identical targets were signiﬁcantly faster at the gap position than at the control position (626ms vs. 597ms, F = 6.96, p < 0.01). This demonstrates that there is a reliable experimental eﬀect at the presumed trace position, as predicted by the TRA. 4.6.3 Discussion The important result here is the interaction of Testpoint and Target type revealing that there is additional priming of the antecedent at the gap position over and above that which would be expected from the baseline conditions. This result is as predicted by the TRA, since PPT predicts that there will a trace at our second testpoint. The other accounts of the trace reactivation data lack an explanation of this result, however. Recall that the DAH and the SPA predict reactivation of displaced arguments only at subcategorisers, and the DOP predicts it only at potential clause boundaries. Since trace position we used in this experiment coincides with neither of these, they are not supported. An additional implication of our data is that short scrambling is indeed a movement structure, not a base-generated one. A possible counterargument to this might be to argue that the reactivation of the DO could be due to the processing of the particle. Since the particle is at least lexically related to the verb, one might suggest that it would behave like a verb and cause reactivation of its arguments. We shall argue here that this account cannot provide an alternative account of the eﬀects shown in Experiment 2. The primary reason why the verbal particle cannot play a role in the reactivation of the direct object is because German VPs are head-ﬁnal and so the trace and testpoint actually precede the particle. Since the particle has not yet been processed at the testpoint, it cannot cause the eﬀects found there. There is experimental conﬁrmation of this: in other CMLP studies investigating empty categories (e.g. Swinney et al. 1989; Hickok et al. 1992) the position

Antecedent Reactivation and Trace

immediately preceding the subcategorizing verb has been used as a control position, and has provided no evidence of eﬀects dependent on the verb. We can thus be conﬁdent that the position immediately preceding the particle is similarly safe from potential association eﬀects. However, we shall also show that there is no motivated mechanism which could cause the particle to cause reactivation of the gap antecedent. There are two ways that this might be argued to occur. The ﬁrst would hold that the ﬁllerverb association cannot be made until the whole verb has been encountered (i.e. here until after the particle) because its argument structure cannot safely be identiﬁed any earlier. This is however contradicted by the plentiful evidence (e.g. Clifton & Frazier 1989) that dependencies are formed at the earliest opportunity, and that these are reanalysed if later information proves them to be erroneous. This can be illustrated by example sentences such as (133) where Swinney et al. (1989) found reactivation of doctor at the verb position #1 even though the very next word ruled this analysis out. (133) The boxer visited the doctor that the swimmer had advised #1 him to see.

In the light of this ﬁrst resort strategy it seems most unlikely that the parser would wait patiently for a sentence ﬁnal particle before making an association. A second way of motivating reactivation of the DO at the particle would be to argue that when the parse reaches the particle, this last is associated with its mother verb and that this operation triggers measurable reactivation of the verb and, indirectly, also its complements. This proposal of indirect reactivation requires us to look in rather more detail at the mechanism proposed within the DAH and its applicability to these circumstances. Pickering & Barry’s DAH suggested that it was the association of the displaced element with its subcategorizer, i.e. a grammatical process, that caused its mental reactivation. It seems natural that the DAH would predict a reactivation of the verb at the particle, since the particle is a syntactic dependent of the main verb stem in HPSG, and the association of these two is essentially parallel to that of an argument with its verb. The second step necessary to reactivate the DO at the particle position is however not predictable by the DAH. This step requires that the activation of the verb by the particle also indirectly causes the activation of all its arguments. Now we might not be surprised by some small eﬀect like that in purely physical terms when looking at activation levels in an associative network; we illustrate this in Figure 14.

131

132

Empty Categories in Sentence Processing

Figure 14.The activation pattern necessary for a verbal particle to prime an argument

In a network of nodes, the activation of a connection between A and V will also cause some increased level of activation of other nodes connected to V, namely B, C and D. It must be clear that this can form no part of the predictions of the DAH, however, since there is no grammatical need to process B when A is associated with V. Additionally it can be demonstrated that this indirect activation cannot possibly occur in our data. This generalised activation would suggest that all existing syntactic dependents would be reactivated every time the parser identiﬁed a new one. Against such a background of repeated reactivation, the localised eﬀects that we and others have found would be quite indistinguishable. It is clear that the indirect reactivation cannot account for the very speciﬁc eﬀects at the gap position which we have found. We conclude that there is no motivated or empirically adequate way that the reactivation we observed can be attributed to the particle. One alternative suggestion which has been made is that these materials cannot exclude verb-based eﬀects as the cause of the reactivation found since the testpoint is at or near a position where we might expect a verb trace to be located, on the assumption that the V2 word order is derived by movement. It is quite true that this cannot be excluded, but let us note that this account requires the verb trace to cause experimentally measurable reactivation of its antecedent, and thus to play a fully active part in sentence processing. Since the aim of the experiment is to test exactly this hypothesis, this alternative explanation cannot defeat the conclusion from these results that the TRA of antecedent reactivation is the most adequate, though it might provide an interesting alternative perspective on what may be behind the result. However, there is a further problem with this idea, namely that the testpoint is not in fact at the presumed verb trace position but precedes the particle which itself precedes the

Antecedent Reactivation and Trace

verb trace position. As we mentioned above, in previous work positions preceding verbs have been used as control testpoints, and no eﬀects have been found. This makes it unlikely that the reactivation we note is due to the verb trace. A more feasible alternative explanation of the eﬀects found in Experiment 2 concerns the proximity of the gap position to the clause boundary. While the gap is separated from the clause boundary by the particle, it might nevertheless be argued that the separation is insuﬃciently substantial to deﬁnitely exclude the eﬀects predicted by the DOP at this position. It was to control for this possibility more conclusively that we carried out Experiment 3.

4.7 Experiment 3: VP in SpecCP To exclude the possibility that priming at our trace position was due to DOP eﬀects at the clause boundary and not to trace reactivation, we tested a sentence type with the VP raised into sentence-initial position, commonly analysed as SpecCP (134). CP

(134) spec

C´ C

IP spec

I´ VP

spec

I V´

obj

V

This leaves the scrambled DO trace position distant from the end of the sentence and thus immune to the DOP account. In the example below the VP has been topicalised into the pre-verbal position in a V2 structure.

133

134 Empty Categories in Sentence Processing

(135) [VP [So ein Haus]i einer völlig normalen #1 Kundin ti #2 verkaufen] [such a house a quite normal customer sell konnte nur Maria could only Maria “Sell a house like that to a quite normal customer? Only Maria could (do it).”

Since this was a follow-up experiment aiming only to exclude an alternative explanation of the data, not establish the existence of a phenomenon, it was decided to test only identical targets. This removes the control condition which establishes the baseline of target recognition speed, but since the previous two experiments had shown no signiﬁcant diﬀerences for unrelated targets at the two testpoints, it was felt to be legitimate to streamline the procedure as a matter of simple practical economy. 4.7.1 Results and discussion The mean lexical decision times collapsed across items and subjects proved to be signiﬁcantly slower at the control testpoint than at the gap testpoint (619ms vs. 591ms, F = 4.25, p = 0.040) (see Figure 15). This is consistent with the suggestion that the antecedent is reactivated at the gap position.

Mean RT in ms

630 620 610

619

600 590 580

591 Control position

Gap position

Figure 15.Experiment 3 reaction times by testpoint

This third structure too shows evidence of antecedent priming. The aim of this supplementary experiment was to rule out any account of the eﬀects found in Experiment 2 in terms of sentential processing at a potential boundary of a saturated clausal structure. The data obtained is evidence against this DOP account. Let us however note two points: ﬁrstly, the lack of unrelated targets leaves open the possibility that the eﬀects observed were simply a result of

Antecedent Reactivation and Trace

background variations in processing complexity, even though no such eﬀect was found in previous experiments. Secondly, since the gap position is adjacent to the verb in this structure, it would be tempting to attribute these eﬀects to the subcategoriser, in line with the suggestions of the DAH and SPA accounts. As we have previously noted, however, this preverbal position has previously been successfully used as a control position in similar studies, and so the attribution of subcategoriser eﬀects to this position in our experiment would necessitate a reanalysis of these earlier studies too.

4.8 Generalized discussion The interpretation of these results as a body of evidence requires some care, as the results from Experiments 2 and 3 provided evidence of antecedent reactivation, while those of Experiment 1 did not. Faced with this contradiction we must seek to determine which data set is the more reliable, and which is more likely to have been aﬀected by external factors. We have already suggested in the discussion of Experiment 1 (see Section 4.5.2) that there are reasons to consider that set of results open to alternative interpretation; indeed it was for that reason that we repeated the basic experiment with revised materials designed to exclude the weaknesses we identiﬁed. We therefore propose to disregard the results of Experiment 1, and place greater weight on the data of Experiments 2 and 3 in the following. Another point which might cause some misgivings in the interpretation of these results is the lack of a control condition in Experiment 3. We should underline here that this was only intended as a supplementary experiment, aiming only to test whether the eﬀects noted would repeat themselves at a position clearly remote from a clause boundary. It is instructive to view the results from the unrelated targets in Experiments 1 and 2 in Table 9. Table 9.Collated results of Experiments 1–3 by Target type and Testpoint Targets

Unrelated targets

Identical targets

Testpoint

Control

Gap

Control

Gap

Experiment 1 Experiment 2 Experiment 3

724 692 –

707 707 –

622 626 619

612 597 591

Means

708

707

622

600

135

136 Empty Categories in Sentence Processing

The mean values of the unrelated targets are extremely even (708ms at the control testpoint and 707ms at the gap testpoint), and there were no signiﬁcant diﬀerences between the four ﬁgures. One-way anovas on each experiment showed no signiﬁcant diﬀerences by Testpoint (Exp1 F = 0.15, p = 0.70; Exp2 F = 0.46, p = 0.50). We also merged the data ﬁles of Experiment 1 and 2 and performed a factorial anova on the data of the unrelated targets only. Neither the factor of Experiment nor that of Testpoint approached signiﬁcance (Experiment: F = 1.03, p = 0.31; Testpoint: F = 0.04, p = 0.40). Since the data of the unrelated targets shows no diﬀerences by experiment or testpoint, and Experiment 3 was identical with Experiments 1 and 2 in everything except that the items had VPS in SpecCP, there can be little reason to speculate that such diﬀerences might have been found. We therefore feel justiﬁed in using the data from Experiment 3 as a source of facts from which to draw conclusions.

4.9 Conclusions The major aim of the experiments here was to try to distinguish between the four major accounts of the antecedent reactivation data. To this end we tested at a position where syntactic theories making use of ECs would standardly posit a trace, but where the other explanations of the data set would not predict any eﬀects. The experimental technique we used, cross-modal lexical priming, taps on-line sentence processing and is sensitive to the reactivation of structurally appropriate antecedents of anaphoric elements. A secondary aim was to determine whether a derivational or a base-generated analysis of short scrambling would be the more empirically adequate. We shall address this ﬁrst, as it can be more brieﬂy dealt with. The syntactic question is answered in favour of movement approaches to the analysis of short scrambling. Recall that analyses of this data may be divided into two groups, depending whether they make use of movement or not. The ﬁrst group (Webelhuth 1989; Saito 1989; Mahajan 1990; Fanselow 1990; Haider 1993, 1996; Deprez 1994; Poole 1996; Müller & Sternefeld 1994; Vikner 1997; Kitahara 1997; Grewendorf & Sabel 1998) assumes that there is a canonical order of arguments in this clause type, namely IO > DO. The scrambled DO > IO order is thus a derived order, and the relationship between the overt position of the DO and its canonical position is of the same type as other antecedent–anaphor relations. This school predicts that there will be a trace at

Antecedent Reactivation and Trace

the base of the dependency, and that this trace will behave similarly to other anaphors when subjected to psycholinguistic experimentation. The alternative view (e.g. Bayer & Kornﬁlt 1991; Kiss 1995; Fanselow 1997) holds that the order of arguments is not determined by strictly syntactic factors, and since both DO > IO and IO > DO occur in identical syntactic environments, then both must be generable, perhaps optionally. While the precise details of the licensing process diﬀer between authors, the common factor relevant here is that this analysis involves no movement and thus leaves no trace. Empirical data suggesting that the trace has a measurable processing reﬂex would thus tend to throw doubt on this group of scrambling analyses. We have obtained evidence of the reactivation of the antecedent at the foot of the dependency. Since this is consistent with the psychological reality of the syntactic construct of traces in this structure type, these must receive some support as a motivated part of the syntactic structure of scrambled sentences. Since only movement approaches to this construction utilize traces, they are favoured by this data. Let us note one speculative point: if our assumption holds that only traces at the foot of dependencies are visible to experimentation (see Section 2.4.1.5, 2.4.1.7 and Chomsky 1995: 300ﬀ) it would follow also that our results would contradict Kayne’s (1994) Linear Correspondence Axiom, since this would entail all objects being generated after verbs and we have robust evidence of object trace in a pre-verbal position. The primary aim of our studies was the question whether there is any psycholinguistic reﬂex of the accessing of antecedents at gap locations. The answer is generally positive. While the data showing antecedent reactivation was originally interpreted as trace reactivation, this explanation of the phenomenon has more recently come into question. In particular, three other suggestions have been made which have attempted to account for the data in ways which have made no reference to traces. The earliest alternative account was the DAH (Pickering & Barry 1991). This put forward the idea that the processing of nonlocal arguments is rather more similar to the processing of local arguments than had previously been assumed. Instead of positing the discrete storage of a ﬁller and its recovery at a gap position, this new account assumes that it is directly associated with its subcategoriser when this is identiﬁed. It is at this position, therefore, that the DAH predicts eﬀects. No additional processing is presumed at the gap position at all. The SPA is an account based on the idea that the antecedent reactivation data could be due to semantic and not syntactic processing. Employing the semantic features of HPSG to implement this, we can see that this approach too can reasonably predict antecedent reactivation

137

138

Empty Categories in Sentence Processing

eﬀects at subcategorisers and sentence ﬁnally, but not at gap positions. The fourth interpretation of the data, the DOP, has two assumptions as its foundation: that more complex structures occasion greater activation of their arguments, and that this background activation level is ampliﬁed at potential clause boundaries by additional sentential level processing. The result is to predict extra activation of nonlocal arguments at potential clause boundaries, but once again not at gap positions. In our experiments we sought to tease apart the diﬀerent predictions of these competing accounts, a complex task since their predictions reveal themselves as very parallel in their empirical realisations. By testing German, a language with a head-ﬁnal VP, we were better able to distinguish between the position of the subcategorising verb and that of the trace. By employing V2 sentences, we were able to examine a DO which is not directly adjacent to its verb. By using sentences with short scrambling, we were able to test a dependency which did not cross its subcategoriser. Finally, by utilizing particle verbs, we were able to prevent our gap position being at a clause boundary. These factors together enabled us to distinguish between the TRA and its competitors. Since the reaction times we observed showed a speeding of responses to the antecedent at the gap position relative to control conditions, it is likely that this represents the reactivation of the antecedent at the trace site. If this is the case, we have evidence that traces are empirically motivated as a psychologically real part of sentence processing, which must oﬀer support to principle-based parsers and analyses of dependencies which use trace. Since the existence of ECs is a non-obvious prediction, evidence that these have actually reﬂexes in on-line processing constitutes strong conﬁrmation of grammatical theories which assume them. Since such evidence would suggest that PPT is at least partially correct in its approach to the description of linguistic phenomena, this result must also lend support to the fundamental assumptions of generative grammar and weaken the position of psychological approaches which deny the reality of a mental grammar. Let us ﬁnish on a more modest note. We must be clear that our results in no way contradict the predictions of the DAH, the SPA and the DOP. These may hold absolutely; our results neither conﬁrm nor disconﬁrm this. Our results merely show that the extent of antecedent reactivation is not as these accounts would predict it to be, which leaves the TRA as the most empirically adequate account of the phenomenon. Even if evidence were to appear disconﬁrming the competitors, it is probable, even likely, that the TRA would need to undergo at very least some revision and clariﬁcation. Just one example of an outstanding

Antecedent Reactivation and Trace 139

question is the status of intermediate trace. In Chapter 7 we put forward the outline of an account in terms of syntactic processing compatible with both PPT and HPSG, but for any ﬁnal answer to be reached on these issues will require more work and more data. Our results here should help to move the debate one step further.

AUTHOR ""

TITLE "NP-trace and PRO"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 5

NP-trace and PRO: Local ECs

5.1 Introduction In the previous chapter we discussed the evidence for movement traces: this chapter addresses the processing of what is referred to in PPT as NP-trace and PRO, which we shall call here local ECs. They have in common that they are constructs used in strictly locally bounded dependencies and appear in constructions which are frequently analysed in rather parallel ways. There are clear diﬀerences between them too: while NP-trace has two main instantiations: in passives and in raising, PRO has only one: as the subject of the inﬁnitive complement clauses of equi verbs. The factor which leads us to treat them separately from the traces in the previous chapter is that they pattern together and diﬀerently from other trace (e.g. wh-trace) in experimental results. PPT and HPSG analyse raising and equi constructions in rather diﬀerent ways (see e.g. Chomsky 1981: 55ﬀ; Pollard & Sag 1994: 138ﬀ).50 PPT assumes that in raising structures the subject of an embedded inﬁnitive complement clause is moved to become the subject of the matrix clause, leaving behind an NP-trace (136). (136) [] seems Cheri to like Toni >>> Cheri seems ti to like Toni

This is necessary because inﬁnitives do not check nominative Case and possible as raising predicates do not assign a theta role to their subjects; the chain formed has just one theta role and checks its case just once. The introduction of PRO has something of the character of a last resort operation. All clauses

50.Because of the ambiguity of the term control in the context of a discussion of experiments, we shall prefer the word equi here, though we shall still use controller to refer to the antecedent of a null subject in an inﬁnitival clause. Note that since HPSG analyses as object raising certain structures which PPT treats as exceptional case marking, we consider only subject raising and subject equi here. In the following, “raising” and “equi” should be understood as “subject raising” and “subject equi” unless otherwise speciﬁed.

142 Empty Categories in Sentence Processing

require a subject, but the inﬁnitive complement clause permits the checking only of the null Case of PRO subjects, so this is inserted from the lexicon. Its reference is determined by its controller, the subject of the matrix clause (137). (137) Cheri tries Cheri to like Gordon >>> Cherii tries PROi to like Gordon

Both these types of embedded subjects are therefore present as covert syntactic constituents, but have diﬀerent characteristics. In HPSG, raising and equi verbs are lexically marked to subcategorize for unsaturated complement clauses, that is, clauses in which the subject is unexpressed. In raising structures the embedded subject is structure-shared with the matrix subject. This means that all SYNSEM features are copied. This allows the embedded subject to be identiﬁed and be syntactically active even though it is unexpressed. In equi structures the two are coindexed. (138) seem [SUBCAT · 1 NP, VP [inf, SUBCAT· 1 NPÒ] Ò] (139) try [SUBCAT ·NP 1 , VP [inf, SUBCAT·NP 1 Ò] Ò]

The diﬀerence between the two cases is the number of features which are shared. For equi it is just the INDEX value, for raising it is all the SYNSEM features. In addition, an equi verb assigns a semantic role to its subject, a raising verb does not. There are no empty categories in either raising or equi. We discuss the syntax of these constructions in greater detail in Sections 5.6.2 and 5.6.4 below. 5.1.1

The early CMLP evidence

During the debate about the Most Recent Filler Strategy (see Chapter 3 above) a great number of studies on empty categories were being carried out and great excitement generated by the results on wh-trace. Nicol & Osterhout (1988, reported in Nicol & Swinney 1989) sought evidence of the identiﬁcation of the controller of PRO using CMLP, following on the demonstration of antecedent activation at wh-trace gaps by Swinney et al. (1989). They tested at three points #1–3 in sentence pairs of the form of (140). They diﬀer in that (140a) contains an object equi verb and (140b) a subject equi verb. Put in terms of PRO, the one in (140a) is coindexed with the head of the RC the actress, while that in (140b) is coindexed with the subject of the RC, the dentist.

NP-trace and PRO 143

(140) There is the actressi that the dentistj from the new medical center … a. in town #1 had invited PROi to #2 go to #3 the party b. in town #1 had planned PROj to #2 go to #3 the party with ei

The Most Recent Filler Strategy would predict activation of dentist at #2 and #3 in both cases since it denies that control information is available on-line. If the antecedent reactivation found at wh-gaps were replicated for PRO, the results should show the real antecedent (i.e. actress in (140a) and dentist in (140b)) activated at #2 and #3. The results showed the following signiﬁcant activations: #1 nothing in either (140a) or (140b) #2 actress in both (140a) and (140b) #3 actress in both (140a) and (140b) #1 is a control testpoint and performs as expected. The eﬀects in the object control (140a) are in line with the real antecedent of PRO, but the eﬀects after the subject control verb are a surprise, and conﬁrm neither set of predictions. We have to assume with Nicol & Swinney (1989) that the eﬀects are due to the wh-trace of the head of the RC, and not the PRO (141). (141) There is the actressi that the dentistj from the new medical center … a. in town #1 had invited ei PROi to #2 go to #3 the party b. in town #1 had planned (ei?) PROj to #2 go to #3 the party with ei

In the ﬁrst example this is unremarkable, since invite is transitive, but this is only explicable in (141b) if plan is being construed as a transitive and a wh-trace is being posited after it too, in spite of the implausibility of a dentist planning an actress. As we have seen in Chapter 2, this ﬁnding that world knowledge is initially ignored has been robustly replicated in other studies. One last point from the results that we should note: there was a trend in each sentence type towards the activation of the antecedent of PRO at testpoints #2 and #3, but this trend did not reach signiﬁcance. We shall discuss this below. In a second attempt to test for the reactivation of the controller of PRO Osterhout & Nicol (1988, reported in Nicol & Swinney 1989) removed the RC from the materials to exclude the confounding wh-dependency. They contrasted sentence types with an active/passive alternation in the matrix verb as in (142). There were ﬁve testpoints; the ﬁrst testpoint was after to, the others were 500, 1000, 1200, and 1500 downstream.

144 Empty Categories in Sentence Processing

(142) a.

The actressj invited the dentisti from the new medical center PROi to #1 go to the pa#2rty at the #3 mayor’s #4 house #5 b. The actressj was invited by the dentisti from the new medical center PROj to #1 go to the pa#2rty at the #3 mayor’s #4 house #5

The results revealed the following signiﬁcant activations: #1 nothing in either (142a) or (142b) #2 nothing in either (142a) or (142b) #3 dentist in both (142a) and (142b) #4 the real antecedent of PRO in each #5 actress in both (142a) and (142b) Testpoints #1 and #2 are located where we should conﬁdently predict reactivation of the antecedent of the PRO if it, like wh-trace, behaved like an overt anaphor. There are no signiﬁcant eﬀects here, and so these results do not show PRO behaving like wh-trace. The #4 result is what we would predict on the basis of antecedent reactivation, but the eﬀects at #3 and #5 show that this is not the only factor. Nicol & Swinney (1989) speculate that multiple antecedents are activated right to left, i.e. backwards, but this is without precedent or conﬁrmation.51 Whatever the cause of the rogue results at #3 and #5, the result at #4 would suggest that there is some measurable reactivation of the real antecedent. Osterhout (1988, reported in Nicol & Swinney 1989) next adapted the experiment to examine NP-trace. He tested using CMLP as before but at two positions: at the presumed trace site and 1000ms downstream (143).

51.In fact this is one of the few published processing results that remains without reasonable explanation. We suspect that unintended lexical priming was involved. The priming of dentist at #3 may be due to the word party, and the priming of actress at #5 due to mayor’s house. This sounds improbable on the basis of this particular example sentence, but let us assume that the sentences were originally written in the active as in (142a), and the (142b) type sentences derived from them simply by passivizing the verb, a natural procedure. The sentence constructor would choose the two participants, say X and Y, and then, to give the sentences some coherence, choose a location or scenario in some way connected with each, say x and y, and then combine them to make the experimental sentence thus: the X invited the Y to go to the x at the y. For example: the doctor invited the pilot to go to the hospital #3 at the #4 airport #5. The unexplained eﬀects at #3 and #5 would be caused by the priming of doctor by hospital and pilot by airport. Since the passivization of the verb causes no change to the word order, the eﬀect is visible in the (142b) sentences as well: the doctor was invited by the pilot to go to the hospital #3 at the #4 airport #5. This approach to sentence construction would thus exactly account for the eﬀects found at #3 and #4. This is of course only speculation, but in the absence of any other account of this data, it seems probable.

NP-trace and PRO

(143) The dentisti from the new medical center in town was invited ei #1 by the actress to go to #2 the party

There was no signiﬁcant priming at the trace position but a non-signiﬁcant eﬀect 1000ms downstream. NP-trace was thus shown to pattern with PRO and not with wh-trace. Osterhout & Swinney (1993) conducted one more CMLP experiment on NP-trace. They tested once more at varying distances after passive verbs and used an active equivalent as a control. The test positions were at the trace position, 500ms later, and 1000ms later (144). (144) a.

The bakeri who had just moved into the neighbourhood was asked ei #1 by the woman #2 to help her #3 out at the party. b. The baker who had just moved into the neighbourhood asked #1 the woman #2 to help her #3 out at the party.

The results showed no eﬀects in the active, but in the passive nonsigniﬁcant priming at #1 (p = 0.1), nearly signiﬁcant priming at #2 (p = 0.08), and ﬁnally signiﬁcant eﬀects at #3 (p = 0.03). Let us review these results as a body of data. The methodology used in these experiments was exactly the same as that which was producing clear and immediate reactivation of the antecedent at wh-gaps, and yet for PRO and NP-trace only much weaker, later and more marginal eﬀects were being found. It does not seem likely that the experiments were in some way ﬂawed, because the results all show the same pattern of feeble and delayed reactivation. Two of the studies showed signiﬁcant eﬀects and the other two statistically marginal trends towards the reactivation of the appropriate antecedents from about one second downstream of the trace site. Since it is unlikely that parallel results over four experiments should be pure chance, we may conclude that this is a reliable experimental eﬀect and that it represents a real processing reﬂex of some feature of constructions containing PRO and NP-trace. It is probable that it represents reactivation by the local ECs posited here by PPT. This processing reﬂex is certainly very diﬀerent from that of wh-trace, but it is not null as has been claimed (e.g. Fodor 1993; Walenski 1998). 5.1.2 The evidence from probe recognition experiments Bever & McElree (1988; McElree & Bever 1989) applied a diﬀerent methodology to the problem. The sentences were presented using phrase-by-phrase selfpaced reading, and were followed by a visual probe recognition task. In sentences

145

146 Empty Categories in Sentence Processing

which contain an overt anaphor this probe is recognized faster than in sentences without. Thus subjects can answer that the word surly appeared in the preceding sentence more quickly in (145a) than in (145b). This, they hypothesized, is because a pronoun reactivates its antecedent and there is a recency eﬀect which speeds responses to more recently activated items. (145) The surly beadle at St Maud’s… a. said he stopped latecomers to matins. b. always stopped latecomers to matins.

SURLY fast SURLY slow

The aim of their experimentation was to ﬁnd out whether covert anaphoric elements such as traces would cause the same speeding of response to the probe. In several experiments, which we conﬂate here, Bever & McElree compared six diﬀerent structures, four experimental conditions and two control conditions (146). (146) a.

Raising The astute lawyer … tended e to cheat … b. Passive The astute lawyer … was easily cheated e … c. PRO The astute lawyer … decided e to cheat … d. Tough movement The astute lawyer … was easy to cheat e … e. Pronoun control The astute lawyer … said he could cheat … f. No priming control The astute lawyer … had cheated …

The ﬁrst (146a) is a raising structure, and is hypothesized by PPT to contain an NP-trace in the subject position of the embedded clause. The second (146b) is a passive and contains an NP-trace after the passivized verb where it was assigned its theta role. The third is an equi construction in which the embedded subject is postulated to be a PRO. (146d) is an example of tough movement: a structure suggested by Chomsky (1977) to contain a null operator. This entails that there is a wh-trace at the empty object position. There follow the two control conditions: the ﬁrst provides a benchmark for recent reactivation as it contains an pronoun referring to the subject. The last contains no anaphoric element, and thus provides a baseline for no recent reactivation.

NP-trace and PRO 147

In each case they tested at two positions: at an early testpoint immediately following the presumed location of the EC, and at a late testpoint at the end of the sentence. The probe was always the adjective on the subject. The results are simple: there were no signiﬁcant diﬀerences at the early testpoint, but the sentence-end probes did show diﬀerential responses. In the control sentence (146f ) with no reactivator overt or covert, subjects took longer to answer than in the other cases. Since this means that the sentence types with ECs patterned with the sentence type with a pronoun, Bever & McElree concluded that ECs reactivate their ﬁllers. Another researcher, MacDonald, working at about the same time conducted a somewhat similar visual probe recognition experiment (MacDonald 1989). She used sentence-by-sentence self-paced reading with a text-ﬁnal probe and comprehension question. Her contrasting structures were passives, adjectival passives and pure adjectives. The passive (147a) is often assumed to have an NP-trace in its object position, unlike an adjective, which has none; the adjectival passive (147b) should thus pattern with the pure adjective (147c). (147) The Irish terrorists wanted to disrupt the ceremonies… a. the new mayor at the centre podium was shot t b. the new mayor at the centre podium was surprised c. the new mayor at the centre podium was furious

MacDonald tested for both the “antecedent” (i.e. mayor) and the “agent” (i.e. terrorists). The important result was that the antecedent probes were responded to signiﬁcantly faster in the passive condition than in the pure adjective or verbal passive conditions but there was no such eﬀect for the agent probes. This would seem to conﬁrm Bever & McElree’s ﬁndings for NP-trace. These studies performed using a visual probe recognition methodology (also Bever, Straub, Shenkman, Kim, & Carrithers 1990) thus all produced results consistent with the psychological reality of local ECs. Unlike the CMLP results, in this methodology local ECs produced the same results as wh-trace. These results were initially taken as fairly solid evidence that reactivation eﬀects found at wh-gaps had some equivalent at local EC gaps. Below we shall review how the TRA can include this data and what other suggestions have been made.

5.2 The Trace Reactivation Account applied to local ECs The experiments we have reviewed above constitute the basic data on the reactivation by PRO and NP-trace which needs to be accounted for. Bever &

148 Empty Categories in Sentence Processing

McElree and MacDonald but also Nicol & Osterhout certainly had reactivation by ECs in mind when they conducted their experiments. However, as we have seen, the evidence for trace reactivation is much weaker for PRO and NP-trace than for wh-trace, and this is not predicted by PPT since it holds local ECs to be equally syntactically and psycholinguistically active.52 A simple form of the TRA will not cover the data. It must be added to in ways which are responses to the data as found, and which are somewhat stipulative. Essentially this will consist of saying that local ECs are diﬀerent from wh-trace in some way, and that this diﬀerence correlates with their diﬀering visibility to the experimental methods. There are (at least) three ways of doing this, two of which are discussed in some detail in Fodor (1993). The ﬁrst was originally raised in Nicol’s (1988) dissertation and consists of the suggestion that wh-trace is easier for the processor to locate as it has an overt ﬁller, which tells the processor to look for a gap. Local ECs do not have such an obvious preceding ﬂag, but are often unambiguously signalled only by information arriving after the gap. It is thus imaginable that these ECs are interpreted only after the gap, causing the delayed reactivation attested. However, Fodor (1993) reports an ingenious CMLP experiment which demonstrates that this cannot be the case. It tests for the interpretation of local ECs at second hand by checking whether a pronoun or Anaphor dependent on NP-trace for its interpretation reactivates its content quickly. In (148a), the reﬂexive himself requires a clausemate antecedent. The only clause-mate antecedent is the subject of the clause, an NP-trace, so if the reﬂexive causes the reactivation of its correct antecedent doctor, it follows that the antecedent of the NP-trace too was immediately and accurately identiﬁed. The same holds for the pronoun him in (148b), except that the correct antecedent is boxer. (148) a.

The boxerj knew that the doctori for the team was sure [ei to blame himselfi # for the injury] b. The boxerj knew that the doctori for the team was sure [ei to blame himj # for the injury].

The reﬂexive himself and the pronoun him each primed just their syntactically appropriate antecedent but not the other’s or an unrelated target. This demonstrates that the NP-trace had already been interpreted because the reference of

52.There is of course an inaccuracy in talking about “trace reactivation” in the case of PRO, since this is not a movement trace, but the term is convenient and we shall permit ourselves this inconsistency here.

NP-trace and PRO 149

himself and him depends upon it. This result cannot entirely exclude slowness of interpretation as the cause of the lack NP-trace reactivation eﬀects, because the test is at the pronoun or Anaphor three words downstream, but it certainly puts ﬁrm bounds upon it. The second possible amendment to the TRA to encompass the CMLP data consists of the suggestion that such eﬀects occur only with ECs which are visible at PF. Perhaps only casemarked traces or alternatively only accusative traces are visible at PF, and thus to CMLP. Since local ECs occur precisely in caseless positions, this would correspond to the data on reactivation. This possibility is addressed by Fodor (1993) in some detail, but without coming to any clear conclusion because the number of uncontrolled syntactic variables is too large to pin any predictions down. Sag & Fodor (1995) look at the best-known reﬂex of this posited diﬀerence, wanna-contraction, and ﬁnd it unconvincing. We shall just mention one further suggestion as to what may be behind the experimental diﬀerence. Bever et al. (1990) suggest that the reason that diﬀerent sorts of traces produce diﬀerent behavioural results is because of their diﬀerent interaction with thematic roles: “wh-gaps” are immediately assigned a thematic role, “raising gaps” necessitate an existing thematic role being changed, and the relationship between PRO and its antecedent is just semantic and not structural. We doubt that diﬀerences between wh-trace and local ECs in terms of PF-visibility, case-marking or thematic assignment can oﬀer much comfort to the TRA, unless some fairly transparent reason can be shown why such features might correlate to a processing diﬀerence. Nevertheless the syntactic distinctions between these classes of ECs (see Chapter 2) are probably the best motivator for their diﬀerent behaviour in CMLP that is available. The argument from the TRA will run: there are a number of syntactic distinctions between these EC types, and evidently one of these factors inﬂuences their visibility to CMLP, but further work would be required to identify which. This is not a very strong position, but it has the advantage of introducing no unmotivated assertions. The weak point in this defence is that on the crucial criterion here, the quality of being an anaphoric element, these ECs do not diﬀer.

5.3 Alternative accounts of the data We have discussed in Chapter 4 above a number of alternative explanations of why there is reactivation at the foot of unbounded dependencies. Some of these apply naturally to the data on local ECs too. However, we shall not repeat

150 Empty Categories in Sentence Processing

arguments presented in Chapter 4 above, but comment only on areas of diﬀerence here. We ﬁrst outline some important considerations bearing on the quality of the data from the two methodologies. 5.3.1 Methodological criticisms of the visual probe recognition data Fodor (1989) makes two very telling points which diﬀerentiate the data from the two methodologies. The ﬁrst is that the time course of antecedent facilitation in the visual probe recognition experiments is very diﬀerent to that in work on wh-trace (e.g. Swinney et al. 1989; Nicol 1988). The usual ﬁnding is immediate activation on encountering the anaphor, with the eﬀects fading over about a second (see Swinney et al. 1979; Swinney et al. 1989), while Bever & McElree’s early testpoint showed nothing, and only their late testpoint gave the signiﬁcant results. This is particularly telling for Bever & McElree’s wh-trace and pronoun conditions, which performed very similarly to the other ECs that Bever & McElree tested, and not at all like the results for wh-trace reported from CMLP (Swinney et al. 1989; Nicol 1988). This clear diﬀerence in the time course of facilitation by wh-trace in the two methodologies makes it fairly plain that their results are not measuring the same thing. Since the response to wh-gaps in CMLP is identical to that of overt anaphors and therefore very credibly syntactic in nature, it follows that we must characterize the visual probe recognition eﬀects as something other than syntactic. This greatly weakens the claim of the visual probe recognition data to be evidence for a syntactic construct. Fodor’s second argument concerns the subjects’ task. In CMLP the task is a lexical decision on an associate of the antecedent. Eﬀects are found which reﬂect the speciﬁc moment of processing when the target is presented: it is an on-line task. There is no conscious connection between the task and the input sentence, and since the overt task is lexical decision, it involves only lexical access. RTs are typically around 600ms. In visual probe recognition the task is to say whether the probe appeared in the preceding sentence, which involves some conscious matching of the probe and the preceding sentence. This requires ﬁrst lexical access and then review of the whole sentence content. It is not speciﬁc to any moment of processing but rather is aﬀected by anything and everything that is part of the sentence representation. It is also much slower, with response speeds typically around 1000ms. Plainly this task can oﬀer only much weaker evidence for any hypothesis about sentence structure as it oﬀers no clue about the origin of diﬀerences observed. Visual probe recognition data cannot provide convincing evidence that local ECs exist as syntactic entities unless the alternative accounts of the phenomena we discuss below can be excluded.

NP-trace and PRO

5.3.2 The Direct Association Hypothesis Recall that the DAH argues that reactivation attributed to eﬀects at gap positions is in reality due to the association of arguments with their subcategoriser. The CMLP data on local ECs can be argued to be consistent with the DAH. In none of the four studies were any eﬀects of the embedded subject found directly at the EC position, but only later. This pattern of late reactivation is directly predicted by the DAH, since these later testpoints follow the verb whose subject the local EC was, and where the DAH would predict its reactivation. To see how this might work, we repeat here the materials of Osterhout & Swinney (1993) as (149). (149) The bakeri who had just moved into the neighbourhood was asked ei #1 by the woman #2 to help her #3 out at the party.

