Merging Features: Computation, Interpretation, and Acquisition

Merging Features This page intentionally left blank Merging Features Computation, Interpretation, and Acquisition ...

Author: Jose M. Brucart | Anna Gavarro | Jaume Sola

74 downloads 900 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

Merging Features

This page intentionally left blank

Merging Features Computation, Interpretation, and Acquisition

Edited by J O S É M . B RU C A RT, A N NA G AVA R R Ó , A N D JAUME SO LÀ

1

3

Great Clarendon Street, Oxford ox2 6dp Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide in Oxford New York Auckland Cape Town Dar es Salaam Hong Kong Karachi Kuala Lumpur Madrid Melbourne Mexico City Nairobi New Delhi Shanghai Taipei Toronto With offices in Argentina Austria Brazil Chile Czech Republic France Greece Guatemala Hungary Italy Japan Poland Portugal Singapore South Korea Switzerland Thailand Turkey Ukraine Vietnam Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries Published in the United States by Oxford University Press Inc., New York © 2009 organization and editorial matter José M. Brucart, Anna Gavarró, and Jaume Solà © 2009 the chapters their various authors The moral rights of the authors have been asserted Database right Oxford University Press (maker) First published 2009 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this book in any other binding or cover and you must impose the same condition on any acquirer British Library Cataloguing in Publication Data Data available Library of Congress Cataloging in Publication Data Data available Typeset by SPI Publisher Services, Pondicherry, India Printed in Great Britain on acid-free paper by Biddles Ltd., King’s Lynn, Norfolk _________________ ISBN 978–0–19–955326–6 1 3 5 7 9 10 8 6 4 2

Contents List of figures List of tables Notes on contributors List of abbreviations 1 Merge and features: a minimalist introduction José M. Brucart, Anna Gavarró, and Jaume Solà

vii viii ix xii 1

Part I Formal features 2 Probing phrases, pronouns, and binding Fredrik Heinat

25

3 Wh-agreement and bounded unbounded movement Patricia Schneider-Zioga

46

4 Universal 20 without the LCA Klaus Abels and Ad Neeleman

60

5 What it means (not) to know (number) agreement Carson T. Schütze

80

6 Number agreement in the acquisition of English and Xhosa Jill de Villiers and Sandile Gxilishe

104

7 Variable vs. consistent input: comprehension of plural morphology and verbal agreement in children Karen Miller and Cristina Schmitt

123

8 Grammatical features in the comprehension of Italian relative clauses by children Fabrizio Arosio, Flavia Adani, and Maria Teresa Guasti

138

Part II Interpretable features 9 When movement fails to reconstruct Nicolas Guilliot and Nouman Malkawi

159

vi

Contents

10 If non-simultaneous spell-out exists, this is what it can explain Franc Marušiˇc 11 Valuing V features and N features: What adjuncts tell us about case, agreement, and syntax in general Joseph Emonds

175

194

12 The diversity of dative experiencers György Rákosi

215

13 Homogeneity and flexibility in temporal modification Aniko Csirmaz

235

14 The syntactically well-behaved comparative correlative Heather Lee Taylor

254

15 Some silent first person plurals Richard S. Kayne

276

16 From Greek to Germanic: Poly-(∗ in)-definiteness and weak/strong adjectival inflection Thomas Leu

293

17 Acquisition of plurality in a language without plurality Alan Munn, Xiaofei Zhang, and Cristina Schmitt

310

References Language Index Subject Index

329 359 361

List of figures 6.1 Sample stimulus for the recorded sentence: /therabbitsnifftheflowers/

107

6.2 Data on plural and singular subject agreement from two- to three-year-old Xhosa speakers

115

6.3 Tree diagram of derivation of subject agreement in Xhosa

119

7.1 Experimental paradigm

127

7.2 Experiment 1: Sample target trial

130

7.3 Experiment 1: Percentage of plural responses

132

7.4 Experiment 2: Sample target trial

134

7.5 Experiment 2: Percentage of plural responses

136

8.1 Overall results from the picture selection task

149

17.1 Sample picture from Experiment 1

318

17.2 Sample picture from Experiment 2

323

List of tables 5.1 Age range and number of recordings for each Swahili child

88

5.2 Proportions of all indicative clause types for each child and for the adults in a subset of these files

91

5.3 Swahili children’s clause types as a function of age: Hawa

91

5.4 Swahili children’s clause types as a function of age: Mustafa

92

5.5 Swahili children’s clause types as a function of age: Fauzia

92

5.6 Swahili children’s clause types as a function of age: Hassan

93

5.7 Total number of object agreement markers produced

93

5.8 Distribution of MIT twins’ subject pronouns: twin pair I

96

5.9 Distribution of MIT twins’ subject pronouns: twin pair II

97

5.10 Distribution of MIT twins’ subject pronouns: twin pair III

97

5.11 Distribution of MIT twins’ subject pronouns: twin pair IV

98

5.12 Distribution of agreeing versus default verb forms as a function of subject phi-features for three French children

100

6.1 Number of utterances and number of samples ( ) by age band

114

6.2 Pilot studies of subject number agreement comprehension in Xhosa

121

9.1 Determiners and pronouns in French

169

17.1 Proportion of No responses

319

17.2 Proportion of No responses Experiment 1b (English)

321

17.3 Proportion of generic responses: discourse order

325

17.4 Proportion of generic responses: canonicity

325

Notes on contributors Klaus Abels received his PhD at the University of Connecticut in 2003. He has since held positions at the Universities of Leipzig and Tromsø and is currently lecturer in linguistics at University College London. He is interested in constraints on syntactic movement operations. Flavia Adani is a graduate student at the University of Milano-Bicocca and she works in sentence comprehension in typically-developing children and children with language disorders. As an undergraduate, she studied at the University of Siena and at the University of Reading. Fabrizio Arosio is a research assistant at the University of Milano-Bicocca where he teaches in the Faculty of Psychology. He has worked in theoretical linguistics on the semantics of tense, aspect and temporal adverbials and on the processing of verbal agreement morphology in child language. Aniko Csirmaz obtained a PhD degree at the Massachusetts Institute of Technology in 2005. Since then, she has been the recipient of an Andrew W. Mellon Postdoctoral Fellowship (at Carleton College), and is currently an assistant professor at the University of Utah. Joseph Emonds has published four books on syntactic and morphological analysis: Transformational Approach to English Syntax (1976), Unified Theory of Syntactic Categories (1985), Lexicon and Grammar: the English Syntacticon (2000), and Discovering Syntax (2007). He is American but moved to England in 1992. He has also taught in France, Holland, Japan, Austria, and Spain. Maria Teresa Guasti is Professor at the Department of Psychology, Università di Milano-Bicocca. She held positions at the University of Siena, at the Department of Cognitive Science, San Raffaele Hospital in Milan, and at the University of Geneva. She is author of one textbook on language acquisition and of several articles on theoretical linguistics, language acquisition, and language impairment. Nicolas Guilliot defended his PhD thesis, Reconstruction at the Syntax-Semantics Interface, in 2006 at the University of Nantes, and is currently Assistant Professor at the University of Toronto (2007–09). Sandile Gxilishe is an Associate Professor at the University of Cape Town. His research is on child language development, second language acquisition and language in education. He has published widely on these aspects and has also published educational material in Xhosa, an indigenous language of South Africa.

x

Notes on contributors

Fredrik Heinat received his PhD in 2006 from Lund University. The title of his thesis is “Probes, pronouns and binding in the minimalist program”. He currently holds a post-doctoral post at the University of Gothenburg, where he is involved in a project investigating the syntax and semantics of Germanic, and particularly Scandinavian, light verbs. The approach is generative in broad terms. Richard S. Kayne is Professor of Linguistics at New York University. He has written French Syntax (1975), Connectedness and Binary Branching (1984), The Antisymmetry of Syntax (1994), Parameters and Universals (2000), and Movement and Silence (2005), and is editor of the book series Oxford Studies in Comparative Syntax. Thomas Leu is a graduate student in linguistics at New York University. His work includes a novel analysis of the Germanic “what for” construction, a bi-nominal analysis of modified indefinite pronouns like “something strange”, and an analysis of the internal syntax of demonstrative determiners, closely related to the present contribution. Nouman Malkawi is a PhD student at the University of Nantes and his thesis on resumption in Jordanian Arabic should be defended in 2008. Franc (Lanko) Marušiˇc was awarded his PhD from Stony Brook University in 2005, when he joined the University of Nova Gorica as an assistant professor. His main areas of interest are Slovenian syntax, comparative Slavic syntax, and syntactic theory. He has published papers in various journals (including Linguistic Inquiry and Natural Language and Linguistic Theory) and co-edited the volume Studies in Formal Slavic Linguistics. Karen Miller is Assistant Professor in Spanish at Calvin College. She obtained her PhD from Michigan State University in 2007 and the title of her dissertation is “Variable Input and the Acquisition of Plurality in Two Varieties of Spanish”. She is director of the Calvin College Language Studies Lab. Her research focuses mainly on first language acquisition. Alan Munn is Associate Professor of Linguistics at Michigan State University. He received his PhD in 1993 from the University of Maryland, College Park. He has taught at the University of North Carolina, the University of Missouri, and Harvard University. He co-directs the Michigan State University Language Acquisition Lab. Ad Neeleman is Professor of Linguistics at University College London. He obtained his PhD from Utrecht University in 1994 (cum laude). He is co-author of two monographs—Flexible Syntax (1998, with Fred Weerman) and Beyond Morphology (2004, with Peter Ackema)—and has published articles on syntax, morphology, PF, and information structure. György Rákosi is a lecturer in linguistics at the University of Debrecen in Hungary. He defended his PhD thesis Dative experiencer predicates in Hungarian in 2006 at the Utrecht Institute of Linguistics. He has an interest in argument structure and related

Notes on contributors

xi

phenomena in general, and he has published articles on experiencer, reflexive, and reciprocal predicates, and on anaphoric dependencies. Cristina Schmitt is Associate Professor in Linguistics at Michigan State University. She works mainly on the syntax-semantics of noun phrases, aspect, and first language acquisition. Patricia Schneider-Zioga is a lecturer in the Department of English, Comparative Literature, and Linguistics at California State University, Fullerton. Recent works include “Anti-Agreement, Anti-Locality and Minimality: the Syntax of Dislocated Subjects” (2007) and “Dyslexia: the temporal-spatial disordering hypothesis and its metrical reflex” (2007). Carson T. Schütze is Associate Professor of Linguistics at the University of California, Los Angeles. He is author of The Empirical Base of Linguistics (1996) and entries on methodology in three encyclopedias. He has published articles on syntax in Linguistic Inquiry, Lingua, Syntax, and The Linguistic Review, on language acquisition in Journal of Child Language and Language Acquisition, and on psycholinguistics in Journal of Memory and Language and Journal of Psycholinguistic Research. Heather Lee Taylor is currently finishing her PhD studies at University of Maryland, College Park. Her research concentrations are in syntactic theory and second language acquisition. Within these subdisciplines she has investigated comparative and degree syntax and semantics, wh-in-situ, A -movement and island effects, age effects in learning, and implicit learning. Jill de Villiers is a Professor at Smith College in Psychology and Philosophy. She received a BSc degree from Reading University and a PhD from Harvard University, both in psychology. The co-author of two books on language development, she has spent over thirty years doing research and publishing on topics around the acquisition of syntax, mostly on English, and is a co-author of the DELV language assessment test. Xiaofei Zhang is a PhD student in the linguistics program at Michigan State University.

Abbreviations A

Adjective

AAE

African American English

Acc

Accusative

AdvP

Adverbial Phrase

AFH

Active Filler Hypothesis

Agr

Agreement

AgrA

Adjectival Agreement

AgrO

Object Agreement

AMP

Accord Maximization Principle

AP

Adjectival Phrase

Appl

Applicative

AspP

Aspect Phrase

ATOM

Agreement/Tense Omission Model

C

Complementizer

Caus

Causative

CC

Comparative Correlative

CED

Condition on Extraction Domains

CFC

Canonical Form Constraint

C-I

Conceptual-Intentional

Cl

Clitic

CP

Complementizer Phrase

CSC

Coordinate Structure Constraint

D

Determiner

Dat

Dative

Def

Definite

Deg

Degree

Dem

Demonstrative

DemP

Demonstrative Phrase

DO

Direct Object

DP

Determiner Phrase

Abbreviations DS

Determiner Spreading

Du

Dual

ECM

Exceptional Case Marking

ECP

Empty Category Principle

EPP

Extended Projection Principle

Exst

Existential

F

Feature

Fem

Feminine

FL

Faculty of Language

FP

Functional Phrase

Fut

Future

Gen

Genitive

H

Head

HPSG

Head-Driven Phrase Structure Grammar

Ind

Indicative

Inf

Infinitive

Ins

Instrumental

IO

Indirect Object

IP

Inflectional Phrase

JA

Jordanian Arabic

KP

Case Phrase

LAD

Language Acquisition Device

LCA

Linear Correspondence Axiom

LF

Logical Form

Loc

Locative

L1

First Language

MAE

Mainstream American English

Masc

Masculine

MCP

Minimal Chain Principle

MDPH

Mismatch Detection Point Hypothesis

MDSH

Mismatch Detection Stage Hypothesis

MP

Minimalist Program

N

Noun

Neg

Negation

xiii

xiv

Abbreviations

Neut

Neuter

Nom

Nominative

NP

Noun Phrase

Num

Number

NumP

Number Phrase

OA

Object Agreement

Obj

Object

OM

Object Marker

P

Preposition

P&P

Principles and Parameters (Model)

Part

Participle

Pauc

Paucal

Perf

Perfect

PF

Phonetic Form

PIC

Phase Impenetrability Condition

PL

Plural

PP

Prepositional Phrase

PredP

Predicate Phrase

Prs

Present

Pst

Past

Q

Quantifier

QNP

Quantified Nominal Phrase

QP

Quantifier Phrase

QR

Quantifier Raising

RC

Relative Clause

Refl

Reflexive

S

Sentence

SA

Subject Agreement

SAE

Standard American English

Sbjv

Subjunctive

Sg

Singular

SLQZ

San Lucas Quiaviní Zapotec

SM

Subject Marker

S-M

Sensorimotor

Abbreviations Spec

Specifier

Suf

Suffix

TP

Tense Phrase

UG

Universal Grammar

v

small verb

vP

small verb Phrase

V

Verb

VP

Verb Phrase

V2

Verb Second

Wh-agr

Wh Agreement

XP

X(Variable) Phrase

1Pl

First person plural

2Pl

Second person plural

3Pl

Third person plural

1Sg

First person singular

2Sg

Second person singular

3Sg

Third person singular

xv


1 Merge and features: a minimalist introduction JOSÉ M. BRUCART, ANNA GAVARRÓ, AND JAUME SOLÀ

This book is about features and merge and, more specifically, about the intricate ways they interact in generating expressions in natural languages. This introductory chapter is divided into two parts. In the first we offer a brief sketch of the tenets of the Minimalist Program (MP), which constitutes the current mainstream version of generative grammar (Chomsky, 1995, 2000, 2001, 2004, 2005, 2008), and in the second we discuss how the contributions included in the present volume address some fundamental questions raised by it. The Minimalist Program can be seen as a natural development of the Principles and Parameters framework established in the 1980s. It inherits most of its basic assumptions in trying to characterize the faculty of language (FL) as a specific component of the mind/brain. It differs from earlier versions of Principles and Parameters in setting as a main programmatic thesis what was already a recurrent theoretical observation reached from several viewpoints: that the faculty of language is simple, elegant, and nonredundant. As has often been clarified, Minimalism does not simply consist in adhering to the generally accepted methodological principle that theories should be formulated in the simplest way compatible with the available evidence, thus minimizing the ontology and complexity of the postulated basic (axiomatic) principles. Minimalism is rather the programmatic claim that an object in the world, the faculty of language at its core, is extremely simple, and that its apparent complexity is to be derived from its interaction with independent constraints. These constraints are plausibly related to biological complexity: general principles regulating possible complex (biological) structures,

2

Introduction

including the faculty of language, and specific constraints due to the interaction of the faculty of language with other subsystems of the mind/brain. 1

1.1 A brief overview of the minimalist model Two classical demands on generative theories of human language have been descriptive adequacy and explanatory adequacy. The former requires that theories of the faculty of language be compatible with the data. The latter poses the further requirement that theories of the faculty of language be compatible with the evidence provided by the process of acquisition. The metric for evaluating these two levels has been internal to the faculty of language. The Minimalist Program introduces an additional, external level of adequacy that goes beyond explanatory adequacy (Chomsky, 2004): theories of the faculty of language must be compatible with conditions that are independent of the language faculty in a strict sense. These include, on the one hand, general principles of processing which limit the class of possible structures and, on the other, conditions imposed by the interaction of the language faculty with the cognitive modules of thought and sound. Since thought and sound are the systems connected by the faculty of language, 2 it is plausible to think that their specific properties and demands must have some influence on the way the faculty of language works: the faculty of language must be able to interact with them. More specifically, it must provide interface representations which are legible to them: the ConceptualIntentional interface and the Sensorimotor interface. As a consequence, some principles that had previously been considered intrinsic to the faculty of language have been reformulated as conditions imposed by the ConceptualIntentional interface (the Theta Criterion, the Projection Principle) or by the Sensorimotor interface (word order, morphological well-formedness). These are “bare output conditions” in the sense that they must be satisfied not by intrinsic requirements of the faculty of language as an autonomous component of the human mind/brain but by virtue of the relation that it establishes with the other cognitive modules to which it is connected. The other external (non intrinsic) factor conditioning the faculty of language includes “principles of structural architecture and developmental constraints that enter into canalization, organic form, and action over a wide 1

See Martin and Uriagereka (2000) and Boeckx (2007) for further clarification of this matter. We are consciously using the terms “sound” and “thought” in a broad sense. Indeed, the notion of “sound” is a simplification, given the fact that the class of natural languages also includes sign languages. To reflect this, we will use the term Sensorimotor (SM) interface, instead of ArticulatoryPerceptual ( A-P ) interface (Chomsky, 1995: 131). On the other hand, “thought” must be strictly conceived here as the module that interprets the conceptual/intentional meaning of linguistic expressions. 2

Merging features: a minimalist introduction

3

range, including principles of efficient computation, which would be expected to be of particular significance for computational systems such as language” (Chomsky, 2005: 6). In accordance with this line of thought, Chomsky (2000: 96) proposes, as a working hypothesis, the strong minimalist thesis: “Language is an optimal solution to legibility conditions”. It should be optimal due to the principles of structural architecture (which lead to a perfect design), and it should be constrained only by interface requirements. As the null hypothesis, Minimalism assumes that an optimal design of language entails that there should be nothing in the FL that is not required by the need to connect the two interfaces. What should there be, then? Minimally, the faculty of language must provide a syntactic procedure that forms complex objects from combining simpler ones: Merge. It also must provide a set of objects to be combined: the lexicon. Merge and the lexicon instantiate the two uncontroversially indispensable aspects of any theory of the faculty of language: there must be units to combine and there must be a combinatory mechanism. The simplest option is that Merge proceeds in recursive steps: at each step, it combines two objects forming a new one that is structurally more complex. If the objects are independent of each other, it is called External Merge. If one of the objects is a constituent of the other, we call it Internal Merge, an operation that subsumes the movement component of earlier formulations of generative grammar. The recursive nature of Merge is therefore responsible for hierarchical structure, a fundamental characteristic of syntactic objects. 3 Merge is a fundamental component of the computational system of human language. Under minimalist assumptions, the faculty of language in a strict sense consists of the lexicon and the computational system (CHL ). Only general principles of optimal design apply to CHL as a mechanism. There are no intrinsic conditions on the objects it creates, except to the extent that these objects become interface representations; hence, there must be no other levels of representation apart from the ones corresponding to the interfaces. The basic architecture that emerges for the faculty of language is: 4 3 In fact, for Merge to work properly, it is necessary that a previous operation of Select is performed. The function of Select with respect to External Merge consists of taking a lexical item before merging it with a previously formed structural object. With respect to Internal Merge, Select acts on a unit already merged before moving it to a new position in the structure. The existence of the operations Select and Merge in a syntactic system is mandatory. Therefore, both can be justified on the grounds of virtual conceptual necessity. For a discussion of the more basic nature of Merge with respect to the compositional operation of Unify advocated in other linguistic models, see the reply of Boeckx and Piattelli-Palmarini (2007: 410) to Jackendoff (2007a: 362). 4 For a different view on the relation between C HL and the systems of thought, see Hinzen (2006).

4

Introduction

C-I interface representation

(1)

lexicon

CHL SM interface representation General principles Interface conditions

As (1) shows, there is a point where the derivation splits into two separate paths: one proceeds to the Conceptual-Intentional interface and the other leads to the Sensorimotor interface. It is important to note that, contrary to the view undertaken in previous generativist models, the dominant idea in the MP is that the derivational diagram in (1) does not correspond to a whole sentence but to certain designated subparts of it known as “phases”, which are cyclically processed bottom-up. Therefore, the derivation of a sentence includes several points of transfer like the one represented in (1). 5 Let us concentrate on the derivation that generates the ConceptualIntentional interface. This part of the computation—which constitutes what has been called narrow syntax—contains two segments, delimited by the point at which the derivation splits into two branches. Before this divide, computational operations feed both interfaces. On the other hand, only the Conceptual-Intentional representation is affected by the operations performed after it. The point in the derivation at which syntactic information is sent to the Sensorimotor interface is known as Spell-Out. The lexicon should be the simplest expression of possible meaning–sound associations that can be combined by Merge in a given language. A reasonable assumption is that Merge is invariant across languages. As a consequence, any source of linguistic variation should be attributed to the lexicon, the component of language that must be specifically learned (Borer, 1984; Baker, 1996, 2001). It is assumed that for a specific linguistic expression to be computed by the computational system of the faculty of language, it is not the lexicon (as a general repository of irreducible meaning-sound associations) that directly provides the elements to be combined, but a specific subset of items taken from it: the numeration, an array of lexical items obtained by accessing the lexicon once before the computation begins. 6 Hence, the lexicon is part of the faculty 5 The standard account of phases in the MP takes for granted that their size is the same for each interface, but see Marušiˇc (this volume) for an interesting proposal in the opposite direction. 6 The effect of numeration is to drastically reduce the number of possible derivations in competition by avoiding successive accesses to the lexicon. Moreover, this device assures that the


5

of language only to the extent that it can provide appropriate numerations to be computed by the computational system. Example (3) reflects the process of forming the DP a lesson on geometry from the corresponding numeration (2): 7 (2) numeration: {a1 , geometry1 , lesson1 , on1 } (3) derivation: a. [geometry] Select b. [ on [ geometry ]] Select and merge c. [ lesson [ on [ geometry ]]] Select and merge d. [ a [ lesson [ on [ geometry ]]]] Select and merge What kinds of objects are appropriate for CHL ? They should be objects that contain features that are either recognizable by CHL , or interpretable by the Conceptual-Intentional or Sensorimotor interfaces. Assuming that these objects consist of bundles of features, we conclude that features in a lexical item must be either computable or interpretable at the interfaces. The former will feed computational mechanisms, the latter will be simply transferred to the interfaces. Let us call the former formal features and the latter semantic features. 8 Formal features are recognized by CHL and interpreted as instructions that trigger computational operations in the derivation. Putting aside phonological features, formal features (such as Tense, phi-features, or Case features) can be interpretable (legible by the interface) or uninterpretable. This dichotomy is based on the observation that some formal features (such as Case) seem to have no interpretive content. The main motivation for the dichotomy, though, is that in many languages some formal features appear in lexical items where they are not interpreted, i.e. phi-features on the verb, number and gender features on adjectives in some languages, etc. (4)

a. She3sg loves3sg jazz b. losmasc,pl periódicosmasc,pl deportivosmasc,pl sport journals the ‘the sport journals’

representations at both interfaces are based on the same lexical choices (Chomsky, 1995: 225). For different views on this concept, see Zwart (1997), Frampton and Gutmann (1999) and Hornstein (2001). 7

The integer subscripted to each lexical item represents the number of tokens of the correspondent unit to be used in the derivation. A derivation can contain more than one token of a given unit, as is the case of the definite article in The student passed the course. The possibility of having more than one token for some lexical unit is what forces us to conceive the lexical array as a numeration. 8 An interesting question is whether features are universal or, on the contrary, languages admit some variation with respect to them. Chomsky (2000) seems to endorse the second possibility, whereas Sigurðsson (2003) or Kayne (2003b) argue for the first approach: universal features are present whether they are phonologically visible or not.

6

Introduction

This corresponds to the traditional agreement relation, which is characterized in minimalist terms as an asymmetric relation between a source and a target of agreement (respectively, the interpretable and non-interpretable manifestations of the same feature). The presence of uninterpretable features seems to deviate from the optimal solution that Minimalism advocates, given that they are unnecessary for interpretation. Theoretically, the very existence of these redundant features could be viewed as an imperfection of the faculty of language. One possible account of this apparent imperfection is that these features are formal triggers that feed computational operations necessary for certain aspects of interpretation at the Conceptual-Intentional interface. 9 Once these operations have been performed, CHL should be able to delete them before reaching the ConceptualIntentional interface in order to satisfy the principle of Full Interpretation, which precludes the presence of any material that is not interpretable at the interfaces (Chomsky, 1986, 1995). There are two computational operations that are induced by formal features: Internal Merge and Agree. As noted above, Internal Merge corresponds to the concept of Move in previous accounts, and is an operation that comes for free once Merge is assumed. 10 Agree, on the other hand, is an operation that establishes a relation between features of the same type in different structural positions. As a result of the operation, the corresponding structural positions are related to each other. Agree can be seen as an asymmetrical relation between an uninterpretable feature in a head position (the probe) and a feature of the same type in a previously merged position (the goal), which is consequently in the c-command domain of the probe. This relation is blindly triggered by the presence of uninterpretable features, and its effect is to eliminate them from the representation to be delivered to the Conceptual-Intentional interface, while keeping them available for the Sensorimotor interface. This dual function can be expressed as a process of match, valuation, and deletion. 11 9 See for instance the analysis of negative complementizers in Hebrew presented in Landau (2002). 10 More specifically, Internal Merge includes two subparts: Copy and Merge. Copy could be justified by the fact that many lexical units in natural languages can simultaneously perform more than one function, as is the case of interrogative words, which, besides their argumental or adjunct nature, include a modality operator. As a consequence, these units can be conceived of as syntactically discontinuous, affecting at least two structural positions at the same time: one that corresponds to their operator status and the other that corresponds to their argumental or adjunct status. In fact, Chomsky (2003: 307) considers that Copy is not a new relation in addition to Merge: “Copy is simply ‘internal Merge’ ”. 11 The term “checking” has been used, and is still used on occasion, to refer to match plus valuation (or some previous alternative to valuation).


7

Uninterpretable features are, by assumption, unvalued when they enter the derivation. In order for their morphological content to be expressed, they must be assigned a value through an Agree relation and, once valued, they must be deleted: Interpretability of features is determined in the lexicon [. . . ]. The natural principle is that the uninterpretable features, and only these, enter the derivation without values, and are distinguished from interpretable features by virtue of this property. Their values are determined by Agree, at which point the features must be deleted from the narrow syntax [. . . ] but left available for the phonology (since they may have phonetic effects). (Chomsky, 2001: 5)

In conclusion, uninterpretable features are not interpretable themselves but they feed necessary computations for CHL to be an optimal solution to interface conditions. Once valued through an Agree operation, they must be deleted in narrow syntax immediately after Spell-Out. Therefore, they are absent when the derivation is transferred to the Conceptual-Intentional interface. Otherwise, the Principle of Full Interpretation would be violated and consequently the derivation would crash, that is, it would not converge at the Conceptual-Intentional interface. Let us now consider Internal Merge. Until recently (Chomsky, 2008), Internal Merge was conceived of as the result of the interaction between Agree and Merge (and possibly Pied-piping), as exemplified in (5). Simplifying details, in (5a) the uninterpretable phi-features of T probe the phi-features of the DP the treasure and establish an Agree relation with it; subsequently, in (5b), the DP is copied and merged with the whole syntactic object (“it moves to Spec, TP”). 12 (5)

[ [t was3sg ] [vp found the treasure3sg ] ] a. b. [tp the treasure3sg [ [t was3sg ] [vp found the treasure —3sg ] ] ]

However, many languages show that movement and agreement have to be kept separate in cases such as these. In fact, even English provides evidence that Agree functions independently from Move in expletive there sentences: 13 12 The phrase struck out stands for the position from which Move has applied: the one corresponding to the internal argument of the participle. 13 As a reviewer notes, evidence for Agree without Move in English is weak, due to the strong restrictions on postverbal subjects in this language (as compared to postverbal subjects in Romance languages—Burzio, 1986—or quirky subjects in Icelandic—Sigurðson, 1991—, that provide more robust evidence). We provide English examples for convenience, ignoring important aspects such as the contrast between There were found two big jewels and There were two big jewels found. See Caponigro and Schütze (2003) for the view that only the latter involves an associate DP in A-position (we thank C. Schütze for pointing this out).

8 (6)

Introduction a. There was3sg found a very great treasure3sg b. There were3pl found two big jewels3pl

In the above examples the probe matches the features of the goal at a distance and the uninterpretable features of the passive auxiliary are subsequently valued. Once an uninterpretable feature has been valued, it is no longer active and cannot perform a new search for a goal. The conclusion that Internal Merge has to be separated from Agree poses the question as to what the driving force is behind movement in natural languages. Chomsky proposes that it is a consequence of a condition imposed by the Conceptual-Intentional interface: C-I incorporates a dual semantics, with generalized argument structure as one component, the other one being discourse-related and scopal properties. Language seeks to satisfy the duality in the optimal way [. . . ], E(xternal) M(erge) serving one function and I(nternal) M(erge) the other. (Chomsky, 2005: 8)

The interaction between movement and discourse properties seems straightforward in contrasts like the ones in (7), where the postverbal position filled by a spy in (7a) can only host indefinite (and non-specific) arguments not previously present in the discourse background: (7)

a. There is a spy in this room. b. The spy is in this room.

On the other hand, the contrast between active and passive sentences (Three students read a book vs. A book was read by three students) is further evidence in the same direction: the internal argument of the transitive verb becomes the topic of the sentence, as opposed to the rhematic nature of the object of the active construction. 14 All these arguments point to the idea that the function of Internal Merge and Agree in CHL is not the same. The latter is an operation internal to the computational system, which allows for the elimination of uninterpretable features by assigning a value to them. The former, on the contrary, is interfacerelated, and its existence is connected to the necessity of coding discourseoriented and scopal relations. However, the fact that both operations are related to formal features and show a common probe-goal pattern in their functioning leads some researchers to conceive them as two different procedures to obtain essentially the same result: the valuation of uninterpretable features as a necessary condition for their deletion before the derivation 14 Moreover, for some speakers there are also interpretive differences affecting quantifiers: the passive subject tends to be associated with wide scope over the agent.


9

reaches the Conceptual-Intentional interface. Thus, for instance, Bošković (2007a) presents an analysis where the difference between them is tied to the fact that the uninterpretable feature can be placed on the probe (giving rise to Agree) or on the goal (causing Internal Merge). Be it as it may, there is evidence suggesting that the constraints that affect both operations diverge to some degree. Take, for instance, the asymmetry in (8), discussed in Cheng and Rooryck (2000) and Bošković (2000): (8)

a. Jean mange quoi ? John eats what ‘What does John eat?’ b. ∗ Jean ne mange pas quoi? John NEG eats not what ‘What does John not eat?’ c. Qu’est-ce que Jean ne mange pas? What is-that that John NEG eats not ‘What does John not eat?’

In colloquial registers, French allows wh-in-situ questions, as in (8a), together with the corresponding wh-extraction variant, as in (8c). The ungrammaticality of (8b) has been related to the presence of negation, which blocks the Agree relation between the head C and the in-situ interrogative. 15 On the other hand, (8c) shows that Internal Merge is feasible in the same context. We will not review the different proposals made to account for these contrasts, but at first sight it seems that whatever causes (8b) to crash—presumably, the intervention effect of the negation alluded to before—does not prevent the convergence of (8c), where Internal Merge of the direct object to the specifier of CP is available, despite the presence of the negation. A possible way to tackle the asymmetry would be to suppose that the feature the probe cannot value in (8b) is different from the one that feeds Internal Merge in (8c). 16 Once Internal Merge takes place, the valuation of the wh-feature in C is possible because of the local relation between C and the interrogative in its specifier. 17 Selective intervention effects, like the one just discussed, show that CHL is sensitive to locality constraints: a probe can agree with a goal in its c-command domain only if there is no intervener. But this mechanism is not sufficient to cope with other locality effects, such as those that preclude extraction from Rizzi’s relativized minimality (Rizzi, 1990) offers an account of these effects. Chomsky (2005) proposes that the element that gives rise to Internal Merge is an uninterpretable edge feature placed in C. 17 In order for this argument to work, it is necessary to assume a notion of selective intervention effects like the one proposed by Rizzi (1990) and Rizzi (2004). 15 16

10

Introduction

syntactic islands, whose study dates back to Chomsky (1964) and Ross (1967). In order to account for these phenomena, Minimalism resorts to a cyclic organization of computational processes: phase theory. Phases are conceived of as lexical subarrays of the numeration that, once merged, project a span of structure that exhibits semantic and phonological autonomy (Chomsky, 2004: 124). Once all the computational processes corresponding to a phase have been performed, it becomes an inert unit whose components are no longer accessible to any operation triggered by subsequent (higher) phases. Therefore, the computation proceeds phase by phase, in a bottom-up fashion. Thus, with respect to diagram (1) above, as Merge proceeds assembling structure in a bottom-up fashion, Spell-Out would take place at several points in the derivation. 18 Phase theory has been the object of considerable discussion within Minimalism. The two basic questions it raises are: what categories are phases, and why? Chomsky (2000) proposes that phases are minimal propositional entities that have interpretive and phonological autonomy. Following this criterion, he proposes CP and vP as the two projections that qualify as phases. More recently, Chomsky (2004, 2005) has proposed that the fundamental criterion in establishing what constitutes a phase is internal to CHL : as CP and vP are the domains in which agreement relations are established, phases are the minimal domains in which uninterpretable features are valued. Once an uninterpretable feature has been valued, its deletion must take place as soon as possible. Therefore, phases can be viewed as the cyclic domain used by CHL to satisfy this requirement. In summary, the task set out by the Minimalist Program is to show that an explanation of the language faculty can be successfully achieved by resorting to the three following factors: (a) necessary mechanisms of CHL : numerations of lexical items (arrays of features), (Internal/External) Merge, Agree, and deletion; (b) general principles minimizing search and computation (minimal search, phases); (c) interface conditions: Full Interpretation, morphology, word order, etc. We will now assess how the proposals and the evidence provided by the contributions to this volume shed light on the Minimalist Program. 18 Root phases are transferred in their entirety, whereas in the case of non-root phases only the domain of the complement is spelled-out. Thus, only the edge of a phase—i.e. the specifier(s) and the head—remains active to be accessed by operations corresponding to the next phase. Its activity, however, is limited: edge categories cannot trigger further computational operations, although they can be goals for superordinate probes.


