The Acquisition of Syntactic Structure: Animacy and Thematic Alignment

T H E AC QU I S I T I O N O F S Y N TAC T I C S T RU C T U R E

This book explains a well-known puzzle that helped catalyze the establishment of generative syntax: how children tease apart the different syntactic structures associated with sentences such as “John is easy/eager to please.” The answer lies in animacy: taking the premise that subjects are animate, the book argues that children can exploit the occurrence of an inanimate subject as a cue to a non-canonical structure, in which that subject is displaced (The book is easy/*eager to read). The author uses evidence from a range of linguistic subfields, including syntactic theory, typology, language processing, conceptual development, language acquisition, and computational modeling, exposing readers to these different kinds of data in an accessible way. The theoretical claims of the book expand the well-known hypotheses of Syntactic and Semantic Bootstrapping, resulting in greater coverage of the core principles of language acquisition. This is a mustread for researchers in language acquisition, syntax, psycholinguistics, and computational linguistics. misha becker is an Associate Professor in the Linguistics Department at the University of North Carolina, Chapel Hill, where she has taught courses in linguistic theory and child language acquisition since 2002.

CAMBRIDGE STUDIES IN LINGUISTICS General Editors: p. austin, j. bresnan, b. comrie, s. crain, w. dressler, c. j. ewen, r. lass, d. lightfoot, k . rice, i. roberts, s. romaine, n. v. smith

The Acquisition of Syntactic Structure Animacy and Thematic Alignment

In this series 106 sharon inkelas and cheryl zoll: Reduplication: doubling in morphology 107 susan edwards: Fluent aphasia 108 barbara dancygier and eve sweetser: Mental spaces in grammar: conditional constructions 109 hew baerman, dunstan brown and greville g. corbett: The syntax–morphology interface: a study of syncretism 110 marcus tomalin: Linguistics and the formal sciences: the origins of generative grammar 111 samuel d. epstein and t. daniel seely: Derivations in minimalism 112 paul de lacy: Markedness: reduction and preservation in phonology 113 yehuda n. falk: Subjects and their properties 114 p. h. matthews: Syntactic relations: a critical survey 115 mark c. baker: The syntax of agreement and concord 116 gillian catriona ramchand: Verb meaning and the lexicon: a first phase syntax 117 pieter muysken: Functional categories 118 juan uriagereka: Syntactic anchors: on semantic structuring 119 d. robert ladd: Intonational phonology, second edition 120 leonard h. babby: The syntax of argument structure 121 b. elan dresher: The contrastive hierarchy in phonology 122 david adger, daniel harbour and laurel j. watkins: Mirrors and microparameters: phrase structure beyond free word order 123 niina ning zhang: Coordination in syntax 124 neil smith: Acquiring phonology 125 nina topintzi: Onsets: suprasegmental and prosodic behaviour 126 cedric boeckx, norbert hornstein and jairo nunes: Control as movement 127 michael israel: The grammar of polarity: pragmatics, sensitivity, and the logic of scales 128 m. rita manzini and leonardo m. savoia: Grammatical categories: variation in Romance languages 129 barbara citko: Symmetry in syntax: merge, move and labels 130 rachel walker: Vowel patterns in language 131 mary dalrymple and irina nikolaeva: Objects and information structure 132 jerrold m. sadock: The modular architecture of grammar 133 dunstan brown and andrew hippisley: Network morphology: a defaults-based theory of word structure 134 bettelou los, corrien blom, geert booij, marion elenbaas and ans van kemenade: Morphosyntactic change: a comparative study of particles and prefixes 135 stephen crain: The emergence of meaning 136 hubert haider: Symmetry breaking in syntax 137 josé a. camacho: Null subjects 138 gregory stump and raphael a. finkel: Morphological typology: from word to paradigm 139 bruce tesar: Output-driven phonology: theory and learning 140 asier alcázar and mario saltarelli: The syntax of imperatives 141 misha becker: The acquisition of syntactic structure: animacy and thematic alignment Earlier issues not listed are also available

T H E AC QU I S I T I O N O F S Y N TAC T I C S T RU C T U R E A N I M AC Y A N D T H E M AT I C ALIGNMENT

MISHA BECKER University of North Carolina, Chapel Hill

University Printing House, Cambridge CB2 8BS, United Kingdom Published in the United States of America by Cambridge University Press, New York Cambridge University Press is part of the University of Cambridge. It furthers the University’s mission by disseminating knowledge in the pursuit of education, learning and research at the highest international levels of excellence. www.cambridge.org Information on this title: www.cambridge.org/9781107007840 © Misha Becker 2014 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 2014 Printed in the United Kingdom by Clays, St Ives plc A catalogue record for this publication is available from the British Library ISBN 978-1-107-00784-0 Hardback Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

List of figures List of tables Acknowledgements

page x xi xiii

1

Introduction

2

The syntax of displacing and non-displacing predicates

14

2.1

Raising-to-subject and subject control: seem vs. claim 2.1.1 The structure of raising 2.1.2 The structure of control 2.1.3 Raising-to-object and object control: expect vs. persuade Tough-constructions: easy vs. eager 2.2.1 Structure of tough-constructions 2.2.2 Related constructions 2.2.3 Structure of control adjective constructions Unaccusatives and unergatives: arrive vs. dance 2.3.1 A semantically-driven syntactic distinction 2.3.2 Formal representations of unaccusativity Passive 2.4.1 Structure of passive 2.4.2 A different displacing predicate The learning problem

16 19 25 28 30 32 36 39 39 40 42 45 46 49 52

3

Argument hierarchies

61

3.1

The Animacy Hierarchy 3.1.1 Linguistic effects of animacy: morphosyntax and argument structure 3.1.2 Animacy, agency, degree of control, and teleological capability The Thematic Hierarchy 3.2.1 A brief history of thematic roles

63

2.2

2.3

2.4

2.5

3.2

1

64 69 75 76 vii

viii

Contents 3.2.2 Formal accounts of thematic role assignment Animacy and thematic roles in opaque constructions 3.3.1 Raising constructions across languages 3.3.2 Tough-constructions across languages Properties of derived subjects 3.4.1 Argument structure universals, and the “problem” of ergativity A learning procedure Summary

79 84 86 103 109

4

Animacy and adult sentence processing

126

4.1

Relative clauses 4.1.1 Reduced relative clauses 4.1.2 Subject vs. object relative clauses Processing of raising and control 4.2.1 Sentence completion 4.2.2 Novel verb learning Psycholinguistic effects of animacy on production of the passive Summary

129 129 136 139 140 147 153 155

5

Animacy and children’s language

156

5.1

Development of the animacy concept 5.1.1 Featural properties of animates 5.1.2 Behavioral properties of animates 5.1.3 Intentional properties of animates 5.1.4 Further conceptual change 5.1.5 Agency 5.1.6 Summary Children’s use of animacy in learning argument structure 5.2.1 The power and limitations of Semantic Bootstrapping 5.2.2 The power and limitations of Syntactic Bootstrapping Children’s acquisition of displacing predicates 5.3.1 Acquisition of raising and control 5.3.2 Acquisition of tough-constructions 5.3.3 Acquisition of unaccusatives 5.3.4 Animacy and the acquisition of the passive Summary

157 158 159 166 169 172 173 174 179 186 190 192 208 227 235 242

6

Modeling the acquisition of displacing predicates

245

6.1 6.2

Displacing predicates in the input to children Computational modeling of language acquisition

250 256

3.3

3.4

3.5 3.6

4.2

4.3 4.4

5.2

5.3

5.4

114 119 124

Contents

ix

6.2.1 Learning as generalization 6.2.2 Restricting the hypothesis space Hierarchical Bayesian Models 6.3.1 A model of learning raising and control 6.3.2 A model of learning tough-constructions 6.3.3 A model of learning unaccusatives and unergatives Summary of modeling results

258 261 265 267 276 279 281

7

Conclusions and origins

283

7.1 7.2

Origins of knowledge of the animacy distinction Origins of knowledge of linguistic animacy and displacing predicates Further questions

286

Appendix Bibliography Index

298 300 322

6.3

6.4

7.3

289 296

Figures

4.1

5.1 5.2 5.3 5.4 5.5 5.6

6.1 6.2 6.3 6.4 6.5 6.6 6.7 6.8

x

Percent “Correct” responses for novel raising (target) and control (filler) verbs; list condition only, averaged across frequency conditions Mean RT in msec (log10) to warm-ups and fillers Mean RT in msec (log10) to hard and afraid Mean RT in msec (log10) for Group 1 Mean RT in msec (log10) for Group 2 Mean RT in msec (log10) for Groups 1 and 2, novel tough-adjectives Percentage of correct categorizations of novel raising (target) and control (filler) verbs in adult experiment, 1-exemplar condition (Becker and Estigarribia, 2013) A power law distribution, illustrating Zipf’s law A power law distribution plotted on a log–log scale Graphical representation of the Hierarchical Bayesian Model Graphical representation of β and λ in Hierarchical Bayesian Model Posterior distribution plots for λ; CHILDES data Posterior distribution plots for λ; Switchboard data Posterior distribution plots for λ for each adjective Posterior distribution plots for λ; unaccusative and unergative verbs

page 152 218 219 220 221 223

224 259 260 267 268 273 273 278 281

Tables

3.1 3.2 3.3 3.4 3.5 4.1 4.2 4.3

4.4 4.5 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 6.1 6.2

Raising and control verbs in German and Italian page 93 Raising and control verb meanings in Polynesian, European 97 Properties of tough-adjectives across languages 108 Properties of “basic” and derived subjects across languages 114 Typical properties of subjects and non-subject arguments 120 Percentage of raising and control responses in Becker (2005), Experiments 1–3 142 Percentage raising and control responses in Becker (2005), Experiment 4 143 Percentage raising and control responses in Becker (2005), Experiment 4 (separated by aspect of embedded verb) 144 Novel verbs and their pseudo-definitions 149 Observed mean percentage “Correct” responses 151 Mean percent correct (with predicate type) 176 Examples of test items in Becker (2006b) 200 Predicted responses depending on child’s assumptions 201 Results of experiment: relative proportion of OK/silly responses 202 Response time in msec to transitive vs. there-construction questions (5-year-olds) 207 Design of tough-adjective study 216 Response Time (msec) for novel tough-adjectives, group 2 222 Children’s use of animate subjects with unaccusative vs. unergative verbs in spontaneous speech 235 Mothers’ use of animate/inanimate subjects with raising and control verbs 251 Adults’ use of animate and inanimate subjects with tough/control adjectives 252

xi

xii 6.3 6.4 6.5 6.6 6.7 6.8

List of tables Mothers’ use of animate and inanimate subjects with unaccusatives and unergatives Distribution of raising and control verbs with animate/inanimate subjects, from CHILDES Distribution of raising and control verbs with animate/inanimate subjects, from switchboard Adults’ use of animate and inanimate subjects with tough/control adjectives Distribution of ready in speech to children Mothers’ use of animate and inanimate subjects with unaccusatives and unergatives

254 270 271 277 279 280

Acknowledgements

While working on my dissertation on the acquisition of the copula in child English, I happened to read an article by Bob Frank on the relative structural complexity of raising and control constructions. In the Tree Adjoining Grammar (TAG) framework, Frank argued, raising constructions involve additional complexity not found in control constructions. Because of this difference in complexity, he surmised that if someone heard the sentence Gabriel glorps to eat gouda, that person would first assume that the sentence had a control structure, and therefore that glorp was a control verb. When I read his prediction I thought, “I bet he’s right that people will assume glorp is a control verb, but I bet it’s for a different reason.” I quickly conducted an informal survey of the members of my department by giving them this very sentence and asking them what glorp meant. Sure enough, everyone offered me control verbs for glorp; not one person said it was a raising verb. Initially, my hunch about why people would have this preference was that raising verbs are in some sense midway between functional and lexical categories (they are like function words, e.g. auxiliaries, in their argument structure, but lexical in their marking of subject agreement and lack of inversion), and I thought that people might be unwilling to assign a novel word to a category that is similar to closed class categories. But as I began to test people’s assumptions about these sentences with different kinds of inputs, in particular, with fill-in-the-blank sentences (where the participant is not assigning a novel verb to any category), I realized the explanation had to be something else. An early clue came from the observation that when I gave participants an inanimate subject, I got a lot more raising verb responses than I did when the subject was animate. This was the beginning of my exploration of how learners come to identify the underlying structures of these kinds of sentences. And so although his influence was indirect and unintentional, I am grateful to Bob Frank for starting me on this path. Over the years my approach has broadened to

xiii

xiv

Acknowledgements

include different constructions (tough vs. control adjectives, and unaccusative vs. unergative verbs), child and adult participants, and different methodologies, all of which have been profoundly influenced by my colleagues around me. My early work on this topic was much influenced and enriched by Lila Gleitman, John Trueswell and the wonderful and insightful folks who made up the Cheese seminar at University of Pennsylvania’s Institute for Research in Cognitive Science. I am deeply grateful to IRCS and NSF for the opportunity to be immersed in that stimulating environment for two years. Several of the studies I report on in this book are the product of collaborations, and I thank Bruno Estigarribia, W. Garrett Mitchener and Jeannette Schaeffer for their wonderful ideas and for being such inspiring and generous collaborators. Garrett Mitchener deserves a particularly big thank-you for his extensive input not only on our previous collaboration but on chapter 6 of this book. That chapter could not have materialized without his expertise and many hours of work. I thank him for developing our learning model, running the simulations, creating the figures for depicting the results, and most of all for his patience in explaining (and re-explaining) the details of the Bayesian model to me. In addition, though not usually given author status, I thank the many undergraduate research assistants (Bianca Bulchandani, Marguerite Cameron, Brian Cansler, Adam Daland, Sam Farquharson, Alex Fine, Duna Gylfadottir, Michelle Hewitt, Molly Jabeck, Taylor Shirley, Lara Stephenson) and graduate students (Susannah Kirby, Inmaculada Gomez Soler) who helped me collect data. I have been very fortunate to have had various sources and forms of support at the University of North Carolina, from the Faculty Partners and University Research Council small grants, to the Associate Professor Support grant through the College of Arts and Sciences. I could not have completed my various psycholinguistic experiments without this financial support. I am particularly grateful to the Institute for Arts and Humanities, whose Preyer Fellowship through the Faculty Fellows program allowed me a semester’s leave in order to complete this book. No less vital than financial and leave support, I have the great fortune of having wonderful, supportive, and challenging colleagues. I am particularly indebted to Elliott Moreton, Katya Pertsova, and Jennifer Smith for feedback on early drafts of this work. I am lucky to have colleagues in other departments with interests in language, whose provocative questions have helped move my research forward: Jennifer Arnold, Peter Gordon, and the other members of the Psychology and Language UNC Group (PLUG), Dorit Bar-On and Dean

Acknowledgements

xv

Pettit in philosophy, and the audience members of the Linguistics colloquium and the Expressive Communication and Origins of Meaning (ECOM) research series for their insightful questions. Parts of this book have been presented to audiences at the Boston University Conference on Language Development, College of Charleston, Generative Approaches to Language Acquisition North America, University of Maryland, Università di Milano, WCCFL, the Workshop on Animacy (Radboud University, Nijmegen), the Workshop on Input and Syntactic Acquisition (Portland, OR), and the Workshop on Sentence Complexity (Universität Konstanz). I thank these audiences for their comments, questions, and feedback, which invariably improved my work. For data, judgments, and general brainstorming help I would like to thank Ash Asudeh, Daniela Blettner, Julie Brittain, Ivano Caponigro, Sandy Chung, Alex Clark, Stephen Crain, Annie Gagliardi, Jean Berko Gleason, Vincent Homer, Yi Ting Huang, Nina Hyams, Kyle Johnson, Elsi Kaiser, Beth Levin, Jeff Lidz, Diane Massam, Yuko Otsuka, Lisa Pearl, Geoffrey Pullum, Susi Wurmbrand, Charles Yang and Hang Zhang. I also want to express my profound appreciation to all of the adults and children (and their families) who have so generously provided the data for my studies. I think 3- to 7-year-olds are pretty much the most fun data sources ever! A slightly less entertaining but absolutely crucial data source is the CHILDES database, and I’m grateful to Brian MacWhinney and all the members of the chi-bolts mailing list for help in using this important resource. I also thank Helen Barton at Cambridge University Press for her guidance, and a reviewer who provided numerous helpful suggestions which have improved this work greatly. Finally, but most important, a huge debt of gratitude is owed to my husband, Scott, and our children, Olivia and Lyra, whose early years have been touched (and hopefully not too much marred!) by the writing of this book. I thank them for their patience, love, and support, and I thank my children for unwittingly providing entertaining and highly appropriate data, some of which has found its way into this book. I’m also very grateful to my parents who have generously helped with childcare so that I could write uninterrupted, and who have provided general encouragement and support throughout my life.

1 Introduction

How do children take a string of speech sounds, chop it up into discrete units (words), and then assign to that chopped up string of sounds a particular meaning? In many sentences, perhaps most, words that are semantically related to one another are also near to each other in the sentence. For example, a verb and its arguments – the nouns or other phrases that the verb selects – are usually in close proximity to each other (at least, they are generally clausemates): in a simple main clause sentence like The student read a book, the verb read selects a subject and a direct object noun phrase (NP), and these NPs are positioned right next to the verb that selects them. This is so regardless of the particular language’s basic word order or even the degree of rigidity of word order. Many theories of language acquisition exploit this fact to explain (part of) how children begin to tackle the challenge of integrating form and meaning in their language. But arguments need not be proximal to their selecting predicate, and adjacent or proximal words need not stand in a semantic, selectional relation to one another. This is because human language allows for the semantic relations between words to span long distances – in principle, infinite distances. This book is about how children begin to figure out how to interpret sentences in which the proximity of words belies their semantic relations – how children determine the underlying syntactic structure of sentences in which semantic relations are long-distance, and how knowing the syntactic structure helps children interpret those semantic relations. I argue that children recruit cues from the conceptual domain, particularly animacy, in solving this puzzle. One very important insight about language acquisition stems from the observation that subjects of basic, canonical sentences are often animate, or more animate than other nouns in the sentence, and that children can exploit this fact to home in on basic sentence structure (Pinker, 1984): find the most animate noun and it will be the subject. My question is how children go beyond these canonical sentences in which subjects are agents or experiencers, and objects are patients or themes, to figure out the 1

2

Introduction

structures of more complex configurations. The answer I offer is that because children expect subjects to be animate in canonical structures, they can exploit deviations from this expectation, in particular encountering an inanimate subject, to learn that in just these kinds of sentences the underlying structure is non-canonical and complex. A predicate that allows its subject to be inanimate does not bear the same type of semantic relationship to its subject as a predicate that requires its subject to be animate; and the non-canonical semantic relation between subject and predicate translates into a non-canonical syntactic structure: one in which the subject is derived, or displaced. The particular type of non-canonical structure I will focus on involves what I’ll call “displacing predicates.” These are predicates that fail to select an external argument (a semantic subject) – that is, there is no “do-er” or experiencer of the predicate’s action or state. An example of such a predicate is the verb seem. When we say John seems to like French fries it doesn’t make sense to say that John is a “seemer” of anything. Instead, the subject is semantically related only to the lower predicate, and thus we can say it is “displaced” (or derived) in the sentence with respect to the locus of its semantic role. The question I seek to answer is how children figure this out – how they identify just the sentences of their language in which the subject is in fact displaced, which in turn allows them to categorize particular predicates as being “displacing” predicates. One might think that this is such a small corner of the grammar – displacing predicates are such a tiny piece of what needs to be learned about language, and they have a rather peripheral feel to them. Surely what matters most in advancing the study of language and its acquisition is to explain how children acquire the canonical parts of grammar, the most well-behaved and earliest learned predicates, so that the exceptional ones can then be accounted for, precisely, as exceptions. How is the study of these unusual predicates relevant in the larger scheme of things? There are two related answers to this question. One is that these very predicates and their non-canonical structures represent one of the core properties of human language itself: the ability to have non-local dependencies. That words can bear structural relations to other words over an (in principle) infinite distance is one of the hallmarks of human language. In this sense, displacing predicates are profoundly non-peripheral. The second and related answer is that these predicates have formed part of the argument for generative grammar from the very beginning. Not only does the learning puzzle addressed in this book involve determining that a given sentence has a displaced subject, but also the learner must distinguish those sentences with displaced subjects from superficially identical sentences whose subject is not displaced, but rather is the semantic subject of the main

Introduction

3

predicate (e.g. in John claims to like French fries John is the “claimer”). This aspect of the question is old and deep, and it forms one of the pillars on which generative grammar was built. What Chomsky (1957) called “constructional homonymity” (John seems/claims to like French fries) was exposed as a fatal flaw in a theory of grammar that did not combine both phrase structure rules and transformational rules. Even though these subclasses of verbs can be distinguished by their distribution in other types of sentences (It seems/*claims that John likes tomatoes; What John claims/*seems is to be the strongest), the fact of their distributional overlap in even one sentence type requires that learners have a means of teasing them apart. It was suggested in Chomsky (1964, 1965) that the challenge presented by constructional homonyms in terms of language acquisition pointed to the need for an account of language learning within the rationalist tradition of epistemology (Chomsky, 1965, p. 25). That is, they bore directly on arguments for innate knowledge of language. During this era, Carol Chomsky (1969) took up the puzzle these constructional homonyms posed in an empirical study of children’s language. She posed the question of how children handle situations in which [t]he true grammatical relations which hold among the words in a sentence are not expressed directly in its surface structure. (Chomsky, 1969, p. 6)

That is, how do children parse a particular word string that is potentially associable with multiple underlying structures? Chomsky focused on sentences of the form in (1). (1)

The doll is daxy to see.

Without knowing what the adjective daxy means, the sentence could have (at least) either of the following interpretations, the first corresponding to the so-called tough-construction (2a) and the second to the control adjective construction (2b), as disambiguated by the familiar English adjectives. (2)

a. The doll is easy to see. (= it is easy for someone else to see the doll) b. The doll is eager to see. (= the doll is eager to see someone else)

The difference between (2a) and (2b) is clearly semantic, but it is also syntactic. Syntax is about not just the ordering of words, but also the logical relations among them: the fact that the relation between the doll and easy is profoundly different from that between the doll and eager is linked to a

4

Introduction

difference in how the structures of these two sentences are represented. The nature of these structural differences will be explained in detail in Chapter 2. For the moment what is significant is that the semantic role of the subject NP is utterly different in sentences (2a) and (2b), and therefore its syntactic relationship to the main predicate is different. While the main clause subject in (2a) is understood as the semantic object (patient) of the embedded clause, in (2b) the main clause subject is interpreted as the semantic subject (agent/experiencer) of the embedded verb. Assuming that the semantic difference between (2a) and (2b) corresponds to an underlying syntactic difference between them, in the terms being used here the subject in (2a) is displaced, but the subject in (2b) is not. The parallel to the earlier seem example is that neither easy nor seem takes an agentive (or experiencer) subject. For both of these predicates, the subject’s semantic ties are to another predicate altogether; in this sense, both seem and easy are displacing predicates. (3)

The girl is daxy to see. a. The girli is easy [PROarb to see t i .] (tough-adjective) b. The girli is eager [PROi to see.] (control adjective)

(4)

Mary gorped to be strong. a. Maryi seemed [t i to be strong.] (raising verb) b. Maryi claimed [PROi to be strong.] (control verb)

The semantic difference between the (a) and (b) pairs in (4) is a little subtler than that in (3). In both (4a) and (4b) Mary is the semantic subject of the lower predicate to be strong. The difference has to do with its semantic relation to the main predicate, seem vs. claim: as noted above, there really is no semantic relationship between Mary and seem, but there is between Mary and claim (she is the “claimer”). Again, this semantic asymmetry corresponds to a syntactic one: the subject is displaced in (4a) but not (4b). So the problem for language learners is to figure out that the subject of seem or easy is not the semantic subject of these predicates, but rather bears a longdistance semantic relationship to another part of the sentence, even though a construal of the strings in (3) and (4) involving a local semantic relationship is possible given the constructional homonyms with claim and eager.1 1 The same surface ambiguity arises in raising-to-object (also called Exceptional Case Marking)

and object control, as in Sue wanted/asked Gordon [to cut the grass]. Since the main focus here is on constructions with derived subjects I will not have a lot to say about these constructions, but they will be discussed briefly in Sections 2.1.3 and 5.3.1.

Introduction

5

This was precisely the question that Carol Chomsky posed. But it was not the question she answered in her empirical work. Rather, her experiments addressed the question of what children know about predicates like easy and when they know it. In fact, nearly all of the literature on children’s acquisition of tough-adjectives and raising verbs has focused on this aspect of the problem, and so in the decades since Chomsky’s seminal work, the deeper question of how children disentangle the respective constructions has not been tackled directly. The purpose of this book is to tackle that how question. The answer I propose is that hearing an inanimate subject in a sentence like (3) or (4) provides a cue that the subject is displaced, and therefore that the main predicate of the sentence is a displacing predicate. This cue is informative in these cases because an inanimate subject is possible with the structure that involves displacement, but not with the other structure: (5)

a.

The rock is easy to lift.

b. # The rock is eager to lift/fall. (6)

a.

The rock seems to be heavy.

b. # The rock claims to be heavy.

Lest readers be concerned that I have missed a more obvious answer to this puzzle, namely displacing predicates’ ability to occur with expletive subjects (it, there), I should state that I do think predicates’ occurrence with expletives is a valuable cue in this learning process, and I have reasons for focusing on inanimate referential subjects instead. These reasons are laid out in detail in Section 2.5 below. The main focus in this book will be on the two constructions in (3/5) and (4/6), those involving tough-adjectives and raising-to-subject verbs. However, there are other constructions that involve subject displacement, such as the passive, and there are other (non-passive) predicates that can be classified as displacing predicates, such as unaccusative verbs. Unaccusative verbs are a type of intransitive verb which, unlike unergative intransitives, select only an internal argument and no external argument. Thus, the subject of an unaccusative verb has been displaced from an underlying object position. But given a surface string containing only a subject and a verb, it is not immediately obvious whether the underlying structure involves an external argument (as in (7b)) or an internal one (as in (7a)).

6 (7)

Introduction John pilked. a. Johni arrived t i . (unaccusative) b. John danced. (unergative)

The asymmetry in (7a–b) is even more subtle than in (4a–b), and fairly unintuitive for English speakers. We will see in Section 2.3 that many languages exhibit more obvious distinctions between unaccusative and unergative verbs, and the distinction between these types of verbs is well supported crosslinguistically. The spirit of the distinction is that in (7a) John is the theme of the verb (in a sense, John “undergoes” the arriving event; he does not have an agentive role), while in (7b) John is the agent of dance: he “enacts” the dancing event. Thus, while arrive and dance are both intransitive verbs, the underlying relationship between the verb and its lone argument is different in each case. Once again, these distinctions map onto structural differences that the language learner must be able to identify in order to be said to have adult-like competence in her language. And similar to the first two constructions, the string in (7) can be associated with the displacing structure in (7a) if the subject of the sentence is inanimate. (8)

a.

The package arrived.

b. # The package danced.

My proposal is primarily about how children solve the mapping and categorization problems: I take the view that predicates, with some important exceptions, are fundamentally either displacing or non-displacing. So encountering an inanimate subject tells the child that the sentence involves a structure with a displaced subject, which in turn tells the child that the main predicate is a displacing predicate. Although I do not try to explain how children figure out exactly what these abstract predicates mean, I suspect that the categorization of a predicate as displacing (or non-displacing) in turn provides a clue to the set of possible meanings the predicate might have. That is, displacing predicates will be largely limited to auxiliary-like meanings – meanings having to do with modality, happenstance, appearance, ease/difficulty, possibility, and likelihood (and non-volitional events, in the case of unaccusative verbs). Non-displacing predicates, on the other hand, will have a volitional, intentional, or emotive aspect to their meaning. As we have seen, sentence strings like (3), (4), and (7) are associable with multiple syntactic structures if the subject is animate and the predicate’s meaning is not known. I will refer to these sentence strings as “opaque”

Introduction

7

sentences rather than use the term “ambiguous.” The reason is that the structural indeterminacy of these sentences is different from the more typical type of structural ambiguity presented by, for example, Prepositional Phrase (PP) attachment. (9)

a. Put the frog on the napkin in the box. b. I saw the man with binoculars.

The string in (9a) is locally ambiguous at the first PP [on the napkin] because this phrase could indicate either a description of where the frog is, or it could indicate the location where the frog should be put (and the string is disambiguated by the second PP). The processing of this type of construction has been explored extensively in both children and adults (Trueswell et al., 1999, i.a.). But the decision about where to attach the first PP does not have an effect on the lexical meaning of the main predicate put – put means the same thing, whether the PP is attached to the NP or the VP. Similarly in the globally ambiguous example in (9b), the meaning of see does not depend on which structure one applies to this string. If a learner encountered an unknown verb in this string (I gorped the man with binoculars) the meaning of gorp would not necessarily depend on whether the PP was attached to the NP or the VP node. (And, correspondingly, knowing the meaning of gorp would not help resolve the attachment puzzle, and so the sentence is truly ambiguous.) In the kinds of constructions under consideration here, on the other hand, the meaning of the main predicate is fundamentally different according to whether the subject is displaced or not. Not only are the verbs seem and claim different verbs (and this extends to the other pairs of predicates we’ve seen: easy/eager, arrive/dance), but if we encounter a novel predicate in a string like (4) the meaning of this predicate will depend on how the string is parsed. On the surface, this might appear to make the learning problem easier. If you know the meaning of the predicate (seem, claim, etc.) you can choose the right structure: if you know that the main verb means ‘seem’ then you know the subject is displaced, and if you know the main verb means ‘claim’ then you know the subject is not displaced. Thus, the sentence Mary seems to be strong is not ambiguous – once you know the lexical properties of seem the underlying structure of the sentence follows. However, this does not solve the learning problem for children, for two reasons. First, most of the verbs and adjectives that participate in these structures have abstract lexical meanings that are not straightforwardly discernable directly from observation of the non-linguistic world (eager and easy are both different from red in this respect; and seem and claim are different from eat similarly). Secondly, a wealth of empirical studies,

8

Introduction

forming the literature surrounding the Syntactic Bootstrapping hypothesis, tell us that children learn the meanings of predicates, in particular abstract ones, in large part via the sentence structure they occur in rather than the other way around (Gleitman, 1990, and considerable work following this). Thus, learners need to rely on the underlying structure of sentences like (4) in order to figure out whether the main verb means something like ‘seem’ or something like ‘claim,’ but how do they first figure out the underlying structure? It is in this sense that the sentence strings in (3), (4), and (7) are syntactically opaque. I define syntactic opacity as follows. (10)

A string is syntactically opaque if the underlying syntactic structure that generates the string cannot be determined unequivocally on the basis of the string and knowing only the grammatical categories of the words, without at least some lexical semantic knowledge of the main predicate.

Strictly speaking, all strings are opaque in this sense, until the lexical semantics of the main predicate is known. Even a string like that in (11) could be associated with various structures including, but not limited to, those in (11a–c). (11)

NP V NP a. [NPsubj [VP [Vtrans NPobj ]]] b. [NPsubj [VP [Vintrans ] NPloc ]] c. [NPsubj [VP [Vditrans NPobj ] ∅ind.obj ]]

However, much previous work on children’s learning of verb argument structure has revealed that children are prone to making assumptions about these strings: a verb with one NP is assumed to be intransitive, a verb with two NPs transitive, and a verb with three NPs ditransitive (see Gleitman et al. 2005 for a good overview of this literature; though see Tomasello and Brooks (1998); Tomasello (2000) for a different view). This is precisely because, as noted at the beginning of this introduction, proximal words are typically semantically related – and related in particular ways. But the constructions under consideration here are special, and especially opaque, because the adjacent NPs are not semantically related to the predicate in the usual way. And so the assumptions learners might make about the underlying structures of sentences like (11) will not apply straightforwardly to constructions with displacing predicates. In addition to tough-constructions, raising verbs, and unaccusatives I will discuss the passive, which, unlike some of the other constructions

Introduction

9

considered here, has been studied fairly extensively in the acquisition literature (Slobin, 1966; Maratsos et al., 1979; Borer and Wexler, 1987; Crain et al., 1987; Lempert, 1989; Fox and Grodzinsky, 1998, i.a.). I include the passive for two reasons. One is that passives involve a displaced subject: the syntactic subject is understood as the semantic object, or patient of the verb’s action. The second is that in English certain passives are ambiguous between a verbal and an adjectival reading, a fact which Borer and Wexler exploited in their account of children’s interpretation of the passive. Thus, a short passive, as in (12b), is structurally ambiguous. (12)

a. John was kicked by Sam. (verbal passive) b. The door was closed. (verbal or adjectival)

Nevertheless, important differences between the passive construction and the others considered here will explain why the solution I propose for raising verbs, tough-constructions, and unaccusatives actually does not extend to the passive. Most pointedly, like in the example of PP-attachment above, the meaning of the main predicate does not change radically depending on whether the sentence has a passive or an active voice. Kick and be kicked by both denote a kicking event. Thus, discovering the structure of a passive sentence requires understanding that the subject is displaced, but it does not involve the task of categorizing the main predicate as an inherently displacing predicate – that is, the passive verb should not be assumed to have an auxiliary-like semantics. While the passive will be discussed in Chapters 2 and 5 this construction will not occupy a focal point in the overall discussion. All of the constructions outlined above that involve a displaced subject, with the exception of unaccusative verbs, have been argued to be difficult for children to acquire (though see Babyonyshev et al. (2001) for claims that young children represent unaccusatives as unergatives). Chomsky (1969) and Cromer (1970), among others, argued this for tough-constructions, Hirsch and Wexler (2007) have argued the same for raising-to-subject verbs, and Borer and Wexler (1987), among many others, argued this for passives. To the extent that children can correctly interpret such structures, for example non-reversible passives, they are said to do so by relying on “real world knowledge” rather than syntax. For example, Slobin (1966) showed that children responded more quickly to non-reversible passives (The flower is being watered by the girl) than reversible passives (The cat is being chased by the dog), presumably because either dogs or cats can chase or be chased, but flowers do not water girls. What I argue in this book is that children do use “real world knowledge,” not as a

10

Introduction

means of avoiding complex syntax, but rather as a means of discovering the complex syntax itself. The evidence comes from experiments of novel word learning, in which children use subject animacy to draw inferences about the argument structure properties of novel predicates. In this book I revisit Carol Chomsky’s original question through the lens of the advances that have been made in the fields of linguistic theory, psychology, language development, and computational modeling, with the goal of integrating cross-linguistic constraints and preferences on argument structure mapping into a theory of how structures involving displaced subjects are acquired. Chapter 2 provides an overview of some formal accounts of the opaque constructions under consideration here, including a description of how these constructions are analyzed under Minimalist approaches. The vocabulary of the Minimalist Program gives us a unified way of talking about raising, unaccusative, and passive constructions: these are predicates whose vP is considered “defective,” allowing an NP argument to move out of their “weak” phase into the main clause. (The spirit of this unification is no different from previous incarnations of the theory, but the language of it is different.) One mechanism that has been proposed within Minimalism to account for passives and subject raising, namely “smuggling,” has also been invoked in an account of tough-constructions (Hicks, 2009). My thesis is not contingent on any particular syntactic framework or formalism, but some kind of formalism is required in order to see why the acquisition question I’m addressing is a matter of acquiring syntax. Chapter 3 then looks at how animacy is realized in grammar along a number of dimensions: how animacy is grammaticalized in various languages, how it relates to thematic roles, and how, in turn, thematic roles relate to argument structure. The emphasis in this chapter is on typological patterns; that is, how animacy surfaces in the world’s languages, and how displacing predicates work in different languages as well. Across genetically diverse languages we can observe two rather clear and consistent patterns. One is that languages tend to organize animacy distinctions between more animate and less animate entities according to a hierarchy, according to which humans are the “most animate,” followed by non-human animals, followed by inanimates. Though there is diversity in the number of distinctions made in the hierarchy, and where dividing lines are drawn, the hierarchy itself is robust: we do not find languages, for example, which treat humans and inanimates alike to the exclusion of animals. The second consistent pattern is that languages prefer non-displaced (i.e. canonical) subjects to be animate but rather liberally allow displaced

Introduction

11

subjects to be inanimate. Relatedly, where we find displacing predicates in other languages (such as several Polynesian languages), these predicates, but not their non-displacing counterparts, permit inanimate subjects. Some of the work reviewed here is old and widely familiar, such as Comrie’s (1989) work on the Animacy Hierarchy and Keenan’s (1976) work on the universal properties of subjects. What is novel is the use of this work in defining a learning strategy for children’s acquisition of displacing predicates. The typology of human languages is important to bear in mind especially in studies of language acquisition because children have to be able to acquire any human language they happen to grow up hearing, and so their learning strategies must be suited to the broad scope of these grammatical patterns. Furthermore, both linguistic typology and child language have the potential to inform us about deep universals of human language: typology because broadly robust patterns are likely to be components of Universal Grammar (UG), and child language because if all children share the same sorts of stages of development, strategies for learning, and types of errors, then these are also likely to be deeply human traits. Animacy is something that is both universally marked in adult languages and salient in children’s developing cognition and grammar, and as such, it appears to be a most fundamental component of human language. We then turn, in Chapter 4, to psycholinguistic studies of how animacy affects language processing in adults. Studies of adult language processing can be highly informative for understanding children’s language development. One reason is that such studies allow us to identify the target state that children are progressing toward in their development, and so we can measure children’s degree of deviance (if any) from the target state. Another is that what leads to processing difficulty in adults is likely to also lead to processing difficulty in children. If children then exhibit problems producing or comprehending the very types of constructions adults are known to have trouble with (for example, object relative clauses with animate relativized NPs), then children’s problems could be due to the difficulty of processing those constructions rather than grammatical deficits due to immaturity. Conversely, if some manipulation leads to better processing outcomes for adults (for example, an inanimate patient subject, as opposed to an animate patient subject), then to the extent that children experience facilitation in the same manipulation, their better outcome could be due to the same kind of facilitation that adults experienced, and not because they used an alternative, non-syntactic mechanism for parsing the construction. Finally, we can expose adult speakers to novel predicates in different kinds of linguistic contexts and measure how adults categorize and

12

Introduction

interpret these novel predicates. This approach provides a means of simulating the learning process in adult participants. One of the studies reviewed in this chapter points to a strong effect of inanimate subjects leading to adults’ categorization of novel verbs as raising verbs. Chapter 5 turns to children’s language. The first part of the chapter looks at children’s cognitive development of the concept of animacy. Here I cover a broad swath of the psychology literature with the aim of establishing that by the time children are acquiring sentence structure, and certainly before they acquire complex biclausal constructions like raising-to-subject or toughconstructions, children have a well-defined concept of animacy. The purpose of this long exegesis on a distinctly non-linguistic aspect of child development is twofold: one is to support the view that the animacy distinction in general can be recruited in the acquisition of basic sentence structure, as argued by the Semantic Bootstrapping hypothesis (Pinker, 1984); the second is to make the case for children being non-animistic. In spite of a vast amount of research showing children’s rather sophisticated and adult-like conception of animate entities having intentions and goal-directed actions but inanimate entities having none of these properties, the view persists that preschoolers “think everything is alive.” It is important to dispel this myth so that we can correctly interpret children’s responses in some of the linguistic tasks that manipulate animacy. In the second part of Chapter 5, I review the important insights of both the Semantic and Syntactic Bootstrapping hypotheses about language learning, showing how both of them go a long way toward explaining how children acquire argument structure. But I also show that neither of them straightforwardly applies to constructions involving displacing predicates. Thus, another innovation of this work is that I extend both of these approaches so as to increase their empirical coverage and explanatory power by suggesting a way to use inanimate subjects to acquire these non-canonical predicates and the opaque structures they occur in. The third part of the chapter reviews experimental work on children’s acquisition of displacing predicates, and evidence pointing to the important role of inanimate subjects in this categorization process. It is interesting that, with the exception of the passive, these constructions have only rarely been studied with respect to language acquisition. Thus, part of this book’s contribution is an expansion of the acquisition literature to include them. Chapter 6 extends the study of children’s acquisition of displacing predicates into the realm of computational modeling. Recent advances in computational modeling techniques have led to new ways of theorizing about what

Introduction

13

information is useful to learners and how they use this information. I will illustrate some ways of using computational models such as Bayesian learners that can exploit statistical patterns of predicates’ occurrence with animate vs. inanimate subjects in child-directed speech in order to learn which predicates are displacing and which are not. As we will see, our models require prior hypotheses about a predicate’s degree of “selectivity” about its subject, and about subject animacy, in order to perform the categorization task. The final chapter, Chapter 7, concludes by addressing questions about innateness: which of the ingredients in my account (knowledge of animacy; preference for subjects to be animate; knowledge of displacing predicates) can be said to constitute innate knowledge on the part of the learner? What does it mean to say that any of these things are innate? In addition, I lay out what I see as the primary open questions left unresolved by the work presented here. The components of this book span a number of distinct subdomains and theoretical debates within linguistics: formal syntactic analyses of sentence structures involving argument displacement, typological studies of the role of animacy in language, psycholinguistic and developmental studies, and computational modeling. Each of them has alone generated considerable literature and debate. For instance, much ink has been spilled over the syntactic analysis of tough-constructions and the correct definition of Agent. And though the field of Bayesian learning models in studies of language acquisition is relatively new, it has been quick to generate a great deal of active research. Necessarily, many details of these analyses and observations will receive only cursory mention or may not be mentioned at all. This book is intended for a generally linguistically-informed audience, and in the interest of accessibility I have attempted to convey the relevant issues without too much technical detail. Experts in each subfield may therefore take issue with my presentation or choice of omissions. Nevertheless, in order to touch on each of the necessary ingredients (formal theory, typology, psycholinguistics, and modeling), some brevity is required. The purpose of this book is to bring together all of these different domains to show how sentences which violate children’s expectations about canonical sentences can help children attain the correct syntactic parse for non-canonical, opaque sentences – that is, sentences in which the subject is displaced. Although a variety of cues are available for language learners to disambiguate these troublesome strings, and learners undoubtedly do make use of multiple cues in the input, one cue that learners should, and do, pay particular attention to is that these predicates uniformly and easily occur with inanimate subjects.