Since the results showed no signiﬁcant priming at #1 or #2 but signiﬁcant eﬀects at #3, it could be argued that this is due to the embedded verb help. As the third testpoint immediately follows this the DAH can attribute eﬀects found here to the association of the non-local argument with this verb, whose logical subject it is. Note that the materials were not designed to control for the relative position of the verb and the later testpoints and so we must assume that there was some degree of variation. In the light of this it is not surprising that the eﬀects found were only weakly signiﬁcant. The DAH does not predict the results found in the visual probe recognition studies as the eﬀects found were not adjacent to the subcategorising verbs, nor does it oﬀer any account of the speeded responses when the subject was reprimed by an anaphor. However, while the DAH can only account for one part of the data, it is important to remember that there is no reason to assume that only one of the traceless accounts may hold. We shall suggest below that the DAH should be taken together with the SPA. 5.3.3 The Semantic Processing Account While the criticisms of the visual probe recognition data as evidence for ECs are valid, it is necessary to put another account of the results forward in order to wholly discount it. The most commonly advanced alternative is the SPA. We have seen in Chapter 4 above that there are a number of problems with its application to the wh-trace data. It cannot, admittedly, be entirely ruled out as an eﬀect in these experiments, but this is partly because it has not been formulated by its

151

152

Empty Categories in Sentence Processing

protagonists in any testable form. Furthermore, the data on wh-gaps seems to bear little resemblance to the lexical priming which is uncontroversially semantic in nature, but does exactly match the clearly syntactic eﬀects at overt anaphors. This double mismatch must cast some doubt on the semantic account of the wh-gap data. The SPA is a far stronger contender as an account of local EC data, however, for the pattern of activation in these experiments corresponds far more closely to the long drawn-out eﬀects of purely semantic priming. It is also concentrated at the end of sentences which is the canonical location of semantic processing. Some evidence for a semantic origin of the eﬀects comes from the nature of the methodology. Fodor (1989, 1993, 1995) has regularly argued that the visual probe recognition methodology is more likely to be sensitive only to a semantic representation of the sentence, unlike the cross-modal priming task which seems more credibly to access on-line syntactic processing. The main reason to think this is the sentence ﬁnal position of the subject’s task. A ﬁnished sentence is rapidly converted into a semantic form, so it is likely that it is this semantic representation which is scanned. It seems likely that this semantic processing of the target sentence has already occurred when the probe recognition task is performed and that it is a semantic representation of the sentence which is consulted. There remains one diﬀerence between the visual probe recognition data and the time course of semantic priming. Recall from Chapter 4 that Dell et al. (1983) examined the semantic priming of antecedents by superordinate lexical items (150). (150) A burglar surveyed the garage set back from the street. Several milk bottles were piled at the curb. The banker and her husband were on vacation. (The #1 criminal)/(A #1 cat) #2 slipped #3 away #4 from the #5 streetlamp. Antecedent: BURGLAR Clausemate: GARAGE

The appropriate antecedent (burglar) and its clausemate (garage) were immediately primed after the subjects encountered the related stimulus (criminal), i.e. already at #2. This was not the case in the visual probe recognition experiments, as they found no eﬀects at a similar early testpoint, but only later. This then is a diﬀerence, but perhaps not too serious a problem for the SPA, as one might always argue that the semantic processing of an EC is a little delayed compared

NP-trace and PRO

to an overt lexical item. It is not strictly necessary for the SPA to suggest that the antecedent of a hypernym is primed identically to an EC. The SPA is thus a realistic alternative to the TRA as an account of the eﬀects found in the visual probe recognition tasks. It can reasonably be the default assumption that an eﬀect found only at the end of a sentence is semantic in nature, since there can be little doubt that this is a location for semantic processing: slowed reading speeds in self-paced reading at sentence ends provide robust evidence of this (e.g. Pickering et al. 1994; Tanenhaus et al. 1989, 1993). Whether the SPA can equally well account for the data from the CMLP experiments is a rather more diﬃcult question. Recall that marginally signiﬁcant trends were found after the trace but before the clause boundary in all four CMLP experiments we reviewed. It is certainly quite possible to attribute these to semantic processing: any parsing model which allows some degree of incremental interpretation must accept that some semantic processing occurs during the sentence itself. This processing could be the vehicle for the weak eﬀects observed in the CMLP experiments. Nevertheless, the semantic account loses some of its persuasiveness here because the CMLP methodology is known to tap into syntactic eﬀects and because of the position of the testpoints closer to the presumed EC position and before the end of the sentence. Recently Bever & Sanz (1997) have carried out a new assault on this semantic account of NP-trace eﬀects. They made no change to the controversial visual probe recognition methodology, but used phrase-by-phrase self-paced reading and a sentence end probe which was as before the attributive adjective on the subject. The novelty was that they tested Spanish unaccusatives, which are in PPT assumed to have a post-verbal subject trace, in sentence patterns like (151b), and compared them to simple intransitives (151a), which have none. (151) a.

El severo critico/ que visitaba/ el museo/ habló con cuidado. SEVERO the strict critic who visited the museum spoke with care b. El severo criticoi/ que visitaba/ el museo/ llegó ti con cuidado SEVERO the strict critici who visited the museum arrived ti with care

In order to defuse the semantic argument, they divided subjects on the basis of their results into two groups: “sequence sensitive” and “sequence insensitive”. The ﬁrst group were the ones whose RTs to a sentence were proportional to its length, and who are thus thought of as reviewing the syntactic form of the sentence. The other group are those whose RTs did not vary in proportion to

153

154

Empty Categories in Sentence Processing

the length of the sentence, and who are therefore assumed to be scanning a semantic representation. The sequence-sensitives answered more quickly to the unaccusatives than the unergatives, and the reverse is true of the sequenceinsensitives. Bever & Sanz’s interpretation of this is that a post-verb subject trace was priming the probe in the unaccusative cases and that this was causing measurable eﬀects in the subjects who used a syntactic not semantic strategy. This controversial study raises many questions. We shall make just two points: ﬁrst it shares the weakness of the earlier visual probe recognition studies in failing to provide any evidence that the locus of the processing diﬀerence is at the presumed trace site. Secondly, the division of subjects into sequencesensitive and sequence-insensitive types on the basis of their results seems methodologically daring.53 It also introduces a new variable into the data which is little understood and requires further investigation. While this study is interesting, it cannot defeat the SPA without further work. Walenski & Fodor (1995) approach the issue from the opposite perspective. They used CMLP with a naming task to compare diﬀerent reﬂexive types and implied semantic arguments. (152a) contains an overt syntactic reﬂexive, (152b) is a passive which is argued to have an NP-trace, (152c) is a transitive control, and (152d) is a semantic but not syntactic reﬂexive containing an implicit argument. There were two testpoints, an early control common to all and a late one immediately following the DO position. (152) Several clever monkeysi that had been #1 caged at the zoo… a. Overt reﬂexive have squeezed themselves #2 through the bars b. Passive were squeezed ti easily #2 through the bars c. Transitive control have squeezed the bananas #2 through the bars d. Semantic reﬂexive have squeezed easily #2 through the bars … after the zookeeper went home.

The results showed only the overt reﬂexive (152a) as signiﬁcantly diﬀerent from the transitive control (152c), from which they draw two conclusions. Since the implicit argument was not accessed, they argue that the task does not reﬂect the

53.We note however that such a procedure is not without precedent; Friederici & Mecklinger (1996: 169) did essentially the same thing.

NP-trace and PRO

semantics, just the syntax.54 Since there was no reﬂex of NP-trace in the passive, they conclude that it is not psychologically real as a syntactic constituent. While this appears solid evidence against NP-trace, one might make two points to counter it. The ﬁrst is that they tested at the wrong place to ﬁnd evidence of NP-trace, since the eﬀects previously found from CMLP (see Osterhout & Nicol 1988; Osterhout & Swinney 1992 etc above) occur after some time lag, about one second. It is thus quite possible that some of the conditions produced reactivation eﬀects, but downstream of where they were tested for. The second objection is that the control test position is in a place that one might reasonably expect to show some degree of priming, namely inside a RC on the subject. It is possible that the NP-trace and perhaps the implicit reﬂexive did indeed show reactivation, but that this was not robust enough to show up when compared to a control position at which there was also some degree of priming (cf. Nicol 1993 exp2 in Section 4.3.2.2 above). This would be supported if the data showed that the transitive control sentence was somewhat slower relative to the others, thus giving the order of RTs: reﬂexive < passive
54.This was anyway argued by Boland (1991), who showed that the lexical decision version of CMLP showed both semantic and syntactic eﬀects while the version with the naming task was more speciﬁcally syntactic.

155

156 Empty Categories in Sentence Processing

5.3.4 The Depth of Processing Account Let us ﬁrst brieﬂy review the DOP (Fodor 1995, p.c.) (for detail see Chapter 4 above). It comes in three parts. The ﬁrst suggests that eﬀects attributed to trace reactivation are in fact caused by the greater processing complexity of the structures they occur in. Since any dependency causes complexity, the class of structures in which PPT posits traces are precisely the more complex structures, and therefore the constituents of these structures are better recalled. This would predict the observed diﬀerences between experimental and control structures. The next part applies to arguments in these structures. It will be exactly those arguments whose surface position and semantic interpretation are separate which will be most intensely processed and thus best recalled. This predicts the diﬀerences between target items and control items from within the same sentence. Lastly, this additional salience will be most obvious at potential clause boundaries. But gap positions too are usually potential clause boundaries, and so this ﬁnal part predicts eﬀects at the precise positions of traces. The DOP account ﬁts very neatly with the both the visual probe recognition and the CMLP data on local ECs. Parts one and two, taken together, provide a very adequate account of the weak eﬀects after the trace position in the CMLP experiments. The eﬀects were weak, sluggish and with poor deﬁnition, slow to appear and slow to fade, all of which corresponds nicely with what the DOP account based on greater cognitive saliency might be thought to predict. The rather more robust end of sentence results that Bever & McElree and MacDonald found are no surprise to the DOP, since these are the locations where most processing is found. A possible objection to the DOP as the unique explanation of all the data might be to suggest that there is an inconsistency between the ﬁrst resort assumption of clause boundary processing which was necessary to motivate wh-gap position eﬀects in Chapter 4 and the true end of sentence position where Bever & McElree obtained their data. Recall that the account of the eﬀects precisely at trace positions relies crucially on clauses being processed at these very speciﬁc points in time. One might counter this objection by arguing that there are two clause boundary deﬁnitions relevant here, the ﬁrst is at the location of the last strictly subcategorized argument, and the second is the end of sentence including adjuncts, and that each has its processing tasks. This is feasible, but does not entirely defeat the criticism that the DOP requires a menu of clause boundary deﬁnitions from which it can choose the one which suits the data. A second question one might raise is whether the DOP actually predicts the

NP-trace and PRO

local EC data as it stands. When we look only at the parts of the DOP required to account for the local EC data, the DOP seems a rather more attractive option than the reactivation of local ECs as a mechanism for the results found. But when we add in the third part, the extra clause boundary processing the DOP requires to account for the wh-trace data and apply it here, the situation changes. For the DOP predicts strong CMLP-visible eﬀects at potential clause boundaries. Since in English the upstairs raising or equi verb and the downstairs verb are almost adjacent to each other, there should be little diﬃculty about identifying the location of predicted eﬀects. In a simple sentence such as (153) below, the clause boundaries are fairly readily identiﬁable, though again the question rather hinges on the deﬁnition of the clause boundary. (153) [Gaspard wanted [to swim]]

There are two options: either there is a clause boundary between wanted and to, or else, if we permit the DOP to make use of a proposition-based clause deﬁnition, then there is none. In either case, we should expect some DOP eﬀects at swim, since the local argument is not overtly present. However, as the CMLP data from Nicol (1988), Nicol & Osterhout (1988), and Osterhout & Nicol (1988, all three reported in Nicol & Swinney 1989) makes clear, there is no such early eﬀect at the potential clause boundary. Thus the mechanism appealed to in the case of the wh-dependency data is not producing detectable eﬀects here. Now this result is not fatal to the DOP: it can argue that the nature of the processing complexity here is diﬀerent to that of unbounded dependencies, the displaced argument is after all, although absent, locally determined. But though these arguments may go through, they are precisely the same arguments that we suggested best permit the TRA to explain the diﬀerence in reactivation between local ECs and wh-trace: essentially the processing diﬀerence is attributed to syntactic diﬀerences between unbounded dependencies and local EC dependencies. So the DOP does not oﬀer any explanatory advantage over the TRA when faced with the full range of data sources, and the diﬃculty of accounting for the diﬀering results. In addition, the TRA is based upon the syntax which requires ECs at the speciﬁed locations, whereas the DOP is merely a response to the evidence. The TRA is thus the theoretically stronger and more interesting position, and has the same degree of empirical coverage as the DOP, since both have to make reference to syntactic diﬀerences to account for a diﬀerence in reactivation which they do not predict.

157

158

Empty Categories in Sentence Processing

5.3.5 Summary The outcome of the investigation into ECs in equi and raising constructions has thus produced two bodies of data, that from visual probe recognition experiments and that from the CMLP methodology. There are three explanations of this data set. The TRA is consistent with both the visual probe recognition and CMLP data but does not predict the clear diﬀerences in response to wh-trace and local ECs. The SPA immediately accounts for the visual probe recognition data, but is only weakly compatible with the CMLP data, while just the opposite holds for the DAH. On this basis the two together form a reasonable account of the data. The DOP readily accounts for both the visual probe recognition and CMLP data but fails to ﬁnd its additional predictions of strong reactivation at potential clause boundaries conﬁrmed. Clearly there is a need for further studies to distinguish between these confounded accounts.

5.4 Experiment 4: Local ECs and end-of-sentence probes In this experiment we sought to do two things. First we wished to extend the range of data which bears on the psychological reality of local ECs. In order to broaden the data base we performed our experiment in German not English, since all previous work on the subject has used English. As the relevant ECs are thought to be part of Universal Grammar, they should function in the same way in the two languages. On the other hand, word order diﬀerences between English and German will make it easier to distinguish between eﬀects due to the processing of ECs and eﬀects due simply to local complexity. In addition we attempted to control for some of the factors which cause the accounts of the data we reviewed above to be currently indistinguishable. To this end we performed a quasi-replication: we adopted the methodology and experimental design of the classic visual probe recognition studies of Bever & McElree (1988, 1989), but combined several of their experiments into one. Additionally we introduced a new condition designed to distinguish clearly between the subcategoriser-based DAH/SPA and the trace-based TRA. The reason for using the rather controversial visual probe recognition methodology was that it provided greater scope for clariﬁcation of previous results. At present the visual probe recognition data cannot contribute eﬀective evidence for the reality of ECs because there are real doubts about the interpretation of the existing results; in particular, many linguists are not convinced that the

NP-trace and PRO 159

eﬀects found at the end of the sentence are really due to the recent activation of the subject. If some of these doubts could be discounted, the visual probe recognition data might be held to be more persuasive and provide real evidence for local ECs. Recall that the basic premise of the visual probe recognition methodology is that one can detect whether an argument has been recently reactivated by measuring the time taken to respond to it in an end-of-sentence probe. Speeded responses after sentences containing ECs are attributed to reactivation by the EC, in the same way as speeded responses after sentences containing overt anaphors are due to reactivation by them. We tested six sentence types examined by Bever & McElree which have similar PPT structures in English and in German, abstracting from surface phenomena such as word order. The ﬁrst (154a) is raising, which has an NP-trace as the subject of the embedded clause. The second (154b) is a passive, which has an NP-trace at the complement position of the verb. The third (154c) is an equi construction in which PRO is thought to appear as the embedded subject. The remaining three conditions were all designed to provide comparison with other data of diﬀerent sorts. The fourth (154d) was a tough movement construction, argued by Chomsky (1977) to involve the raising of a null operator from the empty complement position in the embedded clause, leaving behind a wh-trace. This sentence type should permit us to compare the robustly found wh-trace eﬀects with those of local ECs. The last two are true control conditions. The ﬁfth (154e) was a sentence containing a pronoun at the same position as the ECs in the other sentence types. Since it is uncontroversial that a pronoun primes its antecedent, this provides a primed control condition. The last (154e) is an unprimed control. In this sentence there is no constituent overt or covert that might be thought to reactivate the subject before the end of the clause. (154) a.

NP-trace — raising Der Anwalt schien t den Fall in Karlsruhe leicht zu gewinnen the lawyer seemed the case in Karlsruhe easily to win b. NP-trace — passive Der Anwalt wurde leicht t gewonnen für den Fall in Karlsruhe the lawyer was easily won for the case in Karlsruhe c. PRO Der Anwalt hoﬀte, PRO den Fall in Karlsruhe leicht zu gewinnen the lawyer hoped the case in Karlsruhe easily to win

160 Empty Categories in Sentence Processing

d. wh-trace — tough movement Der Anwalt war leicht t zu gewinnen für den Fall in Karlsruhe the lawyer was easy to win for the case in Karlsruhe e. Pronoun primer Der Anwalt sagte, daß er den Fall in Karlsruhe leicht gewinnt the lawyer said that he the case in Karlsruhe easily will.win f. Control — no primer Der Anwalt wird den Fall in Karlsruhe sehr leicht gewinnen the lawyer will the case in Karlsruhe very easily win

If our German materials produce similar eﬀects to those of Bever & McElree’s experiments on English, then this would support the claim that the eﬀects were due to ECs. If they produce markedly diﬀerent results, this would tend to favour an alternative explanation such as the DOP, which attributes the eﬀects to greater processing complexity in sentences with extracted elements. Superimposed across the last two sentence types, the control conditions without ECs, we introduced two subconditions. These two sentence types existed in two diﬀerent forms: verb ﬁnal (155a,b) and V2 (155a¢,b¢). (155) Pronoun a. Der Anwalt sagte, daß er den Fall in Karlsruhe leicht GEWINNT the lawyer said that he the case in Karlsruhe easily WIN a¢. Der Anwalt sagte, er GEWINNT den Fall in Karlsruhe leicht the lawyer said he WIN the case in Karlsruhe easily Control b. Der Anwalt wird den Fall in Karlsruhe sehr leicht GEWINNEN the lawyer will the case in Karlsruhe very easily WIN b¢. Der Anwalt GEWINNT den Fall in Karlsruhe sehr leicht the lawyer WIN the case in Karlsruhe very easily

The aim of this is to control for a DAH account of our results, according to which eﬀects would be due to subcategorised arguments being associated with the verb. If the speeded responses that Bever & McElree found were due to activation at the verb, then responses should be quicker in those cases where the verb immediately precedes the testpoint. If no diﬀerence is evident, this would constitute fairly ﬁrm evidence against an account based on processing at the verb. The present study therefore replicates Bever & McElree’s study on another language with the hope of being better able to distinguish between real antecedent reactivation and alternative accounts. We also introduce a new condition of

NP-trace and PRO

verb position which should enable us to verify whether or not any eﬀects are due to the association of arguments with their subcategorising verb. Additionally, since the presentation method of the sentences is self-paced reading, we shall have reading speed data with which to test whether there is evidence of slower reading speeds indicative of greater complexity. Recall that in the studies on which the DOP is based (Cairns & Blank 1976; Cairns, Cowart & Jablon 1981) the speeded recognition correlated with slowed reading at the location of the extra complexity. If we should also ﬁnd longer reading times correlating with shorter probe recognition times, this would be consistent with a DOP account of the visual probe recognition results (note that Bever & McElree’s set-up did not record reading times). 5.4.1 Procedure Subjects were tested using a probe recognition task following a text presented to them in a phrase-by-phrase self-paced reading format on a 17” computer screen. When the subject pushes a button, the previous text section is removed and the subsequent one appears. After the last part of the short text has been presented, the probe word appears on the screen, and subjects press one of two buttons to answer whether that word was in the preceding text or not. Subjects are instructed to respond as quickly as possible, subject to maintaining a high rate of accuracy. The technique utilises the recency eﬀect: the more recent reactivation of the probe word enables a faster response. If there are ECs which have a real processing reﬂex, this should be visible in probe recognition times. In our study the probe was always a predicative adjective on the antecedent of the gap, the subject of the matrix verb. These were matched for length (8–10 letters, 2–3 syllables) and frequency (CELEX lemma frequency 1><20 per million) to ensure that these factors did not cause them to be recognized at diﬀerent rates. 60 experimental sentences were constructed in the six conditions detailed above. Each form was presented to one of the six subgroups of subjects, so that every sentence was seen in each form, all subjects saw 10 sentences in each condition and no subject saw the same sentence in more than one form. The subconditions in the control sentence forms were similarly distributed. All experimental sentences had essentially the same format: a three-word subject NP with an attributive adjective, followed by a six-word relative clause dependent on the subject to remove it from short-term memory. This was followed the matrix verb and a ﬁve or six word ending. This structure provides suﬃcient

161

162 Empty Categories in Sentence Processing

distance between the original position of the probe word in the sentence and the trace for there to be a sizeable recency eﬀect. All sentences were divided into portions of about three to six words for presentation as phrase-by-phrase selfpaced reading, using as far as possible intuitive phrase boundaries as breaks. The penultimate section of an experimental item always included unambiguous evidence that the appropriate EC should be assumed. In the example below the / indicates the position of the break between the ﬁnal two sections.55 (156) Der geschickte Anwalt, der mit dem ﬂeißigen Richter traf, … the skilled lawyer who with the hard-working judge met “The skilled lawyer, who met the hard-working judge …” a. raising schien t den Fall/ in Karlsruhe leicht zu gewinnen seemed the case in Karlsruhe easily to win “seemed to win the case in Karlsruhe easily” b. passive wurde leicht t gewonnen/ für den Fall in Karlsruhe was easily won for the case in Karlsruhe “was easily won for the case in Karlsruhe” c. PRO hoﬀte, PRO den Fall/ in Karlsruhe leicht zu gewinnen hoped the case in Karlsruhe easily to win “hoped to win the case in Karlsruhe easily” d. tough movement war leicht t zu gewinnen/ für den Fall in Karlsruhe was easy to win for the case in Karlsruhe “was easy to win for the case in Karlsruhe” e. pronoun sagte, daß er den Fall/ in Karlsruhe leicht gewinnt said that he the case in Karlsruhe easily win “said he will win the case in Karlsruhe easily” f. control wird den Fall/ in Karlsruhe sehr leicht gewinnen will the case in Karlsruhe very easily win “will win the case in Karlsruhe very easily”

55.In fact the structure of the materials is rather more complex than is implied here. In order to minimize the lexical diﬀerences between items, not only the ﬁrst halves but also the second halves of the sentences were identical between subject groups, but these were attached to diﬀerent ﬁrst halves. See Appendix 2 for further details.

NP-trace and PRO 163

The 60 experimental items were mixed into 180 ﬁller sentences of varying length and structure. The ﬁller probes were verbs, nouns or a few adjectives from other locations in the sentence. In two thirds of the ﬁllers the probe was a word from the preceding sentence, so that the overall ratio of correct yes/noanswers to the probes was 75:25. The experiment was carried out at the University of Essex on 25 native German speakers, who were paid to take part. Including practice exercises, the process lasted about 45 minutes and was carried out using the NESU experimental set-up (Baumann et al. 1993). Subjects read written instructions, asked any questions, and were then given the chance to practice. During and after practice subjects were invited to ask any further questions. Candidates were asked about the strategies they used after completing the experiment. The data from this experiment was processed as follows. First the data of one subject was discarded because he failed to follow instructions. Next all RTs of erroneous probe recognition responses were eliminated, and lastly all scores exceeding two and a half times the standard deviation by subject were excluded from further analysis. For input into the anova calculations the reaction time data was converted into logarithms in order to reduce its skewedness, but as before we refer to the unconverted ﬁgures within the text. 5.4.2 Results There are three parts to the results: the reading time data, the probe recognition data analyzed by sentence type, and the probe recognition data analyzed by verb position. The line graph below (Figure 16) shows the reading times of the diﬀerent sentence types. The ﬁrst position is the second half of the relative clause on the subject, which is identical in all sentence types. The next is the location of the matrix verb and provides disambiguation of what sort of EC or overt anaphor is to be assumed; it therefore provides interesting clues as to the relative complexity of the sentence types. The third position is the remainder of the sentence and the fourth the probe recognition time, included here for comparison. The reading times at the RC and at the probe are very close. However there are interesting diﬀerences by sentence type at the other two positions. At the matrix verb the raising and control sentence types are signiﬁcantly slower than the others, which show little change from the RC position. This was shown by a signiﬁcant eﬀect for Sentence Type (F = 13.02, p < 0.0001) at this position. Care must be taken in attributing this slowing to any particular factor, since the

164 Empty Categories in Sentence Processing

Figure 16.Experiment 4 reading times by condition in milliseconds per segment

words presented at this position varied quite considerably between sentence types. Nevertheless, it is interesting to note that a raising string such as schien den Fall takes longer to read than an equivalent equi string hoﬀte den Fall (see exp 5 below). The passive and tough movement sentence types show slowing at the following position, revealed in a signiﬁcant eﬀect for Sentence Type (F = 7.37, p < 0.0001) at this position. This is probably because the extraposition of an phrase after the verb, as occurs in these two constructions, is rather marked. It would be more usual to say wurde für den Fall gewonnen than wurde gewonnen für den Fall, and it seems likely that the eﬀect in these two conditions at this location is due to this. Given these results, we shall inspect the probe recognition results for evidence that either the raising and control, or else the passive and wh-trace sentence types pattern together, and diﬀerently from the other sentence conditions. This, if found, would be evidence that DOP eﬀects are at work in this methodology. The probe recognition data is presented in Table 10. The error rate was large in all conditions (18.5%). While a high error rate is generally considered undesirable in a set of results, it is not clear that this should invalidate this body of data as a source of evidence about the processing of the sentences involved. We note that in Bever & McElree’s studies too the equivalent error rate varied between 7% and 16%. The number of responses excluded because they exceeded 2.5 times the standard deviation by subject was also large (92 data points, 7.8%), but this was probably in part because we set no cut-oﬀ time for probe recognition. Had a maximum response time of 2500ms been set, the 36 correct responses over this time would have been classiﬁed as No Response, and the

NP-trace and PRO 165

data excluded would have been a more normal 4.7%. The reaction times diﬀer little between sentence types. In analyses of variance Subject (F = 24.4, p < 0.001), and Item (F = 40.04, p < 0.01) both proved to be signiﬁcant. Sentence Type (p = 0.18) did not approach signiﬁcance as a main eﬀect. These results therefore do not replicate Bever & McElree’s ﬁndings in which probes recently primed by an EC or overt anaphor were responded to more quickly than the control sentence with no anaphoric element, and produced fewer errors. In fact our results more nearly show the opposite. Notice that our control condition produced the fastest mean response, the second lowest standard deviation, the fewest erroneous responses and the smallest number of points exceeding 2.5 standard deviations. It is unlikely that this combination of facts is pure coincidence, so the probability is that our control condition is showing exactly the opposite eﬀect to Bever & McElree’s. We investigated this with pairwise comparisons between the control condition and the others, following Bever & McElree. The control condition was signiﬁcantly diﬀerent from the passive (F = 5.5, p < 0.02) and the wh-trace conditions (F = 3.7, p = 0.055). It was not signiﬁcantly diﬀerent from the other conditions (p > 0.2). While the data is fairly marginal, there is nevertheless suﬃcient reason to suspect that the control condition is actually faster than the others. Our ﬁrst thought was that this was an eﬀect of complexity: the control sentences are mono-clausal while all other sentence types were bi-clausal. We supposed that the faster responses in the control condition might be an eﬀect of their less complex structure. In order to control for this we constructed a follow-up experiment in which we contrasted mono-clausal and bi-clausal sentence forms of the control sentence type (157).

Table 10.Results of Experiment 4 Sentence type Mean RTs

Raising Passive PRO wh-trace Pronoun Control

949 954 952 956 941 920

Standard Deviation

Excluded data points Errors

Outliers

Total

379 397 365 377 336 345

47 41 49 41 49 40

15 17 15 17 15 13

62 58 64 58 64 53

166 Empty Categories in Sentence Processing

(157) Der geschickte Anwalt, der mit dem ﬂeißigen Richter traf, … the skilled lawyer who met with the hard-working judge a. Mono-clausal wird den Fall/ in Karlsruhe sehr leicht gewinnen will the case in Karlsruhe very easily win b. Bi-clausal meinte, ich werde/ den Fall in Karlsruhe gewinnen thought I will the case in Karlsruhe win

The procedure was identical to that in the main experiment. Each subject saw half of the twenty sentences in the mono-clausal condition and the other half in the bi-clausal condition. After testing eight subjects a preliminary analysis was performed. The results disconﬁrmed our supposition that the less complex mono-clausal structure of the control condition in the main experiment had speeded probe recognition. In fact the mono-clausal sentence form produced slower mean responses than the biclausal (864ms vs. 821ms). Although this diﬀerence was not yet signiﬁcant on the basis of only eight subjects (F = 1.501, p = 0.224), the trend was clear and the follow-up experiment was not continued. If anything, therefore, the fact that the control sentence type was mono-clausal would have slowed it, not speeded it. This structural feature cannot therefore explain its shorter probe response times relative to the other conditions. The third strand of the results of our main experiment is the verb position subcondition in the pronoun and control sentence types. Recall that the DAH would predict activation at the verb, therefore that the verb-ﬁnal condition will be faster. The results are summarised in Table 11. The results in the two sentence types are similar: the V2 condition is slightly faster than the verb-ﬁnal. While subject (F = 6.62, p < 0.001) and item (F = 7.85, p < 0.001) are signiﬁcant on this reduced data set, neither sentence type (F = 1.62, p = 0.2) nor verb position (F = 0.279, p > 0.5) come near. This result shows no sign that the association of the subject with the verb causes measurable reactivation which could inﬂuence probe recognition times.

Table 11.Results of verb position subcondition in milliseconds

Pronoun sentence type Control sentence type

V2 condition

verb ﬁnal condition

932 916

948 926

NP-trace and PRO 167

5.4.3 Discussion The central aim of this experiment was to test whether Bever & McElree’s results showing speeded responses in sentence types with anaphors would be replicated. The results show no sign of this; on the contrary, if anything, they suggest the exact opposite pattern. The sentence type which has no late primer of the subject produced the fastest mean probe recognition time and the lowest error rate. A follow-up experiment designed to test whether this unexpected result was due to the simpler mono-clausal structure of the control sentence type produced data which, if anything, suggested the contrary eﬀect. These results therefore oﬀer no support to the interpretation of Bever & McElree’s results as priming by overt anaphors and ECs. Since this experiment was largely a replication of their studies, it must raise serious questions about the cause of the eﬀects found in their work. While the trace reactivation account is not supported, this data is no more favourable to the alternative accounts. Since we directly tested for verb position in our subcondition, we can immediately comment on the adequacy of the DAH. If the priming eﬀects were due to the association of the subject with its subcategorising verb, we should expect faster responses when this verb is sentence-ﬁnal than when it has already been encountered several words earlier. Our results show no sign of such an eﬀect, in fact if anything the pattern of results suggests the reverse: the V2 condition with an early verb was responded to more quickly than the verb-ﬁnal condition. Furthermore the other sentence types which always had a sentence ﬁnal verb (raising, PRO) produced probe recognition speeds no faster than those conditions which ﬁnished with an extraposed constituent after the matrix verb (passive, tough movement). These results show no sign of the inﬂuence of the factors argued for in the DAH. We did not attempt to exclude the possibility of semantic processing in this experiment but since we noted that the DAH and the SPA can only account for the data on local ECs taken together, the lack of any support for the one must reﬂect badly on the other. The implications for the DOP are perplexing. Since all ﬁve bi-clausal sentence types produced slower responses than the mono-clausal control condition, we ﬁrst hypothesized that this additional processing complexity was slowing the RTs in these conditions. However, in the follow-up experiment designed to test this, the bi-clausal version produced faster responses than the mono-clausal version. The main experiment would therefore lead one to suspect a slowing due to processing complexity, the supplementary experiment

168 Empty Categories in Sentence Processing

rather a speeding due to depth of processing. While the exact relationship of cause and eﬀect in this area is not clear, it does seem possible that length or depth of processing is a relevant factor, perhaps producing immediate slowing but subsequent speeding of responses. There is some reading time data which is consistent with the hypothesis that complexity slows reaction times. The two sentence types which have fastest reading times on the last reading segment (Control, Pronoun) also have the fastest probe recognition times, while the two slowest on the last segment (Passive, Tough Movement) are also slowest in probe recognition. While this can only be suggestive, it does seem likely that there is some connection between the two. The hypothesis that depth of processing can speed probe recognition times does not seem to be reﬂected in this data. We should expect that the two conditions (Raising, Pronoun) with signiﬁcantly slower reading times at the disambiguating segment, which must be the locus of any diﬀerences in depth of processing, should have speeded probe responses. This is however not the case. This data provides some hints that depth of processing may play a role in determining probe recognition speeds, but the diﬀerentiation of complexity and depth of processing must await further examination. One ﬁnal point we shall make concerns the intriguing similarity of our data with that of Bever & McElree’s early testpoint immediately after the EC or disambiguator. Interestingly, the fastest condition here was their control condition. It is therefore possible that our data is responding to the same inﬂuences as their data at their early testpoint. One reason for this might be the relative position of the verbal material in the clause and the testpoint. In Bever & McElree’s studies on English, the early testpoint generally followed the verb, and the late testpoint followed prepositional arguments or adjuncts; in our German sentences our end-of-sentence testpoint follows the clause-ﬁnal verb position. It is conceivable that this late verb position in German is causing the early testpoint eﬀects to appear later. If this were the case it might account for our failure to ﬁnd anything resembling Bever & McElree’s late testpoint results. 5.4.4 Summary The results from this experiment are chieﬂy negative, but shift the balance of evidence away from accounts making use of ECs: the TRA receives no support at all from our data. The DAH is perhaps even more directly disconﬁrmed as an account of the data found in visual probe recognition studies. The data is in part consistent with the DOP, though other parts are not so consistent. The

NP-trace and PRO 169

overall result must be to weaken the TRA and the DAH, and perhaps to provide some support for the DOP. The consequence of this result must be to ﬁrm up the view already represented in the literature (e.g. Walenski 1998) that there is little convincing psycholinguistic evidence for local ECs. It seems increasingly likely therefore that the data argued to represent processing reﬂexes of local ECs is in fact due to other factors such as semantic processing and depth of processing eﬀects. In the next section we outline the importance that this conclusion may have for the position of ECs in the syntax.

5.5 Trace reactivation data and syntactic theories Taking the results of our experiments one to four together, we may say that our results seem to conﬁrm that there are real psycholinguistic eﬀects at trace positions in longer dependencies, but that the experimental evidence for local ECs is weak. In this section we consider what implications this pattern of results has for grammatical theory, in particular comparing our ﬁndings with the predictions of PPT and HPSG. 5.5.1 The empirical adequacy of HPSG and PPT Fodor (1989) suggests that raising and equi constructions are an area of syntax in which processing data might be able to adjudicate between PPT and PSG.56 The relevant diﬀerence between the analyses is the use of local ECs: PPT makes use of them, PSG does not. It is on this basis that Janet Fodor (1989, 1993, 1995a,b) and Ivan Sag (Sag & Fodor 1993, 1995) have made claims about the relative psycholinguistic adequacy of PPT and PSG. On this view, the diﬀerence between the fairly robust wh-trace eﬀects and the paucity of experimental evidence for local ECs provides conﬁrmation of the PSG analysis: “…the data seem most compatible with [the] hypothesis … that wh-trace appears in mental representations of sentence structure, but NP-trace does not. (…) This conclusion … could turn attention towards theories of language such as Lexical Functional Grammar and Phrase Structure Grammar, which are not committed to NP-trace and account for the same language facts by diﬀerent descriptive

56.Note that parts of this discussion concerned GPSG, not HPSG, but the diﬀerences between the two are not relevant here. We shall refer to PSG when the actual discussion mentioned GPSG.

170 Empty Categories in Sentence Processing

means…” (Fodor 1995: 240). In Table 12 below, we can see contrasted the traces assumed by the two frameworks alongside a summary of the experimental evidence for their syntactic reality. Note that Fodor is assuming a version of HPSG which uses traces rather than a traceless version; we comment upon the implications that the removal of traces has for this debate below. Table 12.The empirical adequacy of HPSG and PPT HPSG

PPT

Experimental evidence for syntactic reality?

Raising, Equi

no trace

trace, PRO

None reliable: our exp4 throws doubt on Bever & McElree’s data

Unbounded dependencies

trace (NONLOCAL)

trace

Yes: e.g. Swinney et al. 1989 (conﬁrmed by our exps 2 & 3)

As Fodor and Sag argue, the distribution of experimental evidence seems to match the distribution of traces in HPSG and contradict the assumptions of PPT. “GB clearly predicts that the hypothesized NP-trace and PRO, as well as wh-trace and overt pronouns/anaphors, will exhibit on-line antecedent reactivation eﬀects. PSG predicts that there will be reactivation eﬀects for wh-trace as for overt pronouns/anaphors, but none for control, raising and passive constructions, where no referentially dependent constituents are posited. On current showing at least, it is the PSG prediction which is best supported by the cross-modal priming results” (Sag & Fodor 1993). Our experimental results support this conclusion. In the section below we shall argue that the experimental data is in fact neutral between the two frameworks (see Barss 1993 and Sag & Fodor 1993 for a similar attempt on a diﬀerent basis and a reply). We shall attempt to demonstrate that HPSG’s unexpressed subjects do in fact have syntactic existence, as they show syntactic behaviour, and that they are not theory-neutrally distinguishable from PPT’s ECs. Since this is the diﬀerence from which Sag & Fodor argue for the superior empirical adequacy of HPSG, their claim is dependent on it. 5.5.2 An excursus on covert categories in HPSG The support for HPSG that Sag and Fodor claim from psycholinguistic data is based upon a very clear foundation: PPT makes use of empty categories in

NP-trace and PRO

raising and equi constructions and HPSG does not: “In these cases [raising, equi] the complement subject does not appear in the syntactic structure of the sentence. There is no overt constituent whose index is identiﬁed with another constituent” (Sag & Fodor 1993: 463), “…the facts of […] raising and control constructions are explained in PSG without positing empty categories” (Sag & Fodor 1993: 453). Now let us investigate the nature and representation of the unexpressed subject of an embedded clause in HPSG, the equivalent of local ECs in PPT. We shall call this unexpressed subject USEC (Unexpressed Subject of Embedded Clause) for brevity of reference. Recall that the identity of the subject of the embedded clause of seem or try is deﬁned by its SUBCAT list. Thus in (158) and (159) USEC is deﬁned as sharing some features with the matrix subject, but it appears only on SUBCAT lists. (158) seem [SUBCAT ·NP 1 , VP [inf, SUBCAT·NP 1 Ò] Ò] (159) try [SUBCAT ·NP 1 , VP [inf, SUBCAT·NP 1 Ò] Ò]

Since the features shared in seem are thought to be all SYNSEM features (Pollard & Sag 1994:140), and SYNSEM includes CONTENT (=roughly “truthconditional semantic”) and CONTEXT (= roughly “pragmatic”), it seems reasonable to say theory-neutrally that USEC can have both syntactic and semantic features. It can in addition be said that USEC can be the foot of a dependency, as it is here. USEC can also be the head of a dependency. In a sentence such as (160), the SUBCAT list of the verb behave contains two arguments, an NP subject and an NP of type NP:ana57 which is the reﬂexive. Anaphors such as reﬂexives are bound by a less oblique element on the SUBCAT list. Thus in (160) the reﬂexive himself is obligatorily bound by Robin. (160) Robin behaves himself behave [SUBCAT ·NP, NP:anaÒ]

So in a case where the complement clause of a raising or equi verb contains an anaphor on the subject, this anaphor is bound by USEC (161).

57.This is the standard abbreviation for a sign with a HEAD value noun, and a CONTENT value which is a nominal-object of sort anaphoric.