11

1.2 Formal features 1.2.1 The role of formal features As we saw above, the uninterpretable status of formal features is the trigger of non-local relations between syntactic positions by means of Agree. According to Chomsky (2001: 5), once an uninterpretable (hence unvalued) feature is merged, it must establish, as soon as possible, an Agree relation with a feature in its search domain in order to be valued and deleted (for the ConceptualIntentional interface). In this view, unvalued features are the syntactic correlate of uninterpretable morphology (Case and agreement). The possibility has been explored, however, that the valued/unvalued property of features has an independent status from the (un)interpretable property. Pesetsky and Torrego (2007) propose that Agree is a valuation mechanism, possibly affecting both interpretable and uninterpretable features, while deletion is an independent mechanism related to interpretation at the ConceptualIntentional interface. On this view, unvalued features correspond to lexical underspecification: a given lexical item contains an unvalued feature, and the Agree mechanism must value it so that it can be interpreted. If it is interpreted only at the Sensorimotor interface, it must be deleted in narrow syntax. In this volume, Heinat develops a proposal, based on Pesetsky and Torrego’s approach, to account for binding of reflexives, where these elements are treated as DPs with unvalued phi-features which have to be valued by Agree. Crucially, phi-features of the probe (the antecedent) and of the goal (the reflexive) are both interpretable, only the latter being unvalued. Indeed, reflexive pronouns are a case of agreement in which the features of both the source and target seem to be interpretable. Most remarkably, Heinat includes phrases as possible probes (contra Chomsky’s (2004: 113; 2008) contention that all Agree relations—including reflexive binding—are probed by a head): he argues that any externally merged head or phrase label may be a probe. It might be that behind the controversial point as to whether only heads or also phrases can probe lies the issue as to whether a unitary account in terms of Agree can be provided for two phenomena that have traditionally been kept apart: subject-verb agreement and reflexive binding, 19 both traditionally belonging to the A-binding domain, but only the former being uncontroversially related to movement (Agree and Move in minimalist terms). 19 And its possible extension to pronominal binding. See Lasnik (1999), Hornstein (2001, ch. 5), Kayne (2002), and Zwart (2002), for an analysis of binding in terms of movement.

12

Introduction

If we shift to the A -binding domain, A -movement seems to be the central phenomenon whereby CHL achieves unbounded dependencies. Nevertheless, there is also an important phenomenon that needs to be included in this picture, namely A resumptive structures, not only because of its parallelism to movement but also because of their frequent complementary distribution. Resumption has often been seen as a last resort strategy to rescue movement out of islands, but it also appears to be a parametric option (an alternative to movement) for encoding scope relations in a given language or construction. For instance, English topicalization (John, I didn’t see), which seems to be a case of movement, is the functional counterpart to Romance Clitic Left Dislocation, which involves resumption (Gianni, non l’ho visto). Since movement and resumption usually are not in free variation, one should ask what forces CHL to choose between these options. If we assume that movement is the unmarked option (if locality conditions are met), then in order for resumption to take place some special configuration would be needed. Schneider-Zioga argues that in Kinande, a Bantu language, non local wh-displacement is not achieved by movement but by a resumptive strategy, in view of the fact that it does not feed reconstruction. According to this author, the impossibility of successive movement is due to the special properties of embedded clauses in this language, which lack escape hatches (edge features on the head of the C phase, in Chomsky’s 2004 terms). Specifically, the idea is that Kinande is a V2 language where the V2 property appears in both main and embedded clauses, and consists of a Spechead agreement in the left periphery, which blocks successive movement. In this view, then, Kinande’s resumptive strategy is a last resort alternative to movement. Importantly, resumptive pronouns in Kinande are not standard pronouns in A-position but rather agreement heads in the left periphery (which license a null pronoun in their specifier). It would be interesting to explore whether the presence of obligatory resumption to obtain A -displacement can be explained in terms of some morphological agreement pattern that forces the occurrence of intervening pronouns that block movement. Most work on the role of formal features in Agree and Move has centered on A and A dependencies in the clausal domain. Less well understood is their role in determining other kinds of movement, such as DP internal movement. The antisymmetry hypothesis (Kayne 1994) has led to the postulation of a great number of movements for which no obvious probe (uninterpretable feature) can be invoked. The only motivation for resorting to movement, as in Cinque’s (1996) account of possible word orders within the DP, is to prove


13

that the antisymmetry hypothesis makes the right predictions on the basis of what is possible and impossible movement, even if possible movement is not theoretically motivated by formal features. By now, it seems clear that the Linear Correspondence Axiom is a feasible proposal to account for complex word order alternations (see Cinque 1996, 2005a), but it is harder to establish which (uninterpretable) features should be responsible for triggering the movements one has to postulate. The question is then whether a movement account of word order variation is warranted as the only option. Abels and Neeleman propose to reduce the range of apparently unmotivated movements that are necessary to derive attested orders, by allowing variation in the base-generated word order of adjacent nodes, and postulating that only movement operations obey a right-to-left (antisymmetric) constraint, reminiscent of the Linear Correspondence Axiom. Specifically, they propose an alternative to Cinque (2005a) (where all the attested word orders within a DP are derived from a basic universal structure obeying the Linear Correspondence Axiom). They show that their proposal generates the same set of possible orders as Cinque’s, just with less movements. Abels and Neeleman argue that the set of required movements in their proposal is a proper subset of Cinque’s set of required movements: it excludes very local movement to Spec (Complement of X moves to Spec of X), precisely the kind of movement that derives Cinque’s counterpart to Abels and Neeleman’s base-generated alternations. They argue that this kind of movement is excluded by well-motivated antilocality principles, which leads the authors to conclude that their proposal is empirically superior to an LCA account. 1.2.2 Subject agreement Unvalued features (and uninterpretable features) appear to be key factors for optimally connecting the lexicon with the interfaces: they embody some mismatch between the lexicon and (one of) the interfaces that has to be repaired by Agree and/or delete. It is perhaps in this sense that one may view Schütze’s proposal of uninterpretable features. Schütze’s Accord Maximization Principle, AMP, establishes that there is a requirement to maximize the presence of uninterpretable features in the numeration (Case and agreement): the maximum compatible with a convergent derivation (this is similar in spirit to Chomsky’s 2001 Maximize matching effects). This principle restricts the set of admissible derivations to those stemming from a numeration fulfilling the AMP. Schütze’s view is that Case and agreement errors in child language may have their origin in the child’s inability to satisfy the AMP in numerations, due

14

Introduction

to processing limitations. Crucially, Schütze assumes Distributed Morphology, whereby syntax is fully specified, independently of the morphological richness of the language. This implies that the children’s deficit of feature insertion cannot be due to a morphological deficit (but to a processing deficit in abiding the AMP). This view is in opposition to a morphology-before-syntax approach, where the issue of whether morphological variation might affect uninterpretable feature insertion arises, in accordance with the idea that all variation is in the lexicon, and syntactic computation has to cope with whatever features happen to make up the inserted lexical items. In this case, a child’s lexicon could just happen to be poorer in uninterpretable features, due to poorer morphology. Schütze provides strong evidence against this view, especially from Swahili: children make agreement (and Tense) morphological omissions while they perfectly master this very morphology. This is an excellent argument for the AMP and Distributed Morphology. Yet the general question remains as to whether the AMP plus syntactically innocuous Distributed Morphology can account for variation, a question which is not just about morphological shape. Valuation is a directional mechanism, in that the lexically valued feature is the source for valuing an unvalued feature of the agreement target. For phi-features, it is standardly assumed that they are interpretable (and already valued) in the (subject/object) DP, and uninterpretable (to be valued) in T or v. Two chapters in this volume address the issue of whether there is an asymmetry between number morphology on the subject DP and on the agreeing verb in language acquisition. Specifically, morphology on the subject DP would be expected to be more reliable for retrieving number features, giving rise to fewer errors in either production or comprehension of DP number morphology than in V number morphology. De Villiers and Gxilishe study two- to three-year-old child language production of subject-verb agreement in Xhosa, a rich morphological language where both subjects and verbs are inflected for a rich paradigm of noun classifiers also encoding number. The expected asymmetry is that the morphological expression of phi-features should be more reliable on the DP than on V-agreement, since the latter is contingent on the former for valuation. Their results show that Xhosa children produce few errors (only of omission, none of commission) in subject-verb agreement. They conclude that number agreement is indeed directional, even if the source of agreement is not always spelled out. Miller and Schmitt address a related question in child comprehension of Chilean Spanish. They show that Chilean four- to five-year-old children can


15

recover the number features of subject DPs (whose morphological marking is less reliable due to optional consonant elision) from verbal agreement (whose morphology is steady). This seems to show that morphological robustness and steadiness is decisive in acquiring lexical features (in line with the Variability Delay Hypothesis, Miller 2007), whether they correspond to interpretable or uninterpretable features. The results of Miller and Schmitt contrast with those of Arosio, Adani, and Guasti, who consider how subject-verb agreement in Italian child language is processed in comprehension depending on the position of the subject (pre- or postverbal), in contexts (relative clauses) where the postverbal subject may be misinterpreted as an object (with word orders like: il ragazzo che ha visto il pagliaccio lit. the boy that has seen the clown). They observe that five- to nine-year-old children misinterpret such sentences by disregarding agreement between the verb and the postverbal subject. They conclude that this is the result of a failure in processing chains (due to the Minimal Chain Principle). Here verb morphology is ignored by the child, even though it is quite robust, in contrast to what happens with the Chilean children. Taking both cases into account, child language behavior with respect to subject verb agreement appears to be determined by factors beyond the (un)interpretable status of lexical features. The conclusion seems to be that directionality of valuation need not influence child processing (retrieval of phi-features), a quite expectable situation given that processing may be related to performance rather than to competence. Yet, as shown by Schütze, inherently grammatical phenomena are also at the source of the deviations characteristic of child language.

1.3 Interpretable features 1.3.1 Reconstruction If Agree is the computational device that takes care of uninterpretable (unvalued) features, Internal Merge is, according to Chomsky (2004), an independent device, triggered by an EPP feature. EPP features prompt Internal Merge. 20 As we saw in section 1.1, Internal Merge is a mechanism that makes available the appropriate structures for second-order semantics, which encompasses discourse-oriented and scope-related phenomena. An issue we already introduced above concerning Schneider-Zioga’s chapter is why movement, as the computational basis for this kind of semantics, so 20 An EPP feature in a head position requires Merge of a phrase to the projection of this head (in specifier position), typically Internal Merge. Chomsky (2005) includes EPP features as a subtype of edge-features, which are the necessary triggers for both types of Merge (Internal or External).

16

Introduction

often alternates with resumptive strategies. It has been a traditional assumption that one of the characteristic traits of resumption is that, unlike movement, it does not show reconstruction effects. So, reconstruction is a diagnostic test for distinguishing movement from resumption. In Chomsky’s (1995) terms, only movement leaves a copy, and reconstruction is the interpretation of (part of) the copy. Guilliot and Malkawi challenge this descriptive generalization. These authors provide evidence that in some languages (for example, Jordanian Arabic) the resumptive strategy shows reconstruction effects. They claim that reconstruction is available for both copies of movement and empty categories. Empty categories occur with some resumptive pronouns, which should be analyzed as determiners licensing an elliptic NP. Interestingly, a resumptive pronoun with NP ellipsis is predicted to head a definite DP, in contrast to a copy, which may be interpreted as indefinite. Therefore, reconstruction is not a uniform phenomenon for movement and resumption, a prediction that can be confirmed by interpretative data. Also focusing on movement as a source of reconstruction, Marušiˇc makes an interesting proposal as an alternative for the copy theory of reconstruction. He analyzes cases of total reconstruction under A-movement (low scope of indefinite subjects). Note that total reconstruction is at odds with the idea that movement is driven by the need to assign scope (imposed by the C-I interface). As we showed in section 1.1 (example (7) and ensuing discussion), A-movement can feed wide scope and specific readings. But in languages like English, total reconstruction is also possible, as in Few students are likely to come. Marušiˇc points out that, in cases of total reconstruction, A-movement can be defined as phonologically visible but semantically innocuous. A mirror image of total reconstruction would be QR, in that it feeds interpretation but not phonology. His proposal is to accommodate these two facts under a redefinition of phase theory, whereby there are phonological phases and semantic phases which do not necessarily coincide, so that then phonological Spell-Out and semantic Spell-Out may partially overlap. This mismatch in Spell-Out has the effect that some movement (A-movement) is only phonological and some movement (QR) is only interpretative. Distinguishing phonological and semantic phases provides an explanation for an important range of sound/meaning mismatches in scope. These results are achieved at the cost of postulating two kinds of phases. If phases should be natural domains for processing, the question is whether the proposed phonological and semantic phases are natural domains for semantic and phonological objects respectively.


17

1.3.2 Adjuncts and interpretation Adjuncts appear to fulfill a variety of semantic functions on the category they adjoin to. Syntactically, however, it is standardly assumed that their link to an XP is much looser than that of complements or specifiers. Emonds explores the possibility that such a link is identifiable, beyond the basic adjunct– adjoinee structural relation. He argues in detail for the empirical generalization that all adjuncts are either PPs or agreeing XPs. The explanation for this pattern is that all lexical categories except P (N, V, A) need Case (in a generalized sense), obtained on adjuncts by either agreement or P. He derives this Case requirement from the proposal that positively specified categorial features ([+N] and [+V]) enter the derivation unvalued ([0N] and [0V]) and Case is precisely the device that values [+N]. Only P (which has no positive categorial value) is exempt from Case requirements. The proposal is of wide theoretical scope. It aims at unifying a broad range of apparently unrelated configurations in natural language: a type of Case is generalized for both [+N] or [+V] categories, for both arguments and adjuncts. Valuation by Case is ultimately an interface requirement: unvalued [0N] and [0V] cannot be interpreted at the interfaces, while they are perfectly legible to syntax. Within Chomsky’s Minimalist Program, Case appears to be a special uninterpretable feature, in that it does not accord with the generalization that an uninterpretable feature occurs on a head that acts as a probe: it occurs on the goal of the T/v probes. It is licensed, according to Chomsky (2001: 6), as a byproduct of the Agree operation, as a solution to the Case–Agreement puzzle. Emond’s proposal seems to push this puzzle to a much wider generality. Perhaps one of the interpretative components of grammar that have been most uncontroversially attributed by the MP to interpretation at the interface is Theta Theory: the Projection Principle establishes that syntactic representation of argument structure stems from selectional properties whose expression must be guaranteed throughout the derivation. Yet this simple picture is often obscured by many cases where the argument–adjunct divide is difficult to establish. Rákosi’s chapter proposes to characterize a class of thematic adjuncts to be distinguished from both arguments and adjuncts proper. With respect to datives, to which the chapter is devoted, thematic adjunct datives are distinguishable from argument datives in several syntactic and interpretative respects. Rákosi also provides a theta-theoretical basis for predicting the distribution of the two kinds of datives. The interesting issue that this chapter raises is the relation between the three types of datives (argumental, thematic adjunct, and non-thematic adjunct)

18

Introduction

and syntactic structure, which standardly allows for only one adjunct configuration. Perhaps it might turn out that the distinction Rákosi establishes between thematic and non-thematic adjuncts does not correspond to different types of adjunction structure but to different adjunct positions. In fact, Rákosi’s examples of non-thematic adjuncts are all sentenceinitial. It would then be in the VP area where (dative) adjuncts could be thematic. Adjunct positioning can be viewed as a matter of selection: the adjunct selects an adjoinee of the right (semantic) type. The question arises, then, if the selected type for a given adjunct corresponds to only one category (say, VP for a VP adjunct) or to several categories. The issue arises for adjuncts that appear in two positions with two corresponding interpretations and yet seem to make the same semantic contribution: they are likely not to be two homophonous lexical items. Csirmaz makes the point that durative adverbials (for-phrases) are interpretable in two distinct syntactic positions (under and above negation) without being ambiguous (i.e. their head—for in English—is not lexically ambiguous). This question arises in Cinque’s (1999) proposal on adverb placement in a cartographic view. It so happens that some adverbs can appear in more than one of the dedicated positions. Cinque is cautious in this respect in admitting that, although adverbials are specialized for one position, a given adverb may appear in more than one, possibly due to lexical underspecification. Csirmaz’s view that the same adverbial expression can appear in more than one position, provided it gives an interpretable output, is more in line with the Minimalist Program, where it is the Conceptual-Intentional interface that filters otherwise unrestricted adjunction structures. 1.3.3 Universal functional features CHL generates a set of convergent derivations. The set of convergent representations at the SM interface is obviously not universal, since each language has specific morphological and phonological patterns. The question is whether the set of convergent representations at the C-I interface is, in some sense, universal: whether for all languages CHL can compute the same set of derivations in narrow syntax (abstracting away from phonological features), giving rise to the same set of convergent C-I representations. A particular view on how syntactic representations are universal is the socalled cartographic project (Cinque, 1999; Rizzi, 1997, 2002): syntactic representations are highly uniform across languages both in their structure and


19

essential content. 21 Let us suppose that the cartographic project is descriptively correct. This would imply, in minimalist terms, that there is a universal set of convergent derivations (abstracting away from the meaning of descriptive roots). Since CHL is a blind mechanism for assembling lexical items, a universal set of convergent derivations should stem from the following conditions: (a) the set of formal features available in the lexicon is universal (Sigurðsson, 2003); (b) the set of functional lexical items (as sets of formal features) available in the lexicon is universal, whether they are overt or null; (c) all (relevant) functional lexical items must be used in a derivation; (d) the hierarchical arrangements imposed by the C-I interface are universal. Notice that requirement (b) is independent of requirement (a), since it is conceivable that a specific lexicon might contain lexical items with arrays of features that are not universal, due possibly to inflexional morphology, which creates more or less complex arrays of morphemes. Requirement (c) makes sure that, for instance, CHL cannot derive a representation in which certain projections are lacking, even if it might be convergent (for instance, a sentence without Mood or Aspect projections, if their corresponding lexical items are universally available). Requirement (d) is necessary to ensure that, in accord with the cartographic claims, a certain functional lexical item is always merged in the same hierarchical position, even if it is conceivable, under some semantic proposal, that a different hierarchical arrangement might be interpretable. Even if we do not (fully) adhere to all the tenets of the cartographic project, the generative tradition has mostly adhered to the view that syntactic structures are highly universal. Let us consider a kind of construction that has been claimed not to be uniform across languages: the Comparative Correlative (CC) (The more pizza Romeo eats, the fatter he gets). Taylor’s chapter in this volume argues that this construction is, contrary to what has been claimed by constructionalists, quite uniform across languages in all important respects: it consists of two clauses, the first clause being embedded under the second one; and there is A -movement of the comparative constituents within each clause. The embedded clause is not construction-specific, as it shares properties with conditional clauses. The only source of variation is the exact shape of the complementizers 21 In opposition to the cartographic project stands the constructionalist view, which emphasizes that specific, not universal, syntactic derivations are driven by specific and irreducible construction patterns (see, for instance, Culicover, 1999; Culicover and Jackendoff, 2005).

20

Introduction

(the for both the first and the second clause in English; cuanto for the first clause in Spanish, etc.), one option being that they are null. Taylor shows that, with this minimal source of variation, CCs appear to be universal. We can conclude that the properties of CCs are universal: lexical items (null or overt) of the appropriate kind to build a CC are universally available, and that C-I interface conditions for expression of “comparison” of this kind force the relevant External and Internal applications of Merge. Kayne’s proposal pushes the universalist view even further: he postulates null (silent) elements for a variety of cases where overt morphosyntax would suggest pervasive language variation. He specifically analyzes French nominative pronominals, and proposes empty pronominals that fill the gap between overt syntax and universal syntax. These patterns are clearly underdetermined by overt morphology. We can speculate that their universal character must stem from a universal lexicon with a rich array of (often null) lexical items, and from strict conditions on what counts as an interpretable representation at the C-I interface. Another field in syntax where there might seem to be important variation across languages is the behavior of adjectival modification in DPs and its interaction with definiteness. Leu discusses evidence from several languages (Greek and Germanic languages) where definiteness morphemes or adjectival inflection appear to be scattered on various heads in the DPs containing an adjective. He adopts the view (stemming from Kayne, 1994) that adjectival modification involves a clausal source from which both N and A originate. Taking this structure to be universal, and in view of the highly abstract nature of most of its constituents, again the question reappears: What forces such complex derivations? One possible reason is that “direct” adjectival modification (without the relative structure) is not interpretable because the C-I interface can only read a predicate (such as an adjective) if it is couched in a propositional structure. In this case, too, a universal structure implies that there are empty heads corresponding to empty lexical items from the universal inventory. The cases studied by Leu happen to involve more visible lexical elements expressing (in)definiteness. A specific problem for the theory of a universal lexicon is the existence of overt lexical items whose exact meaning is hard to establish in view of their apparent polysemous or vague content. It is difficult to characterize these items on the basis of the universal set of features, and they pose a problem for acquisition. A typical case is provided by lexical items that are used for the expression of definiteness and genericity. Munn et al. discuss the issue of how plurality, definiteness, and genericity are acquired, considering the intrinsic relation there seems to be between definiteness and genericity (perhaps due to


21

their relatedness at the conceptual interface), and the lexical correlations adult languages establish between these interpretative categories and determiners or classifiers. They address the acquisition of the -men morpheme in Chinese, a marker carrying both plurality and definiteness in adult language. Child errors seem to point to a preference of assigning definiteness and genericity to the same lexical item (-men in Chinese, the in English). If definiteness and genericity are commonly expressed by different arrays of features (as in adult English), why do they tend to be related in child language, like they are in Romance languages (where the definite article can express both definiteness and genericity)? This again raises the issue of whether functional lexical items are universal: the fact that some features (definiteness and genericity) tend to be combined in a single lexical item but need not be in other languages would suggest that functional lexical items are not universal. But highly analytical approaches like Kayne’s might preserve universality by resorting to an (overt or null) lexical item per feature. According to the Minimalist Program, linguistic expressions are derivations proceeding from a lexicon through syntactic computation to interfaces. Features have to be available in the lexicon, computable by syntax and interpretable at the interfaces. In this generative procedure, the derivation creates objects, and the interfaces filter them. Therefore, there must be an appropriate set of lexical features (including formal features) available to satisfy both formal requirements of computation and interpretative requirements of the interfaces. We can conceive of the generative enterprise as committed to establishing how the lexicon and syntax have come to be connected to a (preexisting) conceptual system in the human species. As Hinzen (2007: 15) recently phrased it, It may in fact be that most species have concepts, yet only one uses them to intentionally refer. The challenge is then to explain how concepts are put to use in an occasion. On the story told here, this depends firstly on the evolution of a lexicon which lexicalizes concepts through words, each of which consists in the pairing of a phonetic label with a concept or meaning. Secondly, it depends on embedding in hierarchical structural patterns correlating with specific semantic capacities. Particular kinds of patterns enable particular acts of intensional language use. The ideal is to see semantic complexity track syntactic complexity.

We hope that this volume contributes towards this ideal. The chapters in this volume were presented at the GLOW Meeting celebrated in Barcelona between the 5th and 8th of April, 2006, in the main

22

Introduction

session or the workshops on “Adjuncts” (organized by M. T. Espinal and J. Mateu) and on “The acquisition of the syntax and semantics of number marking” (organized by A. Gavarró and M. T. Guasti). We wish to thank three anonymous Oxford University Press reviewers for their invaluable suggestions on the volume, and especially the second reviewer for the detailed comments on every single chapter. Herewith our acknowledgment to Jon MacDonald and Ángel Gallego for their comments on the introduction. Any remaining errors are our own. Finally, our thanks to John Davey for his excellent editorial support.

Part I Formal features


2 Probing phrases, pronouns, and binding∗ FREDRIK HEINAT

2.1 Introduction The distribution of nominal expressions has given rise to a lot of debate in the literature (Chomsky, 1981; Reinhart and Reuland, 1993; Reuland, 2001; Zwart, 2002, among many others). Traditionally, their distribution has been regulated by the binding principles. The binding principles are not available in a minimalist syntactic theory, the reason being that two key relations in the definitions of the binding principles, government and co-indexing, don’t have any theoretical status in the Minimalist Program (Chomsky, 1995). However, Reuland (2001) claims that the complementary distribution of pronouns and reflexives can be accounted for as an effect of the operation Agree (after movement into checking positions (Chomsky, 1995)). Also, he claims that the notion of bound variable interpretation plays a crucial role. Consider the examples in (1). Co-indexing does not have any theoretical status. (1)

a. b. c.

Maryi saw herselfi . . . . and so did John. Maryi said that John saw heri . . . . and Lisa said so, too. ∗ Maryi saw heri . . . . and so did John.

For some reason a “locally” bound variable must be (spelled out as) a reflexive (1a), whereas a “non-locally” bound variable is (spelled out as) a pronoun (1b). In (1a), a bound variable reading is required, but not in (1b). Reuland suggests that only a reflexive allows the bound variable reading in (1a) because it enters into an Agree relation with its antecedent. A pronoun cannot enter such a relation and consequently does not allow a bound variable reading. I assume, in line with Reuland (2001), that the bound variable reading that is possible in ∗ I am grateful to Eva Klingvall, Satu Manninen, and two anonymous readers for comments on earlier versions of the chapter.

26

Formal features

(1b) is a different kind of relation, especially since only the one in (1a) has a morphological effect. There are two problems with Reuland’s analysis: it requires (overt or covert) movement of the reflexive for feature checking, and it predicts that 1st and 2nd person pronouns can be used as reflexives in all languages. 1 As is well known, this is true in some languages, for example Germanic languages other than English. It holds in Swedish, (2a), but as we can see in (2b) and (2c) it doesn’t hold for English. The reflexive pronoun is obligatory with a bound variable reading. (2)

a.

Jag slog mig. I hurt me ‘I hurt myself.’ b. ∗ I hurt me. c. I hurt myself.

In this chapter I will present an analysis that shows that there is an Agree relation between a reflexive and its antecedent but not between a personal pronoun and its antecedent. This Agree relation is a probe–goal relation, just like other Agree relations. Moreover, in line with Zwart (2002) (but contra Reuland (2001)), the morphophonological form of a pronoun/reflexive is an effect of the syntactic derivation. The analysis is couched in the framework of distributed morphology, which assumes that word formation is syntactic and that there are only roots in the lexicon that feeds the syntactic derivation. These roots are unmarked for reflexivity. The syntactic difference between reflexives and pronouns is a consequence of what category-forming head a pronominal root merges to. The problem of accounting for the complementary distribution of pronouns and reflexives can be split into two parts: r the difference between reflexives and pronouns r the formation of a syntactic relation between antecedent and reflexive:

probes. The outline of this chapter is as follows: in section 2.2, we will briefly look at probes. Section 2.3 deals with the structure of pronouns and reflexives. In section 2.4, we will look at how the agree relation between an antecedent and 1 Reuland claims that all 1st and 2nd person personal pronouns can enter an agree relation with their antecedent. Consequently, they should, irrespective of language, form chains and allow a bound variable reading (see Heinat, 2006: for details).

Probing phrases, pronouns, and binding

27

a reflexive is established. Finally, section 2.5 summarizes and concludes the chapter.

2.2 Probes Syntactic relations are formed via Agree between a probe and a goal (Chomsky, 2001, 2004, 2008). (3) Feature checking, then, resolves to pairs of heads < H, H > [. . . ]. For optimal computation, one member of the pair must be available with no search. It must, therefore, be the head H of the construction · under consideration, · = {H, XP}. Call H a probe P, which seeks a goal G within XP; . . . (Chomsky, 2004: 113) (emphasis in original). In short, a probe is the head of the structure and it searches its c-command domain. According to Chomsky (2001, 2008) the head is a probe because it is available without search. But since the label (or projection) of D and N in (4) and (5) is available for external merge without search, it should be a probe when it is merged to vP. (4)

(5)

ONML HIJK D ??? ?? ?

D N D {D, N}

Therefore, I will assume that all externally merged heads/labels are probes (cf. Epstein et al., 1998: 26–36). In the rest of the chapter I will refer to the label of D as DP, to avoid confusion. Note that only the label, that is DP, is a probe; the objects embedded in DP are not available without search and do not enter a relation with the elements DP agrees with. Now, what are the consequences of letting phrases probe? If we maintain Chomsky’s activation condition (Chomsky, 2001, 2004) that probes and goals are only active if they have unvalued features, there are no unwanted side effects. When the subject in (6) is merged, the label, DP, probes its domain. But since little v and the object have already valued and checked each other’s features, there are no unvalued features left, and no active goals in the domain of DP. So the subject DP does not agree with anything when it probes. When T is merged, the subject gets its unvalued case feature valued and the ê-features

28

Formal features

of T are valued. The conclusion is that there are no unwanted consequences of letting all labels probe. All syntactic objects, heads and phrases/labels with unvalued features are probes when they are externally merged (see Heinat, 2006: for details). (6)

TP A }} AAA } AA } } A }} T

vP A }} AAA } AA } } A }} DP subject

vA }} AAA } AA }} A }} v

inactive

VP

}AA }} AAA } AA } }}

V

DP inactive

To sum up this section, the assumption is that all externally merged syntactic objects with unvalued features are probes. There appears to be no unwanted consequences. Now, let us turn to the structure of the pronominal DP.

2.3 Pronouns The suggestion in this section is that personal and reflexive pronouns are formed from the same root (Zwart, 2002). The differences we see between personal pronouns and reflexives depend on what category-forming head the pronominal root merges to. First, we will look at word formation processes and what they can tell us about roots in general and the pronominal roots in particular. 2.3.1 Word formation On the assumption that word formation is syntactic (Halle and Marantz, 1993; Marantz, 1997; Josefsson, 1998; Julien, 2002; Embick and Noyer, 2001; among many others), the formation of a word proceeds as outlined in (7).


29

(7) Word formation (Josefsson, 1998) a. Hcat A }} AAA } AA } } A }} √ root Hcat b.

NA0 }} AAA AA }} } A }} √ root N0

c.

AA0 }} AAA AA }} } A }} √ root A0

Josefsson (1998) claims that a word is formed as in (7a). A category-neutral √ root, root, is merged to a category-forming head, H. In (7b) a noun is formed and in (7c) an adjective. The root lacks syntactic features—these are on the category-forming head. This means that inflectional morphemes are part of the category-forming head, not the root. Now let us look at compounds. Since we don’t find inflection inside compounds, as in (8) (see, for example, Williams, 1981), Josefsson (1998) claims that the first element in a compound is a bare root without a category-forming head. (8)

a. b. c.

cannonballs cannonsball ∗ cannonsballs ∗

Josefsson’s suggestion is that a compound, such as Swedish knäböja ‘kneebend’, ‘kneel’, is formed as in (9). (9)

Swedish Compounds (Josefsson, 1998) a. VA0 }} AAA AA }} } A }} √ böj V0 -a bend + verbal infl

30

Formal features b.

VA0 }} AAA AA }} } A } } √ knä VA0 }} AAA AA }} } A }} √ böj V0 -a kneebend ‘kneel’

First, the root böj ‘bend’ is merged to a category-forming head V which is instantiated with the morpheme -a, as in (9a). Then the root knä ‘knee’ is merged to the structure, as in (9b). Since the root knä never gets any inflection it is spelled out as a root. 2.3.2 Pronouns are roots If we take a closer look at how compounds and pronouns relate to each other, we see that pronouns occur in compounds. On the basis of the analysis of compounds we can conclude that pronouns are roots. Consider word formation: (10) English (Déchaine and Wiltschko, 2002; Rullmann, 2004) a. the me-decade (the 1970s), the me-generation, we-generation b. you-section, you-factor c. he-goat, she-devil, it-girl (11) Swedish a. jag-känsla, jag-centrerad, jag-föreställning, vi-känsla I-feeling, I-centered, I-image, we-feeling ‘me-feeling, me-centered, self-image, we-feeling’ b. dua, du-skål, du-reform, nia, ni-reform you(verb), you-toast, you-reform, you-pl(verb), you-pl-reform ‘to-say-you(sg.), drop-the-titles, you-reform, to-say-you(pl.), youreform’ It is clear from (10) and (11) that pronouns may occur as the first element in compounds. According to the analysis of compounds outlined above, this means that they are roots that haven’t merged to a category-forming head.


31

A further indication that pronouns are roots is that the root pronouns can merge with an N0 , forming a noun, as in (12), which we see examples of in (13) and (14). (12)

(13)

(14)

NA0 }} AAA AA }} } A }} √ pron. N0 a. Is it a he or a she? b. A whole new me. c. There will never be another you. d. The mes and yous in this world. Swedish a. Är det en hon eller en han? is it a she or a he b. Det egna jaget blir lidande. the own I-the becomes suffering ‘The self suffers.’ c. I den här boken vänder hon sig till ett annat du än in this here book-the turns she refl. to an other you than i sin förra diktsamling. (Teleman et al., 1999) in refl.poss. last collection of poems ‘In this book she turns to another you than in her last collection of poems.’

In the DPs above, the pronoun is the head noun in the noun phrases. On the basis of these data I propose the following structure for pronouns: (15) Referential DP DP D zz DDD z DD z DD zz zz NA0 }} AAA AA }} } A }} √ pron. N0 Further support for an analysis where pronouns are N-heads is the fact that they that can be (moderately) modified by preceding adjectives. Under the assumption that the DP has the structure as in (16) (Abney, 1987), adjectives precede the NP.

32

Formal features

(16) The DP (Abney, 1987: 213) DP }AA } } AAA } AA } }} D AP A }} AAA } AA } } A }} A NP

N (17) Swedish a. lilla jag/mej little I/me b. lyckliga du/dej/han/hon/dem happy you/you/he/she/them c. dumma hon/han stupid her/him (18) Huddleston and Pullum (2002: 430) a. poor old me b. lucky you As we can see in (17) and (18) adjectives may precede pronouns. However, this seems to be possible only when the pronouns don’t raise to D. Since pronouns occur in complementary distribution with determiners, the standard analysis is that referential pronouns raise to D0 . Let us turn to the reflexives and see how they can be incorporated in the analysis of pronouns. 2.3.3 Reflexives If reflexives are formed from the same roots as personal pronouns we don’t expect to see them in word formation: (19)

∗

himself-defense herself-contempt ∗ sej-försvar refl.-defense ∗ d. sej-förakt refl.-contempt a. b. c.

∗

Probing phrases, pronouns, and binding (20)

33

a. self-defense b. self-contempt c. själv-försvar self-defense d. själv-förakt self-contempt

As is clear from (19) and (20) it is impossible to use reflexives in word formation. Note that this is not due to the complex/simplex distinction that is sometimes made (Reinhart and Reuland, 1993). I suggest that reflexives have the following structure: (21)

Reflexive DP DP

0

√

DA }} AAA } AA }} A }}

pron.