2 The syntax of displacing and non-displacing predicates

This chapter presents formal analyses of the constructions of interest in this book: raising-to-subject, tough-movement, and unaccusative constructions. In each of these constructions the main clause subject is displaced with respect to its thematic position, and the main predicate is therefore a displacing predicate. Each of these displacing constructions will be contrasted with a surface-similar construction that does not involve displacement, as discussed in Chapter 1. And in each case, the cue of subject animacy can distinguish the two constructions: only the displacing predicates permit inanimate subjects in an unrestricted fashion. We’ll also discuss the passive, which works like these other displacing constructions in certain respects, but behaves differently in other respects. For each displacing construction I will give a descriptive overview of the construction and some empirical diagnostics for it. I will then present what I take to be standard contemporary accounts of the construction’s underlying structure and derivation, with some references to older accounts or different frameworks as relevance dictates. All of the constructions under consideration here have provoked lively debates over the past several decades, resulting in numerous analyses in the syntax literature. It is beyond the scope of this book to give a complete history of the analyses of each construction. Rather, my goal here is to highlight some of the classic and contemporary treatments of these structures. The accounts I present will lean heavily towards movementbased frameworks, though in fact the main thesis of this book does not hinge on a movement-based treatment of these constructions. In movement- and non-movement-based frameworks alike, displacing predicates lack a semantic relationship to their subject, unlike non-displacing predicates. I use this framework merely as a convenience, since movement-based accounts are the ones I am most familiar with. By the end of this chapter it should be clear in what sense subjects of displacing predicates are syntactically displaced, why each of the constructions discussed here presents a learning puzzle, and how inanimate subjects can offer learners a cue to discover the structure of an opaque string.

14

Syntax of displacing and non-displacing predicates

15

In some cases it will be useful to look back at some of the early accounts of these structures, for example from the Standard and Extended Standard Theory, because many of the analyses of that era were stated in terms of intuitive relations between surface word orders and “deep” semantic relations. These deep semantic relations will be useful for understanding how children might parse and acquire these structures. While the intuitions behind these early analyses sometimes remain, quietly, in modern analyses (such as the “smuggling” approach to passive, raising, and tough-movement, Collins (2005b,a); Hicks (2009)), their more explicit forebears can sometimes provide a clearer picture of the target surface parse of these opaque constructions. What all of these constructions have in common is that the mapping between thematic role and surface position is non-canonical in some sense. In the language of Theta Theory, no external θ -role is assigned by the predicate. For some of these constructions (tough-constructions and raising) there is no thematic relationship at all between the main predicate and its adjacent subject NP, because the subject has raised from a lower clause. That is, these structures are biclausal, and the main (displacing) predicate and main clause subject belong, thematically speaking, to different clauses. For others (unaccusatives, passives) there is a thematic relation between the verb and subject (these constructions are monoclausal) but it is not the “typical” one: the subject is not an external argument (agent or experiencer) of the verb, but an internal argument. Moreover, each of these constructions (excepting the passive; but see discussion of the “adjectival passive” in Section 2.4.2) has a syntactic foil, what Chomsky called “constructional homonyms”: a construction that appears identical on the surface but does have the canonical sort of thematic relation between the subject and predicate. The presence of these foils in the grammar makes the strings opaque and turns the learning of these non-canonical structures into a true puzzle: not only does the child have to determine that the relations between the words in these displacing constructions are not what they appear to be on the surface, but she must also discriminate the non-canonical structures from the nearly identical canonical ones. We will first look at the biclausal constructions, namely, raising-to-subject and tough-movement. It is in these constructions that I think my argument for inanimate subjects signaling a derived subject is the most straightforward. Clausal complements are often selected by verbs whose lexical meaning attributes to the subject some mental attitude about the clausal complement (e.g. it expresses desire, effort, emotional preference or dispreference, or denotes a communicative act) and thus require a sentient subject. Sentient subjects are animate. Therefore, the lack of this crucial property – a non-sentient

16

Syntax of displacing and non-displacing predicates

or inanimate subject – should be unexpected, and thus provides a clue that the sentence is somehow different than one would normally expect. The monoclausal constructions, unaccusatives and passives, do not work quite as neatly. Although the broad generalization holds (unaccusatives and passives easily allow inanimate subjects while unergatives and actives do not allow them as easily), there are several counterexamples among the intransitive verbs, and we will see that passives behave altogether differently from the other displacing predicates.1 2.1

Raising-to-subject and subject control: seem vs. claim

Raising-to-subject (or subject raising) is the name given to constructions in which the main verb selects a clausal complement (usually infinitive, but in some cases it can be finite) and no external argument; the main clause verb bears no thematic (semantic) relation to its adjacent subject, and the main clause subject is thematically related instead only to the predicate of the complement clause. A standard example from English is given in (1). (1)

Johni seems [t i to be a nice guy.]

In this sentence, John is semantically related only to the predicate be a nice guy; he is not an agent or experiencer (or theme) of “seeming.” Evidence for the lack of a thematic relationship between seem and John comes from the fact that seem can host an expletive (semantically empty) subject, such as weather-it or existential there, and idioms can be split around raising verbs. (Idioms are thought to be treated in the syntax as if they were a single lexical item or “chunk,” which is to say they are generated as one unit. The fact that they can be split around raising verbs is explained if their component parts do not each receive separate thematic assignments from other predicates in the sentence.) (2)

a. It seems to be snowing outside. b. There seems to be a problem with your analysis. c. The cat seems to be out of the bag. (= the secret is out)

1 Please note that I will restrict attention to constructions involving main predicates. I will not deal

with copular constructions, which quite easily permit inanimate subjects; e.g. This is a book or The book is heavy. The reason for abstracting away from these types of constructions is that I am interested in how children learn the selectional properties of main predicates, and copular constructions lack main predicates.

Raising-to-subject and subject control

17

In these respects, raising verbs like seem contrast with control verbs, like claim. In (3), John is a thematic argument of claim, and it is also semantically related to the predicate in the complement clause. (3)

Johni claims [PROi to be a nice guy.]

Unlike raising verbs, control verbs cannot take expletive subjects or split idiom chunks, underscoring the fact that these verbs select a thematic subject. (4)

a. * It claims to be snowing outside. b. * There claims to be a problem with your analysis. c. ? The cat claims to be out of the bag. (= the secret is out)

The reason the sentences in (4) are ill-formed is that the verb claim assigns a thematic role to its subject – it requires a “claimer” as its external argument. Expletive subjects cannot bear this (or any) thematic role, and if idiom chunks are generated as a single unit, they cannot have particular thematic roles assigned to subparts of the idiom. That is, in (4c) claim would assign an agent role to the subject the cat, separate from the lower predicate be out of the bag, and this is incompatible with the unitary nature of idioms. Additional differences between raising and control structures were noted in the early literature (Rosenbaum, 1967; Postal, 1974). For example, when the complement clause of the raising or control verb is passivized the semantic relationship between the active and passive versions is different for control than for raising. (5)

a. John seemed to have cooked potatoes. b. The potatoes seemed to have been cooked by John.

(6)

a. John tried to invite Mary. b. Mary tried to be invited by John.

(7)

a. John tried to build a boat. b. # The boat tried to be built by John.

Sentences (5a) and (5b) are truth-functionally equivalent; (6a) and (6b) are not. Moreover, as Davies and Dubinsky (2004) point out, when the embedded object under a control verb is inanimate, the passivized form is semantically ill-formed, as in (7b) (note: no restriction on animacy for raising, cf. (5b)). Relatedly, a further diagnostic for distinguishing these constructions is the ability of the verb to occur with an inanimate subject. Control verbs select an

18

Syntax of displacing and non-displacing predicates

agent or experiencer, which (nearly always) must be animate.2 Raising verbs place no selectional restrictions on their subject. (8)

a. John/The machine seemed to be helpful. b. John/# The machine tried to be helpful.

In the case of raising verbs, the subject of the sentence is subject to the selectional restrictions of the lower predicate only (The machine/#The rock seemed to be helpful), while in the case of control verbs, the subject of the sentence is subject to the selectional restrictions of both the control verb and the lower clause predicate. One final diagnostic I’ll mention for distinguishing raising from control verbs involves an asymmetry of scopal effects. While raising predicates exhibit scope ambiguity, such that the matrix subject can take either wide scope or narrow scope with respect to the raising predicate, control predicates do not admit the narrow scope reading (May, 1977, 1985). (The following examples are based on Wurmbrand (2001, p. 192).) (9)

Someone from New York is likely to win the lottery. a. = There is some person from New York who is likely to win the lottery. (wide scope) b. = It is likely that the person who wins the lottery (whoever that is) will be from New York. (narrow scope)

(10)

Someone from New York claimed to win the lottery. a. = There is someone from New York who claimed to win the lottery. (wide scope) b.  = It is claimed that the person who won the lottery (whoever that is) is from New York. (narrow scope)

The ability of raising predicates to support both scopal interpretations, and the corresponding failure of control predicates to support narrow scope interpretations, has been argued to provide evidence that the subject of a raising verb does not originate in its surface position, but rather raises to that position in the syntax. It can then reconstruct (through Quantifier Lowering) to its base position at the level of Logical Form (LF). Although this explanation is based upon the assumption of movement in syntax, such an assumption is neither necessary nor particularly relevant to the crucial property of 2 A very small number of control verbs permit inanimate subjects. The only such verbs I am

aware of are serve and help; see Rudanko (1989). We will return to the issue of inanimate subject arguments below in Section 2.3.1 and in more depth in Section 3.1.2.

Raising-to-subject and subject control

19

interest here. The failure of raising predicates to select a thematic subject transcends theoretical frameworks (see Asudeh (2005) and Carpenter (1997) for treatments of these scopal effects without movement). For example, in Lexical-Functional Grammar raising verbs like seem are said to project an “athematic” subject argument, and the non-expletive subject of such a verb is instead the thematic argument of a different predicate in the structure. What is important for our purposes is that raising verbs, as a group, share the property of failing to select a thematic subject and thus placing no semantic restrictions on the kind of NP that appears there. Verbs that function like seem include appear, happen, tend, turn out, going (to) (future), used (to) (past), the predicate adjectives be likely and be certain, and the preposition be about (to). These predicates do not behave alike in all respects. For example, seem and appear can take a that complement (11a) and can occur in copy raising constructions (11b) while tend, for example, can do neither (12a–12b). The raising adjective be likely can take a clause in subject position while seem cannot (for discussion of seem vs. be likely see, for example, Olsen (1981)). (11)

(12)

a.

It seems that/like Betty stole the cheese.

b.

Betty seems like she stole the cheese.

a. * It tends that/like Betty steals cheese. b. * Betty tends like she stole the cheese.

(13)

a.

That Betty will steal cheese again is likely.

b. * That Betty stole cheese again seems.

Nevertheless, these predicates all arguably fail to bear a thematic relationship to their subject, hence the patterns in (2). Verbs that function like claim include want, try, decide, love, and hate, and many others. These verbs also display a certain degree of variability (see Section 3.3.1.1), but none of them can take an expletive subject or split an idiom; that is, they behave like claim in the constructions in (4). 2.1.1 The structure of raising To account for the lack of a semantic relationship between raising verbs and their adjacent raised subjects, derivational syntactic theories have employed a mechanism of NP-movement (Chomsky and Lasnik, 1977). The subject NP is generated in the lower infinitive clause at the level of D-structure (which encodes thematic relations), as shown in (14). The main clause subject position

20

Syntax of displacing and non-displacing predicates

is underlyingly empty (indicated by e), reflecting the fact that the main verb (seem) does not select an external argument. IP

(14)

I

e I

VP V

IP

seems

I

NP John

I

VP

to

be a nice guy

At the level of S-structure, which represents other syntactic properties such as Case assigment, the subject of the infinitive raises to the (previously empty) main clause subject position, shown in (15), and leaves behind a trace of movement (t) in its original position. (15)

IP I

NP Johni

I

VP V seems

IP I

ti I

VP

to

be a nice guy

Within Government-Binding (GB) theory this type of movement (known as A-movement, movement into an argument position, as opposed to A movement, which includes wh-movement) was motivated in the following way. NPs are required to bear Case at S-structure according to the Case Filter (Chomsky, 1980). Case was argued to be assigned in a particular structural configuration (called government) by the tensed Infl node (node I in the tree, for Inflection), but not by untensed Infl (for example, the I of an infinitive clause). Since the infinitive is untensed, the subject of the infinitive (John in

Raising-to-subject and subject control

21

the D-structure representation in (14)) lacks Case; it does, however, have a theta (θ ) role, having been assigned its θ -role at D-structure by the embedded predicate. Furthermore, the main clause subject position is empty, there being no external argument of the raising verb, and so the NP John is free to move into that position, where it can now get Case from the tensed Infl of the main clause (and would not get any additional θ -role, since seem does not assign an external θ-role to begin with).3 Thus, both Case Theory (which pertains to the licensing of overt NPs in particular structural positions – NPs that are not assigned Case in their D-structure position must move to get it at S-structure) and Theta Theory (which governs the assignment of θ -roles to NPs and includes the restriction that each NP may be assigned, at D-structure, one and only one θ-role) are central to this account of raising constructions. A further component of GB that is particularly relevant to our discussion of raising and control constructions is the Projection Principle. The Projection Principle requires that the subcategorization requirements of predicates be projected at all levels of structure (D-structure, S-structure, and LF). In other words, if a verb selects a subject argument, that argument is present at all levels of structure (though it may move out of subject position, or it may be unpronounced, etc.). If a verb fails to select a subject argument, it will not have a subject argument at any level of structure, though an NP may move into its subject position at S-structure. In this case, that (surface) subject is still not an argument of the verb, since it was not selected by the verb at D-structure.4 The final (and related) crucial piece of the puzzle is the Extended Projection Principle: the requirement that all clauses have a subject (though it may not be pronounced). This requirement stems from the observation that “not all subjects result from the lexical requirements of verbs” (Davies and Dubinsky, 2004, p. 182). That is, weather predicates and existentials require a subject (it, there), but this pleonastic subject is not really a thematic argument of that predicate. Although Minimalist syntax has moved away from the three-level model of the GB era (D-structure, S-structure, LF), raising structures have a similar treatment within this framework as they did under GB. The raising verb is merged with an infinitive predicate and the subject of that infinitive raises 3 The construction known as copy raising (Richard seems like he is in trouble) must be accounted

for in a different way, since both the matrix subject and the pronominal copy are in Case positions. For more on copy raising, see Section 3.3.1. 4 Chomsky (1980) also discusses the discrepancy between the syntactic and semantic notions of “subject.”

22

Syntax of displacing and non-displacing predicates

by movement into the main clause. However, the introduction of phase theory (Chomsky, 2001) led to a revision in the formal distinction between predicates that fail to project an external argument (raising predicates, unaccusative verbs, and passives – these predicates are “defective”) and predicates that do project an external argument (active transitives and unergatives – these are non-defective) that was not as explicitly present in earlier versions of the theory.5 More specifically, the distinction between non-defective and defective phases splits those predicates that assign an external argument role from those that do not. The intuition behind phases is that phases are “propositional,” that is, they are complete packages of the arguments required by a predicate. Non-defective phases are said to be φ-complete; that is, their argument roles are saturated and their agreement features are valued (“checked” in previous terminology). However, if the light verb v is φ-incomplete (passive), then V is defective, as is Tv selected by V (in raising/ECM constructions) . . . Tv is raising Tense, which is assumed to be defective in having only a person feature, as opposed to finite Tense, which has full φ-features. (Chomsky, 2001, p. 9)

In other words, because unaccusative/passive/raising predicates do not select an external argument (subject), they lack full φ-features, and therefore their vP is a “defective” phase. Therefore, higher projections can be merged to vP before material is sent to Spell-Out. This results in movement being allowed out of this lower clause into the matrix clause. One problem for an analysis of raising within Minimalism is how to raise the subject past an overt experiencer argument, as in (16): (16)

John seems [to Mary] to be a great guy.

The Minimal Link Condition (Chomsky, 1995, p. 264) (similarly Rizzi’s (2001) Relativized Minimality) limits all types of syntactic movement to the “shortest move,” such that movement over an intervening structural position of the same type is blocked (filled intervening A-positions block A-movement, filled intervening A -positions block A -movement). In the case of (16) that intervening position is occupied by the experiencer DP (to Mary), which should block movement of the subject John. Collins (2005a) suggests a solution to this problem by invoking a mechanism he terms “smuggling.” The idea behind smuggling is that if a phrase that needs to move (e.g. an NP/DP) is 5 Chomsky (2001) does not discuss tough-movement.

Raising-to-subject and subject control

23

contained inside a larger phrase of a different type (e.g. a VP, which is not an argument and therefore does not undergo A-movement), it can be “smuggled” past a barrier to movement (the experiencer) by moving within the larger phrase as a whole. In the case at hand, the whole VP moves, carrying with it the NP/DP argument.6 Applying this approach to raising constructions, we have the following series of structures in the derivation, adapted from Collins (2005a, p. 295). Collins proposes that following the application of Merge, we have the structure in (17). ApplP stands for Applicative Phrase, a projection typically associated with benefactives or indirect objects, and Collins argues the experiencer of a raising verb is this sort of argument (for discussion of applicatives see Pylkkänen (2008)). (17)

IP I I

vP v ApplP

v Exp

to Mary

Appl Appl

XP X X

VP V

DP John

V

IP

seems

to be a great guy

Collins designates the XP above the embedded VP as a functional projection that is needed in order to host the lower IP, which is extraposed into the SpecXP position. This is shown in (18). 6 I use the label DP in the trees below, following Collins, but I do not intend any meaningful

distinction between the labels DP and NP.

24

Syntax of displacing and non-displacing predicates

(18)

IP I I

vP v ApplP

v

Appl

Exp to Mary Appl

XP X

IPi to be a great guy

X

VP V

DP John

V



seems

Next the whole VP containing John and seems is raised into SpecvP. (19)

IP I

I

vP

v

VPj DP

V

John

V seems



ApplP

v

Appl

Exp to Mary Appl

XP X

IPi to be a great guy

X

tj

Raising-to-subject and subject control

25

In the final step, the DP John, which has been smuggled out of the embedded clause, can raise into the matrix subject position. (20)

IP I

DP John I

vP

v

VPj V

DP V seems

ApplP

v

Appl

Exp to Mary Appl

XP IPi to be a great guy

X XP

tj

We will see this mechanism employed for Minimalist treatments of toughconstructions (Hicks (2009); Section 2.2.1) and passive (Collins (2005b); Section 2.4.1). While the theory needs a way of allowing raising past experiencers to account for the grammaticality of (16) in languages like English, there is a sense in which the theory should not make these structures “too easy.” There are languages, like Icelandic, which allow raising but not over experiencers, and as we will see in Chapter 5 (Section 5.3.1) this type of raising is difficult for children, while raising without an experiencer is not.

2.1.2 The structure of control The syntactic structure of subject control constructions has been relatively straightforward and stable across the various changes seen by syntactic theory over the decades. Previously known as Equi(valent)-NP Deletion (Rosenbaum,

26

Syntax of displacing and non-displacing predicates

1967), the Extended Standard Theory brought in the projection of the null pronoun PRO as the subject of the infinitive clause, replacing the Equi-NP Deletion rule. This analysis of subject control remained the standard one through the Government-Binding era. (21)

Johni tried [PROi to win the election.]

Here the main clause subject, John, is said to “control” the reference of PRO, as indicated by coindexation. However, both John and PRO are generated independently, each selected by a different predicate (John by try and PRO by win the election) and each assigned its own θ -role by that predicate. Thus, John is generated in the main clause and there is no A-movement. PRO is generally interpreted as coreferent with (“controlled by”) a c-commanding referring expression (John in (21)), but it can also have arbitrary reference (e.g. PROarb to sell the house now would be a mistake). Some of the Minimalist literature on control has explored the question of whether there is in fact a syntactic difference between raising and control structures. For example, Hornstein (1999) analyzed control, like raising, as resulting from movement. One of the main differences between PRO and NP-trace (the trace of an NP that has undergone A-movement, as in raising) is that PRO receives its own θ -role from the embedded predicate, separate from the θ-role of PRO’s controller, while NP-trace and the raised NP share a θ-role. Hornstein argued, however, that the requirement that there be a strict one-toone mapping between arguments and θ -roles (i.e. each argument gets only one θ -role and each θ-role is mapped to only one argument), as was assumed in the Theta Theory of GB, is not a theoretical necessity. (Jackendoff (1972, 1990) also rejected the biunique mapping between thematic roles and NPs, though for different reasons; see discussion in Section 3.2.2.) Instead, if a single argument is allowed to bear more than one θ-role it is no problem for the controller of PRO to be derived by movement: (22)

a. John1,2 tried1 [to leave2 .] b. John1 seemed [to leave1 .]

On Hornstein’s account the difference between (22a) and (22b) is that in (22a) the subject John receives two θ-roles, while the subject of (22b) receives only one (from the lower predicate).7 Importantly, though, Hornstein’s 7 See arguments by Culicover and Jackendoff (2001), and different ones by Landau (2003),

against Hornstein’s account.

Raising-to-subject and subject control

27

argument is orthogonal to the claims I am making. The crucial difference between raising and control predicates, as far as I am concerned, is that control predicates select their subject while raising verbs do not, and this asymmetry is maintained in Hornstein’s analysis of these predicates. Thus, whether both predicates involve movement of the subject, or only raising predicates do, the asymmetry between these sorts of predicates in terms of the semantic relationship they bear to their surface-adjacent subject remains. Just as Hornstein’s unification of raising and control constructions as involving movement is compatible with the claims made here, syntactic frameworks that do not employ movement at all are likewise compatible with my story. In Lexical-Functional Grammar (LFG), which is non-transformational, raising and control constructions do differ in their syntax despite the lack of movement in either one of them. In particular, while LFG posits identical c(onstituent)structures for both types of sentences, shown unannotated in (23), the two types of predicates are associated with different a(rgument)- and f(unctional)structures; the respective a-structures are illustrated in (24) (Bresnan, 1982b, 2001). (23)

S NP John

VP V seems/claims

VP to

V agree

(24)

a. seem (subj) b. claim

The difference between the a-structures in (24) is that the subject of seem is outside of the angled brackets, indicating an athematic relation between it and the verb seem; the only component selected by seem is the propositional complement (xcomp) (seem can also take an oblique argument if there is an overt experiencer or perceiver of the complement clause – seems to X; Bresnan (1982a)). For claim, on the other hand, both the subject and the propositional complement are thematically related to (selected by) the predicate. This asymmetry is reflected also in the f-structure representations for these

28

Syntax of displacing and non-displacing predicates

types of predicates. Thus, a syntactic asymmetry between raising and control is maintained, in spite of the lack of movement. 2.1.3 Raising-to-object and object control: expect vs. persuade Raising-to-object verbs are verbs that take both a direct object and an infinitive complement, as in (25), but they do not assign a θ -role to their direct object. (25)

Murdoch expected Parliament to interrogate him.

Other verbs that behave like expect include want, believe, and need. They have in common that, although they do not select a direct object (Parliament is not the thematic patient of expect in (25) above), an NP appears between the verb and its infinitive complement. Historically, the theoretical debate around these constructions has centered on resolving the question of whether that “object” is truly a syntactic object (i.e. a complement of the V head in the main clause; this is the raising-to-object analysis) or whether it is a syntactic “subject” inside the infinitive clause (this is the Exceptional Case Marking analysis). What is agreed upon, however, is that these verbs do not select a direct object underlyingly. In this, raising-to-object verbs contrast with object control verbs, which do select a thematic object. Persuade is such a verb. (26)

Murdoch persuaded Parliament to leave him alone.

Verbs that behave like persuade in this regard include tell, ask, convince, allow, and force. The main criterial distinction between raising-to-object and object control parallels the distinction between raising-to-subject and subject control: in both cases the distinction turns on whether there is a thematic relationship between the verb and the subject/object of the main clause. In the case of raising-toobject, the matrix verb does not assign a θ-role to the object position underlyingly; rather, the surface object raises to that position from the subject of the infinitive.8 In object control the matrix verb does θ -mark its object, which (in the normal case) controls the reference of PRO, the subject of the infinitive.9

8 In the ECM analysis there is no direct object at all, only an infinitive clause complement with

an exceptionally case-marked overt subject. 9 Some languages, like English, have control verbs that take a direct object argument, but the

subject controls the reference of PRO in the embedded clause. Such a verb is promise (Johni promised Billj [PROi/∗j to cut the grass]).

Raising-to-subject and subject control

29

Some diagnostics for distinguishing the two constructions are given below. All of them follow from the basic thematic difference between the two types of verbs mentioned above. (27)

a.

John expected it [t to rain]. (raising-to-object)

b. * John persuaded it [PRO to rain]. (object control) (28)

a.

John expected there [t to be a solution].

b. * John persuaded there [PRO to be a solution]. (29)

(30)

(31)

a.

John expected Maryi [t i to photograph Bill].

b.

John expected Billi [t i to be photographed t i by Mary]. (=29a)

a.

John persuaded Maryi [PROi to photograph Bill].

b.

John persuaded Billi [PROi to be photographed ti by Mary]. ( =30a)

a.

John expected the shit to hit the fan.

b. # John persuaded the shit to hit the fan.

In (27)–(28) a pleonastic argument is permitted in the object position of expect since no θ -role is assigned by the verb to that position; persuade does assign a θ-role to its object so pleonastic arguments are banned. In (29) the embedded passive is truth-functionally equivalent to the embedded active version of the sentence. This is not true for the object control pair in (30). Finally, like raising-to-subject, raising-to-object constructions allow idioms to be split, while control constructions disallow this (see (31)). A vast literature exists on these constructions, and debates over their correct structural characterization have been contentious. A lucid and detailed summary of this debate is to be found in Davies and Dubinsky (2004), and the reader is referred to that work for a thorough explanation. Although raisingto-object involves the displacement of an argument in the syntax, as well as surface opacity with respect to the underlying structure (due to the constructional homonym of object control), I will not focus on these constructions in this book. The main reason for abstracting away from them is that the position into which the displaced argument is moved is an internal position (complement of the verb) rather than an external position (subject). In this, raising-to-object is distinctly different from all of the other constructions under consideration

30

Syntax of displacing and non-displacing predicates

here. Since the internal argument position is prototypically associated with an inanimate (or less animate) NP, in contrast to the external argument position, the effect of encountering an inanimate NP in this position is less easy to predict. Nevertheless, we do see an asymmetry in object animacy between raisingto-object and object control similar to what we have seen with respect to subject animacy between raising-to-subject and subject control. Raising-toobject predicates always allow an inanimate NP in their object position, while object control predicates do not. (32)

a.

John expected the book to be boring.

b. # John persuaded the book to be boring.

The occurrence of an inanimate NP in object position is nothing unexpected or non-canonical. Instead, what is perhaps unexpected is that object control verbs generally require an animate (sentient) object (cf. (32b)). Thus, although the precise means by which animacy distinguishes raising-to-object from object control predicates is not exactly the one found in the other constructions discussed in this book, it is of a piece with the more general learning strategy advocated here. Patterns of NP animacy are highly informative of predicate subclass, and so learners do well to pay attention to these patterns. For some predicates the restriction will be on the subject NP, for others it will be on the object NP.

2.2

Tough-constructions: easy vs. eager

One of the early observations made within modern linguistic theory was that in order for the theory to be truly comprehensive (that is, to be both descriptively and explanatorily adequate) the theory would have to account for how two sentences that were nearly identical on the surface could differ so radically in meaning. Lees (1960) illustrated with the example of the “multiply ambiguous adjectival construction”: (33)

a. John is easy to please. b. John is eager to please.

Both Lees (1960) and Chomsky (1964) used these examples to illustrate the inadequacy of surface-oriented structures such as (34).

Tough-constructions

31

S

(34) NP John

VP is

Pred Adj easy/eager

to

VP please

The problem with (34) is that this type of structure does not account for the different interpretations or transformational properties of these sentences: (35)

(36)

tough-adjective construction a.

John is easy to please.

b.

It is easy to please John.

control adjective construction10 a. John is eager to please. b. * It is eager to please John.

The set of tough-adjectives is quite small: it is limited (at least in English) to the adjectives hard, easy, simple, difficult, and impossible (Anderson, 2005). Control adjectives include emotive adjectives like happy, excited, glad, willing, and afraid. One of the empirical diagnostics for distinguishing toughadjectives from control adjectives is that the former, but not the latter, permit an expletive it subject as in (35b). Parallel to the raising-to-subject construction already discussed, tough-constructions also permit an inanimate referential subject, while control adjectives do not. As Lees (1960) noted, control adjectives are generally “confined to those which occur as predicates to animate subjects” (emphasis mine). (37)

(38)

a.

John is easy to see.

b.

The mountain is easy to see.

a. John is eager to see. b. # The mountain is eager to see.

10 I place a ‘*’ for ungrammaticality next to (36b) because the it subject is meant to be an exple-

tive subject. If it is interpreted as a referential pronoun the sentence is still ill-formed, but semantically so.

32

Syntax of displacing and non-displacing predicates

Tough-adjectives also permit a gerundive or infinitive clause subject (which itself contains an arbitrary PRO subject), and no infinitive complement. (39)

[Seeing the mountain/To see the mountain] is easy.

(40)

* [Seeing the mountain/To see the mountain] is eager.

In contrast, control adjectives permit a simple referential (animate) NP subject and no infinitive predicate (41), while tough-adjectives do not (42). At least, tough-adjectives cannot occur in this simple matrix construction on the basic meaning of the adjective – other connotations of the adjectives, such as sexual or behavioral connotations, are called to mind in these cases (e.g. Mary is easy ‘easy to get into bed’; That child is difficult ‘behaves badly’). Interestingly, certain simple referential inanimate subjects are acceptable with tough-adjectives without an infinitive complement; in many of these cases there is an implied event that is easy or hard, such as that in (43), which could be paraphrased as Taking that test was easy/hard/difficult.11 (41) (42) (43)

John is eager/happy/anxious/excited. # John is easy/hard/difficult. That test was easy/hard/difficult.

2.2.1 Structure of tough-constructions Early accounts of tough-constructions, such as Lees’ (1960) and Rosenbaum’s (1967), built on the intuition that the surface subject belonged, semantically, to the lower verb (see also Bresnan, 1971; Postal, 1971; Partee, 1977) and therefore started out in that position in the syntax. Rosenbaum’s analysis of this construction posited the subject being base-generated as the object of the embedded verb (please in (44)) and raising up to the subject position of the main clause. The transformation took place in a couple of steps. (44)

a. [to please John] is tough Extra-position of the infinitive and expletive insertion → b. It is tough [to please John]. Object-to-subject raising → c. Johni is tough [to please ti ].

11 Certain referential animate NPs can take a tough-adjective with no embedded complement

(e.g That teacher is hard/easy), where the implication is that the teacher gives hard/easy work.

Tough-constructions

33

This kind of account preserves our intuition about the thematic relationship between the lower predicate and the surface subject (that John is the semantic object of please), but as syntactic theory developed more explicit constraints on the movement of arguments in the course of a derivation a number of problems with this analysis came to light. In particular, NPs are assigned Case under government, so if John originates as the complement of the verb it would get Accusative case there, making its raising to subject position unmotivated (recall that, normally, the purpose of NP-movement is to get Case), and it would leave its bearing Nominative case (rather than Accusative) unexplained. In addition, an NP-trace must be locally bound by its antecedent. Therefore, while (45a) was unproblematic, (45b) presented several problems for the theory: (45)

a. It is tough to please John. b. Johni is tough to please t i .

(45a) satisfies both Theta Theory and Case Theory since please assigns both its θ -role and Case to John under government. There is no movement, so NPtrace is not present, and an expletive subject is inserted at S-structure. However, in (45b), John would not need to raise to get Case, since it could get Case in its D-structure position. Moreover, if it did raise to subject position its trace would not be locally bound. The solution presented in Chomsky (1977) avoided these pitfalls by base-generating John in the matrix subject position and postulating a null wh-operator that moved from the complement of please to a topic position within that lower clause (movement to a topic position constitutes an instance of A -movement). (46)

John is tough [whi [PRO to please t i ]].

While there is some independent evidence for the A -movement account (from island effects; Chomsky (1977)), this analysis is odd in that it generates the subject in a position where it clearly does not get its semantic interpretation (see also Lasnik and Uriagereka (1994) for additional critiques of this analysis). Throughout the 1980s Chomsky’s (1977) analysis of tough-movement was largely retained, in spite of its problematic nature. Other attempts to analyze this construction, such as the movement account by Brody (1993) and the deletion analysis of Lasnik and Fiengo (1974), had their own problems.12 12 Brody’s analysis required A -movement followed by A-movement, known as “improper move-

ment” (Chomsky, 1981), and the deletion account rests on questionable data, such as that

34

Syntax of displacing and non-displacing predicates

As Holmberg (2000) noted, the construction was in principle “unexplainable” in the Government-Binding framework of that time. In fact, the A -movement treatment of tough-constructions was so problematic that it served as one of the catalysts for abandoning D-Structure entirely in early Minimalist work (Chomsky, 1995, p. 188). Removing D-Structure from the formalism may have removed some of the technical reasons why the previous analysis failed, but it did not immediately produce a better analysis. Hicks (2009) has made the most recent attempt at resolving both the problem of why the underlying object does not get Case and how movement is motivated, while still invoking movement as opposed to deletion or surface insertion. Hicks exploits the notion of “smuggling” (Collins, 2005a,b) and argues for a more complex structure for null operators. He returns to the notion that it is a null operator that is raising in the lower clause, but with a crucial difference. Hicks proposes that the null operator is actually a predicate within a complex DP that bears a wh-feature and takes a simple DP (e.g. John) as its single argument. The configuration of the complex DP looks like this (omitting feature notation): DP

(47) D

NP N

DP

Op

John

The DP John does not check its case feature in this configuration because it is not “visible” to V. This complex DP then undergoes A -movement to the SpecCP of the embedded clause (by virtue of its wh-feature). Now that John has been “smuggled” out of the lower vP to SpecCP (the edge of a phase), the simple DP John can raise (via A-movement) to SpecTP of the matrix clause where it gets Nominative case. Thus, Hicks essentially invokes A -movement followed by A-movement but avoids an Improper Movement violation by claiming that it is two different DPs undergoing the respective movements: the complex DP of the operator undergoes A -movement, but only the simple DP John undergoes A-movement.

involving idioms. For example, the deletion account is argued to be supported by the fact that idiom chunks are not uniformly acceptable in tough-constructions. Lasnik and Fiengo (1974) find the sentence Headway is easy to make on problems like these marginal and Tabs were easy to keep on Mary ungrammatical, but I find the first perfectly fine and the second only slightly degraded.

Tough-constructions

35

A simplified structure for the sentence John is tough for us to please, based on that given in Hicks (2009, p. 551), is presented here: (48)

TP T

DPl John T is

aP AP

a toughj

a

A

PP for us

CP

tj

C

DPk D

NP

C

N

DP

Op

tl

TP PRO to please tk

The assignment of the θ -role to the subject potentially remains a problem on Hicks’ account. He argues that the tough-construction subject gets its θ-role directly from the operator itself, not from the verb. On his account the θ -role cannot come directly from the verb for the same reason Case is not assigned by the verb: the subject argument is not “visible” to the verb when it enters the derivation as part of the null operator’s complex DP. One possibility is that the verb’s θ -role reaches the subject via “percolation,” or some other “composite” mechanism between the verb and null operator. Hicks favors the view that the θ-role comes exclusively from the operator, based on evidence having to do with lack of variability in which θ-roles get assigned in this construction. That is, he claims that subjects of tough-constructions are uniformly themes, but there is some counter-evidence. In fact, tough-constructions allow stimulus subjects (John is easy to see), experiencers (John is easy to frighten), and benefactees (John is easy to cook for), among others.13 Moreover, although one of the apparent advantages of the smuggling approach to tough-movement was a return to the intuition that the subject starts out in the object position 13 I thank Ash Asudeh for pointing these out to me, and for discussion on this point.

36

Syntax of displacing and non-displacing predicates

of the lower clause, it is not clear how this particular analysis yields the right semantic relationship. That is, if the subject does not in fact get its θ-role from the embedded verb, what gives rise to our intuition that the syntactic subject is really the semantic object of that verb?14 Although the exact syntactic analysis of tough-constructions remains elusive, it will be sufficient for our purposes to note the semantic asymmetry between the relationship between the subject and main predicate in toughconstructions as compared to control adjective constructions, and to accept that this asymmetry translates into a structural difference between the two constructions. That is, children will have to distinguish the two underlying structures in spite of their surface similarity and will have to determine that the subject of tough-adjectives, but not that of control adjectives, is displaced. 2.2.2 Related constructions There are a number of other constructions which resemble tough-constructions on the surface, but which involve somewhat different structural properties. For example, the sentences in (49) (examples from Lees (1960)) involve an adjectival predicate in the main clause and an embedded infinitive clause. (49)

a. He’s free to go. b. He’s too old to send. c. It’s too hot to eat.

Lees (1960) noted that some of these constructions involve a subject gap (49a), others involve an object gap (49b), and still others are ambiguous (49c) (see also Bolinger (1961) and see Anderson (2005, ch. 2) for an overview of these constructions and their structural properties). Interestingly, in some cases ambiguity can be (partially) resolved by the animacy of the subject. Under normal circumstances (and barring cannibalism), (50a) will be an instance of an object gap, while (50b) will involve a subject gap. Likewise (51a) means that it was good to read this book, not that it was good of the book to read something else, and (51b), though ambiguous without an overt direct object, most naturally means that it was good of the man to help (someone). (50)

a. The sandwichi is ready [PROarb to eat t i .] b. The girli is ready [PROi to eat e.]

(51)

a. This booki was good [PROarb to read t i .] b. The mani was good [PROi to help (the old lady).]

14 See Richards (2001) for a different Minimalist account of tough-movement.

Tough-constructions

37

Our intuitions are partially restricted, of course, by our expectations about the world: our non-cannibalistic culture in the case of (50) and our knowledge of what books and men are capable of doing in the case of (51). Given the right predicate and context an animate subject can have an object-gap reading. (52)

a. Johni is ready [PROi to push e]. b. Johni is ready [PROarb to push t].