171

172 Empty Categories in Sentence Processing

(161) Robini tried/seemed USECi to behave himselfi S [SUBCAT 〈 〉] S [SUBCAT 〈 1 NP3 〉]

1 NP3 Robin

2 VP [SUBCAT 〈 1 NP3 , 2 VP[SUBCAT] 〈NP3 〉]〉] [SUBCAT 〈 4 NP3 〉] tries

to [SUBCAT 〈 4 NP3 〉]

VP

V [SUBCAT 〈 4 NP3 , 5 NP:ana〉] behave

5 NP:ana himself

Note that USEC is obligatorily represented on the SUBCAT list of behave, as its omission would entail HPSG predicting that (162) would be grammatical, since reﬂexives with no o-commanding antecedents are unbound. (162) *Robin tried/tended to behave themselves

We thus conclude that USEC can also be the head of a dependency, exactly as PPT’s trace can. An alternative approach might be to suggest that USEC lacks any precisely locatable position in a syntactic string: if this were the case, one might argue that it theory-neutrally distinguished local ECs in PPT and USEC in HPSG. The position of an EC is dependent on the same factors as those which control the position of an overt element (e.g. the head parameter and direction of theta assignment). Interestingly USEC too can be argued to be subject to HPSG’s Linear Precedence rules. Pollard & Sag (1987) state obliqueness as deﬁned by position on the SUBCAT list as the key to Linear Precedence: “Thus the L[inear] P[recedence] constraint […] constitutes a generalization that relates the surface ordering of complements to a diﬀerent, more abstract, and independently motivated ordering, namely the one imposed by the obliqueness hierarchy” (Pollard & Sag 1987: 175). “The content [of Linear Precedence Constraint 2], then, is that (…) any complement daughter is constrained to temporally precede (…) any of its sisters which is a more oblique complement (in the sense

NP-trace and PRO

that it occurs further to the left on the head daughter’s SUBCAT list…)” (Pollard & Sag 1987: 174). Let us summarise the content of these two quotes. The ﬁrst says that linear precedence is dependent upon the obliqueness hierarchy. The second makes it clear that the order of elements on the SUBCAT list is also directly related to obliqueness. Therefore, anything which appears on a SUBCAT list has a (potential) temporal position in the auditory string. So HPSG’s covert categories, like PPT’s trace, potentially have a deﬁnite position in the phonetic form of the sentence, in the position where the equivalent overt constituent would appear. What therefore is the syntactic status of USEC? Plainly it has no phonetic features, but neither has trace. There is no question that it is syntactically active, because it has syntactic and semantic features, and it can be the head or foot of a dependency. It also has an identiﬁable temporal location. The ﬁnal criterion by which one might seek to distinguish the representation of USEC from trace is the quality of being a “sign”. But being a sign is a theory-internal construct of no clear empirical content, and thus is unsuitable as a basis for comparison between frameworks. We summarise the situation in Table 13. Table 13.Comparison of ECs in HPSG and PPT

has syntactic features has semantic features has phonetic form can be the head of dependency can be the foot of a dependency is ordered relative to other constituents is a sign

HPSG’s USEC

PPT’s trace

Yes Yes No Yes Yes Yes No

Yes Yes No Yes Yes Yes N/A

It is on the basis of this evidence that we reject the conventional view that HPSG does not use any equivalent to local ECs. Sag and Fodor make two claims: “… the complement subject does not appear in the syntactic structure of the sentence. There is no overt constituent…” (Sag & Fodor 1993: 463). We accept the second of these but contest the ﬁrst. We consider it reasonable on the basis of the evidence above to refer to the representation of USEC as a “covert constituent”. These convert constituents resemble PPT’s local ECs in that they carry out the referential functions of locally absent arguments as syntactic placeholders. In our judgement therefore the diﬀerence between HPSG’s covert

173

174 Empty Categories in Sentence Processing

constituents and PPT’s traces is notational, and no basis for claims about comparative empirical adequacy.58 5.5.3 The empirical adequacy of HPSG and PPT: Revised The distribution of ECs and robust experimental evidence now looks as in Table 14. Table 14.The empirical adequacy of HPSG and PPT: revised HPSG

PPT

Experimental evidence for psychological reality of EC?

Raising, Equi

covert constituents

trace, PRO

None reliable: our Exp. 4 throws doubt on Bever & McElree’s data

Unbounded dependencies

trace (NONLOCAL)

trace

Yes, e.g. Swinney et al. 1989 (conﬁrmed by our Exps 2 & 3)

The neat correspondence between the use of ECs in HPSG and the existence of reliable evidence has disappeared and left HPSG in the same situation as PPT, namely with only a subset of the phonetically null constituents it utilizes having their existence conﬁrmed by processing data. The data thus does not support the argument for the superior empirical adequacy of HPSG. Let us note here that the use of NONLOCAL features instead of trace in more recent HPSG does not change the balance of this argument. Essentially the traceless account of unbounded dependencies is an extension of the mechanism used in raising and equi: locally absent arguments are subcategorised for and it is this representation within NONLOCAL subcategorisation (parallel to the SUBCAT list), which engages in local syntactic behaviour. Essentially therefore traceless versions of HPSG use covert constituents in unbounded dependencies too (see revised Table 15). This leaves HPSG with the same problem as PPT: why does one sort of covert constituent/trace cause measurable activation and the other not?

58.One might add that a version of HPSG in which these covert categories did have the status of signs could readily be imagined with only minor adjustments to the mechanisms: it would be less economical than the present version in that it would have more empty categories, but it would be more constrained in that it would eliminate a number of unsaturated projections.

NP-trace and PRO

Table 15.The empirical adequacy of traceless HPSG and PPT HPSG

PPT

Experimental evidence for psychological reality of EC?

Raising, equi

covert constituents

trace, PRO

None reliable: our Exp. 4 throws doubt on Bever & McElree’s data

Unbounded dependencies

covert constituents

trace

Yes, e.g. Swinney et al. 1989 (conﬁrmed by our Exps 2 & 3)

We may summarise by saying that the evidence bearing on the psycholinguistic reality of local ECs, which has been thought to distinguish between PPT and HPSG, cannot in fact do so, because it rested on the claim that HPSG had no equivalent, a claim that we have argued does not hold on closer inspection. This data cannot therefore enable us to adjudicate between the two frameworks. Below we report an approach based on another diﬀerence between the raising and equi analyses of the two theories which we hope will permit us to test their empirical adequacy. The evidence gained this way will bear on the question of the psychological reality of ECs, since it involves testing the processing of local ECs against the alternative account with no ECs. 5.5.4 An excursus on diﬀerences between raising and equi in HPSG Consider Table 15 above again. Although the two frameworks cannot be distinguished by their use of ECs in raising and equi constructions, there is another way in which they can be distinguished, namely in making a distinction between raising and equi at the unexpressed subject position where PPT uses local ECs. There is a clear diﬀerence between the two constructions in PPT. Raising involves the movement from a lower position to a higher position of an argument, while in equi constructions the silent pronoun PRO, which is a unique type of EC checking only null Case, is inserted from the lexicon. HPSG on the other hand draws no such sharp syntactic diﬀerence. For HPSG the diﬀerence between the two is semantic, not syntactic: “..the essence of the equi/ raising distinction: all subcategorized dependents of equi verbs are assigned a semantic role; raising verbs always fail to assign a semantic role to one of the dependents that they subcategorize for.” (Pollard & Sag 1994: 134). There are a number of implications of this diﬀerence in semantic role assignment: Pollard & Sag (1994: 136ﬀ) mention the Raising Principle, the distribution of expletive

175

176 Empty Categories in Sentence Processing

subjects, existential entailments, and passivization diﬀerences; however all of these are ﬁrmly located in the semantic features. The only factor which might be argued to be a syntactic diﬀerence is the set of features shared between the controller and USEC. Pollard & Sag (1994: 136) assume that equi involves only the sharing of the INDEX value, while raising shares all the SYNSEM features. This unambiguously aﬀects CATEGORY and not only CONTENT, and is thus clearly syntactic in nature. However, the diﬀerence is somewhat fuzzy; it is essentially quantitative rather than qualitative. Pollard & Sag themselves say “we have not fully motivated the particular choice of SYNSEM objects cross-linguistically in our analysis” and admit that the features shared in raising might be “some subset of the SYNSEM information” (Pollard & Sag 1994: 140 fn. 40). For equi too the set of features shared is not ﬁxed: “…our theory does not prevent the possibility of SYNSEM sharing in equi constructions.” (Pollard & Sag 1994: 140 fn. 40). Since they are willing to envisage full SYNSEM sharing in equi and only a reduced subset of SYNSEM sharing in raising, it must be clear that the set of features shared in each case is merely a descriptive language-speciﬁc detail to be worked out, not a necessary fact implicit in the analysis. We conclude that there is no syntactic diﬀerence in principle between the two constructions in the HPSG analysis, but that the distinction rests in the semantic features. 5.5.5 A new approach This therefore is the alternative perspective from which we intend to test the competing analyses of raising and equi. In our Experiment 6 below, we measure the electrophysical activity of our subjects brains at the location of the unexpressed subject of the complement of raising and equi verbs. PPT predicts diﬀerences between raising and equi constructions at the position of the unexpressed inﬁnitival subject, since there is a syntactic diﬀerence between the local ECs at this position. HPSG predicts no diﬀerence here because it locates the diﬀerence between raising and equi in the semantics and would thus predict processing diﬀerences at the location of the semantic contrast between the two, namely at the matrix verb (163a,b). (163) a.

Der Sheriﬀ schien […] Ø den Täter verurteilen zu können the sheriﬀ seemed the oﬀender condemn to can [no sem role] *SEEMER[sem role] CONDEMNER

NP-trace and PRO 177

b. Der Sheriﬀ hoﬀte […] Ø den Täter verurteilen zu können the sheriﬀ hoped the oﬀender condemn to can [sem role] HOPER[sem role] CONDEMNER

The matrix subject is assigned no role SEEMER in the raising sentence, but it is assigned the role of HOPER in the equi sentence. The subject of the embedded clause is assigned the role of CONDEMNER in both sentence types identically. It is at this position that we test in Experiment 6, since the ECs that PPT assumes at this position are distinct. We may summarise this section by saying that the lack of reactivation of antecedents by local ECs, which was previously held to support HPSG over PPT, does not, on closer inspection, do so, since HPSG does in fact make use of covert constituents which are not in any theory-neutral way distinguishable from PPT’s ECs. It follows from this that all previous studies designed to bear on the issue of the psycholinguistic reality of local ECs cannot do so, for even traceless accounts make use of representations which are empirically indistinguishable. However, we have found an alternative way of testing the hypothesis that these constructions involve ECs, namely by testing for syntactic diﬀerences between the two at the USEC, since these are predicted by the local ECs assumed in PPT but not by HPSG.

5.6 Experiment 5: Event Related Potentials at local EC sites The aim of this experiment was to ﬁnd out whether equi and raising sentences would be processed identically or diﬀerently at the site of PPT’s local ECs and thus allow conclusions to be drawn about the competing analyses. Additionally this experiment was motivated by the unexplained diﬀerence in reading time between segments like schien den Fall and hoﬀte den Fall in Experiment 4 above. There is only one overt diﬀerence between the two segments, the lexical alternation of schien and hoﬀte. Since the verb scheinen has a higher frequency than the hoﬀen set of verbs, and they are broadly matched for length and phonotactic frequency, it seems unlikely that it is this lexical diﬀerence that would motivate a signiﬁcant reading time diﬀerence. This suggests that we should consider the possibility that there is a covert diﬀerence between the two, for example in the representation of the unexpressed subject of the inﬁnitival complement clause, such as is predicted by PPT’s local ECs. To examine this

178 Empty Categories in Sentence Processing

question we were able to perform a experiment measuring event-related brain potentials (ERPs) in cooperation with Thomas F. Münte and Matthias Gross of the Department of Neurology at Hannover Medical School. We brieﬂy review the previous ERP results relevant to sentence processing and some interpretations which have been put on them below (see Garnsey 1993; Osterhout 1994; and Kutas & Van Petten 1994 for overviews of the methodology applied to linguistics more generally). Event-related brain potentials are minute voltage ﬂuctuations of the electrical activity produced by the neurons in the brain that are recorded from various points on the human scalp. The strength of the methodology is the exceptional time resolution of the results; the measurements are time-locked to the experimental stimulus with great accuracy, producing wave patterns which represent the time course of on-line language processing. In our experiment the results have a 4ms sampling rate: for contrast self-paced reading produces one data point per word and CMLP one per testpoint. In addition ERP results are recorded at many diﬀerent positions on the scalp; this oﬀers another dimension of data. A weakness of ERPs is their relatively poor spatial discrimination. The brain is a three-dimensional object and the each electrode gives merely the response at one point on the surface. The relationship between the surface position at which a signal is strongest and its internal origin is complex and subject to a number of clouding factors. This means that it is still a major goal of ERP studies to isolate those electrical components that are correlated with a particular task or stimulus. Previous ERP-studies have led to the identiﬁcation of components involved in semantic and syntactic processing. In the following, we will present a brief summary; see Kutas & Van Petten (1994) and Osterhout & Holcomb (1995) for further details. 5.6.1 ERP data and sentence processing There are three components thought to be fairly speciﬁc to language processing, but one of these seems to be related most closely to lexical factors. The ﬁrst to be discovered was the N400 component, a negative potential with a peak at 400ms and a centro-parietal distribution (Kutas and Hillyard 1980). A multitude of studies have shown that the N400 varies as a function of semantic plausibility and ease of lexical access, most obviously predictability and frequency. It has nevertheless been used to test sentence processing in studies such as Garnsey et al. (1989) which used the embedded anomaly technique (see Chapter 2 above) to test whether semantically anomalous gaps would be posited.

NP-trace and PRO 179

An early left anterior negativity (LAN) has been reported in a range of experimental circumstances. It appears that LANs can be sensitive to morphosyntactic violations of diﬀerent varieties: phrase structure violations (Friederici et al. 1993, 1996), morphosyntactic incongruencies (Coulson et al. 1998; Münte et al. 1998), overregularization errors in irregular verbs and nouns (Penke et al. 1997; Weyerts et al. 1997). Friederici & Mecklinger (1996) suggest that the LAN might reﬂect failure of ﬁrst-pass parsing (see Section 3.2.2 above). Kluender & Kutas (1993) and Kluender & Münte (1998) compared eﬀects at the ﬁller and gap positions in subject and object wh-questions and noted that object questions were associated with a larger LAN. They suggest that this reﬂects the greater working memory load associated with object questions in which the ﬁller has to be kept active for a longer time. Results that can be viewed as supporting the working memory account of the LAN come from studies using embedded subject and object relative clauses (King & Kutas 1995; Müller et al. 1997; Münte et al. 1998a). Kluender & Münte (1998) also contrasted a condition that taxed working memory (object wh-questions) and a condition which contained a that-trace violation. Anterior negativities were seen for both the object wh-question and the that-trace violation, but with clearly diﬀerent scalp distributions. This would suggest that there are diﬀerent varieties of LAN reﬂecting diﬀerent sorts of processing. Our understanding of its origin is thus really quite fragmentary and it is doubtful if any of the proposals regarding LANs correctly identiﬁes the circumstances which can trigger such eﬀects. The second component which has been associated with syntactic factors is a positive-going trend usually having a peak at about 600ms after the stimulus event, hence P600 (also Syntactic Positive Shift). The P600 has been observed in response to a variety of syntactic incongruencies including morphosyntactic, phrase structure, and movement violations. It has been proposed that the P600 is sensitive to syntactic violations (Osterhout & Mobley 1995) but that its amplitude co-varies more generally with higher processing costs (Osterhout 1994), for example garden paths (Osterhout et al. 1994; McKinnon & Osterhout 1996) and syntactically ambiguous sentences (Osterhout & Holcomb 1992, 1993). This would imply that the P600 is elicited not only by outright ungrammaticality but also by syntactic anomalies. Friederici and colleagues (Friederici & Mecklinger 1996; Friederici et al. 1996) have linked the P600 to the second stage of two-stage parsing models (contrasting with ﬁrst-pass parsing for the LAN); it would thus be a measure of the necessary reanalysis of an initially assigned structure. However, in the light of recent evidence that the P600 exhibits a probability sensitivity (Coulson et al. 1998) and can also occur as a

180 Empty Categories in Sentence Processing

response to semantic and orthographic errors (Münte et al. 1998b), Osterhout et al. (1996) suggest that a narrow view of the signiﬁcance of the P600 is untenable, and that it might be related to the P300, which is linked to decision making and is not language speciﬁc. Münte et al. (1998b) have proposed that reanalysis in a more general sense might be the trigger for the P600 and that the amplitude of the P600 might be related to the processing resources required for the computation of a sentence structure. Numerous reports suggest that the amplitude of the P300 component is related to the resources that are consumed by a certain cognitive task (e.g. Johnson 1986). Here too, therefore, we are some way from a full understanding of the trigger for the eﬀect. As ERP experimentation on sentence processing is still in its infancy and our interpretation of ERP responses to syntactic factors is as yet only tentative, we are unable to make predictions for each of the accounts of the antecedent reactivation data as we have done for previous experiments. The null hypothesis we test is thus that there will be no diﬀerence in ERP response between raising and equi at the site of the unexpressed subject. This lack of speciﬁcity in processing predictions allows us only to contrast the PPT assumptions about the local ECs here versus all other accounts. 5.6.2 Materials In this study we compared three conditions: raising (164a), equi (164b) and transitive structures as a control (164c). The aim of the third sentence type is to control for the eﬀect of positing an embedded clause, as this verb type takes only a nominal complement (164c). The diﬀerence between the ﬁrst two that we focus on here is the sort of EC assumed within PPT to be at the position of the embedded subject. Raising verbs imply an NP-trace, equi verbs a PRO (see Section 5.1). (164) a.

Micheli scheint NP-tracei Gerhard zu mögen Michel seems Gerhard to like b. Micheli versucht PROi Joschka zu mögen Michel tries Joschka to like c. Michel mißtraut Oskar Michel mistrusts Oskar

An important consideration which motivated the use of German in this experiment was the position of the verb relative to the presumed EC location. Previous work on this structure has been on English, where the EC and matrix

NP-trace and PRO

verb are generally adjacent, which could lead to eﬀects found at the EC position being attributed to the verb in line with the DAH. The use of German permits us to exclude this alternative account. Sixty sentence triplets of the form in (165a–c) were constructed. (165) a.

raising condition Der Sheriﬀi schien, als die Witwe plötzlich in das Zimmer kam, the sheriﬀ seemed as the widow suddenly into the room came NP-tracei den Täter endlich verurteilen zu können. the oﬀender at.last sentence to can “The sheriﬀ seemed, as the widow suddenly came into the room, to be able to sentence the oﬀender at last.” b. equi condition Der Sheriﬀi hoﬀte, als die Witwe plötzlich in das Zimmer kam, the sheriﬀ hoped as the widow suddenly into the room came PRO den Täter endlich verurteilen zu können. the oﬀender at.last sentence to can. c. transitive control condition Der Sheriﬀ erkannte, als die Witwe plötzlich in das Zimmer the sheriﬀ recognized as the widow suddenly into the room kam, den Täter endlich im Scheinwerferlicht. came the oﬀender at.last in.the spotlight.

Each consisted of a subject and the verb deﬁning the experimental structure (i.e. raising, equi or transitive) followed by an adjunct clause. The presumed EC position was at the head of the embedded clause, which was unambiguously marked by a overtly marked accusative noun phrase. Since no nominal object of the matrix verb is possible in the raising and equi constructions, this gave clear evidence of an embedded clause whose subject must be unexpressed: it is here therefore that the eﬀect of processing an EC can be measured. The sentences were identical (except for the matrix verb) in the three conditions until after the embedded direct object. The adjunct clause between matrix verb and embedded clause was added in order to amplify any ﬁller-gap processing eﬀect; since local ECs have very short dependencies their processing may not be very salient in the data. 5.6.3 Procedure Eighteen right-handed native speakers of German took part at the Hannover Medical School (11 women, 7 men; age range 21 to 35 years, mean age: 27).

181

182 Empty Categories in Sentence Processing

They participated in two sessions separated by several days. The 180 experimental sentences together with 180 ﬁller sentences of various structures were presented word by word in yellow letters on a blue background. Each word was shown for 300ms and was followed by 200ms interval, in which a yellow ﬁxation dot was present in the centre of the screen. To ensure that subjects paid attention to the sentences, a yes/no-question relating to the preceding sentence appeared on the screen 2500ms after the onset of the ﬁnal word of the sentence. It remained on the screen for 3500ms and was followed by the next sentence after a further 2200ms. The task of the subject was to answer the question by pressing one of two buttons. Subjects were tested in two sessions while seated in an easy chair in a dimly lit room. Subjects were instructed how to minimize movement and eye-blinks during the experiment. We recorded ERPs at all 19 standard scalp-sites of the 10/20 system (Jasper 1958) using tin electrodes mounted in an electrode cap with reference electrodes placed at the mastoid processes. The signals were recorded relative to a reference electrode on the scalp (C1) and subsequently rereferenced to the mean of the activity at the two mastoid processes. Additional electrodes were aﬃxed at the right external canthus and at the right lower orbital ridge to monitor eye movements. The signals were ampliﬁed with a bandpass from 0.01 to 70Hz and digitized at 250 points per second. Artifacts were excluded by an automated procedure using individualized amplitude criteria on the scalp channels. ERP averages were calculated for epochs of 2048ms timelocked to the article following the embedded subordinate clause and included a 200ms prestimulus baseline. The data was subjected to repeated measures analyses of variance. Since eﬀects were diﬀerentially distributed over the scalp, separate analyses were done for the midline (ml; Fz, Cz, Pz), parasagittal (ps; Fp1/2, F3/4, C3/4, P3/4, O1/2), and temporal (te; F7/8, T3/4, T5/6) electrodes with the latter two sets split into an electrode site and a hemisphere factor (left vs. right hemisphere). First overall anovas were computed with Condition (raising, equi, transitive), Site and Hemisphere as factors. Then we conducted pairwise tests (raising vs. equi, raising vs. transitive, equi vs. transitive). 5.6.4 Results The ERPs are time-locked to the onset of the accusative deﬁnite article which demonstrates that an embedded clause must be assumed. The three conditions give rise to diﬀerent ERP patterns starting at around 250ms with the eﬀects

NP-trace and PRO 183

being greatest at central and parietal sites. These diﬀerences are most clearly seen in pairwise comparisons at the Cz site (Figures 18–20). Figure 18 compares ERPs from the transitive and equi conditions. The two conditions diverge at about 300ms with the equi condition being more positive; the eﬀect lasting to about 600ms. Statistically, there was a signiﬁcant eﬀect of condition for the midline electrodes (F = 4.61, p < 0.05) and an eﬀect which approached signiﬁcance for the parasagittal electrodes (F = 2.83, p < 0.11) for the 400–600ms time window. ERPs from the raising and transitive conditions are compared in Figure 19. The ERPs in the raising condition proved to be more positive from early in the epoch. This eﬀect was statistically signiﬁcant in the 400–600ms (midline F = 4.29 p = 0.054, parasagittal F = 7.92 p = 0.015, temporal F = 5.34 p < 0.04), and marginally signiﬁcant in the 600–800ms time window (midline F = 4.29 p<0.54, parasagittal F=3.68 p<0.07, temporal F=2.10 nonsigniﬁcant). For two subsequent time windows, 800–1000ms and 1000–1200ms, the eﬀect of condition approached signiﬁcance on several occasions. The comparison between raising and equi conditions showed a more positive going ERP for the former (see Figure 20), with the two conditions diverging at about 500ms. This was reﬂected in signiﬁcant eﬀects for condition in the 600–800ms (midline, F = 6.1 p < 0.025, parasagittal F = 8.77 p < 0.009, temporal F = 6.56 p < 0.02) and 800–1000ms time windows (midline F = 4.81 p < 0.05, parasagittal F = 8.06 p < 0.015, temporal F = 7.13, p < 0.02). In the

Figure 18.The ERPs responses to equi and transitive control conditions at the Cz site

184 Empty Categories in Sentence Processing

Figure 19.The ERP responses to raising and transitive control conditions at the Cz site

1000–1200ms time window, the eﬀect was marginally signiﬁcant (midline, F = 3.65 p < 0.073, parasagittal F = 4.87 p < 0.05, temporal F = 3.82 p < 0.067). Given that some previous ERP studies had revealed a LAN eﬀect at gap positions in wh-questions, we examined the ERPs at the anterior temporal positions. There were no diﬀerential eﬀects between the diﬀerent conditions at the F7 electrode. However, a comparison of the F7 with the corresponding right hemisphere site (F8) showed a more negative ERP over the left hemisphere, which probably corresponds to a LAN in all three conditions. The left/right diﬀerence only appears for the critical article but not for the control article,

Figure 20.The ERP responses to raising and equi conditions at the Cz site

NP-trace and PRO

suggesting that in fact a LAN was present at this position in all three conditions. Statistically, this was reﬂected in main eﬀects for hemisphere for the temporal electrodes in time windows 200–400ms (F = 17.63, p < 0.001). 5.6.5 Discussion We can identify two hypotheses that this data bears upon. The ﬁrst is simply that there is a diﬀerence in processing between positing a clause as a complement and positing a nominal complement. This is conﬁrmed and is consistent with the view that the P600 co-varies with higher processing costs (Osterhout 1994). The second more controversial hypothesis is the main focus of our interest here. This is that there are processing diﬀerences between raising and equi structures at a position where they are overtly identical but at which the assumption of local ECs would predict a diﬀerence. This too is conﬁrmed by the statistically signiﬁcant diﬀerence between the two in their P600 components. The precise nature of the contrast cannot be identiﬁed on the basis of this data alone. Since the ungrammaticalities and implausibilities which are have most commonly been found to elicit P600s are absent here, our approach must therefore be to consider the range of possible explanations. One possible candidate is the claim by Friederici & Mecklinger (1995) that the P600 represents activity at a second, structure-analysis checking stage of parsing, perhaps caused by the unusual distance between matrix verb and embedded clause brought about by the inserted adjunct clause. On this interpretation therefore we should assume that the distance forced subjects to make some assumption about how the sentence was going to continue, and that the amplitude of the P600 would represent the degree of dispreferredness of the eventual continuation of the sentence. This would be consistent too with the results found by Coulson et al. (1998) showing that P600 was ampliﬁed by the probability of the sentence structure. In order to control for this account of the results in terms of violation of lexical preferences, we performed a supplementary experiment with our materials designed to establish whether the raising sentence type was less preferred than the equi sentence type. A pencil and paper test was administered to 18 native German speakers at the University of Düsseldorf in which subjects were instructed to grade sentence pairs by their naturalness. Ten randomly selected raising and equi items from our ERP experiment were presented among 20 ﬁller items. Each experimental sentence was shown in two forms: with the adjunct clause interposed between matrix verb and embedded object

185

186 Empty Categories in Sentence Processing

(166a), and with the same clause attached externally (166b). Subjects were presented with these two forms together and invited to judge each of these sentences on a ﬁve point scale from 1 (very natural) to 5 (very unnatural). (166) a.

Der Sheriﬀ schien/ hoﬀte, als die Witwe plötzlich in das Zimmer the sheriﬀ seemed/ hoped as the widow suddenly into the room kam, den Täter endlich verurteilen zu können. came the oﬀender at.last condemn to can b. Der Sheriﬀ schien/ hoﬀte, den Täter endlich verurteilen zu the sheriﬀ seemed/ hoped the oﬀender at.last condemn to können, als die Witwe plötzlich in das Zimmer kam. can as the widow suddenly into the room came

Subjects were divided into two groups: each group saw the same sentences, but those that group 1 saw as raising structures, group 2 saw as equi structures and vice versa. This provided us with a measure of the relative naturalness of the equi and raising versions of the sentences without the subjects being aware of the real aim. In addition it allowed us to test whether the somewhat marked internal position of the adjunct was causing diﬀerent eﬀects in raising and equi, and thus exclude this as the cause of our experimental eﬀects. Four out of 720 responses were null and therefore discounted. The results show our raising structures judged to be marginally more natural than the otherwise identical equi items (see Table 6). Table 16.Mean naturalness by structure and adjunct position (n = 716) Mean naturalness (1 = very natural, 5 = unnatural)

Raising structures Equi structures

Internal adjunct

Extraposed adjunct

3.40 3.54

1.92 2.04

Analyses of variance were carried out by Subjects (F1) and by Items (F2). There is a signiﬁcant main eﬀect for Subject (F = 36.71, p < 0.01) but Item was not signiﬁcant (F = 0.535, p > 0.9). Adjunct position was strongly signiﬁcant (F1 = 671.04, p < 0.001, F2 = 327.22, p < 0.001), but there was no interaction of Adjunct position by Sentence Type (F1 = 0.03, p > 0.86, F2 = 0.01, p > 0.92), which demonstrates that the marked adjunct position was not responsible for our experimental eﬀects. Sentence type was signiﬁcant on one analysis and approached signiﬁcance on the other (F1 = 6.02, p = 0.014, F2 = 2.94, p = 0.086),

NP-trace and PRO 187

but since the raising sentences were judged more, not less, natural than the equi, the signiﬁcance of this eﬀect is not critical. The important point is that this result rules out the structural preferredness account of the greater P600 in the raising condition of our ERP study, since this should be reﬂected in our naturalness judgement task too, if such a diﬀerence existed. In order to conﬁrm this result we also conducted the same experiment using self-paced reading methodology. We used the full set of materials from the original ERP experiment but with one minor change. We replaced the internal adjunct clause with a single adverb. The reason for this was practical; it greatly reduced the amount of text that subjects had to read, and as the naturalness judgement test above had shown no diﬀerences between the versions with internal and extraposed adjunct clauses, we were conﬁdent that it would not aﬀect results here either. The sentences were presented in small phrasal segments (as indicated by the slashes in the example below). (167) a.

Der Admiral/ hoﬀte/ trotzdem/ den Text/ nochmal/ im Detail zu überprüfen b. Der Admiral/ schien/ trotzdem/ den Text/ nochmal/ im Detail zu überprüfen c. Der Admiral/ überprüfte/ trotzdem/ den Text/ nochmal/ im Detail

33 native German speakers participated in the experiment at the HeinrichHeine-Universität Düsseldorf and were randomly divided into three groups. Each group saw the same items in counterbalanced scenarios such that everyone saw the same sentences in some form, no subject saw the same sentence twice, and every subject saw the same number of each structure type. The 60 experimental sentences were pseudo-randomly mixed among 140 ﬁller sentences. The results were simple: there were no signiﬁcant diﬀerences between structures types at the direct object, the onset of the embedded clause, the equivalent position to the ERP testpoint. This result too therefore shows no sign of complement preference diﬀerences between the conditions. Let us note here that these results would tend to support the proposal that the P600 reﬂects additional processing complexity due to reanalysis, as suggested by Friederici & Mecklinger (1996). It seems plausible that the reanalysis is the reconstruction of the matrix subject as the embedded subject, and that that adding a raising embedded clause involves more reanalysis than adding an equi embedded clause. Recall that in the self-paced reading data from our Experiment 4 we found signiﬁcantly slower reading times for the string schien den Fall than hoﬀte den Fall, whereas the higher frequency of scheinen than the equi

188 Empty Categories in Sentence Processing

verbs would predict the opposite. These two results converge in showing greater complexity in raising than in equi constructions at the onset of the embedded clause. Notice that while this view that the P600 represents reanalysis is consistent with previous studies, it is not essential to the testing of our hypothesis that the processing of raising and equi diﬀer at the unexpressed embedded subject location, for any diﬀerence will suﬃce to disconﬁrm the null hypothesis that there is no such diﬀerence. It is a striking fact that the results of our ERP experiment, as well as the selfpaced reading data, are immediately consistent with the processing predictions of the PPT analysis of raising and equi with ECs, whether or not we choose to interpret the P600 as a measure of reanalysis. If we do, then we compare the structural dependencies of NP-trace and PRO to determine which is the more complex. It seems clear that raising, which involves an extraction and the formation of a multi-membered chain with a Case position at its head and a theta position at its foot, is a more complex structure than equi, which consists merely of the insertion of a caseless null pronoun from the lexicon and a dependency merely of reference. The markedly greater amplitude of the P600 in the raising condition on this account reﬂect the greater processing load of positing an NP-trace. If we do not draw the conclusion from previous studies that the P600 co-varies with computational complexity, then we simply interpret the results minimally, and admit only that they demonstrate that there is a marked diﬀerence between the processing of the experimental conditions at the onset of the embedded clause. As we have seen, this is quite consistent with the analysis of the two constructions in PPT, since this posits the use of quite distinct mechanisms. This is in contrast to HPSG which employs the same mechanism and makes no such prediction of diﬀerence at this position. We conclude that our ERP result can be seen as more consistent with the approach to raising and equi in PPT than with that in HPSG. Since the PPT account makes use of ECs, these too as a component of the more empirically adequate account can be argued to be supported. Let us note here three possible counter-arguments which might be made. The ﬁrst concerns the diﬀerence in features shared between the matrix and embedded subjects in raising and equi. We pointed out that Pollard & Sag themselves (1994: 140) admit that their analysis does not demand that this diﬀerence exist, and argued that this meant that HPSG does not predict any processing diﬀerence. However, the point could certainly be made that Pollard & Sag’s comment does not alter the fact that the standard analyses of raising and equi do in fact diﬀerentiate here and that this makes the HPSG analysis

NP-trace and PRO 189

consistent with the data. Additionally, there is a fairly standard assumption in HPSG that the sharing of merely the INDEX features, as in equi, represents a dependency of reference, and that this is diﬀerent in kind to a full structureshare, as in raising. A second objection which might be raised concerns our claim that PPT uses two diﬀerent mechanisms in the two structures while HPSG uses the same mechanism. We assumed that the use of two mechanisms corresponds more exactly to the large diﬀerence in the ERP data. In reply one might comment that the reason that HPSG makes use of only one mechanism here is because it is a more parsimonious grammar, and this should not be regarded as an empirical inadequacy (see similar discussion in Barss 1993 and Sag & Fodor 1993). Furthermore, a large part of the motivation for the diﬀerentiation of the mechanisms in PPT is due to technicalities of Government, which has since been abandoned, removing much of the basis for the distinction. Lastly one might argue that the location of the processing of HPSG’s semantic diﬀerences between equi and raising at the matrix verb is not strictly necessary. This view would admit that semantic contrasts are lexically dependent on the character of the verb as a raising or equi predicate, but would point out that the structure is not unambiguously identiﬁable as raising or equi at the matrix verb position, but only at the disambiguator, the onset of the embedded clause. We raise these possible objections, but we shall not reply to each here since it is unlikely that any deﬁnitive answer would be attainable.

5.7 Conclusions In this chapter we have reported two experiments which each had the aim of shedding light on the status of the local ECs made use of in PPT. The experimental techniques currently available oﬀer us the chance to test whether these ECs have some identiﬁable reﬂex in processing. This seems a fruitful approach, for positive evidence for this would confer a strong claim on psychological reality for these ECs and thus assure them a secure position in grammatical theory. Previous experimentation has shown only weak evidence for their psycholinguistic existence, as the data has left a number of questions unanswered. We performed a experiment similar to those of McElree and Bever (1989) but using German exemplars of the structures, and with some additional subconditions. The main result showed rather contradiction than support for Bever and McElree’s interpretation of their ﬁndings in terms of antecedent reactivation. The control condition with no anaphor provided the fastest

190 Empty Categories in Sentence Processing

response times and the fewest errors and outliers: exactly the reverse of Bever and McElree’s result. These diﬀerences were only marginally signiﬁcant, but it is nevertheless clear that our data tells a very diﬀerent story to theirs, and must generally favour an alternative account. The SPA predicts reactivation at the clause boundary but provides no obvious explanation of the diﬀerential activation of the diﬀerent sentence types in the Bever and McElree studies and its reversal in our own data. We tested the predictions of the DAH in our verb position subcondition, and found no support at all for it. This leaves some form of the DOP as the strongest contender. It successfully accounts for the faster responses in the biclausal than in the mono-clausal sentences in our supplementary experiment, and may provide a framework for the explication of the reversal of results in the Bever & McElree studies and in our own, for a diﬀerence in local complexity due to the diﬀerent position of the verb is a credible cause of the contrasting ﬁndings. Nevertheless, the DOP is far from being a fully developed account of the data as a whole for numerous problems remain. In particular we would need to know whether depth of processing is the same as or diﬀerent to the convention idea of processing complexity, and why it is that complexity is generally thought to slow responses while depth of processing can speed them. Such an account is not impossible: one can readily imagine an account in which complexity and depth of processing are indeed the same, but produce slowing and speeding on diﬀerent time scales. At the position of the complexity all processing, including such experimental tasks as probe recognition, is slowed because of the heavier load on the ﬁnite computational resources. Once the complexity is past, an element which was part of the local complexity is responded to faster because it has a higher activation. While the outline of such an account is clear, it is plain that further work is required for it to become more than a post hoc account and make predictions in any detail. Our ﬁfth experiment was motivated by the view that previous approaches to the issue cannot distinguish between syntactic accounts with traces and those without, since these latter actually make use of a construct empirically indistinguishable from an EC. In this light we tested for another diﬀerence between the analysis of raising and equi constructions with and without ECs: the prediction of a processing distinction between the two at the unexpressed embedded subject position. While the exact nature of the diﬀerence cannot be ﬁrmly identiﬁed, the data is consistent with the PPT analyses and the positing of diﬀerent EC types at this position. This result replicates eﬀects5 found in the self-paced reading data from Experiment 4 and is not predicted by traceless accounts of

NP-trace and PRO

the data such as HPSG, though we have seen that these can be argued to be consistent with it. It is possible to take the further step and assume that the P600 found represents processing complexity due to reanalysis, since this has been argued for on the basis of previous studies. Now on fairly natural assumptions about the diﬀerent tasks involved in the processing of raising and equi structures, the former is more complex than the latter. This would open the possibility that the P600 may actually represent the computational cost of the ECs involved. A more conservative conclusion would be to say that the PPT analysis predicts the ERP data, while the HPSG analysis is only consistent with it. To summarise: while our Experiments 2 and 3 provided robust evidence of measurable processing reﬂexes of trace, our Experiment 4 not only does not support any role for local ECs but seriously undermines the credibility of previous visual probe recognition results. We may conclude that it is questionable whether there is any empirical evidence pointing to the psychological reality of local ECs since the CMLP data can readily be accounted for by traceless mechanisms. This result does not, however, lead us to credit HPSG with greater empirical adequacy since, contrary to what is generally supposed, it too makes use of an equivalent construct. However, the conclusion that data argued to be experimental evidence for local ECs was in fact due to other factors in processing must raise the question whether this might not be true of the evidence for other trace too. For it is intuitively unlikely that one sort of phonetically null anaphor should produce such strong results but another none at all. Our Experiment 5 is an attempt to elucidate the issue from a new perspective. Since this methodology indisputably measures on-line processing, any evidence from it weighs heavily in the balance. The strongly contrasting ERP responses to raising and equi at the presumed local EC site place a large question mark over any analysis which treats the two inﬁnitival clauses as equivalent. One interpretation of the ERP diﬀerence is as the processing reﬂex of NP-trace and PRO, though this will require further work to be conﬁrmed.