D0

The prediction of the structure in (21) is that reflexives cannot be modified in any way since they lack all projections below D. This is also, to my knowledge, true. It is impossible to modify reflexives. In the next subsection we take a closer look at why reflexives, in contrast to pronouns, need an antecedent in the clause. 2.3.4 Why does the reflexive DP need an antecedent? In line with Pesetsky and Torrego (2007), I assume that there is a distinction between feature valuation and feature interpretability. Their claim is that there are in fact four kinds of features: (22) Features that are the input to the syntactic derivation 1. uninterpretable valued 3. interpretable unvalued 2. uninterpretable unvalued 4. interpretable valued The difference from a Chomskyan system is that interpretability is separated from feature values. However, in line with Chomsky (2001, 2004) the assumption is that all features must have a value—otherwise the feature

34

Formal features

cannot be deleted if it is uninterpretable, or it cannot be interpreted if it is interpretable. In (23) I give the feature set up of the DP (Julien, 2005). The N-head has uninterpretable and valued ê-features, and an uninterpretable unvalued T. According to Pesetsky and Torrego (2001), T is a case feature, but it differs from Chomsky’s case feature in that it behaves just like any other feature, and it has a valued counterpart. The features on D are interpretable but unvalued ê-features, and an uninterpretable unvalued T-feature. (23) The feature set up in DP (cf. Julien, 2005; Pesetsky and Torrego, 2005) a. N = uninterpretable valued ê-features b. N = uninterpretable unvalued T-feature c. D = interpretable unvalued ê-features d. D = uninterpretable unvalued T-feature So a DP is built as in (24). A root and N merge, (24a). Then D merges to the structure, (24b). Since D, an externally merged head, is a probe, the ê-features of D get their values from N via Agree, as in (24c). (24)

a. √ b.

AN [uT, vê] }} AAA } AA }} A }}

root

N [uT, vê]

}AA }} AAA } AA } }}

D [uT, uê]

NA [uT, vê] }} AAA } AA }} A }}

√ root c.

N [uT, vê]

DP [uT, vê] }AAA } AA } AA }} }} D [uT, vê]

N [uT, vê] }AAA } AA } AA }} }}

√ root

N [uT, vê]


35

All features in the DP get a value except T (case) which gets its value from a head in the extended projection of V. This is also the way a pronoun is formed. The reflexive on the other hand has the structure we see in (25). Since there is no N-head in this structure the reflexive DP will have unvalued ê-features. Only DPs with an N0 have valued ê-features, therefore the reflexive DP must get into an Agree relation with a DP with valued ê-features. (25) √

D [uT, uê] }AAA } AA } AA }} }}

pron.

D [uT, uê]

The different morphological forms we see are inserted after syntax. The two structures have lexical elements with different morphophonological forms inserted. 2.3.5 Cross-linguistic observations The difference between languages regarding reflexive objects seems to boil down to what kind of roots can be merged to D, as in (25). In (26) we see that languages make use of different roots. (26) Sources for reflexivity (from Schladt, 1999: 103) a. Body part names b. Sources denoting person, self, owner, etc. c. Emphatic pronouns d. Object personal pronouns Also, some languages, like San Lucas Quiaviní Zapotec (SLQZ), allow names to function as a reflexive (a bound variable) (from Lee, 2003): (27)

B-gwi’ih Gye’eihlly lohoh Gye’eihlly zë’cy cahgza’ Li’eb perf-look Mike at Mike likewise Felipe ‘Mike looked at himself, and Felipe did, too.’ (i.e. Felipe looked at himself/∗ Mike)

In (27) the name Mike functions as a bound variable and allows a reflexive interpretation. So instead of saying that in SLQZ names are anaphors sometimes and R-expressions sometimes, we can assume that in SLQZ a name root can be merged to either an N-head or a D-head as in (28). (See Barner and Bale (2002) for arguments that names are roots.)

36

Formal features

(28) √

D [uT, uê] }AAA } AA } AA }} }}

name

D [uT, uê]

Before we move on to the technical details of feature valuation, let us just sum up the main points in this section. The difference between pronouns and reflexives is not lexical, it is syntactic. They originate from the same root, but this root is merged to different heads (in line with Zwart 2002, but contrary to Reuland 2001). In (15) the root pronoun is merged to an N-head and we get the morphophonological form of a personal pronoun. In (21), on the other hand, the root pronoun is merged to a D-head and we get the morphophonological form of a reflexive, and the consequence is that he reflexive DP must get into an Agree relation with another DP to get values for its ê-features. (15) Referential DP (valued ê-features) DP zDD zz DDD z DD z z D zz NA0 }} AAA AA }} } A }} √ pron. N0 (21) Reflexive DP (unvalued ê-features) DP

DA0 }} AAA AA }} } A }} √ pron. D0

2.4 Binding This section deals with how the reflexive gets its ê-features valued. Assuming that all labels/heads, with all their features, valued and unvalued, are probes, it is possible to form a relation between a c-commanding DP and a reflexive DP, and all ê-features will get values and can be interpreted or deleted.


37

2.4.1 Feature sharing I assume that features that agree enter feature “chains” (Frampton and Gutmann, 2000; Pesetsky and Torrego, 2005, 2007), as in (29). An alternative analysis would be along the lines of “Multiple Agree” (Hiraiwa, 2001). (29)

Value Sharing Agree The feature F· of a probe · and the feature F‚ of a goal ‚ share the same value if they match and Agree (Agreement can be vacuous). All active/unvalued features F that share a value with ‚ in the c-command domain of · share the value of F· and F‚ . (Heinat, 2006)

The cases where we get feature chains are listed in (30). If a feature on a probe has a value +v and the goal has the same feature unvalued, the features match and Agree. The important case is when no feature has a value, the third case in (30). Then we get a feature chain, but no valuation. (30) feature on probe [+v]F [−v]F [−v]F [+v]F

feature on goal [−v]F [+v]F [−v]F [+v]F

Agree + + + −

Other assumptions are that a probe and a goal need at least one unvalued feature to be active, the activation condition, and that v and C are phase heads (Chomsky, 2001, 2004). In (31) we see the notation for feature valuation. The features with the same number are in a chain. (31) Notation for value sharing ê[2v] . . . ê[2v] . . . ê[5u] In (31) the ês with value [2] share value, the ê-feature with value [5] does not share the value of the other ê-features. The number is just an indication of a shared value and has no significance in the actual valuation of ê-features. The v stands for a valued feature and u stands for an unvalued feature. Interpretability is irrelevant to the feature valuation. 2.4.2 Forming a relation Now, consider (32). In (32a) the hypothesis is that there is an agree relation between the subject DP, Mary, and the reflexive herself. But, at the same time, we don’t want such a relation to form in (32b) and (32c).

38 (32)

Formal features a. Maryi likes herselfi . b. ∗ Herselfi likes Maryi . c. ∗ Maryi likes heri .

We will go through the derivations of the sentences in (32) and after that we will consider some problems that arise. First, we will look at (32a), renumbered as (33). Before little v and the reflexive are in a relation, they have different features and all of them, but the T-feature on v, are unvalued. Remember that the reflexive lacks ê-feature values since it consists of a root pronoun and a D0 , but crucially, it does not contain an N0 . This is shown in (33a). (33) Maryi likes herselfi v [ VP V a. T[2v], ê[2u]

b. c. d.

v

[V P

V

[ vP

v

[v P

v

T[2v], ê[5u]

DP

T[7u], ê[7v]

DP

T[2v], ê[7v]

refl.

]

refl.

]

refl.

]

refl.

]

T[5u], ê[5u] T[2v], ê[5u]

T[2v], ê[5u] T[2v], ê[5u]

T[2v], ê[7v] T[2v], ê[7v]

In (33b) little v and the reflexive are in an Agree relation. The T-feature of the reflexive has been valued and has the same number and value as that of little v, in this case 2. The ê-features, on the other hand, have formed a chain but they don’t have values yet since neither the reflexive nor little v has valued ê-features. In (33c) the subject is merged to the structure. Since the subject has an unvalued T-feature, it is a probe. In (33d) the subject DP has entered a relation with little v, and in the extension with the reflexive, since the reflexive and little v share their values. Also, all features in (33d) are now valued. The fact that the subject DP gets its T-feature valued by little v is a problem that we will return to. Now, consider (34). In this sentence we don’t want a relation to form between subject and object arguments. As is clear from (34a), the relation between little v and the object Mary leaves no unvalued features. The consequence is that there is no active goal available when the reflexive is merged in subject position. Since T doesn’t have any valued ê-features, the derivation will crash.


∗

39

Herselfi likes Maryi .

a.

vP

sKKKK KKK sss s s KK sss

v[T[1v], ê[2v]]

probe b.

VP

sKKKK KKK sss s s KK sss : DP [T[1v], ê[2v]] V

vP


refl.

vP

[T[3u], ê[3u]]

probe


v inactive

VP sKKKK s s KKK s s s KK sss V

DP inactive

The derivation of (32c) proceeds as in (35). The derivation is more or less the same as in (34). Since the root pronoun is merged to an N0 and therefore has valued ê-features, there will not be any active goals available when the subject DP Mary is merged in (35c). However, in contrast to (34), the derivation doesn’t crash, but it doesn’t allow a bound variable interpretation of the pronoun her. 2 (35)

∗

a.

Maryi likes heri v T[4v], ê[4u]

[ VP V

pron. ]

T[8u], ê[8v]

2 This doesn’t rule out the fact that Mary and her can be co-referential. There are well-known contexts that allow this type of co-reference:

(1) Everybody loves Mary. Mary loves her, too. Crucially, this co-reference doesn’t allow a bound variable interpretation.

40

Formal features b.

v T[4v], ê[8v]

c.

DP T[6u], ê[6v]

[V P [v P

V

pron

]

T[4v], ê[8v]

v inactive

pron ] inactive

As we saw above, the valuation of the T-feature on the subject DP by little v in (33) is problematic. This problem arises only when the object lacks valued ê-features. So, for example in (36), the subject DP has valued T from little v. We don’t want this feature value to “trickle down” into the rest of the DP. The main reason is that it would make the subject DP inactive since all its features would be valued, and when T is merged there are no active goals, leaving the ê-features of T unvalued and the derivation crashes. (36)

TP

}AA }} AAA } AA } }}

T

vP }AAA } AA } AA }} }}

DP = === == = D N

v

There are two approaches to this problem: first, in a multiple-Agree analysis, we would have to assume that DP is not an intervener for T and that T gets its ê-features valued by N, and N gets its T feature valued by T. In the featuresharing approach, the solution is to assume that each time a head/label probes it enters a new feature chain. D probes the DP that it heads. This is one chain D is part of. When D probes v and the reflexive, this is a new feature chain it is part of. In (37) we see the feature set-up of the subject DP. DP, the label, is part of two feature chains, one DP internal (labeled a) which gets its T-feature valued by T, and one DP external (labeled b) which gets its T-feature valued by little v. Now, we might expect that this leads to some kind of semantic clash or mismatch since there is a possibility for the two values to be different. But since T is uninterpretable on DP and the only purpose of the value is to make T on DP possible to delete, such a semantic clash will not occur.


ê

DP

[3v] T [1v]a , [9v]b

41

[D NP]

ê [3v]

T [1v]a

Finally, let us look at some other clause types where we find reflexives and where we don’t find them, and see how the proposed analysis can account for them. In general, the Phase Impenetrability Condition (PIC) (Chomsky, 2001), that states that at a phase head only the next lower phase head and its specifiers are available for syntactic computation, prevents reflexives in subject positions in finite clauses. However, in non-finite clauses the CP phase is missing and it should be possible to form a relation between the subject in the matrix clause and a reflexive in an ECM clause. This is also what we find; consider (38) and (39). In (38a) the reflexive has raised to the subject position in the embedded clause. But since the subject in the matrix clause is not merged until the reflexive has been spelled-out, as in (38b), the reflexive will never get its ê-features valued and the derivation crashes. (38)

∗

Elvis claimed that himself left the building. a. CP }AA } } AAA } AA } }} C TP A }} AAA } AA } } A }} refl. TA [T[1v]ê[3u]] }} AAA AA }} } A }} T

[T[1v]ê[3u]]

probe

vP A }} AAA } AA } } A }}

refl.

[T[1v]ê[3u]]

v

}AA }} AAA } AA } }}

v

VP }AAA } AA } AA }} }} left the building

42

Formal features b.

vP A }} AAA } AA } } A }} v phase head

VP

}AA }} AAA } AA } }}

V

CP }AAA } AA } AA }} }}

C phase head

TP −→ S-O }AAA } AA } AA }} }} refl. . . . v

In a non-finite embedded clause, as in (39), the reflexive raises to the subject position in spec-TP of the embedded clause, (39a). In (39b), little v in the matrix clause is merged, it probes and Agrees with the reflexive, just as in the transitive clause we looked at in (33). Then the subject is merged and the ê-features of the reflexive get values, as in (39c). (39) The King saw himself perform (on video) a. Tn f P }AAA } AA } AA }} }} refl. T AAn f } [T[1u]ê[3u]] }} AAA AA }} } } Tn f vP A [ê[3u]] }} AAA } AA } } A }} refl.

[T[1u]ê[3u]]

vA }} AAA } AA }} A }} perform


43

b.

vA }} AAA } AA }} A }}

v

[T[5v]ê[3u]]

VP

}AA }} AAA } AA } }}

probe V saw

c.

Tn Af P }} AAA AA }} } A }} / refl. TnAf [T[5v]ê[3u]] }} AAA AA }} } A }} perform vP A }} AAA } AA }} A }} DP vA [T[1v]ê[6v]] }} AAA AA }} } A }} v VP A [T[1v]ê[6v]] }} AAA } AA } } A }} V Tn Af P }} AAA saw AA }} } A }} refl.

[T[1v]ê[6v]]

TnAf }} AAA AA }} } A }} perform

One prediction of this analysis is that the first c-commanding DP with valued

ê-features must bind the reflexive. This is also what we see in (40). If we assume that vP is a phase that is not spelled out until C is merged, the analysis presented here can account for “chains” of reflexives in non-finite clauses as in (40e). (40)

a. b.

∗

Bart saw Lisai hurt herselfi . Lisai saw Bart hurt herselfi .

44

Formal features c. d. e.

∗

Bart saw herselfi hurt herselfi . Lisai saw herselfi hurt Bart. Lisai saw herselfi hurt herselfi .

It is clear from (40a) that the subject, Lisa, in the ECM clause can bind the reflexive in the object position in the same clause. From (40b) it is clear that the subject position in the matrix clause is not a position that can bind the embedded ECM reflexive. As shown in (40c), it is obvious that the ECM reflexive cannot be bound by another reflexive. In (40e) the chain formed between the two reflexives gets valued by the subject in the finite clause and all three DPs share ê-feature values.

2.5 Summary and Conclusion The first claim in this chapter was that all labels with unvalued features are probes, in other words, phrases, too are probes. Second, the difference between pronouns and reflexives is an effect of the syntactic derivation; personal pronouns are formed as in (15). This structure has all ê-features valued and needs only case, which it gets in a clausal structure. The reflexive is formed as in (21). This structure does not have values for its ê-features, since it lacks an N0 . Therefore, in addition to case, it needs an antecedent that can value its ê-features. The valuation of ê-features is a probe–goal relation. (15) Referential DP DP D zz DDD z DD z DD zz zz NA0 }} AAA AA }} } A }} √ pron. N0 (21) Reflexive DP DP

DA0 }} AAA AA }} } A }} √ pron. D0 The conclusion is that the distribution of reflexives and pronouns can be explained without making reference to binding principles. Instead, their


45

distribution is a consequence of the way probing and Agree work. The fact that pronouns and reflexives have different forms in certain languages, for example English, is a consequence of post-syntactic lexical insertion. This approach to binding not only accounts for the same data that binding theory in GB does but it also gives us a better understanding of why the binding domains of reflexives are restricted; they are a consequence of Agree and phases.

3 Wh-agreement and bounded unbounded movement∗ PATRICIA SCHNEIDER-ZIO GA

3.1 Unbounded dependencies: overview The classical view of movement (see Chomsky, 1977) is that it: (a) leaves a gap; (b) is prevented by islands/minimality; and (c) is subject to reconstruction. Displacement lacking these properties has been analyzed as base-generated displacement. A leading idea concerning movement that is unbounded is that it actually proceeds in a successive series of short, local steps. The phenomenon of wh-agreement has been taken as strong empirical evidence for successive cyclic movement (Chung, 1982: 39–77; McCloskey, 1979). This is because when wh-agreement occurs, a morphophonological reflex is registered on every clause along the path of a long distance dependency, as if movement had proceeded in a series of smaller steps. The phenomenon is illustrated in (1) for Kinande, a Bantu language. The complementizer kyo, which agrees in class with a displaced wh- or focused word, marks the dependency in each clause between the displaced word and the position to which it is thematically related: (1)

[ekihi kyo Kambale asi [nga kyo Yosefu akalengekanaya what wh-agr K know comp wh-agr Y thinks [nga kyo Mary’ akahuka __ ]]] comp wh-agr M cooks ‘What did Kambale know that Yosefu thinks that Mary is cooking (for dinner)?’

I will establish that, contra initial appearances, Kinande has no successive cyclic movement. Instead, resumption is required to accomplish long distance ∗ I am grateful to Yen-Hui Audrey Li, the audience at the 29th GLOW Colloquium in Barcelona, and an anonymous reviewer for valuable discussions and illuminating comments concerning this chapter. All errors and misinterpretations of comments and suggestions are mine.

Wh-agreement and bounded unbounded movement

47

displacement. I discuss the economy implications of resumption under this condition. Finally, this view suggests that minimalism must consider the possibility of base-generating multiple resumptive copies, which, as noted by an anonymous reviewer, poses the non-trivial question of how the relevant links emerge.

3.2 Distribution of wh-agreement In this section, I will establish the distribution of wh-agreement. The examples in (2a and b) illustrate local displacement of a wh-expression. Note that a lexical item glossed wh-agr (wh-agreement) immediately follows the whexpression and agrees in class with it. The wh-expression in (2a) is a member of class seven and the wh-expression in (2b) is a member of class one. (2)

Kambale alangira e j ] a. [Ekihi j kyo j saw Whatj wh-agrj K ‘What did Kambale see?’ Kambale alangira e j ] b. [Iyondi j yo j saw whoj wh-agrj K ‘Who did Kambale see?’

The following examples illustrate the distribution of wh-agreement when displacement is long distance: wh-agreement occurs in every clause along the path of the displacement. In (3), the displaced wh-expression is interpreted as the object of the verb in the most deeply embedded clause and we see that whagreement occurs in the embedded clause, the intermediate clause, and the matrix clause: (3) [ekihi kyo Kambale asi [nga kyo Yosefu akalengekanaya what wh-agr K know comp wh-agr Y thinks [nga kyo Mary’ akahuka __ ]]] comp wh-agr M cooks ‘What did Kambale know that Yosefu thinks that Mary is cooking (for dinner)?’ Not only can wh-agreement occur, it must occur in every clause between the site of extraction/most deeply embedded wh-agreement morpheme and the site of phonological location of the wh-word. The example in (4a) indicates that it is not possible to have a wh-agreement particle in the most embedded clause and in the intermediate clause, but not also in the matrix clause. (4b) establishes that it is also not possible to have a wh-agreement particle in the most embedded and the most superordinate clause, without also having a whagreement particle in the intermediate clause:

48 (4)

Formal features a. [ekihi ∗ (kyo) Kambale asi [nga kyo Yosefu what wh-agr K knows comp wh-agr Y akalengekanaya [nga kyo Mary’ akahuka ____ ]]] thinks comp wh-agr M cooks ‘What did Kambale know that Yosefu thinks that Mary is cooking (for dinner)?’ b. [ekihi kyo Kambale asi [nga ∗ (kyo) Yosefu what wh-agr K knows comp (wh-agr) Y akalengekanaya [nga kyo Mary’ akahuka]]] thinks comp wh-agr M cooks ‘What did Kambale know that Yosefu thinks that Mary is cooking (for dinner)?’ c. ∗ [ekihi Kambale asi [Yosefu ng1 ’akalengekanaya [nga what K knows Y comp’thinks comp (kyo) Mary’akahuka ]]] (wh-agr) M’cooks ‘What did Kambale know that Yosefu thinks that Mary is cooking (for dinner)?’

Complementizers do not have these properties. They do not generally occur immediately following a displaced wh-word. They do not need to occur in every clause along the path of extraction. These generalizations are illustrated in (5a and b) (5)

a. b.

∗

I know who that left. What do you think (that) Bill said (that) Mary read?

Wh-movement, as indicated by wh-agreement, appears to proceed in very short steps. Is this step-by-step intermediate movement feature-driven on every cycle such that the agreeing complementizer is a reflex of featurechecking of the chain involving the displaced expression? The answer proposed here will be “no”.

3.3 Movement The wh-agreement facts seem to indicate that movement in Kinande proceeds in a series of cyclic steps, marked by the appearance of a wh-agreement particle in every clause involved in the dependency. However, the syntactic evidence I examine next indicates that there is no long distance movement of a 1 The complementizer nga is a clitic and thus is found in several different positions with respect to agreement morphemes in the sentence. See Schneider-Zioga (2007) for discussion.


49

wh-expression. Indeed, I will establish that there is no long distance movement of any kind. 3.3.1 Evidence from reconstruction Despite the morphosyntactic facts from wh-agreement, reconstruction facts in Kinande do not support an analysis of successive cyclic movement. Consider reconstruction and bound pronouns. Recall that a bound reading requires ccommand of the pronoun by the relevant QP. This means that in order for a bound reading to obtain, the expression containing the bound pronoun must reconstruct to some site below the relevant QP. Consider local A -extraction in Kinande, here involving focus. 2 The example under consideration, (6), involves a bound pronoun contained within a locally displaced focused phrase. We see it allows for a reconstructed interpretation and thus behaves as if movement has taken place. (7) schematically illustrates the reconstruction that is possible in such a construction: (6) ekitabu kiwej/k ky’ obuli mukoloj akasoma kangikangi. book his wh-agr each student reads regularly ‘(It is) Hisj book that [every studentj/k ] reads regularly.’ (7) QP c-commands pronoun ↓ [ …… [ [every student] reads [his book] regularly ]] ↑ reconstruction Long distance A -movement, however, behaves differently. A reconstructed interpretation of an expression that has been displaced long distance is not possible: (8)

ekitabu kiwek/∗ j kyo ngalengekanaya [C P nga kyo that wh-agr book his wh-agr I:think [obuli mukolo]j akasoma __ kangikangi] every student read regularly ∗ ‘(It is) Hisk/ j book that I think [every student]j reads regularly.’

The ungrammatical possibility is illustrated schematically below:

2 As far as I have been able to determine, focus constructions and wh-question constructions have the same syntax.

50

Formal features

(9) ∗[ …… [ I think

QP c-commands pronoun [ [every student]

reads

↓ [his book] regularly]]] ↑

reconstruction

In the literature on reconstruction, it has been observed that reconstruction can be partial. That is, reconstruction can be to a point midway between the putative extraction site and the phonological location of the displaced expression. To test this possibility, a QP would need to occur in the superordinate clause. This is so that reconstruction to an intermediate position would still put the pronoun contained within the partially reconstructed expression in the c-command domain of a QP which could then bind the pronoun. However, not even partial reconstruction is possible in Kinande. This impossibility is illustrated below in (10) and schematically in (11): (10) ekitabu kiwek/∗ j kyo [obuli mukolo]j alengekanaya [C P nga comp book his wh-agr every student think kyo nganasoma __ kangikangi] wh-agr I:read regularly ‘(It is) Hisk/∗ j book that [every student]j thinks I read regularly.’ QP c-commands pronoun (11) ↓ ∗[…….[ [every student] thinks [[his book] [ I read ___ regularly ]] ↑ partial reconstruction The fact that partial reconstruction is not grammatical also rules out another possible analysis of long distance displacement. Namely, it has been observed in the literature that in some languages and/or under some circumstances an expression that is displaced long distance doesn’t originate in an embedded argument position. Instead, it originates on the left edge of the embedded clause, perhaps in some type of predication structure (see, for example, Iatridou, 1990), and, from the position on the embedded left edge, could move into the left edge of a superordinate clause. If this configuration occurred, it would mean that reconstruction would not be possible to argument position of the embedded clause. This is because the displaced position did not originate in the embedded argument position. However, reconstruction would be possible to the left edge of the embedded clause, since this is the original site of the displaced position. Under this circumstance, reconstruction of expressions displaced long distance would appear to be possible only if partial reconstruction is considered. Since a partial reconstruction interpretation is not possible


51

in Kinande, we can confirm that the displaced expression does not seem to undergo movement at all, not even from a non-argument position. The classical view of reconstruction under the copy theory of movement is to tie it to movement, and hence to properties of opaque versus transparent domains. However, the above contrast (local displacement allowing reconstruction, long distance displacement forbidding it) does not depend on any standard notion of island. This need not compel us to reject the classical view of reconstruction as long as we interpret this as a problem for successive cyclic movement. That is, given the classical view of reconstruction, there appears to be no successive cyclic A -movement in Kinande. 3.3.2 Evidence from superiority effects As pointed out by an anonymous reviewer, superiority effects in multiple whquestions should be able to provide consolidating evidence bearing on the existence of movement in Kinande: superiority effects are generally attributed to some type of requirement that movement be as short as possible. In Kinande, superiority effects are not exhibited when clausemate wh-words in a single clause are involved, so that the Kinande equivalent of what does who like is grammatical with a distributed interpretation (see Schneider-Zioga 2007 for discussion): 3 (12) Ekihi kyo ndi alangira what wh-agr who saw ‘Who saw what?’ In this way, Kinande is like German, Polish, or Spanish, as reported in the relevant literature (see, for example, Fanselow 2004). In fact, superiority effects between clausemate wh-words in the same clause are not observed in a number of languages that clearly have movement. Therefore, the fact that superiority effects do not emerge in Kinande when clausemate wh-words are involved in a single clause is rather equivocal. In any case, it doesn’t contradict the reconstruction data that has been presented in support of local A -movement occurring. However, in Kinande, superiority effects also do not arise when the superior wh-word is in a different clause from the inferior wh-word. In this, it seems 3 A distributed interpretation is also possible if the subject is moved, and the object left in situ. Note the different morphology on the ex-situ wh-words and the presence of wh-agreement, confirming that the subject has indeed moved under these circumstances:

(i)

iyondi yo walangira ki who wh-agr saw what ‘Who saw what?’

52

Formal features

Kinande distinguishes itself from the languages previously mentioned. The following example illustrates a non-clausemate superiority configuration in Kinande; the interpretation is that of a distributed question: (13) ekihi kyo ndi anasi nga kyo Josefu abula what wh-agr who know if wh-agr Joseph bought ‘Who knows if Joseph bought what?’ (Literally: ‘What does who know if Joseph bought?’) The fact that Kinande is impervious to superiority effects in cases of nonclausemate wh-words can be accounted for if long distance wh-displacement does not involve actual movement across a superior wh-word. 3.3.3 Cyclic A-movement is also impossible In this section, I will establish that no type of cyclic movement appears to be possible in Kinande: long distance A-movement is also ungrammatical. However, in parallel to the observation that local A -movement is possible, local A-movement is also grammatical. Sentence (14a) illustrates a monoclausal active sentence and (14b) illustrates its corresponding passive. Note that abakali, the logical and grammatical subject in (14a), is expressed as the object of a preposition in the passive (14b). Note further that the logical object is now expressed as the grammatical subject in (14b), with the subject/verb agreement reflecting the grammatical subjecthood of the logical object. (14)

a. abakali bahuka ebikene. women agr:cooked yams ‘The women cooked the yams.’ b. ebikene byahuka-wa na bakali. yams agr:cooked-pass with women ‘The yams were cooked by women.’

Consider now subject raising, a long distance A-displacement. The example in (15) illustrates a non-raised counterpart of the subject-raising sentences we will subsequently consider in (16a and b): (15) ali-[nga omulume ananzire ekitabu ] agr:is-if man likes book ‘It seems that the man likes the book.’ The putative subject-raising cases do not look like typical subject raising because the putatively raised subject agrees with the verbs of the subordinate and superordinate clauses.

Wh-agreement and bounded unbounded movement (16)

53

a. Omulume a-kavya [ng’a-nanzire ekitabu ] man agr-is if ’agr-likes book ‘The man seems to like the book.’ b. abalume ba-kavya [nga ba-nanzire ekitabu] men agr-is if agr-like book ‘The men seem to like the book.’

It has been reported in the literature that some languages have raising out of agreeing subjunctive clauses (see Boeckx, 2003, for a recent overview). Therefore, the presence of agreement on the verb of every clause along the path of the A-dependency cannot in itself be enough to support the claim that there is no cyclic A-movement. The following data however provide conclusive evidence that movement does not occur in the putative raising constructions in Kinande. In these examples we see all sorts of well-established constraints on movement flagrantly violated. In (17) we see an example of an object “raised” across the subject of the embedded clause and enter into agreement with the superordinate verb. In (18) we see a grammatical example of “super-raising”: the subject of the most deeply embedded clause is “raised” across the subject of an intermediate clause: (17)

ekitabu j ki-kabya [[subject omulume] ng’ana-ky-anzire _____ j ] man if ’ he:object_clitic.likes book agr-is Lit: ‘The book seems as if the man likes it.’ (18) omulume j a-kavya [subject Marya] ng’akalengekanaya [ng’_____ j M if ’agr:thinks man agr-is if ’ anzire ekitabu] agr:likes book ‘The man seems as if Mary thinks he likes the book.’ This is not a problem for these typical diagnostics of movement provided we conclude that movement is simply not involved in these constructions. It seems more plausible to analyze these as a kind of tough movement construction. Note that in (17) the displaced object is actually resumed by a clitic in the subordinate clause. I return to this fact in the next section where I argue that long distance displacement in Kinande involves resumption.

3.4 Resumption In the previous section I established that long distance movement does not exist in Kinande. In this section I will argue that long distance dependencies involve resumption. I will first consider A -dependencies and motivate the

54

Formal features

existence of a null resumptive pronoun that can be immediately followed by wh-agreement under certain conditions. In the next section, I present evidence that wh-agreement can be either (a) preceded by the overtly displaced expression (when head of the dependency), or (b) by a null expression: (19)

a. [Wh/Focusj wh-agr [IP . . . ___j . . . ]] b. [OP(erator)j wh-agr [IP . . . ___j . . . ]]

This suggests the possibility that an “unbounded” dependency has the following structure schematically: (20) [Wh/Focusj wh-agr [IP . . . [CP OP(erator)j wh-agr [IP . . . ___j . . . ]]]] 3.4.1 The pronominal nature of the null expression Wh-agreement can be immediately preceded by a demonstrative pronoun (21a) and receives a focused interpretation in this context. (21b) illustrates that a null expression preceding wh-agreement in an otherwise identical context is interpreted in the same way: (21)

a. ekyo kyo Kambale alangira. that wh-agr K saw ‘THAT (is what) Kambale saw.’ b. [ ] kyo Kambale alangira wh-agr K saw ∗ ‘what did Kambale see?’ ∗ ‘what Kambale saw (free relative)’ ok: ‘THAT (is what) Kambale saw.’

The pronominal nature of the null expression preceding wh-agreement is independently supported by the behavior of wh-agreement in island contexts. (22a) illustrates that gaps may not occur in islands. (22b) demonstrates that once the gap is embedded one clause deeper, the sentence becomes grammatical. 4 Note that wh-agreement occurs in the island only in (22b): (22)

a.

∗

omukali ndi yo wasiga [island embere __ wabuga] woman who wh-agr you:left before spoke ∗ ‘Which woman did you leave before (she) spoke?’

4

For completeness, I include this example of wh-agreement related to the grammatical displacement of an object across an island. As with subjects, displacement of an object across an island is grammatical once it is embedded deeply enough that wh-agreement can occur: kyo wasoma ___ ]] (i) Ekihi kyo uasiga [island embere Marya aminye [nga before M knew compl wh-agr you:read what wh-agr you:leave ‘What did you leave before Mary knew you had read___?’


55

b. omukali ndi yo wasiga [island embere Kambale anasi [C P woman who wh-agr you:left [ before K knew [ ko yo __ wabuga]] that wh-agr spoke ]] ‘Which woman did you leave before Kambale knew that she spoke?’ This property of wh-agreement parallels exactly the rescuing behavior of obvious resumptive pronouns in island contexts where we see that it is ungrammatical for a dependency to exist across an island unless a resumptive pronoun also occurs: Dependency across an island—without resumption: (23)

a.

∗

ekihi j kyoj Constantine abula [island iyondik nga yo k __k what wh-agr C wondered who if wh-agr uagula __j ] bought ∗ ‘What did Constantine wonder who bought?’ b. ∗ ebaruha yahi j yo wasiga [island isi-wu-li letter neg-you-be which wh-agr you:left uasoma ___ j ] you:read ∗ ‘Which article did you leave before you read?’

Dependency across an island—resumption involved: (24)

a. ekihi j kyo Yosefu akabula [island iyondik nga y’ what wh-agr Y wonders who if wh-agr’ uka-ki-gula ] agr:tense-clitic-buy ∗ ?‘What does Yosefu wonder who is buying (it)?’ b. ebaruha yahi yo uasiga [island isi-wu-li letter which wh-agr you:left neg-you-be uasoma-yo ] you:read-clitic ∗ ?‘Which article did you leave before you read it?’

We can summarize the facts schematically as follows: (25)

a. ∗ whj [ . . . . b. ok: whj [ . . . . c. ok: whj [ . . . .

[island .. . . . . . . ___j ]] [island .. . . .. . . proj ]] [island .. . . .. . . wh-agrj ]]

56

Formal features

Therefore, the null expression that precedes wh-agreement can be analyzed as a pronominal operator fulfilling a resumptive function, similar to the view of McCloskey (2002: 184–226) for wh-agreement in Irish or Davies (2003: 237– 59) for wh-agreement in Madurese. Other relevant non-movement analyses of wh-dependencies would be also Adger and Ramchand (2005: 161–93), and Finer (2002: 157–69) for Selayarese. 5 In addition, Boeckx (2007, 2008), based solely on the position his theory forces him to, concludes that long distance dependencies in Kinande cannot involve successive cylic movement. Rather, he concludes that something along the lines of the prolepsis proposed by Davies for Madurese must be at play in Kinande. Bošković (2007b) also comes to a similar conclusion about Kinande based on his theory of feature checking. 3.4.2 A reconsideration of raising Here I reconsider the putative raising constructions we examined earlier. We can observe that long distance A-displacement is parallel to A -displacement where resumptive pronouns and wh-agreement have the same distribution and function. (26) demonstrates that the embedded object can be displaced across an embedded object and wh-agreement occurs in the embedded clause. (27), a repeat of sentence (17), confirms that a resumptive clitic can also occur in this context. (26) ekitabu j ki-ri-[nga kyo [subject omulume] anzire ____ j ] man he:likes book agr-is-if wh-agr Lit: ‘The book seems as if the man likes it.’ (27) ekitabu j ki-kabya [[subject omulume] ng’a.na.ky.anzire _____ j ] man if ’ he.tense.clitic.likes book agr-is Lit: ‘The book seems as if the man likes it.’ In sum, I have introduced a variety of data to provide evidence that whagreement is associated with a null pronominal operator. This provides support for the proposal that “unbounded” movement in Kinande actually involves resumption. I repeat the (schematic) proposed structure (20) here for convenience: (28) [Wh/Focusj wh-agr [IP . . . [CP OP(erator)j wh-agr [IP . . . ___j . . . ]]]]

5 Although a discussion of these interesting and important papers is beyond the scope of this chapter, it is not insignificant that all of the languages noted above, for which non-movement dependencies have been proposed, are wh-agreement languages.