My own reading is that (52a) is the preferred reading, but that (52b) is also possible. In contrast, an inanimate subject, even when used with a semantically neutral lower verb, still allows only the object-gap reading. (53)

a. * The balli is ready [PROi to roll]. b.

The balli is ready [PROarb to roll t i ].

The semantic difference is subtle, but (53a) gives a distinctly unergative feel, as if the ball is ready in a psychological sense, whereas (53b) means only that things are all set for the ball to be rolled by someone. On the other hand, the distinction is dependent not only on animacy but also on the semantics of the lower predicate, and in particular the likelihood of the inanimate NP to be capable of self-propelled motion. I find vehicles quite natural with a subject-gap interpretation. (54)

The bus is ready to leave.

It seems to me that (54) is actually ambiguous, since it could mean that the bus is ready for us to leave it (i.e. an object-gap reading), but I find the subjectgap reading more natural. As we will see in Chapter 6 ready is most frequently used with a subject-gap meaning in input to children. Crucially, ready is not a true tough adjective, as it does not allow an expletive subject. (55)

* It/there is ready to roll the ball.

Why ready allows both a subject- and an object-gap reading, but does not allow an expletive, is an interesting puzzle. If it does not allow an expletive subject, that should mean that it selects an external argument. But if it takes an external argument, how can a lower object raise into its subject position? This position would need to be empty in order for the object to move into it. Note that ready also differs from tough-adjectives in that it easily permits passivization of the internal argument, whereas tough-adjectives do not.

38 (56) (57)

Syntax of displacing and non-displacing predicates The ball is ready to be rolled. * The ball is easy to be rolled.

I will not propose a syntactic analysis of ready constructions, but let it suffice to say that ready and other adjectives like it (good, nice) form a separate subclass from the tough-adjectives. It seems to me that subject animacy could still be a useful cue in guiding learners towards an object-gap or subject-gap reading of sentences containing these adjectives (inanimate subject → more likely object gap; animate subject → more likely subject gap), but it will not necessarily be useful in discriminating this separate subclass of adjectives from either the tough or the control class. As noted above, degree constructions (e.g. too hot) are ambiguous, and here the role of animacy is not so clear. Rothstein (1991a,b) observes that degree constructions allow either a control structure (58a) or a tough-construction type structure (58b) (examples cited in Anderson (2005)). (58)

a. Theresai is too intelligent [PROi to make that kind of mistake]. b. Theresai is too intelligent [Opi [PROarb to select t i for guard duty]].

Again, though, degree constructions differ from true tough-constructions in that they disallow an expletive subject (*It is too intelligent to select Theresa). A further type of adjective that is sometimes grouped together with toughadjectives is pretty, as in Birds are pretty to look at, since only the object-gap reading is possible (this is the object deletion construction; cf. Lasnik and Fiengo (1974)). However, Anderson (2005) cautions against this grouping, since the matrix subject is the thematic subject of the main predicate (Birds are pretty) and thus not underlyingly an empty position. In the syntactic framework in which Lasnik and Fiengo were working the matrix subject of tough-constructions was also argued to be base-generated in that position (see Section 2.2.1), and in fact Lasnik and Fiengo argued that both tough-adjectives and adjectives like pretty involved object deletion. However, since I am adopting the view that tough-adjectives do not select an external argument, this would constitute an important difference between tough and pretty. Based on these considerations it would seem that expletives are the crucial cue for distinguishing the predicates of interest (tough-adjectives) rather than inanimate subjects. It is true that expletives distinguish tough-adjectives from other adjectives that participate in object-gap constructions. But inanimate subjects are broadly suggestive of an object gap in all of these cases. Therefore, while a string like (59) will not unambiguously signal a tough-adjective per se,

Unaccusatives and unergatives

39

it will indicate to the learner that the string is very likely to be an object gap construction and very unlikely to be a subject control construction. (59)

NPinanimate is Adjective [to Predicate]  = NPi is Adjective [PROi to Predicate]

Once this split has been made, further evidence of an adjective occurring with an expletive subject can narrow down the possible categories of the adjective: if it occurs with an expletive subject, then it is a tough-adjective. In Section 2.5 we will come back to the general reasons for focusing on inanimate rather than expletive subjects in this book. 2.2.3 Structure of control adjective constructions Turning now to the tough-construction’s constructional homonym, sentences with adjectives like eager have been fairly straightforwardly analyzed as involving subject control. Just as with subject control verbs, since Chomsky (1981) subject control adjectives involve a base-generated subject “controlling” the reference of PRO, the (ungoverned) empty subject of the embedded clause. (60)

Johni is eager [PROi to please e].

Since John receives its θ -role from eager and Case from finite T there is no movement. The only thing left to explain is the null object of please, which presumably is understood with arbitrary reference, the same way the object of a verb like eat can be phonologically null but understood to mean “something,” just as John ate means John ate something. In other words, this is not an issue peculiar to control constructions. 2.3

Unaccusatives and unergatives: arrive vs. dance

Having discussed the syntax of two biclausal displacing predicates, raisingto-subject verbs and tough-adjectives, let’s now turn to monoclausal constructions. First we will focus on intransitive verbs. Unaccusative verbs are intransitive verbs that select, underlyingly, a direct object rather than a subject. An example of an unaccusative verb is fall. (61)

The girl fell.

The following section presents empirical evidence to justify this claim, which is perhaps not very intuitive for English speakers. Other unaccusative

40

Syntax of displacing and non-displacing predicates

verbs include arrive, go, come, break, and open (the latter two can also be transitive, as will be discussed below). Unaccusative verbs contrast with other intransitive verbs whose lone argument is an underlying subject, such as laugh. These verbs are called “unergatives.” (62)

The girl laughed.

Examples of other unergative verbs are dance, sleep, and jump. Unlike the previous sections of this chapter, I will discuss accounts of unaccusative and unergative constructions side-by-side, first focusing on the semantic intuition behind the distinction and then presenting some formal accounts of the respective constructions. 2.3.1 A semantically-driven syntactic distinction The distinction of intransitive verbs into the two classes of unaccusative and unergative verbs in modern linguistics is originally due to Perlmutter (1978).15 Perlmutter’s dichotomy turned on whether the sole argument of these verbs was internal (“initial 2” in Relational Grammar) or external (“initial 1”). He sought a semantic basis for the distinction. Broadly, he claimed that unergative verbs were “predicates describing willed or volitional acts” including “manner-of-speaking verbs” and “sounds made by animals” as well as “certain involuntary bodily processes.” Relevant to our purposes, these are predicates that would normally take an animate argument. Unaccusative verbs, instead, include verbs of existence, “non-voluntary emission of stimuli that impinge on the senses,” aspectual verbs, and inchoatives (melt, freeze). In addition, Perlmutter included adjectival predicates denoting sizes, colors, and so forth, but also states of mind. Thus, the class is not inherently restricted to taking inanimate arguments, but the preponderance of unaccusative predicates, according to Perlmutter’s characterization of the split, are grammatical with an inanimate subject. In this they contrast with the unergative predicates. Perlmutter also noted that a good many predicates are ambiguous, being able to behave either as unaccusative or unergative predicates. Here the assignment hinges to a large degree (perhaps exclusively, though Perlmutter does not specify this) on the animacy of the subject. 15 Precursors to Perlmutter’s explicit distinction are found in Hall (1965) (cited in Dowty (1991))

and in Case Grammar (Fillmore, 1968; Huddleston, 1970), where the single inanimate argument of certain verbs like open or melt was labeled with Objective case, corresponding roughly to the theme thematic role in later work; please see further discussion in Section 3.2 below.

Unaccusatives and unergatives (63)

41

a. The wheels slid on the ice. (unaccusative) b. Joe slid on the ice. (unergative or unaccusative)

(64)

a. The train roared as it approached. (unaccusative) b. The lion roared as it approached. (unergative)

In (63a), the action is involuntary and therefore the verb unaccusative; in (63b) the action can be volitional or involuntary.16 Similarly, in (64a) the roaring comes from the approach of the train, while in (64b) the roaring comes from the lion itself, not from its approach. Therefore, Perlmutter considers (64a) to be unaccusative and (64b) to be unergative. Although Perlmutter suggested that the Unaccusative Hypothesis was universal, he noted that variation can be found where a particular lexical item in a language admits or does not admit a volitional interpretation. For example, the English verb travel can have a volitional or non-volitional meaning (John travelled around the world, The package travelled for two weeks), but in Dutch the same verb, reizen ‘travel’ can only be volitional (the Dutch equivalent of The package travelled for two weeks is ungrammatical; cf. the English verb journey, which requires an animate subject). Further cases of “mismatches” (verbs that are classified as unaccusative in one language but unergative in another), as well as arguments for classifying verbs as unaccusative on a semantic as opposed to syntactic basis, are discussed extensively in Levin and Rappaport Hovav (1995). See also Van Valin (1990) for a semantically-driven approach. Perlmutter’s distinction is couched in semantic terms, but there are known syntactic effects of this split. His own illustration of impersonal passives in Dutch demonstrates this, where the construction admits unergative verbs but not unaccusatives. (65)

(66)

Er wordt hier veel geskied. it was here much ski-part “It is skied here a lot” * In dit weeshuis wordt er door de kinderen erg snel gegroeid. in this orphanage was it by the children very fast grow-part (“In the orphanage it is grown very fast by the children”)

Crucially, a given verb may admit the impersonal passive or not depending on whether the argument is interpreted as volitional (see also Zaenen (1988)). 16 Note that, in contrast, a sentence like Joe slid into third base (which was Perlmutter’s example)

is unergative, since it is understood that Joe voluntarily slid into the base, not that he slipped and slid by accident, a scenario that is plausible when treading on ice.

42

Syntax of displacing and non-displacing predicates

(67)

a. De kinderen hebben lekker op het ijs gegleden. the children have with enjoyment on the ice slide-part “The children enjoyed sliding on the ice” b. Er werd door de kinderen lekker op het ijs gegleden. it was by the children with enjoyment on the ice slide-part “It was enjoyed sliding on the ice by the children”

(68)

a.

De sneeuw is van het dak afgegleden. the snow is from the roof slide-off-part “The snow slid off the roof” b. * Er werd door de sneeuw van het dak afgegleden. it was by the snow from the roof slide-off-part (“It was slid off the roof by the snow”)

A complicating factor in defining the split between unaccusatives and unergatives has to do with verbs of sound, light, smell, and substance emission (e.g. roar, glow, stink, and ooze). Although these verbs appear to be unaccusatives on Perlmutter’s diagnostic (“non-voluntary emission of stimuli that impinge on the senses”), Levin and Rappaport Hovav (1995) argue that they are in fact unergatives. In particular, verbs of emission pattern with other “internal causation” verbs in not having a transitive counterpart (*The electrician glowed the lightbulb). And yet, these verbs take inanimate subjects, which I have argued is a hallmark of displacing predicates. As we will see in Chapter 3 verbs of emission display some hybrid behavior of unaccusatives and unergatives, making them difficult to classify as either displacing or nondisplacing verbs. If Levin and Rappaport Hovav are correct in categorizing them as unergative (non-displacing), these verbs would appear to cast doubt on the robustness of my claim about inanimate subjects. However, we will see in Chapter 5 that children produce unaccusatives with both animate and inanimate subjects (in almost equal proportions) but limit unergative verbs (not counting verbs of emission) to having animate subjects. Thus, even if verbs of emission are problematic for extending my generalization about subject animacy to intransitive verbs, they do not appear to tamper with children’s own knowledge of which kinds of subjects can be used with which types of intransitive verbs. 2.3.2 Formal representations of unaccusativity Further syntactic effects of the unaccusative–unergative distinction were brought to light by Burzio (1986), who adapted Perlmutter’s dichotomy to the Government-Binding syntactic framework. Burzio argued that the sole argument of unaccusative verbs was generated in the complement of V

Unaccusatives and unergatives

43

(governed by the verb) and raised to matrix subject position (SpecIP). Here we will mainly look at evidence from Italian ne-cliticization, but I will briefly mention two other pieces of evidence for the syntactic distinction between unaccusatives and unergatives (auxiliary selection and participle agreement). Unlike unergative verbs, unaccusatives in Italian allow ne-cliticization: (69)

a.

Ne arrivano molti. of-them arrive-3p many “Many of them arrive” b. * Ne telefonano molti. of-them telephone-3p many (“Many of them telephone (call)”)

The ne-cliticization diagnostic is instructive because it occurs in transitive constructions as well, but with the restriction that the clitic can only refer to direct objects. (70)

a.

Gianni ne inviterà molti. G. of-them invite-fut-3s many “John will invite many of them” b. * Gianni ne parlerà a due. G. of-them talk-fut-3s to two (“John will talk to two of them”) c. * Molti ne arriveranno/telefoneranno. many of-them arrive-fut-3p/telephone-fut-3p (“Many of them will arrive/phone”)

What is crucial about (70c) is that the clitic ne is meant to refer to the subject; this is ungrammatical whether the subject is base-generated or derived. However, ne can refer to an underived argument that is complement to the verb, as in (70a). Example (70b) shows that it is not sufficient to be a non-subject, or for the argument to be within the VP: only the direct object, and not an indirect object, can undergo ne-cliticization. So, the fact that the lone argument of an unaccusative verb can undergo ne-cliticization, as in (69a) (but not the argument of an unergative, as in (69b)) shows that the argument of an unaccusative verb is a direct object. By base-generating the argument of unaccusatives in the object position, Burzio adopts the spirit and intuition of Perlmutter’s Relational Grammar analysis, but it is formalized according to the framework of GB theory (see (71)). The single argument of an unergative verb, on the other hand, is represented as being base-generated in subject position (72).

44

Syntax of displacing and non-displacing predicates

(71)

a. [IP e [VP arriverà [NP Gianni]]

(72)

[IP Gianni [VP telefonerà]]

b. [IP Giannii [VP arriverà t i ]]

Burzio’s analysis extends neatly to several other syntactic phenomena exhibited by unaccusative verbs. Two related ones are the selection of auxiliary essere ‘be’ and gender/number agreement in past participles. By considering the single argument of arrivare ‘arrive’ to be an underlying object, Burzio brings the behavior of the lone argument of arrivare into line with subjects of passives and si reflexives, and objects of transitive verbs. Both passives and si reflexives involve both auxiliary selection of essere (as opposed to avere ‘have,’ which is selected by transitive verbs and unergatives) and they induce past participle agreement (see (73)–(74)). Objects of transitive verbs, when cliticized, also induce past participle agreement (see (75)). (73)

sono stati accusati t i. Gli studentii the student-m-p be-3p been accuse-part-m-p “The students were accused”

(74)

accusata. Mariai sii è Maria self be-3s accuse-part-f-s “Maria accused herself”

(75)

Gli studenti lai hanno accusata t i. the student-m-p her have-3p accuse-part-f-s “The students accused her”

What all of these constructions have in common is that they have an underlying internal argument which raises in the syntax. Their common behavior is then accounted for in a unified way. One of the puzzles about unaccusative constructions is why the lone argument of the verb must become a subject at all. This question is especially relevant for a language like Italian which does not require overt subjects. What Burzio argued was that the argument cannot receive Accusative case from the verb and therefore must raise to get Nominative case under government from tensed Infl. The reason it cannot get Accusative case is that a verb can assign Accusative to its complement only if it also θ-marks its subject. This became known as Burzio’s Generalization. That is, a transitive verb will assign a θ-role to its subject (and object) and therefore can assign Accusative to its internal argument. But an unaccusative verb, not having an external argument at all, cannot then assign Accusative to its internal argument. That argument must raise to get Case, and in so doing it becomes a derived subject.

Passive

45

With the advent of the Minimalist Program the light-verb projection vP was proposed as the locus of the external argument and its causative or agentive θ-role: “Accordingly, the external argument cannot be lower than [Spec, v]. If it is [Spec,v] . . . then the v–VP configuration can be taken to express the causative or agentive role of the external argument” (Chomsky, 1995, p. 315). Verbs that require an external argument, namely transitives and unergatives, will then project vP. Unaccusatives, on the other hand, will project only a VP structure, which contains the internal argument of the verb. Current iterations of the Minimalist framework hold that all predicates project vP (Chomsky, 2001). A formal difference between predicates that take an external argument and those that don’t is maintained, however. Transitive and unergative verbs project v*P, which takes an agent or experiencer external argument. As discussed above in connection with raising-to-subject (see Section 2.1.1), such a projection is φ-complete and therefore constitutes a non-defective phase: merge operations and “valuation” of features (akin to “checking” of features in earlier iterations of the theory) must take place within the phase before a higher phase can be merged. Like raising verbs, unaccusative verbs project a vP (no *) with no external argument; this vP is not φ-complete, therefore “defective,” therefore the operations within this phase are still visible even when higher phases have been merged (e.g. T and C). The defectiveness of the vP phase for unaccusative verbs is what allows NPmovement to take place from within that projection to SpecTP. In other words, because unaccusative vP is a defective phase, its internal argument is still “visible” at the higher CP phase (i.e. the main clause), and so this internal argument is accessible to movement/merge operations. It can be targeted for movement into the main clause. Most relevant for our purposes is that these defective phases (vP and an untensed TP – non-φ-complete in the terminology used in Chomsky (2001)) lack an EPP feature and therefore do not host an external argument. 2.4

Passive

The passive shares with raising, tough, and unaccusative constructions the fact that the syntactic subject is not the semantic subject – it is derived. It also shares with them a willingness to host inanimate subjects. If the “prototypical” transitive sentence has an animate subject and inanimate object (The boy threw the ball) perhaps the prototypical passive has an inanimate subject and animate by-phrase (e.g. The ball was thrown by the boy). And yet, important syntactic and pragmatic differences between the passive and the other displacing

46

Syntax of displacing and non-displacing predicates

predicate constructions lead us to predict an asymmetry in the role that inanimate subjects will play in the acquisition of the passive, as opposed to the other displacing constructions. As we will see in Chapter 5, the acquisition data support this asymmetry. 2.4.1 Structure of passive Until the GB era, passive was accounted for by a “passive rule,” a transformation that applied specifically to this construction to derive the passive form from its active counterpart. The account of passive in Syntactic Structures (Chomsky, 1957) applied an ad hoc rule of inserting the auxiliary be plus the affix -en to the main verb, and a reordering of NPs. (76)

N P 1 − Aux − V − N P 2

→

N P 2 − Aux + be + en − V − by + N P 1

The treatment of passive developed in the Aspects model (Chomsky, 1965) differed from this somewhat by employing a slightly more general rule, but it was similar in that the surface subject was analyzed as an underlying object which raised in the derivation of the structure. The main shift in analyses of the passive came with the advent of GB theory. GB brought the passive into the family of constructions involving NPmovement, which, in turn, was seen as one instantiation of the even more general rule “Move-α.” The surface subject of the passive was still seen as being derived from the underlying object position, like with unaccusatives (see Section 2.3), but the movement rule was of a more general sort, thus no longer requiring a construction-specific rule. In addition, although the active and passive voices were seen as semantically and lexically related, the passive was not derived directly from the active per se. Rather, the active and passive versions of a sentence each had distinct D-structures. According to Chomsky (1981) the main properties of the passive are (1) that the subject position at D-structure is not assigned a θ -role, and (2) the object position at D-structure is not assigned Case. The notion that the underlying object “becomes” the S-structure subject is actually purely incidental and unnecessary, as the surface subject need not be an argument of the verb at all, as in (77). (77)

It was believed [that John was late].

Under GB the account of passives like John was kissed was reduced to an instance of Move-α and was motivated by the requirements of Case Theory, Theta Theory and the Extended Projection Principle (all clauses must have a

Passive

47

subject). In (78) the surface subject raises from its D-structure object position because it is not assigned Case by the verb and therefore must raise to get Nominative case by matrix tensed Infl. It can raise to this position because that position is not theta-marked and therefore available to host a derived NP. Also, the EPP requires a subject and (78) has no subject.17 The product of this movement is (79). The inability of the verb to take a direct object is further exhibited by the ungrammaticality of (80). (78)

[e was kissed John]

(79)

Johni was kissed t i

(80)

* Johni was kissed the girl

One question left open by this account is what happens to the external θ-role. That is, assuming that as part of the verb’s lexical entry a transitive verb specifies an argument structure involving two arguments, one internal and one external, and assuming, following the Theta Criterion and Projection Principle, that each of those arguments must be represented in the structure and assigned a θ-role, the lack of an external argument in a passive like (79) is problematic. Jaeggli (1986) explained the absence of the external argument in terms of θ-role “absorption.”18 According to Jaeggli, the “passive suffix -en. . . functions as the recipient of the external θ-role of the predicate” (Jaeggli, 1986, p. 590). If the external θ -role is assigned to the -en suffix, it is no longer available to be assigned to the subject position. Furthermore, Jaeggli claimed that it is the external role and not the internal role that gets absorbed by the bound affix because only the external θ-role is unlinked to the verb, in the sense of Chomsky (1965) (also Williams (1981); Zubizarreta (1985); and see Baker et al. (1989) for a similar account of the passive within GB). A different approach to the passive is found in non-transformational frameworks like Lexical-Functional Grammar (LFG). In LFG active and passive predicates are lexically distinct (so, kick and be kicked are different lexical items), and no movement occurs in the syntax. Rather, the apparent reordering of arguments falls out from differences between the a- and f-structures of active vs. passive predicates. That is, active and passive verbs have different 17 The EPP can also be fulfilled by the insertion of an expletive when the complement of the verb

is a CP (and therefore immune to the requirement for Case), as in (77). Expletive subjects are not possible with verbs like be kissed (*It was kissed e) because the θ -role of the verb needs to be discharged to a real argument, and expletives are not arguments. 18 See Baker et al. (1989) for an extension of this proposal, where the -en affix is analyzed as an argument.

48

Syntax of displacing and non-displacing predicates

subcategorization frames. An active transitive verb specifies that it takes both a subject and object argument, as in (81). (81)

kick

A passive predicate, on the other hand, specifies an oblique and a subject argument. (82)

kick

In (82) the subject argument role is assigned to the internal argument. But since English requires subjects to precede the main verb, this argument will appear in the canonical subject position (i.e. preverbally), even though it is the internal argument of the predicate (Bresnan, 1982c, 2001). The passive construction has not been a primary focus of work in Minimalism (though see Boeckx (1998) for a formulation of the GB-style treatment of passives within the formal apparatus of the Minimalist Program). Chomsky (2001) groups passive verbs together with raising and unaccusative verbs in projecting a defective vP (not v*P, i.e. no external argument), and in this gives a much more lexicalist-flavored treatment of the construction. However, Collins’ (2005b) Minimalist account departs from previous derivational accounts of the passive in a number of respects. He employs the operation of “smuggling” (see above in our discussion of Minimalist accounts of raising-to-subject and tough-constructions) to combine what he considers to be the best aspects of the Syntactic Structures-type account of the passive (i.e. the intuition that the “deep subject” is doing the same thing in both the active and passive counterparts) and GB-type accounts (i.e. an account of passive which, unlike the one in Syntactic Structures, does not rely on a special rule). Unlike in Jaeggli’s (and Boeckx’s) treatment of passive, the -en morpheme does not have any special status in Collins’ account. In fact, he argues explicitly that the passive participle is equivalent to the past participle. Moreover, Collins argues that the external argument is generated in the same position in a passive construction as in the active counterpart: in SpecvP. However, instead of a DP the external argument is realized as a PP headed by “dummy” by (i.e. Collins analyzes passive by as not having any interpretable features). Thus, Collins draws on the intuition expressed in the Syntactic Structures analysis of the passive that the logical subject and object should be generated the same way in active and passive constructions, and the passive surface order should be derived in some way.

Passive

49

The way Collins derives the surface word order of the passive is through a sequence of movements (see (83), Collins’ (22)). The verb write raises from V to Part, where it is realized as a participle. The direct object raises from the complement of V to SpecPartP. The crucial step needed to obtain the right word order is that PartP raises to SpecVoiceP, which Collins argues is what licenses PartP. At this point in the derivation, the direct object correctly precedes both the participle and the by-phrase, but it needs to raise to SpecIP in order to end up in subject position and precede the auxiliary. It is now in a position to take this step since it has been “smuggled” out of the lower part of the clause. (83)

IP I

DPj the book Infl

VP

[+past] V

VoiceP

be Voice

PartPk Part

Parti written

Voice

vP

VP Vi



v

PP P

DP

by

John

v



Importantly, the thematic role of the surface subject of the passive is obtained the same way as that of the object of the active, namely by being generated as the sister of V (see also Baker et al. (1989)). According to this view the active and passive are derivationally related; they are not lexically distinct predicates. 2.4.2 A different displacing predicate As hinted at in the beginning of this section, we will see arguments in Chapter 5 that inanimate subjects do not facilitate the acquisition of the passive the same way they do raising verbs, tough-adjectives, or unaccusatives. The primary claim of this book is that inanimate subjects indicate a derived or displaced

50

Syntax of displacing and non-displacing predicates

subject. The subject of a passive is displaced, so why should this cue work differently for passives?19 Recall that the learning problem tackled here is not only about discovering the underlying structure of an opaque string, but also about predicate categorization: Which kinds of predicates are displacing predicates? Which kinds of predicates are not? The question of predicate categorization is important for language acquisition because the category, or subcategory, of a predicate lines up with its argument structure properties: displacing predicates don’t select an external argument; non-displacing predicates do. And knowing the argument structure of a predicate is highly informative about the kinds of lexical meanings a predicate might or might not have (Gleitman, 1990). The event or state denoted by a verb that can be either active or passive (e.g. kicking, seeing) has an agent and patient (kicker and kickee) or an experiencer and a stimulus (seer and thing seen), regardless of whether the verb is used in its active or passive form. The denotation of a verb like seem or arrive, or an adjective like easy does not have an agent or experiencer. The meanings of verbs like kick and see are in this way fundamentally different from the meanings of seem, arrive, or easy. Hearing an inanimate subject provides a key to unlocking the lexical meaning of seem, arrive, or easy in a way that it does not for kick or see. Put another way, be kicked by is in some sense a different “version” of kick, but claim is not another version of seem, eager is not a different version of easy, and laugh is not a different version of arrive. The latter pairs of predicates are lexically distinct in a way that the former (active–passive) are not. Passives are unlike the other constructions considered here in another important respect: they are not entirely opaque. That is, there is no other construction that is identical to a full passive on the surface and does not involve a displaced subject. An active voice sentence can have a locative by-phrase as in The girl threw rocks by the river, but the morphology of the verb and lack of auxiliary disambiguate the constructions. Short, or truncated passives, however, do have an alternative underlying structure. As mentioned in the introduction, short passives with auxiliary be (as opposed to get) could have an adjectival structure instead. The door was closed could mean that the door was in the state of being in a closed position (adjectival passive), or that it underwent the action of being closed by someone (verbal passive). Borer and Wexler (1987, 1992) 19 Even on an LFG account the NP assigned to the syntactic subject position is the theme, or

semantic object, and so its syntactic position is not the canonical one given its semantic role. This non-canonical syntactic role assignment is parallel to displacement in movement-based frameworks.

Passive

51

exploited this ambiguity in their account of children’s early passives: they suggested that what appear to be children’s correct productions of the passive are really adjectival in their structure. What is the difference between the verbal and adjectival structures? In the case of verbal passives the internal argument of the verb raises from the complement of V to the subject position. In the case of adjectival passives the subject NP is a theme argument of the deverbal adjective. Following Stowell (1981) predicate adjectives form a Small Clause with their theme, with the theme argument being generated in an external position within the Small Clause. Thus, the movement to matrix subject position is different in the two cases: only in the verbal passive does the subject raise from an internal position. However, since the argument of adjectival passives is a theme it is free to be inanimate, just like the raised subject of verbal passives. So subject animacy will not help disambiguate these structures. Given the important differences between the passive and the other displacing predicates, one might question whether it should be covered in the same volume. My argument for including it is that it involves the same (general) sort of syntactic displacement as the other constructions. Just like in raising, tough, and unaccusative constructions, the syntactic subject of a passive is not a semantic subject. So even though it lacks the extra puzzle of a truly opaque surface form, children still have to figure out that the syntactic subject is not the agent or experiencer of the verb. Moreover, this is a case where language acquisition data can help inform our theory of adult language structure. As we will see in Chapter 5 (Section 5.3.4), inanimate subjects are not as informative in terms of acquiring the underlying structure of the passive as they are for acquiring subject raising, tough-constructions, or unaccusatives. This might suggest that a derivational analysis of the passive is more likely to be correct than either a lexicalist analysis, which considers active and passive predicates to be lexically distinct, or Chomsky’s (2001) approach which aligns the passive more closely with subject raising and unaccusatives. The passive is relevant for a further reason: like most of the other displacing constructions covered in this book the passive has been argued to develop quite late in first language acquisition (Fraser et al., 1963; Slobin, 1966; Bever, 1970; Horgan, 1978; Maratsos et al., 1985; Borer and Wexler, 1987; Hyams et al., 2006). While there is not strict agreement over when the passive is acquired, nor how to account for certain evidence suggesting apparent early acquisition of the passive (Crain et al., 1987; Demuth, 1989; O’Brien et al., 2006), most researchers of child language agree that there is something non-canonical about the passive in terms of development. In this sense, then, it forms a class with raising-to-subject and tough-movement constructions in that much of the

52

Syntax of displacing and non-displacing predicates

acquisition literature places its acquisition at a relatively late point. Again, though, I will argue that passive is different. We will see in Chapter 5 that while there may be a true delay in the development of the passive, contrary to the conventional wisdom children do not experience particular difficulty with subject raising or tough-constructions. Apparent evidence for the late acquisition of raising and tough-constructions will be argued to stem from effects of experimental methodology. 2.5

The learning problem

The question at hand is how children determine the underlying structures of the types of sentences just discussed, and, relatedly, how children categorize predicates as displacing or non-displacing predicates. In the first three constructions discussed (raising-to-subject, tough-constructions, and unaccusatives), the subject is displaced, yet the sentence’s surface form is identical to that of a parallel construction without subject displacement. We saw that in each case, the displaced subject is free to be inanimate, while the parallel non-displaced subject generally must be animate (with a possible exception involving unergative verbs of emission; these are discussed in Section 3.1.2 in more detail). So far we have been assuming that a given predicate is fundamentally either displacing (it never selects a subject) or non-displacing (it always selects a subject). However, the situation is somewhat more complex than this, as there are some predicates that can do both: they are ambiguous between being displacing and non-displacing predicates. These are verbs like begin (Perlmutter, 1970). Other ambiguous verbs include start, fail, continue, need, and have, as in I have to go now, or There has to be some way out of here. The ambiguous verbs can occur in all of the environments that allow both raising and control verbs. (84)

a. John needs to find Susan before she leaves town. (infinitive complement) b. It needs to rain or else the crops will die. (weather-it) c. There needs to be a peaceful resolution. (expletive subject) d. What John needs is to relax a little. (pseudocleft) e. John needs the car this afternoon. (transitive)

It turns out that these predicates provide quite strong support for the feature of subject animacy as the relevant cue for distinguishing both the (nonambiguous) displacing predicates (seem, tend) from the (non-ambiguous) nondisplacing ones (claim, try), and the displacing vs. non-displacing behavior of these ambiguous predicates (begin, need). Perlmutter’s account of these verbs

The learning problem

53

is that there is a raising verb need and a control verb need, and what distinguishes them is whether the subject is volitional or not. In syntactic terms this would boil down to a difference in whether the verb takes an external argument, and external arguments typically bear a thematic role of agent or experiencer. Thus, what distinguishes the interpretations of an otherwise ambiguous string (like (84a)) is whether the subject is agentive or not. Animate subjects, then, can be volitional (agentive), so (84a) can have a control structure. But inanimate subjects cannot be agentive, so a sentence like (85) would have to have a raising structure.20 (85)

The sheets need to be dry before we can put them back on the bed.

Thinking about how these facts apply to the learning procedure for categorizing unknown verbs, let us dispel two naive approaches. First, the learning procedure for discriminating raising verbs from control verbs cannot rest solely on the subject being animate or inanimate. According to this type of procedure if the subject is animate, a verb with an infinitive complement is a control verb; if the subject is inanimate, a verb with an infinitive complement is a raising verb. However, this fails to capture the right generalization. There are raising verbs that are not ambiguous between a raising and a control reading, yet they can occur with both animate and inanimate subjects. For instance, when tend is used with an animate subject, as in (87a), the sentence does not have a control structure. So, while ambiguous verbs like begin and need may have a control interpretation when used with an animate subject, and while certain raising verbs might be coerced into having control-like properties in the presence of an animate subject (as in (86), where emphasis on the verb is needed in addition to the animate subject), others cannot be so coerced (87) (see further discussion of these phenomena in Section 3.3.1, and see Asudeh and Toivonen (2012) for links between copy raising (seems like) and both raising and control). (86) (87)

John wanted to SEEM to be smart (in order to impress Mary). a.

John tends to shop on Tuesdays. (raising)

b. * John wanted to TEND to shop on Tuesdays.

In brief, children must allow begin, need, and other ambiguous verbs to have a control interpretation when they occur with an animate subject, but they must 20 Recall a similar effect of animacy with adjectives like ready and good, although in that case the

inanimate subject merely suggests an object gap in the infinitive; it neither requires an object gap nor does it actually make the adjective a tough-adjective. See Section 2.2.2.

54

Syntax of displacing and non-displacing predicates

not make this allowance for tend (and, in the unmarked case, they should not make this allowance for seem either). The second naive learning account we must dispel has to do with using expletive subjects to discriminate these verb classes. Although raising and control verbs overlap partially in their distribution (both occurring in the frame John verbed to predicate), it is clear that they are distinguished in other contexts. As noted above, raising verbs and tough-adjectives (unaccusatives to a lesser degree) can occur with an expletive subject while control verbs cannot, and (some) control verbs can occur in certain other constructions such as pseudoclefts (What John liked was to vacation in Florida) and transitive expressions (John wants a car) that bar raising verbs (*What John seemed was to be nice/*John seemed his student). Therefore, both inanimate arguments and expletives could serve as learning cues to displacing predicates, and transitive frames and pseudoclefts could point to non-displacing predicates. But it would be naive to propose that categorization of raising vs. control could be wholly explained by a simple default rule like the one in (88). (88)

i. Assume a verb taking an infinitive complement is a control verb. ii. If that verb is heard with an expletive subject change the analysis to raising.

This rule will not work since a verb like begin/need, once it is heard with an expletive subject, will be incorrectly categorized as only a raising verb. The rule could, of course, be expanded so that if a verb is encountered with an expletive subject and in a transitive frame (as might happen with begin or need), the learner then determines that the verb is ambiguous. However, not all control verbs can occur in the transitive frame (*John hoped the winner), and so that frame cannot serve as a general cue for the control verb class. It could only serve the highly specific purpose of signaling an ambiguous verb (in conjunction with the expletive cue). As a matter of course, we want to avoid requiring the learner to assume such specific rules that have relatively little buying power. Ultimately, the learning procedure will need to rely on the cue of subject animacy probabilistically, in conjunction with other available cues (expletive subjects, transitive frames, etc.). While all of these cues undoubtedly play some role in children’s learning of these constructions, the focus of this book will be on inanimate referential subjects. As we will see in Chapter 3 the link between inanimacy and derived arguments is cross-linguistically robust. But beyond that, there are reasons for viewing inanimate subjects as being at least

The learning problem

55

as good a cue as expletive subjects. One of them was just cited; namely, the problem of ambiguous verbs. In addition, there are five further reasons for focusing on inanimate subjects rather than expletives. Let us look at them each in turn. (a) One reason for focusing on inanimate subjects is that expletives do not exist in every language, while inanimate NPs do. If we want our learning strategy to be plausible, it should be available to children learning any language, whether or not the language has expletives. For instance, null-subject languages lack expletives (e.g. Italian, Spanish, Mandarin, Japanese). In these languages, constructions which in English employ expletives simply have an obligatory null subject. (89)

e Sembra che Gianni sia stanco ∅ seems that Gianni be-subj tired-m “It seems that John is tired.”

The difference between languages with expletives and languages without expletives may not be so grave after all, if we consider the fact that expletives are often homophonous with a non-expletive (i.e. referential) pronominal form, as in English for example. Thus, the learner must determine the difference between expletive it and referential it, and between expletive there and locative there, just as they must distinguish between e and pro. Nevertheless, if it is easier to distinguish between two overt forms than between two non-overt forms, children acquiring non-expletive languages would be at a disadvantage for acquiring displacing predicates if their learning strategy depended on expletives. I know of no comparative studies of the acquisition of displacing predicates between expletive and non-expletive languages, and in the absence of data bearing on this I will assume that a strategy relying on inanimate referential arguments is more universally applicable, and therefore more plausible, than one that relies solely on expletives. (b) A second reason for focusing on inanimate subjects is that expletives are exceedingly rare in speech to children. As we will see in Chapter 6, it is quite likely that learners will encounter raising verbs with referential subjects before they encounter them with expletive subjects.21 If this is true, then having a cue like inanimate referential subjects would be highly beneficial.

21 In that chapter we will also see an interesting asymmetry between raising verbs and tough-

adjectives, but the overall prevalence of referential subjects remains true for both types of predicates.

56

Syntax of displacing and non-displacing predicates

(c) Third, psycholinguistic research has demonstrated that animacy can have a powerful influence on how speakers interpret sentences – inanimate NPs tend to be interpreted as patients in locally ambiguous constructions (discussed in Chapter 4). Furthermore, the concept of animacy is something available to children from quite early in development, long before language production or even comprehension begins (Chapter 5). Expletives, on the other hand, are acquired significantly later (Kirby and Becker, 2007). (d) Fourth, an inanimate referential subject is roughly as reliable a cue to the predicate being a displacing predicate as an expletive subject. Verbs that take a sentential complement (either infinitive or tensed) generally select either an agent or experiencer subject (external argument) or nothing at all. I do not know of sentential complement-taking verbs that select a theme external argument. Verbs of communication can take an inanimate source subject, as in The manual says that we have to do step A first, but verbs denoting mental states (believe, know, etc.) cannot have a source subject. Thus, if agents and experiencers are limited to animate NPs, then an inanimate NP could not be an agent or experiencer, and hence in a biclausal structure it is very likely to be derived. (90)

a.

John gorps that it is sunny. (John = experiencer; gorp = non-displacing predicate)

b. # The rock gorps that it is sunny. c. Itexpl gorps that it is sunny. (gorp = displacing predicate) (91)

a. John gorps to be in the yard. (John = experiencer or theme; gorp = displacing or non-displacing) b. The rock gorps to be in the yard. (rock = theme; gorp = displacing) c. It gorps to be sunny outside. (gorp = displacing)

Expletives may in fact be a more informative cue for the narrow class of tough-adjectives (and raising adjectives, e.g. likely) than for raising verbs (see end of Section 2.2.2). There do not seem to be any truly ambiguous adjectives that can function both as tough-adjectives and control adjectives, and both tough and raising adjectives occur with expletives while no other adjectives do. There are some adjectives that admit both a subject-gap and object-gap reading, like ready, but these are not truly tough-adjectives as they do not occur with an expletive subject.22 (92)

* It/there is ready to be a riot.

22 I find the sentence It’s ready to rain natural, but it feels colloquial and somehow not quite

standard, like a use of ready in place of about, which is a raising predicate.

The learning problem

57

While expletives single out true tough-adjectives, as we saw in Section 2.2.2 when the subject is referential the subject- vs. object-gap reading is influenced (at least to some degree) by animacy. For example, consider again the sentences in (93). (93)

a. The girli is ready [PROi to eat e]. b. The sandwichi is ready [PROarb to eat t i ].