191

AUTHOR ""

TITLE "Evidence from sentence matching on wh-trace dependencies"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 6

Evidence from sentence matching on wh-trace dependencies

In Chapter 4 we focused on the evidence for the psychological reality of wh-trace; this chapter focuses on another diﬀerence between PPT and HPSG in their treatment of unbounded dependencies, namely the nature of the dependency. Within the framework of PPT extracted elements are represented at the extraction site by traces. The transfer of syntactic and semantic information from head to foot of the dependency is carried out by chain formation, a form of binding. Successful chain formation depends upon the conﬁguration fulﬁlling the requirements of binding theory and bounding theory, each of which constitutes a separate module of the framework, and which have a structure-checking not structure-building character. EC dependencies are subject to the degree of locality that the applicable principle in the binding theory allows. The relevant deﬁnition of locality thus varies according to the type of dependency (relativized minimality: Rizzi 1990, shortest move Chomsky 1995: 261ﬀ). This character of trace binding as a structure-checking operation with a relativized locality deﬁnition contrasts with HPSG. The HPSG representation of these dependencies is as a system of synchronic and local constraints on possible feature values, which are strictly local between mother and daughters, information being passed through the tree on a cascade model. This is a nonderivational framework in which all constraints are applied simultaneously, and which allows no division between structure building and structure checking. These representations of unbounded dependencies have diﬀerent implications for processing. One diﬀerence is in the divisibility of the processes of phrase structure computation and dependency formation. It is clear that a principle-based processor (e.g. Crocker 1996) implementing PPT, a modular and derivational theory, can readily allow some processing asynchrony between the computation of the underlying phrase structure and subsequent operations upon it. This would be possible because PPT contains both phrase structure modules such as X¢-theory (Jackendoﬀ 1977; Chomsky 1986a: 2, 1995: 172ﬀ, 241ﬀ), and structure checking modules such as binding, and these are plausible

194 Empty Categories in Sentence Processing

candidates for initial and later parsing functions. Such a disjunction would be excluded in an integrated and incremental processor (e.g. Stevenson 1994) in which the computation of a structure-share is merely a subpart of the computation of the constraints which each projection imposes upon its immediate neighbours. A further implication for processing is that a modular parser can account for processing diﬀerences between local and non-local constructions, since for example X¢-theory makes use of strictly local head-complement and speciﬁer-head relations, whereas binding theory applies the looser chain-link conﬁguration. A strictly incremental processor cannot do this, since it implements all constraints step by step as the structure is projected. In this chapter we report a sentence matching experiment designed to probe these syntactic and processing diﬀerences. The processing of unbounded dependencies which violate movement constraints represents a test case on the basis of which we can draw conclusions about these wider issues. We restrict ourselves to discussion of the sentence matching evidence here, since this methodology seems uniquely to be able to distinguish construction types in an interesting way relevant to these issues (for broader discussion see Pickering, Barton & Shillcock 1994 and Chapter 3 above). The questions we address are: Is the notion of locality strictly identical across constructions, or does another broader conception of locality apply to non-local dependencies? The other: is there evidence that non-local dependencies are processed within a separate module of the parsing mechanism? Interestingly, one interpretation of the experimentation using sentence matching we shall discuss below suggests that there are indeed these processing diﬀerences between local and non-local structures. If this were to be conﬁrmed it would constitute evidence supporting models which readily permit such eﬀects. Let us however note here a change in perspective from previous chapters. The debate about sentence matching in the literature has the unusual feature that it has as its unifying theme a methodology and a result, rather than a research question about the parser or its grammar. More usually a hypothesis is formed and experiments are carried out to conﬁrm or falsify it. In this case a particular experimental phenomenon was observed and the issue is rather the correct characterisation of the data set in which it appears. 6.1 Accounts of the sentence matching data It was discovered in the ﬁfties that the mind is better able to perform a task on a set of materials if it is structured. This provides a rich experimental heuristic

Evidence from sentence matching on wh-trace dependencies 195

from which conclusions can be made about language structure. The correspondence between structure and speed of manipulation is very well established and has been replicated over a wide range of tasks and structure types. The eﬀect is robust at phonetic, phonological, lexical, phrasal and sentential levels, but in this context we are only concerned with the sentential level (see Fodor, Bever & Garrett 1974; Forster 1979; Freedman & Forster 1985 for reviews of work on other linguistic levels, Forster 1987 for plausibility). The procedure usually adopted to exploit this eﬀect is sentence matching. A sentence is presented visually on a screen for subjects to read, and a short time later a similar or identical sentence is presented. The subjects’ task is to indicate with a button press whether the second sentence is identical to the ﬁrst. As this involves the subjects’ creating a mental representation of the ﬁrst sentence with which to compare the second, it is sensitive to the amount of structure which they can impose upon it. Since sentence matching is responsive to the subjects’ perception of structuredness, it provides a test of grammaticality; any ungrammaticality prevents the development of a sentential representation of the presented string and slows responses. The reliability of this detector of ungrammaticality has been conﬁrmed in many studies (e.g. Forster 1979; Freedman & Forster 1985). There is one apparent exception: certain very speciﬁc ungrammatical structures have been shown not to give the normal slowing relative to grammatical controls. We report the data and the main contributions to the debate about its signiﬁcance in the following sections. 6.1.1 Constraint violations Freedman & Forster (1985) tested sentences in four conditions (±Movement, ±Constraint)59 in order to factor out the eﬀect of Speciﬁed Subject Condition violation:60 (168) a.

+Movement +Constraint *Who did the duchess sell Turner’s portrait of? b. +Movement −Constraint Who did the duchess sell a portrait of?

1549ms 1268ms

59.The feature CONSTRAINT should be read here as “the conditions exist under which the constraint would apply”: i.e. if there is no extraction from the embedded NP it applies vacuously. 60.Speciﬁed Subject Condition: No rule may relate X and Y in the structure …X…[α…Z…W1YW2…]α… where Z is the subject of α (Chomsky 1973).

196 Empty Categories in Sentence Processing

c.

−Movement +Constraint The duchess sold Turner’s portrait of her father. d. −Movement −Constraint The duchess sold a portrait of her father.

1511ms 1199ms

These materials allow us to exclude lexical diﬀerences and syntactic complexity as variables. The contrast of (168c) and (168d) allows us to calculate the cost of the genitive speciﬁer on the embedded NP (+Constraint = 312ms). The cost of the question form over the declarative is ascertained from (168b) and (168d) (+Movement = 69ms). Adding these two to (168d) gives us the predicted reading time of (168a), ignoring grammaticality (= 1580ms); any additional slowing can be considered the matching time cost of the ungrammaticality. In fact this sum results in a ﬁgure of −31ms as the cost of the constraint violation. There was no signiﬁcant interaction of Constraint and Movement, as would have been expected. There is thus no eﬀect of ungrammaticality. Freedman & Forster conducted a number of follow-up experiments to locate the cause of this lack of eﬀect: they ﬁrst tested whether subjects perceived these constraint violation sentences to be ungrammatical. They compared their acceptability with three types of controls: fully grammatical, fully ungrammatical, and other constraint violations (in the sense of Chomsky 1973; Ross 1967).61 We summarise the results in percentage form in Table 17.62 The experiment showed that the SSC violations were judged as slightly more grammatical than other constraint violations and ungrammaticals, but the overall groupings into acceptable and unacceptable followed the assumptions of grammatical theory. This would seem to exclude the possibility that the lack of eﬀect for the SSC was due to perceived grammaticality. The fact that the SSC sentences were given higher acceptability scores than other ungrammaticals raises another possible explanation of the lack of eﬀect, namely that the sentence matching task is sensitive only to gross ungrammaticality. Freedman & Forster’s third experiment controlled for this by testing cases where the departure from ungrammaticality was relatively minor.

61.Notice that we use the term “constraint violation” here to mean exactly what was meant by it in the papers that we are reviewing. In fact some of the examples have alternative analyses in the light of more recent work in syntax. 62.20 judges saw 10 sentences of each type and chose between “sounds OK”, “sounds funny (i.e. odd)”, and “sounds bad”. We give an acceptability rating by assigning 2 points for “OK”, 1 for “odd” and 0 for “bad”.

Evidence from sentence matching on wh-trace dependencies 197

Table 17.Sentence type acceptability ratings (Freedman & Forster 85 Exp. 2) Sentence type

SSC violations (M+ C+) SSC controls (M+ C−) Fully grammaticals Fully ungrammaticals Other constraint violations

Scores in % OK

Odd

Bad

18 67 93 0 6

52 31 7 64.5 52.5

30 2 0 35.5 41.5

Acceptability rating in % 44 82.5 96.5 32.25 32.25

Agreement errors and quantiﬁer misplacement (169) were chosen as representative of fairly minor slips which might be invisible to sentence matching. (169) a. *Mary were writing a letter to her husband b. *The baby ate his cereal up all

Signiﬁcant ungrammaticality eﬀects were found for both sorts of “minor” ungrammaticality. Another possibility suggested by the grammaticality judgement data was that sentence matching is not sensitive to degrees of ungrammaticality. If so, the control sentences in the experiment, whose grammaticality judgement ratings were also not fully grammatical, could be causing the same eﬀects as the constraint violations, and thus making these violations invisible. In their experiments four and ﬁve (which we conﬂate here), Freedman & Forster constructed sentences with increasing amounts of illegal variation from grammatical sentences like (170a). They ﬁrst exchanged sentence parts at an intuitive major phrasal break to make “phrase structure scrambles” (170b), they next made “word scrambles” by putting the words in random order (170c), thirdly they made word scrambles of lexical substitutes (170d). (170) a. b. c. d.

The guest introduced his wife politely *His wife politely the guest introduced *Politely wife guest his the introduced *Friend a entertained aunt happily my

These sentences showed eﬀects proportionate to their distance from the original grammatical sentence, conﬁrming that sentence matching is responsive to degrees of grammaticality. On the basis of these results Freedman & Forster argued that their data was evidence that the rules and constraints were located in diﬀerent modules of the language faculty. Essentially the argument runs: the sentence matching task is

198 Empty Categories in Sentence Processing

sensitive to the degree of structure which can be put upon the materials presented. If subjects match sentences with constraint violations as fast as they do fully grammatical sentences, this is evidence that there exists a level of representation in the computational system at which these sentences have a fully acceptable representation. Since PPT syntax divides the grammar into modules, they hypothesize that, while doing the sentence matching task, subjects can eﬀectively switch some modules oﬀ. In particular, they may be able to block the application of wh-trace binding, while still allowing the representation to receive an interpretation. Forster (1987) refers to this as the Shallow Sensitivity Hypothesis. Since this can be accounted for in a natural way in a modular grammar, this type is supported, they suggest (Freedman & Forster 1985: 125). A weakness of this argument (admitted by Freedman & Forster) is that it is unclear exactly what mechanism would allow a subject performing sentence matching to access a representation other than that which is normally fed to the performance interface to produce a surface form. 6.1.2 Correctability and coherence Crain & Fodor’s account (1985b, 1987) of the phenomenon puts it down to a particular combination of qualities in the constraint violation sentences. They are coherent but also, while wrong, not easily correctable. The other ungrammatical sentence types in Freedman & Forster’s experiments only ever have one of these two qualities. For each of the sentences in (171) there is a corrected version which is immediately apparent, which has exactly the same interpretation, and would probably be invariant among diﬀerent people asked to provide it. Intuitively, each ungrammatical sentence is an erroneous form of its grammatical equivalent. (171) a. *Mary were writing a letter to her husband Mary WAS writing a letter to her husband b. *The baby ate his cereal up all The baby ate his cereal ALL UP c. *Why didn’t father let I stay up last night? Why didn’t father let ME stay up last night?

Contrast this with the constraint violation sentences in (172):

Evidence from sentence matching on wh-trace dependencies 199

(172) a. b. c. d.

*Who did the duchess sell Turner’s portrait of? *Who did they deny the claim that the burglar shot? *Which politician did you meet the students who worked for? *Which bag were you silent while the oﬃcers searched through?

None of these sentences has the same sort of correspondence to a grammatical sentence, such that they could be said to be an erroneous form of it. Crain & Fodor suggest that it is this feature which causes the diﬀerence in matching times. Correction is a virtually automatic process; when subjects see the erroneous version they compute, and memorize, the corrected form. When the second form of the sentence is presented, the diﬀerence between the memorized form and the newly presented form will “cause confusion” (Crain & Fodor 1987: 148) and result in a longer matching time. To back this up they produced their own sentence matching experiment to contrast the variables of Correctability and Violation Type. Types (173a) and (173d) are the correctable rule violations and the uncorrectable constraint violations we have discussed so far. Type (173b), correctable constraint violations, were a mixture of six illegal constructions whose ungrammaticality has been attributed to constraint violations; type (173c), uncorrectable rule violations, were subcategorization alternations legitimate in one context applied to a diﬀerent illegal context.63 (173) a. b. c. d.

*In the study Mary were writing letters to her husband *Is this the coat which that you want to buy? *The plumber took what the time was to ﬁnish the job *Who did you buy Picasso’s portrait of?

63.Crain & Fodor admit that these uncorrectable rule violations were extremely diﬃcult to construct, because they should be ungrammatical and uncorrectable but also coherent. The best way they found to do this was to use substitution equivalents which were appropriate in one structure in another structure where they were inappropriate. For example, (iii) and (iv) are legitimate alternates, but what the time was cannot be transplanted from (ii) to (i). (i) *The plumber took what the time was to ﬁnish the job (ii) The plumber took the time to ﬁnish the job (iii) The plumber asked what the time was (iv) The plumber asked the time The problem is that such sentences are uncorrectable only as far as they are incoherent; as soon as their intended communicative content is identiﬁed they are readily correctable. In this way they exemplify the confound between coherence and correctability we discuss below.

200 Empty Categories in Sentence Processing

The design, the predictions, and the results are all summarized in Table 18. Freedman & Forster would predict the violation type to be the distinguishing factor; Crain & Fodor predict that it is correctability. Crain & Fodor’s experiment conﬁrms Freedman & Forster’s empirical ﬁndings in (173a) vs. (173d), but the new conditions (173b) and (173c) show that violation type cannot account for all the data. Crain & Fodor’s own correctability account receives some support: there are clear diﬀerences between the correctables and the uncorrectables. But this support is only partial: there is for example no signiﬁcant diﬀerence between the correctable and uncorrectable rule violations (173a) and (173c), and, among the correctables, the more correctable rule violations (173a) are actually faster than the less correctable constraint violations (173b). The experiment itself has weaknesses: the (173b) and (173c) type materials are less than ideal, which is a serious problem, since it is precisely these conditions which carry the burden of the distinction between the two competing accounts.64 Crain & Fodor also confess that there is a confound between correctability and local vs. global ungrammaticality. Nevertheless Table 18.Design and results of Crain & Fodor (1987) Experiment 2 Sentence type

Correctability Matching time Predictions score incrementa Crain & Fodor Freedman & Forster

a. Correctable rule viola- 1.21 tions

181ms

slow

slow

b. Correctable constraint 1.77 violations

221ms

slow

fast

c. Uncorrectable rule violations

2.71

72ms

fast

slow

d. Uncorrectable constraint violations

3.07

−26ms

fast

fast

a

The increment is the diﬀerence between the actual result and the result predicted by factoring out the costs of Movement and Constraint in the control materials.

64.Viewed from the perspective of today’s syntactic theory we should say that the most that the correctable constraint sentences have in common is that a syntactician at some time has stipulated a constraint to deal with them. Structurally, they have little in common and do not form any natural class. The uncorrectable rule violations are more homogenous but we have noted above that they are uncorrectable only as far as they are incoherent.

Evidence from sentence matching on wh-trace dependencies 201

Crain & Fodor do succeed with this experiment to clarify that the simple violation type explanation put forward by Freedman & Forster is not adequate. Crain & Fodor use the factor of coherence to address Freedman & Forster’s experiments ﬁve and six, which attempted to control for correctability. They (Freedman & Forster) had produced two sentence types which were designed to be uncorrectable. The ﬁrst consisted of sentence halves crossed between sentence pairs (i.e. (174a,b) > (174c,d)), the second were the same but with the crossing always taking place at the phrasal boundary between NP and VP (i.e. (175a,b) > (175c,d)), to maximise local grammaticality. (174) a. b. c. d.

Tony was granted bail by the judge The spy admitted that his plan failed >>>>>>> *Tony was granted bail that his plan failed *The spy admitted by the judge

(175) a. b. c. d.

The girl behind you reminds me of my sister The subsequent discussion soon got boring >>>>>> *The girl behind you the subsequent discussion *Reminds me of my sister soon got boring

Both of these sentence types showed full ungrammaticality eﬀects, suggesting that correctability was not the relevant factor. Crain & Fodor admit that these are uncorrectable, but argue that it is their lack of coherence which makes them diﬃcult to represent as sentences and thus slows their matching. Their (Crain & Fodor’s) Experiment 4 tests the coherence of these sentences and ﬁnds it lower than other uncorrectables, and so they put down the eﬀects Freedman & Forster found here to this factor. Forster & Stevenson (1987) make several eﬀective points in response to Crain & Fodor. They show that Crain & Fodor’s correctability and coherence account predicts a saw tooth pattern of RTs with increasing levels of ungrammaticality. No such eﬀect is found when they test for this with a sentences forming a continuum of unacceptability. They also point out that Crain & Fodor’s two-pronged argument can account for any result and thus makes no testable predictions, since there is no way of distinguishing between the two. Lastly they show that Crain & Fodor’s experiment can be seen as a useful extension to Freedman & Forster’s original experiment, ruling out certain classes of constraints from the group which lack an ungrammaticality eﬀect. Seen from this perspective, there is no need for correctability as an explanatory factor: the question is only what class of violations lacks an ungrammaticality eﬀect. Crain & Fodor’s experiment shows that violation type crudely deﬁned could

202 Empty Categories in Sentence Processing

at best be a partial factor, but correctability too is only a weakly supported contender. Even Crain & Fodor’s own experiment provides only weak conﬁrmation of it; the correlation coeﬃcient of correctability and matching time they found was only 0.26. Later experiments have been argued to replicate Freedman & Forster’s result with materials controlled for correctability (Forster & Stevenson 1987; Clahsen, Hong & Sonnenstuhl-Henning 1995). There is no doubt that Freedman & Forster’s constraint violations which produced no ungrammaticality eﬀects are considerably less correctable than the average unacceptable sentence. But Forster & Stevenson’s (1987) experiment used much less confound-ridden materials than Crain & Fodor’s, and found no sign of its inﬂuence, even though the correctability of the items corresponded fairly closely to Crain & Fodor’s rule violations, and was measured by the same technique. 6.1.3 A critique of correctability We question whether the deﬁnition of correctability that Crain & Fodor adopt is an appropriate one for its ability to act as a determinant of sentence matching times. For the eﬀect as Crain & Fodor describe it (1987: 138ﬀ) to work, the sentence must be clearly an illegal variant of a grammatical and coherent sentence. For the process to be automatic and obligatory, the error must be immediately recognizable, uniquely identiﬁable and reparable. With this in mind it becomes clear that correctability in this technical sense (we shall refer to it as “unique correctability”) is not a scalable quality. It therefore seems doubtful whether the metric that Crain & Fodor use, subjective judgements of correction diﬃculty in a paper and pencil task, is an appropriate one. A more suitable measure would be the number of diﬀerent corrected versions that subjects produce, and to support the automatic and mandatory correction that Crain & Fodor specify, this number should be one. Some examples may clarify the point. The illegal agreement, case and prepositional selection examples in (176) are uniquely correctable. (176) *In the study Mary were writing letters to her husband *Jake and Gill goes to the circus with their granny *Aunt Jacqueline cut his thumb with a pruning saw *Agatha fell out of the cherry tree and hurt themselves *She put in some lettuce and three slices at cheese *Why didn’t father let I stay up late

Evidence from sentence matching on wh-trace dependencies 203

However, there are many which are equally plainly not uniquely correctable. Consider the following (177) based on word order (from Crain & Fodor’s (7) and (8)): (177) a. b. c. d.

*The coach can’t remember the names of the players all *Who could possibly think those guys intelligent are? *Barbara lost the race is a shame *Is this the coat which that you want to buy?

In (177a), one might reasonably correct by deleting all completely, making all the players, making all the names of the players, or simply adding at to make at all out of it. In (177b), we could delete are or move it. (177c) could be Barbara losing… or That Barbara lost…. (177d) can legitimately have either which or that or neither. Clearly these items fail the test of being uniquely correctable, and thus the mechanism that Crain & Fodor suggest cannot apply to them. We might further argue that a pencil and paper task is an inappropriate instrument because of the generous time scale that it oﬀers. The correct form must be available almost instantly, otherwise it cannot ﬁt into the timescale of sentence matching. Consider the following ungrammatical items from Crain & Fodor which they claim to be more slowly matched because they are correctable (178a,b). (178) a. *Since the plane is overloaded, would for it to take oﬀ be illegal? b. *What kind of sauce did Sam serve pork chops and?

Their correct forms are not readily identiﬁed. An informal test on 5 adults showed that subjects required about 4 seconds to mentally correct them, but the subjects in the sentence matching experiments took just two seconds to perform the task.65 This suggests that correction of them played no part in their matching, and it seems most unlikely that they would fulﬁll the criterion of automaticity. But there are even less likely cases. Crain & Fodor claim that the following (from Freedman & Forster) are “not fully noncorrectable” (1987: 152): (179) a. b. c. d.

*Tony was granted bail that his plan failed *The spy admitted by the judge *The girl behind you the subsequent discussion *Reminds me of your sister soon got boring

65.We asked people to look at the sentences and say “Yes” when they could correct them.

204 Empty Categories in Sentence Processing

While subjects rated these as more correctable than the correctable constraint violations, there can be no question of them being uniquely correctable as part of an automatic process, since their intended content is extremely obscure. Crain & Fodor do not report the number of diﬀerent corrected versions of each that were constructed, but we may be sure that it was several, and that each subject in the correctability experiment will have thought for some time before answering. In the light of this we suggest that the criterion of correctability can only be of relevance to a small proportion of the sentences that have been tested in this series of sentence matching tasks. This oﬀers a new perspective on Crain & Fodor’s correctable constraint violations: of the six examples they give, only one qualiﬁes as being uniquely correctable. It is therefore quite possible that Crain & Fodor’s subjects actually corrected them (if at all) as rule violations. For instance, given Barbara lost the race is a shame, a subject who corrected to Barbara losing the race is a shame instead of That Barbara lost the race is a shame can give us no information about constraint violations because that is not how they interpreted it. Yet the matching time will nevertheless appear under the heading constraint violations. To summarise: while the basic insight that Freedman & Forster’s constraint violations are less correctable than other error types holds true, we have seen that there are considerable gaps in both the mechanism Crain & Fodor put forward and the experiment claimed to provide evidence of it. 6.1.4 Locality and operator-variable binding Crain & Fodor (1987: fn. 19) note that correctability is largely coextensive with global versus local ungrammaticality. Forster & Stevenson (1987) accept it, pointing out that local must be seen in phrase structure terms not surface terms. Freedman & Forster speculated that it was perhaps trace binding that sentence matching is insensitive to. Forster & Stevenson (1987) take up the issue and conclude it to be operator-variable binding, as no particular deﬁnition of locality appears to work, even structural locality. The problem is anaphor agreement (180), which causes eﬀects although it is not local. (180) *She decided on a holiday for each other. *The birds made the nest itself. *John smiled for he knew herself to be innocent

There are no eﬀects on the other hand when the implausibility is dependent upon trace binding from an operator position. This can be seen in the contrast (181a,b): both are implausible but (181a) caused eﬀects and (181b) didn’t.

Evidence from sentence matching on wh-trace dependencies 205

(181) a. b.

?The mayor ordered them to unfurl her ?Who did the mayor order them to unfurl?

Other examples show that it is not only the plausibility of extracted elements that is ignored. While local subcategorization violations (182a) produce signiﬁcant costs, even an overt argument in a gap position (182b) does not. (182) a. *John disappeared Mary during the party b. *Who did Harry believe that John liked Bill?

This oﬀers strong support for the suggestion that it is the structural relationship described as operator-variable binding in PPT which is invisible to sentence matching. This account of the data as operator-variable binding has been criticised: Eubank (1993) argues that the CP projection is invisible to sentence matching, since the level of representation which sentence matching accesses is prior to the projection of CP. However, while intuitively the relevant issue concerns the relationship between the CP and the rest of the clause, Crain & Fodor (1987) argue convincingly that it is implausible to suggest that sentence matching accesses a particular grammatical level and Clahsen et al. (1995 exp 3) show that ungrammaticalities in an embedded CP can occasion eﬀects. Clahsen et al.’s own experiment widens the data set by testing V2 violations in German, assuming V2 verb movement also to be operator driven (following Platzack & Holmberg 1989). The experimental items were main clauses but with the verb in ﬁnal, not second, position, thus constituting a V2 violation (183). (183) a.

Grammatical main clause Jetzt LERNST du Englisch in einem Sprachkurs 1827ms now learn you English in a language.course b. Illegal verb-ﬁnal structure Jetzt du Englisch in einem Sprachkurs LERNST 1867ms now you English in a language.course learn

The results between the two structure types are not signiﬁcantly diﬀerent, suggesting that sentence matching is blind to violations in this structure type too. We are in some doubt how to interpret these results, however, since verb movement in V2 shows certain diﬀerences from unbounded dependencies. This qualiﬁes as a member of the previously identiﬁed group neither within PPT, since it is head movement from I to C and not XP movement to SpecCP; nor in HPSG, where it is thought of as an alternative clausal word order (see Section 1.4.2 for discussion), not topicalisation of an NONLOCAL argument. The reason for this movement type is not well understood in either framework,

206 Empty Categories in Sentence Processing

which makes it diﬃcult to draw any strong conclusions about its status without further experimentation.66 6.1.5 Summary It is possible that there is no issue of principle between the locality and operator-variable binding accounts, but only a question of description. A look at how the relevant data is described in HPSG makes this clearer, for it largely coincides with the range of constructions handled with NONLOCAL feature dependencies. It is not obvious how one might determine whether the real criterion is locality and this structure is uniquely not suﬃciently local for sentence matching to be sensitive to it, or whether it is actually some aspect of this construction which fails to trigger the ungrammaticality eﬀect.67 In fact the choice of

66.One possibility is that subjects interpreted the strings as legitimate verb-ﬁnal (embedded) clauses, and located the ungrammaticality in the initial position of the adjunct. Now this adjunct would standardly be analysed as occupying the operator position SpecCP (PPT) or being topicalised (HPSG) and thus could qualify as being a non-local violation of exactly the same type as those in other studies. One might establish more clearly what is happening here by conducting a parallel experiment on English. I to C movement in questions would in principle be an appropriate structure to test, except that the results are grammatical embedded questions (e.g. (i)). One might however use illegal negative inversion structures where the auxiliary has not raised to C (ii). (i) *Why you are going to Germany? (ii) *Never I have tasted such scrumptious doughnuts. 67.One set of examples which might deconfound locality and wh-trace binding would be subject wh-trace ((i), (ii)). This might also allow us to draw conclusions about the assumption in HPSG that there are no subject extractions. (i) (ii)

*Who John killed Cock Robin? *Who does the song say John killed Cock Robin?

One might also examine negative inversion, topicalisation or tough movement to see if they pattern similarly ((iii), (iv), (v)). (iii) (iv) (v)

*Never have I seen such a ﬁlthy teatowel today! *This behaviour we can never put up with such a mess! *This kind of sonata is easy to play Rachmaninov on a Stradivarius.

Another interesting test case would be negative polarity items (vi). Will the antecedent in an operator position be ignored? Since this relationship is generally attributed to c-command not operator binding, it might help us distinguish between a clausal level locality account and true operator binding blocking.

Evidence from sentence matching on wh-trace dependencies 207

description makes little diﬀerence for our purposes, since the consequences for processing and syntactic description are unaﬀected. In the presentation of our experiment we shall refer to the contrast as non-local vs. local ungrammaticality because these are convenient terms. We have outlined above possible support that the sentence matching evidence could oﬀer both to the idea that the parser has an internal modular structure, and also to the notion that the bounding applied to non-local dependencies is diﬀerent to that which applies elsewhere. However there are a number of possible reservations and qualiﬁcations about the validity of this evidence (see also Stowe 1992). First, the eﬀects are relatively unstudied and still somewhat controversial. For example Eubank (1993) found no clear eﬀect of the Speciﬁed Subject Condition, and though Clahsen et al. (1995) are able to show that Eubank’s experimental and control sentences were imperfectly controlled, the existence of the eﬀects is still not universally accepted. For this reason there is a need for more studies to be undertaken to demonstrate that the eﬀects are robustly replicable. Second, other languages and structures need to be looked at to sharpen our understanding of what is happening. It is not yet possible to identify with conﬁdence the set of constructions that are aﬀected, and the exact mechanism is poorly understood. Third, some eﬀect of correctability can still not be ruled out, as we have argued above that the measure of correctability previously employed is an inappropriate instrument to measure it. Therefore structures need to be found and tested which are uniquely correctable but for which the lack of ungrammaticality eﬀect holds. It is these issues which we address with our own experiment reported below.

6.2 Experiment 6: Matching multiple wh-questions The structure we tested in this experiment, multiple wh-questions, is a useful extension to the data set, since it allows us to control for correctability and exclude this account of the eﬀects found. Multiple wh-questions permit this as they require only a simple reordering to become grammatical. In addition they are also readily identiﬁable as an erroneous variant of a single grammatical form, which is our own preferred measure of correctability. The experimental materials are illustrated in (184a–c).

(vi)

*Over my dead body will that wretched postman never darken my doors again!

208 Empty Categories in Sentence Processing

(184) a.

Grammatical baseline Wem meint der Archäologe hat er welche antiken Münzen who.dat says the archaeologist has he which ancient coins verkauft? sold b. Non-local ungrammatical *Wem meint der Archäologe welche antiken Münzen er who.dat says the archaeologist which ancient coins he verkauft hat? sold has c. Local ungrammatical *Wem der Archäologe meint hat er welche antiken Münzen whom.dat the archaeologist says has he which ancient coins verkauft? sold

The standard grammatical form of the multiple question is (184a). Both wh-items unambiguously belong to the embedded clause, but one has been cyclically raised into initial position in the matrix clause, while the other has remained in situ. (184b) is the non-locally ungrammatical version, in which the second wh-item has also moved into initial position of its clause. PPT excludes this by the assumption that the ﬁrst wh-item leaves behind a trace in the embedded SpecCP, preventing the second wh-item from occupying it. Example (184c) is the locally ungrammatical control version. In this condition we used analogues of three of the four diﬀerent forms of ungrammaticality chosen by Freedman and Forster (1985), namely: agreement, quantiﬁer placement, and illegal VP movement. Their last type of ungrammatical sentence was non-sentences made up of two legal phrases not constituting a legal whole. This last is not possible to represent as a form of a sentence as it is by deﬁnition a non-sentence, so it was not used. The three diﬀerent ungrammaticality types are illustrated in (185a–c). (185) a.

Verb-subject inversion *Wem der Archäologe meint hat er welche antiken Münzen who.dat the archaeologist thinks has he which ancient coins verkauft? sold “Who does the archaeologist think has he sold which ancient coins?”

Evidence from sentence matching on wh-trace dependencies 209

b. Quantiﬁed VP topicalisation *Wem meint der Archäologe welche antiken Münzen verkauft who.dat thinks the archaeologist which ancient coins sold hat er? has he c. Agreement errors *Wem meinen der Archäologe hat er welche antiken Münzen who.dat think.pl the archaeologist has he which ancient coins verkauft? sold

In (185a) the matrix subject and verb have been inverted, in (185b) the embedded VP from the quantiﬁed object onwards has been topicalized in violation of pied-piping restrictions in German, and in (185c) the number of the matrix verb and subject do not agree. The use of three diﬀerent types of ungrammaticality should allow conclusions to be drawn about diﬀerent violation types on matching times. 6.2.1 Materials The hypothesis to be tested is that subjects will match non-local ungrammaticals (184b) as fast as grammaticals (184a), but that they will require longer to match local ungrammaticals (184c). However, since the two diﬀerent ungrammaticalities are produced by slightly diﬀerent clause conﬁgurations, we also need to ﬁnd out the cost in matching time of these diﬀerences. Notice that (184a) and (184c) have the embedded verb in the V2 position, whereas (184b) has it in ﬁnal position; on the other hand (184b) has wh-movement in the embedded clause, while (184a) and (184c) do not. In order to demonstrate that it is not these structural diﬀerences which are causing any eﬀects, the cost of these must be factored out. This could be achieved by presenting the sentences in ﬁve diﬀerent forms (186a–e), following Freedman & Forster (1985). (186) Experimental items a. Grammatical baseline Wem meint der Archäologe hat er welche antiken Münzen who.dat says the archaeologist has he which ancient coins verkauft? sold

210 Empty Categories in Sentence Processing

b. Non-local ungrammatical *Wem meint der Archäologe welche antiken Münzen er who.dat says the archaeologist which ancient coins he verkauft hat? sold has c. Local ungrammatical *Wem der Archäologe meint hat er welche antiken Münzen who.dat the archaeologist says has he which ancient coins verkauft? sold Factoring out items d. Grammatical with scope marker Was meint der Archäologe wem er welche antiken Münzen SCO says the archaeologist who.dat he which ancient coins verkauft hat? sold has e. Grammatical with complementizer Wem meint der Archäologe daß er welche antiken Münzen who.dat says the archaeologist that he which ancient coins verkauft hat? sold has

Example (186d) is a grammatical form of this question but, instead of cyclic raising of the dative who into the matrix clause, it has merely a scope marker (= SCO) in this position. (186e) is again the same basic sentence but with a complementizer introducing the embedded clause, which licences the verb in clause-ﬁnal position. These extra items will permit us to ﬁlter out the costs of the diﬀerent word orders.68 (187) a. b. c. d. e.

Grammatical baseline Non-local ungrammatical Local ungrammatical Grammatical with scope marker Grammatical with complementizer

(Wh+ V2+)

(Wh− V2−) (Wh+ V2−)

68.It will be apparent that our assumption that there is an additional cost of movement operations owes much to the derivational theory of complexity (e.g. Miller & Chomsky 1963: 481ﬀ) and the concept of canonical sentence forms. While conscious of its weaknesses, we follow previous work in the sentence matching, as it provides an easy shorthand for describing the structural diﬀerences that we wish to compensate for.

Evidence from sentence matching on wh-trace dependencies

The matching time of (187d) subtracted from (187e) will give us the additional cost of a wh-movement. We may subtract the matching time of (187e) from (187a) to ﬁnd the cost of V2. Having thus established the cost of wh-extraction and V2 verb placement, we can feed these ﬁgures into the comparisons of (187a), (187b) and (187c) to remove the eﬀects of these structural diﬀerences. However, this design requires ﬁve groups of subjects, since no subject can see the same sentence in more than one condition, and it necessitates a large number of ﬁllers, since all the sentences are highly complex multiple wh-questions. We therefore adapted this design to achieve the same results more economically. Since we are performing essentially two separate comparisons, we changed the nominal lexical items in the factoring-out conditions, and added a new relexiﬁed baseline condition parallel to (187a). Since the second wh-item plays no role in these conditions, it was removed, thus reducing the number of highly marked sentences that subjects are presented with. Notice that the addition of a semantically shallow denn (“then”) to (188d) balances the addition of was to (188e) and daß to (188f). (188) Experimental items a. Grammatical baseline Wem meint der Archäologe hat er welche antiken Münzen who.dat says the archaeologist has he which ancient coins verkauft? sold b. Non-local ungrammatical *Wem meint der Archäologe welche antiken Münzen er who.dat says the archaeologist which ancient coins he verkauft hat? sold has c. Local ungrammatical *Wem der Archäologe meint hat er welche antiken Münzen who.dat the archaeologist says has he which ancient coins verkauft? sold Factoring out items d. Relexiﬁed baseline Wem meint der Theologe hat er denn die Schriftrollen who.dat says the theologist has he then the scrolls verkauft? sold

211

212 Empty Categories in Sentence Processing

e.

f.

Grammatical with scope marker Was meint der Theologe wem er die Schriftrollen verkauft hat? SCO says the theologist who.dat he the scrolls sold has Grammatical with complementizer Wem meint der Theologe daß er die Schriftrollen verkauft hat? who.dat says the theologist that he the scrolls sold has

Examples (188a–c) are the experimental sentence types as before, but the factoring out is performed by comparing (188d), (188e) and (188f). The new baseline condition (188d) has undergone two wh-movements and has the embedded verb in the V2 position, the next has one fewer wh-movement and the verb in ﬁnal position, while the third has again two wh-movements like (188d) but a ﬁnal verb like (188e). We therefore have: (189) a. b. c. d. e. f.

Grammatical baseline Non-local ungrammatical Local ungrammatical Relexiﬁed baseline Grammatical with scope marker Grammatical with complementizer

(Wh+ V2+) (Wh− V2−) (Wh+ V2−)

The cost of a wh-movement is thus (189f)–(189e), and the cost of V2 is (189d)–(189f). This division of the materials into two sets allows us reduce the number of experimental groups to 3, as each subject may see both one of the (188a–c) set and one of the (188d–f) set. One hundred and ten ﬁller items were also constructed. These were all questions with one or two wh-items. Of these, ten served as overt practice materials, another forty constituted the ﬁrst phase of the experiment to allow subjects to gain familiarity with the sentence types and task, and the remaining sixty ﬁllers were mixed in a pseudo-randomized order with the experimental items. Overall, half the sentences diﬀered by one word at varying positions in the second appearance, but these were exclusively ﬁller items. Cutting across this category, nearly half the items were ungrammatical. 6.2.2 Procedure Thirty sentence sets were produced in the six conditions described above, so that every subject saw ten items in each condition. Subjects were randomly divided into three groups, each of which saw one of the three counterbalanced sets of the materials. The experimentation took place in a dedicated room

Evidence from sentence matching on wh-trace dependencies

without disturbance. Subjects were tested singly and wore ear muﬀs. A sentence was presented at the top left of the 17” monitor and the same or a minimally diﬀerent one was presented two seconds later oﬀset to the right at the bottom of the screen. Subjects were instructed to determine whether the ﬁrst and second sentences were identical, and respond with the appropriate button press. All experimental sentences were presented as matches. Thirty three subjects participated for payment, 17 females and 16 males, mean age 26.2 years, all students at the Heinrich Heine University in Düsseldorf. The data was treated as follows: all erroneous responses were excluded (140 data points) and those exceeding two standard deviations of sentence matching time by condition (69 data points). The remainder of the data was converted into logarithms for input into anova calculations to reduce skewedness, but the raw data will be reported in the text. 6.2.3 Results Sentence matching times by condition are reported in Table 19. Table 19.Mean sentence matching times for experimental conditions (n = 865) Type

Condition

Mean sentence match- Errors and excluded ing times outliers

a. b. c.

Grammatical Non-local ungrammatical Local ungrammatical

2906 2878 3084

30 41 54

Analyses of variance on the main experimental conditions revealed main eﬀects of Subject (F = 68.7, p < 0.001), Item (F = 3.6, p < 0.001), and Condition (F = 9.7, p < 0.001). There were no signiﬁcant interactions. Subsequent pairwise tests revealed signiﬁcant diﬀerences between the grammaticals and the local ungrammaticals (F=14.3, p<0.001), and the non-local ungrammaticals and the local ungrammaticals (F = 14.0, p < 0.001), but not between the non-local and grammaticals (F = 0.04, p > 0.8). These results resemble the previous ﬁndings of no ungrammaticality eﬀect due to non-local violations. However, since the structures are not wholly parallel, it is necessary for us to factor out the eﬀects of these structural distinctions before we can draw any ﬁrm conclusions. We therefore examine the sentence types designed to control for these diﬀerences, shown in Table 20.

213

214 Empty Categories in Sentence Processing

Table 20.Sentence matching times for relexiﬁed factoring out conditions (n = 895) Type

Condition

Mean sentence matching times

Errors and excluded outliers

d.

Grammatical (Wh+ V2+)

2851

30

e.

Grammatical with scope marker (Wh− V2−)

2816

36

f.

Grammatical with complementizer (Wh+ V2−)

2801

29

An analysis of variance on this data produced main eﬀects for Subject (F = 77.6, p < 0.001) and Item (F = 7.1, p < 0.001), but no signiﬁcant eﬀect by Condition (F = 1.2, p = 0.3). Nevertheless we calculated the cost of a wh-movement from (f−e = −15ms) and a V2 verb position from (d−f = 50ms). Since our non-local ungrammatical type had an additional wh-movement within the embedded clause, we add 15ms to its matching times; since the grammatical and local ungrammatical types had the verb in V2 position, we subtract 50ms from their scores. Our revised experimental results are thus as in Table 21. Table 21.Adjusted sentence matching times for experimental conditions (n = 865) Type

Condition

Mean sentence matching times

Errors and excluded outliers

a.

Grammatical

2856

30

b.

Non-local ungrammatical

2893

41

c.

Local ungrammatical

3034

54

The eﬀect of the compensation for structural diﬀerences is to reduce the diﬀerence between the non-local and local ungrammaticals from 206ms to 141ms. We resubmitted the adjusted ﬁgures to statistical analysis and found a reduced but still unambiguous main eﬀect for condition (F = 7.7, p < 0.001). Pairwise tests showed the same pattern as before: signiﬁcant diﬀerences between grammaticals and local ungrammaticals (F = 14.3, p < 0.001), and nonlocal ungrammaticals and local ungrammaticals (F = 5.3, p = 0.022), but no signiﬁcant diﬀerence between grammaticals and non-local ungrammaticals (F = 3.2, p = 0.076). The factoring out process thus does not alter the basic pattern in our raw results but excludes the variable of structural diﬀerences.