57

3.5 Some issues raised by the lack of movement in Kinande The lack of movement in a wh-agreement language has implications for the theory of feature checking. To see this, consider the following. Comp to comp movement, assuming it exists, raises problems for the idea that all movement is last resort. What then motivates the intermediate steps in unbounded movement? Hornstein (2001: 119) suggests that the checking of A agreement features drives movement to intermediate wh-positions. He bases his proposal on the morphology of wh-agreement languages, whose existence he takes as evidence for his position. However, based on what I have established here concerning the syntax of Kinande, wh-agreement cannot be taken as evidence for feature-motivated intermediate checking of links of an A movement chain any more than it can be taken as evidence of successive cyclic movement. I have established that movement in Kinande is clause-bound. Why is movement so restricted in this language? It appears that there is no successive cyclic movement because even embedded clauses in Kinande behave like root clauses in V-2 languages. Following Emonds (1970), a root clause allows maximally one fronting per S, and since root clauses are not the target of further movement, they lack escape hatches. This one-fronting-perS rule operates in every clause in Kinande producing an X-second pattern where the initial expression is not structure-preserving in that it can be either a DP or PP (see Schneider-Zioga, 2005, 2007, for extensive discussion); the second position element is either a verb or complementizer which agrees with the preceding expression. This pattern is even found embedded under non-bridge verbs, affirming the uniform character of clauses in Kinande. I will establish that embedded clauses in Kinande behave as if they were root clauses in that their syntax is insensitive to the superordinate verb. Sentences (29)–(33) illustrate a variety of superordinate verbs. The clauses embedded under these verbs display SVO word order: (29)

Kambale mwakanganirye kwenene Marya ko’mwagula ekitabu. K proved that M that’bought book ‘Kambale proved that Mary bought the book.’ (30) Kambale anasi Marya ko’mwagula ekitabu. K knows M that’bought book ‘Kambale knows that Mary bought the book.’ (31) Kambale anasadikirye Marya ko’mwagula ekitabu. K claims M that’bought book ‘Kambale claims that Mary bought the book.’

58

Formal features

(32) Kambale mwahakikisya Marya ko’mwagula ekitabu. K emphasized M that’bought book ‘Kambale emphasized that Mary bought the book.’ (33) Kambale mwayibulya Marya nga’mwanagula ekitabu. K doubted M if ’bought book ‘Kambale doubted if Mary bought the book.’ It is also possible for an inversion to take place such that embedded XP-V-S word order occurs. The grammaticality of this word order and the fact that its grammaticality is insensitive to the superordinate verb is illustrated below in sentences (34)–(38): (34) Kambale mwayibulya ekitabu nga ky-agula mukali. K doubted book if agr-bought woman ‘Kambale doubted that a woman bought the book.’ (35) Kambale mwayibulya omomulongo nga mo mwanahika mukali. K doubted in:village if agr arrived woman ‘Kambale doubted that a woman arrived in the village.’ (36) Kambale mwahakikisya ati ekitabu kyagula mukali. K emphasized say book agr:bought woman ‘Kambale emphasized that a woman bought the book.’ (37) Kambale anasi omomulongo ko mwahika mukali. K knows in:village that agr:arrived woman ‘Kambale knows that a woman arrived in the village.’ mukali. (38) Kambale anasi ekitabu ko kyagula K knows book that agr:bought woman ‘Kambale knows that a woman bought the book.’ The above data help establish the root-like nature of embedded clauses in Kinande. This supports the claim that all clauses are root clauses in Kinande. This, in turn, sheds light on the impossibility of unbounded movement in Kinande.

3.6 Economy and resumption—some tentative conclusions The classical view of movement is that it: (a) leaves a gap; (b) is prevented by islands/minimality; and (c) is subject to reconstruction. Displacement lacking these properties has been analyzed as base-generated displacement. A minimalist approach must wonder about the existence of two distinct operations to construct displacement: base-generation and movement, although a priori it is not clear whether it is a luxury for the grammar to prohibit one of the types of operations or a luxury to allow both types.


59

Minimalist researchers have proposed the following disparate positions: (i) There is no displacement via base-generated construal of some type. Instead illicit movement is possible and can be repaired by resumptive pronouns, that is, resumption rescues island violations (see Hornstein, 2001). (ii) There is no displacement via base-generated construal. Illicit movement can be enabled via a big DP which has independent status in the lexical array (see Boeckx, 2003), where the resumptive pronoun is a stranding of the determiner of the big DP. (iii) A lexical item with the same status as other pronouns can be used to achieve displacement via base-generation (see Aoun and Benmammoun, 1998; Aoun and Li, 2003). The Kinande facts surrounding “successive cyclic” movement demonstrate the need for a notion of resumption whose characteristics cannot be attributed to any purported properties of illicit movement, since the dependency in question clearly does not involve islands of the typical sort. This argues for the necessity of displacement via base-generation at least as a language-particular option and supports minimalist position (iii). Finally, the Kinande facts indicate that minimalism will need to reconsider resumption itself. That is, we see there is a resumptive strategy that is different from typical pronominal resumption, where a single resumptive pronoun stands in a relation with a displaced expression across an apparently unbounded distance. Instead, we see a language where multiple resumptive copies must be base-generated. Therefore, minimalism must address this possibility and work out how the relevant links of such a resumptive relation emerge.

4 Universal 20 without the LCA∗ KLAUS ABELS AND AD NEELEMAN

4.1 Introduction There is general agreement that linguistic theory should account for linear asymmetries found in language. This chapter presents a case study of one such asymmetry: the facts uncovered in work following up on Greenberg’s universal 20 (Greenberg, 1963; Hawkins, 1983; Rijkhoff, 1990, 2002; Cinque, 2005a). We confront the known facts with a particular theory of linear asymmetry, Kayne’s (1994) Linear Correspondence Axiom (LCA). We distinguish two claims made by the LCA, one concerning base generation and one concerning movement. Regarding base generation, the LCA claims that specifiers universally precede heads and that heads universally precede their complements; regarding movement, the LCA claims that all movement is to the left. Given the LCA, the second claim follows from the first. We will argue, however, that a weaker theory, one that embraces only the restriction to leftward movement and jettisons the idea that base-generation is universally ordered, is to be preferred (see also Ackema and Neeleman, 2002). It may seem paradoxical but the grounds for this preference come from restrictiveness. While our base component is more permissive, it allows the movement component to be restrictive to a degree that strengthens the theory as a whole. How can this be shown? Our argument proceeds as follows. First, we show that, as far as universal 20 is concerned, the restriction to a universal underlying specifier-before-head-before-complement order does not add restrictiveness. In the extended projection of the noun the same range of orders and ∗ Parts of this material were presented at the EGG summerschool (Wroclaw, July 2005), a minicourse “Universal 20 without the LCA” (Leipzig, December 2005), the Left-Right Seminar (Tromsø, spring 2006), a colloquium in Edinburgh (February 2006), and at GLOW 29 (Barcelona, April 2006). We are grateful to the audiences on all these occasions as well as our colleagues in London and Tromsø for their questions, comments, and suggestions. We gratefully acknowledge the financial support of CASTL that allowed us to write this chapter. None of the above are responsible for any shortcomings of this manuscript, which should be blamed entirely on the authors.

Universal 20 without the LCA

61

associated structures can be generated as long as movement is uniformly leftward. The empirical burden of explanation thus rests entirely on restrictions on movement, including the ban against rightward movement. This result generalizes to other linear asymmetries. They cannot be explained without a restrictive theory of movement. However, the base-generation claim made by the LCA makes a restrictive theory of movement impossible since all variation in ordering has to be derived through movement. In Abels and Neeleman (2006) we provide a rationale from parsing for the assumption that movement must (at least in the cases at hand) be leftward while allowing symmetrical base-generation. For reasons of space this part could not be included here.

4.2 Cinque’s Theory Extending earlier work of his on the order of elements in the noun phrase (Cinque, 1996, 2000), Cinque (2005a) argues that the typology of word order in the extended nominal projection can be explained if four assumptions are made, among which Kayne’s (1994) Linear Correspondence Axiom. The empirical domain that Cinque concentrates on concerns the orders in which demonstratives, numerals, adjectives, and nouns appear in the extended nominal projection. On the basis of careful typological work, he argues that of the 24 logically possible orders of these elements, only 14 are attested as unmarked word orders in natural language. The typological pattern is illustrated below: (1) a. (i) (ii) (iii) (iv) b. (i) (ii) (iii) (iv) c. (i) (ii) (iii) (iv) d. (i) (ii) (iii) (iv)

DEM NUM A N these five young lads DEM NUM N A DEM N NUM A N DEM NUM A DEM A NUM N DEM A N NUM DEM N A NUM N DEM A NUM NUM DEM A N NUM DEM N A NUM N DEM A N NUM DEM A NUM A DEM N NUM A N DEM NUM N A DEM N NUM A DEM

attested attested attested attested unattested attested attested attested unattested unattested unattested unattested unattested attested attested attested

62

Formal features e. (i) (ii) (iii) (iv) f. (i) (ii) (iii) (iv)

A DEM NUM N A DEM N NUM A N DEM NUM N A DEM NUM A NUM DEM N A NUM N DEM A N NUM DEM N A NUM DEM

unattested unattested attested attested unattested unattested attested attested

The main theoretical contribution of Cinque’s paper consists in a demonstration that the assumptions given below generate the 14 existing orders, while excluding the ten nonexisting ones: (2) a. The underlying hierarchical order in the extended projection of the noun is Agrw WAgr X XAgrY YN where Y hosts AP in its specifier, X hosts NumP in its specifier, and W hosts DemP in its specifier; b. all (relevant) movements move a subtree containing N; c. all movements target a c-commanding position; d. all projections are modeled on the template (Kayne, 1994): [ XP Spec [ XP X0 Compl ] ] Let us consider how this result obtains. When the noun is final within DP, the prenominal material can occur in only one order. This order is base-generated in line with assumptions (2a) and (2d). No movement can have taken place, because assumptions (2c), (2b), and (2d) imply that movement results in an order where the noun is not final. This rules out the orders in (1ei), (1fi), (1bi), (1di), and (1ci). (3) [ Agrw P Agr0w [ WP DemP [ W0 [ Agr X P Agr0X [ XP NumP [ X0 [ AgrY P Agr0Y [ YP AP [ Y0 NP]]]]]]]]] A second class of structures is generated by moving NP to [Spec, AgrY P], [Spec, Agr X P], or [Spec, AgrW P]. This will generate all orders in which the underlying sequence Dem≺Num≺A surfaces, while the position of the noun varies ((1aii), (1aiii), and (1aiv)). (4) [ Agrw P (NP) [ Agrw [ WP DemP [ W [ Agr X P (NP) [ Agr X [ XP NumP [ X [AgrY P (NP) [ AgrY [ YP AP [ Y t N P ]]]]]]]]]]]] A third class of structures is generated by extending the set of movable projections to include AgrY P and Agr X P. If these additional movements take place


63

in a “roll-up” fashion, i.e. movement of NP to [Spec, AgrY P], followed by movement of AgrY P to [Spec, Agr X P], followed by movement of Agr X P to [Spec, AgrW P], this will derive the mirror image of the base order ((1fiv)), as shown in (5). (5) [ Agrw P [ Agr X P [ AgrY P NP [ AgrY [ YP AP [ Y t N P ] ] ] ] [ Agr X [ XP NumP [ X t Ag r Y P ] ] ] ] [ Agrw [ WP DemP [ W t Ag r X P ] ] ] ] Partial roll-up movement derives the order (1biii) as above without the final step of Agr X P to [Spec, AgrW P] movement. Four more orders are derived by moving agreement phrases but leaving the NP in situ internally to the moving constituent; thus, AgrY P can move to [Spec, Agr X P] and from there to [Spec, AgrW ]P, giving rise to (1bii) and (1eiii) respectively. Agr X P can move to [Spec, AgrW P] which gives rise to (1dii), or to (1fiii) if combined with movement of AgrY P to [Spec, Agr X P]. Three further orders are derived by a combination of movement of agreement phrases and NP movement internal to those phrases. If NP moves to [Spec, AgrY P], AgrY P can surface either in [Spec, Agr X P] or [Spec, AgrW P]. The first of these is a partial roll-up structure discussed above, but the latter gives rise to the new order (1eiv). If Agr X P moves to [Spec, AgrW P], then NP can surface in either [Spec, Agr X P] or [Spec, AgrY P]; both derivations are new and give rise to the orders (1div) and (1diii) respectively. The final admissible derivation in Cinque’s system is one in which AgrY P moves to [Spec, Agr X P] and is subsequently stranded by NP movement to [Spec, AgrW P]. This derives (1biv). Cinque suggests that the order in (6) is possibly spurious, but we argue in section 4.4.3 that it does exist. (6) [ Agrw P NP [ Agrw [ WP DemP [ W [ Agr X P [ AgrY P t N P [ AgrY [ YP AP [ Y t N P ] ] ] ] [ Agr X [ XP NumP [ X t Ag r Y P ] ] ] ] ] ] ] ] This exhausts the orders that can be generated under Cinque’s theory. Consider why. It follows from the assumptions made that all material preceding the noun must come in the base order, essentially because all other orders violate the condition that moved constituents must contain the noun ((2b)). This subsumes the case of N-finality discussed above, but also includes (1cii), (1eii), (1fii). Finally, the orders (1ciii) and (1civ) are excluded because their derivation either requires movement of a non-constituent or, again, violates (2b). This is because any constituent that contains N and Num also contains A. Therefore, there is no way of placing Num and N in a position preceding Dem without also placing A in a position preceding Dem.

64

Formal features

4.3 The alternative The results described in the previous section certainly make an important contribution to our understanding of the syntax of the extended nominal projection. However, we do not think that they provide evidence for Kayne’s antisymmetry hypothesis. As we show in this section, the pattern of attested and unattested orders also falls out from the assumptions in (7). The first three of these are identical or equivalent to the first three assumptions made by Cinque. The fourth assumption replaces the LCA. It is weaker than Kayne’s hypothesis, because it limits the asymmetry of syntax to movement. (7) a. The underlying hierarchical order of Dem, Num, A, and N in the extended nominal projection is DemNumAN, where indicates c-command; b. all (relevant) movements move a subtree containing N; c. all movements target a c-commanding position; d. all (relevant) movements are to the left. If the LCA is abandoned in favor of (7d), we can base-generate eight of the 14 attested linear strings, simply by allowing cross-linguistic variation in the linearization of sister nodes in the hierarchical structure described by (7a). Seven of these orders are derived through movement in Cinque’s system. In our trees the non-terminals in the extended projection of the noun are unlabeled and the demonstrative, numeral, and adjective are not introduced by dedicated functional heads. This is because nothing in our argument hinges on the label of the nodes in the extended projection of the noun or the existence of dedicated functional heads hosting DEM, NUM, and A as specifiers. The reader is thus invited to resolve these issues in his or her favorite way. (8) a. b. c. d. e. f. g. h.

[ DEM [ NUM [ A N ] ] ] [ [ NUM [ A N ] ] DEM ] [ DEM [ [ A N ] NUM ] ] [ [ [ A N ] NUM ] DEM ] [ DEM [ NUM [ N A ] ] ] [ [ NUM [ N A ] ] DEM ] [ DEM [ [ N A ] NUM ] ] [ [ [ N A ] NUM ] DEM ]

The remaining six orders are derived by leftward movement of a constituent containing the noun:

Universal 20 without the LCA (9) a. b. c. d. e. f.

65

[ DEM [ N [ NUM [ A t N ] ] ] ] [ N [ DEM [ NUM [ A t N ] ] ] ] [ [ A N ] [ DEM [ NUM t[A N] ] ] ] [ [ N [ NUM [ A t N ] ] ] DEM ] [ N [ DEM [ [ t N A ] NUM ] ] ] [ [ N A ] [ DEM [ NUM t[N A] ] ] ]

There are other derivations involving movement, but these do not extend the set of linear strings that can be generated. For example, (1biii) can be basegenerated as above or derived on the basis of, for example, (1bii) by short movement of N as in (10). (10) [ Dem [ [ A N ] Num ] ] [ Dem [ [ N [A t N ] ] Num ] ] The impossibility of the ten unattested orders is explained in essentially the same way as in Cinque’s system. This is very clear in the case of noun-final structures. Since movement is uniformly leftward and must affect constituents containing the noun, noun-final orders must be base-generated. But among the base-generated structures, all of which are given in (8), only (8a) is nounfinal. Therefore, every other permutation of Dem, Num, and A is ruled out prenominally. In fact, this argumentation carries over to prenominal material generally. All material preceding the noun must be base-generated there, and its linear order must consequently reflect the hierarchy in (7a). Finally, the orders (1ciii) and (1civ) are excluded because their derivation either requires movement of a non-constituent or, again, violates (7b). This is because any constituent that contains N and Num also contains A. Therefore, there is no way of placing Num and N in a position preceding Dem without also placing A in a position preceding Dem. We have demonstrated that, in order to capture Cinque’s result, it is sufficient to assume that movement is uniformly leftward. The stronger assumption that projections are uniformly right-branching, the LCA, need not be made and does not restrict the typology. Another set of assumptions that play no role in restricting the typology involves the number and the nature of landing sites for movement, but we cannot discuss the issue here. 1 The remaining assumptions, however, are crucial. Dropping any of them rules in unattested word orders. Thus, relaxing (7a) would incorrectly allow prenominal material to appear in permuted orders 1 Formally, given the base hierarchy in (2a) or (7a), or (3), the set of 14 strings representing attested orders in (1) is closed under permutation by movement according to (2c), (2b), and (2d) or (7c), (7b), and (7d).

66

Formal features

as illustrated very simply in (11a). 2 The same problem arises if movement of constituents that do not contain N were allowed; the order (1di) can be derived either by separate movements of Num and A or by movement of N followed by remnant movement of the phrase containing Num, A, and the trace of the noun. This is illustrated in (11b). The c-command requirement on movement is well motivated on independent grounds; dropping it would wreak havoc on the typology, as illustrated in (11d) for (1di). (11)

a. unattested: Num A Dem N [ NUM [ A [ DEM N ] ] ] b. unattested: Num A Dem N [ NUM [ A [ DEM [ t NU M [ t A N ] ] ] ] ] or [ [ NUM [ A t N ] ] [ DEM [ N t[NU M[ AtN ]] ] ] ] c. unattested: N Num Dem A [ Dem [ Num [ A N ] ] ] [ Dem [ [ N Num ] [ A t N ] ] ] [ [ N Num ] [ Dem [ t[N Num] [ A t N ] ] ] ] d. unattested: Num A Dem N [ [ [ NUM [ A t N ] ] DEM ] N ]

So far, the theory we have tested has fewer restrictions than Cinque’s and therefore the derivations that it allows form a superset of the set of derivations allowed in Cinque’s approach. This is in fact the basis for the claim that Cinque’s findings ought not to be construed as an empirical argument for universal Spec≺Head≺Complement order, since the extra derivations allowed on the assumptions in (7) do not give rise to additional orders. Consequently, the assumed universal Spec≺Head≺Complement order does not carry any of the empirical burden. Cinque’s and our explanations rest entirely on the assumption of a universally fixed underlying hierarchy of elements in the extended projection of the noun and on restrictions on movement. The LCA does not contribute anything to the explanation of the linear asymmetry inherent in the universal 20 data. We will demonstrate below that this carries over to the explanation of other linear asymmetries as well. In this context we should note that our proposal requires fewer movements than the LCA-based alternative. This claim can be construed in two ways, 2 Incidentally, Brugè (2002) assumes that the structure in (11a) represents the underlying universal hierarchy and that demonstratives sometimes surface in this low underlying position. The discussion in the main text indicates that Brugè’s hierarchy is incompatible with the cross-linguistic record. Her suggestion, largely motivated by Spanish, where the article and demonstrative may co-occur on opposite sides of the noun, also fails to make sense of the observation by Rijkhoff (2002, chapter 6) that in all languages that allow demonstratives and articles to co-occur prenominally, the demonstrative always precedes the article.


67

only one of which is relevant. First, Cinque’s theory requires movement in 13 of the 14 licit derivations, while our alternative does so only in six. In each of those no more than a single movement is required, while Cinque’s derivations require up to three movements. This distinction is quantitative and cannot be used as a basis for an argument one way or another. Second, as we will demonstrate, the movement types required in Cinque’s theory form a superset of the movement types required in our theory. This is the crucial point. The fewer movement types there are, the more restrictive the theory of movement can be. As we will also demonstrate, this increased restrictiveness in the movement component gives rise to a more restrictive theory overall.

4.4 Comparing the two theories In the previous section we showed that the typology of word order in the nominal domain can be explained without appeal to the LCA. A weaker assumption barring rightward movement suffices. This means that at this point we have two largely equivalent theories that assign very different representations to the various linear realizations of the extended nominal projection. In this section we try to evaluate the two approaches. We begin by showing that the equivalence of the two theories is more dramatic than may seem to be the case at first sight. They assign very similar structures to each of the attested strings in (1). This, of course, further weakens the empirical content of the LCA. It should also dissuade any attempt at arguing against our theory on the basis of constituency; the theories are too similar to be distinguishable in terms of constituency at any level. We then go beyond the nominal domain, demonstrating that for each ungrammatical string derived by movement to the right, there is an LCAcompatible analysis. Consequently, proponents of antisymmetry will still need to make a stipulation banning apparent rightward movement (that is, structures that are the LCA-compatible equivalent of rightward movement). Finally, we show that there are well-motivated constraints on movement that can successfully be formulated in theories that adopt our more flexible structures, but not in antisymmetric theories, which are based on rigid trees and require additional movement operations in order to capture word-order alternations. 4.4.1 Stretching and shrinking trees The main claim of this subsection is that, despite appearances, for each attested word order in the extended nominal projection the gross constituency and command relations on Cinque’s analysis are identical to those in the

68

Formal features

simplest representation allowed under our alternative proposal. (By “gross constituency” we mean the grouping of all audible material in the base structure, as indicated by traces, and on the surface.) We demonstrate this by giving a mechanical procedure to change Cinquean trees into the trees in (8) and (9), as well as a mechanical procedure to do the reverse. We call the first of these procedures shrinking and the second stretching. Shrinking is defined in (12), (15), and (17). The point is the mere existence of such structure-preserving algorithms, not the details of execution. (12) Prune the Cinquean tree by deleting the functional heads (W, AgrW , X, Agr X , Y, and AgrY ) and their intermediate projections maintaining dominance relations. 3 When applied to the tree in (13) this yields (14). (13)

[ Agr F· P X [ Agr F· P Agr0F· [ F·P · [ F·P F0· t X ] ] ] ]

(14) [ Agr F· P X [ F·P · t X ] ] (15)

Delete any trace whose antecedent is the sister of the trace’s mother.

This will transform (14) into (16). (16) [ Agr F· P X [ F·P · ] ] (17) Prune all non-branching non-terminals maintaining dominance. The final step produces (18). (18) [ Agr F· P X · ] It is easy to see that this three-step procedure takes us from the representation in (3) to one that is isomorphic to (8a). It is equally obvious that shrinking will yield a representation isomorphic to (8h) when applied to (5). We will leave it to the reader to check that this procedure works for all admissible derivations in Cinque’s system. (It does.) Here we will only illustrate this result by going through the effects of shrinking in one of the more complicated cases. When the rule in (12) is applied to (6) it yields (19). (19) [ Agrw P NP [ WP DemP [ Agr X P [ AgrY P t N P [ YP AP t N P ] ] [ XP NumP t Ag r Y P ] ] ] ]

3 Starke (2004) argues on independent grounds that functional heads whose only purpose is to host specifiers of a particular type should be jettisoned from the the theory. Since Cinque provides no morphological motivation for any of the proposed functional heads, our algorithm treats them uniformly as empty. See also Koopman’s (1996) generalized doubly filled COMP filter.


69

Applying (15) will delete the traces of NP and AgrY P as in (20a), which can be pruned to (20b), a structure equivalent to our (9e). (20) a. [ Agrw P NP [ WP DemP [ Agr X P [ AgrY P t N P [ YP AP ] ] [ XP NumP ] ] ] ] b. [ Agrw P NP [ WP DemP [ Agr X P [ AgrY P t N P AP ] NumP ] ] ] Of course, the labeling in the representations resulting from shrinking does not adhere to standard requirements. This does not affect the point under discussion, however, since we are interested strictly in properties of tree geometry here. What is important, though, is that the nodes that make up the extended nominal projection bear labels revealing that. This information is indeed maintained under shrinking. It should be clear that shrinking preserves gross constituency. Material shared in the two types of analysis (Dem, Num, A, N, and traces of “long” movement) is grouped in the same way before and after shrinking. For example, the representation in (6) and the shrunken version of it in (20b) are both characterized by the following bracketing: [N P [DemP [[tN P AP ]NumP ]]]. Shrinking also preserves c-command relations in the sense that any ccommand relation that holds at some point during the derivation in the Cinquean analysis also holds at some point in the derivation in our analysis. This is trivial for the c-command relations determined by the functional sequence, given that the functional sequence determines height of attachment in base-generated structures. The movements that remain in the shrunken trees are inherited from the Cinquean input and so the extra c-command relations they give rise to are also present in LCA-based representations. In order to guarantee full preservation of c-command relations among material shared by both analyses, however, it must also be the case that movements that do not survive shrinking do not give rise to new c-command relations. This is indeed true, as the movements that shrinking eliminates are the roll-up ones. A look at (5) reveals that NP, AP, NumP, and DemP are properly contained in the moving constituent. The moving constituent itself, of course, acquires new c-command relations but its proper parts do not. 4 It is true, of course, that c-command relations are not entirely equivalent. Thus, in (6) NumP c-commands AP before movement of AgrY P but not after movement, whereas in our structure NUM c-commands A throughout the derivation. However, we know of no convincing analysis that crucially invokes the loss of c-command through roll-up movement (though see Kayne, 1994; Cinque, 2006; and footnote 6 for discussion). For movement that is not of 4 According to the letter of Kayne’s (1994) definition of c-command this is false since specifiers of specifiers c-command out. We cannot pursue the consequences of this here.

70

Formal features

the roll-up type, such effects do seem to exist, however (witness the failure of reconstruction for various phenomena, the possibility to bind anaphors under successive cyclic wh-movement, etc.). To complete our argument, we present a partial method for stretching trees. It is essentially the reverse of the three operations that constitute shrinking and is formulated in (21). The procedure is only partial because it is designed to stretch trees with right specifiers or adjuncts, as this is the geometry that the LCA prohibits. In order to develop a complete method of stretching trees one would have to add a procedure that enriches structures with left specifers and adjuncts. In the interest of space, we refrain from doing so here. (21) In a structure [ Y X · ] where (i) · is a non-projecting node, (ii) Y is projected from X, and (iii) · belongs to a class mentioned in (7a) a. insert a node F· P between · and its mother; b. insert a trace of X under F· P and to ·’s right; c. relabel Y as Agr F· P. d. For every headless node ‚, insert an identically labeled node „ between ‚ and ‚’s right daughter and the appropriate head for „ as „’s left daughter. Like shrinking, stretching preserves gross constituency and c-command relations between Dem, Num, A, and N, as well as traces of long movement. What the procedures of stretching and shrinking demonstrate is that the LCA-based analysis of the typology of noun phrases is in fact very similar to the symmetrical analysis proposed here. Not only does it give rise to the same typological predictions qua word order (weak generative capacity) but it also generates very similar trees for those word orders (strong generative capacity). This does not mean, of course, that the theories are identical. In fact, we will show below that stretching trees is not innocent. However, we can already conclude at this point that the original motivation for the LCA is undermined since it does not impose additional restrictions on tree geometry. For every shrunken tree that violates the LCA, there is a stretched variant that is LCA-compatible. The LCA would, of course, restrict tree geometry if there were fewer functional projections that could host movements and if there were fewer admissible patterns of movement (e.g. if roll-up movement were to be disallowed). The problem that this would give rise to, however, is


71

that the typology of noun phrases captured by Cinque’s system would then be beyond the reach of the theory. 4.4.2 Rightward movement and the LCA We have claimed above that the LCA does not restrict the class of possible tree shapes. This entails, in particular, that it does not rule out structures in which a trace precedes its antecedent. We demonstrate this by taking a traditional rightward movement structure and applying the procedure of stretching defined above (modulo labels of inserted functional projections). This yields an LCA-compatible correspondent. (22) Rightward Movement (gap-filler orders): XP ... t· ...

·

F2 P XP ... t· ...

F2

F1 P ·

F1

tX P

The trace and its antecedent can be arbitrarily far apart (in both linear and hierarchical terms), because nothing in the theory restricts the depth to which t· is embedded in XP. For example, · could have undergone a number of steps of successive-cyclic movement within XP. The fact that Cinque’s analysis rules out this type of movement in the nominal domain has nothing to do with LCA. Rather it follows from a restriction on movement, the assumption that movement within the extended nominal projection must always target a constituent containing the lexical head. This rules out the occurrence of structure (22) in the nominal realm, since the two moving elements, XP and ·, cannot both contain N. If · contains N, then XP doesn’t, and vice versa. The requirement that every movement pied-pipe the lexical head does not seem to have a counterpart in the extended projections of other lexical categories. 5 In fact, the structure in (22) is not just a hypothetical possibility, it is a widely used analytical tool known as remnant movement (den Besten and Webelhuth, 1987; Müller, 1998; Koopman and Szabolcsi, 2000; Nilsen, 2003) and references cited there). A particularly striking example comes from Kayne (1998), who argues that negative quantifiers in English raise to [Spec, NegP], a movement that is followed by remnant movement of VP to the specifier of an 5 Indeed, even in the NP the restriction holds only if we abstract away from optional movements, movements of arguments of the noun, etc. Given these simplifications, the restriction might actually carry over to other lexical categories.

72

Formal features

as yet unidentified functional projection WP. This derivation instantiates the possibility raised above. 6 (23) [ WP [ VP . . . t Neg D P ] [ W0 [ NegP NegDP [ Neg0 tV P ] ] ] ] In sum, every tree-geometric shape, including rightward-movement structures, have an LCA-compatible counterpart, not just hypothetically but also in analytical practice. Therefore, a number of cross-linguistic generalizations that have been used to motivate the LCA do not follow from the theory, unless a way can be found to block remnant movement in the relevant cases. 7 For example, it has been observed that wh-movement and long NP raising are universally leftward (Bresnan, 1970; Perlmutter and Postal, 1983/1972). 8 These generalizations have been used to motivate the LCA (e.g. Cinque, 1996), but do not in fact follow from it, as just shown. Similarly, Kayne (1994: 50) observes that while there are languages with verb-second, languages with the verb systematically in penultimate position do not seem to exist. This generalization extends to other second-position phenomena such as secondposition cliticization. Kayne argues that this asymmetry is a result of the LCA, as it implies that heads precede their complement and specifiers are unique and precede heads; therefore, both head movement and phrasal movement must be leftward. Verb-second results if the highest functional projection is targeted by both. But in fact there is an LCA-compatible derivation that results in the offensive pattern, as shown in (24). The derivation requires a sequence of two functional heads both of which attract the head of their respective complement. In addition, the lower functional head (F01 ) attracts some maximal projection (YP) out of its complement, as is independently required for verb-second. Finally, the higher head attracts the complement of F01 , as required for roll-up structures. 9 Although we have not provided a 6 Movement of NegDP to [Spec, NegP] is intended to overtly mark scope. This implies that the subsequent movement of VP should not erase c-command relations. Recall that this assumption was part of the argument we made for saying that c-command relations are preserved under shrinking. See Cinque (2006) for an account where underlying c-command relations are, in effect, destroyed by movement just to be recreated later, while binding principle A is computed somewhere along the way. 7 Remnant movement cannot be ruled out across the board, as it is required for the analysis of structures like the following:

(i)

a. Painted by Picasso, this portrait doesn’t seem to be. b. How likely is there to be a riot?

8 Signed languages are sometimes cited as counterexamples to the claim that wh-movement is universally leftward (for diverging points of view see Neidle et al., 1997, 1998; Petronio and Lillo-Martin, 1997). 9 Remnant movement of headless phrases has been motivated for German by Müller (1998) and for Japanese by Koizumi (1995) and Vermeulen (2005), contra Takano (2000). Should it turn out to be


73

procedure for shrinking trees like (24), where the functional head positions contain moved material, we suspect that removal of structure required only for LCA compatibility yields a structure with head and phrasal movement to the right. (24) [ F2 P [ XP . . . t X 0 . . . tY P . . . ] [ X 0 + F 10 + F 20 [ F1 P YP [ t X 0 +F 10 t X P ] ] ] ] What these observations show is that developments in the field subsequent to the introduction of the LCA in Kayne (1994) have led to a situation in which LCA-based theories fail to capture the data that motivated the LCA in the first place. In other words, the LCA provides no insights into the type of generalizations quoted above. 10 4.4.3 Is stretching harmless? So far we have shown that there is no particular advantage in adopting the LCA over the more traditional alternative advocated here. There is an important disadvantage, however, that convinces us that the LCA should be abandoned. This disadvantage is that the movements required to reconcile the LCA with the attested word-order patterns stand in the way of arriving at a restrictive theory of movement. The general problem manifests itself in at least two ways, each one associated with a type of movement required by Cinque’s analysis. The two movements in question are the very local movement that generates roll-up structures and the movement of NP in (6), where pied-piped material is stranded in an intermediate position. Very local movement is problematic in the light of Saito and Murasugi (1993); Bošković (1997); Abels (2003a,b); Boeckx (2008). In those works an anti-locality condition on movement is proposed according to which no complement can recombine through movement with a projection of its selecting head. 11 Abels (2003b) argues that this condition has the following rationale: A head and its complement are in a local relation in the base structure (they mutually c-command each other) and no different relation is established by recombining the complement with a projection of the head.

the case that head movement bleeds remnant movement, then the argument can be reconstructed by simulating head movement through remnant movement a` la Koopman and Szabolcsi (2000) or Nilsen (2003). 10 Kayne (1994: 140 fn. 8) acknowledges that even with the more restrictive theory of movement assumed at the time, the LCA doesn’t by itself rule out the analogue of rightward V-to-C. “It is also essential to rule out derivations involving leftward movement of the finite verb to C0 followed by leftward movement only in root contexts of IP to Spec,CP”. 11 Grohmann (2000) suggests an even more radical anti-locality constraint.