In general, inanimate subjects are roughly as informative as expletives for determining that the subject of a biclausal construction is displaced, but what about monoclausal constructions? There are two things to consider, one which points to expletives being a better cue, and another which points to expletives being a worse cue than inanimate subjects. In English and many other languages certain transitive verbs permit an inanimate subject as the external causer argument (e.g. The poll results influenced voter turnout). As we will see in Section 3.3 not all languages permit inanimate subjects of transitives, and some do not permit the subject to be less animate (lower on the Animacy Hierarchy) than the object. However, in languages that do permit inanimate subjects of transitives, like English, we don’t want children to mistakenly construe these NPs as displaced. If inanimate subjects are a potential red herring in transitive constructions, then in the general grammar expletives might be a more decisive cue to displacing predicates than inanimate subjects. (No transitive verbs occur with expletives, but some – in some languages – can occur with inanimate subjects.) On the other hand, there are two other types of monoclausal sentences that admit inanimate subjects, where the subject is in fact derived: unaccusatives and certain types of psychological verb (psych-verb) constructions. (94)

The letter arrived/The rock fell.

(95)

The noise bothered me.

In English certain unaccusative verbs can occur with an expletive there subject while unergatives cannot. (96)

a.

There arrived three letters in the mail yesterday.

b. * There danced three people at the party yesterday.

But occurrence with an expletive subject seems to be restricted to a subset of unaccusatives (arrive, begin, come, go) and is not a general property of this

58

Syntax of displacing and non-displacing predicates

class (*There fell a book off the shelf ). Moreover, psych-verb constructions do not allow expletive subjects at all. (97)

* There bothered me the noise.

For these kinds of constructions it would seem that inanimate NPs are a better cue than expletives. (e) Finally, it is interesting to observe that expletives might not be an entirely unproblematic cue. There are predicates that can occur with expletive subjects but which do not, in fact, allow a referential NP to raise into that subject position. (These predicates do allow a CP to raise into subject position.)23 (98)

a.

It is likely that John left.

b.

John is likely to have left.

(99)

a. It is probable that John left. b. * John is probable to have left.

(100)

a.

It seems that John left.

b.

John seems to have left.

(101)

a. It sucks that John left. b. * John sucks to have left.

These predicates are potentially problematic because they mean that encountering an expletive subject is not definitive evidence that a lower NP can raise into subject position. Thus, the learner should not generalize from seem and likely to suck and probable. At this point let me bring up a complicating factor relating to animacy: a very limited number of control verbs permit inanimate subjects. I am aware of only two verbs in English that allow this, serve and help (Rudanko, 1989): (102)

a. This pamphlet serves to explain the rules of the organization. b. * It serves to rain. c. * There serves to be a pamphlet.

(103)

a.

Oil helps to make the engine run smoothly.

b. * It helps to rain. c. * There helped to be some oil in the engine. 23 Chomsky and Lasnik (1977) account for the asymmetry between likely and probable by saying

that likely selects a null complementizer while probable does not (p. 445).

The learning problem

59

(102b,c) and (103b,c) show that serve and help do not behave like raising verbs. On the other hand, these verbs may have some kind of hybrid status. Note that help does allow a pleonastic it subject with a tensed complement: It helped that the engine was well lubricated (serve does not have this capacity). But the status of this construction is not quite clear in terms of verb category diagnostics. While some raising verbs can occur in this context (It seems that John is nice), not all of them can (*It tends that John is nice). Further support for a hybrid account of these verbs is that help and serve appear to allow a Quantifier Lowering-type of interpretation (in fact, my judgment is that the QL reading is preferred), which suggests that they may in fact both have some properties of raising verbs.24 (104)

Justifications don’t serve to help your case. = (?)There are justifications that don’t serve to help your case = No justifications serve to help your case

(105)

A ticket doesn’t help to prevent one from speeding. = (?)There is a ticket that does not help prevent one from speeding = No ticket helps to prevent one from speeding

The fact that there are verbs that display some properties of raising verbs but other properties of control verbs, thus having a sort of hybrid status, should not derail the learning process. Instead, these counter-examples to the general trend are a reminder that the learning process needs to be probabilistic and not deterministic: a single counter-example should not throw off the categorization of verbs that adhere to the main trend. The probabilistic nature of the learning strategy will be the focus of Chapter 6. To summarize, we have seen that raising-to-subject, tough-constructions, unaccusatives, and passives all involve a structure in which no external argument is projected. Instead, an NP derived either from an internal argument position (passives, unaccusatives) or from an embedded clause (raising, toughconstructions) occupies the matrix subject position. Due to the lack of a thematic selectional relationship between the subject and the matrix verb, the subject is free to have any semantic features as long as they are compatible with the predicate that selected it, and thus in principle can be inanimate. Superficially similar to these constructions (except the passive) is a set of constructions in which the subject is the external argument of the matrix predicate and is thus subject to the semantic selectional requirements of that predicate. In these cases, the subject generally must be animate. I have 24 I’m grateful to Kyle Johnson for this suggestion and for discussion about this effect.

60

Syntax of displacing and non-displacing predicates

proposed that animacy can be used as a learning cue by children to distinguish these superficially similar strings according to their correct underlying structures. Specifically, hearing a (biclausal or intransitive) sentence with an inanimate subject provides a clue that the subject is likely to be displaced. In the next chapter we will look at how animacy relates to argument structure, and how inanimacy is a frequently attested property of derived subjects across a variety of languages.

3 Argument hierarchies

In the preceding chapter I described the syntactic structures of four constructions whose predicates fail to select an external argument, or “semantic” subject. The syntactic subject these predicates do occur next to is therefore not thematically related to the main predicate in the usual way. In the language of derivational approaches to syntax, these subjects are derived, or displaced. In some of these constructions the displaced subject ends up in a position non-adjacent to the predicate that selected it (tough-movement and raising-to-subject); in the other constructions the displaced subject is still both string-adjacent to, and clause-mates with, its selecting predicate, but its grammatical role has changed, as it moves from an object position to a subject position (unaccusatives and passive). The first three of these constructions can be contrasted with a construction that looks quite similar on the surface, but which does not involve the kind of argument displacement described above. In subject control constructions (irrespective of whether the predicate is a verb or an adjective), the syntactic subject is also the logical, or semantic subject of the main predicate. In unergative verb constructions the surface subject is an underlying subject, not an underlying object. These pairs of constructions look the same on the surface but have different syntactic structures, precisely the situation described by Chomsky (1969) (see quote in Chapter 1). Because of the surface similarity in each pair, determining the correct argument structure representation before knowing the lexical meaning of the main predicate is a non-trivial task. That is, prior to knowing the lexical meaning of easy, eager, seem, or claim, the string is opaque to the underlying structure.1 In order to arrive at the correct adult grammar, children have to tease these constructions apart.

1 As will be addressed more fully in Chapter 5, the well-supported claim of the Syntactic Boot-

strapping hypothesis is that children learn the meanings of verbs and other predicates, especially

61

62

Argument hierarchies

The thesis of this book is that children accomplish this in large part by noticing that in each of the opaque strings, the structure that involves a derived subject can host an inanimate NP in the derived subject position. There are two central motivations for using inanimacy as the instructive cue. One comes from the early availability of the (non-linguistic) concept of animacy, as will be discussed in Chapter 5. There I present evidence showing that conceptual animacy (i.e. the knowledge of whether an entity is alive and sentient or not) is available in advance of complex syntactic representation. The other motivation comes from the cross-linguistically robust link between animacy, agency, and underlying subjecthood, and the fact that, universally, displaced subjects can lack these typical semantic properties of subjects. Thus, I argue that subjects that violate the typical pattern (e.g. by not being animate or agentive) can provide a cue to a non-canonical underlying structure. It is important to keep in mind that this is a probabilistic strategy, not a deterministic one. That is, animate NPs are not always agents and certainly not always subjects, and themes are not always inanimate. And while derived subjects need not be inanimate, non-derived subjects are almost exclusively animate. Thus, a learner needs the following ingredients to make a guess that an inanimate subject is a displaced subject: 1. animate NPs are distinguished from inanimate NPs; 2. subjects are generally animate, or more animate than other arguments in the clause/sentence; 3. NPs bearing the agent or experiencer thematic role are animate and canonically mapped to subject position; 4. displacing predicates do not select a subject → they place no semantic restrictions on the subject → the subject can be inanimate; 5. displaced subjects generally have the syntactic, but not the semantic, properties of non-displaced (canonical) subjects. ⇒ an inanimate subject is likely to be displaced The inference about the underlying structure of a sentence depends upon knowledge of thematic roles of NPs, not physical or biological properties. But this inference is based upon the ability to determine the likelihood that

abstract ones, via the syntactic structures they occur in, not the other way around. Thus, these lexical meanings are argued not to be available to learners in advance of parsing raising and control (etc.) sentences correctly.

The Animacy Hierarchy

63

a given NP will have a particular thematic role on the basis of its degree of animacy/sentience. In order for this line of reasoning to work for language learners, we must establish that the relationship between animacy and thematic role just outlined in fact holds true of language in general. Likewise, in order to make the claim that children can use knowledge of how thematic roles map to particular syntactic positions in acquiring non-canonical sentence structures, we must show that this mapping is robustly attested in human language. It is these topics that we will address in the present chapter. Hierarchies abound in formal representations of human language. Not only is language structure itself represented as hierarchical (i.e. in terms of syntactic structures, syllable structures, and word structures), but also components of grammar that relate to grammatical or semantic features are organized into hierarchies of prominence. The Animacy and Thematic Hierarchies, discussed here, are joined in the literature by the Accessibility Hierarchy/Relational Hierarchy (Subject > Direct Object > Indirect Object > Oblique > Genitive > Object of Comp; Keenan and Comrie, 1972; Gary and Keenan, 1977), Transitivity Hierarchy (Hopper and Thompson, 1980, 1982), Ergativity Hierarchy (Silverstein, 1976), Prosodic Hierarchy (Selkirk, 1980), and the Subject Promotion Hierarchy (Keenan, 1976), among others (see Croft (2003) for a good overview). Here we will focus mainly on two of these hierarchies: the Animacy Hierarchy since, as we will see in Chapter 5, animacy is a property conceptually available to very young children, and the Thematic Hierarchy since it gives us a way of linking animacy to both semantic/conceptual and grammatical/syntactic structure. We will also discuss the Subject Promotion Hierarchy as it relates to properties of derived subjects. 3.1

The Animacy Hierarchy

Broadly speaking, animacy relates to whether an entity is alive or not. Although the quality of animacy itself is a natural feature, and therefore not intrinsically linguistic (unlike, for example, definiteness, or Nominative), it triggers syntactic or morphosyntactic reflexes in a great many languages. Animacy as it pertains to linguistic effects is not so much a binary distinction (alive–not alive) as a continuum, ranging from human to non-human animals to inanimate objects. Moreover, within the human end of the spectrum many languages distinguish pronominal from non-pronominal NPs in terms of animacy, and we find various distinctions within both pronominal and non-pronominal NPs (Silverstein, 1976; Dixon, 1979). For example, some languages treat 1st

64

Argument hierarchies

and 2nd person as more animate than 3rd person, and some languages treat proper names as more animate than other nouns. (1)

Animacy Hierarchy Human (1/2p > 3p; pronoun/proper N/kinship > other Human N) > Nonhuman Animal (“higher” animal > “lower” animal) > Inanimate

Both Comrie (1989) and Croft (2003) note that these finer-grained human distinctions relate to dimensions of person and referentiality, and seem to interact with the dimension of (biological) animacy. For example, 1st person and 2nd person referents are necessarily human, while 3rd person referents need not be; proper names generally refer to humans while common nouns need not. Some of these finer-grained distinctions will be exemplified below. 3.1.1 Linguistic effects of animacy: morphosyntax and argument structure Like other conceptually-based properties that yield linguistic effects, such as gender or number marking, the linguistic effects of animacy do not always fall directly in line with the actual physical underpinnings of this feature, which in this case are biological properties of animacy. Most blatantly, many languages distinguish human-denoting NPs as more animate than non-human-denoting NPs (such as horses), even though both people and horses are equally alive (the Australian language Arabana marks this distinction; Blake (1977)). In some languages, 1st and 2nd person pronouns are considered more animate than 3rd person pronouns, even if 3rd person pronouns can be used for humans as well. Moreover, it is important to bear in mind that sometimes animacy effects overlap with other features such as definiteness. For example, the Hindi postposition ko can be used with human NPs regardless of whether they are definite, but with inanimates only if they are definite. Nevertheless, it is clear that languages generally make distinctions between “more animate” and “less animate” NPs in ways that affect morphosyntactic marking or argument structure. In this section I list some of these patterns for the purpose of giving an overview of some of the ways animacy effects surface morphosyntactically in human language. Anticipating the relation to language learners taken up in Chapters 5 and 7, there is no claim of an a priori expectation about the precise link between animacy and morphosyntax. Given the wide cross-linguistic variation in exactly how animacy is marked morphosyntactically there could be no such expectation. Rather, the view is that learners should entertain the possibility that if different NPs are distinguished morphologically in their language, degrees of animacy may be the relevant feature underlying these distinctions. But in terms of how animacy relates to argument structure we find a high

The Animacy Hierarchy

65

degree of uniformity across languages. Establishing the cross-linguistically robust association between animacy and subjecthood will set the stage for making the stronger claim (in Section 3.5, fleshed out in Chapter 7) for an a priori expectation that animate NPs are preferentially linked to canonical subjecthood. Cross-linguistically there is a tendency in transitive structures for the subject to be higher in animacy than the object, and for certain grammatical cases, such as Nominative, to be associated with animate NPs. In many languages with an Ergative/Absolutive case system animate NPs may bear Ergative case (e.g. Jacaltec, Basque, among others) while other languages reserve Ergative marking for inanimate NPs (e.g. Ritharngu (Australian)). Where the canonical pattern is broken (i.e. the subject is less animate than the object, or the object is animate) some languages employ differential marking on the subject or object. For instance, many languages employ differential object marking according to animacy (Aissen, 2003). Spanish marks animate direct objects with the preposition a if the object is definite (el empleado ‘the clerk’) or if it is a specific indefinite (non-specific indefinite animate objects are not marked with a), as shown in (2); from Comrie (1989). (2)

El director busca el carro/ al empleado/ a un empleado/ un empleado the director seeks the car/ to-the clerk/ to a clerk/ a clerk “The director is looking for the car/the clerk/a (specif.) clerk/a (any) clerk”

Many Australian languages employ a special Accusative case when the transitive object is higher in animacy than the subject (Blake, 1977): Dyirbal employs this special marking when the direct object is 1st or 2nd person (and the subject is anything else); Gumbainggir uses special Accusative marking for objects that are personal pronouns, proper nouns, or kinship terms; Arabana applies the special Accusative to all human NP objects; and Thargari uses special Accusative on any animate object. Russian also has a special Accusative case that is used for certain masculine nouns if they are animate, but not otherwise, and Russian marks all plural animate (but not inanimate) objects with a special Accusative case (Comrie, 1989). The reverse situation, where a subject has low animacy, can also receive special marking. Dyirbal marks 3rd person pronominal and non-pronominal subjects according to an Ergative/Absolutive system, but 1st and 2nd person pronouns according to a Nominative/Accusative case system (Dixon, 1979). As noted previously, non-pronominal NPs are considered lower on the animacy hierarchy than pronominal NPs, and 3rd person pronouns are considered lower in animacy than 1st or 2nd person pronouns. Dahl and Fraurud (1996) note that

66

Argument hierarchies

in split-ergative languages, “accusative marking tends to go with animate NPs and ergative marking with inanimate ones” (p. 48). Another language that employs different case systems according to the level of animacy of an NP is Ritharngu (Australian). Nominative–Accusative is used for pronouns, a mixed system of Nom/Acc and Erg/Abs for (non-pronominal) NPs denoting humans and higher animals, and an Ergative/Absolutive system for lower animals and inanimate objects (Heath, 1976). Rather than using distinctive case marking, in some languages deviations from the canonical pattern of subjects being more animate than objects yield differences in word order. Navajo requires the subject to be more animate than the object, so that if the agent or causer of an event is more animate than the theme, active voice is used (with prefix yi- on the verb) (3a). If the theme is more animate than the agent or causer, passive must be used (with verbal prefix bi-) (3b). Where both agent and theme are equal in their animacy status, either active or passive is allowed (4a–b) (Hale, 1972; Comrie, 1989). (3)

a. At’ééd nímasi yi-dííłid girl potato burnt “The girl burnt the potato” b. At’ééd nímasi bi-dííłid girl potato burnt “The girl was burnt by the potato”

(4)

a. Diné ’ashkii y-oo’í˛ man boy see “The man sees the boy” b. ’Ashkii diné b-oo’í˛ boy man see “The boy was seen by the man”

Chamorro, like Navajo, requires the subject to be higher on the animacy hierarchy than the object (Chung, 1983). Jacaltec not only requires the subject to be higher in animacy than the object, but inanimate subjects are banned in transitive sentences (Craig, 1976, 1977). (For more on Jacaltec, please see Section 3.3 below.) Japanese (Kuno, 1972), Blackfoot (Ritter and Rosen, 2010), and Tlapanec (Oto-Manguean; Corbett (1991)) exhibit similar bans on inanimate subjects of transitive verbs, and Russian has a strong dispreference for inanimate subjects of transitives (Comrie, 1989). The following examples from Japanese are taken from Kuno (1973, p. 31). Notice that in (6) the inanimate causer is expressed as the object of a preposition (de ‘by/in’).

The Animacy Hierarchy (5)

a. * Zidoosya-ziko ga teen-ager o korosita traffic accident Nom teenager Acc killed (“A traffic accident killed a teenager”) b. * Taihuu ga mado o kowasita typhoon Nom window Acc broke (“A typhoon broke glass windows”)

(6)

a. Zidoosya-ziko de, teen-ager ga sinda traffic accident in teenager Nom died “A teenager died in a traffic accident” b. Taihuu de mado ga kowareta typhoon by window Nom broke “Because of a typhoon, windows broke”

67

Dahl and Fraurud (1996) note, after Foster (1979), that Ozark English has a strong dispreference for inanimate subjects of transitive sentences, akin to that observed in the languages just cited, and see Ransom (1977) for a discussion of similar restrictions in standard American English. The same pattern is found in Southern Tiwa, though the restriction involves 1st and 2nd person versus 3rd person, rather than strictly animate versus inanimate (Allen and Frantz, 1983). That is, the subject must be higher in “person” than the object. Comrie (1989) considers person effects to be a type of animacy effect, as noted above, but see DeLancey (1981) for arguments that person effects are a distinct phenomenon. Some languages exhibit differential agreement marking according to the animacy of the subject. Ancient Greek, Persian, and Georgian are among the languages in which the verb agrees with a plural NP only if it is animate; if the plural NP is inanimate the verb bears the singular form. Other languages, such as Guaraní and Tiwi, have differential number marking according to the animacy of the noun phrase (Croft, 2003, p. 134). While English displays none of these particular animacy effects linked to differential case or agreement marking (not having much of a case or agreement system to begin with), there are some (perhaps mundane) animacy restrictions to be found. Certain verbs, for example, require an animate or sentient subject (think, believe, know) except in metaphorical contexts, while other verbs do not place such a restriction (fall). A number of studies have linked animacy to NP ordering in double object vs. prepositional dative constructions (Ransom, 1977; Bresnan et al., 2007; Bresnan and Hay, 2008). Ransom extends her analysis to passives, arguing that subjects of passives that are higher in animacy and definiteness will be more acceptable than subjects that are lower in animacy and definiteness. We will return to the apparent preference for animate subjects of passives in Chapter 5, Section 5.3.4.

68

Argument hierarchies

These cross-linguistically widespread effects are relevant for learners of English (or any language), for two reasons. First, if universal patterns in language are good candidates for innate linguistic expectations, then animacy marking (in some way or other) is a very good candidate. Not only is the marking of animacy extremely widespread, either through morphosyntactic means or in terms of restrictions on word order/argument structure, but the pattern of linking between degree of animacy and likelihood of being a subject versus object is uniform across languages. (7)

⎧ ⎫ ⎨ more animate ⎬ more likely subject ⎩ ⎭ less likely object

←→

⎧ ⎨

⎫ less animate ⎬ less likely subject ⎩ ⎭ more likely object

While it would not make sense to propose that children are born expecting animate NPs to bear a particular case or trigger a particular agreement pattern, children would be at a distinct advantage if they expect animate NPs to line up with the subject and/or agent role and/or topic, and inanimate NPs to line up with the object and/or patient/theme role and/or non-topic, since this is clearly the unmarked pattern (see discussion in Section 3.4.1 for possible counter-examples to the unmarked pattern). This is the essence of the Semantic Bootstrapping Hypothesis (Pinker, 1984), discussed in detail in Chapter 5. With such an expectation in place, a violation of this pattern (i.e. an inanimate subject) could indicate a more marked structure. Secondly, even in languages that do not display overt animacy effects such as those discussed above, there are other ways in which animacy effects can surface. In Chapter 4 we will discuss psycholinguistic effects of animacy, but here it will suffice to note that there may be statistical animacy effects. Dahl and Fraurud (1996) report that in a corpus of over 3,100 written utterances of Swedish, only 2.6% contained a non-human-denoting subject and a humandenoting direct object. The most common pattern involved a human-denoting subject and a non-human-denoting direct object (47.7%), and in the remaining utterances either both subject and object denoted non-human referents (40.8%) or both denoted persons (8.9%). Dahl (2000) reports that in spoken Swedish, 93% of subjects are animate and 89% of objects are inanimate. Similarly, Øvrelid (2004) found that Norwegian transitive sentences heavily favor the subject having a greater or equal degree of animacy as compared to the object: this was the case in 97% of 1,000 sentences analyzed. Searching a dataset of childdirected English speech (parental utterances), Scott and Fisher (2009) found that 99% of transitive sentences contained an animate subject.

The Animacy Hierarchy

69

It is important that the preference for subjects to be animate and for animate NPs to be subjects is not a hard and fast rule, but a general tendency. Clearly, it is not a hard and fast rule across languages. But the very fact that it is a strong tendency allows for the violation of this tendency to convey important information. In many languages, as we’ve seen, the violation is marked in a special way (e.g. Accusative on animate objects, or Ergative on inanimate subjects) – these languages use this special marking to signal a break from the typical pattern. 3.1.2 Animacy, agency, degree of control, and teleological capability Ultimately we want to know how children categorize NPs as animate or inanimate, and knowing how adult languages make this distinction is helpful in framing the target state that children are moving toward. And yet, the precise nature of the linguistic animate–inanimate distinction appears to involve a number of different variables and features. As we have seen already, clearly it is not strictly a matter of biological category; it cannot be reduced to a feature like [±living]. Not only are some living NPs considered to be “more animate” than other living NPs (e.g. humans vs. animals, or pronominal vs. nonpronominal NPs), but in some languages inanimate NPs capable of physical force, such as wind, can be considered “more animate” than other inanimate NPs. For example, in the following Navajo example (taken from Comrie (1989); see also Hale (1972)), lightning functions as having equivalent animacy with horse, and thus either NP can be the subject in (8). Old age, on the other hand, is unequivocally less animate than horse, and therefore requires the passive marker bi in (9). (8)

a. ’Ii’ni’ ł´˛ı˛í’ yi-yiisxí˛ lightning horse killed “Lightning killed the horse” b. Ł´˛ı´˛ı’ ’ii’ni’ bi-isx´˛ı horse lightning killed “The horse was killed by lightning”

(9)

Shi-ł´˛ı´˛ı’ s´a˛ bi-isx´˛ı my horse old-age killed “My horse was killed by old age”

In the other direction (namely, of living entities being “not animate enough” to be subjects) we have examples such as these, from English: (10)

a. # The plant wonders whether it will rain. b. # The bacterium believes itself to be beneficial.

70

Argument hierarchies

If the animate–inanimate distinction does not hinge on being alive, perhaps the appropriate criterion is agency or sentience: the ability to perceive things, to have mental attitudes, and to act volitionally upon others. In contrast to animacy per se, agency is a quality that relates an entity directly to an action it performs. As Yamamoto (1999, p. 149) notes, “whereas animacy is concerned with the intrinsic features and ontological status of the entities themselves, the notion of agency characterizes the entities according to what they are ‘doing’.” Thus, apparently inanimate “agents” like natural forces (which are sometimes labeled with the terms “cause” or “force”; see discussion in Section 3.2.1), are accounted for on this split (The gust of wind felled the tree). But while inanimate forces can be agents in certain transitive contexts, they are not tolerated with control predicates, such as try or eager. Rather, control predicates require a subject with some degree of sentience and/or volition. (11)

a. # The plant is eager to grow. b.

(12)

The dog is eager to go outside.

a. # The bacterium tried to be infectious. b.

The otter tried to swim after the fish.

Plants and bacteria are alive (they metabolize, reproduce, and die) and can surely be construed as agents under certain circumstances (The tall plant shaded the strawberry patch; Harmful bacteria overtook her healthy gut flora), but they are incompatible with control predicates like try and eager, which require a subject NP that has more agentive properties (i.e. sentience, volition) than bacteria or plants have. Clearly, though, the subject of these predicates need not denote a human: dogs and otters evidently have enough of whatever agentive/sentient quality is needed to be good subjects of these predicates. Naturally, all of these NPs can be displaced subjects. (13)

a. The plant is easy to grow. b. The dog is easy to train. c. The bacterium seemed to be infectious. d. The otter seemed to swim after the fish.

A problem with using the term agency to define the criterion for subjects of control predicates, however, is that traditional tests for agency, such as modification by intentionally, do not make the right cut. These tests of agency often work only on eventive verbs (cf. the observation from Gruber (1965); Lyons (1968), noted below, that only eventive (nonstative) verbs can select an

The Animacy Hierarchy

71

agent subject). Thus, while raising verbs and tough-adjectives generally disallow such adverbs (14), stative control verbs like want and control adjectives (be afraid) also disallow modification by intentionally (15a,b). In contrast, eventive control verbs like try allow it (15c).2 (14)

a. * John intentionally used to commute to New York. b. * John is intentionally easy to please.

(15)

a. ?? John intentionally wants to be the winner. b. * John is intentionally afraid to walk alone at night. c. John intentionally tried to lose the race.

Further underscoring the fact that agency in this sense is not at stake, inanimate subjects of tough-adjectives and raising verbs cannot be modified by accidentally, though unaccusative verbs can. (16)

a. * The book is accidentally easy to read. b. * The book accidentally seems to be heavy. c. The book fell accidentally.

Adverb modification as a test of agency requires the intersection of agency and volition, since predicates of inanimate subjects cannot be modified by intentionally, even if the subject is apparently agentive with respect to the event (#The gust of wind intentionally felled the tree; cf. The woodsman intentionally felled the tree). The conflicting results of such diagnostic tests get at one of the inherent perils in attempts to pin down exactly the criteria for agency as it relates to constraints on predication: we are dealing not only with an intersection of properties that overlap substantially but need not be completely coextensive (agency, volition, animacy, sentience), but also with an intersection between conceptual categories and linguistic constraints. It is the agentive or volitional attribute of animates that appears to be the relevant criterion for determining which NPs can be subjects of control predicates, which makes it seem like agency is indeed the core relevant property. However, to go back to Yamamoto’s distinction between animacy and agency, it is the ontological status of an entity that determines whether it is agentive in the right sense, such 2 The raising verb seem is interesting in this respect, since in some cases it allows a kind of

“intentional” or purposive construal, where it acts more like a control verb (see discussion in Section 3.3.1). Under this kind of construal, I find modification by intentionally somewhat acceptable (?John intentionally seemed to be hurt (so he could get out of playing the rest of the match)).

72

Argument hierarchies

that it is permitted with control predicates. Therefore, it seems to me that an entity’s ontological status (animate) is more important than “what it is doing” (agentive) in these cases. Another candidate term for the sense of animacy we are after might be “degree of control.” Certainly, in different situations an animate entity might have more or less control over events. Tsova-Tush marks 1st and 2nd person subjects differently according to the degree of control they have over an event. For example, in I fell the subject would be marked with Ergative case if I fell deliberately, but with Absolutive case if the falling was accidental (Holisky, 1987).3 Thus, degree of control is a property that animates typically have, though it is not an inherent property of individuals but rather a property of particular situations. In contrast, inanimate entities generally have little or no control over events. Even in cases where inanimate NPs can be causers (e.g. The wind/branch broke the window), such entities have no control over the unfolding of the event, and, clearly, they remain nonvolitional (see discussion in Comrie (1989) of the distinction between control and animacy). Although animate entities can have control over their actions and inanimate entities can’t, it is not clear that degree of control appropriately distinguishes possible from impossible subjects of control predicates. I can say John claims to be the fastest one but not #The red car claims to be the fastest one, and people but not cars are in control of their claims, but here I think the issue is not so much having control over one’s claims but rather being able to make a claim at all. If we want to say that the ability to make a claim about something stems from “degree of control,” then it is not really clear how degree of control is different from animacy in any meaningful way. Folli and Harley (2008) argue that the relevant feature that underlies animacy effects is “teleological capability,” which they define as the internal capacity of an entity to perform the act denoted by the predicate, rather than animacy per se. One of their examples involves inanimate agentive subjects of verbs of sound emission. These verbs, such as whistle, squeak, or ring, are unergative and thus take an external rather than an internal argument. But unlike most unergative verbs (e.g. laugh), these verbs either permit or require the subject to be inanimate (recall the mention of these verbs in Section 2.3.1). (17)

a. John whistled. b. The train whistled.

3 See also Anderson (1976), who does not consider the difference in subject marking to indicate

a true ergative pattern.

The Animacy Hierarchy (18)

a. # John squeaked. (But cf. The mouse squeaked) b.

(19)

73

The door squeaked.

a. # John rang. (N.B. Not meaning ‘to telephone,’ but to emit a ringing sound.) b.

The bell rang.

Thus, these verbs are atypical in that they take external arguments that are inanimate. However, Folli and Harley argue that animacy is not exactly the right feature for describing how subjects of these verbs are restricted. Support for teleological capability as the relevant feature comes from a comparison between the equally inanimate NPs train and bullet, where (20b) is ungrammatical without the locative PP. (20)

a. The train whistled. b. The bullet whistled *(into the room).

It is in the capacity of trains to emit the whistling sound themselves, while bullets only create a whistling sound by their motion. Therefore, bullets lack the capability that trains have to emit a whistling sound of their own accord, and a subject like bullet therefore requires the directed motion PP into the room. This is compatible with Levin and Rappaport Hovav’s (1995) designation of these verbs as involving internal causation. Thus, on this view we could say that external arguments must be capable of instigating or bringing about some event by their own capacity, and this would subsume the set of animate entities. I think there are two ways to reconcile these facts with the general claim I am making. One way is to recognize that semantic selection between predicates and arguments is governed not only by broad features such as animacy, but also by narrow requirements of specific predicates: the predicate squeak requires something that can squeak just as eat requires something/someone that can eat, and fall requires something/someone that can fall (though this is an internal argument) (#The idea fell). So we should not be surprised to find the kind of selectional restriction exhibited here by these verbs of emission. The second way to reconcile these patterns with the larger picture is to question whether verbs of emission are really uniformly unergative. In fact, verbs of emission display quite interesting and varied patterns. According to Johansson and Brittain (2012) verbs of emission in Northern East Cree (Algonquian) exhibit properties of both unaccusative and unergative verbs. They are like unaccusatives (and unlike unergatives) in that they disallow an agent-oriented

74

Argument hierarchies

adverb (e.g. intentionally), and they disallow a purpose clause, as illustrated in (21) (purpose clauses are associated with the presence of an implicit agentive argument; Manzini (1983)). For comparison, (22a) shows an unaccusative with purpose clause as ungrammatical, and (22b) shows an unergative with a purpose clause as grammatical. (21)

(22)

* Shâshwâwâ-piyiu â-wî-pihkhtâ-t chîkîyiu jingle-dyn:3 cj.pvb-desid-win-3 prize (“It is making a jingling sound to win a prize”) a. * Chwân chî-kwâtipi-piyiw châchî-tûtûwât awâsa John pst-roll.over-dyn:3 comp-make.x.do.y children:obv châchî-pâhpiyichî comp-laugh:obv (“John rolled over (unintentionally) to make the kids laugh”) b. Chwân chî-pâ-piyiw châchî-wâpâhtiyiwât utût John pst-come-dyn:3 comp-show.tr.obv 3:canoe “John came (paddling) over to show her his canoe.”

On the other hand, verbs of emission in Cree pattern like other unergatives (and unlike unaccusatives) in that they have no causative alternation, and they take a single external argument that is the cause of the verb’s event.4 Furthermore, recall that verbs of “non-voluntary emission of stimuli that impinge on the senses” are considered unaccusative on Perlmutter’s criteria (see Section 2.3.1). It may be that close scrutiny of all of these components of animacy will reveal inconsistencies and problems. It seems that a precise formal definition of animacy, as well as the precise role of animacy in grammar, has been historically difficult to pin down. Comrie concludes that “the animacy hierarchy cannot be reduced to any single parameter, including animacy itself in its literal sense, but rather reflects a natural human interaction among several parameters, which include animacy in the strict sense, but also definiteness . . . and various means of making an entity more individuated such as giving it a name of its own, and thereby making it also more likely as a topic of conversation” (Comrie, 1989, p. 199). In the introduction to a special issue of Lingua devoted to animacy, de Swart et al. (2008) explain that although pervasive effects of animacy on argument structure, morphology, and language processing have been widely observed, there is little agreement about the exact nature of these effects (see also discussion in Chen (2012)). But rather than export the definition problem onto other terminology, I favor the term animacy because its core meaning contains 4 I’m grateful to Julie Brittain for bringing these facts to my attention.

The Thematic Hierarchy

75

all of the essential ingredients of the distinction I want to highlight. Chen (2012) points out that the adjective animate, from which the noun animacy is derived, is itself derived from the Latin animatus which includes in its meaning not only “alive” and “living” but also “spirited, disposed, inclined” (Oxford English Dictionary). Thus, the term animacy as it is used in linguistic studies goes beyond the simplistic binary biological distinction (alive vs. not alive) which is obviously insufficient; it allows for agency (without necessarily invoking problems with tests of adverb modification); it implies the capacity for control without requiring that such control be exercised. Taking a cue from Dahl (2008), “the capacity for perceiving and acting upon the environment is more or less what one would see as the defining criterion for being animate.” When it comes to hypothesizing about cues children use in learning language we should bear in mind that to the extent that these features overlap to a very large degree (high degree of animacy ↔ high degree of control/agency ↔ high degree of teleological capability) it may not matter which one of them is the true source of the linguistic effect. Rather, whichever of these properties is most easily determined by children either via innate expectations about the world or observation of the world (or both) could be used to exploit the robust association between animate (agentive/controlling/capable) NPs and agent role/subject position on the one hand, and between inanimate (nonagentive/non-controlling/non-capable) NPs and theme role/object position on the other. We will say more about children’s concept of animacy in Chapter 5. For now we will turn to the other argument hierarchy that bears on interactions between animacy and argument structure, the Thematic Hierarchy.

3.2

The Thematic Hierarchy

In this section we will look at how thematic relations (such as agent, theme, etc.) have been argued to be organized into a hierarchy of relations. I will capitalize on the fact that even researchers who disagree about the formalism of thematic roles, their place in the structure of language, and even the inventory and definitions of the roles themselves, still largely agree on the following generalization, quoted here by Marantz (1984). In English and many other languages, it is generally true that if one of the inherent roles associated with a verb is an agent role (the role of an active, animate being who intentionally causes something), then this role will be assigned to the logical subject of the predicate . . . It is also generally true in these languages that theme inherent roles (roles of objects which

76

Argument hierarchies the verb specifies to undergo a change of state or location or of which the verb indicates a location) and patient inherent roles (roles of objects that bear the brunt of the action described by the verb) are . . . borne by logical objects. (Marantz, 1984, p. 32)

First I provide some historical context for treatments of thematic relations in linguistic theory, at the end of which we will look at two proposals for formal analyses of thematic, or theta, roles. Following that we will look at cross-linguistic properties of displacing predicates and displaced subjects. At the end of the chapter I will put forward an explicit proposal for how the mapping between the Thematic Hierarchy and argument structure could be used by children to acquire the structure of opaque constructions. 3.2.1 A brief history of thematic roles Thematic roles, or thematic relations, are labels applied to arguments, which encode the semantic role the arguments play with respect to a predicate. Since the particular role label given to an argument is often considered to depend on both semantic considerations (semantic features of the argument and lexical semantics of the predicate) and syntactic considerations (whether the argument is the subject vs. direct object vs. oblique), the study and explication of thematic roles sits squarely at the interface between syntax and semantics. One of the intuitive justifications for having a system of thematic roles is that the meaning of certain arguments can be independent of syntax in a strict sense. Thematic roles allow us to capture the intuition that John plays the same semantic role across the sentences in (23a–d) (and similarly for the book and Bill across the same sentences), even if it occupies different syntactic positions in each case. The same can be said for Mary and the door in (24a–c). Such groups of sentences are termed “thematic paraphrases” by Baker (1988).5 (23)

a. John gave the book to Bill. b. John gave Bill the book. c. The book was given to Bill by John. d. Bill was given the book by John.

5 In this discussion I will be chiefly concerned with the thematic roles borne by arguments as

opposed to adjuncts; this distinction is also made explicitly by Dowty (1991) and is the relevant distinction for Chomsky’s Theta Criterion.

The Thematic Hierarchy (24)

77

a. Mary closed the door. b. The door closed. c. The door was closed by Mary.

In this way, thematic roles appear to be independent of syntax. In another sense, though, they are quite dependent on syntax. For example, in (25a) the subject of the sentence can be labeled the agent, and in (25b) the subject can likewise be labeled the agent, even though the two NPs have very different roles in the exchange denoted by the verbs (and, the two individuals, John and Bill, have essentially the same role in both sentences; sometimes the respective subject roles in (25a) and (25b) are distinguished as goal vs. source). Similarly, in (26) the NP the truck is a theme in (26a) but a location in (26b), even though the relation between the loading event and the physical truck is essentially the same in both cases.6 (25)

a. John bought a book from Bill. b. Bill sold a book to John.

(26)

a. Mary loaded the truck with hay. b. Mary loaded hay onto the truck.

The first attempts in modern linguistics to compile a set of thematic roles were made both by Gruber (1965) (published as Gruber (1976)), later developed in Jackendoff (1972), and by work in Case Grammar (Fillmore, 1968).7 Within Case Grammar NPs were labeled with a Deep Case (Agentive, Instrumental, Objective, Dative, Locative, etc.) according to the semantic role played by the NP within the proposition, irrespective of how or whether the NP was marked overtly with any case morphology. This Deep Case of the NP remained the same under transformations, such as those above (see (23) and (24)). 6 According to Dowty (1991), the NP hay occupies the same theme role in both versions of (26) in

both Gruber’s and Jackendoff’s approaches to thematic roles. Dowty provides arguments against this view. Note that both hay and truck qualify as themes for different reasons: the hay is the “thing that moves” yet the truck is the “thing that undergoes a change of state.” 7 Dowty (1991) cites an even earlier attempt by Blake (1930), who used the term case to refer to these roles, terminology which clearly influenced Fillmore’s own use of the term (Fillmore, 1968, p. 21). Considerable work in Generative Semantics also dealt with thematic roles (Katz and Fodor, 1963; Katz and Postal, 1964; Lakoff, 1971; McCawley, 1968, among many others). This body of work will not be discussed here.

78

Argument hierarchies

Fillmore defined Agent as “the case of the typically animate perceived instigator of the action identified by the verb” (Fillmore, 1968, p. 24), and a hierarchy akin to that found later in Jackendoff’s work was effected by his rule of Case-to-position correspondence (op.cit., p. 51): (27)

If there is an A [Agentive argument], it becomes the subject; otherwise if there is an I [Instrumental argument], it becomes the subject; otherwise, the subject is the O [Objective argument (=Theme)].