Evidence from sentence matching on wh-trace dependencies

6.2.4 Discussion The ungrammaticality eﬀect for locally ungrammatical sentences shows that the sentence matching task is sensitive to grammatical violations, as the local ungrammatical sentences are matched 178ms more slowly than the grammatical baseline sentences. Three diﬀerent sorts of ungrammaticality were used, but there was little variation between them. The mean score for matrix subject-verb inversion was 3062ms, for illegal VP topicalisation 3042ms and for agreement errors 3132ms, and there were no signiﬁcant diﬀerences between them (F = 0.285, p = 0.75). It comes as a surprise therefore that the non-local ungrammatical sentences are matched considerably faster and pattern together with their grammatical equivalents, from which they are only marginally diﬀerent. In this ﬁnding we replicate the results of previous studies using this methodology investigating other types of non-local ungrammaticality in English. However, due to syntactic requirements, our three experimental sentence types actually vary slightly in their word order, and it might be argued that it is these diﬀerences which caused the dissociation of the results from local and non-local ungrammatical sentences. However, our factoring out conditions allow us to put a ﬁgure on the costs in matching time related to these ordering variations, and to adjust our data to compensate for them. The eﬀect of this is to reduce the observed diﬀerence between the local and non-local ungrammatical sentences, but not to eliminate it. While a part of the diﬀerential eﬀect may have been caused by the structural diﬀerence, it is clear that the larger part cannot be attributed to this and must be explained by other factors. An alternative account of our data might be in terms of correctability. Recall that Crain & Fodor (1985b, 1987) suggested that it was the unique combination of being readily interpretable but hardly correctable that allowed non-locally ungrammatical sentences to be quickly matched, arguing that ungrammatical but correctable violations were slowed because subjects automatically correct them, which causes delay when they subsequently seek to match their internal corrected representation with the second presentation of the ungrammatical sentence. This possibility has already been controlled for by Forster & Stevenson (1987) and Clahsen et al. (1995) using the pen and paper measure of correctability that Crain & Fodor themselves adopted. We have argued above that this is an inappropriate instrument to test correctability, and that for the process to function as Crain & Fodor describe it, a stricter criterion of unique correctability must be applied. To be uniquely correctable, an ungrammatical sentence must be unambiguously related to a particular

215

216 Empty Categories in Sentence Processing

immediately identiﬁable grammatical form; no other relationship can trigger the automatic correction on which the account is based. Both our non-local and local ungrammatical sentences fulﬁll this criterion, for their violations never alter contentive lexical items, nor move any item out of its subcategorizing clause. Each ungrammatical sentence is thus equally uniquely identiﬁable as an erroneous version of a single particular sentence, and in this way all are equally correctable. This criterion, therefore, cannot hold the key to the diﬀerences in matching time between them. Our results support the view that the sentence matching task is sensitive to local ungrammaticality and not to non-local ungrammaticality. Let us note here a problem for this conclusion, namely that we are unable to advance any independently motivated mechanism for the non-local grammaticality blindness eﬀect that we posit. One possibility is that subjects access a speciﬁc level of linguistic representation where the majority of the computational process has taken place, but a subset of it has not. If, for example, one assumed within PPT that the binding of wh-trace and similar processes occurred at LF, one might argue that the representation of the sentence that the sentence matching task accesses is that immediately preceding LF. However it is by no means clear why subjects in this task should have access to a representation which is an intermediate stage of an automatic and unconscious process, nor does the set of constructions which are invisible to sentence matching correspond to any level argued to exist, since even in a derivational theory such as PPT there is no unique level at which all processes have occurred except this one. We follow Crain & Fodor (1987) in considering this option implausible. Another possibility is that the eﬀects found are simply the result of a chance interaction of tolerances within the parser. On this view there is no absolute lack of an ungrammaticality eﬀect, but merely a very weak one. Since the degree of ungrammaticality of the non-local ungrammaticals is typically slightly lower than that of the local ungrammaticals (cf. Crain & Fodor 1987:149), we should only expect a small eﬀect anyway. In addition, let us note that the processor has the capability to partially ignore any sort of ungrammaticality in the search for incremental comprehension, a supposition which is reliably supported by our ability to understand sentences even with quite serious errors. It is possible that we are better able to ignore these non-local error types than we are more local ones, and that this reduces the grammaticality eﬀect again. On this view, therefore, the diﬀerence is not an absolute one but merely one of degree. This account too has weaknesses, however. It is ﬁrstly very unspeciﬁc and untestable. It also fails to correspond to the subjective evidence about happens when one

Evidence from sentence matching on wh-trace dependencies 217

matches local and non-local ungrammatical sentences. When encountering a local error, one is pulled up short with a bump. The high-speed scan of the sentence stops and one is forced to examine the error location. There is no such eﬀect in the non-local ungrammatical type because there is no point at which the scan is halted. This account oﬀers us no explanation of this admittedly subjective experience. In addition, this account would predict a range of values for ungrammaticalities with varying local and non-local error types producing a continuum of values, but the data does not show this. On the contrary, the evidence shows a remarkable uniformity in the eﬀects found for non-local ungrammaticality and a clear dissociation from local ungrammaticality. A more likely candidate is that there is something of the overgenerate-andﬁlter character to the functioning of the parser. This need not necessarily reﬂect the distinction drawn in Government and Binding theory between rules and ﬁlters, but might just imply that the parser has more than one module, as has been suggested by a number of researchers (see discussion in Chapter 3 above). In a comprehension device with this structure, the input would be subjected to multiple processes and in certain circumstances the module responsible for resolving the relevant non-local dependencies could be either switched oﬀ or its output ignored. Now this too is admittedly not a suggestion that has independent motivation, nor would it have been made if the sentence matching evidence had not come to light, for it too leaves questions unanswered. One might ask how it is that just this module and no other is subject to something like conscious control and why the processor has developed such a capability which has no obvious use or purpose. Nevertheless, it does seem feasible that diﬀerent modules should have varying degrees of obligatoriness in the application to the input. One might imagine a layout of information ﬂow where the non-local dependency module was a separate subroutine of the parser, perhaps as illustrated in Figure 21. In such a model, the input is given some sort of initial parse, perhaps using categorial and subcategorization information, and at a later stage the semiprocessed signal is tested for the existence of non-local dependencies. If any are found, the signal enters the non-local dependency processor loop, where the relevant computation is performed, and is returned to the same point in the overall cycle of sentence processing. If there are no dependencies of the relevant sort, the subroutine is by-passed. While this is a very sketchy model lacking many details, the outline should be clear. It is imaginable in such a system that in certain circumstances this subroutine is by-passed in spite of the presence of unresolved dependencies, for example when the subject’s aim is not really to

218 Empty Categories in Sentence Processing

Figure 21.The structure of the non-local dependency subroutine

comprehend the input sentence, but merely make use of their ability to structure it in order to compare it more quickly. It remains somewhat surprising that such a capacity should be available, but of the available mechanisms for the eﬀects found, it seems to us that the assumption of a modular parser is the best founded and the most consistent with the data. Let us note some objections which might be raised to any strong conclusions based on results such as ours. Firstly, it could be argued that the method applied to control for structural diﬀerence is ﬂawed. It makes the assumption that the reading time cost of any structural diﬀerence such as a wh-movement is a numerical constant, which seems unlikely. In addition it assumes that these constants can be added together and the additional processing of each will be performed successively, not simultaneously. A possible reply to this would be to accept that there are imperfections in the materials, but point out that the procedures adopted control for structural diﬀerences as well as is currently possible. Perhaps a more dangerous criticism is that the data can be argued to be showing only a small slowing for non-local ungrammaticality, but not none. This could be due to the interruption of the scan of the sentence in the local ungrammaticals, which does not occur in the non-local ungrammaticals. In the latter, the fact that the sentence is ill-formed becomes increasingly likely as the sentence is read; remember that subjects are exposed to a large number of ungrammatical sentences in the course of an experiment, and so the subject knows that this is quite probable. Over the course of the parse therefore, the subject is able to compute the ungrammatical structure of the sentence and discount the violation when it predictably arrives. This is not possible for the local ungrammaticals since the location is unpredictable and all the processing to deal with the ungrammaticality must occur at the location of the violation before the remainder of the sentence is parsed. This alternative account is not

Evidence from sentence matching on wh-trace dependencies 219

without its own problems; there are examples of non-local ungrammaticals which do not seem to allow any warning or thinking time. For example, Forster & Stevenson (1987) report no slowing in matching time in examples like (190) even though there is no clue to the ungrammaticality until the last word. (190) *Who did Harry believe that John liked Bill?

A ﬁnal criticism might be that it is too early to draw any strong conclusions since we have only a partial understanding of the mechanism which causes the eﬀect. This cautious approach is not unreasonable, but lack of certainty as to the mechanism should not prevent us from discussing the implications of our results, while accepting the need for further work. 6.2.5 Conclusions In this experiment our research aim was to clarify a question concerning the mental representation and processing of non-local constituents and the dependencies which they head. PPT assumes that ﬁllers occupy a speciﬁc position in the phrase structure conﬁguration and that they are represented at their canonical location by a trace. The syntactic dependency between these is one of binding, which forms a separate module within the grammar. In processing terms, this can be seen as a structure-checking operation on already constructed clausal representations, and necessitates a modular parser. In addition the PPT approach makes use of diﬀerent concepts of locality for diﬀerent dependency types, including one for non-local dependencies which relates any position in a clause to the initial “operator” position in that clause. The alternative view has the information about ﬁllers integrated into the local phrasal information, constraining the parsing of the adjacent constituent throughout the dependency. The grammaticality of a non-local dependency is thus as local as all other grammatical relationships. This approach is strictly synchronic and would demand that a parser implementing it be incremental, constructing at any one time all structure formed by the lexical features already encountered. It would not permit a modular parser, or one in which possibly illegal structures could be projected and only subsequently checked for legality. The contribution of the data from sentence matching is to throw some doubt on the empirical adequacy of this second model. For, lacking any obvious alternative account, the suggestion that sentence matching is sensitive to local grammaticality but not non-local grammaticality, however this might be deﬁned, would appear to hold. The consequence of this must be to undermine

220 Empty Categories in Sentence Processing

models of ﬁller dependencies which do not incorporate such distinctions. There are a range of further implications of this result for some parsers proposed in the literature. First, it would seem to show that structure building and structure checking operations can be distinguished, at least optionally. This would support at least some degree of internal modularity in the parser (e.g. Crocker 1996) and put in question any model of a homogenous interactive parser (e.g. McClelland & Kawamoto 1986). Next, it would seem to show that constraints on movement are not integrated into the parser and used on-line (e.g. Fodor 1983, 1985), but rather applied separately at a separate stage of the parsing process (e.g. Frazier 1985). But this result also has consequences for the nature of the embedded grammar. It would appear, for example, not to be consistent with a grammar which is strictly local in its dependencies. HPSG, for example, cascades NONLOCAL information down in a series of local feature inheritances; NONLOCAL dependencies are in fact strictly local in their implementation. It is non-local, not local, ungrammaticality that sentence matching is insensitive to, but this does not exist in HPSG: there are no non-local constraints in the relevant structures which might be violated. It would require some assumptions to be changed to make HPSG consistent with this data. There are also implications for the structure of theories of grammar. It is clear that PPT as a modular and derivational theory can readily allow some processing asynchrony between the computation of the underlying phrase structure and subsequent operations upon it, such as binding, since these are located in diﬀerent modules of the grammar. Such a disjunction is inconsistent with the structure of HPSG, in which the computation of a structure-share is merely an indivisible subpart of the computation of the constraints which each projection imposes upon its neighbour. The evidence we presented here does not of course force the immediate abandonment of mechanisms such as the Nonlocal Feature Principle, but can merely contribute one source of evidence, when it is weighed against its competitors. In this way our study can contribute towards a resolution of the debate about the most adequate description of non-local dependencies.

AUTHOR ""

TITLE "Conclusions"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Chapter 7

Conclusions

The intention of the studies we have reported here was to test certain hypotheses about the processing of structures containing extraction dependencies. We addressed three major questions. The ﬁrst concerned the experimental evidence which has been argued to provide empirical conﬁrmation that the syntactic construct of movement trace has psychological reality. We attempted to tease apart the predictions of the diﬀerent accounts of this data set and establish whether there is any data which can be uniquely attributed to trace reactivation. The second issue we addressed was the empirical evidence for the ECs NP-trace and PRO. The ﬁnal aspect of EC processing we discussed concerned the question of the structure of the parser, and whether there is evidence that unbounded dependency resolution constitutes a separate module. We review our experimentation and results on each of these questions in turn. There is a signiﬁcant body of evidence which is held to be experimentally captured reﬂexes of the processing of trace; most commonly this consists of data which can be interpreted as showing that an antecedent is accessed and mentally reactivated at a trace site. This would support the assumption that ECs exist within human linguistic competence, since this would imply that they behave in the same way as overt anaphoric elements. This construal of such eﬀects we have termed the Trace Reactivation Account (e.g. Swinney et al. 1989). However there are alternative interpretations of the same data set which attribute the reactivation of antecedents to other causes. The Direct Association Hypothesis (Pickering & Barry 1991) argues that the eﬀects found are in fact due to the association of the argument with the verb, and that the confusion of this with trace reactivation is due to the position, adjacent to the verb, at which direct object traces are posited. Ivan Sag and Janet Fodor (e.g. Sag & Fodor 1995) have repeatedly suggested that the eﬀects noted might be a result of semantic not syntactic processing, and that since this cannot be ruled out, the data cannot be held to constitute evidence for a syntactic construct. We referred to this as the Semantic Processing Account. The third alternative account of the antecedent reactivation data is based upon insights about the eﬀect that longer

222 Empty Categories in Sentence Processing

processing of a constituent has upon the speed at which it is recognized as a probe (Fodor 1995, p.c.). It has been shown that a constituent which undergoes more complex computation is better recalled than others from the same structure. Additionally, there is some suggestive evidence that this diﬀerence is magniﬁed at potential clause boundaries, perhaps as a result of extra clausal processing at this position. Since these potential clause boundaries frequently coincide with trace positions, the Depth of Processing account holds that presumed trace reactivation eﬀects are in fact due to summative computation at the potential clause boundary. The antecedent reactivation data can thus be interpreted as providing support for the use of ECs in parsing or alternatively as reﬂecting other separate parsing phenomena. Since these diﬀerent accounts cannot be readily distinguished on the basis of English data, experimental results can currently only oﬀer limited support to the syntactic construct of ECs. We therefore attempted to dissolve the confound using German materials.

Experiments 1–3 Our experiments one to three approached this problem using the cross-modal lexical priming methodology. We constructed sentences containing object scrambles in V2 structures in order to distinguish between the alternative accounts. This was possible because this structure produces an extraction site which is not adjacent to the subcategorizing verb (191). (191) Die Frau gab ihre Münzei dem Mädchen ti the woman gave her coin the.dat girl

In Experiment 1 we tested a sentence-ﬁnal trace position, in Experiment 2 a trace position preceding a sentence-ﬁnal particle, and in Experiment 3 a trace position in a topicalized VP. There was no eﬀect of antecedent reactivation in our Experiment 1 but signiﬁcant eﬀects in Experiments 2 and 3. We argued that there are both syntactic and processing reasons to think that the Experiment 1 result is the least reliable of the three, and we therefore based our conclusions on the results of Experiments 2 and 3. In these two experiments we were able to contrast the predictions of the four accounts by testing at locations where the TRA, but no other account, would lead us to expect activation to be found. The data clearly suggests that the mechanism of trace reactivation does indeed account for some of the experimental data, since no other account predicts the

Conclusions 223

range of eﬀects observed. This result could lead us to pay greater attention to grammatical and processing models which make use of trace, since their empirical adequacy receives support. Additionally, this result suggests that short object scrambling, as in our experimental materials, is indeed what is analysed in PPT as a type of extraction and not simply base generation of arguments in an non-canonical order. However, it is advisable to consider the precise import of this result before drawing strong conclusions. Let us ﬁrst note that none of the alternative traceless accounts of reactivation has been disconﬁrmed; they may all still hold and account for a proportion of the published results, even if they do not appear to account for this one. We cannot show that the mechanism involved in our ﬁnding was necessarily also responsible for all previous results showing antecedent reactivation. Nor can we demonstrate that the cause of the eﬀect we found was indeed reactivation by a covert anaphor; we can only state that of the accounts known to us, only this one would predict the pattern of results we found. It is possible that some other cause is responsible for our ﬁndings, perhaps one speciﬁc to the somewhat marked scrambling structures that we tested. It should be remembered that the very status of scrambling as a movement construction is controversial even within a derivational framework such as PPT, and that therefore even many PPT syntacticians would be surprised by evidence of trace in such structures. Further data would be required, perhaps from other types of scrambling or other languages, in order to allow ﬁrm conclusions to be drawn. In spite of these necessary equivocations, we may state that this result oﬀers substantial support to the idea that the antecedent reactivation can be attributed to the construction of covert constituents at gap positions. The evidence is persuasive since the Trace Reactivation Account predicts the result while none of others can even oﬀer an explanation. We have mentioned above some reasons to be cautious, but these are not really speciﬁc to this result but are rather constraints on all scientiﬁc enterprise: psycholinguistics is a good example of Popper’s dictum that there is no conﬁrmation in science, only ever falsiﬁcation. The view that there is some representation of displaced elements at their canonical location is strengthened by our ﬁnding, since it becomes more likely that it has not merely theoretical but also empirical underpinning. The implication of this result for HPSG is that it needs to consider again the traceless mechanism for dealing with the foot of non-local dependencies, essentially the subcategorisation of locally absent arguments. This approach is descriptively adequate and is from some perspectives superior in explanatory force to the

224 Empty Categories in Sentence Processing

PPT equivalent. It is, for example, more theoretically parsimonious in that it avoids the necessity of assuming lexemes which are unrepresented in the phonetic input. However, this traceless mechanism does not predict the robust evidence of a processing reﬂex at extraction sites. There are a number of responses that an HPSG syntactician might make to this. One possibility would be to reintroduce movement traces, since these are virtually an optional feature of the framework. Another might be to argue that the data represents a parsing phenomenon not a grammatical one, and that a syntactic theory does not need to represent every aspect of mental processing. After all, a grammar is a model of the structure of linguistic signs, not a description of the way that humans arrive at an analysis of a given input. Just as principle-based parsers implementing PPT have to permit predictive parsing at the subcategorizer, HPSG can permit some checking at the gap position of the dependency formed at the subcategorizing verb. This approach would locate the experimental eﬀect inside the parser but outside the embedded grammar. There is however one more possible response, and this is in our opinion the strongest. It consists of building on our arguments in Section 5.6.2 that HPSG too has a representation of an extracted element. We showed that the (covert) category on the SUBCAT list of the subcategorizer is syntactically real and active, taking part in grammatical relations just as a sign does, and that it has a (potential) deﬁnite location in the phonetic form. Given this, it can be plausibly argued that even traceless HPSG can already account for the data showing antecedent reactivation at gap positions. A parser implementing HPSG might well, at the gap position, check on the SUBCAT list to see what constituent is expected, note that this constituent is non-locally subcategorized for under NONLOCAL, and perform the necessary computation. Now it seems quite feasible that this computation, taking place as it does at the temporal position of the gap, could be the cause of the reactivation visible to CMLP. Let us look again at how this is eﬀect might be made up. A verb V has three items on its SUBCAT list ABC which participate in the obliqueness hierarchy for binding reasons. B is also a NONLOCAL value as it is topicalized. On meeting and processing all those elements which the linear precedence rules state should precede B, the parser will next look for B on the SUBCAT list and note the structure-share with a NONLOCAL value. It will then process the NONLOCAL dependency causing measurable reactivation of the antecedent. One might call this ordered NONLOCAL processing. This approach is perhaps the best of both worlds for HPSG as it accounts for the data apparently showing trace reactivation but does not require the assumption of additional ECs in the grammar.

Conclusions 225

Note that ordered NONLOCAL processing rather resembles the SPA in that it is based upon the idea that it is the computation at the gap position which causes the eﬀect, not a phonetically null element in the gap as in the TRA, or some merely coincidental eﬀect as in the DAH and the DOP. It diﬀers from the SPA in that there is no need to make the currently untestable and consequently much weaker claim that this is semantic processing. All that is required to account for increased activation of the antecedent at the gap position is some syntactic representation of it in the structure which identiﬁes its potential temporal position. We have argued that all of this information is already available at the subcategorizer in the form of the SYNSEM value on the SUBCAT list, (potentially) ordered by virtue of its participation in the obliqueness hierarchy. This result has very diﬀerent implications for PPT. ECs are fundamentally rather improbable; so improbable indeed that we must assume that their existence is part of the human genetic inheritance, since it can hardly be imagined that children could learn of them from the input. Having successfully predicted an empirical result on the basis of internal theoretical considerations, PPT syntacticians can feel vindicated. They should not feel too relaxed, however, since the opposite result would have raised very serious questions about the adequacy of PPT. Since frameworks such as HPSG have shown that there is no need for ECs in an adequate syntactic theory, and the assumption of ECs is theoretically unparsimonious, a result showing that supposed trace reactivation data can be accounted for by alternative accounts would deal a real blow to PPT. This is particularly the case because of the multiplication of trace and particularly intermediate trace in current PPT models (see Section 1.2.1). These models depend crucially on traces, but it seems most unlikely that empirical reﬂexes could be found for many of these, in particular those which do not correspond to an intuitive gap. In the light of this PPT syntacticians would be wise to consider whether all of the movement operations they assume are psychologically real. We ﬁnish therefore with a situation where both of our generative grammars can account for our trace reactivation results: PPT with its traces and HPSG via our ordered NONLOCAL processing hypothesis. Let the situation be not misunderstood, however, for ordered NONLOCAL processing is no threat to PPT’s traces. For our discussion in Chapter 5 on which the ordered NONLOCAL processing account is based showed that HPSG uses covert categories which are empirically indistinguishable from PPT’s traces. The ordered NONLOCAL processing explanation of reactivation data is only

226 Empty Categories in Sentence Processing

possible if our arguments about the syntactic status of HPSG’s covert categories go through. Both HPSG and PPT can thus account for the trace reactivation data, each with their own silent categories which we must regard as mere notational variants. Its seems that our study aiming to contrast PPT and HPSG has tended rather to highlight their commonalities.

Experiment 4 In Chapter 5 we considered the evidence for the existence of two local EC types, PPT’s NP-trace and PRO. Since the CMLP data relating to these can fairly readily be thought of as reﬂecting semantic processing or DOP eﬀects, the burden of demonstrating their existence rests ﬁrmly on the data from visual probe recognition studies. This methodology is somewhat controversial since the testpoint is not at the posited EC position and the task performed is a conscious recall operation. We performed a quasi-replication of Bever & McElree (1988, 1989) with German equivalents in order to test whether language-speciﬁc factors could be excluded as an alternative explanation of their results. Additionally we included a subcondition to test the applicability of the DAH to this methodology and also recorded reading times during the selfpaced phrase-by-phrase input to test the predictions of the DOP. None of the four competing accounts we compared in Chapter 4 above was fully supported by the results. The DAH perhaps came oﬀ worst of all, since it was most directly tested for. There was no sign that the proximity of the verb to the testpoint played any role in speeding responses, but the results were equally inconsistent with the TRA, since the only condition with no reactivator, overt or covert, recorded the fastest responses. Since this is precisely the opposite result to that reported by Bever & McElree, this must raise questions about the cause of the eﬀects they report. No part of this result oﬀers any conﬁrmation of the interpretation of their results as evidence supporting the reality of local ECs. The predictions of the DOP are also left unconﬁrmed, since there was no clear correlation between length of processing at the disambiguating segment and probe recognition speed. This is perhaps surprising since the most natural account of Bever & McElree’s data is that the more complex structure always shows the faster response time. While our follow-up experiment, designed to test whether complexity was slowing responses, revealed this eﬀect, the pattern of data across the rest of the experiment does not show this since it also appears that in some places complexity has had the opposite eﬀect and slowed responses.

Conclusions 227

Nevertheless the data does seem to suggest that diﬀerences in complexity can aﬀect response time downstream of the complexity, and that this can be either slowing or speeding, depending on the circumstances. It is tempting to try to link this to the correspondence between our sentence-ﬁnal results and Bever & McElree’s early test point results, in both of which the control condition is fastest. It seems possible that there is a speciﬁc eﬀect causing this correspondence, such as a potential clause boundary causing local slowing due to processing load, but more distant speeding of responses due to depth of processing. One last aspect of this set of results that we shall comment on is that the tough movement condition, generally assumed to contain a wh-trace and included as a reliably primed condition, was no faster than the others, indeed it was among the slower sentence types. Since reactivation at wh-gaps has been robustly shown in other work, this result could lead us to speculate that the visual probe recognition methodology is completely blind to reactivation by ECs and can therefore provide no evidence in either direction. To summarise, while the explication of these results must remain incomplete and some questions unanswered, it does not oﬀer any support to the idea that local ECs play an active role in processing in the same way as our earlier experiments suggested that non-local trace does, but it oﬀers no clear support to any of the alternative accounts either. Nevertheless, our data has contributed to the continuing debate on the processing of extraction dependencies by providing evidence that some accounts of the visual probe recognition data set do not, when tested, seem to be conﬁrmed.

Experiment 5 This further study of raising and equi constructions grew out of consideration of the implications of the previous set of results. Our pattern of results in Experiments 1–4 reﬂects previous ﬁndings, fairly robust experimental evidence for wh-trace but weak evidence for local ECs. Since this pattern of results reﬂects the distribution of ECs in some versions of HPSG but not that in PPT, Fodor & Sag (e.g. 1995) argue that this shows the former framework to be more empirically adequate. We have argued that the basis of this claim, namely that HPSG uses no equivalent of local ECs, does not in fact hold when subjected to closer scrutiny, and that HPSG in fact makes use of covert categories which are empirically indistinguishable from PPT’s local ECs. Given this, we sought an alternative approach to test the two accounts and formulated an ERP experiment

228 Empty Categories in Sentence Processing

which could tell us whether the distinction between NP-trace and PRO drawn by PPT was empirically justiﬁed. We tested German sentences in three conditions, raising, equi and transitive controls and measured brain potentials at the onset of the embedded clause, where subjects had unambiguous evidence of the continuation of the sentence. In two of the three cases, this would lead them to posit an EC. (192) a.

raising condition Der Sheriﬀi schien, [adjunct], NP-tracei den Täter verurteilen the sheriﬀ seemed [adjunct] NP-tracei the oﬀender sentence zu können. to can b. equi condition Der Sheriﬀi hoﬀte, [adjunct], PRO den Täter endlich verurteilen the sheriﬀ hoped [adjunct] PRO the oﬀender at.last sentence zu können. to can c. transitive control condition Der Sheriﬀ erkannte, [adjunct], den Täter endlich im the sheriﬀ recognized, [adjunct], the oﬀender at.last in.the Scheinwerferlicht. spotlight.

The results showed signiﬁcant diﬀerences between all three, but particularly so between raising and equi, where the diﬀerence was plain in three successive 200ms time windows. This result is not predicted by the HPSG analysis of the diﬀerence between raising and equi, since it locates the contrast in the semantic role assignment of the matrix verb. In this way the principled distinction between raising and equi implemented in PPT by the use of diﬀerent ECs is supported. This leaves a confused picture for local ECs. On the one hand it is possible to discount the visual probe recognition results as revealing nothing about the status of local ECs and point instead to the clear diﬀerence in ERP response to raising and equi, and argue this to be solid evidence for PPT’s analysis using local ECs. On this view our Experiments 2 and 3 suggest that the traceless mechanism of non-local subcategorisation for locally absent arguments is of questionable empirical adequacy and Experiment 5 raises questions about this traceless approach in raising and equi constructions too. Since our experimentation has shown measurable processing diﬀerences between them not predicted

Conclusions 229

by HPSG, our results throw some doubt on the treatment of all of these dependencies in HPSG. Alternatively one could emphasize the point that our visual probe recognition result showed exactly the opposite result to Bever and McElree’s, which entails that reactivation by ECs is not the only factor involved in producing diﬀerences between sentence types in visual probe recognition. Not only does our result show no evidence of ECs, therefore, but it additionally demonstrates eﬀectively that previous results using this methodology may also be interpreted without reference to reactivation by ECs. This undermines any claim that visual probe recognition data supports the reality of ECs. One might follow this by maintaining that our ERP result cannot with conﬁdence be attributed to any particular factor because of our patchy knowledge of what a P600 reﬂects. This approach argues that for empirical data to provide any convincing evidence for ECs, it would have to hold equally for all of them, since they are all covert anaphors. Since only a subset, roughly non-subject wh-trace, can be argued to produce reactivation, and it has been shown that a range of factors can produce eﬀects which resemble this, it remains only to identify the real cause of the eﬀects mistakenly attributed to wh-trace for the whole issue of empirical reﬂexes of ECs to be laid to rest. It is no easy matter to decide between these views and a deﬁnitive answer must await further work, but we shall make a few comments here. The visual probe recognition evidence does not disprove the existence of local ECs but merely removes one plank of support for them. What remains is the CMLP data (Nicol & Osterhout 1988; Osterhout & Nicol 1988; Osterhout 1988 [all reported in Nicol & Swinney 1989] and Osterhout & Swinney 1993; see Section 5.2.1) and our own ERP evidence. Neither of these oﬀer anything like proof of a processing role for local ECs but neither of them are null, and both require further investigation. The pool of evidence for local ECs has lost the visual probe recognition data but gained a new recruit in the form of our ERP result. Since the former is the methodologically least sophisticated and the latter the most sophisticated, we must conclude that the case for local ECs has been strengthened by our studies. Let us take the data from our Experiments 1–5 together and try to draw a balance on the issue of evidence for ECs. CMLP Experiments 2 and 3 are strong results supporting the psychological reality of movement trace since these results are not predicted by any traceless account; their only major weakness being that parallel eﬀects were not found in Experiment 1. Our ERP Experiment 5 also oﬀers robust support for the existence of local ECs, since the

230 Empty Categories in Sentence Processing

alternative possible causes of the P600 we found are excluded by our two follow-up experiments. In sum we have a signiﬁcant body of data suggesting that ECs play a role in on-line processing as well as in the grammar. Since this hypothesis is a strong and testable claim, we feel that any evidence in its favour is a surprising fact and should be treated as a success for linguistics.

Experiment 6 In Chapter 6 we reviewed the evidence from one particular methodology which, it has been claimed, is diﬀerentially sensitive to the output of diﬀerent modules in the human sentence processing mechanism. Data from the sentence matching task shows progressive slowing of matching times proportional to increasing distance from grammaticality, with one exception. Illegal wh-trace binding does not show the same diminution in grammaticality and behaves like fully grammatical sentence types. It is diﬃcult to avoid the conclusion that the language processor makes some distinction here, perhaps between local and non-local ungrammaticality. This phenomenon has been argued to support models of parsers which divide the computation into subtasks and assign these to diﬀerent parsing modules. The data would also seem to conﬁrm that grammatical models should have a modular structure, since the diﬀerence is not otherwise capturable except by pure stipulation. However, there are alternative accounts of the phenomenon. One possibility is that the eﬀects are linked to the particular construction and are perhaps language-speciﬁc. Another suggests that the uncorrectability of some examples plays a role. We argued that the oﬀ-line concept of correctability as previously controlled for was an inappropriate measure and put forward our own alternative, unique correctability, which we suggested would be necessary for the correctability account to operate as Crain & Fodor (1987) describe it. In our experiment we utilised materials which control for this revised concept of correctability, for our non-locally ungrammatical sentence types are readily correctable, and have nothing of the ineﬀable quality that Freedman & Forster’s (1985) original examples did. Our experimental sentence types were multiple wh-questions and the non-local ungrammatical sentence types had the second wh-element illegally raised into initial position in the embedded clause.

Conclusions

(193) a.

Wem meint der Archäologe hat er welche antiken Münzen who.dat says the archaeologist has he which ancient coins verkauft? sold b. *Wem meint der Archäologe welche antiken Münzen er who.dat says the archaeologist which ancient coins he verkauft hat? sold has

Our data parallels previous results: sentence matching times are clearly responsive to local ungrammaticality, but show no ungrammaticality eﬀect in the sentence types with non-local violations. This conﬁrms the interpretations of this result as due to a speciﬁc aspect of the structure concerned and disconﬁrms any role of correctability, even on our revised more adequate deﬁnition. It would seem therefore that the binding of wh-trace in these structures can at some level be distinguished from most other grammatical processes. Our sentence matching result thus contributes towards the empirical adequacy of parsing models in which certain non-local processes take place at a later stage of processing than some others. This is consistent with the idea that the parser has both structure-building and structure-checking operations, and that these are performed in sequence. Similarly, it suggests that the division of grammatical models into separate blocks of rules performing diﬀerent tasks may reﬂect some aspect of the way that the grammar is implemented. Lastly, it throws some doubt on the single deﬁnition of local employed in frameworks such as HPSG, and would favour a relativized locality notion. If conﬁrmed, this would suggest that HPSG should revise the cascade-model Nonlocal Feature Principle and adopt a truly non-local mechanism for extracted constituent dependencies. However, it is still necessary to exercise some caution with these sentence matching results since some questions about the validity of the conclusions must still remain. In particular we should note that the mechanism which brings about the observed results is still poorly understood and largely a matter of speculation, and that any conclusions drawn on such a basis are necessarily speculative. The evidence from our results must therefore be taken as only suggestive until we have gained a greater understanding. It cannot on its own provide suﬃciently strong grounds for any strong inferences, but can only contribute to a pool of evidence on the basis of which judgements can be made, when all the indications in each direction are weighed.

231

232 Empty Categories in Sentence Processing

Taking our results together we see a strong pattern of support for a role for linguistic competence in the human sentence processing system, since a nonobvious syntactic prediction, that ECs should form part of the representation of sentential input, receives considerable conﬁrmation. Since it is part of this successful account that these ECs are processed on the basis of syntactic information alone, and this functioning has characteristics not shared by the remainder of the cognitive system, our data is quite consistent with a language faculty which is a separate module of the mind. We also see some evidence favouring a modular internal structure of the processor, with successive structure-building and structure-checking functions. This again is in line with the claims of some grammatical theories and can be seen as a surprising fact successfully predicted by linguistics. No data stands in a vacuum, and such major questions will be answered on the basis of far more evidence than we oﬀer here. Nevertheless, we hope that even our own modest studies of the psychological reality and processing of empty categories can contribute to a better understanding of the position of language in the mind.

AUTHOR ""

TITLE "Appendix 1: Materials for CMLP Experiments 1–3"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Appendix 1 Materials for CMLP Experiments 1–3

Notice that we oﬀer only one translation of each example sentence here for the sake of brevity, but we attempt to give both a reasonably clear idea of the meaning and of the structure with this one translation. Naturally this involves compromises in both the word for word correspondence of the translation with the original and the idiomaticity of the English, but the discussion of the experimental structures in the text should always give suﬃcient detail for the relevant structure to be clearly identiﬁable. Note that (fem) after an person indicates that the German morphologically marks her as a woman, but there is no English equivalent; thus “Lehrerin” is glossed “teacher (fem)”.

Experiment 1: Gap ﬁnal 1

Der Professor erleichtert jedoch die Prüfung den etwas weltfremden Mediävisten, indem er ihnen seine enzyklopädische Kenntnis der Examensordnung zur Verfügung stellt. However the professor simpliﬁes the exam for the somewhat ivory-towered mediaevalists by putting his encyclopaedic knowledge of the exam regulations at their disposal.

2

Auf dem Flensburger Wochenmarkt verkaufte er die ganze Ladung einem dänischen Kaninchenzüchter. At the weekly market in Flensburg he sold the whole load to a Danish rabbit breeder.

3

Verständlicherweise verrät sie das Geheimnis nur ihren zwei erwachsenen Töchtern, die auch in der Küche tätig sind. Understandably she reveals the secret only to her two grown-up daughters who are also employed in the kitchen

4

Die Frau gab ihre Münze dem bitterlich weinenden Mädchen und ging dann leichten Herzens weiter. The woman gave her coin to the bitterly crying girl and went on her way with a light heart.

5

Die neue Bauaufsichtsrätin erteilt den Auftrag dem billigsten Bauunternehmen, und damit hat es sich. The new buildings inspector (fem) awards the contract to the cheapest building company and that’s the end of the matter.

234 Appendix 1: Materials for CMLP Experiments 1–3

6

Sie übergab das Paar einem unbewaﬀneten Vermittler, und die zwei erreichten sicher das Terminal. She handed.over the pair to an unarmed mediator, and the two reached the terminal in safety.

7

Gerade reich ist sie auch nicht, aber sie ﬁnanzierte den Urlaub ihrem beredsamen Neﬀen. She isn’t exactly rich either, but she ﬁnanced the holiday for her persuasive nephew.

8

Die Äbtissin ihres Klosters erlaubte die Reise der sehr frommen Ordensschwester; allerdings unter der Bedingung, daß sie mit niemandem redet. The abbess of her nunnery permitted the journey to the very pious nun; admittedly on the condition that she spoke to noone.

9

Darum verzieh er diese Lüge seiner ungezogenen Enkelin. So he forgave his naughty grand-daughter this lie.

10 Doch kurz bevor er sich zurückzog, überließ der ehrwürdige Missionar das Amt einem eingeborenen Nachfolger. But shortly before he retired, the venerable missionary left the oﬃce to a local successor. 11 Ein nachsichtiger Hauptmann erlaubte den Besuch dem aufgeregten Unteroﬃzier, obwohl das eigentlich während der Dienstzeit nicht üblich ist. An indulgent captain allowed the excited NCO the visit, although this is not usual during time on duty. 12 Er wollte nicht länger warten und überließ den Brief der Stellvertreterin des Direktors, obwohl das vom Auftraggeber ausdrücklich untersagt worden war. He didn’t want to wait any longer and left the letter with the director’s deputy, although that had been expressly forbidden by the client.

Experiment 2: Particle ﬁnal 13 Die Schreinermeisterin hing den Fehler einem sechzehnjährigen Lehrjungen an, aber keiner glaubte es. The master joiner (fem) put the blame on a sixteen-year-old apprentice, but nobody believed it. 14 Sie bietet jetzt die Wohnung einem langjährigen Arbeiter an, als Belohnung für seine Treue. She is now oﬀering the ﬂat to a worker with long service, as a reward for his loyalty. 15 So gewöhnte sie diese Neigung ihrem jungen Assistenten ab. In this way she weaned her young assistant oﬀ this habit. 16 Eines Morgens, als er etwas spät dran war, steckte sie das Brot dem Leiter der deutschen Delegation zu, ohne daß er sich dessen bewußt war. One morning, when he was running a bit late, she put the sandwich into the pocket of the leader of the German delegation, without him being aware of it.

Appendix 1: Materials for CMLP Experiments 1–3 235

17 Sie lieh das Gerät keinem der vergeßlichen Handwerker aus, denn sie fand, so was sollten sie sich selber besorgen. She lent the machine to neither of the forgetful tradesmen, for she felt that they she get hold of such things themselves. 18 Sie führte die Burg auch dem Landesminister für Bildung vor, als er die Schule oﬃziell besuchte. She showed the regional minister of education round the castle too, when he oﬃcially visited the school. 19 Sie kündigte ihre Aktion nur ihren vertrauten Mitstreitern an, unter denen sofort eine Totenstille ausbrach, als sie fragte, wer sie begleiten wollte. She only let her trusted conﬁdants know of her scheme, amongst whom there was a deathly hush when she asked who wanted to accompany her. 20 Er geht in eine Bank, verlangt DM 3000 oder so, und schon zahlen sie das Geld dem hocherfreuten Trickbetrüger aus. He goes into a bank, asks for DM3000 or thereabouts, and straight away they pay the money out to the delighted fraudster. 21 Schließlich aber wies man den Betrug dem kinderreichen Familienvater nach, und da ﬂog er raus. Finally they proved that the father of the large family was swindling, and he was thrown out. 22 Nach zwei Tagen Streit sprach der Richter das Geschäft dem ziemlich überraschten Andreas zu, und die Entschädigung der Susi. After two days of argument the judge awarded the shop to the rather surprised Andreas, and the compensation to Susi. 23 Die Psychologin verkündet laut, sie wende ihr Talent von nun an der Zauberpraxis zu, was stürmischen Beifall auslöst. The psychologist (fem) proclaimed loudly that she would apply her talent in future to the practice of magic, which triggered a storm of applause. 24 Sie macht die Tür grundsätzlich nur Freunden und Bekannten auf; sogar die Müllabfuhr muß sich ausweisen, bevor sie das Grundstück betreten darf. As a matter of principle she only opens the door to friends and acquaintances; even the dustmen have to show identiﬁcation before they are allowed on her premises.