74

Formal features

For any category that only permits extraction through an escape hatch, the prediction is that the complement of that category cannot be extracted. There is good evidence that extraction from CP must proceed through just such an escape hatch. It is therefore predicted, correctly, that IP will resist movement when generated as the complement of C0 (den Dikken 1995 calls this the IP Immobility Principle). This pattern is striking, since extraction out of IP is possible. Both facts are illustrated in (25). (25) a. What do you think that Mary has read? b. Nobody thought that anything would happen. c. That anything would happen, nobody thought. d. ∗ Anything would happen, nobody thought that. The general pattern extends to a number of other categories; thus, preposition stranding is blocked in languages where movement out of PP needs to proceed through an escape hatch while movement out of the complement of PP is unproblematic. Similarly, VP can never strand v, although extraction out of VP is, of course, allowed. Beyond these cases, which are discussed at length in Abels (2003b), patterns of extraction that parallel the data in (25) are found with several other categories. This would follow if these other categories also require extraction through an escape hatch. Two of these can be found in English. Extraction of NP stranding the determiner is ungrammatical but extraction from NP is unproblematic. Furthermore, it can be shown that English has two types of degree expression, one of which is a functional head selecting AP and the other a modifier that adjoins to AP as well as to other categories (see Neeleman et al., 2004). Extraction of AP stranding modifying degree expression is possible, but similar movement stranding degree expressions that are functional heads is ruled out. As before, extraction out of AP is fine in both cases. The structures to be ruled out, then, are given in (26a–e). (26) a. ∗[ CP IP [ C0 t I P ]] b. ∗[ PP DP [ P0 t D P ]] c. ∗[ vP VP [ v0 tV P ]] d. ∗[ DP NP [ D0 t N P ]] e. ∗[ DegP AP [ Deg0 t AP ]] f. [ XP · [ X0 [ YP . . . t· . . . ]]] where X ranges over C, P, v, D, and Deg. The structures discussed so far are all head-initial, but the same patterns are found with complements that precede the head. Thus, the IP Immobility Principle applies to Japanese, Korean, and Turkish as well as to English, and


75

the ban on preposition stranding is as common with postpositions as it is with prepositions. Needless to say, extraction from IP and from the complement of postpositions is unproblematic (see, for example, Sener (2006) for an illustration based on Turkish postpositions). The set of structures to be ruled out should therefore be extended to (27a–e). (27) a. ∗[ CP IP [ t I P C0 ]] b. ∗[ PP DP [ t D P P0 ]] c. ∗[ vP VP [ tV P v0 ]] d. ∗[ DP NP [ t N P D0 ]] e. ∗[ DegP AP [ t AP Deg0 ]] f. [ XP · [ [ YP . . . t· . . . ] X0 ]] where X ranges over C, P, v, D, and Deg. As far as we can tell, the anti-locality constraint on movement must remain toothless in theories that assume syntax to be antisymmetric. 12 To account for head finality, LCA-based theories can adopt one of the following two structures. The problem with adopting (28a) (suggested in Kayne (1994)) is that it violates the anti-locality condition. However, giving up the anti-locality condition implies that the immobility of IP can only be stipulated. If, on the other hand, the anti-locality thesis holds, then (28a) must be rejected in favor of (28b). This implies that the escape hatch for extraction from the CP domain cannot be [Spec, CP] but must be [Spec, AgrC P], which in turn has the unfortunate consequence that the account of IP’s immobility is lost. Thus, either the anti-locality constraint must be abandoned, or it must be voided of its empirical content. 13 (28) a. [ CP IP [ C0 t I P ] ] b. [ AgrC P IP [ AgrC0 [ CP C0 t I P ] ] ] Essentially, the same problem arises in the case of unstrandable postpositions. 12 Thus, Kayne (2005: 272, 331) assumes that the “complement of a given head H can never move to the Spec of H”, but given the proliferation of silent functional heads we do not see what empirical predictions follow from this assumption. 13 Proponents of LCA-based theories face an additional question in this area. Kayne (1994) cites lack of obligatory wh-movement in complementizer-final languages as possible evidence for (28a). However, if (28b) is adopted, there is an additional potential position following IP and preceding C0 . If that position were used for wh-movement, head-final languages would have rightward wh-movement. Thus, the question must be answered why [Spec, CP] is systematically empty. In Kayne (1999, 2004), (see Borsley (2001) for discussion) a different account of head-finality is proposed. Under this proposal certain prepositions and complementizers are merged in a VP-external position and combine with their apparent complements through movement. Along the way, a number of remnant movements occur. We leave it to proponents of such analyses to demonstrate how the IP Immobility Principle, the ban against P-stranding, and the word order typology in the NP can be made to follow. The proposals are not sufficiently worked out to allow general evaluation yet.

76

Formal features

We now turn to a second restriction on movement that cannot be reconciled with the derivations proposed in Cinque’s paper. In general the movements that derive neutral orders are assumed to target the noun. This explains why only movements of constituents containing the noun are admissible. Such movements can be construed as instances of pied-piping. Cinque is actually quite explicit about this; indeed, it is hard to see what else could explain the limitation to subtrees containing the noun. With this in mind, consider again the structure in (6). On the proposed analysis this must be a derivation in which material pied-piped in an initial step of movement (of AgrY P to [Spec, Agr X P]) is stranded by a subsequent one (movement of NP to [Spec, AgrW P]). The problem is that such derivations seem to be systematically ruled out in other domains. Thus, Postal (1972) observed that prepositions piedpiped under wh-movement cannot be stranded in intermediate positions, as shown in (29). 14 Movement under relative clause formation is subject to the same restriction, as (30) illustrates. (29) a. [ P P With which friend] did you say t P P that she went home t P P ? b. [ D P Which friend] did you say t D P that she went home with t D P ? ∗ c. [ D P Which friend] did you say [PP with t D P ] that she went home tP P ? (30) a. the rock [DP pictures of which] I think t D P that Bill has seen t D P b. the rock [PP of which] I think t P P that Bill has seen pictures t P P c. ∗ the rock [PP of which] I think [DP pictures t P P ] that Bill has seen t D P At the very least, the derivation Cinque assumes complicates the generalization that pied-piped material cannot be stranded. Therefore, it may well make it harder to develop an explanation of the relevant data. In contrast, the more conservative analysis of the N-Dem-A-Num word order advocated here ((1biv)) does not rely on stranding and therefore does not give rise to the same complication. This is a second example, then, where tree-stretching is potentially harmful. 15 14 Du Plessis (1977) claims that such derivations do exist in Afrikaans, but the analysis is dubious according to Den Besten (p.c.) who reanalyzes the relevant data as involving parentheticals. Den Besten suggests that Du Plessis’ analysis cannot account for the verb placement in the examples involving putative intermediate stranding. 15 There is one class of analyses according to which quantifier float may give rise to stranding of pied-piped material. Sportiche (1988), in particular, argues that quantifier and DP are generated as a constituent, and that the quantifier can be stranded, not only in its base position but in any Aposition through which the DP passes. However, there are several alternative accounts of quantifier float. Bošković’s (2004) proposal comes very close to Sportiche’s without allowing stranding of piedpiped material. Other authors have argued that floating quantifiers are base-generated as adverbs,


77

Cinque suggests that the crucial order (N-Dem-A-Num) could be spurious. Since our argument rests on its existence, we should take a closer look at the languages that display it. Cinque mentions Pitjantjatjara, Nkore-Kiga, and Noni. Noni has the relevant order as an alternate to N-Dem-Num-A, which suggests that we should put it to one side. For Pitjantjatjara, Bowe (1990: 29–54, 111, 146–50) claims that the order in question is the only admissible one. Indeed, in Eckert and Hudson’s 1988 textbook, examples like those in (31) can be found. The morpheme glossed SubjT is an ergative case marker appearing at the end of a subject DP and repeated in case of apposition, as shown in (31c). Therefore, material to the left of this morpheme can safely be taken to belong to a single extended nominal projection. The examples in (31a) and (31b) establish the sub-orders N-Dem-A and N-A-Num respectively. An example of the N-Dem-A-Num can be found in (31c), on the reasonable assumption that many behaves like a numeral. (Eckert and Hudson (p. 130–4) treat numerals and quantifiers as adjectives of quantity and give a single rule for positioning them among the nominal modifiers.) 16 (31)

a. Tjitji pala tjukutjuku -ngku -ni ungu child that small -SubjT -me gave ‘That small child there gave (it) to me.’ (Eckert and Hudson, 1988: 89) b. Kulata wara kutjara nyara mantjila! spear long two yonder get ‘Get the two long spears over there!’ (Eckert and Hudson, 1988: 132) pulka tjuta -ngku c. Tjitji tjuta -ngku katingu, tjitji panya child that.known big many -SubjT child many -SubjT took ‘The children took it, you know those big children.’ (Eckert and Hudson, 1988: 139)

Nkore-Kiga has been studied even less than Pitjantjatjara. There appears to be only one source, Taylor (1985), to which all claims about the language can be traced. 17 Taylor (p. 55) characterizes word order in the Nkore-Kiga noun phrase as follows (we omit categories not relevant to the present discussion): rather than as part of the associated DP (see, for instance, Bobaljik 1995 and Janke and Neeleman 2005). 16 The morphemes analyzed as demonstratives can stand alone. They are therefore not clitics and hence not subject to special ordering restrictions. 17 Cinque bases his claims about Nkore-Kiga on Dryer (2007) and Lu (1998). Dryer (2007) cites Taylor (1985) as his source and Lu (1998) cites Rijkhoff (1990), who in turn cites Taylor (1985).

78

Formal features

(32) noun—demonstrative adjectives—pure adjectives/appositives—quantifiers—verbal adjectives Some of the orders mentioned in (32) are illustrated in the following example. (We have adjusted Taylor’s gloss slightly.) (33) ekitabo kyawe ekyo eki- hango ekimwe eki- rikutukura book your that 7- large one 7- partic.contin- be.red ekiri aha meeza nikyo which.is on table itself ‘that selfsame single large red book of yours on the table’ Taylor’s (p. 55) discussion of the order in the noun phrase clearly suggests that the order in (32) is the neutral order. It is not “rigidly adhered to”, but it is “preferred”. Taylor (p. 75) further explains that for pure adjectives the alternate N-Dem-Num-A serves to emphasize these elements. This order is therefore marked, and hence irrelevant to our concerns. In view of these data, we think that discarding N-Dem-A-Num is not justified. The predictions of our theory (assuming that pied-pipers never strand pied-pipees) diverge from those of Cinque’s theory once more than four elements are taken into consideration. Consider a structure in which there are five hierarchically ordered elements. Cinque’s theory would then allow the order in (34), but our theory would rule it out, as every derivation consistent with the assumptions we make would require stranding of a pied-pipee (namely “4” in (34)). (34) [ 5 [ 1 [ [ 4 t5 ] [ 2 [ 3 t[4 5] ] ] ] ] ] We do not think that such orders exist, but the issue is worth exploring in some depth as this provides an opportunity to test the theories empirically. Notice that these diverging predictions underline the fact that our theory is overall more restrictive than its LCA-based competitor.

4.5 Concluding remarks Two main conclusions can be drawn from the discussion in this chapter. First, the claim that base-generated structures are anti-symmetric, as stated in Kayne’s 1994 Linear Correspondence Axiom, is empirically vacuous, at least within the nominal domain. The structures allowed by Cinque’s 2005a LCAbased theory are identical in gross constituency to those generated by our


79

more conservative alternative. Although we cannot demonstrate this here, we believe that this conclusion holds more generally. Second, in order to capture the typological patterns uncovered by Cinque, (certain types of) movement must be exclusively leftward. Although a ban on rightward movement was originally argued to follow from the LCA, we have shown that this is not true, except in the most legalistic sense. Every rightward-movement structure can be paired with an LCA-compatible remnant-movement structure that shares its gross constituency. These conclusions lead us to reject the LCA, especially in view of evidence that the LCA stands in the way of a restrictive theory of movement. However, whether we reject the LCA or not, the question presents itself why movement in the nominal domain should be leftward. In work in progress (Abels and Neeleman, 2006), we argue that a parsing explanation might be available on fairly uncontroversial assumptions.

5 What it means (not) to know (number) agreement∗ CARSON T. SCHÜTZE

5.1 Introduction My goal in this chapter is to consider from a theoretical standpoint what it could mean for a child acquiring some language to not (yet) “know” or “have” agreement, and then examine a number of sets of data that may be instances of this. I begin by making explicit my terminological and theoretical assumptions about agreement. I then enumerate the places in this system where children could in principle be different from adults. Next I put forward a specific model that instantiates (at least) one such difference and shows how it accounts for child data from a number of languages. I conclude with some open issues and questions for future research.

5.2 Theoretical background I take expressions of cardinality or numerosity {1, 2, 3 . . . } to be those that specify the number of members in a set; these are distinct from expressions of (grammatical) number {singular, dual, trial, paucal, plural}, which specify the semantic type of what they modify, that is, an atomic individual versus a plural individual, a set with one member versus more than one member (in some languages, {exactly two/exactly three/a small number of/more than a small number of} members). I assume every referential Determiner Phrase (DP) (but maybe not every nominal predicate) has an “interpretable” number specification (in roughly ∗ This chapter was originally presented as a paper at the GLOW workshop. I would like to thank the workshop organizers, Anna Gavarró and Maria Teresa Guasti, for the opportunity to present this work at GLOW, and the participants for their feedback. This chapter has benefited from comments by the volume editors and an anonymous reviewer. This work was supported by a grant from the UCLA Academic Senate Council on Research.

What it means (not) to know (number) agreement

81

the sense of Chomsky 1995). 1 Less obviously, I assume that each instance of number that occurs in the interpretation of a sentence (at Logical Form or perhaps beyond) corresponds to one and only one number feature in the syntactic representation. Any other number markings are uninterpretable, therefore must arise by checking or copying an interpretable number feature, or by default. (I do not discuss theories based on Probe–Goal relationships and the Agree operation.) There are two broad classes of such uninterpretable number markings (the same applies to gender, case, and perhaps even person markings): (1) those internal to a DP (or whatever category represents the top of the extended nominal projection) whose “head” they agree with, which we call concord; (2) those outside a DP, typically on the extended projection of a predicate (agreeing with a DP that it takes as argument), which we call (index) agreement (cf. Wechsler and Zlatić 2000, 2003). A distinction needs to be made between morphophonological expressions of a number feature (be it interpretable or uninterpretable), versus a morphophonological change in some other morpheme that is triggered by a number feature, that is, number features as content versus as context. For example, the stem suppletion rule in (1a) is not an instance of number marking, although there could be situations, perhaps even across an entire language, where the only audible manifestation of distinctions of number meaning are of this type (the forms of the language would reflect number distinctions but would not be overtly marked for number). By contrast, (1b) describes how number is actually expressed on this stem (by a null suffix) and on regular stems: this is a rule of vocabulary insertion or exponence. (1)

a. person → people / __+pl b. pl ↔ Ø / people+__, . . . , pl ↔ s elsewhere

This is consistent with central tenets of Distributed Morphology (Halle and Marantz 1993): morphemes are pieces, not processes, and syntax is fully specified; only vocabulary items can be underspecified. Thus, even if a noun is never marked for number (a particular noun, as in (2), or all nouns of the language, as in (3)), this strong theoretical stance requires a number feature to be part of the DP nonetheless. For Amele we would posit a single vocabulary entry under the Number head, informally num ↔ Ø, underspecified for the value of Number. The number inflections on the verb in (3) must be uninterpretable because they reflect nominal, not verbal, meaning differences. Syntactically, 1 Apparently predicates (event-denoting expressions) can also have interpretable number inflections, i.e. for a single event vs. a set of multiple events, as in Rapanui (Oceanic, Easter Island), where ruku means ‘(to) dive’ and ruku ruku means ‘(to) go diving’ (involving multiple dives, though possibly just one diver). In what follows I ignore this rare possibility.

82

Formal features

(3a) actually has a subject marked singular, (3b) has one marked dual, and (3c) has one marked plural. (2) a. The sheep drinks. b. The sheep drink.

(DP is specified [−pl]) (DP is specified [+pl], number feature is spelled out by -Ø)

(3) Amele (Papuan) 2 a. Dana ho-i-a man come-3sg-pst ‘The man came.’ b. Dana ho-si-a man come-3du-pst ‘The two men came.’ c. Dana ho-ig-a man come-3pl-pst ‘The men came.’ I assume mass nouns permit no specification for Number (unless coerced into count interpretations), hence they cannot actually trigger any number agreement, so they will occur with default index agreement (a feature filled in by the Spell-Out component). In some languages one can even use count nouns while leaving out number marking, being noncommittal about number (hence, on my assumptions, lacking an interpretable feature specification for DP number altogether): (4) Bayso (Cushitic, Lake Abaya, Ethiopia) a. lúban foofe lion watched.1sg ‘I watched lion.’ (one or more than one) b. lubán-titi foofe lion-sg watched.1sg ‘I watched a lion.’ (usually specific) c. luban-jaa foofe lion-pauc watched.1sg ‘I watched a few lions.’ d. luban-jool foofe lion-pl watched.1sg ‘I watched a lot of lions.’ 2 Unless otherwise noted, all examples and facts from non-European languages are taken from Corbett (2000).


83

5.3 What is there (not) to know? When we see children producing non-adultlike utterances with respect to number marking, or agreement more generally, what sorts of things in principle could be different in their minds from those of adults that might explain this? We should distinguish (at least) six possibilities for where a difference may lie: (A) the underlying conceptual knowledge or perceptual information they are trying to encode, i.e. something “prior” to the language system (if a child does not realize that what she is pointing at is actually two objects as opposed to one, we cannot reasonably expect her to use plural marking when referring to it); (B) the knowledge of, access to, or compliance with universal principles; (C) the knowledge of their language’s parametric choices on matters such as i. which categories are marked for number—be it interpretable, concord, or agreement (e.g. do adjectives show number contrasts?); ii. which elements agree or show concord with what (e.g. do V and/or P agree with their complement?); iii. which structural configurations trigger agreement (e.g. Baker (2008) proposes that some languages restrict the agreement trigger to be structurally higher than the head on which the uninterpretable agreement marking appears, while others allow it also to be lower, c-commanded by that head); iv. how many number values the language contrasts (e.g. does it mark dual distinctly from plural?); 3 v. what agreement is sensitive to—pure morphosyntax or also semantics (e.g. some languages, including some varieties of English, allow grammatically singular nouns to trigger plural agreement when they refer to collectives, as in The committee are meeting now, whereas other languages, e.g. German, never allow this); vi. how different S-structure configurations have their features spelled out (e.g. in many languages subject agreement is reduced, optional, or impossible when the subject is in a low structural

3

The greatest number attested seems to be five, e.g. in Sursurunga (New Ireland, Melanesian).

84

Formal features position such as below the inflecting predicate, but richer and/or obligatory when the subject is higher). 4 (D) the knowledge of the particular morphophonological items used to express particular features in a particular context (typically thought of as involving items of type (1b), although non-adultlike knowledge of items like (1a) could also give the appearance of number “errors”); (E) the knowledge of, or ability to carry out, relevant grammatical computations; (F) production abilities (at various levels).

Which of these possibilities actually arise with respect to (number) agreement? Although I have not exhaustively searched the acquisition literature, I am not aware of any cases of type (A). In the next section I will suggest a possible instance of type (B). I am not aware of any cases of types (Ci–v), but Guasti and Rizzi (2002) have proposed that some of the child English data discussed below reflect children’s uncertainty on point (Cvi). (Although they do not actually characterize it as a point of parametric variation, everything they say is consistent with this interpretation.) The literature suggests that errors due to point (D) are attested, at least with respect to interpretable number marking on nouns (and stem allomorphy triggered by such number features), for example in child German (e.g. Clahsen, Rothweiler, and Woest 1992). In the next section I will suggest a possible instance of (E), and allude to others not involving agreement. Finally, errors attributable to (F), for example articulatory limitations, are well documented (Demuth, Song, and Sundara 2007), but children are usually screened for such difficulties and excluded when one is looking for evidence of the other kinds of divergence from adult knowledge or abilities. In laying out the possibilities in (A)–(F) above I am assuming that the elements invoked there in fact constitute the mental machinery responsible for children’s (and adults’) production (and comprehension) of agreement. Different hypotheses would arise under the assumption that children at the relevant ages are not using anything resembling the agreement principles of generative grammar but instead are inducing constructions on the basis of item-specific knowledge derived from input (Wilson (2003) and work cited there). That approach would seem to predict that errors should occur only in the case of insufficient or distorted input, and that they should follow prevailing patterns in that input. Space restrictions do not permit addressing 4 From this summary it might appear that (Ciii) and (Cvi) are referring to different settings of a single-multivalued parameter. I have kept them separate in part because the sources I take these proposals from have very different conceptions of the syntax involved.


85

these predictions here, but the data in §5.5 do not strike me as particularly promising for this perspective.

5.4 The ATOM Model 5.4.1 Omission of agreement (and case) The model to be outlined in this section depends on some further details of minimalist syntax, most closely following chapter 4 of Chomsky (1995). Because interpretable features must by definition survive to Logical Form, although they need never be checked, they can be checked any number of times, and can hence trigger the deletion of several sets of matching uninterpretable features. As a result, unlike in earlier minimalism where one-toone checking constrained the distribution of uninterpretable features, in this newer version uninterpretable features are not needed for convergence. Take your favorite convergent derivation, remove all the uninterpretable features, and you are guaranteed to get another convergent derivation, albeit possibly with a different word order. 5 You could also remove just a subset of the uninterpretable features, as long as you do not leave any with no way to be checked. In particular, case features are uninterpretable on both DP and the corresponding head (canonically T or V), so you cannot leave the case feature off one member of a case-checking pair unless you also leave it off the other. How, then, do we ever ensure that uninterpretable features get into our derivations? By pure stipulation: “UG requires that there is always some choice of Case, phi-features . . . Case and phi-features are added arbitrarily as a noun is selected for the numeration” Chomsky (1995: 236). (The case part of this stipulation effectively implements the Case Filter. Among the phi-features, only gender is uninterpretable on DP; the rest of the stipulation involves an arbitrary choice about meaning, which I would rather not assume.) More generally, as discussed in detail in Schütze (1997), some additional principle beyond the need to converge is required in order to ensure that uninterpretable features become part of a sentence. In that work I proposed that for index agreement this is a principle that operates on the syntax, taking 5 Detailed exploration of the possible word order consequences of uninterpretable feature omission is beyond the scope of this chapter. While it is well known that the positions of finite versus nonfinite verbs in many root infinitive (child) languages are demonstrably different, evidence for distinct positions of case-marked versus non-case-marked DPs has been scarce and controversial at best. Moreover, empirical and theoretical considerations have led most minimalist syntacticians to believe that the features driving overt movement should not be tied to inflectional morphology, instead being treated as “generalized EPP features”. My suspicion is that these are not omissible in the same way that the features I discuss in the main text are, possibly because they are not truly uninterpretable, but I cannot defend these speculations here.

86

Formal features

advantage of Chomsky’s proposal that the set of admissible derivations is a subset of the set of convergent derivations (see below). For concord I proposed that the relevant principle applies in the Spell-Out component; I shall have no more to say about this. Concerning index agreement, I argued in Schütze (1997) that syntax needs to have a relativized or violable, not absolute, requirement on the insertion of uninterpretable features, which says: insert these features in a syntactic representation wherever you can, i.e. to the maximal extent possible given other constraints. I dubbed this requirement the Accord Maximization Principle (AMP), and proposed it as a universal. Subsequently, Chomsky has noted the need for something along these lines: “A natural principle, which has been suggested in various forms . . . Maximize matching effects” (2001: 14). This amounts to saying roughly that, among a set of convergent derivations that differ only in their uninterpretable feature content, the admissible one will be the one that contains the greatest number of uninterpretable features. My central claim about child grammar (at the relevant stage) is that children do not differ from adults concerning the convergent derivations, but they do differ on which convergent derivations turn out to be admissible (or at least admitted). Specifically, they do not (always) choose the optimal derivation (in particular, that one with maximal insertion of uninterpretable features) for a given configuration of interpretable features. (Any derivation they choose is subject to the same Spell-Out procedures.) We can then reduce the question of what is different about children to the question of why they are using these suboptimal structures that adults cannot use. Possible answers in principle could include maturation of a constraint such as the AMP (an answer of type (B) from §5.3), or inability to carry out the computations that AMP requires (an answer of type (E)). A resource-based problem is plausible because implementing the AMP requires, at least conceptually, a comparison among different candidate representations: i.e. version A of a sentence is bad if there is a well-formed version B that contains more phi- and case features than A. It has been suggested in several domains that children have difficulty with such comparisons, for example, in so-called delay of Principle B effects (e.g. Chien and Wexler 1990; Grodzinsky and Reinhart 1993), scalar implicatures (Chierchia, Crain, Guasti, Gualmini, and Meroni 2001; Gualmini, Crain, Meroni, Chierchia, and Guasti 2001), and stress shift and focus sets (Reinhart 2004a). 6 6 Perhaps relatedly, one proposal for why children have trouble with the interpretation of focus operators (e.g. the meaning of only) is that they have difficulty computing the set of alternatives to the focused constituent, and therefore may not have available the element of meaning that only adds to the rest of the sentence, i.e. the set of things that are not true alongside the main assertion (Paterson,


87

5.4.2 Omission of tense The full proposal about acquisition in the so-called Root Infinitive stage, as outlined in Schütze (1997), posits that, independently of children’s failure to obey the AMP just discussed, they also have the option of underspecifying or omitting tense in roughly the sense of Wexler (1994), at least in many languages. Space precludes a full discussion of the motivations and consequences, but an important observation that this is intended to capture is the fact that English children include among their apparently nonfinite utterances some that have nominative subjects and some that have nonnominative subjects (e.g. He cry vs. Him cry, I tired vs. My tired). The idea is that a nonfinite utterance with a nominative subject is missing a tense specification, but the nominative subject is evidence of a case- and agreement-checking relationship between the subject and Infl, hence subject agreement features are not missing. (An additional claim, which Schütze (1997) argued for at length, is the universality of a requirement that structural (morphological) case and agreement must be checked together as part of a single operation, dubbed Accord. 7 ) Conversely, a nonfinite utterance with a nonnominative subject must not manifest case- and agreement-checking with Infl, hence must be missing these features from Infl, but could still bear a tense specification, as evinced by the existence of utterances of the form Him cried. (In this situation a default case form may surface; depending on assumptions, this might also subsume genitive subjects like my—see Schütze (1997, 2001).) This two-factor theory is called the Agr/Tense Omission Model (ATOM). 8 For data supporting this model for English, the reader is referred to Schütze (1997) and Wexler, Schütze, and Rice (1998). New English data are presented

Liversedge, Rowland, and Filik 2003). Computing the set of alternatives to the focused element does not involve comparing alternative representations per se, but it does involve generating a set of related representations that differ in a specific locus, an operation that is also required for the cases in the main text (Principle B, scalar implicature) and hence a potential Achilles’ heel for children’s computation of all of them. 7 Not observed previously is the fact that this restriction could be used to reduce the number of possible derivations that the AMP needs to consider. 8 I follow many older accounts of Root/Optional Infinitives in assuming that the possibility of underspecifying Tense arises from its semantics, and hence is fundamentally unlike that of omitting agreement features; but cf. Hoekstra and Hyams (1995). Proposals inspired by ATOM have attempted to identify a unified underlying mechanism, e.g. Wexler’s (1998) Unique Checking Constraint, or Legendre, Hagstrom, Vainikka, and Todorova’s (2002) Optimality Theory analysis that uses two constraints of the same type for Agr and Tense, combined with a general constraint against functional categories. One prediction made directly by ATOM that requires an extra stipulation in the unified models is that Tense and Agreement could become obligatory at different times in a single child’s development. In fact, a double dissociation has been found: Schütze (1997) observes Agr becoming obligatory before Tense, while Ingham (1998) observes the opposite.

88

Formal features

Table 5.1 Age range and number of recordings for each Swahili child Child Hawa Mustafa Fauzia Hassan

Age range

Starting and ending MLU

Number of recordings

2;2.01–2;6.05 2;0.16–2;10.10 1;8.19–2;2.07 2;10.13–2;11.25

1.54–2.46 1.52–3.57 2.97–3.35 3.15–4.23

7 23 10 4

in §5.6 below. First, however, let us consider a language where the ATOM’s predictions can be tested much more transparently.

5.5 Swahili verbal inflections Background information on the children whose data appear in this section can be found in Deen (2005a, 2005b). 9 The data below have not previously been published in this form. The children were audio-taped by Kamil Ud Deen during spontaneous interaction in Nairobi, Kenya, over a period of several months; each recording session that he transcribed lasted about one half-hour. Pertinent details are summarized in Table 5.1. 5.5.1 Grammatical background Swahili is a pro-drop language with the basic word order SVO; some details below are specific to the Nairobi dialect being acquired by these children. (5)

Morphological structure of the Swahili verb word SubjectAgreement—Tense/Aspect—ObjectAgreement—V— (Suffixes 10 )—Mood

(6) Sample adult utterances a. Juma a-li-m-fuat-a Mariam J 3sg.sbj-pst-3sg.obj-follow-ind M ‘Juma followed Mariam.’ b. Tafadhali ni-pat-i-e kalamu. please 1sg.obj-give-appl-sbjv pen ‘Please give me a pen.’

[indicative]

[subjunctive]

9 Thanks to Kamil Ud Deen for extensive help with this section. All data and grammatical information derives from his works cited in the text or from personal communications. Standard disclaimers apply. 10 Although this slot is traditionally described as containing (up to three) derivational suffixes, the term is somewhat misleading. Morphemes found here include the passive, stative, causative, and applicative. The latter two appear in the data below.

What it means (not) to know (number) agreement c. A-na-tak-a ku-fu-a 3sg.sbj-prs-want-ind inf-husk-ind dafu coconut ‘He wants to husk a coconut.’ d. Som-a! read-ind ‘Read!’

89

[complement infinitive]

[imperative]

5.5.2 Child data on subject agreement and tense The obvious appeal of Swahili for testing the ATOM is that agreement and tense morphemes are separate. Thus, for Swahili the ATOM predicts that children should produce all four combinations of use vs. omission of subject agreement (SA) and tense (T). (Complications involving object agreement (OA) are left for the next subsection.) They do, as exemplified in (7). (7) Child Swahili examples a. Full Form n-ta-ku-on-esh-a mw-ingine ni-ta-ku-on-esh-a mw-ingine 3sg.sbj-fut-2sg.obj-see-caus-ind 1-other ‘I will show you another (person).’ b. Dropped SA ta-ku-pig-a ni-ta-ku-pig-a (1sg.sbj-)fut-2s.obj-hit-ind ‘I will hit you.’ c. Dropped T a-timam-a hapa a-me-simam-a hapa 3sg.sbj-(perf-)stand-ind loc ‘He has stood up here.’ d. Bare Stem (Dropped SA and T) baba end-a hoa baba a-me-end-a hoa B (3sg.sbj-perf-)go-ind home ‘Father has gone home.’

(Fau, 1;11) [adult form]

(Mus, 2;9) [adult form]

(Mus, 2;1) [adult form]

(Has, 2;11) [adult form]

For completeness, it is worth noting that none of the above forms corresponds to the infinitive in Swahili, which is marked with an overt prefix ku-, as in (6c).

90

Formal features

A possible example of a child root infinitive with ku- is given in (8), but as we are about to see these were exceedingly rare. (8) Root Infinitive (with prefix) ku-chez-a ni-ta-chez-a 1sg.sbj-fut-play-ind ‘I’m going to play.’ [However, the child could have intended: ni-na-tak-a ku-chez-a 1sg.sbj-prs-want-ind inf-play-ind ‘I want to play.’]

(Has, 2;10) [adult finite form]

Also, consistent with findings for all other languages where this has been carefully studied, the children make almost no agreement errors, i.e. errors where the form of the agreement inflection mismatches the (overt or intended) corresponding argument—the rate of such errors is less than 1.5%. What makes the Swahili data surprising, however, is that it contrasts with what is known about rich-agreement null-subject languages such as Italian, where root nonfinite verbs are virtually unattested in children’s speech. 11 I return to this contrast in §5.8. Table 5.2 provides counts of the clause types illustrated above (ignoring OA), along with corresponding numbers for the adults interacting with these children. One thing that is immediately clear is that the children are not mirroring the distribution they are hearing: each child is producing vastly fewer fully inflected verbs, and substantial numbers of all three finite forms with missing affixes. Particularly striking are the two utterance types marked with dagger signs in the bottom row, which are virtually unattested in adult speech. A form with subject agreement but no tense morpheme is impossible unless the mood marker is subjunctive, which it was not in any of the child utterances. 12 A form with neither subject agreement nor tense is also possible with subjunctive mood, as in (6b), and as an imperative, as in (6d), but the counts in Table 5.2 11 This characterization of child Italian is not universally agreed upon. Phillips (1995) notes a rate of 13% root infinitives for one Italian child in the earliest recording sessions, and Salustri and Hyams (2006) show that Italian children produce an unexpectedly high rate of imperatives, which they argue to be the analog of root infinitives in certain respects. Nevertheless, child Italian clearly does not look like child Swahili. 12 Kamil Ud Deen (p. c.) informs me that it is unlikely that these utterances were attempts to produce a subjunctive while not reliably knowing the subjunctive suffix, because in context a subjunctive meaning would not have been appropriate. For similar reasons, they are unlikely to be attempted infinitives: those would entail a modal meaning that was not present.

91


Table 5.2 Proportions of all indicative clause types for each child and for the adults in a subset of these files INFL Child

[+SA, +T]

[−SA, +T]

[+SA, −T]

[−SA, −T]

Infin

Hawa Mustafa Fauzia Hassan MEAN Adults

13 (13%) 136 (27%) 183 (52%) 225 (60%) 38% 1,380 (94%)

20 (20%) 225 (44%) 135 (38%) 104 (28%) 33% 72 (5%)

18 (18%) 53 (10%) 17 (5%) 26 (7%) 10% †14 (1%)

47 (48%) 92 (18%) 17 (5%) 15 (4%) 19% †4 (0.3%)

0 7 (1%) 0 7 (2%) 1% 0a

a

Note: There were two or three elliptical infinitives licensed by the preceding utterance.

exclude utterances that may have been intended as imperatives. Effectively, then, the children are producing two ungrammatical clause types. This is not surprising from the perspective of the ATOM, which makes no reference to adult clause types, but it would be very surprising on input-driven accounts, because they are producing as errors forms they never hear. (They are also producing virtually no genuine root infinitives, suggesting that theories of early root nonfiniteness that hinge on special properties of infinitives per se may be on the wrong track.) The month-by-month breakdowns for each child in Tables 5.3–5.6 show that s/he knew (some) SA and T morphemes at the earliest stage, and that all the combinations of omission/nonomission of inflections are attested at virtually every stage. (These tables pool two recordings per month.) This is important for ruling out alternative explanations of type (D) (§5.3), and explanations that, unlike ATOM, would not require two elements to be concurrently “optional” in the grammar. Table 5.3 Swahili children’s clause types as a function of age: Hawa Age Features [+SA, +T] [−SA, +T] [+SA, −T] [−SA, −T]

Verb form

2;2

2;3

2;4

2;5

2;6

TOTAL

SA-T-V-IND T-V-IND SA-V-IND V-IND TOTAL

5 1 0 2 8

2 6 2 12 22

1 5 5 19 30

3 3 5 8 19

2 5 6 19 32

13 20 18 60 111

92

Formal features

Table 5.4 Swahili children’s clause types as a function of age: Mustafaa

Age Features

2;0 2;1 2;2 2;3

2;4

2;5

2;6 2;7 2;8 2;9 2;10 TOTAL

[+SA, +T] [−SA, +T] [+SA, −T] [−SA, −T]

1 0 1 2

2 3 0 0

5 23 14 16

18 14 9 2

33 71 16 18

14 43 4 10

1 6 0 4

5 17 4 13

5 17 1 10

9 16 3 9

43 15 1 8

136 225 53 92

TOTAL

4

5

58

43 138

71

11

39

33

37

67

506

a

Note: Verb form column omitted for space reasons; see Table 5.3.