Yet it was recognized by others working in that framework that the notion of agency probably included several distinct properties, such as Volitive, Effective (for inanimate instigators, such as stones, for example), and Initiative (Cruse, 1973), or was constructed out of sets of distinctive features, such as Intent, Cause, Controller, Animate, Concrete, and so forth (Nilsen, 1973). Whether Agentive case assignees had to be animate was the subject of debate. In early work Fillmore took Agentive case to be (typically) restricted to animate NPs, and he employed Instrumental case for inanimate causers, including forces of nature (Fillmore (1968); these are distinguished as Force by Huddleston (1970)). However, Huddleston (1970) raised some arguments against distinguishing Agentive case from Force/Instrumental case purely on the basis of animacy. He noted that something like “volition” is a more appropriate feature, but that even this raises problems because animate entities can but need not act volitionally. In later work Fillmore argued against using labels like Force and Instrument for inanimate instigators, since he reasoned that the labels of Deep Cases should be categorical and not subject to the particular relations of particular lexical items (Fillmore, 1977). Thus, in the end his view was that Agentive case need not be restricted to animate referents. Gruber defined an agent as “the entity which willed the action and effected it” (Gruber, 1976, p. 43), and the theme as “the entity which is conceived as moving or undergoing transitions” (op.cit., p. 38). Gruber’s characterization of thematic relations centered on properties of verbs that enable them to take certain types of arguments. That is, he spoke of verbs being “agentive” if they take an agent argument as subject. Gruber’s agents were basically animate, but he conceded that instruments could stand in for an agent as the subject of an agentive verb in certain cases (e.g. Poison finally killed Rasputin; op.cit., p. 206). He also noted that adverbs like intentionally could modify agentive verbs (John knocked over the vase intentionally), although this criterion seems at odds with the ability of instruments (inanimates) to substitute for agents (*Poison killed Rasputin intentionally; recall problems with the adverb modification test for identifying predicates that require animate subjects – see Section 3.1.2).

The Thematic Hierarchy

79

In the preceding section (Section 3.1), and to be discussed further below, we saw that there is a general tendency to relate an agent or animate NP with the subject position in a clause. Interestingly, Gruber seemed to link, instead, the theme role with the subject position: he pointed out that “there is some association between the Theme and the subject of deep structure,” and that the theme “has the significance of being an obligatory element of every sentence” (Gruber, 1976, p. 38). Further, he noted that the “meaning relationships among words were often complicated by the fact that certain subjects had the peculiarity that they were also Agents” (op.cit., p. 157; emphasis mine). Nevertheless, in Jackendoff’s (1972) development of Gruber’s groundwork, the subject position was again tied to the agent role: “Agents are invariably found in subject position” (Jackendoff, 1972, p. 42). Jackendoff’s Thematic Hierarchy is given in (28): (28)

Agent > Location/Source/Goal > Theme

3.2.2 Formal accounts of thematic role assignment While the general intuition behind most basic thematic roles and their correspondence to structural positions in simple sentences is fairly well agreed upon, formal definitions of these roles and their precise relation to syntactic structure are as varied as the characterizations of “Agent.” Here we will examine the frameworks of Chomsky (1981) and Jackendoff (1990), followed by an alternative approach which views the roles as clusters of entailments rather than as discrete categories (Dowty, 1991). Chomsky’s (1981) Theta theory was meant to formalize the distinction between “arguments” of a predicate, i.e. those expressions that have a particular reference (names, pronouns, traces, etc.), and expressions that lack such a reference, such as idiom chunks and pleonastic forms (expletive it and existential there). Arguments serve to satisfy the subcategorization features of predicates. That is, a verb “subcategorizes” for a certain number of arguments according to its valency, and the structural positions these arguments occupy are called theta (θ )-positions. θ -roles are then what get assigned to arguments in θ-positions. Thus, θ -roles in Chomsky’s framework served as (semantic) indices of syntactic arguments. As part of their argument-indexing function, θ -roles could be used to justify constraints on the number of arguments projected in the structure (*John kicked the ball the stone). This could only work if θ-roles bore a one-to-one correspondence with arguments (e.g. the ball and the stone cannot both bear the patient role). Thus, Chomsky’s Theta Criterion included a bi-unique relation

80

Argument hierarchies

between arguments and θ -roles, as spelled out in (29) (Chomsky, 1981, p. 36). (Recall that Hornstein (1999) dispensed with this requirement in his syntactic unification of raising and control; see Section 2.1.1.) (29)

Each argument bears one and only one θ -role, and each θ-role is assigned to one and only one argument.

The complement of a verb is always θ -marked by the V head, while the subject position is not necessarily θ -marked. Its θ-role is determined by the VP complex. This conception of the difference between the subject and any complement(s) of the verb is consistent with Williams’ (1984) distinction between external (subject) and internal (object) arguments, and with Marantz’s (1984) asymmetry between selection of the direct object by the verb and selection of the subject by the predicate (verb plus complement). In other words, the role assigned to the subject is assigned by the whole predicate (verb plus object), rather than just the verb. The verb and its assigned roles can be represented in a theta grid, which indexes the number and types of roles that must be saturated for a particular verb (Stowell, 1981). An example for the transitive verb kick is given in (30).8 (30)

Theta grid for kick Agent (NP) Patient (NP) i j

(31)

Impossible grids for kick a.

Agent (NP) i

Patient (NP) j

b.

Agent (NP) i

Patient (NP) ?

Goal (NP) ?

A further detail of GB Theta Theory that is relevant for displacing predicates is that θ-roles are assigned at D-structure, and NPs may under certain circumstances move from their D-structure position to a different position at S-structure (see discussion in Section 2.1.1 above). More precisely, A(rgument)-movement, or NP-movement, is motivated when an NP receives a θ-role but no Case in its D-structure position, since that NP must then move in order to get Case. An example of such a circumstance is given in (32) with the raising verb tend. 8 This example is adapted from Haegeman (1991, p. 45).

The Thematic Hierarchy (32)

81

a. e tends [John to rent horror movies] b. Johni tends [t i to rent horror movies]

The NP John receives its (agent) θ-role at D-structure from the predicate rent horror movies (he is the renter). But as the subject of an untensed clause, according to Case Theory, John lacks Case and is therefore unlicensed. Furthermore, the raising verb tend has no external argument indicated by the e (for “empty”) in subject position. John then raises into the empty subject position at S-structure. Crucially, though, John’s θ-role was assigned in its pre-movement position by its (then) local predicate. Jackendoff’s (1972; 1983; 1990) approach to thematic roles placed a premium on conceptual structure rather than syntactic structure (though syntactic structure was not ignored), and thematic roles were seen as “nothing but particular structural configurations in conceptual structure.” In contrast to the strict bi-unique mapping between θ -roles and arguments of Chomsky’s Theta Criterion, Jackendoff (1990) argued that it is necessary to allow both individual arguments to bear more than one thematic role, and multiple (distinct) arguments to bear the same thematic role. He notes that in a buying exchange, not only is the object bought transferred from one participant to another, but money is also exchanged in the opposite direction between the same two parties. Thus, in (33), Bill and Sue each have two thematic roles. Bill is the goal of the object transfer and source of the money transfer, while Sue is the source of the object transfer and goal of the money transfer. (33)

Bill bought a book from Sue.

The opposite situation arises in examples like (34), where the pronoun it refers to the same entity as the box, and this entity has only a single role in the situation (location). (34)

The box has books in it.

Jackendoff avoids these problems by developing a revised θ -Criterion, termed the Linking Condition, which states that “each syntactic argument is linked uniquely with a conceptual argument” (Jackendoff, 1990, p. 64). In the case of (33) above, both the movement of the book and the exchange of money are indicated in the conceptual structure, but the entities engaged in the secondary transfer are indicated as variables bound by the arguments introduced in the primary transfer.

82 (35)

Argument hierarchies ⎡





⎤

F ROM[SU E]α ) T O[BI LL]β ⎣  ⎦

F ROM[β] )]] [EXCH [GO Poss ([MON EY ], T O[α]

GO Poss ([BOOK],

Presumably the situation in (34) would be dealt with in a similar way, so that both the box and it would bear the same index. Dowty (1991) noted that the very fact that the assignment of thematic roles is influenced by both syntax and semantics makes them very hard to pin down precisely in all but the most ordinary cases. He lamented that [t]he variety of semantic distinctions that correlate with syntactic and lexical patterns in one way or another is surely enormous. To postulate thematic role types for each of them is . . . to dilute the notion beyond its usefulness, but what we lack is a principled way to decide what kind of data motivates a thematic role type. (Dowty, 1991, p. 561)

Instead, Dowty proposed a set of semantic entailments that yield two “fuzzy” proto-roles, Proto-Agent and Proto-Patient. The potential characteristics of a P(roto)-Agent are those in (36), and the characteristics of a P(roto)-Patient are those in (37) (from Dowty (1991, p. 572)). (36)

P-Agent entailments a. volitional involvement in the event or state b. sentience (and/or perception) c. causing an event or change of state in another participant d. movement (relative to the position of another participant) e. exists independently of the event named by the verb9

(37)

P-Patient entailments a. undergoes change of state b. incremental theme10

9 With verbs of creation (build, write, etc.) the direct object exists by virtue of the creative event

while the subject exists independently. In The student wrote a poem the poem’s existence is tied to the writing event, but the student’s existence is not. 10 An incremental theme is an object that is created or destroyed by the verb’s event, or otherwise involves what Dowty calls a “definite” change of state. Examples include the objects in build a house, eat a sandwich, and proofread an article. Incremental themes contrast with “traditional” themes which are understood to undergo some change of state, but it is, in Dowty’s terms, “indefinite.” Examples include the objects in push a cart or raise a thermostat. If an article has been proofread, the implication is that it has been wholly proofread, while a thermostat could be raised by smaller or larger degrees.

The Thematic Hierarchy

83

c. causally affected by another participant d. stationary relative to the movement of another participant e. does not exist independently of the event, or not at all

Interestingly, properties of animacy are not used in any of these entailments. Presumably, animacy is already entailed by the properties of volition and sentience, and it entails existing independently of an event. Since Dowty’s roles are prototypes, and since the properties in (36–37) are not to be considered as distinctive features, it is possible for an NP argument both to have a subset of the properties of P-Agent or P-Patient, and to have some P-Agent properties and some P-Patient properties without contradiction. Indeed, this is Dowty’s whole point: setting up the prototypical characteristics of agents and patients in this way allows us to capture cases that are difficult to account for under more traditional, strict categories of thematic roles. For example, the subject in His loneliness caused his unhappiness is neither volitional nor even tangible (in a physical sense), does not move, etc., yet it causes something. The object his unhappiness in the same sentence cannot be said to be “stationary” with respect to other participants in the sentence, yet it can be said to be “causally affected.” Dowty also notes that the various semantic entailments can be combined to effectively define the traditional thematic roles: an agent has the properties of volition, causation, sentience, and movement, while an experiencer has sentience without volition or causation, for example. Because these distinctions fall out from the entailments themselves, the roles do not need to be stipulated independently. While Dowty’s characterization represents, in some sense, a radical departure from traditional discrete definitions of thematic roles (whether Chomsky’s, Gruber/Jackendoff’s or that of Case Grammar), the intuition behind Jackendoff’s Thematic Hierarchy remains apparent in Dowty’s Argument Selection Principle. This principle states that the argument with the most P-Agent properties will be the subject, and the argument with the most P-Patient properties will be the object (Dowty, 1991, p. 576). This principle captures the intuition that subjects will have more prototypical agent-like properties while objects will have more prototypical patient/theme-like properties, but because his roles are characterized in terms of entailments, rather than discrete features, apparent exceptions can be accounted for. In addition to the unusual cases exemplified above, Dowty accounts for the behavior of psychological predicates that take a theme subject and experiencer object (e.g. frighten). In these cases the theme subject causes an emotional reaction in the (experiencer) object, and causation is a P-Agent property of the subject (also sometimes called a stimulus (Talmy,

84

Argument hierarchies

1985) or target/subject matter (Pesetsky, 1995)); the experiencer object, in turn, is causally affected by the theme, and this is a P-Patient property. Even though the theme subject still retains various P-Patient properties and the experiencer object retains various P-Agent properties, the opposite entailments appear to cancel them out and make both arguments more or less equally good subjects or objects. Interestingly, these verbs have counterparts that take an experiencer subject and theme object, e.g. fear, where the entailments of the subject and object arguments are reversed (Dowty, 1991, pp. 579ff.). One further piece of formalism that will be useful for our purposes is Baker’s (1988) proposed Uniformity of Theta Assignment Hypothesis (UTAH). According to this hypothesis, “identical thematic relationships between items are represented by identical structural relationships between those items at the level of D-structure” (op.cit., p. 46). The relevance of UTAH for the claims made here is that it formalizes the intuition we have seen expressed repeatedly: that agents are canonically mapped to the external argument position, while themes and patients are canonically mapped to the internal argument position. Assuming that this mapping takes place at some underlying level of representation (for Baker it was D-structure; in Minimalist terms it would happen at the point of Merge), then NPs that are subjects on the surface but are poor candidates for agents (by virtue of being inanimate) are likely to have originated elsewhere. This means they are derived, or displaced subjects. A second reason for recalling UTAH is that there is a strong implication in Baker’s work that UTAH is innate – children expect, a priori, that certain thematic relationships will appear in certain syntactic configurations. If true, then UTAH is a powerful tool for explaining how children begin to tackle the problem of parsing sentence strings. Agents will typically be subjects, and themes will typically be direct objects, and most of the time this assumption will yield a correct analysis of argument structure (it will work, for example, for Keenan’s “basic sentences,” as discussed below). According to UTAH, if a child encounters a non-agent/non-causer subject, he should be inclined to assume that that subject NP is not the external argument of its adjacent verb. Thus, UTAH is essentially the reason, from the perspective of Universal Grammar, that inanimate subjects should serve as an informative cue for learners that the sentence has a non-canonical syntactic structure. 3.3

Animacy and thematic roles in opaque constructions

The preceding two sections have established that (a) there is a general tendency across languages for the subject to be more animate (higher on the Animacy

Animacy and thematic roles in opaque constructions

85

Hierarchy) than the object, and (b) there is a general tendency across languages for agents, which are canonically animate, to be projected in the syntactic subject position. In order to make the case that the violation of these tendencies can lead learners to acquire the classes of raising verbs, tough-adjectives, and unaccusative verbs, we must now show that, cross-linguistically, these sorts of predicates have either a derived subject or no subject at all. Focusing mainly on raising-to-subject and tough-constructions, we will first examine the argument structure properties of the lexical items corresponding to English raising verbs and tough-adjectives in different languages. That is, what kinds of syntactic constructions do verbs meaning ‘seem’ occur in, and how do they compare with those of verbs meaning ‘want’ or ‘try’? What kind of argument structures do sentences with adjectives meaning ‘easy’ or ‘hard’ have? And to turn the question around, what lexical meanings are associated with predicates that allow NP-movement in different languages? It is not so important to show that sentences containing a predicate meaning ‘seem’ or ‘easy’ have constructional homonyms containing non-displacing predicates crosslinguistically. Rather, what is important is that, universally, these predicates select a clausal complement and no external argument – they are, therefore, the sorts of predicates that could in principle host a derived subject, if other aspects of the language’s syntax allow NP-movement. Following that, in Section 3.4, I will present evidence that derived subjects typically are “less animate,” or otherwise less “subject-like” than underived (underlying) subjects. These arguments come mainly from Keenan (1976), but we also see that in one particular language that restricts transitive subjects to being animate (Jacaltec), derived subjects are allowed to break this restriction. Finally, in Section 3.5 I will propose the Derived Subject Condition, which spells out a learning procedure that children could utilize in order to begin identifying raising predicates. An important qualification must be made here. In Chapters 1 and 2 I have characterized displacing and non-displacing predicates as forming a dichotomy, although we also saw in Section 2.5 one way in which this dichotomy breaks down (namely, the existence of predicates that are ambiguous between the two classes). A further way in which this dichotomy breaks down will be revealed in the second part of Section 3.3.1, where we see evidence that raising and control predicates form more of a continuum, from “more raising-like” to “more control-like” rather than a strictly binary distinction. Clearly, the characterization of these predicates as forming a dichotomy vs. a continuum has repercussions for the type of learning procedure needed. For simplicity, I will still generally talk about these predicates

86

Argument hierarchies

as if they constituted two distinct classes, but when it comes to laying out the actual learning process, I will argue that the probabilistic strategy I propose is necessitated by the more continuous nature of these predicates and their degree of displacing behavior. 3.3.1 Raising constructions across languages Though many descriptive grammars provide examples of passive (and antipassive) constructions, few explicitly discuss patterns of raising or toughmovement. Nevertheless, where cross-linguistic data on these constructions exists, the constructions exhibit broadly similar patterns concerning the underlying argument structures. That is, although languages may differ in the surface representations of these constructions, contingent upon certain languagespecific properties of syntax, in general the lexical items that correspond to the English raising verbs (verbs meaning ‘seem,’ ‘turn out,’ and so forth) select a clausal complement and no external argument. As we will see in Section 3.3.2 tough-constructions will involve roughly the same set of lexical meanings, again selecting a clausal complement and no external argument. Crucially, the predicates that fail to select an underlying external argument will allow an inanimate subject. Thus, while surface patterns may result in these constructions not being opaque in other languages they way they are in English, the argument structure properties that I argue lead to learning of the constructions in English will hold true in other languages as well. Here I will discuss data from two European languages, German and Italian, and the Polynesian languages Maori and Tongan. Additional data, where relevant, will be introduced from Niuean, Chamorro, and Indonesian. Both German and Italian have raising-to-subject constructions, exemplified in (38) ((38a) is from Olsen (1981)). (38)

a. Die Sache scheint ihm über den Kopf zu wachsen the/f/Nom issue seem-3s him/Dat over the/m/Acc head to grow-inf “The issue seems to be over his head.” b. Spesso i documenti dei lavoratori stranieri risultano non often the/p document-p of-the/p worker-p foreign-p turn out-3p neg essere in regola be-inf in order “The documents of foreign workers often turn out not to be in order.”

Let us turn first to German raising constructions. There are several notable differences between English and German raising constructions. Since German allows topicalization much more liberally than English, not only the embedded

Animacy and thematic roles in opaque constructions

87

subject, but also other phrases within the embedded clause may raise to the matrix subject position (Olsen, 1981).11 (39)

a. Ihm scheint die Sache über den Kopf zu wachsen him/Dat seems the issue over the head to grow b. Über den Kopf scheint ihm die Sache zu wachsen over the head seems him/Dat the issue to grow c. Mir scheint ihm die Sache über den Kopf zu wachsen me/Dat seems him/Dat the issue over the head to grow “The issue seems (to me) to be over his head.”

But in their basic argument structure German and English raising verbs are alike. Wurmbrand (2001) argues that the set of raising verbs, including scheinen ‘seem’ and pflegen ‘used to,’ are functional verbs that lack thematic properties. That is, German raising verbs, like English raising verbs, fail to select an external argument, and allow weather-it and inanimate subjects. (40)

a. Es scheint zu schneien it seem-3s to snow-inf “It seems to be snowing” b. Der Kuchen scheint gut zu schmecken the/m/Nom cake seem-3s good to taste-inf “The cake seems to taste good”

In this respect, German raising verbs clearly differ from German control verbs. (41)

a. * Es versucht zu schneien it try-3s to snow-inf (“It’s trying to snow”) b. * Der Kuchen versucht gut zu schmecken the cake try-3s good to taste-inf (“The cake is trying to taste good”)

Wurmbrand (2001) distinguishes these verbs both by their thematic properties (control verbs assign a θ -role to their subject – they select an external argument; raising verbs are “nonthematic”) and the type of “restructuring” they allow. Restructuring refers to an analysis of certain infinitival complements in which the complement clause is argued to be less than a full clause: it is either 11 Related to the verb-second property of German main clauses, many accounts of German syntax

place the preverbal phrase in SpecCP rather than SpecIP/SpecTP position, after den Besten (1977); Koster (1978).

88

Argument hierarchies

reduced (e.g. only IP or vP instead of CP), or it is not a separate clause from the main clause at all (Evers, 1975; Napoli, 1981; DiSciullo and Williams, 1987; Rochette, 1988; Rosen, 1989; Cinque, 2000, among many others). That is, certain infinitival complement clauses behave as if they are, syntactically, part of the main clause. Such clauses are said to undergo “restructuring,” which renders them transparent to the main clause in terms of syntactic operations that depend on locality, such as A-movement. Wurmbrand distinguishes two types of restructuring, lexical and functional. Lexical restructuring verbs are those that select an argument (e.g. versuchen ‘try’ selects an external argument) and thus have some thematic weight to them. Functional restructuring verbs, on the other hand, include modals, auxiliaries, and raising verbs, as well as some verbs of motion and perception; in general these verbs either never assign any θ -roles (such as raising and auxiliary verbs) or are at least thematically “lighter” than regular transitive or intransitive verbs (cf. English go in He went swimming). Wurmbrand’s distinction between lexical and functional restructuring lines up with several syntactic asymmetries between raising and control verbs. For example, only verbs whose complements involve lexical restructuring allow extraposition of the infinitive, as in (42) (from Wurmbrand (2001, pp. 156–157)). (42)

a.

dass Hans versuchte [den Kuchen zu essen] that John try-pst the/m/Acc cake to eat-inf “that John tried to eat the cake”

b. * dass Hans schien [den Kuchen gegessen zu haben] that John seem-pst the/m/Acc cake eat-part to have-inf (“that John seemed to have eaten the cake”)

According to Wurmbrand, the reason for this asymmetry stems from different properties of the restructured clause. In the case of lexical restructuring, the infinitival complement of the lexical verb (e.g. versuchen ‘try’) is an independent prosodic phrase, and as such it is a candidate for extraposition. Functional verbs, on the other hand, behave prosodically as if they are part of the infinitive, and therefore the infinitive cannot be separated from the main verb by extraposition. A further difference between raising and control verbs concerns their behavior with passives. While control verbs allow a “long passive” construction (43a), raising verbs do not, but instead allow their embedded clause to be passivized (44b) (examples from Wurmbrand (2001, p. 178)).

Animacy and thematic roles in opaque constructions (43)

Der Wagen wurde zu reparieren versucht the/m/Nom car was to repair-inf try-part “The car was tried to be repaired” b. * Der Wagen wurde zu reparieren gemusst/geschienen the/m/Nom car was to repair-inf must/seem-part (“The car was had-to/seemed to be repaired”)

(44)

a. * Der Wagen versuchte repariert zu werden the/m/Nom car tried repair-part to become-inf (“The car tried to be repaired”) b. Der Wagen muß/scheint repariert (zu) werden the/m/Nom car must/seems repair-part (to) become-inf “The car must/seems (to be) repaired”

89

a.

Wurmbrand notes expicitly that these facts “follow straightforwardly from the assumption that verbs like try establish a thematic relation with their subjects, whereas verbs like must [or seem] are true raising verbs that do not assign a theta-role to the subject” (Wurmbrand, 2001, p. 178). One interesting point about the set of verbs that are categorized as raising or control in German is that the verbs drohen ‘threaten,’ versprechen ‘promise,’ and beginnen ‘begin’ display hybrid behavior, just as the corresponding verbs do in English (see discussion of begin in Section 2.5). These three verbs can function as control verbs – thus permitting extraposition, as exemplified in (45a) – but only when the subject is animate (cf. (45b); these examples are from Olsen (1981, pp. 128–129)). (45)

a.

weil man mir gedroht hat, mich in ein because one me/Dat threaten-part have-3s me/Acc in a Irrenhaus zu sperren insane asylum to lock-inf “Because they threatened to lock me in an insane asylum” b. * weil das Spiel droht, aus den Fugen zu geraten because the/n/Nom game threaten-3s out the/m/Acc track to fall-inf (“because the game threatens to get off track”) c. Das Spiel droht aus den Fugen zu geraten the/n/Nom game threaten-3s out the/m/Acc track to fall-inf “The game threatens to get off track”

When the verbs corresponding to threaten, promise, and begin have an animate, agentive subject, they behave as control verbs and, correspondingly, as lexical restructuring verbs (thus allowing the infinitive to be extraposed, as in (45a)). With an inanimate subject, they do not show this behavior (cf. (45b)).

90

Argument hierarchies

But these verbs can also take an expletive subject, as seen clearly in (46), a hallmark of raising verbs. (46)

Es begann/drohte/versprach zu schneien it began/threatened/promised to snow-inf “It began/threatened/promised to snow”

The preceding data show that the underlying argument structure of scheinen and other German raising verbs is identical to that of English raising verbs (in particular seem, but the facts extend to appear, happen, turn out, and used-to). In both languages, these verbs select no external argument and select a clausal complement (47).12 (47)

a. e scheint [die Sache ihm über den Kopf wachsen] b. e seems [the issue to be over his head]

As for Italian, we find some underlying similarities between Italian and English raising verbs, as well as some differences. We saw above in (38b) that, like English, Italian permits certain verbs (e.g. risultare ‘turn out’) to take a non-thematic subject and an infinitive complement. We can also observe that Italian sembrare ‘seem,’ like English seem, fails to select an external argument and does take a finite clausal complement, as in (48). (48)

e sembra che Gianni sia stanco seem-3s that Gianni be-subj tired “It seems that John is tired”

Also indicative of its raising-like argument structure is the ability of sembrare to occur in weather constructions (cf. (46) above) and with inanimate subjects. In this respect, sembrare differs from Italian control verbs, such as provare ‘try.’ (49)

a. e sembra che e piova / e sembra e piovere seem-3s that rain-subj/3s / seem-3s rain-inf “It seems to be raining” b. Questo libro sembra troppo lungo this book-m seem-3s too long-m “This book seems (to be) too long”

12 It might be argued that Wurmbrand’s structure commits us only to a VP complement of

scheinen. The point, however, is that these verbs take no direct object and no thematic subject.

Animacy and thematic roles in opaque constructions (50)

91

a. * e provava e piovere try-impf/3s rain-inf (“It was trying to rain”) b. * Questo libro provava essere troppo lungo this book try-impf/3s be-inf too long (“This book tried to be too long”)

However, sembrare and seem are not alike in all respects. In particular, when the complement of sembrare is not finite it cannot host an infinitive verb plus adjective (51a), though a simple adjective with no infinitive verb is allowed (51b). (51)

a. * Gianni sembra (di) essere contento Gianni seem-3s (to) be-inf happy (“Gianni seems to be happy”) b.

Gianni sembra contento Gianni seem-3s happy “Gianni seems happy”

Moro (1997) points out that sembrare can take an infinitive complement, but movement out of that complement is not allowed. (52)

pro sembra [CP di PRO volare] pro seems to PRO fly-inf “It looks like flying”

After Kayne (1984), Moro notes that di (like French de) is a Complementizer that blocks government from the higher clause. Thus, the subject of the infinitival clause is PRO rather than trace. In this sense, sembrare is not properly a “raising” verb. Moro (1997) argues that Italian sembrare is actually lexically ambiguous between English ‘seem’ and English ‘resemble,’ as it can take a direct object. That (53a) involves a direct object and not a clausal predicate can be seen in the ungrammaticality of (53b) (examples based on Moro (1997, p. 209)). (53)

a.

Mariai sembra [DP la suai maestra] Maria seems the her teacher “Maria resembles her teacher”

b. * Mariai è [SC t i [DP la suai maestra]] Maria is teacher the her (own) (“Maria is her teacher”)

92

Argument hierarchies

Another possible difference between Italian and English, according to Cinque (1995), is that Italian lacks raising adjectives. On the other hand, Ivano Caponigro (personal communication, 7/6/2011) judges (54b) to be perfectly good, with a slight lexical preference for sicuro ‘sure’ over certo ‘certain.’ (54)

a. Johni is sure/certain [t i to win the election] b. Berlusconi è sicuro/?certo [di vincere le elezioni] B. be-3s sure/certain to win-inf the-f/p election-p “Berlusconi is sure/certain to win the election”

Since raising adjectives like likely, sure, and certain behave differently from the raising verbs (seem, appear, happen, turn out) in several ways, and since the Italian judgments are not clear, I will not dwell on them here.13 Thus, there are various language-specific differences between Italian sembrare and English seem, but crucially for our story sembrare has the same underlying argument structure as seem in the relevant respects: it takes a clausal complement and no external argument, in other words, no underlying subject. Table 3.1 provides a summary of the behavior of raising and control predicates in German and Italian. Next, let us examine patterns of raising in Polynesian languages. Here I will present data from two Polynesian languages, focusing on subject raising and subject control. The data come from Maori, which is a morphologically accusative language, and Tongan, which is a morphologically ergative language. Both Maori and Tongan have an underlying VSO word order (Chung (1978); see also work by Anderson and Chung (1977) who report that Samoan, also underlyingly VSO, exhibits subject-to-subject raising and subject-to-object raising). Subject raising in Maori is triggered by a negative verb (glossed as not, don’t) which takes a clausal complement and no external argument. The subject of the complement clause raises to the higher clause containing the negative verb. Raising is optional for full NPs but reportedly extremely common, and it is obligatory for pronominal NPs. (55)

a. K¯ihai a Tamahae i haere not det T. pst go “Tamahae didn’t go” (raised)

13 Olsen (1981) devotes several chapters to analyzing the differences between predicates like

likely as opposed to seem. To give two small examples, likely allows preposing of the whole complement clause while seem does not: That John has quit smoking is likely, *That John has quit smoking seems; and likely can occur in a pseudocleft while seem cannot: What is likely is that John has quit smoking, *What seems is that John has quit smoking.

Animacy and thematic roles in opaque constructions

93

Table 3.1: Raising and control verbs in German and Italian Language

Raising verb properties

Control verb properties

German

fail to select external argument occur with inanimate, expletive subjects functional restructuring disallow extraposed infinitive disallow “long” passives allow passive embedded clause

examples

scheinen ‘seem’ pflegen ‘used to’

select external argument disallow inanimate, expletive subjects lexical restructuring allow extraposed infinitive allow “long” passives disallow passive embedded clause versuchen ‘try’ vergessen ‘forget’

Italian

fail to select external argument occur with inanimate subjects take weather predicate complements disallow raising with infinitive complement sembrare ‘seem’ risultare ‘turn out’

examples

select external argument disallow inanimate subjects disallow weather predicate complements disallow raising with infinitive complement provare, cercare ‘try’ affermare ‘claim’

b. K¯ihai i haere a Tamahae not pst go det T. “Tamahae didn’t go” (unraised) (56)

a.

kaua t¯atou e haere don’t we nonpst go “Let’s not go!” (raised; from Biggs (1969), cited in Chung (1978))

b. * kaua e haere t¯atou don’t nonpst go we (“Let’s not go”) (unraised)

Chung (1978) gives extensive arguments for analyzing the negative verbs as verbs, and these structures as involving two clauses. As is expected if these negative verbs do not select an external argument, inanimate subjects can occur here. (57)

a. K¯aore an¯o te r¯a kia whiti not yet the sun subj shine “The sun hasn’t risen yet” (raised)

94

Argument hierarchies b. K¯aore an¯o kia whiti te r¯a not yet subj shine the sun “The sun hasn’t risen yet” (unraised)

Not only can raising apply to inanimate subjects, but it can also apply to semantic patients and experiencers of stative predicates. An example showing the raising of a semantic patient is given in (58). (58)

Kaua ahau e whakar¯ere-a don’t I nonpst abandon-pass “Let me not be cast aside” (Biggs, 1969)

In contrast to raising, control in Maori applies only to NPs that are both syntactic subjects and semantic agents, i.e. not to passivized subjects (59b), subjects of stative predicates (59c), or oblique agents (e.g. in a passive; (59d)). (59)

a.

Ka hiahia / whakaaro au ki te haere tam want / think I comp go “I want/decided to go”

b. * I hiahia koe ki te riro i te wahine pst want you comp taken caus the woman (“You wanted to be taken away by the woman”) c. * I hiahia au ki te patu-a e Rewi pst want I comp hit-pass agt Rewi (“I wanted to be hit by Rewi”) d. * I hiahia au ki te patu-a te poaka pst want I comp hit-pass the pig (“I wanted the pig to be killed (by me)”)

Although no data show explicitly that inanimate NPs are ungrammatical controllers, the fact that semantic patients cannot be controllers, even if they are animate, suggests that inanimate controllers would be disallowed. Raising and control in Tongan are quite similar to the corresponding constructions in Maori, with some minor lexical differences. Tongan raising occurs with the matrix modal verb lava meaning ‘be possible/be able.’ (60)

a. ’E lava ’a e p¯ep¯e ’o lea tam can Abs the baby comp talk “The baby can talk” (raised) b. ’E lava ke lea ’a e p¯ep¯e tam can subj talk Abs the baby “The baby can talk” (unraised)

Animacy and thematic roles in opaque constructions

95

Chung presents various pieces of evidence that the structure in (60a) involves NP-movement. For example, unlike controllers, whose case is determined by the matrix predicate (see below), raised NPs bear the case assigned by the embedded predicate. Thus, if the embedded verb is intransitive, a raised NP will bear Absolutive case, and if the embedded verb is transitive, a raised NP will bear Ergative case. (NB The verb meaning ‘eat’ is canonically intransitive in Tongan.) (61)

a. Te ke toe lava (*’e) koe ’o ki’i kai? tam you again can Erg you comp small eat “Can you eat just a little more?” b. Te ke lava *(’e) koe ’o lau ’eku mata’itohi? tam you can Erg you comp read my handwriting “Can you read my handwriting?”

Control constructions in Tongan behave somewhat differently from the control constructions we have seen so far, and I will argue that this can be linked to the lexical semantic properties of the verbs that induce control. Tongan control verbs are verbs of “motion or directed action” (Chung, 1978, p. 116), such as ‘go’ or ‘stand up,’ but not verbs of effort or desire.14 (62)

a.

Na’a ne tu’u hake ’o ’alu pst he stand up comp go “He stood up and went”

b. * Na’a ku feinga ’o u mavahe ’aneafi pst I try comp I leave yesterday (“I tried and left yesterday”)

Consistent with this semantic difference, Tongan control structures lack some of the semantic restrictions found in Maori: Tongan controllers can be semantic patients, and they can be inanimate: (63)

a. Na’a ku taa’i ia ’o mate pst I hit him comp die “I hit him and he died”

14 The correct form of (62b) would be that in (i), where the embedded clause is headed by the

subjunctive marker ke rather than the complementizer ’o: i.

Na’a ku feinga ke u mavahe ’aneafi pst I try subj I leave yesterday “I tried to leave yesterday”

It is worth noting that in Samoan, whose lexical meanings of control verbs are similar to those of Maori – verbs relating to desire and effort – control constructions are limited to having animate subjects.

96

Argument hierarchies b. Na’e t¯o mai ha matangi ’o haveki ’a e falé pst fall here a wind comp break Abs the house “Some wind came and blew down the house”

However, as expected according to both lexical semantics and the patterns we have seen so far, inanimate subjects are also allowed in raising constructions. (64)

’E lava ’e he m¯ahiná ’o tamate’i ’a e la’á? tam can Erg the moon comp shut Abs the sun “Is it possible for the moon to block out the sun?”

Chung (1998) also notes that Chamorro (Austronesian) has raising constructions with the aspectual verbs meaning ‘begin’ and ‘stop,’ in which no semantic restriction is placed on the subject ((65); examples from Chung (1998, p. 317)). In this respect they differ from control predicates such as ‘be afraid,’ which do select a semantic subject ((66), from Chung (1983, pp. 235–236)). (65)

a. Anai p-um-ara häm [um-atgumentu t] when agr-stop we inf-argue “When we stopped arguing” b. P-um-ara pro [um-uchan t] agr-stop inf-rain “It stopped raining”

(66)

a.

Ma’a’ñao yu’ [p-um-o’lu i salape’-hu gias Rita] be-afraid I inf-put the money-my Loc R. “I’m afraid PRO to leave my money with Rita”

b. * Ma’a’ñao [um-uchan] be-afraid inf-rain (“It’s afraid to rain”)

A final note of interest about the classes of raising and control verbs in the Polynesian languages is that, although individual languages appear to have quite a limited set of raising or control verbs, looking across the different languages there is significant overlap with the corresponding lexical meanings for these predicates in English, as well as the meanings we saw above in German and Italian. (NB These lists are not necessarily exhaustive; for the Polynesian languages they are all the examples I came across in the literature, and for the European languages they are a subset of the examples I obtained from my sources.)

Animacy and thematic roles in opaque constructions

97

Table 3.2: Raising and control verb meanings in Polynesian, European Language

Raising verb meanings

Control verb meanings

Maori (Chung, 1978)

not, don’t

Tongan (ibid.) Samoan (ibid.)

be able, be possible be able, be necessary, begin, be (done) thus

Niuean (Seiter, 1980)

be possible, begin, not, usual (customary), almost, nearly begin, stop

want, decide (think), go, be able, agree, prepare go, come, stand up, turn, hit want, try, encourage, go, think, (be) tired (of), (dis)like try, want, choose

Chamorro (Chung, 1998) German (Wurmbrand, 2001) Italian (Caponigro, p.c.)

be afraid

seem, used-to, must

try, forget, forbid

seem, turn-out

try, claim, pretend

Thus, the lexical meanings of raising predicates tend to cluster around the types of meanings we see in English modal verbs, the ambiguous verb begin, and other functional heads like negation, though there are interesting crosslinguistic differences as well. For instance, none of the raising verbs in the Polynesian or Austronesian languages have meanings related to appearance. The lexical meanings of subject control predicates likewise overlap to a large degree with the control predicate meanings in Indo-European languages, but again with some differences, especially in Tongan. 3.3.1.1 Bleached control verbs, copy raising, and the sliding scale of raising–control Cross-linguistically there is both relative uniformity in the semantic features of raising and control verbs, and also within languages we can see that, broadly speaking, raising and control verbs can be distinguished according to particular diagnostics (see Section 2.1). But it turns out that neither raising verbs nor control verbs are completely uniform either syntactically or semantically. In this section I’ll discuss some of the syntactic and semantic variation we find, within English and certain other languages, in these classes of verbs.

98

Argument hierarchies

The unacceptability of inanimate subjects with control predicates is semantic in nature, hence the # diacritic (rather than *) in sentences like #The book tried to be interesting. They are ill-formed in the same way as a sentence like #The paperweight devoured a rainbow. Metaphorical extension permits language users to bend the canonical semantic rules, either by attributing animate qualities to an inanimate object or by modifying a predicate’s typical meaning (the latter is known as the Verb Mutability Effect; Gentner and France (1988)). In the present case, modification of a control predicate’s meaning to make it more like a raising verb would involve a bleaching of its lexical meaning, i.e. making it more auxiliary-like. To give a concrete example, many speakers of English find the sentence in (67) marginally to perfectly acceptable, even though it violates the ban on control verbs taking a weather-it subject. (67)

It seems like it wants to rain.

This example actually brings to light two independent issues that raise problems for the dichotomy between raising qua non-thematic verbs and control qua thematic verbs. The first problem has to do with the ability of want to lack a thematic subject. The second has to do with the ability of seem (and similar raising verbs) to take a thematic subject. We will look at each of these briefly in turn. The sense of want in (67) lacks the usual meaning of ‘desire’ that want has when used with animate subjects. Instead it has more of a sense of meaning ‘potential.’ Note that such a meaning is often associated with modals (could, might), which are arguably raising verbs (Wurmbrand, 2001). Thus, want in this case could be said to have a comparatively semantically “bleached” meaning. As an aside, I think it plausible that the acceptability of want with a weather-it subject is enhanced by being embedded under seem; ??It wants to rain strikes me as less acceptable than (67). If this intuition is correct, I suspect the reason is that the appearance-denoting matrix predicate allows the embedded proposition more flexibility in terms of its own denotation. Note that since the verbs seem and appear allow both true and false appearance interpretations, one can easily embed absurd or patently false propositions under them and maintain plausibility of the entire proposition ((While John was tripping on acid) It seemed (to him) like the sky was filled with floating elephants). (I don’t mean to imply that people who find (67) acceptable are hallucinating; rather, my point is simply that embedding a predicate under seem could enhance its mutability.) On the other hand, the ability of control verbs to take on this type

Animacy and thematic roles in opaque constructions

99

of bleached or auxiliary-like meaning, even when embedded under a verb of appearance, seems to be limited. Note that try seems somewhat more marginal in the same context, and claim, decide, and forget are outright impossible. (68)

a. ?? It seems like it’s trying to rain. b. # It seems like it’s claiming to rain. c. # It seems like it decided/forgot to rain.