Experiment 3: VP in SpecCP Notice that the structure of these sentences is quite simply not grammatical in English, and so the translation is rather loose. The word order in the original is always DO > IO > nonﬁnite verb in the topicalized VP (see Chapter 4 for more detailed structural analysis). 25 Diesen Mut dem völlig durchschnittlichen Beamten zugetraut hat aber keiner. “This courage nobody had expected of the quite average civil servant”

236 Appendix 1: Materials for CMLP Experiments 1–3

26 Solche Drogen einem Verdächtigten zuschieben würde kein anständiger Rechtsbeamter. “No decent law oﬃcer would plant such drugs on a suspect” 27 Eine solche Strafe der gesamten Unterstufe auferlegen würde wohl kaum ein anderer Lehrer. “Surely no other teacher would impose such a punishment on the whole lower school” 28 Meine Maschine einem unerfahrenen Neuling zeigen tu ich nicht. “No way will I show my machine to an inexperienced newcomer” 29 So ein Haus einer völlig normalen Kundin verkaufen konnte nur Herr Reibel. “Sell such a house to a quite normal customer? Only Herr Reibel could do it!” 30 Seine Zustimmung der hochrangigen Stadträtin verweigern hieße sich von seiner Karriere verabschieden. “To refuse his agreement to the high-ranking alderwoman would mean saying good-bye to his career”. 31 Die paar Mark einer hochintelligenten Frau verweigern ﬁnde ich schändlich. “I ﬁnd it disgraceful to deny a highly intelligent woman those few Marks.” 32 Den Fernseher einer alten Pfarrerswitwe schenken war für ihn selbstverständlich. “It was natural for him to donate the TV to an old parson’s widow.” 33 Solche Musik meiner konservativen Frau vorspielen wäre doch sinnlos. “It would be pointless to play such music to my conservative wife” 34 So eine Summe dem verdammten Ausbeuter bezahlen ﬁel mir gar nicht ein, also bin ich vor Gericht gegangen. “I wouldn’t dream of paying such a sum to that damned exploiter, and so I went to court.” 35 Diesen Angriﬀ seinen erschöpften Soldaten ersparen schien Tilly nur menschlich und humanitär, weil sie ja schon so schwach und hungrig waren. “It seemed only human and humane to Tilly to save his soldiers this assault, since they were already so weak and hungry.” 36 Die Ausstellung einem unbekannten Künstler widmen ist für die großzügige Sponsorin Ehrensache. “It is a point of honour for the generous sponsor (fem) to devote the the exhibition to an unknown artist.” Below is an example text showing how the experimental sentences were embedded in a coherent context. The words in capitals show the positions at which visual targets were presented; here two ﬁllers and one experimental identical target are included. “»Einmal ist ONKEL keinmal« sagte der Opa, als er entdeckte, daß die kleine Rosi ihn wegen der verschwundenen Bonbons beschwindelt hatte. Darum verzieh er diese Lüge seiner ungezogenen Enkelin LÜGE. Im ganzen Ort kennt man ihn schon seit eh und je als herzensguten und kinder KNALLEN lieben Mann. Manchmal ist er allerdings zu nachsichtig.” “‘It could happen to anybody once’ said Grandad when he found out that little Rosi had lied to him about the sweets which had vanished. So he forgave his naughty grand-daughter this

Appendix 1: Materials for CMLP Experiments 1–3 237

lie. He has always been known in the whole village as a child-friendly man with a heart of gold. Sometimes he is a bit too easy-going, however.” Related and unrelated targets with frequency (CELEX lemma frequency based on a corpus of 6 million written and spoken words) and length: Exp. 1: Pure V2 Freq

Length

Lüge “lie” Ladung “load” Geheimnis “secret” Münze “coin” Auftrag “contract” Paar “pair” Urlaub “holiday” Reise “journey” Prüfung “exam” Amt “oﬃce” Besuch “visit” Brief “letter”

192 83 211 105 602 158 274 273 302 701 1003 847

4 6 9 5 7 4 6 5 7 3 6 5

Mean

396

Freq

Length

Kern “kernel” Spende “donation” Festigung “fortress” Kanal “canal” Tochter “daughter” Tanz “dance” Export “export” Hafen “harbour” Pfennig “penny” Zug “train” Gesetz “law” Wagen “car”

190 81 212 101 576 167 260 273 310 717 1049 860

4 6 9 5 7 4 6 5 7 3 6 5

Mean

400

Exp. 2: Particle ﬁnal Freq

Length

Fehler “mistake” Wohnung “ﬂat” Neigung “tendency” Brot “bread” Gerät “device” Burg “castle” Aktion “action” Geld “money” Betrug “fraud” Geschäft “shop” Talent “talent” Tür “door”

341 903 132 174 337 99 497 1204 54 647 108 739

6 7 7 4 5 4 6 4 6 8 6 3

Mean

436

Freq

Length

Modell “model” Kollege “colleague” Bindung “ligature” Zins “interest” Ebene “plane” Rind “bovine” Kultur “culture” Erde “earth” Bistum “bishopric” Hoﬀnung “hope” Panzer “tank” Arm “arm”

342 943 132 176 340 98 499 1137 54 600 109 731

6 7 7 4 5 4 6 4 6 8 6 3

Mean

430

238 Appendix 1: Materials for CMLP Experiments 1–3

VP in SpecCP Freq

Length

Mut “courage” Summe “sum” Mark “Mark” Maschine “machine” Zustimmung “assent” Haus “house” Strafe “punishment” Fernsehen “TV” Musik “music” Drogen “drugs” Angriﬀ “assault” Ausstellung “exhibition”

253 246 1022 862 335 2000 195 532 717 14 442 99

3 5 4 8 10 4 6 9 5 5 7 11

Mean

560

AUTHOR ""

TITLE "Appendix 2: Materials for probe recognition Experiment 4"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Appendix 2 Materials for probe recognition Experiment 4

Experiment 4 compared sentence-end probe recognition times in six diﬀerent structures. For this task the internal word order variation between structures does not particularly matter, though it is naturally desirable to keep it as small as possible. However, we also wanted to compare reading times per segment, a measure for which diﬀerences between structures matter very much. For this reason we did not construct this materials set by the standard method of producing every sentence in each condition, but tried to reduce the variation between items by sharing sentence ends between items. The conventional approach consists of constructing 60 sentences in six diﬀerent conditions each. This gives 60 sentence beginnings but 360 sentences endings, since these will be diﬀerent for each condition and for each item. We constructed 60 sentence beginnings, but only 20 sets of sentence ends, since each could be attached to any one of three diﬀerent sentence beginnings. In theory only ten sets of endings are necessary, but this involves each being applicable to six diﬀerent sentence beginnings, which causes diﬃculties. The Test Item Generator is thus used as follows: Group 1 uses the PRO endings for Items 1–10, the Raising endings for Items 11–20 and so on. By Items 21–30 Group 1 has returned to the ﬁrst block of endings but uses the Tough form. And so on. Item nos.

1–10

11–20

21–30

31–40

41–50

51–60

Group 1 Group 2 Group 3 Group 4 Group 5 Group 6

PRO Control Pronoun Passive Tough Raising

Raising PRO Control Pronoun Passive Tough

Tough Raising PRO Control Pronoun Passive

Passive Tough Raising PRO Control Pronoun

Pronoun Passive Tough Raising PRO Control

Control Pronoun Passive Tough Raising PRO

Let us here note a weakness of these materials: the mean length should be approximately equal across conditions. Bever and McElree achieved this by matching length in words, and we adopted this measure too. However, a letter or syllable count would have been more appropriate for German, and in fact the application of the word count measure actually increased the variation between conditions and forced us to make a number of assumptions. For example abzusägen “to saw oﬀ” is one orthographic word but zu sägen “to saw” is two. Since the former consists of the latter plus a particle, intuitively the former is longer than the second. We assumed that zu+verb was one word in all cases, and generally applied an absolute length criterion in addition to the word count.

240 Appendix 2: Materials for probe recognition Experiment 4

The Test Item Generator 1. 21. 41. PRO 1 Raising 1 Tough 1 Passive 1 Pronoun 1 Control 1 2. 22. 42. PRO 2 Raising 2 Tough 2 Passive 2 Pronoun 2 Control 2 3. 23.

Der gnadenlose Richter, der mit dem ﬂeißigen Anwalt zusammentraf, The merciless judge, who met with the hard-working lawyer Der schlaksige Notar, dem Petra ihre Probleme anvertraute, The lanky notary, who Petra conﬁded her problems to Der ängstliche Sachbearbeiter, der für den Bauvertrag zuständig war, The fearful case worker, who was responsible for the building contract hat abgelehnt, den Fall/ ohne eine Verabredung zu besprechen. refused to discuss the case without an appointment erschien uns, die Gefahr/ der Zeugenbeeinﬂüssung vergessen zu haben. seemed to us to have forgotten the danger of jury nobbling war schwer zu erkennen/ bei dem Dämmerlicht im Kreissaal. was diﬃcult to recognize in the half-light of the birthing room wurde oﬃziell gefragt,/ wann der unheilvolle Prozess stattfände. was oﬃcially asked when the terrible trial would happen hat gemeint, er sollte/ die staatliche Rechtshilfe verweigern. thought he should refuse state legal aid hat Herrn Wagenhuber/ Zugang zu höheren Instanzen verweigert. refused Herr Wagenhuber the right to appeal Der idiotische Schauspieler, der mit der reizbaren Diva arbeitete, The idiotic actor, who worked with the irritable star (fem) Der angesehene Kontrabaß, den die Chorleiterin zu vorlaut fand, The respected double bass, whom the choir leader (fem) found too cheeky Der unbeholfene Schüler, der schon zweimal hat sitzenbleiben müssen, The awkward pupil, who has already had to repeat a year twice hat gewählt, den ganzen Abend/ das Hauptstück zu proben. chose to rehearse the main play all evening erschien mir, das Hauptstück/ nicht mehr proben zu wollen. seemed to me to want to rehearse the main play no longer war leicht zu gewinnen/ für die Rolle des Erzengels. was easy to win for the role of archangel wurde endlich ausgewählt/ für die Rolle des Erzengels. was ﬁnally chosen for the role of archangel glaubte fest, er könnte/ die anderen wirklich beeindrücken. ﬁrmly believed he could really impress the others hätte gern die Rolle/ des Erzengels Gabriel gespielt. would have liked to play the role of the archangel Gabriel Die musterhafte Lehrerin, die uns vorhin freundlich gegrüßt hat, The exemplary teacher (fem), who greeted us so warmly just then Die anständige Witwe, deren Haus im exklusiven Vorort lag, The respectable widow, whose house was in an exclusive suburb

Appendix 2: Materials for probe recognition Experiment 4 241

43. PRO 3 Raising 3 Tough 3 Passive 3 Pronoun 3 Control 3 4. 24. 44. PRO 4 Raising 4 Tough 4 Passive 4 Pronoun 4 Control 4 5. 25. 45. PRO 5 Raising 5

Die nüchterne Wirtin, die in dem Nachbardorf geboren wurde, The sober landlady, who was born in the neighbouring village hatte vor, irgendwann/ in die schicke Stadtmitte umzuziehen. intended to move into the chic town centre some time scheint jetzt eine Wohnung/ im Zentrum kaufen zu wollen. seems now to want to buy a ﬂat in the centre war kaum zu verstehen/ für manche unerfahrene Norddeutsche. was barely comprehensible for some inexperienced north Germans wurde oft gewarnt, nachts/ seien abgelegene Landstrassen gefährlich. was often warned remote country roads were dangerous at night dachte, daß sie vielleicht/ ins Grüne ziehen sollte. thought that she should move out into the country ging recht selten/ ohne modische Handtasche zum Einkaufen. went shopping very rarely without a fashionable handbag Der vergnügte Elektrotechniker, dessen Frau das alte Testament studiert, The cheerful electrician, whose wife studied the old testament Der bedächtige Weinbauer, der selber keinen einzigen Tropfen trinkt, The thoughtful vintner, who never drinks a drop himself Der entspannte Monteur, der im Skatspiel Hamburger Meister wurde, The relaxed ﬁtter, who became Hamburg skat champion hat versprochen, den Schrebergarten/ seiner Eltern bald umzugraben. promised to dig his parents allotment over soon erschien jedoch, zu viel Zeit/ am Flußufer zu verschwenden. seemed however to spend too much time on the river bank war schwer wegzubekommen/ aus der Versammlung des Kegelvereins. was diﬃcult to get out of the bowling club meeting wurde plötzlich erschossen,/ was allen äußerst leid tat. was suddenly shot, which upset everyone very much berichtete uns, er sollte/ im Kirchenchor Tenor singen. reported to us he should sing tenor in the church choir arbeitet jede Woche zwei Abende/ im geliebten Gemüsegarten. worked two evenings a week in his beloved vegetable garden Die geschickte Mutter, die in das überfüllte Abteil einstieg, The skillful mother, who got into the packed compartment Die gebildete Maklerin, die den Wert von Luxusvillen einschätzte, The educated broker, who estimate the value of luxury villas Die humorlose Schreinerin, die wir auf der Ausfahrt kennenlernten, The humourless joiner (fem), whom we got to know on the trip versuchte ständig, mit anderen Passivesagieren/ ins Gespräch zu kommen. kept trying to enter a conversation with other passengers erschien uns die Reise/ zur Safari machen zu wollen. seemed to us to want to turn the trip into a safari

242 Appendix 2: Materials for probe recognition Experiment 4

Tough 5 Passive 5 Pronoun 5 Control 5 6. 26. 46. PRO 6 Raising 6 Tough 6 Passive 6 Pronoun 6 Control 6 7. 27. 47. PRO 7 Raising 7 Tough 7 Passive 7 Pronoun 7 Control 7

war nicht auszuhalten/ während der drei Stunden Fahrzeit. was unbearable during the three hours travelling time wurde schließlich beruhigt,/ als der Nachtisch aufgetragen wurde. was at last calmed down when the dessert arrived erklärte, daß sie unbedingt/ in Fahrtrichtung sitzen mußte. explained that she absolutely had to sit facing the front of the train wollte unbedingt/ in der Fahrtrichtung des Zuges sitzen. absolutely insisted on sitting facing the front of the train Der originelle Schriftsteller, der eine Menge Aktien geerbt hat, The excentric writer, who inherited a lot of shares Der begüterte Bibliothekar, von dem das Personal Pünktlichkeit lernte, The landed librarian, from whom the staﬀ learned punctuality Der begeisterte Wanderer, der nie Goethes Werther gelesen hat, The enthusiastic hill walker, who has never read Goethe’s Werther bereut sehr, das Rauchen/ nicht früher aufgegeben zu haben. greatly regrets not giving up smoking earlier scheint mir auch wohlgemeinte Kritik/ nicht gut zu vertragen. seems to me not to be able to take even well-meant criticism ist nicht zu befreien/ aus den Händen der Entführerbande. cannot be freed from the hands of the gang of kidnappers wurde endlich befreit/ aus den Händen der Entführerbande was at last freed from the hands of the gang of kidnappers hat klargemacht, er möchte/ die Schlösser Nordfrankreichs besuchen. made it clear he would like to visit the castles of northern France schreibt Novellen und Gedichte/ mit einem psychologischem Geschmack. writes novels and poems with a psychological taste Der gutmütige Forscher, den die meisten Kollegen “Strohkopf” nennen, The good-natured researcher, whom most colleagues call “hairbrain” Der erschöpfte Bergsteiger, dessen Auto einen Totalschaden hatte, The exhausted climber, whose car was written oﬀ Der tiefsinnige Historiker, der jahrelang keine Arbeit gefunden hat, The meditative historian, who spent years looking for a job hatte Angst, ein neues Auto/ ohne Führerschein zu fahren. was scared to drive a new car without a licence scheint mir immer nur/ mit öﬀentlichen Verkehrsmitteln zu fahren. seems to me always only to take public transport war schwer zu begeistern/ für die Regeln der Börse. was diﬃcult to inspire with the stock exchange rules wurde auch geliebt/ mit einer heißen, innigen Leidenschaft. was also loved with a burning, fervent passion hatte Angst, er würde/ das große Motorrad kaputtfahren. was scared he would smash up the big motorbike besuchte einen Verwandten,/ der im Osten Afrikas wohnte. visited a relative who lived in the east of Africa

Appendix 2: Materials for probe recognition Experiment 4 243

8. 28. 48. PRO 8 Raising 8 Tough 8 Passive 8 Pronoun 8 Control 8 9. 29.

49. PRO 9 Raising 9 Tough 9 Passive 9 Pronoun 9 Control 9 10. 30. 50.

Die störrische Sekretärin, deren Sohn einen tödlichen Unfall hatte, The troublesome secretary, whose son had a fatal accident Die bildhübsche Stiefmutter, die in eine kleinere Wohnung umzog, The beautiful stepmother, who moved into a smaller ﬂat Die rundliche Ansagerin, die den Fotoapparat hat fallen lassen, The chubby announcer (fem), who had dropped the camera war bereit, das Fahrrad/ einem beliebigen Fremden zu geben. was prepared to give the bicycle to a complete stranger erschien uns nicht so schnell/ zu fahren wie früher. seemed to us not to drive as fast as before war sehr zu bewundern/ während der ganzen verhängnisvollen Aﬀäre. had to be admired during the whole disastrous aﬀair wird jetzt geplagt/ von einer Art bösartigen Krampfader. is now plagued by a sort of malignant varicose vein bezweifelt, ob sie etwas/ dagegen hätte machen können. doubts whether she could have done anything about it legt immer noch frische Blumen/ aufs uralte Familiengrab. still lays fresh ﬂowers on the ancient family grave Der besonnene Beamte, der kürzlich in Rente gehen mußte, The sensible civil servant, who recently had to retire Der gefrässige Postbote, der frühmorgens erstaunlich guter Laune ist, The greedy postman, who is in an astonishingly good mood in the early morning Der kränkliche Familienvater, über den so oft Gerüchte umgehen, The ailing family father, about whom so often rumours circulate neigt dazu, andere/ nicht zu Wort kommen zu lassen. tends not to let others get a word in erschien jedem im Staatsdienst/ zuverlässig und unermüdlich zu sein. seemed to everyone in the civil service to be reliable and indefatigable ist leicht zu übersehen/ in einer Gruppe von Gleichgesinnten. is easily overlooked in a group of like-minded people wurde amtlich bedankt/ bei einer feierlichen Abschiedszeremonie. was oﬃcially thanked at a formal ceremony of departure erwähnte, daß er gerne/ bei der Bundespost arbeitete. mentioned that he liked working for the post oﬃce erzählt gern von der/ einzigartigen Arbeitweise der Bundespost. likes to tell stories of the unique working methods of the post oﬃce Der feinfühlige Künstler, der im Armenviertel eine Wohnung baut, The sensitive artist, who built a house in the deprived quarter Der gemütliche Förster, der sehr stimmungsvolle Aquarelle malt, The genial forester, who paints atmospheric water colours Der weltoﬀene Bildhauer, der sich für den Naturschutz engagiert, The internationalist sculptor, who was committed to environmental protection

244 Appendix 2: Materials for probe recognition Experiment 4

PRO 10

hat beschlossen, an der Besetzung/ leerer Häuser teilzunehmen. decided to take part in the occupation of empty houses Raising 10 erschien zunächst uns ein leeres Zweifamilienhaus beschaﬀen zu können. seemed at ﬁrst to be able to obtain two empty semi-detached houses for us Tough 10 ist nicht zu unterschätzen/ wenn von Ökopolitik gesprochen wird. should not be underestimated when ecological policy is discussed Passive 10 wurde mehrfach ausgestellt/ von den verschiedenen Mäzenen hierzulande. has frequently been exhibited by the various art sponsors in this country Pronoun 10 bestreitet, daß er an/ gewaltätigen Demonstrationen teilgenommen hat. disputes that he has taken part in violent demonstrations Control 11 benutzt Brennholz nur/ wenn Heizöl übertrieben teuer wird. uses ﬁre wood only when heating oil becomes excessively expensive 11. 31. 51.

Die gutwillige Köchin, die die geheime Gruppenaufgabe verpfuscht hat, The good-willed cook (fem), who messed up the secret group task Die jugendliche Gutachterin, deren Ehemann zu viel Alkohol trinkt, The youthful consultant (fem), whose husband drinks too much alcohol Die niedliche Reiseleiterin, die am Brennerpaß plötzlich Durchfall bekam, The cute courier (fem), who suddenly got diarrhoea on the Brenner Pass

PRO 11

vergißt nie, die Kinder/ morgens um sechs zu wecken. never forgets to wake the children at six am Raising 11 scheint nun ein neues Leben/ alleine bilden zu müssen. seems to have to build a new life on her own Tough 11 war sehr zu bedauern/ in einer so jämmerlichen Lebenslage. was very much to be pitied in such a terrible situation in life Passive 11 wird jetzt beraten/ von einer staatlichen Eheberatungsstelle. is now being advised by a state marriage guidance clinic Pronoun 11 gibt zu, sie erﬁndet/ neue Kochrezepte aus Zufall. admits she invents new recipe by chance Control 11 macht leckere Suppen/ und Süßspeisen, die nie mißglücken. 12. 32. 52. PRO 12 Raising 12 Tough 12 Passive 12

Die aufrichtige Musikerin, die die Statistikprüfung nur knapp bestand, The upright musician (fem), who only scraped a pass in the statistics exam Die muskulöse Rugbyspielerin, die ihre Kindheit in Belgien verbrachte, The muscular rugby player (fem), who spent her childhood in Belgium Die eigenartige Verkäuferin, die Kamelien immer zu vulgär fand, The individual sales assistant (fem), who always found camellias too vulgar leugnete ab, die Genossen/ im Stich gelassen zu haben. denied having left the comrades in the lurch erschien uns von der Hungersnot/ ein falsches Bild zu haben. seemed us to have a false picture of the famine ist nur zu ﬁnden,/ wenn alle Studentenkneipen zu haben. can only be found when all student bars are closed wurde kürzlich bestraft/ wegen eines banalen Verkehrsdelikts. was recently punished for a banale traﬃc oﬀence

Appendix 2: Materials for probe recognition Experiment 4 245

Pronoun 12 behauptet immer, sie wäre lieber/ Vermessungsbeamtin geworden. always maintains she would rather have become a public surveyor (fem) Control 12 ist mit den strengen Gymnasiallehrern/ selten gut ausgekommen. rarely got on well with the strict grammar school teachers 13. 33. 53.

Der neurotische Schulbusfahrer, der auf Glätte schnell fahren konnte, The neurotic school bus driver, who could drive fast on icy roads Der unbeugsame Gutsherr, der allen Knechten heftige Ehrfurcht einﬂösste, The inﬂexible lord of the manor, who inspired all the farm hands with awe Der türkische Obsthändler, der im Garten ein Sommerhäuschen baute, The Turkish fruit merchant, who built a summer house in the garden

PRO 13

beschloss dann, endlich etwas/ Außerordentliches im Hof zu tun. then decided to do something extraordinary in the farmyard Raising 13 scheint uns ein Spinettspieler/ ersten Ranges werden zu können. seemed to us to be able to become a spinett player of the ﬁrst order Tough 13 war nicht zu erkennen/ während der trüben Wintermonate. could not be recognized during the gloomy winter months Passive 13 wurde einmal ausgelacht,/ wegen eines Muttermals im Gesicht. was once laughed at because of a birth mark on his face Pronoun 13 hörte, daß er einen modernisierten/ Bauernhof geerbt hatte. heard that he had inherited a modernized farm Control 13 erbte einen urigen Bauernhof/ an der österreichischen Grenze. inherited an authentic farm on the Austrian border 14. 34. 54.

Die sanftmütige Hausfrau, die ständig etwas Neues zu nörgeln hatte, The gentle house wife, who always had something new to moan about Die kompetente Jägerin, die für Rehwild kein Mitgefühl empfand, The competent hunter (fem), who had no sympathy with deer Die auﬀällige Handwerkerin, deren Aktien plötzlich steil gefallen sind, The conspicuous tradeswoman, whose shares suddenly fell sharply

PRO 14

hat beabsichtigt, ein illustriertes/ Buch über Baumarten zu kaufen. intended to buy an illustrated book about species of trees Raising 14 erschien damals nur die jüngste/ Trivialliteratur lesen zu wollen. seemed at that time to want to read only the most recent bestsellers Tough 14 war leicht aufzuregen/ wenn man etwas schmutzig machte. was easily upset if one made something dirty Passive 14 wurde schließlich angezeigt/ von einem aggressiven Polizisten. was ﬁnally oﬃcially reported by an aggressive police oﬃcer Pronoun 14 bestätigte, daß sie alle Teppiche/ täglich putzen mußte. conﬁrmed that she had to clean all the carpets every day Control 14 möchte einmal im Leben/ mit einem Fallschirm springen. would like once in her life to do a parachute jump 15.

Der zimperliche Student, dessen Eltern sich für China interessierten, The squeamish student, whose parents were interested in China

246 Appendix 2: Materials for probe recognition Experiment 4

35. 55.

Der rothaarige Pfarrer, dem ein saftiger Schweinsbraten immer schmeckte, The red-headed parson, who always liked a juicy pork roast Der ehrgeizige Schnapsbrenner, der Silvester gern mit Feuerwerk feierte, The ambitious distiller, who liked to celebrate the new year with ﬁreworks

PRO 15

zog vor, im chinesischen/ Restaurant mit Eßstäbchen zu essen. preferred to eat with chopsticks in a Chinese restaurant Raising 15 erschien danach chinesische/ Gerichte nur mit Eßstäbchen zu essen. seemed afterwards to eat Chinese food only with chopsticks Tough 15 war kaum zurückzuhalten,/ wenn man indisch essen ging. could hardly be restrained, when we went for an Indian meal Passive 15 wurde sofort abgelehnt/ als Redner vor den Burschenschaften. was immediately refused as a speaker by the student associations Pronoun 15 hoﬀte vergebens, er würde/ in Ostasien Arbeit ﬁnden. hoped in vain he woud ﬁnd work in east Asia Control 15 verbrachte einen Teil/ der Sommerferien in lebhaftem Shanghai. spent a part of his holidays in busy Shanghai 16. 36. 56.

Die urwüchsige Geschäftsfrau, die vom Golﬃmmel angesteckt wurde, The earthy business woman, who was bitten by the golf bug Die lautstarke Gymnastikerin, die Samstags bis zum Morgen ausbleibt, The vociferous gymnast (fem), who stays out until morning on Saturdays Die geistreiche Gesellschaftsdame, bei der wir nachmittags Kaﬀee trinken, The witty society lady, at whose house we drink coﬀee in the afternoon

PRO 16

lehnt ab, am Wochenende/ diesen unnötigen Versammlungen beizuwohnen. refuses to attend these unnecessary meetings at the weekend Raising 16 scheint wirklich diese unnötigen/ Versammlungen am Wochennde zu hassen. seems to really hate these unnecessary meetings at the weekend Tough 16 ist schwer hinzuschleppen,/ wenn es keinen Champagner gibt. is diﬃcult to drag along when there is no champagne Passive 16 wird oft eingeladen/ zu diesen ﬂauen Eröﬀnungsfeiern is often invited to these drab opening ceremonies Pronoun 16 sagt uneingeschüchert, sie hat/ vorm Bürgermeister kein Respekt. says openly she has no respect for the mayor Control 16 redet oﬀen von dem schändlichen/ Amtsmißbrauch im Rathaus. talks openly about the disgraceful abuse of power in the town hall 17. 37. 57. PRO 17

Das jämmerliche Waisenkind, das wir im Baggersee baden sahen, The piteous orphan child, that we saw bathing at the ﬂooded quarry Das schmutzige Mädchen, dessen Stimme so ungewöhnlich schrill war, The dirty girl, whose voice was so unusually piercing Das frühreife Nachbarskind, das letzen Sommer am Mittelmeer war, The precocious child next door, who was at the Mediterranean last summer versprach uns, Hans und Maria/ nicht mehr zu einzuschüchtern. promised us not to intimidate Hans and Maria any more

Appendix 2: Materials for probe recognition Experiment 4 247

Raising 17

scheint uns die armen Kinder/ der Umgebung einzuschüchtern. seemed to us to intimidate the children in the neighbourhood Tough 17 war nicht wegzudenken/ aus der Gruppe der Schwimmer. without whom the swimming group would be unimaginable Passive 17 ist geschwind abgetrocknet worden/ mit einem riesigen Badetuch. was quickly dried with an enormous bath towel Pronoun 17 begriﬀ, daß sie mit dem/ Schiessbogen Erwachsene erschrak. understood that she frightened adults with her (archery) bow Control 17 schwamm hinaus zu einer Badeinsel/ im azurblauen Wasser. swam out to a bathing platform in the azure water 18. 38. 58.

Die vielseitige Zoologin, die der Lastfahrer zum Tanz auﬀorderte, The versatile zoologogist, who invited the lorry driver to dance Die barmherzige Klosterschwester, die einmal im Monat Besuch bekommt, The merciful nun, who received a visit once a month Die gefühlvolle Journalistin, die endlich von Slowakien zurück ist, The sentimental journalist, who has at last come back from Slovakia

PRO 18

ist imstande, wegen/ der Berichterstattung die Taz zu lesen. is capable of reading the Taz (a radical newspaper) for its reporting Raising 18 scheint jetzt wegen/ der Berichterstattung die Taz zu lesen. seems to read the Taz now for its reporting Tough 18 ist kaum ernstzunehmen,/ wenn es um Beruﬂiches geht. can hardly be taken serious in the professional ﬁeld Passive 18 ist einmal operiert worden/ wegen eines chronischen Darmgeschwürs. was once operated on for a chronic stomach ulcer Pronoun 18 erfuhr, daß sie fürs rastlose/ Zigeunerleben geschaﬀen war. found out that she was made for the restless gypsy life Control 18 hat selten so spannende Minuten/ erlebt wie diese. has rarely experienced such tense minutes as these 19. 39. 59. PRO 19

Der gediegene Höhlenforscher, der eine Gletscherschlucht hinunterrutschte, The upright cave explorer, who slipped down a glacier crevasse Der schmissige Buchhalter, der beim Schifahren einen Halsbruch erlitt, The dashing bookkeeper, who suﬀered a broken neck whilst skiing Der runzelige Schornsteinfeger, der an unserem Kamin gescheitert ist, The wrinkled chimney sweep, who failed on our chimney

vergißt nie, trotzdem ein/ sehr positives Weltbild zu verbreiten. never forgets nevertheless to propagate a very positive world view Raising 19 scheint uns trotzdem ein/ sehr positives Weltbild zu verbreiten. seems to us nevertheless to spread a very positive world view Tough 19 war nicht zu überreden,/ Risiken darf man nicht eingehen. could not be persuaded one should not take risks Passive 19 wurde nicht überredet,/ Risiken darf man nicht eingehen. was not persuaded one should not take risks Pronoun 19 verkündet oﬀen, er möchte/ es gerne nochmal versuchen. says openly he would like to try it again

248 Appendix 2: Materials for probe recognition Experiment 4

Control 19 20. 40. 60. PRO 20

hat später Trost gefunden/ bei einer verständlichen Freundin. later found consolation with an understanding girl friend Die fröhliche Braut, die ein blutrotes Hochzeitskleid getragen hat, The cheerful bride, who wore a blood-red wedding dress Das warmherzige Bauernmädchen, die dann den jungen Pfarrer heiratete, The warm-hearted peasant girl, who then married the young parson Die gesprächige Managerin, die werktags eine goldene Brille trug, The talkative manager (fem), who wore golden glasses on work days

träumt davon, im Frühling/ korbweise gelbe Schüsselblumen zu pﬂücken. dreams of picking basketfuls of yellow primroses in spring Raising 20 scheint also im Herbst/ korbweise Steinpilze ﬁnden zu können. thus seems to be able to ﬁnd basketfuls of ceps in autumn Tough 20 war unheimlich zu beneiden/ an dem aufregenden, mitreissenden Glückstag. was greatly to be envied on the exciting, thrilling day of joy Passive 20 wurde sehr beneidet/ an dem aufregenden, mitreissenden Glückstag. was much envied on the exciting thrilling day of joy Pronoun 20 schrieb später, sie vermisste/ das gesunde Leben draußen. wrote later she missed the healthy outdoor life Control 20 bemalte das Innere des Hauses/ mit moldauischen Wandgemälden. painted the inside of the house with Moldavian wall paintings

AUTHOR ""

TITLE "Appendix 3: Materials for ERP Experiment 5"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Appendix 3 Materials for ERP Experiment 5

These materials are in sets of three: the equi version, the raising version and the transitive version. In each case the three conditions are variants of one basic sentence, which is most usefully represented as the equi version, since the raising form can be derived by replacing the equi verb by schien (“seemed”), and the transitive form can be derived simply by removing the equi verb and making the main verb ﬁnite. We illustrate this by providing all three conditions in the ﬁrst sentence set. Note that occasionally a word or phrase was sometimes added or changed at the end of a sentence, after the testpoint, to match the sentence lengths or to make it read more naturally. Q1 Der Admiral drang darauf, während die Managerin draußen das ungeduldige Publikum aufheiterte, den Text nochmal im Detail zu überprüfen. – The admiral demanded, while the manager (fem) warmed up the impatient audience outside, to check the text over again in detail R1 Der Admiral schien, während die Managerin draußen das ungeduldige Publikum aufheiterte, den Text nochmal im Detail zu überprüfen. – The admiral seemed, while the manager (fem) warmed up the impatient audience outside, to check the text over again in detail T1 Der Admiral überprüfte, während die Managerin draußen das ungeduldige Publikum aufheite, den Text nochmal im Detail. – The admiral checked, while the manager (fem) warmed up the impatient audience outside, the text over again in detail 2 – 3 – 4 –

Der Anfänger lehnte ab, da die Geschäftsleitung leider nur die Hauptbeteiligten bezahlte, den Choral vorher einzuüben. The beginner refused, as the management unfortunately paid only the main participants, to rehearse the choral beforehand Der Anführer wagte es, während die Stabsoﬃziere ernst den mißratenen Feldzug diskutierten, den Saal mit einer geladenen Pistole zu betreten. The ring leader dared, while the staﬀ oﬃcers were sombrely discussing the failed campaign, to enter the room with a loaded pistol. Der Assistent beharrte darauf, nachdem die Anklägerin nochmal vor eventueller Zeugenbeeinﬂüssung warnte, den Brief nicht zu beantworten. The assistant persisted, after the plaintiﬀ (fem) again warned of possible jury nobbling, in not answering the letter

250 Appendix 3: Materials for ERP Experiment 5

5 – 6 – 7 – 8 – 9 –

Die Bäuerin weigerte sich, während die Familie komplett ins städtische Freibad fuhr, den Garten von vorne bis hinten umzugraben. The peasant woman refused, while the whole family went to the municipal open-air baths, to dig the garden over from end to end. Der Bildhauer neigte dazu, als die Bundesrepublik stolz die neuen Länder übernahm, den Journalisten als naiv einzuschätzen. The sculptor tended, when the federal republic proudly took over the new Länder, to rate the journalist as naive. Der Bräutigam wagte es, während die Gastgeberin erregt von der Bundespost erzählte, den Saal auf Zehenspitzen zu verlassen. The bridegroom dared, while the hostess was excitedly telling stories about the post oﬃce, to leave the room on tiptoe. Der Chronist beharrte darauf, nachdem die Abgeordnete vor mutwilliger Geschichtsverfälschung warnte, den Vorwurf erst nach sorgfältiger Überlegung zu beantworten. The chronicler persisted, after the MP warned of gratuitous misrepresentation of historical facts, in answering the reproach only after careful consideration. Der Dirigent plädierte dafür, obwohl die Ansteckungsgefahr natürlich den Besuch riskant machte, den Professor im Krankenhaus zu besuchen. The conductor made a plea, although the danger of infection naturally made the visit risky, to visit the professor in hospital.

10 Der Ehegatte strebte danach, während die Kinder langsam den unangenehmen Schock überwanden, den Ausdruck eines Trauernden nicht zu verlieren. – The husband strove, while the children slowly got over the unpleasant shock, not to lose his mournful expression. 11 Der Ehrengast lehnte ab, da der Einzug schon am folgenden Tag stattfand, den Choral im Detail zu proben. – The guest of honour guest refused, as the procession was already the following day, to rehearse the choral in detail. 12 Der Exporteur traute sich, als die Koalition erfolgreich an die Macht kam, den Kontakt mit dem Ausland auszubauen. – The exporter dared, when the coalition successfully came to power, to extend contacts with abroad. 13 Der Fabrikant traute sich, als die Grünen bundesweit in die Landtage kamen, den Handel mit Tropenholz nach fünfzig Jahren einzustellen. – The manufacturer dared, when the Greens got into regional parliaments across the country, to cease the trade in tropical wood after ﬁfty years. 14 Der Filmstar drang darauf, während die Schreinerin noch das große Podium anfertigte, den Text nochmal zu proben. – The ﬁlm star demanded, while the joiner (fem) was still ﬁnishing the large stage, to practise the text once again.

Appendix 3: Materials for ERP Experiment 5

15 Der Fotograf liebte es, während die Schwefelsäure weiterhin die deutschen Nadelhölzer abtötete, den Gewerkschaftsbund durch freizügige Spenden zu unterstützen. – The photographer loved, while sulfuric acid continued to kill oﬀ the German pine forests, to support the union federation with generous donations. 16 Die Fürstin plante ein, als die Hochzeitsgäste freudig vor der Kirche ankamen, den Jüngling in seiner weißen Pferdekutsche zu begrüßen. – The princess planned, when the wedding guests arrived joyfully in front of the church, to greet the youth in his white carriage. 17 Der Gärtner weigerte sich, während der Arbeitgeber vergnügt in den Westalpen wanderte, den Wagen mit der üblichen Sorgfalt zu pﬂegen. – The gardener refused, while the employer happily trekked in the west Alps, to look after the car with the usual care. 18 Der Gutachter genoß es, während der Gehilfe stillschweigend in dem Tagebuch weiterschrieb, den Fall des gestohlenen Fahrrads auf die Tagesordnung zu bringen. – The consultant enjoyed, while the assistant continued writing in the diary without saying a word, putting the case of the stolen bicycle on the agenda. 19 Der Hauptmann hütete sich, während die Kollegen schon die bevorstehende Frührente erwarteten, den Posten einen weiteren Monat zu behalten. – The captain took care to avoid, while the colleagues were already looking forward to the forthcoming early retirement, keeping the post for a further month. 20 Der Hotelier rühmte sich, da die Angestellten überwiegend für den Vorschlag stimmten, den Arbeitstag verkürzt zu haben. – The hotel owner prided himself, since the employees voted overwhelmingly for the suggestion, on having shortened the working day. 21 Der Jüngling haßte es, während der Bürgermeister streng die ganze Gruppe tadelte, den Erfolg bekannt zu machen. – The youth hated, while the mayor told oﬀ the whole group, to announce the success. 22 Die Kaiserin schlug vor, obwohl die Lage ungünstigerweise die erfolgreiche Verteidigung erschwerte, den Park in einem schönen Tal nahe der Grenze zu bauen. – The empress suggested, although the position unfortunately hindered successful defence, to build the park in a beautiful valley near the border. 23 Die Kaiserin plante ein, als die Gesandten würdevoll in den Saal eintraten, den Gast aus Burgund gebührend zu begrüßen. – The empress planned, as the emissaries digniﬁedly entered the room, to greet the guest from Burgundy in the ﬁtting manner. 24 Die Kellnerin hatte vor, während die Preise weiterhin stark nach unten tendierten, den Altbau in der Stadtmitte zu kaufen. – The waitress intended, while the prices continued to have a strong downward trend, to buy the old building in the town centre.