5.5.3 Child data on object agreement Recall that the proposed explanation of children’s omission of uninterpretable agreement features in terms of failure to comply with the AMP is not specific to subject agreement—all instances of index agreement are predicted to be potentially omitted in the same way. We would therefore like to know whether these Swahili children are omitting object agreement (OA), given that they generally know how to produce it, as shown in Table 5.7. However, OA is syntactically “optional”, in the sense that its presence versus absence is governed by specificity, animacy, and other semantic properties of the object. There are only two contexts where we can determine that object agreement would be obligatory without being able to unequivocally determine discourse status: when the object is a proper name and when it is syntactically topicalized. Rates of OA omission in these contexts were quite low: of 27 proper name objects, only two lacked OA (7%), as in (9); of 12 topicalized object contexts, two lacked OA (17%). These figures are Table 5.5 Swahili children’s clause types as a function of age: Fauzia Age Features [+SA, +T] [−SA, +T] [+SA, −T] [−SA, −T]

Verb form

1;8

1;9

1;10

1;11

2;0

2;1

TOTAL


6 12 1 4 23

34 44 3 4 85

4 20 2 1 27

33 21 8 6 68

64 12 1 1 78

42 26 2 1 71

183 135 17 17 352


93

Table 5.6 Swahili children’s clause types as a function of age: Hassan Age Features [+SA, +T] [−SA, +T] [+SA, −T] [−SA, −T]

Verb form

1;10

1;11

TOTAL


125 53 19 8 205

100 51 7 7 165

225 104 26 15 370

consistent with ATOM’s prediction, though they seem unexpectedly low compared to the SA omission rates in the previous subsection. It is possible that the constraints that require OA in these particular configurations effectively cause derivations without it to crash, independently of the AMP. (9)

OA Omission h-u-beb-a Fauzia? neg-2sg.sbj-carry-ind F h-u-m-beb-a Fauzia? neg-2sg.sbj-3sg.obj-carry-ind F

[child utterance] [adult form]

‘You don’t carry Fauzia?’

5.6 English twin data The data in this section have not been previously reported. Background information on the study from which these data were derived, which was conducted at MIT, can be found in Ganger (1998). 13 The children are monolingual Table 5.7 Total number of object agreement markers produced Child Hawa Mustafa Fauzia Hassan

Number of OAs 35 40 95 65

13 Deepest thanks to Jenny Ganger for the counts reported in this section and for comments on an earlier version. Twin speech data collected by Jennifer Ganger, with financial support from Kenneth Wexler and the Research Training Grant “Language: Acquisition and Computation” awarded by the National Science Foundation (US) to the Massachusetts Institute of Technology (DIR 9113607).

94

Formal features

English-learning same-sex twin pairs, either mono or dizygotic, recorded in spontaneous interactions; their ages during the period of recording are specified in the tables below. Data are reported only for the subset of twin pairs who produced substantial numbers of nonnominative subjects, because only they allow the specific predictions of ATOM to be tested; this resulted in four of the eight twin pairs from the study being included. As in all ATOM studies, we only examine recordings after a point at which the child has demonstrated productive (i.e. non-imitative) use of both a nominative pronoun and its nonnominative counterpart, and to avoid diluting error rates we stop when nonnominative subjects disappear. 5.6.1 Data classification The children’s utterances are classified according to the form of the subject (for pronouns that show subject/object case distinctions) and to the nature of Infl. For the latter, three categories are used: “Agreeing” forms include finite forms of auxiliary and copular be in present or past tense, 3rd person singular main verbs, and present tense dummy do. “Ambiguous” forms include all modals and past tense forms of main verbs and dummy do. “Uninflected” forms include omission of obligatory 3rd person singular -s ; omission of dummy do, auxiliary or copula be; uses of the word do or have with a 3rd person singular subject; uses of the word be in place of a finite form; omission of past tense marking in a past tense context. This last subcategory has not been employed in previous investigations of the ATOM because pre-existing corpora did not have rich enough contextual information to allow confident assessment of the intended time reference of children’s utterances. In the twin study, however, the experimenter was present at the recording sessions and cognizant of the need to track this property of the situation as well as possible. The logic behind these three groupings of forms is as follows. The agreeing forms are taken to be unambiguously specified for agreement features. That is, am, is, etc. could not surface unless the subject person and number features were there for vocabulary insertion to refer to. Likewise, -s appears only with a particular person/number combination, with -Ø being the elsewhere form of (nonpast) inflection, so 3rd person and singular must be specified for it to surface. (The assumptions about do are somewhat more complicated— see Schütze (in prep).) Ambiguous forms are ones that would sound the same whether agreement features were specified or not. That is because they show no agreement contrasts. Uninflected forms are forms that are definitely missing some feature specification, though we may be unable to determine which one. For example, when we hear Mary like ice cream, something that


95

conditions -s is missing, but we do not know if it is the tense feature or agreement. Similarly, Mary run home in a past tense context is obviously missing the tense specification, but it may or may not also be missing the agreement specification. The reason for dividing verb forms up in this way is to derive more finegrained predictions from the ATOM, specifically concerning rates of nonnominative subjects, which (because of the assumption about Accord as a unified checking operation) are predicted to occur only when agreement has not taken place, that is, when uninterpretable phi-features are not inserted in Infl. Agreeing forms definitely have these phi-features. Ambiguous forms could be fully featurally specified, including phi-features, although they might also be lacking them. Uninflected forms are definitely underspecified for either agreement or tense, hence have a good chance of lacking phi-features (the former possibility), ceteris paribus. The ATOM therefore predicts that the rate of nonnominative subjects should increase as we move from agreeing to ambiguous to uninflected forms, as the proportion of missing agreement increases (by hypothesis). 5.6.2 Results The counts for the four relevant twin pairs are presented in Tables 5.8–5.11, with twins, arbitrarily labeled A and B, shown next to each other for easy comparison; because of the requirements for prior production of forms, the recordings that could be used were not necessarily the same for each member of a given pair. Casual inspection suggests that there are noncoincidental similarities within pairs of twins in the rates and possibly the distribution of nonnominative subjects. This is not surprising, given that twins share both genes and environment. Given this lack of independence, the data cannot be interpreted as if they were coming from eight unrelated children. For purposes of this chapter I summarize the data from the A members of the twin pairs, and compare this sample to the data from the B members of the twin pairs, as a sort of confirmation of the findings. Each child’s data consist of four sets of pronouns, for a total of 16 opportunities to test the predictions. Among the A twins, four of these contexts are uninformative because no nonnominative subjects were produced with any verb form. Of the remaining 12, let us consider first the clearest prediction, namely that agreeing verbs should have fewer nonnominative subjects than uninflected verbs: this is true in 11 cases (92%). The picture is more mixed when we add the ambiguous forms, and ask whether the rate of nonnominative subjects is nondecreasing across the three categories: that is true in eight of

96

Formal features

Table 5.8 Distribution of MIT twins’ subject pronouns: twin pair I

Subject

A (2;5–3;5)

B (2;0–2;11)

Verb form

Verb form

Agreeing Ambiguous Uninflected Agreeing Ambiguous Uninflected

I me my %nonNOM

15 0 0 0%

54 2 0 4%

24 0 2 8%

23 0 0 0%

43 1 0 2%

20 3 0 13%

he him %nonNOM

17 3 15%

4 3 43%

10 9 47%

10 1 9%

13 1 7%

23 6 21%

she her %nonNOM

9 0 0%

5 0 0%

7 7 50%

12 0 0%

4 0 0%

14 11 44%

they them %nonNOM

13 0 0%

7 3 30%

8 4 33%

28 0 0%

3 1 25%

7 0 0%

the 12 cases (67%). Independent of these relative rate predictions, the ATOM also predicts that, except for errors due to noise, e.g. production errors (type (F)), the rate of nonnominative subjects with Agreeing verbs should be zero. It is exactly zero in three of 12 cells, and below 10% (a standard noise threshold) in seven of 12 (58%). 14 Among the B twins, only one of the 16 test environments lacks nonnominative subjects altogether. Of the remaining 15, agreeing verbs have fewer nonnominative subjects than uninflected verbs in nine cases (60%). The rate of nonnominative subjects is nondecreasing across the three categories in seven cases (47%). The rate of nonnominative subjects is exactly zero in nine of 15 cells, and below 10% in 12 of 15 (80%). 14 One might argue that the data patterns for different pronouns produced by the same child should not be thought of as statistically independent, if they are seen as alternative ways of approximating the same quantities, namely the child’s actual rates of agreement production as a function of case. But this view holds only if an approach like ATOM is on the right track. It has been suggested instead that children’s early choices of case and agreement are tied to particular lexical items and that “pronoun case-marking errors reflect the absence of abstract knowledge of Case and Agreement” (Pine, Rowland, Lieven, and Theakston 2005), so that there is nothing that these separate pronoun data sets could all be reflections of. Thus, to the extent that they actually pattern similarly, this is part of what I am trying to establish; it is not self-evident a priori.

97

What it means (not) to know (number) agreement Table 5.9 Distribution of MIT twins’ subject pronouns: twin pair II

Subject

A (2;6–3;9)

B (2;8–3;8)

Verb form

Verb form


I me my

77 1 0

135 10 1

45 3 0

65 0 0

144 4 0

20 7 0

%nonNOM

1%

8%

6%

0%

3%

26%

he him

33 2

9 3

22 14

23 1

13 2

5 7

%nonNOM

6%

25%

39%

4%

13%

58%

1 1

0 1

1 9

2 1

0 1

2 0

50%

100%

90%

33%

100%

0%

7 0

4 0

7 0

6 0

1 0

0 0

0%

0%

0%

0%

0%

—

she her %nonNOM they them %nonNOM

Table 5.10 Distribution of MIT twins’ subject pronouns: twin pair III

Subject

A (2;7–3;6)

B (2;7–3;8)

Verb form

Verb form

Agreeing Ambiguous Uninflected Agreeing Ambiguous Uninflected 15 0 0

64 0 0

32 0 0

19 0 0

95 0 0

43 1 0

%nonNOM

0%

0%

0%

0%

0%

2%

he him

17 1

11 0

47 2

12 1

7 0

44 3

%nonNOM

6%

0%

4%

8%

0%

6%

she her

40 1

56 2

145 13

29 1

36 6

120 13

%nonNOM

2%

3%

8%

3%

14%

10%

7 0

9 0

36 0

10 0

3 1

6 3

0%

0%

0%

0%

25%

33%

I me my

they them %nonNOM

98

Formal features

Table 5.11 Distribution of MIT twins’ subject pronouns: twin pair IV

Subject

A (3;1–4;3)

B (2;6–4;4)

Verb form

Verb form


I me my

14 0 0

163 0 0

63 0 0

20 0 0

127 0 4

70 0 0

%nonNOM

0%

0%

0%

0%

3%

0%

8 1

33 3

44 15

3 0

20 7

20 12

11%

8%

25%

0%

26%

38%

1 4

2 21

0 62

0 3

2 24

1 61

80%

91%

100%

100%

92%

98%

5 4

2 20

0 22

1 0

1 9

2 8

44%

91%

100%

0%

90%

80%

he him %nonNOM she her %nonNOM they them %nonNOM

In summary, while there are certainly counterexamples that deserve further investigation, overall the ATOM is faring reasonably well—putting aside the “exactly zero” counts, within each of the two samples all of its predictions are correct more often than would be expected by chance. (Nondecreasing rates across the three inflectional categories would be expected to arise about 17% of the time by chance.)

5.7 Other languages 5.7.1 French The three children whose data are discussed here are Grégoire (1;9.14–2;3.0, Christian Champaud corpus) and Philippe (2;1.19–2;6.21, Suppes, Smith, and Léveillé 1974), both from CHILDES (MacWhinney 2000), and Daniel (1;8.1– 1;11.1, Lightbown 1977). For full details of this study see Ferdinand (1996)— glosses and translations from Ferdinand; for more on the present analysis see Schütze (1997). These children produce a nontrivial number of what appear to be agreement errors, i.e. finite verb forms that do not match the phi-features


99

of the subject, at a stage when correct agreeing forms are also being produced. These errors are illustrated in (10). (10)

a. des motos fait du bruit [adult form: font] some.pl motors makes(3sg) some noise ‘Motors are making noise.’ b. moi a tout bu [adult form: (j’)ai] me has(3sg) all drunk ‘I have drunk everything.’ c. je va les retrouver [adult form: vais] I goes(3sg) them find_back ‘I am going to find them back.’ d. les bulles elles s’en va [adult form: vont] the bubbles they(fem.pl) goes(3sg) away ‘The bubbles are going away.’ e. est dedans. moi est dedans [adult form: (je) suis] is(3sg) inside me is(3sg) inside ‘I am inside. I am inside.’

In all such cases, however, the verb form in question is what Ferdinand identifies as the default form, that is, the (present tense) finite form with the widest distribution across subject features. There are crucially no errors involving misuse of a marked finite form, e.g. ∗ papa vont ‘daddy go.3pl’ or ∗ des motos vais ‘some motorbikes go.1sg’. The frequency of correctly agreeing forms versus incorrect default forms is as shown in Table 5.12. Note that Ferdinand counted not only utterances where the subject was overt but also those where it was null but the intended meaning was clear from context. Prima facie, these data look like counterexamples to the widespread claim mentioned above that children do not misuse agreeing verb forms. It would be hard to claim that they are really infinitives with a phonologically dropped ending, i.e. Root Infinitives, because Ferdinand claims that “no positional difference can be observed between elsewhere forms and specified forms”, that is, the elsewhere forms apparently undergo verb raising, unlike Root Infinitives (Pierce 1989). However, under my hypothesis these forms have Infl in which present tense is specified but agreement is not. If phi-features are missing from Infl, there is nothing to trigger insertion of the correct agreeing verbal suffix, but since the form is tensed, the infinitival suffix cannot be inserted either. Instead, the agreement slot will be filled with the agreement affix that has no phi-features specified, the elsewhere affix. In the case of -er verbs, this will be

100

Formal features Table 5.12 Distribution of agreeing versus default verb forms as a function of subject phi-features for three French children Child Form Grégoire Elsewhere form Specified form Daniel Elsewhere form Specified form Philippe (early files) Elsewhere form Specified form Philippe (later files) Elsewhere form Specified form

Environment Elsewhere context

Specified context

>50 0

14 8

>100 0

27 5

>100 0

16 5

>100 0

20 89

zero, but with irregular verbs like those illustrated in (10) that affix may be audible. Pratt and Grinstead (2007) report analogous findings for child Spanish, supplementing them with grammaticality judgment data that show four- to five-year-olds accept root nonfinite forms (including counterparts to Ferdinand’s default forms) over 25% of the time.

5.7.2 Nonsubject case errors ATOM in conjunction with the hypothesis that case and agreement must be checked together makes the following prediction. In languages where nominative (i.e. subject case) is the default case, lack of agreement should manifest itself as overextension of nominative case marking to nonsubject DPs. That is, in all languages I assume that agreement with all arguments will fail some proportion of the time during the ATOM stage, but in the case of arguments that trigger no agreement on the predicate, the only evidence of this would be case-marking errors, and these will only be detectable if the correct case in a given position is different from the default case of the language. This prediction has been shown to be true in German (Schütze 1997; Berger-Morales 2005a, 2005b) and Russian (Babyonyshev 1993). In both languages, case marking in subject position is essentially perfect, but in nonsubject positions


101

errors are relatively common, and the most frequent error is the use of the nominative. 15

5.8 Concluding remarks The data and theoretical proposals I have presented leave a number of issues in need of further investigation: r What makes child Swahili different from child Italian, such that Swahili

children can omit finite inflections but Italian children apparently cannot? Is it the difference between portmanteau vs. agglutinative realization of features? That is, omitting any of the three relevant morphemes from the Swahili verb word yields a pronounceable and morphotactically wellformed word. By contrast, omitting finite inflection from an Italian verb would yield a bare stem, which can never surface as a word in any context, and because tense and (subject) agreement are generally fused, it is also not possible to omit just one of them and be left with a well-formed word. Future research will be needed to determine whether this proposed explanation for the cross-linguistic contrast is too “superficial”. Ideally we would like to find a language (perhaps in the mountains of Northern Italy) that is syntactically just like Standard Italian but in which bare stems happen to be used somewhere in the adult language. According to my hypothesis, children should produce these in substantial numbers in declarative contexts, with a variety of subjects. r Why does number (marking) show the particular developmental profile that it does? Specifically, why does it appear to be acquired later than word order but earlier than, for example, Principle B, 16 and around the same time as finiteness? Does the kind of computation discussed in this chapter (the AMP) simply become reliable at a certain developmental point relative to parameter setting and to other computational abilities (e.g. derivational comparisons involving meaning)? 15 This is straightforwardly true for Russian. German has an extra complication, however. Certainly it evinces an asymmetry between subjects and objects that is the opposite of that found in English: in Berger-Morales’s (2005a) sample of three-year-olds, out of 312 total case errors, four were object forms in subject position, while 19 were subject (nominative) forms in direct object position. In positions where dative case was required, most errors involved forms that are syncretic (homophonous) nominative/accusative, but among unambiguous forms, accusatives outnumbered nominatives 96 to 11. However, there is good reason to think that many of these errors are due to the phonological similarity of the masculine accusative and dative endings, and hence do not represent a general overextension of accusative case. 16 Unless Conroy, Takahashi, Lidz, and Phillips (2006) turn out to be correct that the Delay of Principle B Effect is an artifact of experimental methodology and processing difficulties.

102

Formal features

r Would an account of “optional” (possibly conditioned) agreement in

adult grammars, as found in Russian (11), be a more parsimonious approach for accounting for optional agreement marking in child language? This might be hard to answer before we are sure what the former really is (cf. van Gelderen 1997 for comparison of approaches), but Guasti and Rizzi (2002) propose exactly this (cf. also Meisel 1994). (11) Russian (Corbett 2000: 213) a. vošl-o pjat’ devušek came_in-sg.neut five.nom girl.pl.gen ‘Five girls came in.’ b. vošl-i pjat’ devušek came_in-pl five.nom girl.pl.gen ‘Five girls came in.’ r Given that adult languages employ representations that require more

computation than is necessary for convergence, and that children produce less costly (but still convergent) counterparts some proportion of the time, why have all languages not evolved to look like Chinese, with no case or agreement marking, and possibly no case or agreement features either? 17 I have no answer to this question, but I do not believe it arises just under my analysis. One could equally well ask why the need to learn paradigms of inflections does not disfavor inflected languages. Perhaps the computation and learning are actually so easy, once humans reach a certain stage in development, that there is really no pressure to avoid them. To summarize my proposals: The distribution of uninterpretable features in syntax is governed by the Accord Maximization Principle, which requires that among convergent derivations that differ only on the presence/absence of such features, the admissible one will be the one that first satisfies any independent grammatical requirements and then includes as many uninterpretable features as possible. The distribution of such features is constrained by the stipulation that case and agreement features on the same head must be checked simultaneously as part of a single operation (Accord). This much was independently needed to solve issues in the syntax of adult languages. Children are hypothesized to differ from adults in their ability to consistently enforce the AMP, plausibly because it requires comparisons among competing derivations and there is independent evidence that this is difficult for children. This is the respect in which they “do not know” agreement; there may well 17

Thanks to an anonymous referee for raising this issue.


103

be other things they do not know about agreement, particularly at earlier stages of development, and I have enumerated what some of these might be. Separately, for reasons not relevant to the acquisition of agreement, children in the relevant stage are assumed to be able to underspecify tense features in matrix declaratives. The combination of these two claims about child language constitute the ATOM. As we have seen, its observable predictions differ across languages, depending on their morphological properties. For example, different sorts of verb forms arise as a result of lack of agreement (default finite forms in French, bare stems in English, transparently missing agreement prefixes in Swahili), as do different kinds of case errors as a reflex thereof. Clearly the ATOM makes a great many empirical predictions that remain to be tested.

6 Number agreement in the acquisition of English and Xhosa JILL DE VILLIERS AND SANDILE GXILISHE

6.1 Overview How do children come to understand number agreement, that is, how does an inflection on the verb carry information about subject number? The puzzling fact that emerges is that children acquiring English control number agreement in production quite early, while failing to use the information from the verb in comprehension tasks until several years later. Two proposals are compared to account for this asymmetry: a formal one from modern linguistics accounts that predicts that features on the target of morphological agreement should not be accessible to interpretation, and a conceptual one that verbs, unlike pronouns, do not carry notional number. English is compared with other languages in which agreement on the verb is not redundant, for example in pro-drop languages. Xhosa, one of the Bantu languages, is taken as a test case: it has a rich noun class system and correspondingly rich subject agreement on the front of the verb. The status of subject agreement in Bantu languages is a topic of several decades of debate, in particular, whether it should be treated as a clitic pronoun or as an agreement marker, and whether the different members of this family of languages differ along the continuum of possibilities. A proposal is made in which data from children’s comprehension might contribute to this discussion.

6.2 Basic number agreement Agreement has been considered in several different ways under different theories. The example to be considered here is from the agreement between the subject and the verb in number, though there are many languages that have number agreement also with adjectives, determiners and so forth. How

Number agreement in the acquisition of English and Xhosa

105

does it happen when the subject and verb agree in number in Standard (or Mainstream) English? (1) He goes to the store. (2) They go to the store. The classical account is that the number agreement on the subject noun is copied to the verb, in a unidirectional fashion; that is, the verb’s number is dictated by the number of the subject noun. Person, number, and gender are known as phi-features, and in a variety of languages they enter into agreement dependencies with elements in the clause distant from their source. The mechanism by which the source and target are connected varies with the linguistic theory. In modern generative accounts, the verb may raise in the syntactic structure to check number in some node called Agreement, which may or may not coincide with Tense (Pollock, 1989). In other accounts such as HPSG (Pollard and Sag, 1994), the copying is non-directional and initiates from a “referential index” in the world that dictates the number both to the subject and the verb (see also Murphy, 1997, on Bantu languages). A constraint then operates to ensure that both elements have the same number. In the copying model, the subject is the controller that has the semantics attached (e.g. number). In generative syntax, the subject raises out of the VPmerged position into a higher node, which triggers the verb to move to a subject agreement node to check its phi-features. In the constraints model, the “reference” properties are dictated to both subject and verb, so neither has priority. It is possible that data on child languages could be used to differentiate these approaches, a question to which we will return in considering languages with very rich agreement systems.

6.3 Number agreement in AAE (African American English) and MAE (Mainstream American English) How do children acquire number agreement on verbs? Although it has been established for decades now that young children speaking Standard English reliably produce agreement (3rd person /s/ on verbs) at about age three and a half years, less was known about comprehension of the form. The process of pursuing comprehension of agreement led us to some surprising findings. Our work began with an innocent question about AAE, which has “optional” number agreement (Green, 2002; Labov, 1969; Myhill and Harris, 1986). (3) He go to the store. (4) They go to the store.

106

Formal features

As part of a large project to investigate the normative course of development of AAE (Seymour, Roeper, and de Villiers, 2003a,b), it became necessary to understand whether children acquiring that dialect might be missing the 3rd person /s/ on the verb for purely phonological reasons. It is well established (Labov, 1969; Seymour and Pearson, 2004) that AAE has a different set of phonotactic rules for the ends of words than other dialects of English, especially in final consonant clusters (producing “tes” for “test” etc.), and the question arose about the child’s sensitivity to the final /s/ in comprehension. If it could be shown that AAE speakers could understand the information carried by the /s/ even though they did not say it for phonotactic reasons, then the prospect was good for using a comprehension test rather than the usual production test in assessment of children suspected of language disorders. A chronic problem recognized for many years is that the inventory of morphemes used for language assessment on standardized tests is at variance with the inventory of morphemes in AAE, potentially leading to misdiagnoses of language disorder in AAE-speaking children (Seymour, Bland-Stewart, and Green, 1998). In previous work on comprehension of the information conveyed by the number agreement on the verb (e.g. Fraser, Bellugi, and Brown, 1963), the trick has been used of making the subject an irregular plural with no overt marking so all the information on number is carried by the verb ending: (5) The deer run in the park. (6) The deer runs in the park. Results suggested that this was difficult even for Standard-English-speaking children (e.g. Fellbaum, Miller, Curtiss, and Tallal, 1995), but it is also well known that children aged four to six have problems with irregular plurals such as “deer” and “sheep” and “feet” (Brown, 1973). In our pilot testing, four-yearolds often asked us “Do you mean the deers?” suggesting that this was a source of confusion. As an alternative, one can disguise the plurality of the noun by ensuring that the following verb begins with /s/, in which case if the sentences are pronounced as in running speech, the existence or not of the plural /s/ on /cat/ in (7) and (8) is indeterminate: (7) The cat sleeps on the bed. (8) The cats sleep on the bed. Johnson (2005) devised a set of stimuli of this sort and used a picture comprehension test of the kind shown in Figure 6.1 to ask whether AAE-speaking children aged four to six years could use the number agreement on the verb to identify the number of the subject noun. Since no other information was


107

Figure 6.1 Sample stimulus for the recorded sentence: /therabbitsnifftheflowers/ Source: From studies of Johnson (2005) and Johnson, de Villiers, and Seymour (2005)

available from the referential context, the assumption was that a child who had mastered the grammar of subject number agreement would be able to use the 3rd person /s/ to determine the number of the subject even if AAE phonological constraints meant that it was not produced in their own speech. In fact, AAE-speaking children showed no sensitivity to the information in 3rd person /s/ by age six. Johnson concluded that the 3rd person /s/ is not present as number agreement in the grammar of AAE-speaking children at this age, and maybe not in adult AAE either (Green, 2002; Myhill and Harris, 1986). The complication with the story is that Johnson also had MAE-speaking children as participants, originally considered a “control” for the AAE speakers (Johnson, de Villiers, and Seymour, 2005). It was expected that these children, who have full control over 3rd person /s/ in their speech from about three years of age (Brown, 1973), would have no difficulty detecting the /s/ as a clue to subject number in comprehension. However, three-, four-, and even many five-year-old MAE-speaking children failed to use the 3rd /s/ as a clue to subject number. Several alternative explanations were explored, for example, that something about the pictured stimuli might have been responsible. Perhaps spontaneous speech offers more clues than these simple pictures, clues that somehow support the 3rd person /s/ production. As a precaution, we ran a study with children of three and four using single pictures, and asking them to describe what they showed. To ensure the use of 3rd person /s/ rather than

108

Formal features

past tense or progressive, we said the picture showed “what the animals do everyday”, and we taught the children to start every sentence with “Every day” to legitimize the use of the generic tense that 3rd person also marks. After providing a sample of both morphological forms on different verbs, e.g. (9) Every day the raccoon washes in the pool. (10) Every day the pigs roll in the mud. the children were then presented with the series of pictures again, in which the number of raccoons or pigs sometimes varied from the original. The preschoolers had no difficulty producing matched number agreement in this study. The same subjects were tested on the comprehension test used in Johnson (2005) and Johnson et al. (2005), but with the words “Every day” inserted before test sentences like (7) and (8). Despite these methodological improvements, the results were the same, namely the children showed no discrimination of subject number based on the cue from the 3rd person /s/. Thus, even MAE speakers show a two-year gap in performance between production and comprehension, with comprehension lagging behind production. The question becomes, why? Why is it that the information in the agreement feature on the verb is “bleached” of its numeric content? Before offering some theoretical alternatives, consider another question: why does anyone over six years old succeed? The speculation is that six-year-olds might be capable of comparing multiple representations, that is, they may succeed by comparing the output of their own production given the scene with the test sentence. This process would entail the ability to hold the sentences in working memory and to compare multiple representations. Others have proposed that this capacity for comparing representations is late-developing and may be responsible for the delay of Principle B effect, in which ambiguity in Principle B interpretation is also resolved only after age six or so (Reinhart, 2004b).

6.4 Alternative theories Here we compare two basic alternatives for why agreement features on the target, i.e. the verb, may be inaccessible in interpretation without special effort (of the sort that requires comparing representations). The first is more linguistic and the second more conceptual. In explaining the problem in interpretation of 3rd /s/ on the verb, Johnson et al. (2005) borrowed a concept from the discussion in Chomsky’s Minimalism (1995), namely that, once the number is checked off in agreement, it no longer survives as semantic information at LF. This would allow automatic procedures to arrange number agreement in production, but in


109

comprehension the agreement marker on the verb would carry no semantic information. Bobaljik (2006) makes a more specific proposal that agreement is a late operation, part of the post-syntactic morphological component. He uses data from a variety of languages to address the prediction that it should be possible for an NP to control agreement on a predicate, even if it bears no syntactic relationship to that predicate other than being “close enough” (for which the technical details are not needed here). The conclusion he reaches then forces a second prediction, namely that “agreement features on the target of agreement do not contribute to interpretation”. But, if there is something right about the failure in interpretation of agreement targets, then it should not just be a feature of child language but should also appear in adult processing of language. No direct test of this has been carried out, though there is a different literature on how adults resolve number agreement with ambiguous nouns such as collectives. For example, in a study by Bock, Nicol, and Cutting (1999), adult subjects participated in a production task in which they either had to produce verb agreement or pronoun agreement with a subject noun. The subject nouns were of different types but included forms such as “committee”, which is a collective noun that is notionally plural, but grammatically singular: (11) The committee meets on Tuesdays. (12) ∗ The committee meet on Tuesdays. Adults responded differently when asked to provide verb agreement or pronoun agreement. The verb agreed with grammatical number: (13) The committee meets. . . . but pronouns agreed notionally with the subject: (14)

and they said. . . .

Bock et al. contend that when an agreement “controller” (namely the subject) carries a grammatical number that is not the same as its notional number, verb agreement targets generally match the grammatical number and pronoun agreement targets instead generally match the notional number. However, these authors also draw a larger conclusion about why nouns and verbs may behave differently with respect to number, which brings us to the second, more conceptual account. “Verbs denote things whose number properties are at best slippery. As a property of states and events, number is abstract (Shipley & Shepperson, 1990; Wynn, 1990) and often indeterminate. Is hand-shaking singular or plural? Is kissing singular or plural? Is football playing singular or plural? It may well be that the syntactic work of indicating

110

Formal features

what goes with what in a string of words is more readily accomplished by using the number features of the subject to mark the verb, especially since English verbs usually occur with a morphologically explicit subject” (1999, p. 341). The implication is that the information on the verb is secondary, or derived, for a very good conceptual reason. In more recent work, Eberhard, Cutting, and Bock (2005) provide a synthesis of several experiments in this vein to argue that, in speech production, the status of number on pronouns and verbs is derived from different sources, and repeat the general conclusion from their model that pronouns match the notional number of the subject more readily than verbs. Nonetheless, it remains to be investigated whether or how the verb information on sentences such as (7) and (8) might be accessed during online processing in adults. Furthermore, the conceptual account provides a different angle on the phenomena but is not incompatible with the generative account. These authors raise interesting historical questions about how pronouns and agreement relate. Eberhard et al. (2005) bring in evidence that verb agreement in English arose by a process in which antecedent–pronoun number agreement was linked to subject–verb number agreement. Historically, there is evidence that topicalizing constructions move the subject into an initial topic position, often introducing a pronoun repeat of the subject (Givón, 1976): (15) The girl, she like candy. The argument is that in the earliest Germanic origins of English, the topicalization results in a post-posed subject: (16) The girl, like-she candy. Phonological reduction and assimilation processes over time then reduce the pronoun to a verb inflection such as the one in Mainstream English dialects: (17) The girl likes candy. Once the form has grammaticized, it becomes an obligatory verb inflection insensitive to discourse requirements such as those that give rise to topicalization. Eberhard et al. (2005) argue that verb inflections and pronouns share a common sensitivity to number historically, but in present-day English, “singular and plural verb forms are comparatively numb to number meaning” (p. 538). In their model pronouns achieve their number via concord, a semantic process of coreference with the subject noun. But verbs get their number via syntactic agreement. Eberhard et al. admit that their model so far is meant to account for agreement-lean languages such as English, and they make no claims for generality to languages that might bear richer agreement. The developing picture is one in which pronouns carry agreement features by a different process than does verb agreement, and that will become an


111

interesting issue for us when dealing with a class of languages where the distinction between pronoun and agreement is not so evident.

6.5 Number agreement in languages other than English English is a difficult language from which to reach a broad conclusion. Agreement in general is very weak: there is no marking of case or gender on nouns or verbs, and the verb number agreement on regular verbs is only for third person subjects, and only in the case of the so-called and misnamed (Sauerland, 2002) “present tense”. Furthermore, the circumstances are rare in which the notional plurality of the subject number is disguised, as with abstract collectives (“committee”) or when the following verb starts with an /s/ and there are no other contextual or linguistic (e.g. pronoun) clues. Writing about this problem, Brown (1973) argued that the clue from number agreement was not salient to children because it is rare in English to have to rely on it. It is very important to consider data from languages in which the verb provides a more consistent and important cue to number, namely pro-drop languages. If the subject is not there, then the only clue to its number (and/or gender, etc.) comes from the morphology on the verb. Looking at children speaking Dominican Spanish, Pérez-Leroux (2006) found strikingly parallel results to those described above for English (Johnson et al., 2005). When the children were exposed to sentences in which the agreement morphology on the verb was the only cue to detecting which picture to choose (singular or plural subjects), the three-, four-, and five-year-old Spanish speakers were no better than the English-speaking preschoolers at the task. The verb agreement morphology was not used as a clue to the number of the subject even though in Spanish the subject must frequently be absent because of prodrop. At least for children, it seems that verb agreement is not more salient as a marker of notional number even when its “cue validity” is increased by pro-drop. In the chapter by Arosio, Adani, and Guasti (this volume), a similar result is found in Italian with more complex structures involving relative clause interpretation. In that case it is not until children are well advanced in years, perhaps nine years of age, before they will use the verb agreement properties to differentiate whether something is a subject or object relative clause. The control case demonstrates that they know the relative clause structures, in that they can use information from structural position to make the right interpretation, but not from agreement. The authors argue that, in parsing, agreement is a post-syntactic operation, a position potentially compatible with the formal grammar model by Bobaljik (2006) discussed earlier.

112

Formal features

6.6 The case of Xhosa Our goal is to enlarge the discourse even further by considering a very rich agreement language, Xhosa, and to consider the ramifications for how children learn number agreement in that language. One of nine national African languages of South Africa and one of the Nguni group of Bantu languages, Xhosa is primarily agglutinative, with morphology accumulating on the verb stem. There are nine positions on the verb into which a grammatical morpheme might slot, and they include markers of agreement with both subject and object noun class. Like many Bantu languages, Xhosa has numerous noun classes that are relatively arbitrary, but may have historically semantic roots. For example, noun class 1 is mostly names for humans. However, in the present-day language there are many exceptions and overlaps (e.g. names for humans also appear in seven other noun classes) so the semantics of the referent are only weakly associated with noun class membership. A partial ordering of morphemes on the verb in Xhosa is as follows: (18)

U - ya - m - fund - is - a. SM-1a - T - OM-1a - Verb - CAUS - M1 She present him learn cause indicative She causes him to learn ‘She teaches him.’