The control verb try appears to be quite natural in other contexts with an inanimate subject, provided the inanimate subject bears certain P-Agent properties in Dowty’s sense. For example, consider a situation in which one’s car won’t start. One can easily utter (69), but neither of the sentences in (70). (69)

The engine didn’t even try to turn over.

(70)

a. # The engine didn’t even claim to turn over. b. # The engine didn’t even remember to turn over.

The relative acceptability of these control verbs with inanimate or weatherit subjects appears to support a continuum, with verbs like claim, decide, and remember/forget being more canonically control, try being less so, and want even less canonically control. Relatedly, although English want is standardly categorized as a control verb, it also has certain unusual properties. Its controllike behavior is seen in its inability to take expletive subjects (modulo the discussion about (67) above) or to exhibit the scopal ambiguity found in raising constructions. (71) (72)

* There wants to be a solution to this problem. Someone from New York wants to win the lottery. a. = There is someone from New York who wants to win the lottery. b.  = It is wanted that the person who wins the lottery (whoever that is) is from New York.

On the other hand, it deviates from more canonical control or transitive verbs in some respects. Although it can function as a transitive verb it is awkward under passivization, even as near synonyms easily permit passivization. (73)

a.

We want that outcome.

b. * That outcome was wanted. c.

That outcome was desired.

100

Argument hierarchies

In many Romance and Germanic languages the verb meaning ‘want’ is a modal, and modals are often argued to be raising verbs – Wurmbrand (2001) specifically argues that all modals are raising verbs, though she gives a special treatment of the German modal wollen ‘want.’ Even in English it exhibits certain modal-like properties such as taking a bare verb complement in wanna contraction (I wanna go) (Postal and Pullum, 1978; Pullum, 1997). Indonesian also offers an example of this kind of variable behavior, as the verbs in this language meaning ‘want’ can function as raising verbs. Polinsky and Potsdam (2008) demonstrate that the Indonesian verbs mau and ingin ‘want’ take a clausal complement and allow an unusual interpretation when that complement clause is passivized. In contrast to English, both (a) and (b) are possible interpretations of (74) (from Polinsky and Potsdam (2008, p. 1618); the same type of reading is permitted with ‘want’ verbs in other Austronesian languages). (74)

anak itu mau/ingin di-cium oleh ibu child that want pass-kiss by mother a. “The child wants to be kissed by the mother” b. “The mother wants to kiss the child”

Polinsky and Potsdam argue that the interpretation in (74b) comes about through syntactic raising, rather than through the operation of backwards control (meaning that the controller is c-commanded by the controllee, attested in, for example, Malagasy (Potsdam, 2006)) or clause union (‘want’ and the lower verb occur together in a single clause, e.g. Aissen and Perlmutter (1983)). The raising analysis appears problematic from the view that a verb meaning ‘want’ selects an external experiencer argument, i.e. it needs to have a “wanter.” Polinsky and Potsdam’s solution, briefly, is that the Indonesian ‘want’ verbs are semantically similar to the so-called “subject-oriented adverbs” (deliberately, intentionally), and thus semantically require an animate (in this case, experiencer) argument but do not select this argument syntactically. For our purposes, what is important to note is that, consistent with the construction in (74) involving A-movement, the derived subject may be inanimate (75). And consistent with mau/ingin having a semantic requirement for an experiencer, (76b) is ill-formed. (75)

rumah itu mau/ingin di-hancurkan oleh mereka house that want pass-destroy by 3p “They want to destroy that house”

Animacy and thematic roles in opaque constructions (76)

a.

101

kota ini di-hancurkan oleh api town this pass-destroy by fire “This town was destroyed by fire”

b. # kota ini mau/ingin di-hancurkan oleh api town this want pass-destroy by fire “#Fire wants to destroy this town”

In addition to control verbs forming a continuum from “more control-like” to “more raising-like,” we will now see that a symmetrical continuum is appropriate for raising verbs. Sentence (67) is an instance of “copy” raising (Rogers, 1974; Potsdam and Runner, 2001; Asudeh, 2002; Asudeh and Toivonen, 2012, among many others). In a copy-raising construction, a pronominal “copy” of the matrix subject (in (67) it is weather-it) appears in the finite lower clause, typically in the subject position of the lower clause. The copied pronominal subject can also be referential, as in (77). (77)

Richard seems like he is in trouble.

Sentences like (77) have the same truth conditions as the standard raising sentences like Richard seems to be in trouble, as well as the unraised version It seems (like/that) Richard is in trouble.15 The challenge that the copy-raising construction poses is this: if Richard is raised from the embedded clause (a) what motivated that movement, since it moved out of a finite clause and therefore out of a Case position, and (b) why is its trace expressed as an overt pronoun? One way to solve this conundrum is to argue that in these cases, seem actually selects its own (thematic) subject and therefore does not involve raising at all. For example, in (78a–b) seem appears to be able to assign a θ -role to PRO, and seem has a meaning along the lines of “give the impression of” (or, “intentionally seem”). (These examples are from Potsdam and Runner (2001), but note that Potsdam and Runner do not in fact argue that seem is a thematic verb in copy-raising constructions.) (78)

a. The workersi want [PROi to at least seem like they are busy.] b. It is important [PROarb to seem like you want the job.]

On the other hand, seem can clearly be non-thematic in a copy-raising construction, as in (79). Here the copy raising seem is able to host an expletive 15 This is not to say that the copy raised and non-copy raised sentences are semantically iden-

tical. See Rett et al. (2013) for discussion about asymmetries in the evidential status of these constructions.

102

Argument hierarchies

subject or part of an idiom in its subject position ((79b) is due to Horn (1981)).16 (79)

a. % There seem like there are problems. b. It seems like it’s raining harder than it is. c. The shoe looks like it’s on the other foot.

To reconcile these apparently conflicting properties, some researchers have suggested that seem is ambiguous between being thematic (assigning a θ -role to its subject) and non-thematic. If seem is lexically ambiguous between being raising and control, we would expect this ambiguity to surface even in more canonical raising constructions, i.e. when it takes an infinitive complement. In fact, this appears to be the case. My own judgment is that the infinitive counterparts of (78) also allow this intentional sort of meaning, though emphasis on seem is needed (cf. The workers want to at least SEEM to be busy; see also Section 2.5). Potsdam and Runner (2001) argue that even though seem need not assign an external θ-role, sentences like (77) involve a base-generated subject that is co-indexed with the lower pronoun. Evidence for base-generation comes from scope effects. Recall from Chapter 2 that raising verbs allow scopal ambiguity not allowed by control verbs, and this is attributed (on derivational accounts) to the fact that a raising verb’s subject is derived and can therefore reconstruct to its underlying position at LF. (80)

Two people seem to have won the lottery. two > seem; seem > two

(81)

Two people seem like they have won the lottery. two > seem; *seem > two

(82)

Two people tried to win the lottery. two > try; *try > two

Thus, seem behaves like a control verb in taking an external argument when it occurs in a copy-raising construction (the same holds for appear). Note, however, that not all raising verbs have this property. For example, tend and happen are completely ungrammatical in a copy-raising type construction. This is probably related to the fact that seem and appear are the only raising verbs that permit a tensed clause complement at all. 16 The symbol % indicates inter-speaker variation in acceptability. In my own dialect (79a) is

well-formed.

Animacy and thematic roles in opaque constructions (83)

103

a. * John tends like he eats fish on Fridays. b. * It tends that John eats fish on Fridays.

So certain raising verbs (notably those whose meanings relate to appearance) appear to be able to function as control verbs, taking a thematic subject. This kind of intermediate behavior is suggestive of a continuum in which some verbs have more canonical raising or control properties while others display hybrid properties, rather than a strict dichotomy. As previewed at the beginning of this section, such a continuum will have repercussions for the learning procedure proposed in Section 3.5. And as we will see in Chapter 5, children’s interpretations of these predicates is in line with their continuous nature: for example, children appear to grant intermediate status to the verb want, allowing it to have a more modal-like, or semantically bleached meaning that is compatible with its functioning as a raising verb. (84)

control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . raising decide try want seem/appear turn out happen/tend

3.3.2 Tough-constructions across languages Cross-linguistic data on tough-movement is harder to come by than data on raising and control, but the construction is attested in several languages besides English. Both Italian and German have tough-movement ((85b) is from Cinque (1995)). (85)

a. Dieses Buch ist schwer/einfach [PRO zu lesen t] this-n/Nom book be-3s hard/easy to read-inf “This book is hard/easy to read” b. Il problema non è facile [PRO da risolvere t subito] the/m problem neg be-3s easy to solve-inf immediately “The problem is not easy to solve immediately”

Wurmbrand (2001) notes that while tough-movement in English permits embedding of additional predicates within the complement of the tough-adjective, German disallows such embedding (from Wurmbrand (2001, p. 29)). (86)

* Dieses Buch ist schwer Hans zu überzeugen zu lesen this book is hard John to convince-inf to read-inf (“This book is hard to convince John to read”)

104

Argument hierarchies

Wurmbrand concludes from this fact that German tough-movement involves A-movement, while English tough-movement involves A -movement (after Chomsky (1977); see Section 2.2 above). There are two questions regarding the typology of tough-constructions that are relevant for the present work. One question is whether the predicates with lexical meanings like ‘easy,’ ‘tough’ etc. select a thematic external argument, or whether they universally fail to select such an argument and take only a clausal complement. Secondly, if such predicates are like the English counterparts in not having an external argument and like English in licensing a gap in the embedded clause (i.e. some kind of movement takes place out of the embedded clause), the question is whether such constructions are universally associated with an object gap in the embedded clause, as opposed to a subject gap. That is, are there languages in which a sentence like John is easy to please can mean that John is the pleaser? While in both German and Italian the interpretation of tough-movement is the same as in English (the surface subject can only be interpreted as the referring to the embedded object – thus, an object gap is licensed but not a subject-gap), it turns out that the subject-gap interpretation does occur in some languages. In Cantonese the tough-construction is ambiguous, as the following example shows (this example is from Anderson (2005, p. 426)). (87)

keoi(5) hou(2) laan(4) ming(4) gaa(3) he very difficult understand (particle) “He is very difficult (for someone) to understand” or “It is very difficult for him to understand something”

Mandarin Chinese also permits both a subject-gap and object-gap reading of a tough-construction, as shown in (88). The subject-gap reading can be disambiguated by either the use of an intransitive predicate in the lower clause (e.g. become angry) or the use of the verb xing ‘perform, do an action,’ as shown in (89). However, when the subject is inanimate some speakers license only the object-gap reading, even if the lower verb is semantically compatible with a subject-gap interpretation, illustrated in (90) (Stephen Crain, personal communication 1/10/2012).17 17 I have elicited conflicting judgments from native speakers regarding the ambiguity of this

construction in Mandarin Chinese. Anderson (2005) reports that the ambiguity is found in both Cantonese and Mandarin, but one native speaker I consulted reported the two meanings to be necessarily disambiguated on the surface by the use of an overt NP in the lower clause (Y. Huang, personal communication 7/20/11). In addition, while some speakers find the subject-gap reading of (90) ungrammatical, others allow it, even if the subject is inanimate.

Animacy and thematic roles in opaque constructions (88)

Zhangsan hen rongyi hui-lu Zhangsan very easy bribe “It is easy for John to commit bribery” or “It is easy to bribe John”

(89)

Zhangsan hen rongyi xing-hui Zhangsan very easy do-bribe “It is easy for John to commit bribery” (*“It is easy to bribe John”)

(90)

Zhe ge qiu hen rongyi gun this CL ball very easy roll “It is easy to roll the ball” (*“It is easy for the ball to roll itself”)

105

Finnish allows both a subject and an object gap in a tough-construction, but it marks the subject differently according to which interpretation is intended. Specifically, the subject is marked with Partitive or Nominative case if it is associated with an object gap, and with Accusative case if it is associated with a subject gap (Anderson (2005, p. 427)).18 (91)

Marja-a on vaikea opettaa Maria-Part is hard-Nom teach-inf “It’s hard to teach Maria”

(92)

Marja-n on vaikea opettaa Maria-Acc is hard-Nom teach-inf “Maria has a hard time teaching”

However, as is expected under our assumptions, when the subject is inanimate only the object-gap reading is licensed (E. Kaiser, personal communication 3/26/11). That is, the subject can be marked with Partitive or Nominative but not Accusative case. (93)

(94)

Kirja-a on vaikea lukea book-Part is hard-Nom read-inf “It’s hard to read the book” * Kirja-n on vaikea lukea book-Acc is hard-Nom read-inf (“The book has a hard time reading”)

However, the subject-gap reading disappears, as would be expected, when the lower predicate denotes an event for which the inanimate NP could not plausibly be the subject, as in The ball is easy to play (with) or The book is easy to read. 18 Anderson glosses the -n case marker as Genitive, but Elsi Kaiser (personal communication, 11/8/12) suggested to me that Accusative is a more accurate gloss in this case. The choice of Nominative vs. Partitive on the subject in the object-gap construction depends on the telicity of the main verb.

106

Argument hierarchies

Although (94) could be ruled out on purely semantic grounds, importantly an inanimate subject cannot license a subject gap even if it is semantically plausible. That is, (96) is ungrammatical – it cannot have an unaccusative type of meaning.19 (95)

(96)

Ikkuna on helppo rikkoa window-Nom is easy break-inf “The window is easy (for someone) to break” * Ikkuna-n on helppo rikkoa window-Acc is easy break-inf (“It’s easy for the window to break (itself)”)

Similar to Finnish, Seiter (1980) reports that the verbs uka ‘difficult’ and mukamuka ‘easy’ in Niuean exhibit raising and allow either a subject gap or an object gap. However, since the experiencer argument appears to be obligatorily overt, there is no ambiguity as to whether the raised NP is the underlying subject or object of the complement clause. Example (97a) shows the unraised version; (97b) shows a subject gap (the subject of the embedded clause is raised; literally “We are hard to read that story”); and (97c) shows the object gap version.20 Like other Polynesian languages, Niuean word order is VSO. (97)

a. Uka ke totou e mautolu e tala ia difficult subj read Erg we/p/ex Abs story that “It’s hard for us to read that story” b. Uka a mautolu ke totou e tala ia difficult Abs we/p/ex subj read Abs story that “It’s hard for us to read that story” c. Uka e tala ia ke totou e mautolu difficult Abs story that subj read Erg we/p/ex “That story is hard for us to read”

Chung (1976) presents data from Bahasa Indonesian (Austronesian), showing that tough-adjectives take a clausal complement and allow extraction from 19 The verb rikkoa ‘break’ can be made morphologically unaccusative by adding an inchoative

affix (yielding rikkoontua), but using this form of the verb in a TC is ungrammatical (E. Kaiser, personal communication 11/9/12). 20 Seiter’s analysis of these constructions in Niuean posits movement akin to raising-to-subject (presumably A-movement), as distinct from English-style tough-movement (standardly A movement). The details of the analysis are not relevant here; crucial is only the fact that Niuean, like other languages we have seen, allows raising of an inanimate NP to a derived subject position with the verbs meaning ‘hard’ and ‘easy.’

Animacy and thematic roles in opaque constructions

107

that complement, but there are strict limits on what can be extracted. Interestingly, neither underlying objects nor underlying subjects can undergo this extraction, but derived subjects of the embedded clause can. (98a) shows the basic form of this construction, without extraction; (98b) shows ungrammatical subject extraction; (98c) shows ungrammatical object extraction; and (98d) shows grammatical extraction of a derived (passive) subject. (98)

a.

Sulit bagi kami untuk mem-perbaiki mobil ini hard for us for tr-repair car this “It’s hard for us to repair this car” b. * Kami sulit untuk mem-perbaiki mobil ini we hard for tr-repair car this (“We are hard to repair this car”) c. * Mobil ini sulit (bagi kami) untuk (mem)-perbaiki car this hard for us for tr-repair (“This car is hard (for us) to repair”) d. Mobil ini sulit untuk di-perbaiki (oleh) kami car this hard for pass-repair by us “This car is hard to be repaired by us”

Thus, it is the semantic object that is raised, similar to English, with the restriction that this argument must have already undergone raising-to-subject within its embedded clause before it is available for tough-movement. Chung lists, in addition to sulit ‘hard,’ the adjectives mudah ‘easy’ and menjenangkan ‘nice’ as adjectives that occur in this particular construction. Baker (1988) provides another example of a tough-adjective having the same sort of argument structure we have seen in the preceding examples. In Labrador Inuttut the predicate meaning ‘easy’ takes a clausal complement and no external argument, and it permits incorporation of the verb from the lower clause (p. 157). (99)

Angutik muuta-mik siqumi-tsi-sagai-juk. man-Abs boat-instr break-apass-easy-3ssub “It was easy for the man to break the boat”

While this example does not provide evidence for NP-movement of the underlying object, it does demonstrate that the argument structure of Labrador Inuttut ‘easy’ parallels that of English easy. Table 3.3 summarizes the cross-linguistic properties of tough-adjectives. I have encountered one language that may lack the sorts of constraints on tough-constructions we have seen so far: the language Lisu (Lolo-Burmese) apparently allows any NP to be the subject of a tough-predicate and this

108

Argument hierarchies

Table 3.3: Properties of tough-adjectives across languages Language

Example tough-adjectives

External argument?

Object or subject gap?

German

schwer ‘hard,’ einfach ‘easy’ difficile ‘hard,’ facile ‘easy’ laan ‘difficult’ rongyi ‘easy’ vaikea ‘hard,’ helppo ‘easy’ uka ‘difficult,’ mukamuka ‘easy’ sulit ‘hard,’ mudah ‘easy’ menjenangkan ‘nice’ sagai ‘easy’

no

object

no

object

no no no

both both* both**

no

both

no

derived subject

no

?

Italian Cantonese Mandarin Finnish Niuean B. Indonesian

L. Inuttut

*Speaker variation; **Subject/Object gap disambiguated by case marking

subject can be linked with either the agent or patient role of the embedded verb (examples from Li and Thompson (1976, p. 474)).21 (100)

a. ánà nya làma dzà hí-a” dog top tiger eat difficult-decl “As for dogs, they are difficult for tigers to eat/tigers are difficult for them to eat” b. làma nya ánà dzà hí-a” tiger top dog eat difficult-decl “As for tigers, they are difficult for dogs to eat/dogs are difficult for them to eat”

Li and Thompson (1976) argue, however, that Lisu does not have subjects at all; rather, they argue that the leftmost NP in this language is a topic, which according to them is not derived. Without a fuller set of data and a more detailed syntactic analysis of this language, it is hard to draw firmer conclusions about whether this indeed presents a counter-example to 21 The diacritic ” is used in the grammar of Lisu by Hope (1974) to indicate laryngealization.

This can take the form of either glottalization or tenseness, depending on the tonal features of the vowel.

Properties of derived subjects

109

the general pattern (we will return to the distinction between subjects and topics in Chapter 7). 3.4

Properties of derived subjects

The Animacy Hierarchy places human-denoting NPs above animal NPs and animal NPs above inanimate NPs, with more animate NPs being more likely to be subjects. The Thematic Hierarchy places agents, with their canonically human properties of volition, sentience, and/or causation, at the top of that hierarchy and again as likely candidates for subjecthood. About inanimate NPs not much is said, except that they are either rarely or never agents, and in some languages they may not occur as the subject of a transitive verb (Jacaltec, Japanese, Tlapanec). In the previous section we saw that across diverse languages raising and tough-predicates permit inanimate subjects while their control predicate counterparts usually do not. Here we will examine more general evidence supporting the potentially important role that inanimate NPs can play in indicating a displaced (derived) subject. The main argument involves a distinction between the properties of “basic” versus “derived” subjects as detailed by Keenan (1976). Namely, derived subjects typically possess (some of) the structural properties of external arguments while retaining certain properties generally associated with internal arguments (including inanimacy). Keenan (1976) defines a “basic subject” as an NP having some subset of a cluster of properties relating to the position of the NP in the sentence, case marking, syntactic properties concerning control of another NP’s reference (i.e. reflexives, PRO), and semantic properties such as agency and autonomous reference. (The criterion of autonomous reference is parallel to Dowty’s “independent existence” entailment: in a sentence like The worker built a house, the house does not exist separately from the building event, but the worker does.) Non-basic, or derived subjects, are typically “less subject-like,” that is, they usually have fewer of these subject properties. Keenan gives an example from passive constructions. The subject of a passive is not an agent, nor does it exist independently of the verb’s event (cf. A house was built by John); it also does not control the reference of a reflexive (*John was insulted by himself). A derived subject may possess few or many of the properties of basic subjects, but it does so according to a hierarchy which Keenan dubs the Promotion to Subject Hierarchy (Keenan, 1976, p. 324).

110

Argument hierarchies

(101) Promotion to Subject Hierarchy Coding Properties > Syn. Behavior and Control Properties position > case markcontrol of reference, ing > verb agreement etc.

>

Semantic Properties agency, autonomous existence selectional restrictions

Thus, a derived subject may have the “Coding Properties” of basic subjects (e.g. English passive subjects occupy the same syntactic position as active subjects and bear Nominative case – He was hit by him), but may lack the properties lower on the hierarchy. On the other hand, if a derived subject has the “Semantic Properties” of basic subjects, then it will have the higher properties as well. Semantic properties, such as agency, are lowly ranked, and Keenan notes that “we know of no clear cases in which derived subjects become e.g. agents” (op. cit.). In other words, it is quite easy for a derived subject to lack the canonical semantic properties of subjects (like agency), but much harder for a derived subject to lack canonical subject properties relating to position or case marking. Craig’s (1976) study of Jacaltec provides concrete support for Keenan’s features of basic and derived subjects, and for his Promotion Hierarchy. As noted above, Jacaltec transitive verbs permit only animate subjects; inanimate subjects are ungrammatical, though they may occur with intransitive verbs. Craig notes that two types of derived subjects (clefted instrumentals and indirect causative agents) violate the ban on inanimate subjects. Other derived subjects (passive and “promoted” subjects) do not – these subjects must be animate.22 Since Jacaltec is a VSO language basic subjects are positioned immediately after the main verb. Transitive subjects trigger an Ergative case marker which appears as an agreement marker on the verb (seen in (102a)). The two derived subjects that take on only the highly ranked subject properties (i.e. Keenan’s “coding” properties) are instrumental subjects and indirect causative agents. Both of these derived subjects violate the typical semantic properties of subjects in that they permit inanimate subjects. Craig shows that 22 Curiously, although Craig (1976) claims that passives in Jacaltec disallow inanimate subjects,

in her 1977 book, The Structure of Jacaltec, Craig gives an example of a passive sentence with an inanimate subject: (i)

x-∅-tz’ah-ot/lax/lo/cha te’ nah y-u naj asp-Abs3-tr-pass cl./the house Erg3-by cl./him “the house was painted by him”

I am not sure what to make of this apparent contradiction.

Properties of derived subjects

111

in a basic transitive sentence, an instrumental NP must occur in an oblique phrase, marked by the preposition Erg-u ‘with’ (102a). However, instrumental NPs can function as a subject in a cleft construction, as in (102b), where the agent becomes oblique. The clefted subject NP exhibits deletion of the Ergative case marker (indicated by e within the main predicate), and the verb bears the added suffix (-n(i)). Sentence (102c) shows that the inanimate instrument cannot function as the subject of a basic transitive sentence. (102)

a.

x-∅-in-tzoc’i-c’oj te’ te’ y-u ch’en machit an asp-Abs3-Erg1-cut-dir cl. tree Erg3-with cl. machete 1p.part “I cut the tree with the machete” b. ha’ ch’en machit ti’ x-e-tzoc’-ni-c’oj te’ te’ w-u cleft cl. machete this asp-∅erg-cut-suf-dir cl. tree Erg1-by “It is this machete that cut the tree by me” c. * x-∅-tzoc’i-c’oj ch’en machit te’ te’ asp-Abs3-cut-dir cl. machete cl. tree (“The machete cut the tree”)

Note that the instrumental subject does not take on the canonical (postverbal) subject position; however, it does retain the “coding” properties of Ergative marker deletion and verb suffixation (-n(i)) associated with subject movement operations. Indirect causative agents bear some similarities to instrumental subjects. For example they may be inanimate, as we see in the following example. (103)

a.

cake’ x-e-’a’-ni tajoj xil kape wind asp-∅erg-make-suf dry cl. clothes “The clothes dried because of the wind; the wind dried the clothes” b. * x-∅-s-taj-Ne cake’ xil kape asp-Abs3-Erg3-dry-tr.suf wind cl. clothes (“The wind dried the clothes”)

Both instrumentals and indirect causative agents can be derived subjects, and both can be inanimate, thus violating the ban on inanimate subjects of transitives. Sentences (102b) and (103a) do not appear to be opaque in the same way that English raising or tough-movement constructions are. The verbs are simple transitive verbs, and although the inanimate subjects retain some of the “coding” properties of typical subjects, they do not retain all of them. Notably, they do not occur in the canonical subject position (immediately postverbal). Nevertheless, these examples illustrate the fact that even in a language that generally bans inanimate subjects of transitive verbs, in certain derived contexts the subject may in fact be inanimate.

112

Argument hierarchies

Craig discusses two other types of derived subjects, subjects of passives and “promoted” subjects. By “promoted” subjects, Craig refers to constructions such as (104), where the subject is promoted over the aspectual verb ’ichi ‘begin.’23 (104)

xc-ach ’ichi ha-munlayi asp-Abs2 begin Erg2-work “You began to work” (lit. ‘you began you to work’)

Both passive subjects and promoted subjects must be animate (though see the caveat about passives in footnote 22), and they possess all of the “coding” and syntactic behavior properties of basic intransitive subjects (e.g. Absolutive case marking, control of reflexives, etc.). It is interesting that the verb meaning ‘begin’ has more control-like than raising-like properties in Jacaltec.24 In Samoan, Niuean, and Chamorro the verbs meaning ‘begin’ are raising verbs (see Table 3.2), and begin is ambiguous in English (see Section 2.5). I find this variation particularly interesting in view of the language learning problem. This type of variation means that even if children had unequivocal access to these predicates’ lexical meanings in advance of knowing the syntactic structure they occurred in (which is highly unlikely given the abstractness of the predicates; Gleitman et al. (2005)), it still would not unambiguously indicate what kind of structure the verb occurred in: that is, it would not determine whether the verb selected an external argument or not. This further underscores the need for children to figure out the sentence’s structure independently of the predicates’ lexical meanings, and, thus, the need for some cue that can help 23 Craig notes explicitly that the “promotion” operation is not a case of “raising,” but rather

“copying,” perhaps akin to English copy raising. As discussed in Section 3.3.1 copy-raising constructions may involve a thematic subject in the main clause. 24 The verb x’ichi ‘begin’ in Jacaltec can occur with inanimate subjects but only when it is used intransitively, as in (i). When it has any complement, whether clausal (as in (104)) or nominal (as in (iii)) its subject must be animate. i.

x’ichi caNal began dance “The dance began”

ii.

x’ichi naj began cl./he “He began”

iii.

xiche naj munil began cl./he work “He began the work”

Properties of derived subjects

113

them figure out the structure, like subject animacy. Animacy, in fact, lines up with the syntactic argument-taking properties of these verbs both within and across languages: languages in which ‘begin’ functions as a raising verb permit it to occur with inanimate subjects, and languages in which ‘begin’ functions as a control verb require that it occur with animate subjects. An additional property of internal arguments that derived subjects typically retain is their ability to undergo noun incorporation in incorporating languages (Tomlin, 1986; Baker, 1988). Baker notes that while objects of transitive verbs freely undergo noun incorporation, subjects generally do not. The examples in (105) from Mohawk illustrate this asymmetry (from Postal (1962), cited in Baker (1988, p. 81); see also extensive examples in Tomlin (1986)). (105)

a.

Yao-wir-aija ye-nuhweij-s ne ka-nuhs-aij pre-baby-suf 3f.sub/3n-like-asp the pre-house-suf “The baby likes the house”

b.

Yao-wir-aija ye-nuhs-nuhweij-s pre-baby-suf 3f.sub/3n-house-like-asp “The baby house-likes”

c. * Ye-wir-nuhweij-s ne ka-nuhs-aij 3f.sub/3n-baby-like pre-house-suf (“Baby-likes the house”)

While some intransitive subjects do allow incorporation, it appears to be limited to non-agentive arguments, i.e. arguments of unaccusative verbs. (106) shows an example of incorporation of a theme subject in Onondaga, and (107) shows an example of ungrammatical incorporation of an agentive intransitive subject in the same language (both from Baker (1988, pp. 87–89)). (106)

a. Ka-hi-hw-i neij o-hsaheijt-aij 3n-spill-caus-asp the pre-bean-suf “The beans spilled” b. Ka-hsaheijt-ahi-hw-i 3n-bean-spill-caus-asp “The beans spilled”

(107)

a.

H-ate-ijse:-ij neij o-tsiijkt-aij 3ms-refl-drag-asp the pre-louse-suf “The louse crawls”

b. * H-ate-tsiijkti-ijse:-ij 3ms-refl-louse-drag-asp (“The louse crawls”)

114

Argument hierarchies

Table 3.4: Properties of “basic” and derived subjects across languages Basic subject property

Coding properties position (highest NP)

Found in derived subjects?

Languages

yes

English, all languages in Sec. 3.3.1, 3.3.2 Jacaltec English, all (case-marking) languages in Sec. 3.3.1, 3.3.2, Jacaltec English, all (agreement-marking) languages in Sec. 3.3.1, 3.3.2, Jacaltec

case marking

no yes

verb agreement

yes

Semantic properties agency Derived Subject Property noun incorporation

no

all languages

Found in basic subjects? no

Language Mohawk, Onondaga

Thus, a further clue that a subject is derived is its ability to undergo noun incorporation. Interestingly, the example given here in (106) involves an inanimate NP. A summary of the cross-linguistic patterns discussed in this section is found in Table 3.4. 3.4.1 Argument structure universals, and the “problem” of ergativity If the use of an inanimate subject to signal a derived subject is to be used as an acquisition strategy by children learning language, we expect the strategy to be available universally. The typological literature reviewed above provides evidence that, cross-linguistically, non-derived subjects tend to be high on the Animacy and Thematic Hierarchies, and derived subjects tend to be lower on both hierarchies (Keenan, 1976). Thus, lack of animacy on the subject should work cross-lingusitically as a general cue, with some variation to be expected between languages and between constructions within a language (e.g. depending on particular verbs). However, if children innately expect a non-derived (i.e. logical) subject to be more (or at least as) animate and/or agent-like than

Properties of derived subjects

115

the (logical) object, and, presumably, than a derived subject, this raises the question of how this strategy could work in syntactically ergative languages. We have already mentioned displacement phenomena in languages that are morphologically ergative, for example Jacaltec. To say that a language is morphologically ergative means that subjects of intransitive verbs and objects of transitive verbs receive the same morphological case (Absolutive), but subjects of transitive verbs receive a different case (Ergative). But the argument structure as it maps onto the syntax is (potentially) the same as in a Nominative/Accusative language: the transitive and intransitive subjects are still external arguments, and the transitive object is still an internal argument. However, Marantz (1984) claims that Dyirbal (Australian) and Arctic Eskimo are syntactically ergative languages (Marantz uses the term “true ergative”) in that they assign the agent role (and Ergative case) to the syntactic object, not the syntactic subject, and the subject bears the patient/theme role (and Absolutive case). Before proceeding, I should note that in work on ergativity the symbol O is standardly used to mark the transitive object, A is used to mark the transitive subject, and S is used to indicate the sole argument of an intransitive verb (Dixon, 1979). Marantz’s analysis is that syntactically ergative languages have an underlying structure like that in (108), where the meaning is “Mary hugged John” (example from Pye (1990)). This structure contrasts with an accusative syntactic structure, shown in (109).

(108)

Ergative

S NPabs

VP

(109)

NPerg

V

Mary A

hugged

Accusative

John O

S NPnom Mary A

VP V

NPacc

hugged

John O

116

Argument hierarchies

While morphologically ergative languages are not problematic for the story presented here, languages with an underlying structure like that in (108) would be problematic, not only for my hypothesis, but also more generally for the view that children have certain fundamental expectations about language structure, and that these expectations help guide language learning (such as Semantic Bootstrapping, Pinker (1984); see Section 5.2.1). Pye (1990) notes that if children need to figure out on the basis of language input whether their language places agents or patients as the direct objects of verbs, it is not clear what evidence could be used to make this determination. The system of case marking is not completely informative, since many (in fact, most) morphologically ergative languages have an accusative-type syntax (i.e. direct object is patient; Dixon, 1979), while Dyirbal, which is reported to be syntactically ergative, in fact has a split morphological system, as noted above in Section 3.1.1. Word order is also problematic as a cue to syntactic ergativity. One can imagine a strategy by which the NP closest to the verb is represented as the object, so that the order AOV or VOA (i.e. Agent-Object-Verb/Verb-Object-Agent) would signal an accusative syntax, and OAV (or VAO) would signal ergative syntax. But there are in fact VAO languages which have accusative syntax (e.g. Welsh (Sproat, 1983), Jacaltec (Craig, 1977) and several Polynesian and Austronesian languages (Chung, 1978)). Moreover, verb-medial orders would then be uninformative, since an AVO string (one of the most common word orders across languages (Tomlin, 1986; Comrie, 1989)) could correspond to either an A[VO] or an [AV]O structure. Pye (1990) claims that K’iche’, a Mayan language with ergative morphology but accusative syntactic structure, has a particular construction that exhibits properties of syntactic ergativity. In wh-questions the O or S argument can be questioned directly, but the A argument cannot be; in order to question the A argument it must be promoted to an S argument and the verb is marked with antipassive focus (abbreviated FA in the examples below). Thus, O and S function together (as subjects) to the exclusion of A (an object); and A in fact becomes an S when questioned. An example follows (from Pye (1990), his examples (14) and (15)). (110)

jas k-∅-u:-b’an le: patax what incompl-3Abs-3Erg-do the duck “What is the duck doing?”

(111)

jachi:n x-∅-ya’-ow le: su’t chi-aw-e:ch who compl-3Abs-give-FA the cloth to-2Erg-poss “Who gave the cloth to you?”

Properties of derived subjects

117

In (110) the object of do is questioned directly, whereas in (111) the agent is marked with Absolutive (via agreement on the verb), and the verb bears the focus antipassive (FA) suffix. According to Pye, this grouping of O and S resembles a syntactically ergative pattern since both O and S arguments are “more accessible” to wh-questioning (and therefore more subject-like, following the Accessibility Hierarchy of Keenan and Comrie, 1972) than the A argument. In his study of children acquiring K’iche’, Pye found that although the children made very few errors in their production of ergative morphology (fewer than 10% of their case markers were produced incorrectly), they exhibited significant difficulty with the focus antipassive construction used in questions.25 When the children (aged 2 to 3 years) were asked wh-questions about the A argument, as in (111) above, they failed to answer the question (responding, for example, “what?”) or responded with a verb but without the required focus antipassive form. Only after age 3 do the children begin to respond reliably to these questions, and even then they often make errors by using the active voice form of the verb. Based on these data Pye suggests that individual constructions that involve syntactic ergativity are acquired late and only on a construction-specific basis, after the child has pinned down various properties of the target syntax, such as word order and case marking properties.26 As for languages that are supposedly wholly syntactically ergative, Marantz concedes the learning problem implied by such languages, and he suggests that their extreme scarcity (there are only two such languages) underscores their linguistic markedness. If the Animacy and Thematic Hierarchies are to do any work in helping to constrain children’s hypotheses about language structure (this is essentially the point of the Semantic Bootstrapping hypothesis), children must be able to map the most animate/agentive NP onto the subject position (i.e. external argument) and not wonder whether their language has a structure like (108) or (109).

25 The relative ease of acquisition of ergative case marking is also reported in Schieffelin (1979,

1985) for Kaluli. These reports contrast with Ochs’s (1982) finding that ergative case marking is acquired late in child Samoan. However, Ochs attributes the late acquisition in Samoan to variable use of the case markers in the input (contingent on sociolinguistic factors), rather than morphosyntactic difficulties. See also discussion of these facts in Section 5.2.1. 26 Another aspect of this construction to consider, however, is that it is dependent upon the person of the subject and object: if both are 1st person and/or 2nd person, active voice is used; if one is 1st or 2nd person and the other is 3rd person, the verb is marked for 1st/2nd person rather than 3rd (Croft, 2003). Thus, in addition to the ergative-like nature of this construction there may be further complicating factors that delay its acquisition.

118

Argument hierarchies

Marantz conducted his own study with children (Marantz, 1982), in which he taught 4- and 5-year-old English-speaking children novel verbs, some of whose argument structure properties lined up with the canonical argument structure mapping found in English (the subject was an agent and the object was a patient or location) but others were reversed in their role assignment (the subject was a patient or location and the object was an agent). The latter type of verbs have a syntactically ergative argument structure. This experiment will be discussed in more detail in Chapter 5, but here I will point out the following main result. Marantz found that 4-year-olds took longer to learn verbs with a patient subject and agent object than the more canonically mapping verbs, and they produced more “reversals” of these verbs than the more canonical verbs. That is, when presented with an action in which a person was pounding a book with his elbow, described as “The book is puming Larry,” 4-year-olds would often reverse the order of agent and patient in their own spontaneous productions of the verb pume (e.g. “He is puming the book”) in describing an action of pounding a book with one’s elbow. Five-year-olds, on the other hand, took about the same amount of time (measured in number of training exposures) to learn patient-agent verbs as agent-patient verbs, and they produced virtually no reversals. This result appears to support Pye’s claim that syntactically ergative constructions are acquired only later in development. The data from K’iche’ and Marantz’s own data from child English appear to cast doubt on the analysis of Dyirbal and Arctic Eskimo as purely syntactically ergative languages, since they would be expected to be either unlearnable (as argued by Pye) or learned on a very different (i.e. delayed) time-course than all other languages. I am not aware of any studies of the acquisition of these two particular languages.27 Pye’s suggestion is that these languages are not fully syntactically ergative at all – indeed, that no language is. Instead, they contain several constructions that exhibit traits of syntactic ergativity, but underlyingly have an accusative structure. Supporting this kind of view, Baker (1988, p. 428) points out that these languages display patterns of noun and verb incorporation quite similar to those found in all other languages, namely, incorporation of (logical) objects (as patients) but not (logical) subjects (as agents). This pattern suggests that even in these languages objects/patients form a complex with the verb to the exclusion of subjects/agents.

27 Fortescue (1984) describes a detailed study of the acquisition of Greenlandic Eskimo. How-

ever, Marantz argues that, although the two languages are closely related, Greenlandic Eskimo appears to have an accusative structure rather than an ergative one, different from Arctic Eskimo. See Marantz (1984, pp. 215–216).