251

252 Appendix 3: Materials for ERP Experiment 5

25 Der Kurfürst schlug vor, als der Feldherr endlich den blutigen Bürgerkrieg gewann, den Park an einem schwer zugänglichen Berghang wieder aufzubauen. – The elector suggested, as the ﬁeld commander at last won the bloody civil war, to rebuild the park on a poorly accessible mountain side. 26 Der Landrat bot an, als der Produzent bitter die bisherige Leistung kritisierte, den Vortrag noch einmal zu wiederholen. – The District Commissioner oﬀered, when the producer bitterly criticized the previous performance, to repeat the lecture once more. 27 Die Lehrerin hatte vor, während der Mittelstand zunehmend hinaus ins Grüne zog, den Altbau in der Stadtmitte zu kaufen. – The teacher (fem) intended, while the middle class increasingly moved out into the country, to buy the old building in the town centre. 28 Der Lokführer neigte dazu, wenn der Stammtisch geräuschvoll das schwarz-grüne Bündnis diskutierte, den Kanzler als Birne zu bezeichnen. – The train driver tended, when the drinking pals noisily discussed the CDU/Green alliance, to call the Chancellor “pear” (common nickname for Helmut Kohl). 29 Der Matrose verzichtete darauf, während der Junge unten die zerfetzten Kleider ﬂickte, den Abend mit Kartenspiel zu verbringen. – The sailor did without, while the lad downstairs patched the tattered clothes, spending the evening playing cards. 30 Der Mediziner bemühte sich, als der Staatshaushalt schließlich in die Zeitungen kam, den Tabakgenuß sofort zu reduzieren. – The medical student tried hard, when the national budget statement at last got into the papers, to reduce his tobacco consumption immediately. 31 Der Metzger rühmte sich, da die Kunden überwiegend die neue Lösung akzeptierten, den Arbeitstag zu verkürzt zu haben. – The butcher boasted, as the customers overwhelmingly accepted the new solution, to have shortened the working day. 32 Der Missionar kündigte an, während der Choleraausbruch bereits nahezu epidemische Ausmaße annahm, den Chor neu zu gründen. – The missionary announced, while the cholera outbreak was already almost an epidemic, that he was refounding the choir. 33 Der Monteur träumte davon, während die Fachwelt heutzutage nur aufs Geld sieht, den Kunsthändler mit seinen stimmungsvollen Aquarellen zu beeindrücken. – The ﬁtter dreamt, while the specialists nowadays only consider money, of impressing the art dealer with his atmospheric water-colours. 34 Der Pförtner haßte es, während der Bauer enthusiastisch die ganze Besatzung anfeuerte, den Angriﬀ des Feindes verschweigen zu müssen. – The doorman hated, while the farmer enthusiastically ﬁred up the whole garrison, having to keep quiet about the enemy’s attack.

Appendix 3: Materials for ERP Experiment 5 253

35 Der Philosoph überlegte sich, obwohl die Passagiere ungeduldig die übliche Nachtruhe erwarteten, den Richter neben ihm in eine heftige politische Debatte zu ziehen. – The philosopher considered, although the passengers were impatiently expecting the usual night-time quiet, dragging the judge next to him into a violent political argument. 36 Die Putzfrau ﬂehte darum, als die Kinder lautstark in dem Abteil herumtobten, den Roman in Ruhe lesen zu dürfen. – The cleaning lady pled, as the children noisily rampaged around the compartment, to be allowed to read the novel in peace. 37 Der Radfahrer träumte davon, obwohl die Ölkonzerne immer die vielen Nachteile hervorhoben, den Bürgermeister von den Vorteilen des Fahrradfahrens zu überzeugen. – The cyclist dreamt, although the oil companies always highlight the many disadvantages, to convince the mayor of the advantages of cycling. 38 Die Reisende ﬂehte darum, während die Fußballfans laut in dem Wartesaal herumbrüllten, den Krimi in Ruhe lesen zu dürfen. – The traveller (fem) pled, while the football fans yelled loudly around the waiting room, to be allowed to read the detective novel in peace. 39 Der Schlosser liebte es, während die Gattin abends vor dem Fernseher schlummerte, den Gewerkschaftsbund mit viel freiwilliger Arbeit zu unterstützen. – The locksmith loved, while his wife slumbered in front of the TV in the evening, to support the trade union federation with a lot of voluntary work. 40 Der Schneider tendierte dazu, wenn die Familie zusammen um den Tisch saß, den Wein mit etwas Wasser zu mischen. – The tailor tended, when the family sat around the table together, to mix the wine with a little water. 41 Der Sheriﬀ hoﬀte, als die Witwe plötzlich in das Zimmer kam, den Täter endlich verurteilen zu können. – The sheriﬀ hoped, when the widow suddenly came into the room, to be able to sentence the oﬀender at last. 42 Der Skiläufer unterliess es, während der Wetterbericht warnend einen großen Sturm ankündigte, den Hang zu bewältigen. – The skier forbore, while the weather forecast gave warning notice of a large storm, to master the slope. 43 Der Soziologe tendierte dazu, obwohl die Fachpresse eher die deutsche Braukunst pries, den Wein zu jeder Mahlzeit zu genießen. – The sociologist tended, although the specialist press sooner praised German brewing, to enjoy wine with every meal. 44 Der Steinmetz hoﬀte, als das Dach unglücklicherweise in den Keller ﬁel, den Fehler in seinem Entwurf in kürzester Zeit zu erkennen. – The mason hoped, when the roof unfortunately fell into the cellar, to recognize the error in his design as soon as possible.

254 Appendix 3: Materials for ERP Experiment 5

45 Die Studentin gab zu, nachdem die Freundin knapp den schweren Unfall überlebte, den Sportwagen zu schnell gefahren zu haben. – The student (fem) admitted, after the friend (fem) just survived the serious accident, to have driven the sports car too fast. 46 Der Teenager gab zu, als der Bauer knapp den schweren Zusammenprall überlebte, den Sportwagen gestohlen zu haben. – The teenager admitted, when the farmer just survived the serious crash, to have stolen the sports car. 47 Der Theologe bot an, als die Klasse zusammen die zerstörte Burg besuchte, den Rundgang durch das Münster zu wiederholen. – The theologist oﬀered, when the class visited the destroyed castle together, to repeat the tour round the minster. 48 Der Theologe überlegte sich, obwohl die Studenten unruhig die schwere Prüfung erwarteten, den Arbeiter in eine tiefgehende akademische Diskussion zu ziehen. – The theologist considered, although the students restlessly waited for the diﬃcult exam, drawing the worker into an in-depth academic discussion. 49 Der Tischler bemühte sich, als die Gesundheit weltweit zur ersten Priorität wurde, den Tabakgenuß erheblich zu reduzieren. – The joiner made great eﬀorts, as health was becoming the ﬁrst priority world-wide, to reduce tobacco consumption considerably. 50 Der Torwart unterliess es, da die Stadtluft immer zu starken Halsschmerzen führte, den Marathonlauf durch die Innenstadt zu beenden. – The goalkeeper forebore, as urban air always led to bad sore throats, to ﬁnish the marathon through the centre of town. 51 Der Tourist zwang sich, als die Gruppe wieder von dem Bergkamm zurückkam, den Mietwagen trotz seines miesen Zustands zu reparieren. – The tourist forced himself, when the group came back from the mountain ridge, to mend the hired car in spite of its rotten condition. 52 Der Uhrmacher plädierte dafür, obwohl die Malaria natürlich das ganze Unternehmen gefährdete, den Betrieb in Ostafrika zu besuchen. – The watchmaker argued in favour, although malaria naturally endangered the whole undertaking, to visit the plant in East Africa. 53 Der Urlauber pochte darauf, während die Familie entschlossen für das Mittelmeer argumentierte, den Fernwanderweg zum Zentralmassiv zu erkunden. – The holidaymaker insisted, while the family ﬁrmly argued for the Mediterranean, on exploring the long distance path to the central massif. 54 Der Verleger genoß es, während der Sekretär düster in den Unterlagen kritzelte, den Fall der mißratenen feierlichen Eröﬀnung zu besprechen. – The editor enjoyed, while the secretary scribbled darkly in the papers, discussing the case of the failed formal opening ceremony.

Appendix 3: Materials for ERP Experiment 5 255

55 Der Verleger strebte danach, während die Börsenkurse langsam den anfänglichen Abwärtstrend überwanden, den Ausdruck eines Gewinners anzunehmen. – The editor strived, while the stock exchange indices slowly overcame the initial downward trend, to adopt the expression of a winner. 56 Der Verwalter kündigte an, obwohl der Pfarrer meistens die alte Orgel spielte, den Chor im Dorf neu zu gründen. – The administrator announced, although the parson mostly played the old organ, that he was going to refound the village choir. 57 Der Vorsitzende hütete sich, während die Kollegen bereits auf eventuelle Reformen hoﬀten, den Posten eines Ersatzstaatssekretärs neu zu besetzen. – The chairman took care not to, while the colleagues were already hoping for possible reforms, ﬁll the post of replacement national secretary. 58 Der Wanderer pochte darauf, während die Mehrheit eifrig für die Alpen schwärmte, den Gipfel der Pyrenäen zu erkunden. – The walker insisted, while the majority keenly enthused for the Alps, on exploring the peak of the Pyrenees. 59 Der Wildhüter verzichtete darauf, während die Frauen ﬂeißig für die Kinder strickten, den Abend im Gasthaus um die Ecke zu verbringen. – The gamekeeper did without, while the women knitted hard for the children, spending the evening in the bar round the corner. 60 Der Zahnarzt zwang sich, als die Spritze schon die gewünschte Wirkung zeigte, den Schaden am Backenzahn schnell zu reparieren. – The dentist forced himself, when the injection had achieved the desired result, to repair the damage to the molar quickly.

AUTHOR ""

TITLE "Appendix 4: Materials for sentence matching Experiment 6"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Appendix 4 Materials for sentence matching Experiment 6

These materials are in two sets of three sentences: the ﬁrst three contain the primary experimental conditions and the second three additional conditions which are necessary to factor out matching time diﬀerences caused by the diﬀerent word orders in the experimental conditions. The (b) and (c) conditions are derivable from the (a) sentence types by rule, as are the (e) and (f) types from the (d) type. See Chapter 6 for further details of the structural diﬀerences between them. We provide all six types for the ﬁrst set only as an example. Notice that many of the translations, particularly of the ﬁrst sentence type (a), are superiority violations in English and therefore ungrammatical. The equivalent German sentences are not ungrammatical, and are much less marked than in English because the overt case-marking on the wh-items makes the sentences readily comprehensible, and because German prefers pied-piping to preposition stranding. There are two constraints on the materials which provide the major reason for the sentences having the form they do: ﬁrstly the wh-item should be unconstruable as dependent on the matrix clause so that the cyclic raising construction is the only possible one, but secondly the experimental wh-items should be reasonably varied, so that subjects do not develop strategies. These wh-element plus preposition compounds such as woran, wovon, and wh-element plus particle compounds such as wohin and woher provide a ready means of fulﬁlling these constraints. We use the somewhat archaic whither and whence here to give a closer rendition of the original wohin and woher, though these are not at all marked in German. In order to limit the use of paper, we provide idiomatic translations only where the intended meaning of the German is not clear from the literal translation. 1a – 1b 1c 1d – 1e 1f

Wem meint der Archäologe hat er welche antiken Münzen verkauft? To whom does the archaeologist think he sold which ancient coins? Wem meint der Archäologe welche antiken Münzen er verkauft hat? Wem der Archäologe meint hat er welche antiken Münzen verkauft? Wem meint der Theologe hat er denn die Schriftrollen verkauft? To whom does the theologist think he sold the scrolls? Was meint der Theologe wem er die Schriftrollen verkauft hat? Wem meint der Theologe daß er die Schriftrollen verkauft hat?

2a – 2d –

Wen sagt die dreißigjährige Psychologin möchte sie wann heiraten? Whom does the 30-year-old psychologist say she would like to marry when? Wen sagt Brigitte will sie denn vielleicht doch heiraten? Whom does Brigitte say she perhaps does want to marry?

258 Appendix 4: Materials for sentence matching Experiment 6

3a – 3d –

Woran sagte der Ornithologe erkennt man sofort welchen Raubvogel? By what did the ornithologist say one recognizes which bird of prey immediately? Woran sagte der Polizist erkennt man denn sofort den Heroinsüchtigen? By what did the police oﬃcer say one recognizes a heroin addict immediately?

4a – 4d –

Wohin glaubst du hat Boris Becker welche deiner Töchter begleitet? Whither do you think Boris Becker has accompanied which of your daughters? Wohin sagtest du ist Jürgen Klinsmann denn mit deiner Frau entﬂohen? Whither did you say Jürgen Klinsmann has ﬂed with your wife?

5a – 5d –

Wem meint die Stadträtin hat sie welche Korruption nachgewiesen? Of whom does the town councillor (fem) think she has proven what corruption? Wem dachte der Pfarrer hat er denn die Kerzen für die Kirche anvertraut? To whom did the parson think he had entrusted which church’s candles?

6a – 6d –

Woran behauptet der Lazarettarzt soll welcher Soldat krank sein? From what does the military hospital doctor claim which soldier is suﬀering? Woran behauptet der Hals-Nasen-Ohren-Arzt soll dein Vater denn leiden? From what does the ENT doctor claim your father is suﬀering?

7a – 7d –

Wovon sagt der Hilfsarbeiter sei welcher Manager geistig abhängig? On what does the manual worker say which manager is mentally dependent? Wovon sagt der kleine Alex sei seine Mutter physisch abhängig? On what does little Alex say his mother is physically dependent?

8a – 8d –

Wen sagt die ehrgeizige Ärztin will sie in welcher Klinik behandeln? Whom does the ambitious doctor (fem) say she wants to treat in which clinic? Wen sagt die Ärztin wird sie denn mit Laserstrahlen behandeln? Whom does the doctor (fem) say she will treat with laser beams?

9a – 9d –

Mit wem sagt Svenja wird sie welche Wüste mit dem Motorrad durchqueren? With whom does Svenja say she will cross which desert by motorbike? Mit wem sagt Katrin wird sie denn die Ostsee im Kayak überqueren? With whom does Katrin say she will cross the Baltic by canoe?

10a – 10d –

Woher schreibt Die Zeit hat Gerhard Schröder welche Kollegin geholt? Whence does Die Zeit say Gerhard Schröder has fetched which colleague (fem)? Woher schreibt Der Stern hat Klaus Kinkel denn seinen Whisky geholt? Whence does Der Stern say Klaus Kinkel has fetched his whisky?

11a – 11d –

Wen behauptet Golo Mann hat Günther Grass wann zum Mord angestiftet? Whom does Golo Mann claim Günther Grass incited to murder when? Wen behauptet Christa Wolf hat Anna Seghers denn zur Flucht angespornt? Whom does Christa Wolf claim Anna Seghers incited to ﬂee? (from East Germany)

12a – 12d –

Wen teilte die Bundespost mit darf man künftig wann nicht anrufen? Whom did the federal post oﬃce announce one is no longer allowed to phone when? Wen teilt Reuters mit kann man denn als nächsten CDU-Chef erwarten? Whom did Reuters announce one can expect as the next head of the CDU?

13a Woher sagt der Bäcker hat er wann das Sauerteigrezept gestohlen? – From where does the baker say he has stolen the sour dough recipe when?

Appendix 4: Materials for sentence matching Experiment 6 259

13d Woher sagt Roman Herzog hat er denn seinen Vornamen geerbt? – From where does Roman Herzog say he has inherited his ﬁrst name? 14a – 14d –

Wofür glaubt die Polizei hat der Bankrotteur wieviel Geld ausgegeben? On what does the police believe the bankrupt spent how much money? Wofür glaubt der Gerichtsvollzieher hast du denn das Geld ausgegeben? On what does the bailiﬀ believe you have spent the money?

15a – 15d –

Wen sagt der Reitlehrer hat welches Pferd im Galopp abgeworfen? Whom does the riding instructor say which horse threw while galloping Wen sagt der Skilehrer hat Bernd denn über den Haufen gefahren? Whom does the ski instructor say Bernd ran over?

16a – 16d –

Woran sagt die Chirurgin wird sie dich morgen weshalb operieren? For what does the surgeon (fem) say she wants to operate on you tomorrow? Woran sagt der Pfarrer wird er denn trotz seines Ordens nie glauben? In what does the parson say he will never believe in spite of his (religious) order?

17a – 17d –

Wie weit meinst du muß die Schwalbe im Winter wohin ﬂiegen? How far do you think the swallow must ﬂy in winter whither? Wie weit meinst du muß die Lerche denn im Herbst nach Süden ﬂiegen? How far do you think the lark must ﬂy south in autumn?

18a – 18d –

Welchen Baum sagt der Botaniker muß er wie oft täglich giessen? Which tree does the botanist say he has to water how many times a day? Welchen Hebel sagt der Mechaniker muß er denn jeden Tag schmieren? Which lever does the mechanic say he has to oil every day?

19a – 19d –

Wem dachte deine Frau wolltest du weshalb die roten Rosen schenken? To whom did your wife think you wanted to give the red roses for what reason? Wem dachte dein Mann wolltest du das schöne Bild von dir schenken? To whom did your husband think you wanted to give the nice picture of yourself?

20a – 20d –

Wen sagt Uwe hat er in seinem Trabbi wo auf der Autobahn überholt? Whom does Uwe say he overtook in his Trabant where on the motorway? Wen sagt Anja hat sie denn auf ihrer Honda im Verkehrsstau überholt? Whom does Anja say she overtook on her Honda in a traﬃc jam?

21a – 21d –

Womit meint der Zollbeamte haben wir welchen Koﬀer vollgestopft? With what does the customs oﬃcer think we have stuﬀed which suitcase full? Womit meint die Polizei hat der Fahrer denn den Lastwagen vollgeladen? With what do the police think the driver has loaded the lorry full?

22a – 22d –

Wen behauptet Schumacher hat Frentzen wann in der Kurve überholt? Whom does Schumacher claim Frentzen overtook when on the bend? Wen behauptet Brecht hat Kurt Weill denn mit seiner Musik übertroﬀen? Whom does Brecht claim Kurt Weill surpassed with his music?

23a – 23d –

Wohin meint der Reiseleiter sollten wir wann mit dem Dampfer fahren? Whither does the courier think we should go by steamer when? Von wo aus meint Kerstin kann man denn den Kölner Dom sehen? From where does Kerstin think one can see Cologne cathedral?

260 Appendix 4: Materials for sentence matching Experiment 6

24a Mit wem warnt Lafontaine muß man weshalb als nächsten Kanzler rechnen? – With whom does Lafontaine warn one must for what reason reckon as next chancellor? = “Whom does L warn (us) that we should we be prepared for as the next chancellor?” 24d Mit wem warnt Waigel muß man denn als nächsten Bundespräsidenten rechnen? – With whom does Waigel warn one must reckon as next federal president? “Whom does Waigel warn (us) that we should be prepared for as next federal president?” 25a – 25d –

Wem behauptete der Zeuge hat der Landrat wieviel Pilsner bezahlt? Whom did the witness claim the district commissioner paid for how many beers for? Wem behauptete der Wirt hat der Fremde die Flaschen Weißbier bezahlt? Whom did the publican claim the stranger paid for the bottles of wheat beer for?

26a – 26d –

Wie tief meint der Geologe sei die Erdölschicht wo in Westfahlen? How deep does the Geologist think the crude oil deposits are where in Westphalia? Wie groß meint der Geologe sei denn das Risiko eines Erdbebens in Ulm? How big does the Geologist think the risk of an earthquake is in Ulm?

27a – 27d –

Wem behauptet der Journalist hat Graf Lambsdorﬀ weshalb gelogen? To whom does the journalist claim Count Lambsdorﬀ lied for what reason? Wen sagt der Kanzler hat die Fraktion denn zum Rücktritt aufgefordert? Whom does the chancellor say the parliamentary party called upon to resign?

28a – 28d –

Mit wem behauptet Die Bildzeitung hat Steﬃ Graf wo schamlos geﬂirtet? With whom does Die Bildzeitung claim Steﬃ Graf ﬂirted shamelessly where? Mit wem behauptet Burda möchte Claudia Schiﬀer denn gerne arbeiten? With whom does Burda claim Claudia Schiﬀer would like to work?

29a – 29d –

Wohin erzähltest du will dich welcher Bischof als Missionar schicken? Whither were you saying which bishop wants to send you as a missionary? Wohin erzähltest du will dich denn dein Vorgesetzter zur Strafe schicken? Whither were you saying your superior wants to send you as a punishment?

30a Woran behauptet Pavarotti ist Fischer-Dieskau weshalb gescheitert? – On what does Pavarotti claim Fischer-Dieskau foundered for what reason? “What does Pavarotti claim was Fischer-Dieskau’s undoing (and why)? 30d Woran behauptet Karajan ist Karl Böhm denn letzlich gescheitert? – On what does Karajan claim Karl Böhm foundered in the end? “What does Karajan claim was Karl Böhm’s undoing in the end?

AUTHOR ""

TITLE "References"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

References

Ades A. & Steedman M. 1982. “On the order of words”. Linguistics and Philosophy 4, 517–558. Altmann G. & Steedman M. 1988. “Interaction with context during human sentence processing”. Cognition 30, 191–238. Baayen H., Piepenbrock R. & Rijn H. van 1996. The CELEX2 Lexical Database (CD ROM). Linguistic Data Consortium, University of Pennsylvania. Bader M. & Lasser I. 1994. “German verb-ﬁnal clauses and sentence processing: Evidence for immediate attachment”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 225–242. Hillsdale: Erlbaum. Barss A. 1993. “Transparency and visibility: Sentence processing and the grammar of anaphora”. In Cognitive Models of Speech Processing: The Second Sperlonga Meeting, G. Altmann & R. Shillock (eds), 401–452. Hove: Erlbaum. Bates E., McNew S., MacWhinney B., Devescovi A. & Smith S. 1982. “Functional constraints on language processing: A cross-linguistic study”. Cognition 11, 245–299. Baumann H., Nagengast J. & Klaas G. 1993. New Experimental Setup (NESU). Ms. MaxPlank-Institut für Pyscholinguistik, Nijmegen. Bayer J. & Kornﬁlt J. 1991. “Against scrambling as Move-Alpha”. Proceedings of NELS 21, 1–16. University of Massachusetts. Behaghel O. 1909. “Beziehung zwischen Umfang und reihenfolge von Satzgliedern”. Indogermanische Forschungen 25, 110–142. Belletti A. 1990. Generalized Verb Movement. Turin: Rosenberg & Sellier. Berwick R. & Weinberg A. 1984. The Grammatical Basis of Linguistic Performance: Language Use and Acquisition. Cambridge MA: MIT Press. Berwick R. & Weinberg A. 1985. “Deterministic parsing: A modern view”. Proceedings of NELS 15, 15–33. University of Massachusetts. Bever T. & McElree B. 1988. “Empty categories access their antecedents during comprehension”. Linguistic Inquiry 19, 35–43. Bever T & Sanz M. 1997. “Empty categories access their antecedents during comprehension of Spanish Unaccusatives”. Linguistic Inquiry 28, 69–91. Bever T., Straub K., Shenkman K., Kim K., & Carrithers C. 1990. “The psychological reality of NP trace”. Proceedings of NELS 20, 46–60. University of Massachusetts. Bobaljik J. 1995. “Morphosyntax: the syntax of verbal inﬂection”. PhD dissertation, MIT. Boland J. 1993. “The role of verb argument structure in sentence processing: Distinguishing between syntactic and semantic eﬀects”. Journal of Psycholinguistic Research 22, 133–152.

262 References

Boland J., Tanenhaus M. & Garnsey S. 1990. “Lexical structure and parsing: evidence for the immediate use of verbal argument and control information in parsing”. Journal of Memory and Language 29, 413–432. Borsley R. 1989. “PSG and the Barriers conception of clause structure”. Linguistics 27, 843–63. Borsley R. 1991. Syntactic Theory: A Uniﬁed Approach. London: Edward Arnold. Borsley R. 1996. Modern Phrase Structure Grammar. Oxford: Blackwell. Boškovic´ Z. 1994. Categorial status of null operator relatives and ﬁnite declarative complements. PhD dissertation, MIT. Bouchard D. 1984. On the Content of Empty Categories. Dordrecht: Foris. Bourdages J. 1992. “Parsing complex NPs in French”. In Island Constraints, H. Goodluck & M. Rochemont (eds), 61–88. Dordrecht: Kluwer. Bresnan J. 1978. “A realistic transformational grammar”. In Linguistic Theory and Psychological Reality, J. Bresnan, M. Halle & G. Miller (eds), 1–59. Cambridge MA: MIT Press. Bresnan J. & Kaplan R. 1982. “Introduction: Grammars as mental representations of language”. In The Mental Representation of Grammatical Relations, J. Bresnan (ed), xvii–lii. Cambridge Massachusetts: MIT Press. Cairns H. & Blank M. 1976. “Word recognition latency and the duration of clausal processing”. Working Papers in Speech and Hearing Sciences, Vol II. Graduate School and University Center of the City University of New York. Cairns H., Cowart W. & Jablon A. 1981. “Eﬀects of prior context on the integration of lexical information during sentence processing”. Journal of Verbal Learning and Verbal Behaviour 20, 445–53. Chomsky N. 1965. Aspects of the Theory of Syntax. Cambridge: MIT Press. Chomsky N. 1973. “Conditions on transformations”. In A Festschrift for Morris Halle, S. Anderson & P. Kiparsky (eds), 232–86. New York: Holt, Reinhart & Winston. Chomsky N. 1977. “On wh-movement”. In Formal Syntax, P. Culicover, T. Wasow & A. Akmajian (eds), 71–132. New York: Academic Press. Chomsky N. 1981. Lectures on Government and Binding: The Pisa Lectures. Berlin: Mouton de Gruyter. Chomsky N. 1982. Some Concepts and Consequences of the Theory of Government and Binding. Cambridge MA: MIT Press. Chomsky N. 1986a. Barriers. Cambridge MA: MIT Press. Chomsky N. 1986b. Knowledge of Language: Its Nature, Origin and Use. New York: Praeger. Chomsky N. 1991. “Some notes on economy of derivation and representation”. In Principles and Parmeters in Comparative Grammar, R. Freidin (ed), 417–454. Cambridge, Massachusetts: MIT Press. Chomsky N. 1993. “A minimalist program for linguistic theory”. In The View from Building 20, K. Hale & S. Keyser (eds), 1–52. Cambridge, Massachusetts: MIT Press. Chomsky N. 1995. The Minimalist Program. Cambridge: MIT Press. Clahsen H., Hong U. & Sonnenstuhl-Henning I. 1995. “Grammatical constraints in syntactic processing: Sentence matching experiments on German”. The Linguistic Review 12, 5–33. Clifton C. & Frazier L. 1986. “The use of syntactic information in ﬁlling gaps”. Journal of Psycholinguistic Research 15, 209–224.

References 263

Clifton C. & Frazier L. 1989. “Comprehending sentences with long distance dependencies”. In Linguistic Structure in Language Processing, G. Carlson & M. Tanenhaus (eds), 273–317. Dordrecht: Kluwer. Clifton C., Frazier L. & Connine C. 1984. “The use of syntactic information in ﬁlling gaps”. Journal of Verbal Learning and Verbal Behaviour 23, 696–708. Cloitre M. & Bever T. 1988. “Linguistic anaphors, levels of representation, and discourse”. Language and Cognitive Processes 3, 293–322. Corbet A. & Chang F. 1983. “Pronoun disambiguation: Accessing potential antecedents”. Memory and Cognition 11, 283–294. Corver N. & van Riemsdijk H. (eds) 1994. Studies on Scrambling: Movement and Nonmovement Approaches to Free Word Order Phenomena. Berlin: Mouton de Gruyter. Coulson, S., King, J. W., Kutas, M. 1998. “Expect the unexpected: Event-related brain response to morphosyntactic violations”. Language and Cognitive Processes 13, 21–58. Crain S. & Fodor J. 1985a. “How can grammars help parsers?” In Natural Language Parsing: Psychological, computational and theoretical perspectives, D. Dowty, L. Kartunen & A. Zwicky (eds), 94–128. Cambridge: CUP. Crain S. & Fodor J. 1985b. “Rules and constraints in sentence processing”. Proceedings of NELS 15, 87–104. University of Massachusetts. Crain S. & Fodor J. 1987. “Sentence matching and overgeneration”. Cognition 26, 123–169. Crain S. & Steedman M. 1985. “On not being led up the garden path: The use of context by the psychological parser. In Natural Language Parsing: Psychological, computational and theoretical perspectives, D. Dowty, L. Kartunen & A. Zwicky (eds), 320–358. Cambridge: CUP. Crocker M. 1994. “On the nature of the principle-based sentences processor”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 245–266. Hillsdale: Erlbaum. Crocker M. 1996. Computational Psycholinguistics: An Interdisciplinary Approach to the Study of Language. Dordrecht: Kluwer. De Vincenzi M. 1991. “Filler-gap dependencies in a null subject language: Referential and nonreferential WHs”. Journal of Psycholinguistic Research 20, 197–213. De Vincenzi M. 1996. “Syntactic analysis in sentence comprehension: Eﬀects of dependency types and grammatical constraints”. Journal of Psycholinguistic Research 25, 117–133. Dell G., McKoon G. & Ratcliﬀ R. 1983. “The activation of antecedent information during the processing of anaphoric reference in reading”. Journal of Verbal Learning and Verbal Behaviour 22, 121–132. Deprez V. 1994. “Parameters of object movement”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 101–152. Berlin: Mouton de Gruyter. Elman J. & McClelland J. 1984. Speech perception as a cognitive process: The interactive activation model. In Speech and Language vol 10, Lass N. (ed), 337–374. New York: Academic Press. Eubank L. 1993. “Sentence matching and processing in L2 development”. Second Language Research 9, 253–280. Fanselow G. 1988. “German word order and universal grammar”. In Natural Language Parsing and Linguistic Theories, U. Reyle & C. Rohrer (eds), 317–355. Dordrecht: Reidel.

264 References

Fanselow G. 1990. “Scrambling as NP-movement”. In Scrambling and Barriers, G. Grewendorf & W. Sternefeld (eds), 113–140. John Benjamins: Amsterdam. Fanselow G. 1997. Features, θ-roles and free constituent order. Ms. Farmer A. 1980. On the interaction of morphology and syntax. PhD dissertation, MIT. Featherston S. 2000 wh-constructions in German: Superiority and that-trace eﬀects. Ms., University of Tübingen. Ferreira F. & Clifton C. 1986. “The independence of syntactic processing”. Journal of Memory and Language 25, 348–368. Fiengo R. 1977. “On trace theory”. Linguistic Inquiry 8, 35–62. Flores D’Arcais G. 1990. “Parsing principles and language comprehension during reading”. In Comprehension Processes in Reading, D. Balota, G. Flores D’Arcais & K. Rayner (eds). Hillsdale: Erlbaum. Fodor J.A., Bever T. & Garrett M. 1974. The Psychology of Language. New York: McGraw-Hill. Fodor J. A. 1983. Modularity of Mind. Cambridge, Massachusetts: MIT Press. Fodor J. D. 1978. “Parsing strategies and constraints on transformations”. Linguistic Inquiry 9, 427–473. Fodor J. D. 1983. “Constraints on gaps”. Linguistics 21, 9–34. Fodor J. D. 1988. “On modularity in syntactic processing”. Journal of Psycholinguistic Research 17, 125–168. Fodor J. D. 1989. “Empty Categories in Sentence Processing”. Language and Cognitive Processes 4, 155–209. Fodor J. D. 1993. “Processing empty categories: A question of visibility”. In Cognitive Models of Speech Processing: The Second Sperlonga Meeting, G. Altmann & R. Shillock (eds), 351–400. Hillsdale: Erlbaum. Fodor J. D. 1995. “Comprehending sentence structure”. In Language: An Invitation to Cognitive Science. Vol 2, L. Gleitman & M. Liberman. Cambridge Massachusetts: MIT Press. Ford M., Bresnan J. & Kaplan R. 1982. “A competence based theory of syntactic closure”. In The Mental Representation of Grammatical Relations, J. Bresnan (ed) 727–796. Cambridge, Massachusetts: MIT Press. Ford M. & Dalrymple M. 1988. “A note on some psycholinguistic evidence and alternative grammars”. Cognition 29, 63–71. Forster K. 1979. “Levels of processing and the structure of the language processor”. In Sentence Processing, W. Cooper & E. Walker (eds), 27–86. Hillsdale: Erlbaum. Forster K. 1987. “Binding, plausibility and modularity”. In Modularity in Knowledge Representation and Natural Language Processing, J. Garﬁeld (ed), 63–82. Cambridge Massachusetts: MIT Press. Forster K. & Stevenson B. 1987. “Sentence matching and well-formedness”. Cognition 26, 171–186. Frazier L. 1985. “Modularity and the representational hypothesis”. In Proceedings of NELS 15, 131–145. University of Massachusetts. Frazier L. 1987a. Theories of sentence processing. In Modularity in Knowledge Representation and Natural Language Processing, J. Garﬁeld (ed), 291–308. Cambridge Massachusetts: MIT Press.

References 265

Frazier L. 1987b. Processing syntactic structures: Evidence from Dutch. Natural Language and Linguistic Theory 5, 519–559. Frazier L. 1993. “Processing Dutch sentence structures”. Journal of Psycholinguistic Research 22, 85–108. Frazier L. 1995. “Constraint satisfaction as a theory of sentence processing”. Journal of Psycholinguistic Research 24, 437–468. Frazier L. & Clifton C. 1989. “Successive cyclicity in the grammar and the parser”. Language and Cognitive Processes 4, 93–126. Frazier L., Clifton C. & Randall J. 1983. “Filling gaps: Decision principles and structure in sentence comprehension”. Cognition 13, 187–222. Frazier L. & Flores D’Arcais G. 1989. “Filler-driven parsing: A study of gap ﬁlling in Dutch”. Journal of Memory and Language 28, 331–344. Frazier L., Flores D’Arcais G. & Coolen R. 1993. “Processing discontinuous words: On the interface between lexical and syntactic processing”. Cognition 47, 219–249. Frazier L. & Rayner K. 1982. “Making and correcting errors during sentence comprehension: Eye movements in analysis of structurally ambiguous sentences”. Cognitive Psychology 14, 178–210. Freedman S. & Forster K. 1985. “The psychological status of overgenerated sentences”. Cognition 19, 101–131. Friederici A., Hahne A. & Mecklinger A. 1996. “Temporal structure of syntactic parsing: Early and late event-related brain potential eﬀects”. Journal of Experimental Psychology: Memory and Cognition 22, 1219–1248. Friederici A. & Mecklinger A. 1996. “Syntactic parsing as revealed by brain responses: Firstpass and second-pass parsing processes”. Journal of Psycholinguistic Research 25, 157–178. Friederici A., Pfeifer E. & Hahne A. 1996. “Event-related brain potentials during natural speech processing: Eﬀects of semantic, morphological, and syntactic violations”. Cognitive Brain Research 1, 183–192. Friedman D., Simson R., Ritter W. & Rapin I. 1975. “The late positive component (P300) and information processing in sentences”. Electroencephalography and Clinical Neurophysiology 38, 255–262. Garﬁeld J. (ed) 1987. Modularity in Knowledge Representation and Natural Language Processing. Cambridge, Massachusetts: MIT Press. Garnsey S. 1993. “Event-related brain potentials in the study of language: An introduction”. Language and Cognitive Processes 8, 337–356. Garnsey S., Tanenhaus M. & Chapman R. 1989. “Evoked potentials and the study of sentence comprehension”. Journal of Psycholinguistic Research 18, 51–60. Gibson E. & Hickok G. 1993. “Sentence processing with empty categories”. Language and Cognitive Processes 8, 147–161. Gibson E., Hickok G. & Schütze C. 1994. “Processing empty categories: A parallel approach”. Journal of Psycholinguistic Research 23, 381–405. Gibson E. & Pearlmutter N. 1994. “A corpus-based analysis of psycholinguistic constraints on prepositional phrase attachment”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 181–198. Hillsdale: Erlbaum. Goodall G. 1987. Parallel Structures in Syntax. Cambridge: CUP.

266 References

Goodluck H., Finney M. & Sedivy J. 1991. “Sentence completeness and ﬁller-gap dependency parsing”. In OTS Yearbook 1991, P. Coopmans, B. Schouten & W. Zonneveld (eds), 19–31. Utrecht: Utrecht University Press. Goodluck H. & Rochemont M. 1992. Island Constraints: Theory, acquisition and processing. Dordrecht: Kluwer. Gorrell P. 1989. “Establishing the loci of serial and parallel eﬀects in syntactic processing”. Journal of Psycholinguistic Research 18, 61–73. Gorrell P. 1992. Syntax and Perception. Ms., University of Maryland. Gorrell P. 1993. “Evaluating the Direct Association Hypothesis: a reply to Pickering & Barry (1991)”. Language and Cognitive Processes 8, 129–146. Gorrell P. 1995. Syntax and Parsing. Cambridge: CUP. Grewendorf G. 1988. Aspekte der deutschen Syntax. Tübingen: Narr. Grewendorf G. 1995. “Syntactic Sketches: German”. In Syntax: An International Handbook of Contemporary Research, Volume 2, J. Jacobs, A. von Stechow, W. Sternefeld & T. Vennemann (eds), 1288–1319. Berlin: De Gruyter. Grewendorf G. & Sabel J. 1999. Scrambling in German and Japanese: Adjunction versus multiple speciﬁers. Ms., University of Frankfurt. Grewendorf G. & Sternefeld W. 1990. (eds) Scrambling and Barriers. Amsterdam: John Benjamins. Gunji T. 1986. Japanese Phrase Structure Grammar. Dordrecht: Reidel. Gunter T., Vos S., & Mulder G. 1995. Syntactic violations and ERPs: P600 or P3b?. Poster presented at the CUNY Sentence Processing Conference, Tucson, Arizona. Haegemann L. 1994. Introduction to Government and Binding Theory (2nd edition). Oxford: Blackwell. Hagoort P. & Brown C. 1994. “Brain responses to lexical ambiguity resolution and parsing”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 435–80. Hillsdale: Erlbaum. Hagoort P., Brown C. & Groothusen J. 1993. “The syntactic positive shift as an ERP measure of syntactic processing”. Language and Cognitive Processes 8, 439–483. Haider H. 1992. Branching and discharge. Working papers of SFB 340 No.23, Universität Stuttgart. Haider H. 1993. Deutsche Syntax — Generativ. Tübingen: Narr. Haider H. 1994. “Scrambling: Locality, economy and directionality”. Proceedings of the Tokyo Festival on Scrambling, S. Tonoike (ed). Haider H. 1997. “Precedence among predicates”. Journal of Comparative Germanic Linguistics 1.1, 3–41. Hale K. 1980. “Remarks on Japanese phrase structure: Comments on the papers on Japanese syntax”. In Theoretical Issues in Japanese Syntax, Y. Otsu & A. Farmer (eds). MIT Working Papers 2. Hale K. 1981. On the position of Warlpiri in a typology of the base. Bloomington: Indiana University Linguistics Club. Hale K. & Keyser S. 1993. “On argument structure and the lexical expression of semantic relations”. In The View from Building 20, K. Hale & S. Keyser (eds), 53–110. Cambridge, Massachusetts: MIT Press.