There are arguments that the verb in Xhosa is not a complex head (See Buell (2005) on the closely related Nguni language, Zulu), but that each morpheme in fact has its own head in the hierarchy (Du Plessis, 1997; Deen, 2005a) and the full morphologically complex verb is only created at Spell-Out (Julien, 2000, cited in Deen, 2005a). Xhosa has SVO word order but other variations of this order occur frequently. The subject noun can be dropped (pro-drop), leaving only the subject agreement on the verb appropriate to the class of the absent subject noun. The basic sentence form is thus: (19) I-si-lumko si-thanda iincwadi. 7-genius 7-SM-likes 9-books ‘The genius likes books.’ but post-verbal subjects occur, as in: (20)

Si-thanda ii-ncwadi isilumko. 7-SM-like 9-books 7-genius ‘Likes books the genius.’

1 Noun classes and agreement are marked with numbers, according to convention. TNS = Tense, SM = Subject Marker, OM = Object Marker, REL = Relative Marker, PERF = Perfective.


113

or with pro-drop: (21)

Si-thanda ii-ncwadi. 7-SM-like 9-books ‘Likes books.’

What about number agreement? Number is not associated with a single morpheme but instead the form changes by noun class. Of the 15 noun classes, eight are singular and seven are plural; however, the formation of the plural is not straightforward in morphology. For example, in the following examples, the change from singular to plural is different for each noun class, unlike languages such as English, which have only tiny irregularity in plural formation (man/men, child/children, foot/feet). (22)

Singular Class 1: um-ntwana 1a: u-tata Class 7: isi-lumko

Plural Class 2: aba-ntwana 2a: oo-tata Class 8: izi-lumko

The assumption can be made that the plural morphology on the noun arises, as with noun class, in the lexicon. When it comes to subject and object verb agreement with the noun class, once again it is not a straightforward copy of an agreeing prefix, rather the plural form of agreement varies with class: (23)

Oosisi ba-hlala phezu kwesofa. 2a-sisters 2a-SM-sit on top sofa ‘The sisters sit on the sofa.’ (24) Izinja zi-hlala phezu kwesofa. 10-dogs 10-SM-sit on top sofa ‘The dogs sit on the sofa.’

How does a child acquire such a system, and does the child learn it in a piecemeal fashion, verb by verb and morpheme by morpheme?

6.7 Acquisition of subject and number agreement in Xhosa The initial data to answer this question come from a sample of children studied longitudinally by the second author. This group consists of six children growing up with Xhosa as their first language in the township of Gugulethu outside of Cape Town, South Africa. Beginning at age two they were audiotaped approximately once every two months in natural interaction in their homes with family members and with a native Xhosa-speaking researcher who

114

Formal features

Table 6.1 Number of utterances and number of samples ( ) by age band Age

C6

C7

C8

C9

24–30m 30–36m 36–39m Total

80 (3) 124 (3) 69 (2) 273 (8)

152 (3) 132 (3) 92 (3) 376 (9)

142 (4) 56 (2)

45 (3) 75 (3)

198 (6)

120 (6)

C10

C11

Total

149 (4) 86 (3) 104 (3) 339 (10)

72 (4) 54 (2) 50 (2) 176 (8)

640 (21) 527 (16) 315 (10) 1,482 (47)

also transcribed the tapes. The transcriptions were checked by a second native Xhosa speaker. Table 6.1 shows the number of samples from each child and the number of utterances collected for each sample over the period between approximately 24 to 39 months of age for each child. Samples were combined into three age bands to provide enough utterances in each age band for reliable developmental analysis. The transcripts consist of the child’s utterance, a gloss of the intended utterance as it would be produced by an adult speaker, and an English gloss. The research assistant’s speech is also transcribed and provided with an English gloss. These data and also data from an even younger group of Xhosa speakers collected in the same manner reveal that subject agreement was well established by age two years (Gxilishe, de Villiers, and de Villiers, 2007). The group of children aged two to three years used subject agreement appropriately, with practically no substitution errors. The finding of omission but no substitution errors has been reported commonly for Bantu language acquisition (Deen, 2005a; Demuth, 2003; Suzmann, 1982). That is, children do omit the subject agreement at age two years but almost never use the wrong form, despite the complexity of the agreement paradigms. It does not seem to be the case that the children are using many rote utterances, in fact quite the contrary: like children everywhere, they are using novel utterances. As an index of productivity, the number of noun classes per transcript were tallied, and they averaged three to five different noun classes per transcript. This means that children were not restricting their talk to one or two familiar noun classes and hence achieving success by limited productivity. Neither is it the case that the children use only a few verb roots to achieve success: the different verb roots on which subject agreement is reliably provided vary from five to 15 per transcript. There is thus ample opportunity for errors that nevertheless do not occur. The question that is significant for the present chapter is, how well do children mark number agreement from the subject to the verb? In the current data, plurals represent only 13% of the potential cases of subject agreement


115

Number and subject agreement 100 90

Percentage supplied

80 70 60 50 40 30 20

SingularSubjAgr

10

PluralSubjAgr

0 24−30m

30−36m Age

36−39m

Figure 6.2 Data on plural and singular subject agreement from two- to three-year-old Xhosa speakers

from the children aged 24 to 39 months. Nevertheless, they are very well supplied. Figure 6.2 shows the graph of subject agreement averaged across the six children by age, and it is clear that plural agreement is better supplied than singular subject agreement. Most of the plural agreements are from noun classes 2 and 10, and most of the singulars are from the corresponding noun classes 1a and 9. There is nothing particularly transparent about the plural/singular marking for these classes. Again, omission is the only source of error, not substitution. Again the question can be asked, what if the subject is present versus absent? Does it make a difference in the likelihood that the children will produce the number agreement? Unfortunately, the data are limited, given that the number of obligatory contexts for plural agreement is only 13% of the total subject agreement opportunities. The plural subject noun was present in only six cases, and verb agreement was appropriately supplied in five out of those six (83.3%). The plural subject noun was absent in 21 cases, with plural subject agreement provided in 19 of those 21 instances (90.9%). It appears to be the case that the subject does not have to be overt in the sentence for the child to supply correct number agreement on the verb. This would not be surprising

116

Formal features

in a familiar language like English, where the form of irregularly pluralized lexical nouns has no influence on the form of number agreement on the verb: (25) The man dances. (26) The men dance. But in Xhosa the form of agreement is dictated by the noun class of the subject. If the missing subject noun had a noun class that was semantically transparent (say human) then the referent properties could dictate the form of number agreement. But in Xhosa, noun classes are not so semantically transparent, and noun class is provided in the lexicon, not directly in the world of reference. It seems to be an inescapable conclusion that the subject noun had to be present to dictate the subject agreement on the verb before undergoing deletion. The one qualification necessary to this conclusion is that the predominant noun classes used at this age are classes 1/2 and 9/10, which are mostly humans and artifacts, so it is possible that semantic properties could assist the child at the beginning. But since both humans and artifacts occur in other noun classes, such a hypothesis will soon lead to substitution errors, which, as we have seen, are virtually nonexistent. The puzzle arises when one considers not just production but comprehension, which must be the route by which the system is acquired. The adults around the child do not necessarily restrict themselves to four of the 15 noun classes, so the input will provide evidence counter to a simple semantic mapping. In order to make sense of the input, the child must recover any deleted subject nouns in the input to figure out the relation of noun class and number agreement marker. Little is yet known about the naturalistic input from caregivers to very young children acquiring Xhosa, but it would be interesting to see if “motherese” somehow makes the subject nouns more accessible, either by less deletion, or by having adjacent utterances with and without the subject noun. This question remains a subject for further research.

6.8 The nature of subject agreement in Xhosa There is considerable work on the issue of whether subject agreement in Bantu languages should be treated as a kind of pronominal clitic attached to the verb, or an affix like English /s/ or Italian verb endings. As discussed, there is often a historical move from free pronoun, to clitic, to agreement morpheme, and it must be borne in mind that different languages in this group could be at different points in this progression. The classic work on this question is by Bresnan and Mchombo (1987), who raise the question about subject agreement in Chichewa, another Bantu


117

language. In their typology the status of agreement markers is based on the co-occurrence possibilities of person markers and their controllers in the same construction rather than on the morphophonological form of the agreement markers (Turunen, 2007). A syntactic agreement marker cannot occur without an overt controller in the same sentence; however, a pronominal marker cannot occur with an overt local controller or it would violate Principle A. For that reason, they classify Chichewa subject agreement as pronominal in type when an overt subject is absent (pro-drop), and as agreement when an overt subject is present. However, object agreement in Chichewa obligatorily occurs when the object is dropped or displaced beyond the phrase, and so the OM is classified as pronominal in form. Even for Chichewa, however, there are proponents of the view that the SM is also a pronominal clitic. Baker (2001, 2005) raises the possibility that Bantu languages like Chichewa have a parameter setting of “Optional Polysynthesis” in which crucial parts of the event are incorporated into the verb. In particular, he argues that, in Bantu languages such as Chichewa, the overt subject must be moved outside the clause (before or after) when there is subject agreement. That is, the language may have a grammar like the historical stage of early English verb agreement discussed in Eberhard et al. (2005): (27)

The girl, she like candy.

In such an analysis, the subject agreement marker occupies the subject position, namely Spec-AgrS, and behaves more like a clitic pronoun. Baker argued that the subject in Chichewa is displaced outside the phrase by the presence of SM, i.e. SM occupies the subject position. In that way, there would no longer be a Principle A violation with an overt subject, since it is displaced. One of the convincing rationales for a subject displacement into topic is that whquestions can never be asked directly in Chichewa, but only using a cleft or passive construction. This is because topics cannot be directly replaced with wh-questions. On a view such as this, perhaps both SM and OM markers, being pronominal clitics, lead to dislocation of the corresponding arguments outside the clause, rescuing Principle A. Zeller (2008) puts forward a complex proposal that in Zulu, closely related to Xhosa, the SM is a pronominal clitic that forms a constituent with the subject DP, and that “agreement” in SVO constructions is a form of cliticdoubling like that found in Northern Italian dialects. In contrast to the pronominal view, Buell (2005) puts forth evidence that Zulu has SM markers that do not behave like pronouns. One of his pieces of evidence is that Zulu has a range of compound tenses with a lexical verb embedded under a variety of modals, auxiliaries, and aspect markers. All of

118

Formal features

these auxiliary forms are marked also with subject agreement, in positions that pronouns would not usually occupy. In Buell’s view, subjects are in the specifer of Agr-S (but null in pro-drop) and contribute their features to the verb when it raises to the head of Agr-S, as in generative accounts of English, hence are not displaced. Buell also adduces evidence that the OM in Zulu is an agreement marker, departing from other linguists of Bantu who argue that OM is pronominal (Bresnan and Mchombo, 1987). His argument is an interesting one that raises more questions about possible interpretive differences between agreement markers and pronouns, but it is not clear that it is decisive. In Zulu, the second conjunct of a coordinated sentence must take OM such as: (28)

a. Ngi-dl-e a-mahhabula a-mabili no Sipho u- wa1S. eat PERF 6-apple 6.REL and 1-Sipho 1.SM - 6.OM dl- ile. eat- PERF (lit. ‘I ate two apples, and Sipho ate them, too’) (Buell, ex 82, 2005, p. 52)

Buell points out that this does not mean Sipho ate the same two apples, in fact, it means Sipho ate his own. Compare this with the overt English pronoun: (29)

I ate two apples and Sipho ate them too.

Comparing the status of subject agreement markers in Chichewa and Nairobi Swahili, Deen (2006) argues that, unlike in Chichewa, subject questions are possible in Nairobi Swahili, along with other diagnostic differences between the two languages. Deen concludes that the SA marker in Nairobi Swahili is not a pronominal clitic, though it may be ambiguous in other dialects of Swahili (Keach, 1995). There are many complex arguments about the nature of SM in Bantu languages, therefore it is impossible to do justice to them here. However, it is clear that the matter is not settled for even one language at this point, and the languages may indeed differ. The literature suggests at least three possible mechanisms of SM for Xhosa, each with their strengths and problems. One mechanism, proposed by Du Plessis and Visser (1998), is that the subject noun is merged VP internally, and the noun class marker is provided in the lexicon. The subject noun then moves to the Specifier of AgrS. The verb moves from its base position into Tense and then to AgrS, where subject agreement is then dictated by the noun class of the subject noun (see Figure 6.3). In keeping with that proposal are the facts in Xhosa about subject agreement on compound verbs, as in Zulu in Buell (2005).

119

Number agreement in the acquisition of English and Xhosa CP

AgrSP

AgrS

Spec Isilumkoi

AgrS

Abbreviated to avoid AgrO complexities

TP

sithandaj

VP

V

NP ti V tj

NP iincwadi

Figure 6.3 Tree diagram of derivation of subject agreement in Xhosa

In contrast, a pronominal account like that given for Chichewa might be that the SM itself occupies Spec of Agr-S, displacing the subject noun into a topic position. That would be compatible with the lack of ordinary subject wh-questions in Xhosa (Du Plessis and Visser, 1998; Zeller, 2008). In a very recent paper, Zeller (2008) proposes that the SM is part of a “big DP”, in which the SM is the head of the preverbal subject, and takes the subject DP as its complement (which could be null). The SM then incorporates with the verb in T, as a pronominal clitic. But without agreement, what motivates subject movement? Zeller claims that SM is an anti-focus marker in order to rationalize its move from VP. This account attempts to explain a variety of complex facts about Bantu languages, for Zulu in particular and by extension, Xhosa.

120

Formal features

As can be seen from this condensed review, the matter of the nature of subject agreement as pronominal or agreement marker is far from clear in Xhosa. Can child data help in distinguishing the alternatives?

6.9 What can child language tell us? In Gxilishe et al. (2007) a question is explored about the nature of subject agreement in Xhosa child language. Du Plessis and Visser (1998) argue that the morpheme in Xhosa is a subject agreement marker with either an explicit or deleted subject. The children’s data are certainly compatible with the latter position, in that there is no difference in the likelihood of supplying the subject agreement on the verb as a function of whether the subject is overt or deleted. Deen (2006) draws a similar conclusion for the status of the agreement prefix in child data from the Nairobi dialect of Swahili. Yet with respect to subject wh-questions, Xhosa behaves like Chichewa, and it remains possible that the status of SM may be pronominal in character. If so, it may more readily carry notional number, in concord with the subject noun, perhaps by virtue of its position in Spec-Agr (see also Zeller, 2008). In contrast, the suffixes in a language like Italian or Spanish cannot raise to Spec-Agr, and can only carry grammatical number by syntactic agreement. Furthermore, Bobaljik’s claim about the post-syntactic nature of morphological agreement is assumed not to apply to pronominal clitics. Nothing in the interesting work of Eberhard, Cutting, and Bock connects their production model to particular alternative grammatical configurations in a generative framework, but a bridge may be possible. Recall that young English children cannot seem to use the agreement marker on the verb to establish subject number. The possibility that this is because the information is usually redundantly marked is contradicted by data from Pérez-Leroux (2006) on Spanish-speaking children, who also cannot use the verb marking even when the subject noun is pro-dropped (see also Arosio et al., this volume). But it is as yet unexplored whether children speaking Xhosa could use the subject agreement marking on the verb to determine subject number. It is a much more difficult problem in the case of Xhosa, given the variety of forms and their dependence on noun class. However, if the forms behave more like pronouns, perhaps children will be able to judge number from these forms. We have begun pilot research with stimuli very like those used in Johnson et al. (2005) and Pérez-Leroux (2006). One such study is designed as picture choice as in the study in English by Johnson (2005). All stimuli are common enough nouns and verbs that three- to six-year-old Xhosa speakers should

121

Number agreement in the acquisition of English and Xhosa Table 6.2 Pilot studies of subject number agreement comprehension in Xhosa Xhosa sentence

English gloss

Choices: correct in bold

1. a-dlala kuswingi 2. ba-thetha efonini 3. i-dlala emanzini 4. li-tshaya estratweni 5. zi-lala ebhedini 6. u-nukisa amablomu

‘they swing on the swing’ ‘they talk on the phone’ ‘he plays in the water’ ‘he smokes in the street’ ‘they sleep on the bed’ ‘it sniffs at the flowers’

2 girls 2 women 2 boys 2 police 2 cats 2 bunnies

1 girl 1 woman 1 boy 1 police 1 cat 1 bunny

know (see Table 6.2). The question is, will young Xhosa-speaking children be able to retrieve the subject number in keeping with their production? If they behave like English speakers and show a production/comprehension asymmetry, then perhaps that would count as evidence that the markers are indeed agreement affixes, and possibly 2 their targets therefore inaccessible to interpretation of number. If children can retrieve number from the markers, despite the complexity of form mapping, then perhaps it can add to the arguments on behalf of SM as a pronominal clitic in some Bantu languages. We have also begun to explore the potential contrast with object agreement, for which there is much more consensus that the OM is pronominal in form. For instance, the object in a transitive sentence is displaced outside the prosodic envelope of the verb phrase when OM is present (see Van der Spuy, 1993; Buell, 2006; also Gxilishe, de Villiers, and de Villiers, 2007 on Xhosa children’s language). If children can retrieve number information about the object from OM, but not from SM, that will give support to other linguistic arguments that the two forms are different in type. The task is very similar. The child sees for instance two pictures, one in which a woman is watering a single flower, and one in which she is watering three flowers. After saying about the pair of pictures: (30) Jonga . . . Oomama, . . . amablomu See 2a-women, 6-flowers ‘See? Women, flowers.’ the child is asked to “show the picture where”: (31) Umama u-ya-wa -nkcenkceshel-a 1a-Woman 1a-SM-TNS- 6-OM waters - M ‘The woman waters them.’ 2 Only possibly, because, for example, Buell (2005) would have subject markers heading their own projections in the syntax, and therefore not being generated post-syntactically as in Bobaljik (2006).

122

Formal features

versus (32) Umama u-ya-li-nkcenkceshel-a 1a-Woman 1a-SM-TNS- 5-OM waters - M ‘The woman waters it.’ Since the lexical object is dropped, the only clue to object number is contained in the object agreement marker (Class 5 or 6) in preverbal position. Will it behave like a pronoun and allow number to be accessed? We 3 have tested eight children aged four and five years, all native Xhosa speakers, in a small day care center in a township in the Western Cape Province of South Africa, three girls and five boys. Each received six test examples of subject agreement and six of object agreement, after checking that they would respond by pointing to pairs of simple pictures. In no case did any child show mastery of the number properties of subject marking, that is, they did not use the marker to recover the subject number (average score 3/6 correct, no discrimination between singular and plurals as in Table 6.2). However, neither did they use the object marking as a cue to object number (average score 3.36/6) on examples like (31) and (32). It is premature to draw a firm conclusion at this stage, especially about the status of object agreement, as there are no data on when object agreement enters Xhosa children’s speech production. However, it seems likely that young Xhosa speakers, like their English, Spanish, and Italian counterparts, may produce subject agreement successfully at age three but fail to interpret it as a cue to number of the subject for several years thereafter. Such a finding might then give credence to arguments that the subject marker in Xhosa is indeed a post-syntactic, morphological agreement, and not a pronominal. Although child language data is rarely used to arbitrate between different theories of the adult language, it is the continual hope of child language researchers that data from children may play some useful role in theoretical accounts. 3 Many thanks are due to Dr. Rose Mantoa Smouse, Thabisa Xhalisa, and Nolubabalo Tyam of the University of Cape Town, Clara Feldmanstern of Smith College, and the staff and children of the Kaya Mandi crèche. A full study is under way.

7 Variable vs. consistent input: comprehension of plural morphology and verbal agreement in children∗ KAREN MILLER AND CRISTINA SCHMIT T

7.1 Introduction The process of language acquisition is often represented in terms of the following equation: Language Acquisition Device + Input = L1. The term input refers to the speech of speakers in the child’s language community while the term Language Acquisition Device refers to the innate component that allows humans to acquire language. While it is generally assumed that the innate component is invariable across typically-developing human populations (all typically-developing humans have the ability to acquire language), we know that the input (i.e. the speech of the speakers with whom the child interacts) varies within and across speakers. Exactly how the input determines the grammar that children initially construct is not yet well understood. Several studies have examined the effect of frequencies in the input on language acquisition (Brown 1973; Valian 1991; ∗ This study was funded by NSF Doctoral Dissertation Improvement Grant #0446769, NSF Grant BCS-0126502, and the Michigan State University Graduate Student Research Enhancement Award. We thank the following schools in Punta Arenas, Chile: Colegio Alemán, Colegio Británico, Colegio Pierre Faure, Junta Nacional de Jardines Infantiles de Chile (JUNJI), Jardín Bambi, Jardín Las Charitas, Escuela 18 de Septiembre, Colegio Miguel de Cervantes, and Universidad de Magallanes, and also in Mexico: Centro de Desarrollo Infantil (CENDI) and the Universidad Autónoma Metropolitana (UAM), Campus Iztapalapa of México, D.F. We especially thank John Grinstead and Antoinette Hawayek for their assistance in Mexico City and for their comments and suggestions on this work. Thanks to the following research assistants: Rodrigo Cárdenas, Cynthia Corona, Marena García, Katerina French, Edgardo Mansilla, Erika Mendoza, Andrew Sanford, Heriberto Sierra, and Pascale Schnitzer. Finally, we thank Ana Pérez-Leroux, Alan Munn, and the members of the Michigan State University Language Acquisition Lab.

124

Formal features

Wang et al. 1992; Kupisch 2003); however, studies examining the effect of different types of input (e.g. inconsistent input vs. consistent input; variable input vs. consistent input) on language acquisition are very few in number, in spite of the fact that, as Wilson and Henry (1998) note, the input into the emerging linguistic system is variable, even within a monolingual context. Any theory of language acquisition must account for the fact that a key part of the language acquisition device (LAD) is designed to enable it to cope with this variability, which may cause unreliability in the input. Unreliability in the input can arise from at least two sources. It could be caused by inconsistency in the speech of adult speakers. Inconsistent input is neither linguistically nor extra-linguistically predictable (see Hudson Kam and Newport 2005). One finds this type of input coming from non-native speakers of a language to their children. On the other hand, unreliability could be caused by sociolinguistic variation in the speech of adult speakers. In both cases the unreliability arises when the input provides evidence both for (the adult produces a particular form) and against (the adult omits the form) a particular form in the grammar the child is acquiring. Studies have reported that learners show a tendency to regularize inconsistent input (see Hudson Kam and Newport 2005 and also Singleton and Newport 2004). However, at least in production, studies have shown that children do not regularize variable input but rather tend to show patterns of variability in their own speech (Kovac and Adamson 1981; Labov 1989; Roberts 1994; Smith et al. 2006). Furthermore, a study by Johnson (2005) suggested that variable input causes a delay in the comprehension of grammatical morphology. While these studies provide insight on the effect of different types of input on language acquisition, much more work is needed in this area if we are to clearly understand the language acquisition device. In this chapter, we present two experimental studies that test children’s comprehension of plural morphology in the nominal and verbal domain in the context of variable input. Previous work on the acquisition of number morphology in the nominal domain has revealed that English-speaking children as young as three years of age are sensitive to plural morphology on the noun. They can use it to distinguish between “one” vs. “more than one” (Kouider et al. 2006) and mass vs. count nouns (Barner and Snedeker 2005), although they begin to produce the plural morpheme as early as two years of age (Ferenz and Prasada 2002). Johnson et al. (2005), however, suggest that children do not use the 3rd person singular /s/ as an indication that the subject is to be interpreted as denoting a single individual until much later, at around five years of age (see also de Villiers and Gxilishe, this volume). The finding of interest for the present chapter is that comprehension of plural morphology

Comprehension of plural morphology and verbal agreement

125

in the noun phrase appears to precede comprehension of number agreement on verbs in English-speaking children. In the studies mentioned above, however, English-speaking children were exposed to an input that consistently marked nominal plural morphology and verbal agreement; hence, these studies provide an overview of how acquisition proceeds when the input is consistent and reliable. The goal of the present chapter is to examine whether similar patterns are found in a language where the input is variable; in other words, where the plural morpheme in the noun phrase is often omitted in the speech of adult speakers, making the input with respect to the plural morpheme unreliable (the plural morpheme is sometimes present and sometimes absent in semantically plural noun phrases), yet verbal agreement is consistently produced. Given this different type of input, we ask whether children will still use plural morphology in the noun phrase before verbal agreement in comprehension tasks similar to children exposed to consistent input. This chapter is organized as follows. Section 7.2 provides background information on plural morphology and verbal agreement in Mexican and Chilean Spanish and discusses previous research on the acquisition of verbal agreement and plural morphology in Spanish-speaking and English-speaking children. Section 7.3 presents an experimental study (Experiment 1) on the comprehension on plural morphology in the noun phrase in Chilean and Mexican Spanish-speaking children. Section 7.4 presents an experimental study (Experiment 2) that tests whether Chilean Spanish-speaking children can use verbal agreement to interpret number on the subject. Finally, Section 7.5 provides a summary of the results and a conclusion.

7.2 Linguistic and acquisition background 7.2.1 Number marking in Chilean Spanish vs. Mexican Spanish The experimental studies presented in this chapter examine language acquisition in two varieties of Spanish: Mexican Spanish (of Mexico City) and Chilean Spanish (of Punta Arenas, Chile). In the first variety, plural morphology is consistently produced on all elements within the noun phrase (in D, N, and A). This is shown in (1). (1)

a. La niña está saltando. The.sg girl.sg is.3.sg jumping ‘The girl is jumping.’ b. Las niñas están saltando. The.pl girls.pl are.3.pl jumping ‘The girls are jumping.’

Pronunciation [la]/[niña]

[laz][niñas]

126

Formal features

The examples in (1) illustrate that in Mexican Spanish the plural morpheme in the noun phrase undergoes a process of assimilation where /s/ occurs as [z] before voiced consonants and as [s] before voiceless consonants, vowels, and pauses. The distribution of [s] and [z] is categorical. What is important here is that in the Mexican Spanish (Mexico City) variety the plural morpheme is always pronounced as an alveolar fricative and is never omitted in the speech of adult Mexico City speakers. On the other hand, the phonological form of the plural morpheme in Chilean Spanish undergoes a process of lenition. In this dialect all syllable final /s/ is pronounced as [s], [h], or is omitted (zero). The phonological variant ([s], [h], or zero) that surfaces is dependent on both linguistic and extra-linguistic factors. Some of the linguistic factors include phonological environment and syntactic category and some of the extra-linguistic factors include socioeconomic status, gender, and age. Because the plural morpheme occurs as /s/ in syllable-final position, this process of lenition affects the pronunciation of the plural morpheme in the noun phrase as well (Cepeda 1995; Miller and Schmitt 2006; Miller 2007). This is shown in (2).

(2)

a. La niña está saltando. The.sg girl.sg is.3.sg jumping ‘The girl is jumping.’ niñas están saltando. b. Las The.pl girls.pl are.3.pl jumping ‘The girls are jumping.’

Pronunciation [la] [niña]

[las/lah/la] [niñas/niñah/niña]

The examples in (2) illustrate that there is possible overlap in the pronunciation of semantically plural and semantically singular determiners and nouns, as indicated in bold and underlining (i.e. both [la] and [niña] can be used to describe semantically singular and semantically plural nouns). This creates unreliability in the input that Chilean children are exposed to. However, in both Mexican and Chilean Spanish verbal agreement is consistently produced. 1

1 In Chilean Spanish verbal morphology is consistently produced in adult speech except when the verbal morpheme is represented as /s/ and occurs in word-final position. This is the case with the 2nd person singular (e.g. estás ‘be.2.sg’ can be pronounced as [estas], [estah], or [esta]. The pronunciation [esta] overlaps in form with the 3rd person singular (está ‘be.3.sg’ is also pronounced as [esta]). However, this fact is not relevant for the present chapter, as we did not test child comprehension of the 2nd person singular vs. 3rd person singular verb forms.


127

Figure 7.1 Experimental paradigm Source: adapted from Pérez-Leroux 2005

7.2.2 Nominal number and verbal agreement in child language Research has indicated that Spanish-speaking children who are presented with consistent input for plural morphology in the noun phrase begin producing the plural morpheme at about 2;0 years of age (Kvaal et al. 1988; Marrero and Aguirre 2003), which parallels findings for English-speaking children (Cazden 1968; Mervis and Johnson 1991; Ferenz and Prasada 2002). Research on children’s comprehension of plural morphology has shown that Spanish-speaking children exposed to consistent input associate the plural morpheme in noun phrases to an interpretation of “more than one” by at least 3;5 years of age (Miller and Schmitt 2006; Munn et al. 2006; Miller 2007), which is consistent with what has been found for English-speaking children (Kouider et al. 2006, Munn et al. 2006). However, as far as we know, there is no research examining comprehension of plural morphology in Spanish-speaking children who are younger than 3;0 years of age and, for this reason, we do not know whether production of the plural morpheme precedes comprehension in Spanish as it appears to do in English. With respect to verbal agreement, we know that Spanish-speaking children begin to produce verbal agreement by at least 2;5 years of age (Durán 2000; Grinstead 2000; Félix-Brasdefer 2006). Similar findings were reported for English-speaking children (Brown 1973). However, comprehension studies have suggested that Spanish-speaking children are unable to use verbal agreement to interpret number on the subject until around five years of age (PérezLeroux 2005), similar to what has been reported for English-speaking children (Johnson et al. 2005). Pérez-Leroux (2005), for example, tested 23 three- to six-year-old Dominican Spanish-speaking children on their comprehension of sentences as in (3) in the context of Figure 7.1. The task of the child was to choose the picture that best represented the experimental sentence.

128 (3)

Formal features a. Duerme en la cama. sleeps.3.sg in the bed ‘(It) sleeps in the bed.’ b. Duermen en la cama. sleep.3.pl in the bed ‘(They) sleep in the bed.’

Note that in Spanish the subject can be null, as illustrated in (3). For this reason, in (3) Spanish-speaking children must rely solely on verbal agreement when interpreting number on the subject. The results of this study revealed that, while 3;2–4;5 year old children did not use verbal agreement to determine number on the subject noun phrase, 4;8–6;6 year old children did so 67% of the time when the verb was inflected for 3rd person plural (but not 3rd person singular). 2 These data are comparable with findings for English-speaking children (Johnson et al. 2005), which indicate that four- to five-year-old Englishspeaking children are able to use 3rd person singular (but not 3rd person plural) verbal morphology to determine number on the subject noun phrase between 74% and 79% of the time. It is important to note that in Spanish it is the 3rd person plural that is realized morphologically, while the 3rd person singular has no overt morphological realization (e.g. duerme ‘sleeps.3.sg’ vs. duermen ‘sleep.3.pl’). The opposite is true for English, where the 3rd person singular is realized morphologically, while the 3rd person plural is not (e.g. he sleeps vs. they sleep). This may explain why English-speaking children performed better on the 3rd person singular and Spanish-speaking children on the 3rd person plural. The results of this study suggest that Spanish and English-speaking children cannot use verbal morphology in comprehension until about five years of age. Taken together, the above studies indicate that Spanish-speaking children, who are exposed to consistent input, can use plural morphology in the noun phrase to distinguish between “one” vs. “more than one” much earlier than they can use verbal agreement to make this distinction. The purpose of the following two experimental studies is to test the Variability Delay Hypothesis (Miller 2007). The idea is that children exposed to variable input will have a delay in their comprehension of grammatical morphology that is affected by this variability. The Variability Delay Hypothesis is stated in (4). (4)

Variability Delay Hypothesis (based on Yang 2002): Variability in the input will delay child comprehension of grammatical morphemes when

2 These results have to be interpreted with caution because plural morphology, subject–verb agreement and the use of null subjects may be subject to variation in Dominican Republic Spanish (Poplack 1980; Lipski 1994b; Toribio 1994; Morgan 1998).


129

the variability causes unreliability in the input (the input involves a zero form) and is constrained not only by linguistic (phonological, grammatical) but also extra-linguistic (SES, age, sex) factors. This hypothesis is adapted from Yang’s (2002) Variation Model of language acquisition, which proposes that the cumulative effect of the input combined with a theory of a restricted search space can explain language acquisition. According to Yang, children make hypotheses within the limits of UG that are punished or rewarded depending on their ability to account for particular properties of the input. If the input is reliable and frequent, acquisition happens early. If input is unreliable, the child may take longer to set a parameter. If the Variability Delay Hypothesis is supported, we may find, contrary to what has been reported previously in the literature, that Chilean Spanish-speaking children can use verbal agreement before they can use nominal plural morphology to make the distinction between “one” vs. “more than one” because plural morphology in the noun phrase is variable and unreliable in the input to Chilean children.

7.3 Experiment 1: Comprehension of plural morphology in the noun phrase Experiment 1 is designed to test Mexican and Chilean Spanish-speaking children’s ability to use plural morphology in the noun phrase to distinguish “one” vs. “more than one”. It is possible that unreliable input has no effect on the acquisition of plural morphology, in contrast to what seems to have happened in a previous verbal agreement study (Johnson 2005). In other words, as long as the adult speaker produces the plural morpheme on semantically plural nouns some of the time in their speech to children (as is the case for the Chilean children, see Miller 2007), the child will initially construct a grammar that associates the plural morpheme to an interpretation of “more than one”. In this case, no differences would be found between Mexican vs. Chilean children in their comprehension of plural morphology. On the other hand, it may be the case that unreliable input causes a delay in the comprehension of plural morphology because the child is receiving evidence both for ([s] and [h]) and against (zero marking) nominal plural morphology in the grammar they are acquiring. This prediction is consistent with the Variability Delay Hypothesis and Yang (2002). With respect to plural morphology, we would expect that the child may initially depend on some other element in the input that is more reliable for number marking (e.g. quantifiers, numerals) and may not initially associate the plural morpheme

130

Formal features

Figure 7.2 Experiment 1: Sample target trial

in the noun phrase to an underlying representation of [PLURAL] because it is not a reliable marker in the input the child is exposed to. Hence, we would predict that the Mexican child would associate the plural morpheme to an interpretation of “more than one” before the Chilean child. While both alternatives may seem equally plausible: (1) variable and unreliable input may have no effect on language acquisition vs. (2) variable and unreliable input may cause a delay in language acquisition, there is some empirical evidence for the latter alternative (Moore 1979; Johnson 2005; Miller and Schmitt 2006; Miller 2007). The goal of Experiment 1 is to test the Variability Delay Hypothesis by determining whether both Chilean and Mexican children associate the plural morpheme /s/ in indefinite plural noun phrases (e.g. unos “some.m.pl”; unas “some.f.pl”) to an interpretation of “more than one”. 7.3.1 Method and design Experiment 1 used a picture matching task to examine child comprehension of singular and plural indefinites, as in (5), in the context of Figure 7.2. (5)

a. ¿En cuál de las dos tarjetas hay una botella? On which of the two cards exst a/one.sg bottle.sg ‘On which of the two cards is there a/one bottle?’ botellas? b. ¿En cuál de las dos tarjetas hay unas On which of the two cards exst some.pl bottles.pl ‘On which of the two cards are there some bottles?’