A learning procedure

119

To summarize, languages that are morphologically ergative (i.e. they have an Ergative-Absolutive case system) present no problem for the learning procedure proposed here, since animacy functions within the same parameters in morphologically ergative languages as in morphologically accusative languages. Languages with a truly ergative syntactic system would potentially present a problem, on the other hand, since my claim is that an inanimate subject should be unusual and therefore signal a derived subject. In order for this to work as a learning cue, the learner must presuppose that subjecthood, animacy, and agency are properties that cluster together, and these properties do not cluster together in a syntactically ergative language. However, syntactic ergativity (if it in fact exists) presents problems for most accepted approaches to language acquisition. Studies of the acquisition of the reportedly syntactically ergative languages should be taken up in the future. 3.5

A learning procedure

In this chapter we have explored cross-linguistic (semantic and syntactic) properties of canonical subjects, displaced subjects, and displacing predicates, and we have seen that across diverse languages the semantic feature of animacy plays an important role in discriminating the constructions and predicates of interest here. In all of the languages we’ve looked at, displacing predicates can occur with inanimate subjects while non-displacing predicates generally cannot. The purpose of this typological study within a book about language acquisition is to shore up support for the idea that children should pay attention to the animacy of NPs, because subject animacy is strongly indicative of predicate selectional properties (and therefore of predicate subclass). The claim is that if children do pay attention to animacy, this will help them figure out how to parse non-canonical syntactic structures that involve the movement of NPs, and to categorize the predicates occurring in these structures. Underlying this claim is the view that children have certain innate expectations about how language works. A detailed discussion of innateness will be taken up in Chapter 7. For now we will assume that patterns shown in this chapter to be highly robust cross-linguistically will form (part of) children’s innate knowledge about language: primarily, canonical subjects are typically animate, and there are predicates that fail to select a subject. With these expectations in place, encountering an inanimate subject can provide evidence of a non-canonical structure. Based on the cross-linguistic generalizations about the semantic properties of displaced and non-displaced subjects, we can now specify a learning

120

Argument hierarchies

procedure children could use to figure out the argument structures of the opaque constructions outlined in Chapter 2. The ingredients for this procedure were sketched at the beginning of this chapter as follows (with section numbers indicating the source of the supporting evidence): 1. animate NPs are distinguished from inanimate NPs (Section 3.1.1); 2. subjects are generally animate, or more animate than other arguments in the clause (Section 3.1.1); 3. NPs bearing the agent or experiencer thematic role are animate and canonically mapped to subject position (Sections 3.2, 3.2.2); 4. displacing predicates do not select a subject → they place no semantic restrictions on the subject → the subject can be inanimate (Section 3.3); 5. displaced subjects generally have the syntactic, but not the semantic, properties of non-displaced (canonical) subjects, such as animacy (Section 3.4) ⇒ an inanimate subject is likely to be displaced A brief synopsis of the main syntactic and semantic features of canonical subjects as opposed to non-subjects is given in Table 3.5. Now, if an NP with the semantic hallmarks of a non-subject (e.g. inanimate, theme role) turns up with the syntactic hallmarks of a subject (e.g. Nominative case, triggering subject-verb agreement), there are two possibilities: either (a) the NP is an inanimate external argument (i.e. it is a true subject but lacks some of the semantic properties of canonical subjects), or (b) the NP is a derived argument, either as an internal argument of the verb or as an argument Table 3.5: Typical properties of subjects and non-subject arguments Subject high(er) on Animacy Hierarchy high(er) on Thematic Hierarchy more P-Agent entailments (= agent/ experiencer) Nominative/Ergative case triggers subject-verb agreement

Non-subject Semantic properties low(er) on Animacy Hierarchy low(er) on Thematic Hierarchy more P-Patient entailments (= theme/ patient) Syntactic properties Accusative/Absolutive case does not trigger subject-verb agreement

A learning procedure

121

of a predicate in another clause. Option (a) is plausible in copular constructions (e.g. The book is heavy; see mention of these constructions in Chapter 2, footnote 1) where there is no verbal predicate. We must also keep this option available for verbs of emission if they are correctly categorized as unergative verbs (see Sections 2.3.1 and 3.1.2). However, in the biclausal constructions in which we find raising verbs and tough-adjectives, an inanimate subject is nearly always indicative of a derived subject. This is because the sorts of predicates that do select their subject in these sentence frames (i.e. control predicates) select an external argument that is animate. So if the subject is instead inanimate, option (b) is more likely. Thus, the general approach can be summarized in the following two tree structures. In the first structure the syntactic subject bears the semantic properties of an external argument (non-derived subject), while a second argument, either within the main VP or within an embedded clause, bears the semantic properties of a non-subject, or internal argument. In the second structure the syntactic subject bears the semantic properties of a non-subject and is shown having raised from a lower position in the structure; again, either within the VP or within an embedded clause. (112)

IP I

NP agent/exper. P-Agent entail. +human/+animal

I

VP V

Pred theme/patient P-Patient entail. -human/-animal

(113)

IP I

NP theme/patient P-Patient entail. -human/-animal

I

VP V

Pred NP

We can formalize this notion in the following way:

122 (114)

Argument hierarchies Derived Subject Condition An inanimate subject is (likely to be) an external (underived/non-displaced) argument if a. there is another NP argument in the same clause that is equal or lower on the Animacy or Thematic Hierarchy (has more P-Patient entailments), or b. the role played by the inanimate subject in the verb’s denoted event is that of Source or Causer.28 Otherwise, the inanimate subject is (likely to be) a derived (displaced) argument.

The upshot of the Derived Subject Condition is that if the subject of the sentence is the least animate NP argument of the main predicate, and/or has the most P-Patient entailments and fewest P-Agent entailments, it is very likely to be derived. In some sentences (e.g. simple intransitive sentences, but also very simple biclausal sentences like The book seems to be heavy) an inanimate subject will vacuously be the least animate NP in the sentence. In other sentences there may be another NP, but if there is it will be marked with a preposition, signalling an adjunct phrase (as in (115a), (116a), and (117)), or it will be a direct argument of another predicate (as in (115b) and (116b)). (115)

a. The book seems to be heavier than the [pebble/mouse/infant.] b. The book seems to [give instructions for basketweaving]/[be a gift from Aunt Gerty.]

(116)

a. The book is easier to read than the magazine. b. The book is easy to do a handstand on.

(117)

The book fell on [the floor]/[me].

Notice that if an experiencer NP is present it would be an argument of the displacing verb/adjective (though still marked with a preposition, or 28 This qualifier is meant to subsume cases like verbs of light/sound emission, which are unerga-

tive but generally take inanimate subjects (The bulb glowed) and theme-experiencer psych verbs (frighten; Belletti and Rizzi’s Class II). These verbs take a (possibly inanimate) theme subject and a (normally animate) experiencer object (The sudden noise frightened the baby). Although this class of psych verbs is argued by Belletti and Rizzi (1988) to be unaccusative, and as such would not present a problem for (114), Landau (2010) argues that for certain types of predicates of this class (eventive ones) the theme argument is external.

A learning procedure

123

Dative case, as in German), but it will always be higher in animacy than the inanimate subject. (118)

a. The book seems to John to be pretty heavy. b. The book is easy for Sarah to read.

Naturally, full passives will have another NP (marked by by, thus an adjunct), but according to Jackendoff’s (1972) Thematic Hierarchy Condition, the by-phrase will always be higher in animacy than the theme subject.29 Thus, passives will fail to satisfy parts (a) and (b) of (114), and the subject is displaced. It is important to keep in mind that the structure of the passive need not be known beforehand; rather, all the learner needs to know is that if the syntactic subject is the least animate of the arguments in the sentence, it is very likely to be displaced. Viewed as a means of parsing the structure of a string, this procedure will work fairly straightforwardly, with exceptions needing to be made only for copular constructions, verbs of emission, and possibly (in languages that allow them) inanimate subjects of transitive verbs (although these should be taken care of by Dowty’s proto-role entailments). Even though there are verbs that are ambiguous between being raising and control verbs (see Section 2.5), in general these verbs behave as raising verbs when their subject is inanimate and as control verbs when their subject is animate. Furthermore, although we saw arguments in Section 3.3.1 for raising and control predicates forming a continuum, rather than two discrete categories, the variable behavior of these predicates aligns almost perfectly with subject animacy: control verbs that fall towards the middle of the continuum can be coerced to have a raising function when the subject is inanimate (or non-referential, as in weather-it), and raising verbs that fall towards the middle of the continuum can be coerced to have a control function when their subject is animate. However, I have also suggested that this is a learning procedure for categorizing predicates, and viewed in this way the procedure must allow for some flexibility: that is, it should be probabilistic rather than deterministic. The very fact that there are predicates that can function either as selecting a subject or not selecting a subject, while other predicates disallow this flexibility, means that predicates occurring with an inanimate subject should be deemed likely 29 Couper-Kuhlen (1979) argues for the same pattern in Prepositional Passives (The bed was

slept in): agentive/volitional NPs and sometimes inanimate causers can be the (implied) agent (by-phrase), but themes, locations, and instruments cannot be.

124

Argument hierarchies

to be displacing predicates, and predicates occurring with an animate subject should be deemed likely to be non-displacing predicates. As a final note, although we are mostly disregarding raising-to-object and object control constructions (see Section 2.1.3) it is interesting that these constructions also pattern with the others with respect to the Derived Subject Condition (if construed more broadly as a condition on displaced arguments). (119)

John expects the booki [t i to be heavier than the pebble/infant.]

In (119) the NP in question is the book. The subject of the sentence is animate (thus higher in animacy than the book), and the other inanimate NP is in a prepositional phrase headed by than. Thus, sentence (119) meets the criteria for a displaced argument. Of course, raising-to-object involves a displaced object rather than a displaced subject. But the relevant comparison here is that object control constructions require not only an animate subject but an animate object. So in the case where that object is inanimate, it is highly likely to be displaced.

3.6

Summary

In this chapter I laid out some of the main ways in which animacy is marked in human language, from having effects on case or other morphological marking to affecting argument structure relations. We also looked at how argument structure relations are captured in terms of thematic relations, and some different ways of formalizing these notions. The remainder of the chapter focused on how displacing predicates work cross-linguistically, focusing primarily on raising-to-subject and tough-constructions. We saw that while there is interesting variation, both in the semantic properties and syntactic behavior of raising and control predicates, there is broad cross-linguistic uniformity in that raising predicates (likewise tough-predicates) select a clausal complement and fail to select an external argument. In languages that permit raising (such as German, Italian, several Polynesian languages, and Chamorro), raising verbs take a derived subject, and that derived subject may be inanimate. Moreover, even languages that generally disallow inanimate subjects in basic transitive sentences, like Jacaltec, nevertheless permit inanimate subjects if they are displaced (in certain constructions, at any rate). Broad typological patterns inform the scope of possible grammars to be attained in principle. With this range of possibilities in mind, we turn now to

Summary

125

psycholinguistic studies, which help inform the “psychological reality” of our mental representations of grammar. By exploiting the wealth of knowledge in both of these domains we can combine the wisdom of Greenberg’s universals with that of Chomsky’s Universal Grammar to give us a more comprehensive picture of the role that something like animacy can play in the acquisition of language.

4 Animacy and adult sentence processing

In Chapter 3 we saw a robust cross-linguistic tendency for non-displaced (underived) sentence subjects to be high on both the Animacy and Thematic Hierarchies, and for displaced (derived) subjects to be lower (than non-displaced subjects) on both scales. We saw this pattern both as a general property of grammar and as a specific property of the displacing predicate constructions, where evidence was available. In the present chapter we will look at experimental evidence showing that manipulating the animacy of the main clause subject can affect how adults interpret the structure of a sentence, as well as their assumptions about the meaning of the main clause verb. The main finding is that adults tend to interpret an animate subject agentively, and an inanimate subject as displaced: they take it to be either a patient of the matrix verb or an argument of an embedded predicate. Studies of adult sentence processing are relevant for understanding child language acquisition for two primary reasons. The first is that adult linguistic knowledge is the target state that children are moving toward in their development. If children exhibit deviations from the pattern of behavior observed in adults (e.g. they make more errors in comprehension or production), these deviations can inform us about which aspects of grammar (or processing) are difficult for children. So differences between children and adults can be informative for understanding the path of development. But similarities between children and adults can be highly informative too, and this is the second reason it is useful to look at adult sentence processing. Adults have a great deal more knowledge about how the world works than young children do. If we find that some experimental manipulation makes one condition harder for adults than another, and children exhibit the same increased difficulty in that condition, we should not chalk up children’s difficulty in the harder condition to lack of knowledge of how the world works (Gleitman et al., 2005). That is, similarities in behavior across children and adults can indicate deep properties of the relevant component of linguistic knowledge; in the present case, animacy. 126

Animacy and adult sentence processing

127

To give an example of early work on psycholinguistic effects of animacy in adults, I will describe an experiment by Clark and Begun (1971), which looked at adult native speakers’ rating of well-formedness of transitive sentences according to the relative animacy of the subject. The study was carried out in several parts. In the first part, forty nouns of each of five semantic types were culled from the Thorndike and Lorge (1944) word list. The five semantic types were Human (e.g. teacher), Animal (spider), Inanimate-count (Ns that can be pluralized; book), Concrete-mass (snow), and Abstract-mass (growth). These 200 nouns were then presented to adult participants (college students) in the subject position of a transitive sentence frame: The – – ed the – . Participants were asked to complete the sentences in a sensible way by filling in the verb and object (Clark and Begun note that the definite determiner was not used for the abstract mass nouns). Once the sentences had been constructed, the sentences were systematically altered by replacing the predicates (verb plus object) originally created for nouns of one semantic type with predicates originally created for nouns of a different semantic type. For example, a predicate created to go with a Human noun (e.g. considered the problem) was reassigned to an Abstract noun (e.g. noise), and a predicate created to go with an Abstract noun (e.g. frightened the baby) was reassigned to a Human or Animal noun (e.g. man, spider). These exchanges created potentially anomalous new sentences, e.g. The noise considered the problem.1 The resulting rearranged sentences were then presented to a new set of adult participants, who were asked to rate each sentence on a scale from 1 (very nonsensical) to 7 (very sensible). Two main generalizations can be made about the rating of these sentences. The first is that sentences with Human subjects were on average rated as more sensible than other sentences (mean 5.29) regardless of the predicate. All other subject types received lower acceptability scores, with Inanimate-count nouns and Concrete-mass nouns receiving the lowest average score (3.37). The second generalization is that Human predicates (i.e. those predicates that were originally created to go with Human nouns in the first stage of the study) were the least flexible in taking different subject types. Sentences with a Human predicate (considered the problem) received an average score of 3.38 (across subject types), and as predicates get “less” human the average scores 1 The resulting sentences would not have been anomalous in all cases, as in [The man] Hum [frightened the baby]Abs , but Clark and Begun describe the resulting sentences as semanti-

cally anomalous. Examples are not given, but Clark and Begun’s description implies that they attempted to make the sentences as anomalous as possible.

128

Animacy and adult sentence processing

get higher, with Abstract-mass predicates receiving the highest average score (4.5). Thus, these predicates (those created to go with Abstract-mass subjects) are the most flexible in allowing different semantic types of subjects. Human predicates paired with a Concrete-mass subject received the very lowest score of any pairing (2.14). The upshot is that human-denoting subjects can occur with a wide range of predicates, but predicates that need a human-denoting subject are very choosy about their subjects (this asymmetry in the “choosiness” of predicates will be a key component of the computational models introduced in Chapter 6). The next phase of Clark and Begun’s study was the following: a third group of adult participants were given the anomalous sentences and were asked to “improve” them by changing one word (other than the) in the sentence. Clark and Begun report the percentage of the time the subject was changed (as opposed to the verb or object), and they found that subject alterations were, on average, most frequent for Human predicates and least frequent for Abstract predicates. Furthermore, the direction of change was generally towards a Human subject. That is, when subjects were replaced, they were replaced with a noun higher on the animacy hierarchy. The resulting sentences had a much higher proportion of Human subjects overall than any other semantic type (38% compared to 20% each for Animal and Inanimate-count nouns, 12% for Concrete-mass nouns, and 10% for Abstract-mass nouns). This experiment provides an example of how the preference for subjects to be animate found in cross-linguistic typological patterns (see Section 3.1) is also “psychologically real.” It is possible that the cross-linguistic association between animacy and subjecthood in fact results from speakers’ psychological preferences, though it is also possible that the direction of influence goes the other way: people come to associate animacy with canonical subjects because this is what they encounter in language. Nevertheless, what is clear is that there are measurable psycholinguistic effects of this linguistic preference. In the following sections we will review other types of psycholinguistic experiments with adults, experiments which employ different methodologies and target different syntactic constructions but show the same relationship between subject animacy and speakers’ intuitions about sentence structure. That is, animacy effects are pervasive in adult speakers’ processing of language. We will finish the chapter by discussing a psycholinguistic study with adults that directly investigated the effect of subject animacy on speakers’ intuitions about novel raising and control verbs, thus showing that animacy effects extend not only to adults’ interpretations of known words, but also to their learning of

Relative clauses

129

novel forms. This influence will have repercussions for the next chapter, which focuses on child language learners. 4.1

Relative clauses

4.1.1 Reduced relative clauses The influence of NP animacy on sentence parsing and interpretation has played a significant role in debates over the underlying mechanism of sentence processing. As a sentence unfolds, the sentence perceiver (reader or listener; henceforth “reader”) must construct the correct structure for the sentence or else the sentence is misparsed.2 In parsing the sentence the reader takes into account lexical information from the individual words as they are encountered, structural information about how the words can or must be put together, as well as discourse information and knowledge of how things work in the world. Animacy is an interesting feature of the sentence parsing process since it relates to a number of points along the trajectory of sentence building: it relates to real-world knowledge (e.g. the knowledge that rocks are inanimate but people are animate), and as we saw in Chapter 3 there can be syntactic effects of NP animacy: in some languages animacy relates to special case marking or other morphosyntactic reflexes. However, in the psycholinguistic literature animacy is generally viewed as being primarily (lexico-)semantic in nature (for example, rock has a [−animate] feature while person has a [+animate] feature). It is this semantic aspect of animacy that has made it relevant in identifying the precise point in sentence processing at which semantic information is used in the construction of the sentence parse, and in determining the way in which lexical semantics influences the on-line representation of syntactic structure. Researchers agree that animacy plays a role in sentence processing. Because the selectional relationship between a predicate and its arguments is partly semantic in nature (see Section 3.2), a subject’s (+/−) animacy feature can be predictive of the thematic properties of the predicate that follows. But there is disagreement about the way in which animacy influences sentence processing. One historical difference is that proponents of the Garden Path model (Frazier, 1978; Frazier and Fodor, 1978; Frazier and Rayner, 1982; Rayner et al., 1983; Ferreira and Clifton, 1986; Clifton, 1993) claim that lexical semantic 2 Most studies employ experimental methods that measure processing of written, rather than spo-

ken language. Therefore, a sentence “perceiver” in these cases is really a reader, not a listener. The assumption is that similar mechanisms are at work in reading and hearing language.

130

Animacy and adult sentence processing

information like animacy is available only after the basic sentence structure has been built; it is said to affect interpretation (eventually) but not to guide the initial construction of a parse (Ferreira and Clifton, 1986). Proponents of an interactive, constraint-based system, instead, argue that both syntactic and semantic information are available simultaneously (Tyler and Marslen-Wilson, 1977; Trueswell et al., 1994; Trueswell and Tanenhaus, 1994). Thus, on this view semantic information about lexical items, or contextual information from the discourse, is available during the course of a reader’s initial construction of the syntactic parse. Therefore, semantic or contextual information has the potential to guide the building of syntactic structure along the way. So while these two approaches differ in the explanation of exactly how and when information like animacy affects sentence parsing, both approaches agree that subject animacy affects sentence interpretation. To see how such an effect is assessed, let’s look at a fairly well-established example. In many sentences there are points prior to the end of the sentence that are compatible with various different types of continuations, where the different continuations vary in the underlying sentence structure, not just choice of lexical items. These are called points of local or temporary ambiguity. For example, in (1) the subject could be simply the (agent) subject of a main clause, but it could also be the head of a relative clause, as the two different continuations show. (1)

The defendant examined . . . a. . . . the evidence. b. . . . by the lawyer turned out to be unreliable.

Several factors conspire to create the local ambiguity at the verb examined: (a) the likelihood of the subject to be an agent or a theme (defendant is a plausible agent, though it could also be a theme), and (b) the ambiguity of the verb (examined) between being a simple past form (therefore the main verb of a transitive clause) and a passive participle (therefore being part of a reduced relative clause). Manipulating these factors can affect a person’s likelihood to choose one parse over the other. For example, if the subject in (1) is replaced with an inanimate NP, now the continuation with the reduced relative clause becomes the more natural of the two. (2)

The evidence examined . . . a. #. . . the NP. b.

. . . by the lawyer turned out to be unreliable.

Relative clauses

131

Ferreira and Clifton (1986) recorded participants’ eye movements while reading sentences displayed on a computer screen. Each sentence of interest contained a relative clause, either reduced (as in (1b)/(2b)) or full (The defendant that was examined . . . ), with an animate or inanimate subject. A longer reading time at the point of disambiguation (e.g. compared to the same point in the sentence with no ambiguity) would indicate a garden path effect, i.e. a revision of the original parse. The Garden Path model predicts that since the verb examined could be a simple past verb (unlike, for example, shown), and since a transitive matrix clause involves fewer branching nodes than a sentence with a reduced relative clause (its syntax is in some sense less complex), readers should begin constructing a simple transitive matrix clause at the point of reading examined. The fact that a subject like defendant is a possible agent does not come into play. Then, when readers encounter the by-phrase they will have to revise their parse to make it fit with the required reduced relative structure, and this will result in a slowing down in reading time. Thus, readers should slow down at the point of by in the reduced relative compared to the full relative version. Crucially, the Garden Path model makes the further prediction that since syntactic structure is computed independently of semantic/thematic information, the choice of the transitive matrix clause parse (following Minimal Attachment; Frazier, 1978; Rayner et al., 1983) will be made in the reduced relative version regardless of the animacy of the subject; so, in (2) just like in (1). Even though readers will recognize that evidence is not a possible agent, they will nevertheless construct a transitive matrix clause parse initially upon reading The evidence examined . . . , and will show equal slowing down in reading time at the word by, as compared to the version with an animate subject. On the other hand, if semantic information is taken into account from the beginning, informing the very construction of the sentence structure, then no slow-down is expected at the word by when the subject is inanimate. Reading times for the reduced relative sentences were generally longer (significantly so) compared to the unreduced relatives. As expected, Ferreira and Clifton found a significant slow-down specifically at the by-phrase for both animate-subject and inanimate-subject sentences in the reduced relatives as compared to the unreduced relatives. They interpret this slow-down as indicating a garden path effect regardless of subject animacy – participants were relying on the syntactic preference for Minimal Attachment (i.e. the verb takes an NP complement) in spite of the semantic cue provided by the inanimate subject. Furthermore, they argue that the semantic anomaly of the sentence participants were building (The evidence examined . . . as the start of a matrix

132

Animacy and adult sentence processing

transitive sentence) is revealed in a slow-down found at the point of the main verb – this additional slow-down was found only in reduced relatives with an inanimate subject. In other words, Ferreira and Clifton argue that the transitive structure being built was anomalous to participants, but they built it anyway – until they reached the disambiguating by-phrase and were forced to revise their structure. Trueswell and Tanenhaus (1994) and Trueswell et al. (1994) point out that a constraint-based approach is also compatible with a slow-down in reading at the point of by in both types of sentences. According to this approach the construction of a parse is influenced by many factors, including the frequency of individual lexical items and the frequency of the relevant syntactic structures, as well as contextual information that might favor one parse over another. In the case at hand, a sentence beginning with an NP and a verb with suffix -ed is predicted to be more likely to be the start of a matrix transitive clause than a reduced relative clause, all else being equal, since sentence-initial matrix clauses are predicted to be more frequent than sentence-initial reduced relative clauses. Although the relative frequency of these syntactic frames has not been established empirically,3 there is evidence that the frequency of a given verb as a participle has effects on readers’ sentence parsing. For example, Trueswell (1996) showed that the reading time for a reduced relative clause sentence is reliably faster when the -ed verb has a higher participle frequency, as compared to -ed verbs with lower participle frequencies (thus, relatively higher simple past frequencies). Presumably, then, a verb that is more frequently a simple past verb (as opposed to a participle) competes with the thematic information from the subject (i.e. its likelihood to be an agent) and weighs in favor of the matrix transitive parse. In addition, Trueswell and Tanenhaus argue that contextual bias also plays a role in readers’ choice of sentence parse. Without a supporting discourse context for a relative clause, a main clause at the start of a sentence is the most neutral option. Thus, both frequency (for certain verbs) and context weigh in favor of the matrix transitive parse. An animate subject (The defendant examined . . . ) also favors this structure, since it is a possible agent, and so this version of the sentence sets the reader up to expect a transitive main clause with three different cues. An inanimate subject (The evidence examined . . . ) suggests an alternative structure based on the thematic information of the subject (i.e. a poor agent), but the frequency and contextual cues still favor the 3 Several articles cite Tabossi et al. (1994) as substantiating this claim. Although Tabossi et al. had

planned to address this issue, in fact they did not provide evidence for it.

Relative clauses

133

matrix transitive parse. Thus, both approaches predict at least some slow-down at the by-phrase for both animate and inanimate subjects when the sentence is continued as a reduced relative clause. Trueswell et al. (1994) observed that some of the inanimate NPs used as subjects in Ferreira and Clifton’s experiment are not actually unlikely subjects, either because the NP denoted something that could act upon another entity (The car towed . . . ) or because the verb used was unaccusative (The trash smelled . . . ; recall from Section 2.3 that unaccusative verbs easily permit inanimate subjects). Burgess (1990) had tested the likelihood of different subject + past-tense-verb combinations to be beginnings of matrix transitive sentences or reduced relatives by presenting adult native speakers with a determiner, noun, and past tense verb, and asking participants to complete the sentence. Two examples of sentence fragments are given in (3). (3)

a. The spy watched . . . b. The building watched . . .

Participants were instructed to complete the sentence using “good English,” and Burgess analyzed the sentence completions as involving either a simple noun phrase (e.g. the man), which indicated that the verb watched had been analyzed as a simple past tense verb and the clause was a main clause, or a by-phrase (e.g. by the agent), which indicated that watched had been analyzed as a passive participle, and the NP was the head of a reduced relative clause. Burgess found that, in general, participants were more likely to use a simple past tense completion with animate nouns (the spy watched . . . the man) and a passive participle completion with inanimate nouns (the building watched . . . by the agent . . . ), though there were some deviations from this pattern. In particular, in some of the examples taken from Ferreira and Clifton (1986), Burgess’ participants gave simple past completions in spite of the inanimate subject (numbers in parentheses are the percentage of simple past completions): (4)

a. b. c. d.

The skin felt (100%) The ship sighted (70%) The trash smelled (100%) The meal brought (100%)

The results indicated that while inanimate subjects are often analyzed as underlying objects (themes), clearly not all of them have this effect on readers. Although Burgess did not offer an explanation for why these particular verbs

134

Animacy and adult sentence processing

are compatible with an inanimate subject argument, it is likely to be partly due to the fact that smell and feel are unaccusative (Trueswell and Tanenhaus, 1994). As for (4b) it is possible that the NP ship was understood metonymically, as in the crew of the ship (which is then animate). (4d) is perfectly natural in the (perhaps fixed) expression The meal brought me/him/her great pleasure (though I don’t know whether this was the sort of continuation participants offered). Moreover, there could be idiosyncratic variation in the degree to which verbs require an animate subject, just as there is variation in the degree to which verbs require an agent subject (as opposed to experiencer, causer, etc.). Taking Burgess’ results as evidence, Trueswell and Tanenhaus suggested that inclusion of such items (though there were relatively few of them) in Ferreira and Clifton’s experiment could have increased reading times at the by-phrase by favoring a matrix clause parse. In their own study they avoided this potential conflict by using inanimate subjects that were not followed by a matrix clause more than 30% of the time in Burgess’ sentence completion study. Trueswell et al.’s results were in fact different from Ferreira and Clifton’s in that they found a reliable effect of subject animacy: readers slowed down significantly at the by-phrase only when the subject was animate and was the head of a reduced relative clause; no slowing was found when the subject was inanimate or the sentence involved an unreduced relative.4 Thus, Trueswell et al. concluded that subject animacy did play a significant role in the shaping of readers’ on-line sentence parses. Curiously, even with the stricter inclusion criteria for inanimate subjects, Trueswell et al. found that there were particular reduced relatives on which participants appeared to have some difficulty, even though the subject was inanimate. That is, on these items participants slowed down somewhat at the by-phrase, as compared to their reading time in the unreduced relative version. Trueswell et al.’s explanation for this pattern relies on the fact that in order for an inanimate subject to be a (semantically) likely candidate as the head noun of a reduced object relative clause, the NP must be both a poor agent and a good patient. In their construction of stimuli the authors had considered only whether the NPs were good or poor agents. In a follow-up study, they asked adults to rate on a scale how good certain NPs were as either patients or agents. For example, they asked adults, “How typical is it for evidence to be examined by someone?” (patient question) (Trueswell et al., 1994, p. 302), and 4 There was a further methodological difference between the Trueswell et al. and Ferreira and

Clifton experiments having to do with how reading times were calculated. I will not address this aspect of the studies here.

Relative clauses

135

corresponding agent questions (presumably, “How typical is it for evidence to examine something?”). The result was that while inanimate NPs always received a low agent score (on average 1.4 out of 7), they varied in the patient score, ranging from 1.8 to 6.5 (out of 7). After isolating those NPs that received a “good patient” score, Trueswell et al. reported that these items yielded no slowing down at either the verb or the by-phrase; in fact, the reading times associated with these items were nearly identical to those for the unambiguous verbs (verbs with a passive participle form that differs from the simple past form, e.g. thrown). In contrast, those NPs that received a “poor patient” score yielded slower reading times at both the verb and the by-phrase, though the increase in reading time at the by-phrase was less dramatic for these items compared to items with animate NP subjects. It is worth noting as an aside that the notion of “plausibility,” i.e. an NP being a plausible agent or plausible patient, is related to the feature of animacy but is not identical to it. For example, an object NP that is inanimate may or may not be a plausible direct object according to the selectional restrictions of the verb it follows (e.g. . . . stopped the car vs. #. . . stopped the moon). Clifton (1993) manipulated the plausibility of an NP being a direct object where both plausible and implausible NPs were inanimate (Before the police/truck stopped [the moon]/[the Datsun]. . . ), and argued that a postverbal NP is initially analyzed as being a direct object, even if it is an implausible object, if what is measured is first fixation durations (initial reading times), rather than total fixations, which, Clifton argues, may include later re-analysis of the sentence. The important point is that animacy manipulations in sentence processing experiments tell us something about how semantic information is integrated in the process of formulating a sentence parse, but they do not necessarily distinguish the influence of thematic/selectional restrictions as opposed to world knowledge restrictions. In summary, these psycholinguistic experiments show that NP animacy has an effect on parsing at some point in the parsing process, and recent evidence suggests that this effect happens quite early in that process. In general, most inanimate NPs make poor agents, and so their occurrence in subject position can suggest to readers that the sentence will not have a matrix transitive structure, but rather some alternative structure that allows the subject to be construed as the patient of the verb, e.g. a reduced object relative structure. A complication with this general picture arises from details of semantic compatibility between NPs and the individual verbs that select them. As we saw, certain inanimate NPs make perfectly good subjects/agents when they are paired with certain verbs (e.g. the car towed . . . ), and certain

136

Animacy and adult sentence processing

other inanimate NPs make terrible patients when paired with certain verbs (. . . stopped the moon). In order to truly isolate the effects of NP animacy on sentence parsing we would need to construct sentences that differed only along this dimension and kept all other parts of the sentence the same. The experiment described in Section 4.2.1 includes pairs of sentences that meet this stricter criterion (though in the context of a sentence completion task, not an eye-tracking study). Putting this issue aside for the moment, the experiments described above illustrate a general preference, all things being equal, for the subject of the sentence to be construed as an agent and the sentence structure that of a main clause. It is only when things are profoundly unequal, as when one encounters an inanimate subject, that readers are pushed to consider alternatives. This tendency is underscored in the studies described in Section 4.2. 4.1.2 Subject vs. object relative clauses In the studies on relative clauses just mentioned, the two competing analyses involve a matrix clause (i.e. no displacement of the subject) or a reduced relative clause (i.e. displacement of the subject from the object position inside the relative clause). More recent work on the role of NP animacy in sentence processing has focused on the relative ease of processing subject vs. object unreduced relative clauses (Mak et al., 2002; Traxler et al., 2002, 2005). In these constructions, the subject is unambiguously displaced, but the position where it originated from depends on the sentence parse. In a subject relative, it is displaced from the subject position inside the relative clause, and in an object relative it is displaced from the object position in the relative clause. These studies allow a finer-grained distinction among different types of displacement, and underscore the link between subject animacy and an NP’s thematic role, as opposed to its surface position. The traditional observation is that it is harder to parse object relatives (OR), like (5), than subject relatives (SR), like (6), and this asymmetry has been demonstrated in English (Wanner and Maratsos, 1978; Ford, 1983; King and Just, 1991) as well as other languages (for example, French (Frauenfelder et al., 1980), German (Mecklinger et al., 1995) and Dutch (Frazier, 1987)).5 (5)

The lawyer [that [the banker irritated t]] filed a hefty lawsuit. (OR)

(6)

The lawyer [that [t irritated the banker]] filed a hefty lawsuit. (SR)

5 The opposite preference was claimed to hold in Mandarin by Hsiao and Gibson (2003), but

see Lin and Bever (2006) for counter-evidence and arguments that the SR preference holds in Mandarin as well.

Relative clauses

137

Earlier work on this topic focused on explanations involving either working memory load or differences in syntactic structure between OR and SR structures. The argument for working memory notes that OR sentences impose a greater working memory load because the referent of the “filler” (the lawyer) must be kept in memory for a longer stretch in (5) than in (6) (Wanner and Maratsos, 1978), or because more other referents intervene between the filler and its gap (t) in (5) than (6) (Gibson, 1998). Syntactic explanations point out that readers have an easier time with SR structures either because the filler NP has the same grammatical/thematic role in both the main and relative clauses (Sheldon, 1974) or because readers have a bias to analyze the main clause subject as also the subject of the relative clause (Clifton and Frazier, 1989, i.a.).6 There is some recent evidence that manipulating the animacy of the sentence subject (which is also the head of the relative clause) and the noun phrase inside the relative clause (the banker in the above examples) affects the ease with which these constructions are parsed. In two studies of OR processing, Weckerly and Kutas found a significant advantage for OR constructions containing inanimate subjects in terms of reading times (Weckerly and Kutas, 1998) and Event-Related Potential (ERP) effects (Weckerly and Kutas, 1999). Traxler et al. (2002) (Experiment 3; see also Traxler et al. (2005)) presented adults with four types of sentences in an eye-tracking study, manipulating both the animacy of the sentence subject and the type of relative clause: (7)

a. The director [that [t watched the movie]] received a prize at the film festival. (animate–SR) b. The director [that [the movie pleased t]] received a prize . . . . (animate– OR) c. The movie [that [t pleased the director]] received a prize . . . . (inanimate– SR) d. The movie [that [the director watched t]] received a prize . . . . (inanimate– OR)

In measuring the “quasi-first pass” reading times for the relative clause region of the sentence7 these authors found a main effect of clause type 6 Other accounts of the OR/SR asymmetry point to a shift in perspective that is required by OR

constructions but not SR constructions (MacWhinney, 1977; MacWhinney and Pléh, 1988). These accounts will not be described here; a brief summary is given in Traxler et al. (2002, pp. 71–72). 7 “Quasi-first pass reading times” refers to the sum of people’s fixations on a given region – for example the relative clause region – from the beginning of that region until the person fixates on anything to the right of that region.

138

Animacy and adult sentence processing

(OR sentences were generally slower than SR sentences) but also a significant interaction between animacy and clause type. Sentences with an animate subject and an OR type of relative clause (7b) yielded slower reading times than any other type of sentence (7a,c,d). The same effect was found in the total time for the relative clause region, and in the quasi-first pass reading times for the matrix verb region. Mak et al. (2002) conducted a similar study on Dutch, and also found that OR constructions containing an inanimate relativized NP appeared to be easier to parse than OR constructions containing an animate relativized NP. One of the problems in interpreting the findings in both the Traxler et al. experiments and the one by Mak et al., is that both the subject NP and the embedded verb were manipulated (please vs. watch in Traxler’s examples above). These verbs were obviously chosen in order to make the sentences sound natural: please is a psych-verb that takes a theme subject and experiencer object. Recall that themes can easily be inanimate while experiencers are animate (The student/gift pleased the teacher/*The student/gift pleased the book). Watch, on the other hand, takes an agent subject, which must be animate.8 But given these verb choices, it is not possible to tease apart the possible influences of NP animacy and verb choice on subjects’ parsing strategies (see King and Just (1991) for evidence that the semantic pairings between nouns and verbs matter for sentence processing).9 A more recent study by Lowder and Gordon (2012) did tease these factors apart by comparing the reading times of sentences that differed only in the animacy of one NP. They compared sentences like the following:10 (8)

a. The sheriff [that [t injured the cowboy]] persuaded the members of the jury. (SR) b. The sheriff [that [the cowboy injured t]] persuaded the members of the jury. (OR–Animate) c. The sheriff [that [the pistol injured t]] persuaded the members of the jury. (OR–Inanimate)

8 Inanimate NPs are acceptable where an animate NP is generally required if the inanimate

NP is understood metonymically, i.e. standing in for a semantically related entity such as an institution, group of people, etc. For example the word college can refer metonymically to the college administration, and in this sense The gift pleased the college is not ill-formed. Here we will abstract away from these cases. 9 I thank Carson Schütze for discussion about this point. 10 These sentences were modeled on the ones used in Traxler et al. (2005).

Processing of raising and control

139

Notice that here it is not the animacy of the head noun that is manipulated, but the animacy of the subject of the relative clause. Lowder and Gordon found that both of the OR sentences were more difficult to parse than the SR sentence. They found no difference between the animate and inanimate OR sentences at the point of the matrix verb (persuaded), suggesting that neither type of OR sentence was globally easier to parse. But they did find a significant difference between the two types of OR sentences at the embedded verb (injured), with the inanimate version taking longer to parse. That is, pistol injured took longer to parse than cowboy injured. These researchers surmised that the delay in the former case was related to difficulty integrating an inanimate subject with an eventive verb (what they call an action verb) when the two are clausemates. Indeed, a further experiment revealed that this local slowing of reading time surfaced when an inanimate subject and matrix verb were adjacent within the same clause (as in (9a)) but not when the verb was part of an adjunct (as in (9b)). (9)

a. The pistol injured the cowboy last night in the saloon. b. The pistol that injured the cowboy was known to be unreliable.

Although the results of Lowder and Gordon’s experiments appear to cast doubt on the robustness of the animacy effect for parsing object relative clauses espoused by Traxler, Mak, and their colleagues, an important animacy effect was nevertheless revealed: people have a harder time parsing a sentence when an inanimate subject is paired with a verb that allows both animate and inanimate subjects, than when an animate subject is paired with the same verb, at least when the two are clausemates. This asymmetry, in turn, might indicate that people generally expect subjects to be animate, and an inanimate subject breaks this expectation. One view of how this affects sentence processing is that, “following an initial inanimate noun [i.e. an inanimate subject] . . . there may be no singular strong expectancy” about how the sentence will continue (Weckerly and Kutas, 1999, p. 566). That is, inanimate subjects force readers to abandon their normal expectations about sentence structure. The remainder of this chapter will describe studies showing this effect in main clauses with displacing predicates.

4.2

Processing of raising and control

The preceding discussion has focused on readers’ use of animacy in interpreting the semantic role of the subject of a sentence, and the influence of this

140

Animacy and adult sentence processing

interpretation on the syntactic structure readers assign to the sentence. Readers tend to interpret an initial animate NP as the subject of a matrix transitive clause, and must slow down when the sentence turns out to have a reduced relative clause structure. On the other hand, they interpret an inanimate subject more readily as the head of a reduced object relative clause. While unreduced object relative clause structures are generally harder to parse than unreduced subject relative structures, some evidence suggests that this difficulty is lessened, or even eliminated, when the subject is inanimate, as opposed to animate. Although the apparent animacy effect in mitigating the difficulty of unreduced object relatives may be due to a confluence of features, including both noun animacy and verb selectional properties manipulated in the experimental stimuli, other evidence points to a real hesitancy on the part of readers to pair an inanimate subject thematically with an adjacent verb. These are useful observations because they tell us that people tend to interpret sentence-initial animate NPs as non-displaced (semantic) subjects and inanimate NPs as displaced subjects (semantic objects). To return our discussion to the topic of the displacing predicates covered in Chapter 2 I will discuss the results of two studies, a sentence completion (fillin-the-blank) task (Becker, 2005), and a novel verb learning study (Becker and Estigarribia, 2011, 2013) that directly targeted raising and control verbs. 4.2.1 Sentence completion In a 2005 study, I asked adult participants to write a verb in a blank to complete a sentence. Across four experiments, different aspects of the target sentences were manipulated, with the purpose of finding out which such manipulations drove participants to write in more raising verbs as opposed to control verbs. Pilot work had revealed that adults were quite reluctant to write raising verbs in sentences, even though these sentences were felicitous with a raising verb completion. For example, (10) could felicitously take a raising verb or a control verb in the blank. (10)

The batter

to hit the ball hard.

a. The batter seemed to hit the ball hard. (raising) b. The batter tried to hit the ball hard. (control)

However, participants generally wrote control verbs in these sentences. A full 95% of responses to sentences like (10) were control verbs; 0% of responses were raising verbs. The remaining 5% of responses were either “ambiguous” verbs, like begin, or intransitive verbs like jump whose use

Processing of raising and control

141

implies a purpose clause interpretation. It seemed, in fact, that control verbs were a kind of “default” choice, and piloting suggested that adults could be led away from the default choice by altering certain semantic and syntactic properties of the sentences. Thus, in Becker (2005) I sought to manipulate each of these properties independently in order to find out their relative strength in eliciting raising verb responses. A reasonable way to interpret the results of this task is that if a person writes a control verb in the blank, then they have represented the sentence as having a control structure, and therefore they assume a thematic relationship between the subject and main verb. On the other hand, if a person writes a raising verb in the blank, then they have represented the sentence as having a raising structure and assume no thematic relationship between subject and main verb. Some of the experiments in this study manipulated the syntactic structure of the sentence. For example, in Experiment 1 participants were given sentences like those in (11). (11)

a. Barry b. It

to know the answer even before she finished the question. that the principal believed her excuse for being late.