References 267

Hickok G. 1993. “Parallel parsing: Evidence from reactivation in garden path sentences”. Journal of Psycholinguistic Research 22, 239–249. Hickok G., Canseco-Gonzalez E., Zurif E. & Grimshaw J. 1992. “Modularity in locating wh-gaps”. Journal of Psycholinguistic Research 21, 545–561. Hinrichs E. & Nakazawa T. 1994. “Linearizing AUXs in German verbal complexes”. In German Grammar in HPSG, J. Nerbonne, K. Netter & C. Pollard (eds), 11–37. CSLI: Stanford. Holmes V., Stowe L. & Cupples L. 1980. “Lexical expectations in parsing complement-verb sentences”. Journal of Memory and Language 28, 668–689. Hukari T. & Levine R. 1994. “Adjunct extraction”. Papers of the West Coast Conference on Formal Linguistics 12, 283–298. Jackendoﬀ R. 1977. X¢ Syntax. Cambridge, Massachusetts: MIT Press. Jackendoﬀ R. 1990. “On Larson’s treatment of the double object construction”. Linguistic Inquiry 21, 427–456. Jacobson P. 1987. “Phrase structure, grammatical relations, and discontinuous constituents”. In Syntax and Semantics 20, G. Huck & A. Ojeda (eds). New York: Academic Press. Jasper H. 1958 The ten-twenty electrode system of the International Federation. Electroencephalography and Clinical Neurophysiology 10, 371–375. Johnson R. 1986. “A triarchic model of P300 amplitude”. Psychophysiology 23, 367–384. Just M. & Carpenter P. 1987. The Psychology of Reading and Language Comprehension. Boston: Allyn & Bacon. Kaplan R. & Zaenen A. 1989. “Long-distance dependencies, constituent structure and functional uncertainty”. In Alternative Conceptions of Phrase Structure M. Baltin & A. Kroch, 17–42. Chicago: University of Chicago Press. Kathol A. 1990. “A uniform approach to V2 in German”. Proceedings of NELS 20, 244–254. University of Massachusetts. Kayne R. 1984. Connectedness and Binary Branching. Dordrecht: Foris. Kayne R. 1989. “Facets of Romance past particle agreement”. In Dialect Variation in the Theory of Grammar, P. Benincà (ed). Dordrecht: Foris. Kayne R. 1994. The Antisymmetry of Syntax. Cambridge, Massachusetts: MIT Press. Kennedy A., Murray W., Jennings F. & Reid C. 1989. “Parsing complements: Comments on the generality of the principle of minimal attachment”. Language and Cognitive Processes 4, 51–76. King J. & Kutas M. 1995. “Who did what when? Using word- and clause-level ERPs to monitor working memory usage in reading”. Journal of Cognitive Neuroscience 7, 376–395. Kiss K. 1994. “Scrambling as the base generation of random complement order”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 221–256. Berlin: Mouton de Gruyter. Kitahara H. 1997. Elementary operations and optimal derivations. Cambridge, Massachusetts: MIT Press. Kluender R. & Kutas M. 1993. “Bridging the gap: Evidence from ERPs on the processing of unbounded dependencies”. Journal of Cognitive Neuroscience 5, 196–214.

268 References

Kluender R. & Münte T. 1998. Subject/object asymmetries: ERPs to grammatical and ungrammatical wh-questions in German. Paper at CUNY98 conference on sentence processing. Koster J. 1978. Locality Principles in Syntax. Dordrecht: Foris. Kurtzman H. 1985. Studies in syntactic ambiguity resolution. PhD dissertation, MIT. Kutas M. & Hillyard S. 1983. “Event-related brain potentials to grammatical errors and semantic anomalies”. Memory and Cognition 11, 539–550. Kutas M. & Petten C. van 1994. “Psycholinguistics electriﬁed”. In Handbook of Psycholinguistics, M. Gernsbacher (ed), 83–143. San Diego: Academic Press. Larson R. 1988. “On the double object construction”. Linguistic Inquiry 19, 335–391. Larson R. 1990. “The double object construction revisited: A reply to Jackendoﬀ”. Linguistic Inquiry 21, 589–632. Lasnik H. & Saito M. 1984. “On the nature of proper government”. Linguistic Inquiry 15, 235–289. Lee Y.-S. & Santorini B. 1994. “Towards resolving Webelhuth’s paradox: Evidence from German and Korean”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 257–300. Berlin: Mouton de Gruyter. Lenerz J. 1977. Zur Abfolge nominaler Satzglieder im Deutschen. Tübingen: Narr. Lewis J., Shapiro L., Afton R. & Tuller B. 1996. The magnitude of CMLP as a function of the verb-object relationship in a sentence. Poster given at the 9th CUNY conference on sentence processing. Love T. & Swinney D. 1996. “Coreference processing and levels of analysis in object-relative constructions: Demonstration of antecedent reactivation with the cross-modal priming paradigm”. Journal of Psycholinguistic Research 25, 5–24. MacDonald M, Pearlmutter N. & Seidenberg M. 1994. “Syntactic ambiguity resolution as lexical ambiguity resolution”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 123–154. Hillsdale: Erlbaum. MacDonald M. 1989. “Priming eﬀects from gaps to antecedents”. Language and Cognitive Processes 4, 35–56. MacWhinney B. & Bates E. 1989. The Crosslinguistic Study of Sentence Processing. Cambridge: CUP. Mahajan Anoop 1990. The A/È distinction and movement theory. PhD dissertation: MIT. Mahajan A. 1994. “Toward a uniﬁed theory of scrambling”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 301–330. Berlin: Mouton de Gruyter. Marslen-Wilson W. 1975. “Sentence perception as an active parallel process”. Science 189, 226–228. Marslen-Wilson W. & Tyler L. 1980. “The temporal structure of spoken language understanding”. Cognition 8, 1–71. Marslen-Wilson W. & Tyler L. 1987. “Against modularity”. In Modularity in Knowledge Representation and Natural Language Processing, J. Garﬁeld (ed), 37–62. Cambridge, Massachusetts: MIT Press. Mazuka R. & Nagai N. 1995. Japanese Sentence Processing. NY: Erlbaum.

References 269

McClelland J. 1987. “The case for interactionism in language processing”. In Attention and performance Vol XII: The Psychology of Reading, Coltheart M. (ed), 3–36. Hillsdale: Erlbaum. McClelland J. & Kawamoto A. 1986. “Mechanisms of sentence processing: Assigning roles to constituents of sentences”. In Parallel Distributed Processing Part II: Psychological and Biological Models, J. McClelland & D. Rumelhart (eds), 272–325. Cambridge, Massachusetts: MIT Press. McElree B. & Bever T. 1989. “The psychological reality of linguistically deﬁned gaps”. Journal of Psycholinguistic Research 18, 21–35. McKinnon R. & Osterhout L. 1996. “Constraints on movement phenomena in sentence processing: evidence from event related brain potentials”. Language and Cognitive Processes 11, 495–524. McKoon G. & Ratcliﬀ R. 1994. “Sentential context and on-line lexical decision tasks”. Language, Memory and Cognition 20, 1239–1243. Mecklinger A., Schriefers H., Steinhauer K. & Friederici A. 1995. “Processing relative clauses varying on syntactic and semantic dimensions: An analysis with ERPs”. Memory and Cognition 23, 477–494. Mergel B. 1994. Wortstellungsalternation im deutschen Mittelfeld. Arbeitspapiere des Sonderforschungsbereichs 340 Nr. 56. Meyer D. & Schvanenveldt R. 1971. “Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations”. Journal of Experimental Psychology 90, 227–234. Miller G. & Chomsky N. 1963. “Finitary models of language users”. In Handbook of Mathematical Psychology vol 2, R. Luce, R. Bush & E. Galanter (eds), 419–492. New York: Wiley. Mitchell D. 1987. “Lexical guidance in human parsing: Locus and processing characteristics”. In Coultheart M. (ed) Attention and performance XII: The Psychology of Reading, 601–618. Hillsdale: Erlbaum. Mitchell D. 1994. “Sentence Parsing”. In Handbook of Psycholinguistics, M. Gernsbacher (ed), 375–410. New York: Academic Press. Müller H., King J. & Kutas M. 1997. “Event-related potentials elicited by spoken relative clauses”. Cognitive Brain Research 5, 193–203. Müller G. & Sternefeld W. 1994. “Scrambling as È-movement”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 331–386. Berlin: Mouton de Gruyter. Münte T, Heinze H., Matzke M., Wieringa B. & Johannes S. 1998. “Brain potentials and syntactic violations revisited: No evidence for speciﬁcity of the syntactic positive shift”. Neuropsychologia 36, 217–226. Münte T, Matzke M. & Johannes S. 1998. “Brain activity associated with syntactic incongruencies in words and pseudowords”. Journal of Cognitive Neuroscience 9, 318–332. Nagel H., Shapiro L. & Nawy R. 1994. “Prosody and the processing of ﬁller-gap sentences”. Journal of Psycholinguistic Research 23, 473–485. Nakano Y. 1998. Two experiments testing for antecedent reactivation in Japanese scrambled structures. Ms., University of Essex.

270 References

Neeleman A. 1994. “Scrambling as a D-structure phenomenon”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 387–430. Berlin: Mouton de Gruyter. Netter K. 1992. “On non-head non-movement”. In Proceedings of KONVENS 92, G. Gorz (ed), 218–287. Berlin: Springer. Nicol J. 1993. “Reconsidering Reactivation”. In Cognitive Models of Speech Processing: The Second Sperlonga Meeting, G. Altmann & R. Shillock (eds), 321–350. Hove: Erlbaum. Nicol J., Fodor J. & Swinney D. 1994. “Using cross-modal lexical decision tasks to investigate sentence processing”. Journal of Experimental Psychology: Learning, Memory and Cognition 20, 1229–1238. Nicol J. & Pickering M. 1993. “Processing syntactically ambiguous sentences: Evidence from semantic priming”. Journal of Psycholinguistic Research 22, 1993. Nicol J. & Swinney D. 1989. “The rôle of structure and co-reference assignment during sentence comprehension”. Journal of Psycholinguistic Research 18, 5–19. Osterhout L. 1994. “Event-related brain potentials as tools for comprehending language comprehension”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 15–44. Hillsdale: Erlbaum. Osterhout L. & Holcomb P. 1992. “Event-related brain potentials elicited by syntactic anomaly”. Journal of Memory and Language 31, 785–806. Osterhout L. & Holcomb P. 1993. “Event-related potentials and syntactic anomaly: Evidence of anomaly detection during the perception of continuous speech”. Language and Cognitive Processes 8, 413–437. Osterhout L. & Holcomb P. 1995. “Event-related potentials and language comprehension”. In Electrophysiology of Mind: Event-related Brain Potentials and Cognition, M. Rugg & M. Coles (eds), 175–215. Oxford: OUP. Osterhout L., Holcomb P. & Swinney D. 1994. “Brain potentials elicited by garden path sentences: Evidence of the application of verb information during parsing”. Journal of Experimental Psychology: Learning, Memory and Cognition 20, 786–803. Osterhout L., McKinnon R., Bersick M. and Corey V. 1996. “On the language-speciﬁcity of the brain response to syntactic anomalies: Is the syntactic positive shift a member of the P300 family?” Journal of Cognitive Neuroscience 8, 507–526. Osterhout L. & Mobley L. 1995. “Event-related brain potentials elicited by failure to agree”. Journal of Memory and Language 34, 739–773. Osterhout L. & Swinney D. 1989. “On the role of the simplicity heuristic in language processing: Evidence form structural and inferential processing”. Journal of Psycholinguistic Research 18, 553–562. Osterhout L. & Swinney D. 1993. “On the temporal course of gap-ﬁlling during the comprehension of verbal passives”. Journal of Psycholinguistic Research 22, 273–286. Penke M., Weyerts H., Gross M., Zander E., Münte T. & Clahsen H. 1997. “How the brain processes complex words: An ERP study of German verb inﬂections”. Cognitive Brain Research 6, 37–52. Pickering M. 1993. “Direct Association and sentence processing: a reply to Gibson & Hickok”. Language and Cognitive Processes 8, 163–196. Pickering M. 1994. “Processing local and unbounded dependencies: a uniﬁed account”. Journal of Psycholinguistic Research 23, 323–352.

References 271

Pickering M. & Barry G. 1991. “Sentence processing without empty categories”. Language and Cognitive Processes 6, 229–259. Pickering M., Barton S. & Shillcock R. 1994. “Unbounded dependencies, island constraints, and processing complexity”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 199–224. Hillsdale: Erlbaum. Pinker S. & Prince A. 1988. “Language and connectionism: Analysis of a Parallel Distributed Processing model of language acquisition”. Cognition 28, 78–193. Platzack C. & Holmberg A 1989. “The role of AGR and ﬁniteness”. Working Papers in Scandinavian Syntax 43, 51–76. Pollard C. 1996. “On head non-movement”. In Discontinuous Constituency, H. Bunt & A. van Horck (eds), 279–306. Berlin: Mouton de Gruyter. Pollard C. & Sag I. 1987. Information-based Syntax and Semantics. Stanford: CSLI. Pollard C. & Sag I. 1994. Head-driven Phrase Structure Grammar. Chicago: University of Chicago Press. Pollock J.-Y. 1989. “Verb movement, Universal Grammar, and the structure of IP”. Linguistic Inquiry 20, 365–424. Poole G. 1996. “Optional movement in the Minimalist Program”. In Minimal Ideas: Syntactic Studies in the Minimalist Framework, 199–216. W. Abraham, S. Epstein, H. Thráinsson & J.-W. Zwart (eds). Amsterdam: John Benjamins. Pritchett B. 1988. “Garden path phenomena and the grammatical basis of language processing”. Language 64, 539–576. Prichett B. 1991. “Head position and parsing ambiguity”. Journal of Psycholinguistic Research 20, 251–270. Prichett B. 1992a. Grammatical Competence and Parsing Performance. Chicago: Chicago University Press. Prichett B. 1992b. Parsing with grammar: Islands, heads and and garden paths. In Island Constraints: Theory, Acquisition and Processing, H. Goodluck & M. Rochemont (eds), 321–350. Dordrecht: Kluwer. Pullum G. & Postal P. 1979. “On an inadequate defence of trace theory”. Linguistic Inquiry 10, 689–706. Radford A. 1997. Syntactic Theory and the Structure of English. Cambridge: CUP. Ratcliﬀ R. & McKoon G. 1981. “Automatic and strategic priming in recognition”. Journal of Verbal Learning and Verbal Behaviour 20, 204–215. Rayner K., Carlson M. & Frazier L. 1983. “The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences”. Journal of Verbal Learning and Verbal Behaviour 22, 358–374. Reape M. 1994. “Domain Union and word-order variation in German”. In German Grammar in HPSG, J. Nerbon, K. Netter & C. Pollard (eds), 151–97. Stanford: CSLI. Regan D. 1989. Human brain electrophysiology: Evoked potentials and evoked magnetic ﬁelds in science and medicine. New York: Elsevier. Reis M. 1980. “On justifying topological frames: Positional ﬁeld and the order of non-verbal constituents in German”. Revue de Linguistique 9, 5–82. Reis M. 1985. “Satzeinleitende Strukturen im Deutschen”. Über COMP, Haupt- und Nebensätze. W-Bewegung und die Doppelkopfanalyse. In Erklärende Syntax des Deutschen [Studien zur Deutschen Syntax 25], W. Abraham (ed), 271–311. Tübingen: Narr.

272 References

Reis M. 1986. Die Stellung der Verbargumente im Deutschen. Stilübungen zum Grammatik: Pragmatik-Verhältnis. Paper at 5th Lunder Symposium on Speech and Pragmatics, 12–16 May 1986. Rizzi L. 1990. Relativized Minimality. Cambridge, Massachusetts: MIT Press. Ross J. 1967. Constraints on variables in syntax. PhD dissertation, MIT (Reproduced by the Indiana University Linguistics Club). Rothstein S. 1983. The Syntactic Forms of Predication. PhD dissertation, MIT (Reproduced by the Indiana University Linguistics Club). Rumelhart D., McClelland J. & the PDP Research Group (eds) 1986. Parallel Distributed Processing, Part I: Foundations. Cambridge, Massachusetts: MIT Press. Sag I. & Fodor J. 1993. “Agreement, coindexing and reactivation: a reply to Barss”. In Cognitive Models of Speech Processing: The Second Sperlonga Meeting, G. Altmann & R. Shillock (eds), 453–466. Hillsdale: Erlbaum. Sag I. & Fodor J. 1995. “Extraction without traces”. Proceedings of the 13th West Coast Conference on Formal Linguistics, 365–384. Saito M. 1989. “Scrambling as semantically vacuous È-movement”. In Alternative Conceptions of Phrase Structure, M. Baltin & A. Kroch (eds), 182–200. Chicago: Chicago University Press. Saito M. 1992. “Long-distance scrambling in Japanese”. Journal of East Asian Linguistics 1, 69–118. Samar V. & Berent G. 1991. “Be is a raising verb: Psycholinguistic evidence”. Journal of Psycholinguistic Research 20, 419–443. Scherpenisse W. 1986. The connection between base structures and linearization restrictions in German and Dutch. European University Studies Series 21, Linguistics vol. 47. Frankfurt am Main: Lang. Schreuder R. 1990. “Lexical processing of verbs with separable particles”. In Yearbook of Morphology 3, A. Jongman & A. Lahiri (eds), 81–86. Dordrecht: Foris. Schwartz B. & Vikner S. 1989. “All verb second clauses are CPs”. Working papers in Scandinavian syntax 43, 27–50. Shapiro L., Nagel N. & Levine B. 1993. “Preferences for a verb’s complements and their use in sentence processing”. Journal of Memory and Language 32, 96–114. Shillock R. 1982. “The on-line resolution of pronominal anaphora”. Language and Speech 25 385–401. Sportiche D. 1988. “A theory of ﬂoating quantiﬁers and its corollaries for constituent structure”. Linguistic Inquiry 19, 425–449. Stechow A. von 1991. Lexical decomposition and scope. Ms., University of Konstanz. Stechow A. von & Sternefeld W. 1988. Bausteine syntaktischen Wissens. Opladen: Westdeutscher Verlag. Steedman M. 1987. “Combinatory grammars and parasitic gaps”. Natural Language and Linguistic Theory 5, 403–439. Stern C., Prather P., Swinney D. & Zurif E. 1991. “The time course of automatic lexical access and aging”. Brain and Language 40, 359–372. Stevenson S. 1994. “Competition and recency in a hybrid network model of syntactic disambiguation”. Journal of Psycholinguistic Research 23, 295–322.

References 273

Stiebels B & Wunderlich D. 1994. “Morphology feeds syntax: The case of particle verbs”. Linguistics 32, 913–968. Stowe L. 1986. “Parsing wh-constructions: Evidence for on-line gap location”. Language and Cognitive Processes 1, 227–246. Stowe L. 1992. “The processing implications of syntactic constraints: The sentence matching debate”. In Island Constraints: Theory, Acquisition and Processing, H. Goodluck & M. Rochemont, 419–444. Dordrecht: Kluwer. Stowell T. 1981. Origins of phrase structure. PhD dissertation, MIT. Swinney D., Ford M. & Bresnan J. 1989. “Coreference assignment during sentence processing”. In Language Structure and Processing, B. Grosz, R. Kaplan, M. Macken & I. Sag (eds). Stanford: CSLI. Swinney D., Onifer W., Prather P. & Hirschkowitz M. 1979. “Semantic facilitation across sensory modalities in the processing of individual words and sentences”. Memory and Cognition 7, 159–165. Tanenhaus M., Boland J., Garnsey S. & Carlson G. 1989. “Lexical structure in parsing longdistance dependencies”. Journal of Psycholinguistic Research 18, 37–49. Tanenhaus M., Boland J., Mauner G. & Carlson G. 1993. “Lexical information: thematic structure in parsing and interpretation”. In Cognitive Models of Speech Processing: The Second Sperlonga Meeting, G. Altmann & R. Shillock (eds), 297–320. Hillsdale: Erlbaum. Tanenhaus M., Carlson G. & Seidenberg M. 1985. “Do listeners compute syntactic representations?” In Natural Language Parsing: Psycholinguistic, Theoretical and Computational Perspectives, A. Zwicky, L. Kartunen & D. Dowty. (eds), 359–408. London and New York: CUP. Tanenhaus M., Stowe L. & Carlson G. 1985. “The interactions of lexical expectation and pragmatics in parsing ﬁller-gap constructions”. In Proceedings of the Seventh Annual Cognitive Science Society Meeting, 361–365. London: Erlbaum. Tonoike S. 1994. “On scrambling: Scrambling as a base-generated scopal construction”. In Proceedings of the Tokyo Festival on Scrambling, Tonoike S. (ed). Thiersch C. 1978. Topics in German syntax. PhD dissertation MIT. Thiersch C. 1982. “A note on scrambling and the existence of VP”. Wiener Linguistische Gazette vol 27–28, 83–95. Travis L. 1984. Parameters and eﬀects of word order variation. PhD dissertation, MIT. Travis L. 1991. “Parameters of phrase structure and V2 phenomena”. In Principles and Parameters in Comparative Grammar, Freidin R. (ed), 339–364. Cambridge, Massachusetts: MIT Press. Trueswell J. & Tanenhaus M. 1994. “Towards a lexical framework for constraint-based syntactic ambiguity resolution”. In Perspectives on Sentence Processing, C. Clifton, L. Frazier & K. Rayner (eds), 155–180. Hillsdale: Erlbaum. Trueswell J., Tanenhaus M. & Kello C. 1993. “Verb-speciﬁc constraints in sentence processing: Separating eﬀects of verbal preference from garden paths”. Journal of Experimental Psychology: Learning, Memory and Cognition 19, 528–553. Tyler L. & Marslen-Wilson W. 1977. “The on-line eﬀects of semantic context on syntactic processing”. Journal of Verbal Learning and Verbal Behaviour 16, 683–692.

274 References

Ueyama A. 1994. “Against the A/È-movement dichotomy”. In Studies on Scrambling: Movement and Non-movement Approaches to Free Word Order Phenomena, N. Corver & H. van Riemsdijk, 459–486. Berlin: Mouton de Gruyter. Uszkoreit H. 1984. Word Order and Constituent Structure in German, [CLSI Lecture notes no. 8]. Stanford: CSLI. Uszkoreit H. 1986. “Constraints on order”. Linguistics 24, 883–906. Vikner S. 1997. “The interpretation of object shift, Optimality Theory and Minimalism”. Working Papers in Scandinavian Syntax 60, 1–24. Walenski M. 1998. On the syntactic status of passive NP-trace. Paper presented at CUNY conference on sentence processing. Walenski M. & Fodor J. 1995. On the syntactic status of passive NP-trace. Poster presented at 19th CUNYForum. Wanner E. & Maratsos M. 1978. “An ATN approach to comprehension”. In Linguistic Theory and Psychological Reality, M. Halle, J. Bresnan & G. Miller (eds), 119–161. Cambridge, Massachusetts: MIT Press. Wasow T. 1979. Anaphora in Generative Grammar. Ghent: Editions Story-Scientia. Webelhuth G. 1989. Syntactic saturation phenomena and the Germanic languages. PhD dissertation, University of Massachusetts. Weinberg A. 1992. “Parameters in the theory of sentence Processing: Minimal Commitment Theory Goes East”. Journal of Psycholinguistic Research 2, 339–64. Weyerts H., Penke M., Dohrn U., Clahsen H. & Münte T. 1997. “Brain potentials indicate diﬀerences between regular and irregular German noun plurals”. NeuroReport 8, 958–962. Zwart J. 1997. Morphosyntax of verb movement: A minimalist approach to the syntax of Dutch. Dordrecht: Kluwer.

AUTHOR ""

TITLE "Index"

SUBJECT "Linguistik Aktuell/Linguistics Today, Volume 43"

KEYWORDS ""

SIZE HEIGHT "220"

WIDTH "150"

VOFFSET "4">

Index

A Active Filler Strategy 60, 62, 75 adjunct 21, 30, 46, 47, 91, 112, 156, 168, 181, 185-187, 228 AgrIOP 113 AgrOP 31, 40-42, 113 ambiguity 44, 50-53, 58, 60, 62, 74, 76, 78, 89, 90, 94, 97, 98, 122, 179 anaphor 5, 14, 16, 19, 22, 78, 81-84, 100, 101, 119, 136, 144, 146, 148, 150, 151, 152, 155, 159, 163, 165, 167, 170, 171, 189, 191, 204, 221, 223, 229 antecedent reactivation (see also TRA) 4, 43, 86, 88, 94, 99, 103, 104, 107, 108, 109, 116, 123, 125-127, 130, 132, 133, 135-138, 143, 144, 148, 149, 156, 160, 167, 170, 180, 189, 221, 222, 224, 226 B Basic Branching Conjecture 39 binary branching 28, 37, 44, 115 binding 5, 6, 9, 11, 12, 14-17, 22, 25, 26, 54, 57, 76, 82, 87, 193, 198, 205, 206, 216, 219, 220, 224, 230, 231 C Case 12, 15-18, 21, 38, 40, 42, 54, 87, 89, 113, 115, 141, 175, 188 chain 12, 15-17, 62, 84, 85, 100, 141, 188, 193 CMLP (cross-modal lexical priming) 72, 75, 78, 82-84, 88, 95, 105, 108, 119, 121, 130, 144, 147, 148, 150, 153,

155, 156, 158, 178, 191, 224, 226, 229 correctability 200, 202, 204, 207, 215, 230, 231 D Depth of Processing Account (DOP) 104, 105-107, 122, 125, 127, 130, 133, 134, 138, 156, 157, 160, 161, 164, 167, 168, 190, 225, 226 Direct Association Hypothesis (DAH) 89, 90, 91, 93, 94, 96, 98, 99, 105, 108, 109, 121, 124, 130, 132, 135, 137, 138, 151, 155, 158, 160, 166, 168, 181, 190, 226 E ECs in HPSG 21, 26, 27, 171, 173, 175 ECs in PPT 15, 16, 19, 20, 170, 173 embedded anomaly technique 60, 178 empirically adequate 9, 109, 124, 132, 136, 138, 174, 188, 191, 219, 223, 227, 228, 231 Empty Category Principle 18 English 4, 12, 26, 28-31, 34, 38, 41, 46, 89, 109, 125, 127, 157-159, 168, 180, 222 equi 5, 12, 21, 22, 141, 142, 146, 157, 158, 159, 164, 169, 171, 174, 177, 180, 181, 182, 185, 187, 188, 190, 191, 227 Event-Related Potentials (ERP) 72, 116, 126, 178, 180, 182, 185, 187, 188, 191, 227, 229

276 Index

F ﬁlled gap eﬀect 61, 62, 64-66, 68, 70, 73, 91 French 12, 41, 67, 127 functional category 40, 113 G garden path 51, 52, 55, 61, 64, 67, 70, 76, 91, 179 Generalised Phrase Structure Grammar (GPSG) 77, 115, 169 government 15, 39 grammaticality judgement 58, 59, 65, 70, 197 H Head-driven Phrase Structure Grammar (HPSG) 1, 2, 6, 20-22, 24, 26, 27, 45, 47, 48, 54, 85, 87, 89, 102, 106, 108, 115, 128, 131, 137, 141, 169-173, 175-177, 188, 191, 193, 206, 220, 223, 224, 225, 227, 231 I interactive parser 51, 55, 56, 62, 63, 220 intermediate trace 18, 40, 43, 44, 139, 225 island constraint 6, 64-66, 68, 69, 76 L lexical expectation 59 light verb 37, 40 Linear Correspondence Axiom 34, 38, 39, 137 linear precedence 22, 46, 93, 115, 173, 224 locality 193, 204, 206, 219, 231 M methodology 2, 5, 68, 71, 83, 88, 96, 108, 116, 119, 126, 145, 147, 152, 153, 158, 164, 178, 187, 191, 194, 215, 222, 226, 229 Minimal Chain Principle 62, 100 Minimalist Program 2, 9, 12, 13, 15, 17, 19, 38, 113, 115

modular parser 6, 51, 55, 56, 60, 62, 64, 74, 79, 194, 198, 207, 218, 219, 232 Most Recent Filler Strategy 76-78, 142 N NONLOCAL features 23, 25, 87, 89, 170, 174, 205, 220, 224, 225 NP-trace 5, 14, 16, 18, 21, 27, 44, 81, 104, 144, 145, 147-149, 153, 155, 159, 169, 180, 181, 188, 191, 226, 228 O operator 16, 19, 29, 41, 76, 119, 159, 205, 219 ordered NONLOCAL processing 224, 225 P parallel processing 51, 52, 55, 56, 63 particle verb 127, 128, 131, 133, 138, 222 passive 5, 13, 16, 27, 44, 78, 104, 141, 143, 145-147, 154, 159, 162, 164, 165, 167, 168, 170 phonetic features 15, 18, 20, 21, 45, 75, 81, 84, 86, 90, 173, 174, 195, 224 pragmatic information 60, 73, 74, 82, 88, 112, 115, 171 Principles and Parameters Theory (PPT) 1, 2, 5, 9, 10, 13, 16, 19, 21, 22, 24, 27, 28, 36, 40, 47, 48, 56, 84, 85, 87, 108, 109, 112, 115, 127, 128, 130, 138, 141, 145, 148, 153, 156, 159, 169, 171, 172, 174, 175, 177, 180, 188, 190, 193, 198, 205, 208, 216, 219, 223-225, 227 PRO 5, 11, 12, 14, 19, 21, 22, 81, 142, 144, 145, 147, 149, 159, 162, 165, 167, 170, 174, 175, 180, 188, 191, 221, 226, 228 probe recognition 5, 104, 105, 116, 147, 151-156, 158, 161, 163, 164, 166, 168, 190, 191, 226, 227 processing complexity 4, 50, 52, 63, 73, 92, 100, 104, 108, 123, 135, 156, 158, 160, 161, 163, 165, 167, 187, 188, 190, 196, 226, 227

Index 277

pronominal 14, 112 psychological reality 2, 19, 56, 78, 79, 82, 94, 137, 138, 147, 155, 158, 174, 189, 191, 225, 229 R raising 5, 13, 16, 22, 27, 30, 31, 38, 41, 42, 44, 141, 142, 146, 149, 157, 158, 159, 162, 163, 167, 169, 171, 174, 177, 180-182, 185-188, 191, 227 reanalysis 53, 61, 95, 99, 135, 179, 180, 187, 188, 191 reﬂexive 12, 15, 22, 100, 148, 154, 171 relative clause 16, 53, 61, 74, 83, 161, 163, 179 S scrambling 4, 110-112, 114-116, 119, 122, 124, 130, 136, 138, 223 semantic features 18, 47, 85, 86, 102, 103, 137, 171, 173, 176 Semantic Processing Account 4 Semantic Processing Account (SPA) 100, 103, 108, 110, 121, 124, 130, 135, 137, 151, 153, 155, 158, 167, 190, 225 serial parser 51, 53, 60 Spanish 22, 26, 153 structure-sharing 20, 21, 23, 47, 103, 142, 189, 194, 220, 224 T Theta Criterion 11, 15, 16, 37, 63, 73, 87, 89, 115, 141, 146, 172, 188

time course of processing 50, 58, 96, 100, 150, 152, 178 Trace Reactivation Account (TRA) 84, 85, 86-88, 91, 95, 96, 98, 99, 104, 107, 109, 121, 124, 130, 132, 138, 147, 149, 153, 155, 157, 168, 222, 225, 226 transparency 3, 27, 50 U unbounded dependency 23-25, 27, 60, 149, 157, 174, 193, 205, 221 Universal Grammar (UG) 10, 14, 35, 50, 89 unsaturated projection 2, 20, 21, 142 V verb second (V2) 4, 5, 29, 33, 45, 47, 108, 109, 111, 121, 124, 126, 127, 132, 133, 138, 160, 166, 205, 209, 210, 212, 214, 222 W wh-trace 5, 6, 14, 16, 18, 23, 81, 100, 101, 104, 141-151, 157, 159, 164, 169, 198, 216, 227, 229 word order 4, 12, 29, 31, 34, 44, 46, 50, 60, 110, 111, 115, 127, 132, 158, 159, 203, 205, 210, 215

In the series LINGUISTIK AKTUELL/LINGUISTICS TODAY (LA) the following titles have been published thus far, or are scheduled for publication: 1. KLAPPENBACH, Ruth (1911-1977): Studien zur Modernen Deutschen Lexikographie. Auswahl aus den Lexikographischen Arbeiten von Ruth Klappenbach, erweitert um drei Beiträge von Helene Malige-Klappenbach. 1980. 2. EHLICH, Konrad & Jochen REHBEIN: Augenkommunikation. Methodenreflexion und Beispielanalyse. 1982. 3. ABRAHAM, Werner (ed.): On the Formal Syntax of the Westgermania. Papers from the 3rd Groningen Grammar Talks (3e Groninger Grammatikgespräche), Groningen, January 1981. 1983. 4. ABRAHAM, Werner & Sjaak De MEIJ (eds): Topic, Focus and Configurationality. Papers from the 6th Groningen Grammar Talks, Groningen, 1984. 1986. 5. GREWENDORF, Günther and Wolfgang STERNEFELD (eds): Scrambling and Barriers. 1990. 6. BHATT, Christa, Elisabeth LÖBEL and Claudia SCHMIDT (eds): Syntactic Phrase Structure Phenomena in Noun Phrases and Sentences. 1989. 7. ÅFARLI, Tor A.: The Syntax of Norwegian Passive Constructions. 1992. 8. FANSELOW, Gisbert (ed.): The Parametrization of Universal Grammar. 1993. 9. GELDEREN, Elly van: The Rise of Functional Categories. 1993. 10. CINQUE, Guglielmo and Guiliana GIUSTI (eds): Advances in Roumanian Linguistics. 1995. 11. LUTZ, Uli and Jürgen PAFEL (eds): On Extraction and Extraposition in German. 1995. 12. ABRAHAM, W., S. EPSTEIN, H. THRÁINSSON and C.J.W. ZWART (eds): Minimal Ideas. Linguistic studies in the minimalist framework. 1996. 13. ALEXIADOU Artemis and T. Alan HALL (eds): Studies on Universal Grammar and Typological Variation. 1997. 14. ANAGNOSTOPOULOU, Elena, Henk VAN RIEMSDIJK and Frans ZWARTS (eds): Materials on Left Dislocation. 1997. 15. ROHRBACHER, Bernhard Wolfgang: Morphology-Driven Syntax. A theory of V to I raising and pro-drop. 1999. 16. LIU, FENG-HSI: Scope and Specificity. 1997. 17. BEERMAN, Dorothee, David LEBLANC and Henk van RIEMSDIJK (eds): Rightward Movement. 1997. 18. ALEXIADOU, Artemis: Adverb Placement. A case study in antisymmetric syntax. 1997. 19. JOSEFSSON, Gunlög: Minimal Words in a Minimal Syntax. Word formation in Swedish. 1998. 20. LAENZLINGER, Christopher: Comparative Studies in Word Order Variation. Adverbs, pronouns, and clause structure in Romance and Germanic. 1998. 21. KLEIN, Henny: Adverbs of Degree in Dutch and Related Languages. 1998. 22. ALEXIADOU, Artemis and Chris WILDER (eds): Possessors, Predicates and Movement in the Determiner Phrase. 1998. 23. GIANNAKIDOU, Anastasia: Polarity Sensitivity as (Non)Veridical Dependency. 1998. 24. REBUSCHI, Georges and Laurice TULLER (eds): The Grammar of Focus. 1999. 25. FELSER, Claudia: Verbal Complement Clauses. A minimalist study of direct perception constructions. 1999. 26. ACKEMA, Peter: Issues in Morphosyntax. 1999.

° 27. RUZICKA, Rudolf: Control in Grammar and Pragmatics. A cross-linguistic study. 1999. 28. HERMANS, Ben and Marc van OOSTENDORP (eds.): The Derivational Residue in Phonological Optimality Theory. 1999. 29. MIYAMOTO, Tadao: The Light Verb Construction in Japanese. The role of the verbal noun. 1999. 30. BEUKEMA, Frits and Marcel den DIKKEN (eds.): Clitic Phenomena in European Languages. 2000. 31. SVENONIUS, Peter (ed.): The Derivation of VO and OV. 2000. 32. ALEXIADOU, Artemis, Paul LAW, André MEINUNGER and Chris WILDER (eds.): The Syntax of Relative Clauses. 2000. 33. PUSKÁS, Genoveva: Word Order in Hungarian. The syntax of È-positions. 2000. 34. REULAND, Eric (ed.): Arguments and Case. Explaining Burzio’s Generalization. 2000. 35. HRÓARSDÓTTIR, Thorbjörg. Word Order Change in Icelandic. From OV to VO. 2000. 36. GERLACH, Birgit and Janet GRIJZENHOUT (eds.): Clitics in Phonology, Morphology and Syntax. 2000. 37. LUTZ, Uli, Gereon MÜLLER and Arnim von STECHOW (eds.): Wh-Scope Marking. 2000. 38. MEINUNGER, André: Syntactic Aspects of Topic and Comment. 2000. 39. GELDEREN, Elly van: A History of English Reflexive Pronouns. Person, ‘‘Self’’, and Interpretability. 2000. 40. HOEKSEMA, Jack, Hotze RULLMANN, Victor SANCHEZ-VALENCIA and Ton van der WOUDEN (eds.): Perspectives on Negation and Polarity Items. 2001. 41. ZELLER, Jochen : Particle Verbs and Local Domains. n.y.p. 42. ALEXIADOU, Artemis : Functional Structure in Nominals. Nominalization and ergativity. 2001. 43. FEATHERSTON, Sam: Empty Categories in Sentence Processing. 2001. 44. TAYLAN, Eser E. (ed.): The Verb in Turkish. n.y.p. 45. ABRAHAM, Werner and C. Jan-Wouter ZWART (eds.): Issues in Formal German(ic) Typology. n.y.p 46. PANAGIOTIDIS, Phoevos: Pronouns, clitics and Empty Nouns. ‘Pronominality’ and licensing in syntax. n.y.p.

Empty Categories in Sentence Processing

Bilingual Sentence Processing

Sentence Processing: A Crosslinguistic Perspective

Bilingual Sentence Processing: Relative Clause Attachment in English and Spanish

Empty

Empty

Working Memory in Sentence Comprehension: Processing Hindi Center Embeddings

Empty

Categories

Categories

Categories

Empty Nuclei in Korean

Categories

Categories

Deaf Sentence

Deaf Sentence

Deaf Sentence

Death Sentence

Cut Elimination in Categories

Coherence in Categories

Death Sentence

Death Sentence

Death Sentence

Death Sentence

Deaf Sentence

Death Sentence

Categories in continuum physics

Coherence in categories

Cut Elimination in Categories

Deaf Sentence

Deaf Sentence

Empty Categories in Sentence Processing

Bilingual Sentence Processing

Sentence Processing: A Crosslinguistic Perspective

Bilingual Sentence Processing: Relative Clause Attachment in English and Spanish

Empty

Empty

Working Memory in Sentence Comprehension: Processing Hindi Center Embeddings

Empty

Categories

Categories

Categories

Empty Nuclei in Korean

Categories

Categories

Deaf Sentence

Deaf Sentence

Deaf Sentence

Death Sentence

Cut Elimination in Categories

Coherence in Categories

Death Sentence

Death Sentence

Death Sentence

Death Sentence

Deaf Sentence

Death Sentence

Categories in continuum physics

Coherence in categories

Cut Elimination in Categories

Deaf Sentence

Deaf Sentence

Recommend Documents