131

The indefinite in (5a) is singular and the indefinite in (5b) is plural. The existential verb hay (“there is/there are”) was used because it does not carry number information that could be associated with the subject. It can be used with both plural and singular nouns. For this reason, the only number information in (5a) and (5b) is the plural morpheme in the indefinite noun phrase. The plural morpheme was always pronounced as [s]. Chilean children who consistently chose the singular picture in the plural condition were tested one to two weeks later with the plural morpheme pronounced as [h]. There were four trials of the plural condition, four of the singular condition, and four fillers from another experiment testing child comprehension of the Spanish copulas ser and estar. In the eight experimental trials the initial sound and gender of each target word was controlled for Chilean subjects: burros ‘donkeys’, monos ‘monkeys’, barcos ‘boats’, martillos ‘hammers’, bolitas ‘marbles’, manzanas ‘apples’, botellas ‘bottles’, monedas ‘coins’. The same words were used for Mexican children except changos was used for ‘monkeys’ and canicas was used for ‘marbles’ so that we could continue to use the same materials yet accommodate to the Mexican Spanish lexicon. In addition, half of the indefinites were feminine and half were masculine. In the feminine indefinites, only the plural morpheme provides number information (e.g. una bolita ‘a/one.f.sg marble.f.sg’ vs. unas bolitas ‘some.f.pl marbles.f.pl’). In masculine indefinites the form of the determiner is also different in the singular vs. plural conditions (e.g. un burro ‘a/one.m.sg donkey.m.sg’ vs. unos burros ‘some.m.pl donkeys.m.pl’). All subjects were tested by native speakers of Spanish who lived in the same city as the subjects. Controls were un solo ‘only one’ and muchos ‘many’. The controls were administered after the target questions so that un solo ‘only one’ would not provide any information to the child about the interpretation of un ‘a/one’. In addition, placement of cards (singular card above plural card vs. plural card above singular card) was controlled for. 7.3.2 Subjects Fifty children participated in this study: 19 Mexican working-class (4;11–6;2, Mean Age: 5;4), 17 Chilean working-class (4;9–6;4, Mean Age: 5;5), 10 Chilean middle-class (4;10–6;4, Mean Age: 5;5) children. In addition, 22 Chilean adults and 8 Mexican adults participated in this study. Both working-class and middle-class Chilean children were tested because previous research on syllable-final /s/ lenition in Chilean Spanish has found that working-class adults omit syllable-final /s/ more often than middle-class adults (Cepeda

132

Formal features 100

100 79

80 60

unas (‘some. pl’) una (‘a/one.sg’)

40

35

33

20 0

6

0

0 Adults

Mex WC

ChMC

1 ChMC

Figure 7.3 Experiment 1: Percentage of plural responses

2005; Miller and Schmitt 2006; Miller 2007). Only working-class Mexican children were tested because syllable-final /s/ lenition does not occur in the speech of Mexican adults from Mexico City (Canfield 1982; Lipski 1994a; Morgan 1998). The Chilean children were recruited from schools in Punta Arenas, Chile, and the Mexican children were recruited from a daycare in Mexico City. All children were in preschool and kindergarten. Chilean adults were undergraduates at the Universidad de Magallanes in Punta Arenas, Chile, and the Mexican adults were undergraduates at the Universidad Autónoma Metropolitana de Iztapalapa in Mexico City. 7.3.3 Results and Discussion Although all three child groups performed the same on the controls, always associating un solo (‘only one’) with an interpretation of “one” and muchos (‘many’) with an interpretation of ‘more than one’, they did not perform the same in the target conditions. The dependent variable was the number of plural responses children gave. Choosing the card with more than one item was considered a plural response. Choosing the card with only one item was considered a singular response. The mistakes that children made in the plural conditions were always the same, they chose the picture with only one item. Figure 7.3 shows the percentage of plural responses in the singular and plural indefinite conditions when the plural morpheme was pronounced as [s]. Within every child group there were children who associated the plural indefinite unos (‘some.pl’) to an interpretation of “more than one”; however, the groups differed significantly in how many children treated unos as plural.


133

It is important to note that most children were systematic in their response patterns, either always associating the plural indefinite to “more than one” in all four trials or never doing so. The number of plural responses in the plural indefinite unos condition for each child was entered into a one-way ANOVA (adults, MexWC, ChMC, ChWC). The results showed a significant difference between the four groups (F (3,74) = 20.210, p < .001). Post hoc Bonferonni tests showed that only ChMC (p < .001) and ChWC ( p < .001) children, but not MexWC ( p = .092) children, differed significantly from adults in the number of plural responses assigned to the plural indefinite. MexWC children also differed significantly from ChWC ( p < .05) and ChMC ( p < .05) children but there were no significant differences between the two Chilean child groups ( p = 1.0). Between one and two weeks after this initial experiment was carried out with the plural morpheme pronounced as [s], Chilean children who systematically assigned a singular interpretation to the plural indefinite were tested again but this time the plural morpheme was pronounced as [h]. Eleven ChWC children and 7 ChMC children participated in this part of the experiment. The behavior of each child remained the same. The 11 ChWC children continued to choose the singular card in the plural condition 95% of the time and the 8 ChMC did so 97% of the time. Paired samples t-test showed that there was no significant improvement either for the 11 ChWC children (t(1,10) = −1.00, p = .343) or the 7 ChMC children (t(1,8) = −.552, p = .598). The results show that, given the same experimental conditions, five-yearold Mexican children associate the plural indefinite to an interpretation of “more than one” much more often than five-year-old Chilean children, regardless of whether the plural is pronounced as [s] or [h] for the Chilean children, which suggests that several five-year-old Chilean children match neither [s] nor [h] to an underlying representation for [PLURAL]. This data supports the Variability Delay Hypothesis that variable and unreliable input causes a delay in the acquisition of grammatical morphology. Given that several five-year-old Chilean children associate both the plural and singular indefinite (una ‘a/one.sg’ and unas ‘some.pl’) to an interpretation of “one”, the next question is whether Chilean children can use verbal agreement in a subject relative clause to determine the number of the head of the relative clause.

7.4 Experiment 2: Comprehension of verbal agreement The goal of Experiment 2 is to test whether Chilean children can use verbal agreement, which is much more reliable in the input, to make a distinction

134

Formal features

Figure 7.4 Experiment 2: Sample target trial

between “one” vs. “more than one”. According to the Variability Delay Hypothesis we predict that Chilean children will be able to use verbal agreement earlier than nominal plural morphology because it is more reliable in the input that they are exposed to. If Chilean children pattern like Englishspeaking and Dominican Spanish-speaking children, then they too should be able to use verbal agreement in comprehension by five years of age. In this case, Chilean comprehension of verbal agreement would precede comprehension of plural morphology because verbal agreement is consistently produced in Chilean adult Spanish but plural morphology in the noun phrase is variable and unreliable. 7.4.1 Methods and design Experiment 2 used a picture matching task to examine child comprehension of verbal agreement, as in (6), in the context of Figure 7.4. (6)

a. ¿En cuál de las dos tarjetas hay una niña que On which of the two cards exst a/one.sg girl.sg that está saltando? is.3.sg jumping ‘On which of the two cards is there a/one girl that is jumping?’ niñas que b. ¿En cuál de las dos tarjetas hay unas On which of the two cards exst some.pl girls.pl that están saltando? are.3.pl jumping ‘On which of the two cards are there some girls that are jumping?’

In (6a) the head of the subject relative clause is singular and the verb agrees with it. In (6b) the head of the relative clause is plural and the verb shows


135

agreement with the 3rd person plural. Unlike the previous experiment, now we have both verbal agreement and plural morphology on the noun phrase marking an interpretation of “more than one”. There were three target trials of the singular sentence, as in (6a), and three target trials of the plural sentence, as in (6b). The plural morpheme was always pronounced as [s] on the noun and determiner and the indefinite noun phrases were always feminine. All subjects were tested by native speakers of Spanish who lived in the same city as the subjects. Controls were una sola ‘only one’ and dos ‘two’. 7.4.2 Subjects Thirteen Chilean working-class children (4;5–6;0, Mean Age: 5;1) participated in this study. Only Chilean working-class children were tested because they overwhelmingly treated the plural indefinite in Experiment 1 as singular and because previous research has indicated that Chilean working-class adults omit syllable-final /s/ more often than Chilean middle-class adults (Cepeda 1995; Miller and Schmitt 2006; Miller 2007). In addition, 12 Chilean undergraduate students from the Universidad de Magallanes participated in a written version of this test. 7.4.3 Results and Discussion Chilean children and adults performed the same on controls, always treating una sola (‘only one’) as singular and dos (‘two’) as plural. With respect to the target trials, Figure 7.5 shows the percentage of time children and adults assigned a plural interpretation to the plural sentence (5b) and the singular sentence (5a). The dependent variable was the number of plural responses. Choosing the card with the plural set of characters was considered a plural response. Choosing the card with only one character was considered a singular response. The results showed a significant difference between the children and adults (F (2,23)=19.546, p < .001), which indicates that, similar to Experiment 1, ChWC children did not reach adult levels. However, the number of plural responses in the plural indefinite condition were also tested for chance behavior and the results revealed that, unlike Experiment 1, ChWC children chose the plural picture in the plural condition significantly more often than chance (t(1,12)=2.607, p < .05) (chance = 50%). Although ChWC children did not reach adult levels in Experiment 2, the data, taken together with the findings of Experiment 1, nevertheless indicate that ChWC children are able to use verbal agreement in comprehension tasks

136

Formal features 100

100 77

80 60

Adults ChWC

40 20 0

5

0 están.3.pl

está.3.sg

Figure 7.5 Experiment 2: Percentage of plural responses

before they can use plural morphology in the noun phrase. In addition, the results for Chilean Spanish-speaking children pattern with those for Englishspeaking and Dominican Spanish-speaking children, which indicated that by five years of age children can use verbal agreement to interpret number on the subject. However, unlike the results found for Dominican Spanish-speaking children, Chilean children performed well on both the plural and singular forms of verbal agreement (e.g. está ‘is.3.sg’, están ‘are.3.pl’). One important difference between the experimental design of the Dominican Spanish study (Pérez-Leroux 2005) and the study presented in this chapter is that the former tested definite noun phrases, while the present study tested indefinite noun phrases. Given that Chilean children overwhelmingly prefer a singular reading for both plural and singular indefinite noun phrases (as revealed in Experiment 1), it is not surprising that they correctly associate the singular indefinite to a singular interpretation in Experiment 2, especially given the fact that there is no verbal morphology overtly realized on the verb. What is interesting is that the presence of the 3rd person plural marker (están ‘are.3.pl’) causes Chilean children to associate the plural indefinite to an interpretation of “more than one” and this suggests very strongly that Chilean children can use verbal agreement to interpret the appropriate interpretation of number on the subject by 4;5 years of age.

7.5 General discussion The data presented in Experiment 1 indicate that five-year-old Chilean children have a delay in their comprehension of plural morphology in indefinite


137

noun phrases, which is consistent with the Variability Delay Hypothesis. The data presented in Experiment 2 revealed that 4;5-year-old Chilean children can use verbal agreement to assign number to indefinite subjects. Contrary to what has been previously reported, the data from Chilean Spanish-speaking children suggest that verbal agreement is used more efficiently in comprehension before plural morphology in indefinite noun phrases. We can speculate that the reason children exhibit a delay in their comprehension of plural morphology in the noun phrase is because they entertain for a longer period the hypothesis that they are acquiring a grammar without overt number marking in the noun phrase. The interesting question for further research is how much variability and unreliability can the learner tolerate before they will settle on a grammar that is different from the adult grammar.

8 Grammatical features in the comprehension of Italian relative clauses by children FABRIZIO AROSIO, FLAVIA ADANI, AND MARIA TERESA GUASTI

8.1 Introduction 1 Relative clauses are widely studied in the psycholinguistic literature since they represent an interesting challenge for parsing strategies. Given the left to right incremental course of parsing, it is assumed that, whenever the parser hits a relative pronoun or the complementizer that following an NP, it postulates a relative clause (RC) containing the trace of the relative pronoun as given in (1) and (2) below. (1) The woman [who/that twho is watching the clown]RC (2) The woman [who/that the clown is watching twho ]RC

SUBJECT OBJECT

One important question addressed over the years concerns the question of which strategies the parser follows in reconstructing the movement relation or the filler gap dependency between the relative pronoun and its trace, especially in light of the fact that (1) and (2) are locally ambiguous, being identical up to the relative pronoun. Studies on adult processing have established 1 The facts discussed in this chapter were presented at Gala 2005 in Siena, at the Glow Workshop 2006 held in Barcelona, at the GLOW Summer school held in Stuttgart, at the University of TrentoRovereto, at the Symposium for Marica De Vincenzi held at the University of Chieti, and at the University of Siena. We would like to thank the audience of these events. We would also like to thank Lyn Frazier for insightful discussion, Adriana Belletti, Ivano Caponigro, Carlo Cecchetto, Francesca Foppolo, Carlo Geraci, and Luigi Rizzi, Stavroula Stavrakaki for various comments. Finally, we would like to express our gratitude to Francesca Citron for assistance in data collection. This research was supported by a grant from the Italian Ministry of University and Research (PRIN 2003). Although the experiments have been conceived jointly and the chapter has been written jointly, for the purposes of the Italian Academy, Fabrizio Arosio takes responsibility for sections 8.4.3, 8.4.4, 8.4.5, 8.5.1, and 8.6, Flavia Adani for sections 8.2, 8.3, 8.4, 8.4.1, 8.4.2, and Maria Teresa Guasti for sections 8.1, 8.5.2, 8.5.3.

Grammatical features in Italian relative clauses

139

that subject RCs are easier to comprehend than object RCs in a variety of languages regardless of whether they are temporally ambiguous or not (e.g. Frauenfelder, Segui, and Mehler, 1980, for French; King and Kutas, 1995, for English; Schriefers, Friederici, and Kuehn, 1995, for Dutch). According to serial syntactic-based parsing strategies motivated by human limitations of the computational resources, these findings have been explained in terms of economy principles of gap prediction that drive the analysis of filler gap dependencies (Frazier and D’Arcais, 1989; De Vincenzi, 1991). As we see in (1), the trace of the relative pronoun in a subject RC is adjacent to the pronoun, while in the object RCs it is separated at least by the embedded subject and an embedded verb, as shown in (2); therefore, the filler gap distance in subject RCs is shorter than in object RCs in English. Since shorter dependencies are computationally less demanding than longer dependencies and parsing strategies are driven by principles of economy, when the parser sees a relative pronoun following an NP it postulates a RC containing the trace of the relative pronoun in the embedded subject position as in (1), in agreement with the Active Filler Hypothesis (AFH, Frazier and D’Arcais, 1989) or the Minimal Chain Principle (MCP, De Vincenzi, 1991). So far so good for subject RCs, where the input following the relative pronoun is compatible with the parsed structure; but what happens when the following input contradicts the ongoing parsed structure as in the case of the object RC in (2)? Assuming a serial model of syntactic sentence processing (Frazier, 1978; Frazier and Rayner, 1982), a garden path effect will arise. The presence of a NP in preverbal position in (2) will trigger a reanalysis that results in the placement of the trace of the relative pronoun in the embedded object position. The detection of the (temporal) incongruity and the revision of the structure in object RC processing is computationally costly; moreover, the filler gap distance in object RCs is longer than in subject RCs and therefore computationally more demanding. These facts explain why object RCs are more difficult to process than subject RCs. To reiterate, as we can see in (1) and (2), it is the position of the embedded NP that says to the English listener/reader that the sentence is a subject RC or an object RC: when the embedded NP is in the preverbal position we have an object RC, when it is in the postverbal position we have a subject RC. In languages with a relatively free word order like German or Italian, morphology has a fundamental role in the subject–object RC distinction and in these languages a proper analysis of the verbal agreement morphology is crucial for a correct interpretation. Consider the German examples (3) and (4) where the relative pronoun die and the plural definite article die (e.g. in die Kinder, ‘the children’) are morphologically case-ambiguous between nominative and accusative case.

140

Formal features

(3) Die Frau die tdie die Kinder sieht The woman [who twho the children watches]RC ‘The woman who is watching the children.’ (4) Die Frau die die Kinder tdie sehen The woman [who the children tdie watch]RC ‘The woman who the children are watching.’

SUBJECT

OBJECT

Despite their identical word order, 2 (3) should be interpreted as a subject RC since the embedded verb and the head of the RC (henceforth head-NP) die Frau (and not the embedded NP die Kinder) share the same number features, while (4) should be interpreted as an object RC since the embedded verb and the embedded NP die Kinder (and not the head-NP die Frau) share the same number features. Due to the left to right incremental course of parsing, when the parser encounters the relative pronoun die in (3) and (4), it will postulate a RC with the trace of the relative pronoun in the embedded subject position. According to this analysis, the embedded NP die Kinder will be analyzed as the direct object of the embedded verb. When the parser finally sees the embedded verb in (3), it will complete the processing of the sentence by analysing the agreement morphology on the verb as matching the agreement features of the head-NP and its trace. This will fail to happen for (4). In this case, the agreement mismatch (or a temporary ungrammaticality) between the headNP and the embedded verb results in a garden path effect from which the parser can recover through a reanalysis that takes the trace of the head-NP to occupy the embedded object position and reinterpreting the embedded NP as the embedded subject. Agreement morphology is not the only morphological information that German listeners/readers can avail themselves of in order to achieve a correct RC interpretation. Consider the German RCs in (5) and (6) below where the relative pronoun die is again morphologically case-ambiguous between nominative and accusative and the morphology of the embedded verb is always 3rd singular. (5)

Die Frau die den Clown sieht The woman [who theACC clown watches]RC ‘The woman who is watching the clown.’ Clown sieht (6) Die Frau die der The woman [who theNOM clown watches]RC ‘The woman who the clown is watching.’

SUBJECT

OBJECT

2 German subject and object RCs usually share the same word order when the embedded verb is transitive; the order is: head NP < relative pronoun < embedded NP< embedded verb.


141

Again, though (5) and (6) share the same word order, they do not share the same interpretation. While (5) is a subject RC, since the accusative case morphology on the definite article den says that den Clown is the embedded object, (6) is an object RC, since the nominative case morphology on the definite article der says that der Clown is the embedded subject. Given the incremental course of parsing, when the parser encounters the relative pronoun die in (5) and (6), it will postulate a subject RC with the trace in the embedded subject position. This analysis will be compatible with the accusative morphology borne by the embedded NP den Clown which will be analyzed as the direct object of the embedded verb as in (5), but it will be incompatible with the nominative morphology borne by the embedded NP der Clown in (6). In this case, a temporal ungrammaticality due to a case mismatch is detected that results in a garden path effect and triggers a reanalysis according to which the trace of the head-NP occupies the embedded object position and the embedded NP der Clown is interpreted as an embedded subject. Based on German, Bader and Meng (1999), Meng and Bader (2000) observed that, in processing subject–object ambiguities, garden path effects are stronger when disambiguation is obtained through agreement morphology than when it is obtained through case information (see also Fodor and Inoue, 2000). At the same time, ungrammatical sentences displaying an agreement mismatch (e.g. the girls is running) are more easily detected than ungrammatical sentences including a case mismatch. According to the authors, there is a relation between garden path strength and detection of ungrammaticality: the easier it is to detect ungrammaticality in a downright ungrammatical sentence, the harder it is to recover from a garden path induced by the same local ungrammaticality. For example, in parsing a locally ambiguous structure, such as the RCs disambiguated by number agreement in (3) or (4), the parser initially commits to a subject RC analysis. When it encounters the inflected verb, it finds a local ungrammaticality in (6) due to an agreement mismatch with the head-NP that was initially assumed to stand for a subject. Two solutions can be pursued at that point: either the parser reanalyzes the sentence into an object RC or it judges it as ungrammatical and does not start to reanalyze it. Bader and Meng’s results show that it is easier to initiate reanalysis when the relevant disambiguating information is case than when it is agreement. However, agreement errors are more reliably detected than case errors in downright ungrammatical sentences and thus an increased error rate is observed in the latter case with respect to the former during the speeded grammaticality judgment task. Elaborating on this view, which is based on German, we conjecture that this difference in garden path strength, which is also observed in RC processing,

142

Formal features

might depend: (i) on the different points of the RC analysis at which the temporary ungrammaticality triggers reanalysis, the Mismatch Detection Point Hypothesis (MDPH), or (ii) on the different stage at which ungrammaticality is detected, the Mismatch Detection Stage Hypothesis (MDSH). Concerning the first hypothesis, we note that the parser encounters a temporary ungrammaticality due to a case mismatch relatively soon, as shown in (6), where the nominative case on the NP der Clown following the relative pronoun triggers reanalysis. By contrast, a temporary ungrammaticality due to an agreement mismatch is encountered only at the end of the clause, as shown in (4). Thus, the computational cost of the reanalysis due to agreement disambiguation will be larger in comparison to the reanalysis cost due to case disambiguation, since in the former situation the parser should entirely revise the structure parsed so far. This hypothesis, which we called the MDPH, predicts that garden path effects in processing object RCs should be relatively weak in languages in which agreement disambiguation occurs earlier in the clause than in German. The second hypothesis, the MDLH, holds that what counts is the stage of the grammar at which the temporary ungrammaticality is detected. When disambiguation is obtained by case, ungrammaticality detection takes place together with thematic role (re)assignments, a process which is crucial for the semantic interpretation of the clause. When disambiguation is obtained by agreement, a temporary thematic structure for the clause has already been built, and the detection of ungrammaticality takes place during agreement checking, a morphosyntactic operation that does not contribute to the semantic interpretation of the clause. This difference concerning the stage at which case and agreement play a role in the disambiguation of German object RCs (during or after theta role assignment) determines the strength of garden path effects: being case expressed on the argument and directly dependent on the argument structure of the clause has an immediate effect and will trigger reanalysis; being agreement relational and more indirectly dependent on the argument structure of the clause is taken into account at a point when the interpretation of the sentence has already been established; in this case, the parser tends to judge an agreement mismatch as an error and possibly correct it by repairing the verbal agreement morphology. The MDSH predicts that garden path effects due to agreement mismatches should not be affected by the point in the clause in which the mismatch is detected. In this chapter, we will test these two predictions based on the comprehension of Italian RCs by children. Looking at this problem in children might provide some data not available in a fully developed system. Studies on the acquisition of relative clauses have concentrated on the availability of the mechanisms underlying the formation of these clauses. Some studies have shown that children have a hard time comprehending RCs (e.g.


143

Tavakolian, 1981) and on this basis have concluded that children do not build RCs as adults do; other studies have shown that children’s difficulties with RCs can be alleviated if the presuppositions for their use are satisfied (e.g. Hamburger and Crain, 1982; Crain, Mckee, and Emiliani, 1990). Less investigated is the question of how children exploit different grammatical devices towards understanding RCs. In the light of the hypotheses discussed above, we might expect different developmental patterns depending on whether children are more affected by the point at which disambiguation occurs in the sentence or by the stage of the grammar in which disambiguation operates.

8.2 Italian relative clauses: the problem Italian RCs with an embedded NP in postverbal position can be ambiguous between an object and a subject reading. In fact, the Italian RC in (7) can be interpreted as a subject RC or as an object RC with the embedded subject in the postverbal position. (7)

Il ragazzo che guarda il pagliaccio The boy that watch3sg the clown ‘The boy who is watching the clown.’ ‘The boy who the clown is watching.’

SUBJECT OBJECT

Clearly, when the head of the RC and the embedded NP do not share the same number features, the sentence unambiguously conveys a subject or an object reading depending on the number agreement morphology on the embedded verb, as shown in (8) and (9) (8) Il ragazzo che guarda i pagliacci The boy that watch3sg the clowns ‘The boy who is watching the clowns.’ (9) Il ragazzo che guardano i pagliacci The boy that watch3pl the clowns-SUBJ ‘The boy who the clowns are watching.’

SUBJECT

OBJECT

In (8), the head-NP, but not the postverbal NP, agrees with the embedded verb and therefore we have a subject RC; in (9), the postverbal NP, but not the head-NP, agrees with the embedded verb and therefore we have an object RC. Italian speakers can additionally convey the object reading by placing the embedded subject in the preverbal position, as shown by the example below: (10) Il ragazzo che il pagliaccio guarda The boy that the clown watch3sg ‘The boy who the clown is watching.’

OBJECT

144

Formal features

It is the preverbal position of the embedded NP that makes (10) an object RC. As we can see from (9) and (10), object RCs can be unambiguously conveyed by making use of (i) a structural strategy, i.e. the position of the embedded subject as in (10), or (ii) a morphological strategy, i.e. number agreement between the embedded verb and the post verbal NP subject as in (9).

8.3 Italian RCs: predictions Italian RCs are particularly interesting since in Italian two different grammatical devices can cue the “object meaning”. In our study, we examined how children use these cues during development and how our data bear on the two hypotheses discussed in section 8.1. According to the MDPH, the strength of garden path effects depends on the surface point in the clause at which the mismatch or temporal ungrammaticality is encountered. In Italian, disambiguation is provided by the position of the embedded subject or by number agreement morphology. These two pieces of information are available in exactly the same surface position, i.e. after the complementizer che (that), the point at which the parser has engaged in a subject RC analysis. Thus, it is expected that no difference be found in the processing of the two kinds of object RCs in (9) and (10) and in particular that the garden path effect caused by agreement disambiguation, which occurs earlier in the clause than in German, should be relatively weak and similar to the garden path effect caused by case or position. 3 Under the assumption that children’s processing reflects adults’ processing and that what is difficult for adults is difficult for children, too, perhaps in a magnified way, the MDPH predicts that the developmental pattern of the two kinds of object RCs should be very similar, as what counts is just the point at which the temporary ungrammaticality is detected. By contrast, the MDSH anticipates differences in the processing of the two kinds of object RCs. In (10), it is the position of the embedded NP that says to the Italian listener/reader that the sentence is an object RC; in this case, the detection of ungrammaticality takes place together with thematic role (re)assignments. 4 In other words, as happens for case, the relevant disambiguating information is directly coded on the argument. Thus, we expect garden path effects to be mild, like garden path effects induced by case in German, and reanalysis to be promptly triggered. Analogously to German, 3 As a matter of fact, this prediction would also concern adult processing. However, in this chapter, we will not be concerned with adults. 4 In current approaches, the subject receives its thematic role inside the VP and the subject in the preverbal position is in Spec, IP. However, given the configurational definition of subject, once an NP in that position is found, thematic (re)assignment can occur and the thematic role is assigned to the chain between the NP in Spec IP and its trace in the VP.


145

disambiguation in (9) is achieved through number agreement on the verb and the detection of ungrammaticality takes place at a point where an interpretation of the sentence has been established. Garden path effects will be very strong and it will be more costly to perform reanalysis than in the previous case. If what is difficult for adults is also difficult for children, but to a higher degree, we expect the developmental pattern of the two object RCs to be different and RCs disambiguated by number agreement to be more difficult than RCs disambiguated by position.

8.4 Our study To investigate children’s comprehension of restrictive relative clauses, monolingual Italian-speaking children from four age groups and a control group of adults were tested in a series of experiments. Children and adults participated in a picture selection task testing the comprehension of subject and object RCs; three groups of children were also tested in a grammaticality judgment task whose aim was to establish whether participants were sensitive to number agreement mismatches between the subject and a lexical inflected verb. Finally, children underwent a backward repetition span test (Ciccarelli, 1998) to control for memory effect. 8.4.1 Participants In our study we tested 139 Italian monolingual children divided into four age groups and 24 adults, as represented in the table below: 32 children 35 children 36 children 36 adolescents 24 undergraduate students

Mean Age 5;3 Mean Age 7;3 Mean Age 9;1 Mean Age 11;3

The children were recruited from infant schools in Milan, Modena and Como (Italy). Adults were undergraduate students at the Università degli Studi di Milano-Bicocca. 8.4.2 Procedure and materials Before getting started with the experiments, children were familiarized with a puppet who was learning Italian and was asking for their help. They would be asked to play different games that we had prepared with the puppet. Experiments were all carried out in a quiet room in which children were tested

146

Formal features

individually. Adults were tested with the same procedure used for children, except that the interaction was not mediated by the puppet. 8.4.3 The picture selection task: the comprehension of relative clauses In the picture selection task, subjects were presented with an auditory RC, which was previously pre-recorded, and delivered through loudspeakers connected to a portable computer. Immediately after, two pictures appeared on the computer screen, only one of which represented the sentence being heard. Children were asked to indicate the picture that matched the sentence. In practice trials, they were told that the sentences were recorded by the puppet who was asking their help to find out how to use them appropriately. The stimuli included three practice sentences, during which children were given feedback if this was needed (none of these sentences contained an RC), 54 experimental sentences, and 18 fillers. The experimental sentences were introduced by the lead in fammi vedere (show me) and consisted of 18 subject RCs, 18 object RCs with the embedded subject in the postverbal position and 18 object RCs with the embedded subject occupying the preverbal position. All sentences were unambiguous. These resulted in three conditions exemplified below. SUBJECT (SVO) (11) Fammi vedere il cane che insegue i cavalli. Let-me see the dog that chase3sg the horses ‘Show me the dog that is chasing the horses.’ (12) Fammi vedere il cane che OBJECT (OVS): AGREEMENT Let-me see the dog that inseguono i cavalli. chase3pl the horses ‘Show me the dog that the horses are chasing.’ (13) Fammi vedere il cane che OBJECT (OSV): POSITION Let-me see the dog that il cavallo insegue. the horse chase3sg ‘Show me the dog that the horse is chasing.’ Given the left to right incremental nature of parsing, the parser at the che (that) engages in a subject RC analysis in agreement with the AFH or the MCP. This analysis is confirmed in (11), but not in (12) or (13). After the complementizer che (that) disambiguation towards an object RC analysis occurs and this is brought out by the number agreement morphology on the embedded verb and by the preverbal position of the embedded NP respectively. In order to control for a plural versus singular number effect, six subject RCs had the


147

head-NP in the singular and six in the plural form; likewise, 12 object RCs had the head-NP in the singular and 12 in the plural form. Three different experimental lists were created, each containing six items for each condition plus 18 fillers, for a total of 36 sentences. Items were rotated across lists such that each item occurred only in one of the three conditions above and all conditions were equally present in each list. Items in each list were randomly ordered and subjects were randomly assigned to lists.

8.4.4 The grammaticality judgment task The grammaticality judgment task was designed following McDaniel and Cairns (1996) and was administered to five-, seven-, and nine-year-olds, but not to 11-year-olds since it was judged to be too easy for that age. Its aim was to establish whether children could detect ungrammaticality due to an agreement mismatch. Participants were instructed to listen to a series of pre-recorded sentences delivered through loudspeakers connected to a personal computer. Children were told that the sentences had been recorded by the puppet and they were asked to say whether the puppet spoke correctly or incorrectly. For sentences that children judged to be incorrect, they were asked to tell the experimenter what the puppet should have said instead. If children wanted, they could listen to the sentence a second time. The material comprised six practice sentences in which children could receive feedback and 18 experimental sentences. Usually, the child was administered only three practice sentences, unless she/he had problems in understanding the task. The experimental trials consisted of sentences in which the lexical inflected verb displayed correct or incorrect number agreement with the subject, which could be located either in a preverbal or in a postverbal position. In half of the experimental items the subject was in the singular, and in the other half the subject was in the plural. Eight sentences were grammatical and ten were ungrammatical because of number agreement violations. Sentences were presented in a pseudo-random order. (14)

∗

I cuochi cuoce la pasta. The cooks make3sg the pasta ‘The cooks is making pasta.’ (15) I bambini mangiano la mela. the apple The children eat3pl ‘The children are eating the apple.’

148

Formal features

8.4.5 Backward repetition span test In the backward repetition span test (Ciccarelli, 1998) children were required to maintain an ordered sequence of words in memory and repeat it back in the reverse order.

8.5 Results 8.5.1 Picture selection task A clear subject/object asymmetry is found, with subject RCs being easier to comprehend than object RCs for all groups of children; in addition, no development is observed in the comprehension of subject RCs. Moreover, object RCs disambiguated by position are easier to comprehend than object RCs disambiguated by number agreement. These findings are confirmed by a repeated measure Anova on the percentage of correct responses with type of sentence as a within-subject variable displaying three levels (Subject RC, object RC with a preverbal subject, object RC with a postverbal subject) and age as a between-subject variable displaying five levels. We found an effect of age (F(4,158)=34,16, p

Merging Features: Computation, Interpretation, and Acquisition

Organizational cognition: computation and interpretation

Features

Objects, Agents, and Features

Acquisition

Behavioral Activation: Distinctive Features (CBT Distinctive Features)

E-merging Media

DSP Processor Fund Archit Features: Architectures and Features

DSP Processor Fund Archit Features: Architectures and Features

French Dislocation: Interpretation, Syntax, Acquisition (Oxford Studies in Theoretical Linguistics)

Language acquisition and learnability

Language Acquisition and Learnability

Language Acquisition and Learnability

Merging Processes in Galaxy Clusters

Games, Puzzles, and Computation

Complexity and Computation

Nanotechnology: Science and Computation

Content, Computation and Externalism

Games, Puzzles, and Computation

Complexity and Real Computation

Nanotechnology: Science and Computation

Games, puzzles, and computation

Computation, Cognition, and Pylyshyn

Symbolic computation and education

Cognition, Computation, and Consciousness

Rewriting, Computation and Proof

Cognition, Computation, and Consciousness

Computation, Physics and Beyond

Computation and Asymptotics

Orthogonal Polynomials: Computation and Approximation (Numerical Mathematics and Scientific Computation)

Luger machanical features

Merging Features: Computation, Interpretation, and Acquisition

Organizational cognition: computation and interpretation

Features

Objects, Agents, and Features

Acquisition

Behavioral Activation: Distinctive Features (CBT Distinctive Features)

E-merging Media

DSP Processor Fund Archit Features: Architectures and Features

DSP Processor Fund Archit Features: Architectures and Features

French Dislocation: Interpretation, Syntax, Acquisition (Oxford Studies in Theoretical Linguistics)

Language acquisition and learnability

Language Acquisition and Learnability

Language Acquisition and Learnability

Merging Processes in Galaxy Clusters

Games, Puzzles, and Computation

Complexity and Computation

Nanotechnology: Science and Computation

Content, Computation and Externalism

Games, Puzzles, and Computation

Complexity and Real Computation

Nanotechnology: Science and Computation

Games, puzzles, and computation

Computation, Cognition, and Pylyshyn

Symbolic computation and education

Cognition, Computation, and Consciousness

Rewriting, Computation and Proof

Cognition, Computation, and Consciousness

Computation, Physics and Beyond

Computation and Asymptotics

Orthogonal Polynomials: Computation and Approximation (Numerical Mathematics and Scientific Computation)

Luger machanical features

Recommend Documents