Sentence (11a) is compatible with either a raising or a control verb completion (Barry seemed to know . . . , Barry claimed to know . . . ), but (11b) has an expletive subject and a sentential complement and was therefore expected to require a raising verb for syntactic reasons.11 For sentences like (11a) participants wrote control verbs 52.5% of the time, and raising verbs 32.5% of the time (the remaining responses were ambiguous verbs like begin), so although both types of verbs were offered there was a clear preference for control verbs (see Table 4.1).12 One interesting result was that the it subject in (11b) was not uniformly interpreted as an expletive subject, as seen in the relatively low (55%) rate of raising verb responses for this type of sentence. In 45% of cases participants interpreted the it subject as referential it and thus offered a variety of responses consistent with that interpretation (e.g. It assumed that Barry knew the answer even before she

11 The study reported in Becker (2005) was run twice, first with twenty participants in each exper-

iment, and a second time with ten participants in each experiment. The second group of ten participants per condition was run in order to replicate the initial study with different stimuli. The overall patterns were the same in the two groups. Here I report only the figures from the initial group. Please see Becker (2005) for the full set of results. 12 The column labeled “Ambig./Other” contains responses that are ambiguous verbs like begin as well as other types of responses, such as intransitive verbs and mental verbs (e.g. know).

142

Animacy and adult sentence processing

Table 4.1: Percentage of raising and control responses in Becker (2005),

Experiments 1–3 Sentence type

Example

% Raising % Control % Ambig./ Other

NP + infin. (11a)

Barry to know the answer even before she finished the question. that the principal It V that S It believed her excuse for being (11b) late. to be raining for most of expl. it (12a) It the morning. expl. there There to be five dozen (12b) cartridges strewn about the floor.

32.5

52.5

15

55

0

45

90

0

10

97.5

0

2.5

finished the question – note that such responses appear to violate mental verbs’ selection of a sentient subject!). Experiments 2 and 3 demonstrated that when a subject is unambiguously non-referring, as when it is weather-it or existential there, participants were at ceiling in providing raising verbs in the blanks. (12)

a. It b. There

to be raining for most of the morning. to be five dozen cartridges strewn about the floor.

The results from the first three experiments are shown in Table 4.1. Experiment 4 explicitly manipulated the animacy of the subject, rather than the category of the complement (Experiment 1) or the referentiality of the subject (Experiments 2 and 3). Using the opaque construction in (11a), half of the target sentences had an animate subject and half had an inanimate subject, and the embedded infinitive predicate was kept the same across conditions. (13)

a. His campaign manager candidate. (animate) b. The extramarital affair candidate. (inanimate)

to remain a problem for the mayoral to remain a problem for the mayoral

I predicted that inanimate subjects would lead to more raising verb responses, since raising verbs, which do not select their subject, are not picky about what goes there: raising verbs can tolerate animate and inanimate subjects equally. Control verbs, on the other hand, generally select animate

Processing of raising and control

143

Table 4.2: Percentage raising and control responses in Becker (2005), Experiment 4 Subject type

Example

Animate (13a)

His campaign manager to remain a problem for the mayoral candidate. to The extramarital affair remain a problem for the mayoral candidate.

Inanimate (13b)

% Raising

% Control

% Ambig./ Other

18.8

52.5

28.8

43.8

17.5

38.8

subjects. In principle, sentences with animate subjects could have yielded either raising or control verb responses, since no ungrammaticality would have resulted from either choice. But animate subjects in fact triggered significantly more control verb responses than raising verb responses. Thus, participants were significantly more likely to write a control verb in the blank given a sentence like (13a) than (13b), and they were significantly more likely to write a raising verb for (13b) than (13a). Thus, inanimate subjects influenced participants’ choice of raising verbs to complete the sentence. This result is shown in Table 4.2. Two further points about the study warrant attention. One is that participants also wrote in a fair number of ambiguous verbs (e.g. begin, start), which Perlmutter (1970) argued can be either raising or control verbs (see discussion in Section 2.5). Although Perlmutter argued that when ambiguous verbs are used with an animate subject they function as control verbs, and when used with an inanimate subject they function as raising verbs, I kept these responses separate. These responses are grouped in the above tables under the “Other” category.13 Note, however, that if we provisionally add ambiguous verbs supplied for sentences with an inanimate subject to the raising tally, and ambiguous verbs supplied for sentences with an animate subject to the control tally (in line with Perlmutter’s suggestion about how these verbs should be categorized), the percentage of raising responses for inanimate subject sentences rises to 68%, and the percentage of control responses for animate subject sentences rises to 70%. 13 The “Other” category also includes responses in which the infinitive clause appeared to be

construed as a purpose clause (e.g. “It scrambled to scurry along the edge of the field, as if pursued by something”), and a few ungrammatical responses (“It may to be too foggy to drive safely”).

144

Animacy and adult sentence processing

Table 4.3: Percentage raising and control responses in Becker (2005), Experiment 4 (separated by aspect of embedded verb) Subject type

Pred. type

Example

% Raising

% Control

% Ambig./Other

Animate Animate Inanimate Inanimate

Eventive Stative Eventive Stative

(14a) (14b) (14c) (14d)

5 32.5 17.5 70

65 40 32.5 2.5

30 27.5 50 27.5

The second point is that another property of the sentences was manipulated in this experiment, in addition to subject animacy. That property was the lexical aspect of the infinitive predicate – whether the embedded verb was stative or eventive. I had noticed in pilot work that this seemed to have an effect on participants’ choice of which verb to write in the blank, and so it was systematically manipulated here. The aspect of the lower clause was crossed with subject animacy to yield four possible sentence types: (14)

a. animate subject, eventive predicate to hit the car on the passenger’s side. The driver b. animate subject, stative predicate to remain a problem for the mayoral His campaign manager candidate. c. inanimate subject, eventive predicate to hit the car on the passenger’s side. The boulder d. inanimate subject, stative predicate to remain a problem for the mayoral The extramarital affair candidate.

Participants were indeed strongly influenced by the aspect of the embedded clause: an eventive lower predicate led to more control verb responses in the main clause, while a stative lower predicate led to more raising verb responses (see Table 4.3).14 In addition, there was a significant interaction between the two properties of the sentence: the combination of an inanimate subject and a stative lower predicate, as in (14d) led to 70% raising verb responses (if ambiguous verbs are added to the tally, the percentage rises to 93%), while 14 In fact, the eventive/stative manipulation at least partially accounts for why participants wrote

so many more control verb responses (95%) in the pilot sentence (10) which contains an eventive predicate, as compared to the 52.5% control verb responses in sentences like (11a) or (13a), which contain stative predicates.

Processing of raising and control

145

the combination of an animate subject and an eventive lower predicate led to 65% control verb responses (again, if we add in the ambiguous verbs that could have been control verbs here, the percentage rises to 80%). It is not entirely clear why the aspect of the lower clause verb influences the choice of the matrix verb (and, correspondingly, the structure of the sentence), since there is no selectional relationship between these two parts of the structure. One possibility is that the effect of the lower verb’s aspect has to do with the general relationship between lexical aspect and thematic roles. As observed by Lakoff (1966); Lyons (1968); Gruber (1976) and mentioned above in Section 3.2.1, eventive (Gruber called them “active”) predicates are more likely to take an agent subject, as opposed to an experiencer or theme subject. Stative predicates, on the other hand, do not allow an agent subject and instead take a theme or experiencer (remain takes a theme subject, know takes an experiencer). Many control verbs also take an agent subject (claim), some an experiencer (want), but none take a theme subject. Recall that in both raising and control constructions the embedded clause subject (t or PRO) refers to the matrix clause subject, and even though in subject control constructions the matrix clause subject gets its θ -role independently of the embedded clause subject (PRO), it is essentially the semantic subject of the embedded clause too. Perhaps if an eventive lower predicate assigns an agent (or causer) θ-role to its subject (t or PRO), then the matrix subject, which must be co-referential with the embedded subject, is also associated with the agent/causer role. Since control verbs can assign an agent role, readers are inclined to think of a control verb in the matrix clause (possibly due to priming, though this has not been tested specifically). In the examples above, then, [. . . to hit the car] needs an agent (or causer), and if readers associate the agent role with the matrix subject, they will think of a control verb. (In the case of sentences like (14c), The boulder – to hit the car. . . , the matrix subject could not really bear an agent role, but it could be considered a causer. Since in a control construction the matrix and embedded subjects get independent θ-roles this is not problematic.) On the other hand, stative predicates may take a theme or experiencer subject. Focusing on theme subjects for a moment, these are often displaced subjects, as discussed in Chapter 3. Recall, too, that in raising constructions the subject of the embedded clause (t) and matrix clause share a single θ -role (Section 2.1.1). Therefore, it may again be a case of associating the θ -role of the embedded subject with that of the matrix subject: the stative verb evokes a theme argument, which is likely to be a displaced argument, which in turn elicits a main predicate that has no external argument, i.e. a raising verb.

146

Animacy and adult sentence processing

Notice that I am not claiming that an eventive verb implies an animate subject, or a stative verb implies an inanimate subject. In this experiment the animacy of the matrix subject was given to participants, and when an inanimate matrix subject was paired with an eventive lower predicate, participants still offered more control verbs (32.5%) than raising verbs (17.5%). Thus it seems like the eventive lower predicate implies an agent/causer subject, which in turn elicits a control (non-displacing) representation. Interestingly, when an animate matrix subject was paired with a stative lower predicate participants offered marginally more control verbs (40%) than raising verbs (32.5%), though they still offered more raising verbs in this condition than they did in the inanimate subject–eventive predicate condition (17.5%). Thus, a stative lower predicate may, by selecting a theme/experiencer subject, elicit a displacing structure representation. However, the exact nature of the relationship between the eventive/stative aspect of predicates, subject animacy, and displacing/non-displacing clause structure is far from clear and I hesitate to draw firm conclusions about how these various cues might interact in the learning process. Another possible reason for the influence of the aspect of the lower verb relates to frequencies of combinations of raising vs. control verbs with eventive vs. stative lower predicates in natural speech. Garrett Mitchener and I reported that in the Switchboard corpus (Taylor et al., 2003), which contains naturalistic speech from adult-to-adult conversations, only 55% of sentences with raising verbs (with an animate subject) had an eventive lower predicate, while 80% of sentences with control verbs (with an animate subject) had an eventive lower predicate (Mitchener and Becker, 2011).15 Of course, the frequencies of raising plus eventive vs. control plus eventive pairings may help to explain the pattern of adult responses in Becker (2005), but it begs the question of where these asymmetric frequencies come from. I must leave a further fleshing out of this effect to future work. To summarize, sentence completion tasks can be used to find out about speakers’ structural representations of sentences. A speaker’s choice of how to complete the sentence reveals the kind of structure they attribute to the string. My experiment, which manipulated subject animacy, found that inanimate subjects triggered a representation involving a derived subject: an inanimate subject in a sentence frame compatible with both raising and control verbs 15 If we include all utterances with raising or control verbs, not just those with animate subjects,

the proportion of raising verb utterances with an eventive lower predicate drops to 46%, and the proportion of control verb utterances with an eventive lower predicate remains at 80%.

Processing of raising and control

147

yielded significantly more raising verb responses. Animate subjects in the same sentence frame yielded significantly more control verbs, though we also saw complicating factors involving the lexical aspect (stative vs. eventive) of the lower predicate, factors which are not currently well understood. Thus, subject animacy influences adult speakers’ intuitions about sentence structure in raising vs. control sentences just as it did in relative clause structures, as discussed in Sections 4.1.1 and 4.1.2. This shows that subject animacy is pervasive in influencing adults’ interpretation of sentence structure. Quite uniformly, inanimate subjects are preferentially interpreted as derived subjects, while animate subjects are not. 4.2.2 Novel verb learning While tasks that measure on-line sentence comprehension and production give us information about how adults process sentences composed of words already in their lexicon, the real question, with respect to learning, is how new words and the sentences they occur in are encoded in the first place. In the next chapter we will look at studies of the role of animacy in children’s learning of sentence structure in general, and of displacing predicates in particular. Before proceeding to that literature, let us look at how language learning can be simulated in adult speakers. There are two important reasons for simulating language learning in adults. One is that, as pointed out by Gleitman et al. (2005) (and mentioned at the beginning of this chapter), a possible explanation for children’s errors or difficulties with certain constructions is their lack of conceptual wherewithal. For example, mental verbs, like think or believe, and the sentential complement constructions they often occur in, are acquired later than more concrete verbs that denote actions and physical states. Perhaps this is so because children have difficulty representing these mental states. On the other hand, if it turns out that adults also have difficulty “learning” these verbs without particular types of cues (such as sentential-complement syntax), then we wouldn’t normally suppose this is so because adults lack the relevant conceptual ability. Then, we can hypothesize that children’s late learning of these predicates also does not rest solely on their conceptual deficiencies, and that the same kinds of syntactic cues that help adults “learn” these verbs also help children learn them. The same would apply to opaque constructions. If we can show that adults make use of inanimate subjects in “learning” novel raising verbs, or tough-adjectives, then, presumably, child learners could make use of this cue as well. The second reason for testing learning simulations in adults is more practical. Adults are generally easier to recruit for studies than children, and they

148

Animacy and adult sentence processing

have considerably longer attention spans and so can be asked to respond to many more stimuli than children. Fortunately, it is fairly easy to simulate language learning on a very small scale by presenting adults with novel (made-up) words in various contexts and asking them to guess what the novel word means. This kind of experimental approach is known as the Human Simulation Paradigm (HSP; Gillette et al. (1999)). Many variations on this paradigm exist, but the general idea is to present adults with varying degrees of situational and linguistic information, and ask them to make a guess about the meaning of a word occurring in this situational or linguistic context. The original HSP studies presented adults with a video of a mother-child play scene without audio, and adults heard a beep whenever the mother uttered a particular word in her speech (e.g. the verb call). Upon hearing each beep, adults had to guess which word the mother had uttered, and at the end of the video they had to make one final guess. Across different conditions Gillette et al. provided participants with additional cues, such as a list of sentences used in the video scene but with all words turned to nonsense, or only the content words turned to nonsense (function words, like determiners and auxiliaries, remained in their English forms), or only the verb turned to nonsense. In some conditions the audio-less video was provided; in others there was no video, and only linguistic cues (such as lists of words used in the conversation) were available to participants. By manipulating these different linguistic and situational cues the researchers were able to measure the relative strength of these different cues in helping adults make correct guesses about the target verb. My colleague Bruno Estigarribia and I have conducted a study in the spirit of HSP with adults, to test the effect of subject animacy on the categorization of novel verbs as raising verbs (Becker and Estigarribia, 2011, 2013). The Becker (2005) study showed that inanimate subjects in the opaque sentence frame elicited known raising verbs, and animate subjects in the same frame elicited known control verbs, but we wondered whether the same effect would be found in the categorization of unknown verbs. Our study used a slightly different methodology from most of the traditional HSP work, in that there was no situational context (video scene) provided to participants. Instead, they were given only linguistic input, in the form of lists of sentences, or sentences embedded in a (very) short story. These sentences contained a novel verb (for example, The old man joops to be very tired), and after reading the stimulus sentences participants had to make a forced choice grammaticality judgment about the same novel verb used in two further sentence constructions (constructions that were not used in the stimulus sentences). The two test sentence constructions were a there-sentence (There

Processing of raising and control

149

Table 4.4: Novel verbs and their pseudo-definitions Target (raising-like):

Definition

joop meb trollick

to look a certain way to probably be a certain way to be some way most of the time

Fillers: Control-like zid rickle sart

to really enjoy being someplace to really dislike being someplace to make a big effort to be some way

Transitive/Intransitive zorp ballop flim

to touch something very gently to swing your arms in circles to breathe fire and be scary

joops to be a computer on the desk) and a pseudocleft (What the fairy joops is to be small). Since there-sentences allow raising but not control verbs (There seems/*claims to be a man in the garden), and pseudoclefts allow control but not raising verbs (What John claims/*seems is to be a nice guy), we predicted that if a participant had categorized the novel verb as a raising verb, she would choose the there-sentence as “sounding better.” If she had categorized the novel verb as a control verb, she would choose the pseudocleft. For this experiment, we constructed three novel raising verbs and two types of novel filler verbs, one type resembling control verbs and the other type being novel transitive or intransitive verbs. Each novel verb was assigned a made-up definition that was semantically compatible with its category (see Table 4.4). Participants were not always privy to these definitions (according to the condition they were in; see below), but we constructed our sentence stimuli so that the verbs were always used in a way that was sensible given their assigned definition. All participants saw all verbs, so verb type was manipulated within participants. Between participants we manipulated a number of different cues in our stimulus sentences. The two cues of primary interest were (a) one which manipulated subject animacy (i.e. novel raising verbs were used with only animate subjects, or with at least one inanimate subject), and (b) a cue to the lexical meaning of the verb (the pseudo-definitions shown in Table 4.4). These two cues were partially crossed, so that some participants received the animacy cue (raising verbs occurred with at least one inanimate subject) but did

150

Animacy and adult sentence processing

not receive any definitions, some participants received definitions but all verbs were presented with an animate subject (NPanim verb to S), and other participants received both cues together for each verb.16 We hypothesized that for the novel raising verbs, getting a sentence with an inanimate subject would be more informative than the definition cue alone, but that the two cues together would have an additive effect. In addition, we manipulated the frequency with which participants encountered each verb (1, 3 or 5 exposures to each novel verb), and the context of the sentences containing the novel verbs (either independent sentences or part of a story). Furthermore, all stimuli with an infinitive complement (which was all of the stimuli for our novel raising verbs) contained a stative predicate. This was done in order to avoid potential influences of eventive verbs found in the Becker (2005) study (see Section 4.2.1 above). As already mentioned, verb type was manipulated within participants, so each participant was exposed to three novel raising verbs and six novel fillers, three of which were novel control verbs, and three were novel transitive or intransitive verbs. All other cues were manipulated between participants, so with 3 (animacy only vs. definition only vs. both) × 3 (1, 3, or 5 exposures) × 2 (list vs. story) variables we ended up with eighteen different conditions to which participants were randomly assigned. The observed mean “correct” responses for our target verbs (i.e. choices of the there-sentence) and our control-like fillers (correct means choosing the pseudocleft) are shown in Table 4.5 (this table shows responses in the list condition only, i.e. not the story condition).17 Particularly low or high correct response rates for the novel raising verbs are in boldface. There were three main results of this experiment. One was that participants were at ceiling in categorizing our control-like filler verbs as control 16 As an analogue to the subject animacy manipulation with novel raising verbs, we introduced

a similar manipulation with the control-like filler verbs in which the novel control verbs were used either in typical control frames (i.e. with an infinitive complement) or in a transitive frame (NP complement). The condition where novel control verbs occurred with at least one transitive frame was designed to be analogous to a novel raising verb’s occurrence with at least one inanimate subject, because certain control verbs can be used in transitive frames (e.g. want). Thus, we reasoned that observing the same novel verb used in both of these frames (infinitive and NP complements) might have a facilitative effect on participants’ categorizations of these verbs. However, since participants were already essentially at ceiling for correct categorizations of novel control verbs there was no room to measure any facilitative effect of this manipulation. 17 We found a small but significant interaction between the animacy cue and the story/list manipulation. The animacy cue appears to have been more helpful in the list context than the story presentation, but the animacy cue remained significant even in the full model. See Becker and Estigarribia (2013) for discussion.

Processing of raising and control

151

Table 4.5: Observed mean percentage “Correct” responses 1 Exemplar

Raising (target) Control (filler)

3 Exemplars

5 Exemplars

def.

anim.

both

def.

anim.

both

def.

anim.

both

53 100

93 70

73 97

60 90

77 100

70 93

40 100

91 100

73 100

verbs (they picked the pseudocleft as sounding better). Categorization of our novel “raising” verbs was less than perfect and in fact exhibited quite extreme variance according to the condition. Participants gave significantly more correct responses for novel raising verbs when they saw that verb with at least one inanimate subject, as compared to when they were told the definition of the verb but saw the verb only with animate subjects. Thus, the second main result was that the animacy cue was highly significant for the raising verbs. This was true across the different frequency conditions, but we can see it most clearly in the 1-exemplar and 5-exemplar conditions. Notice that correct raising categorizations occurred in only 53% of cases (in the 1-exemplar condition; 40% of cases for 5 exemplars) if a lexical definition was provided but the verb was used only with an animate subject. But removing the lexical definition and replacing it with an inanimate subject made the proportion of correct raising categorizations jump to 93% (1-exemplar; 91% for the 5-exemplar condition). There was no significant effect of frequency, and so in the following graphic depiction of the results, I have collapsed the data across the different numbers of exemplars (see Figure 4.1). (Like the table above, this graph shows responses only from the list condition; see Becker and Estigarribia (2013) for results of the story condition.) The third interesting result, which we can observe in Figure 4.1, is that when the two cues were presented together (the “both” condition), they did not have an additive effect for novel raising verbs. In fact, their effects appear to be averaged. That is, although the percentage of correct responses for raising verbs was significantly better in the “both” condition than the “definition” condition, it was worse in the “both” condition than the “animacy” condition. Thus, for novel raising verbs, the animacy cue was much more powerful in guiding raising categorizations than the definition cue, and the presentation of both cues together did not make raising categorizations more likely than presentation of the animacy cue alone. We do not know why this is, but we can imagine two possible reasons. One possibility is that the presence of the lexical definition interfered with participants’ ability to use the animacy cue (inanimate subject)

152

Animacy and adult sentence processing

100 90 80

Percent Correct

70 60 50 Raising (target) 40

Control (iller)

30 20 10 0 Deinition

Animacy Condition

Both

Figure 4.1. Percent “Correct” responses for novel raising (target) and control (filler) verbs; list condition only, averaged across frequency conditions

when both cues were presented together. Another possibility is that these different cues are weighted by participants probabilistically, so that if both cues are presented together each cue contributes proportionally less weight to the categorization process.18 We did not find a significant main effect of frequency of presentation, but this cue did have a somewhat surprising effect on novel verb categorizations. We had predicted that more exemplars would serve to solidify participants’ mental categorization of verbs as raising verbs, so that we would find more raising categorizations in the 5-exemplar condition than the 1-exemplar condition, and that the rate of raising categorizations in the 3-exemplar condition would fall in the middle. However, this was not what we found. Instead, participants became less likely to categorize a novel raising verb as raising the more exemplars they saw. To explain this somewhat surprising result we must look more carefully at the types of frames presented in the subject animacy condition. In particular, in the 1-exemplar condition, participants saw a single sentence with an inanimate subject (and infinitival complement), and had to make their judgment on this basis (see (15)). 18 I am grateful to a reviewer for this suggestion.

Psycholinguistic effects of animacy on production of the passive (15)

153

The book joops to be very long.

So, 100% (1/1) of the sentences they saw had an inanimate subject. In the 3-exemplar condition participants saw three sentences, one with an inanimate referential subject, one with a weather-it subject, and one with an animate subject, as in (16). (16)

a. The old man joops to be very tired. b. The book joops to be very long. c. It joops to be sunny outside.

Here, only 66.7% (2/3) of the sentences contained an inanimate (or at least non-animate) subject. The 5-exemplar condition presented one animate subject, two inanimate referential subjects, and two weather-it subjects, yielding 80% (4/5) sentences with inanimate subjects. Our reason for including one sentence with an animate subject in the 3- and 5-exemplar conditions was to simulate the fact that raising verbs do occur quite often with animate subjects in speech (though they also occur sometimes with inanimate subjects, unlike control verbs; see Chapter 6), and so we thought it natural to have these verbs occur some (relatively small) proportion of the time with animate subjects. This experiment would need to be compared to a version in which all three (or all five) sentences have only inanimate subjects, but the data at hand seem to suggest that even a single exemplar of a verb with an animate subject weakens (or perhaps serves as a counterexample to) the categorization of that verb as a raising (displacing) verb. To summarize, we found a strong effect of subject animacy on adult speakers’ categorization of novel verbs as raising as opposed to control. Although speakers’ default categorization was control, meaning that they tended to choose the pseudocleft as sounding better with our novel verbs, they were led away from that choice significantly if they encountered a novel verb with an inanimate subject. Thus, inanimate subjects, more than raising-like lexical definitions (see our pseudo-definitions in Table 4.4), led adults to categorize the novel verb as a raising verb.

4.3

Psycholinguistic effects of animacy on production of the passive

Before concluding this chapter I want to mention a study that might appear to contradict the main thrust of my argument: that inanimate subjects are more

154

Animacy and adult sentence processing

likely to be displaced subjects than animate ones. In a sentence production study by Ferreira (1994) adults were presented with two nouns and a past tense verb (e.g. MANAGER LAYOFFS WORRIED) and were asked to come up with a full sentence using just those words (plus any necessary function words). At issue was whether participants would produce an active or a passive sentence, and whether the choice of active or passive voice would be influenced by the argument structure of the verb (agent/experiencer–theme or theme– experiencer verbs) or the animacy of the nouns. Some of the verbs were theme–experiencer verbs like worry or frighten (since themes can be inanimate, these verbs allow an inanimate subject) and some were what Ferreira called “normal” verbs: agent–theme or experiencer–theme verbs (e.g. kick, fear). These are verbs that require an animate subject, since agents and experiencers are animate (see Chapter 3). Ferreira found that passives were produced overall less frequently than actives, but that theme–experiencer verbs yielded significantly more passives (31% when one of the NPs was inanimate, as in the example above) than “normal” verbs did (4% in the same condition). She also found that when both NPs were animate, theme–experiencer verbs still triggered more passive productions (than “normal” verbs), but the asymmetry was far weaker. Having one inanimate NP and a theme–experiencer verb yielded the most passives. However, the passive of a theme–experiencer verb is a sentence that has an animate subject! So the participants in this experiment were preferentially placing animate NPs in a derived subject position. It is in this sense that these results appear to contradict my claims, and the other results reviewed in this chapter. But in fact these results are not contradictory with my general claims. Ferreira explains her results in terms of the overall preference for the subject to be animate, an explanation that is in keeping with the picture laid out in Chapter 3. Recall from Chapter 2 (see Section 2.4.2) that passive verbs are not really a separate category of verbs from actives. Rather, they are the same verb but expressed in a different voice (at least, on non-lexicalist accounts of the passive). Learning to parse a passive construction is not a matter of verb categorization, the way distinguishing raising from control, for example, involves categorizing different verbal predicates. Thus, Ferreira’s results are fully in line with the general picture presented here and underscore the fact that subject animacy serves a different function in psycholinguistic processing of the passive as opposed to raising and other displacing predicates. In addition, it is important to keep in mind that this was a production study, not a comprehension study. It is possible that reading a passive sentence with an inanimate subject, rather than an animate subject, might be facilitative

Summary

155

in terms of processing time (similar to the apparent advantage of inanimate subjects in OR constructions seen in Section 4.1.2; such an effect was found in Slobin’s (1966) study of the passive, discussed below in Section 5.3.4). Since Ferreira’s participants were told to construct sentences, not read them, her results do not bear on this possibility. 4.4

Summary

In this chapter we established that subject animacy has measurable effects on adult speakers’ representation of sentence structure. Specifically, adults have an easier time construing an inanimate subject as a displaced argument than an animate subject, which tends to be construed as the external argument of the main predicate. Clark and Begun’s (1971) study showed that speakers rate NPs as “better” subjects if they are higher on the Animacy Hierarchy, and they improve nonsensical sentences by making the subject more animate. The sentence processing literature shows that speakers are generally unsurprised to find out that an inanimate subject is really the head of a reduced (object) relative clause, while they are surprised when an animate subject turns out to have this role (pace Ferreira and Clifton (1986)). Moreover, the well-established asymmetry between Subject Relative and Object Relative clauses is mitigated when the head of the object relative clause is inanimate (though see caveats about these results in the discussion above). More specific to the raising/control classes of verbs, in my fill-in-the-blank study (Becker, 2005) I showed that animate subjects with infinitive complements generally elicited control verb responses (His campaign manager – to remain a problem. . . ), but that this preference could be overridden by using an inanimate subject (The extramarital affair – to remain a problem. . . ). The novel verb learning study extended the scope of this influence from eliciting known verbs to novel verb categorization. With these results framing the target state that young children are moving toward, let us now examine the development of the concept of animacy in children, and evidence bearing on the role that NP animacy plays in children’s acquisition of grammar in general, and opaque constructions in particular.

5 Animacy and children’s language

So far we have established the following. (a) There are constructions in natural language in which the surface subject of a declarative sentence is not semantically related to its adjacent predicate, and it is therefore displaced with respect to its logical place in the sentence. These constructions involve raising verbs, tough-adjectives, and unaccusative verbs. I have referred to these constructions as “opaque” constructions, in part because the structural relationship between the main clause subject and main clause predicate is not transparent from their surface adjacency, and in part because these constructions bear a striking resemblance, on the surface, to constructions that have a very different underlying structure (those involving control predicates or unergative verbs). In these opaque constructions, to a much greater degree than canonical sentences, languages permit the displaced/derived subject to be inanimate. (b) The linguistic marking or encoding of animacy is found quite broadly in human language, and animacy is universally linked to argument structure in that animacy and agenthood generally imply subjecthood, while inanimacy generally implies non-subjecthood, and, furthermore, an inanimate subject generally implies a derived subject. Finally, (c) even in languages that lack extensive morphological marking of animacy, such as English and Dutch, there are effects of animacy on the processing of language in adult speakers. Specifically, adults more readily treat inanimate syntactic subjects, as opposed to animate subjects, as themes – either as the head of an object relative clause or as the subject of a raising verb. At this point we turn to children’s use of animacy in the acquisition of sentence structure, and the acquisition of displacing predicates in particular. We will see quite robust evidence that young children take an inanimate subject as an implication that the structure involves displacement, and therefore that the main predicate is a displacing predicate. In order for the animate–inanimate distinction to inform the process of learning language, this distinction must be conceptually available to young children at the appropriate age. If, as proposed by Piaget (1929) (and still 156

Development of the animacy concept

157

widely believed among non-psychologists), children are animistic, meaning they attribute animate properties such as sentience and feelings to inanimate things, then inanimate NPs could not be of much use in constraining the representation of grammatical structures. That is, if tables and books are alive and sentient, then hearing a sentence like The table gorps to be heavy or The book is daxy to read could not tell children that in these sentences table and book are not external arguments. So in order to make the case that this kind of learning procedure is even plausible, we must establish that children are not animistic, and that they have quite good abilities to correctly distinguish animate from inanimate things in the relevant properties. First we will review some of the literature on children’s perceptual and conceptual development of the (non-linguistic) animate–inanimate distinction. We will then focus on language and look at previous examinations of children’s use of animacy in acquiring argument structure more generally. Finally, turning to the opaque constructions, we will look at experimental evidence showing how children can exploit the occurrence of inanimate subjects to discover the non-canonical underlying structures of opaque constructions and, relatedly, the syntactic properties of displacing predicates. 5.1

Development of the animacy concept

As any parent of a young child can attest, children are fascinated by animals. When there are pets in the home children are drawn to them; they interact with animals and laugh at their playful antics, just as they do with siblings and other human family members. Children’s words for animals come in early and form a large portion of their early vocabularies. At age 12 months one of my own children, with a vocabulary of about ten words, had just one general-purpose human term (mama), three object terms (ball, water/milk, and banana) and six distinct animal terms (dog, bird, chicken, frog/rabbit, lamb, and cat). Although many of her words took the form of the English onomatopoeia for the animal’s sound rather than the English noun label ([ba:] for ‘lamb’), her semantic reference for each label was clear. Gleason et al. (2009) note that it is not at all unusual for children’s early vocabulary to include a preponderance of animal terms. According to these authors, 50% of 14-month-olds have common animal words like duck, bear, cat, and dog in their lexicons, and by 30 months the majority of English-speaking children understand words for more unusual animals such as alligator and zebra. In their own study of spontaneous speech production, Gleason et al. found that children between ages 2 and 5 years use animal terms in between 4 and 14.5 per 100 utterances, a rate that rivals the

158

Animacy and children’s language

rate at which children talk about vehicles and other commonly encountered objects. Thus, animals figure prominently in a child’s world. What I’d like to establish in this section is that even before acquiring language, children have quite striking abilities to cognitively distinguish agentive, sentient entities from non-agentive, non-sentient things. In general this dichotomy refers, in practice, to humans vs. inanimate objects, though we will also consider non-human animals. That is, in order for the animacy distinction to do the work of helping children acquire argument structure in general, and the structure of opaque constructions in particular, it is important to justify the claim that children have the ability, early in development, to perceptually and conceptually distinguish animate from inanimate entities. An enormous literature is devoted to experimental studies of infants’ and children’s perception and conception of animacy. A brief but fairly comprehensive overview of the psychological literature can be found in Gelman and Opfer (2002), and a more detailed one in Rakison and Poulin-Dubois (2001) (see also Legerstee (2001)). Here I will highlight studies that indicate an early ability to distinguish living from non-living things along different dimensions: featural properties, behavioral properties (e.g. movement and causation), and intentional states. Despite the enormous breadth this literature covers, there are several gaps in our knowledge. One is that many of them contrast humans with inanimate objects, so that little is yet known about how infants regard non-human animals in terms of certain animate properties, in particular intention. Another is that some of the experiments that impute babies with the ability to perceive causality use experiments with only inanimate objects (e.g. moving balls), thus muddying the ability to determine the relationship between causality, agency, and animacy in children’s representations. Despite these gaps, it seems well established that even very young babies can perceptually distinguish animates from inanimates, have expectations about how people vs. objects move, and form categories of animates even in the absence of movement information or shape similarity. 5.1.1 Featural properties of animates One of the most salient physical features of biological animals (humans and non-human animals), as opposed to inanimates, is that animals typically have a face, and very young babies show heightened attention to faces. Within the first month of life infants prefer to look at (schematic) drawings of faces than drawings of non-faces (Fantz et al., 1975). And even within the first hour of life neonates attend significantly more to moving pictures of faces with a normal

Development of the animacy concept

159

configuration (eyes aligned symmetrically above the nose and nose centered above the mouth) than faces with a mixed-up configuration or a blank outline of a head shape (Goren et al., 1975; Johnson et al., 1991).1 Curiously, this preference for a normal face configuration actually disappears after about 30 days, only to re-emerge several months later. Johnson et al. (1991) speculate that an initial face perception system is in place immediately after birth to draw newborns’ attention toward faces. This mechanism lays the neural groundwork for a second, more sophisticated, adult-like face recognition system which emerges later in development. As Gelman and Opfer (2002) note, other physical attributes may also be recruited in discerning the difference between animate and inanimate kinds: the presence of legs vs. wheels, for instance, or shape features such as rounded edges as opposed to hard angles, or textural differences, such as fur vs. metal. Comparatively less experimentation has been devoted to exploring infants’ reliance on these other physical properties for distinguishing animates from inanimates. However, a study by Massey and Gelman (1988), which we return to below, shows that preschoolers rely on some of these other physical attributes for distinguishing animate from inanimate pictures in the absence of any evidence about movement properties. 5.1.2 Behavioral properties of animates A major behavioral property of animate entities, distinct from most inanimate entities (excepting – at least, superficially – vehicles) is the ability to engage in self-propelled motion. A number of experiments with young babies have demonstrated a difference in babies’ expectations of how objects vs. people move. Woodward, Phillips, and Spelke (1993) conducted one such experiment, in which 7-month-olds were habituated to a scene in which either two objects (large geometric shapes) or two people moved behind an occluding screen. (Object vs. person was manipulated between participants.) For example, a large rectangular box would enter the scene from the left (moving on small wheels, pushed by a hidden person), and go behind a screen from the left side of the screen. At the same time, another object (a large cylinder) would be partially visible at the right edge of the screen (i.e. partially occluded by the screen). After a short pause, the second object (cylinder) would move out from behind the screen and exit the scene to the right-hand side. In the person condition, two adults moved in the same way as the inanimate shapes. The point 1 The pictures moved slowly in an arc in front of the infant; the components of the face did not

move with respect to one another.

160

Animacy and children’s language

of the occluding screen was to prevent the infant from seeing what happened in the middle of the scene; that is, the infant could not see whether the objects or people made contact or not. After habituation the babies were shown a test scene, which involved the same objects or characters and same movements, but this time without the occluding screen. Each infant witnessed three videos involving contact between the two entities (i.e. the shapes or people collided in the middle of the scene) and three videos with no contact: the first shape or person approached the second in the middle of the scene but stopped short of contact, leaving one foot of space between them, before the second object or person moved off screen. Previous experimental work had shown that babies at this age (and younger; Spelke (1991)) have expectations about how inanimate objects move through space: namely, objects move along continuous trajectories, they do not occupy the same space at the same time, and they move only if propelled by some external force. Given these findings, the expectation in Woodward et al.’s experiment was that babies would look longer (indicating surprise) at the test videos in which the geometric objects did not contact one another in the middle of the scene, but the second object nevertheless moved as if it had been pushed by the first. This expectation was borne out. The real question was whether these babies would respond the same way to the people videos. In fact, the 7-month-olds responded very differently. They looked longer at the videos in which the people collided in the middle of the screen than the videos in which the first person approached the second but stopped short of contact. The results support the view that babies at this age reason differently about the movement properties of people vs. objects.2 Along similar lines, Poulin-Dubois et al. (1996) showed 9- and 12-montholds scenes in which either a robot or a person was given a verbal command. Babies attended longer to the scene in which the robot then moved in response 2 Seemingly in contradiction to Woodward et al.’s result was Golinkoff and Kerr’s (1978) finding

that 14–24-month-olds did not appear to discriminate scenes in which a chair pushed a man from scenes in which a man pushed another man. Using heart rate as a measure of habituation, Golinkoff and Kerr found that babies dishabituated when the agent–patient roles were reversed (A pushes B in habituation and then B pushes A in the test phase), but they did so both when both A and B were human and when A was a human and B was a chair. In fact, the degree of dishabituation was slightly greater when both A and B were human. Although this result appears to run counter to the view that children expect only animate entities to act upon others, it is possible that there is something inherently more interesting about two humans contacting one another. Recall that in the Woodward study, infants also looked longer when two humans were seen to bump into each other.

Development of the animacy concept

161

to the command (i.e. without being caused to move through physical contact with another entity), than to the parallel scene in which a person moved in response to the command. The studies just cited point to very early expectations about certain physical properties of animate (qua human) vs. inanimate entities, and the ability to discriminate novel animate from inanimate objects on the basis of movement behaviors. Two important questions remain open, which we will address in turn. One question is whether (and at what age) babies understand goaldirected and causal movement. That is, are babies’ expectations about the motions of people and objects based on understanding that people often act for a particular reason (e.g. they reach for an object because they want the object), and can cause events to occur, but objects generally cannot? The second question, related to the first, is whether (and at what age) babies discriminate on the basis of a conceptual distinction rather than a perceptual one. Rakison and Poulin-Dubois (2001) define conceptual knowledge as meaning that babies “have disassociated perceptual information from its source and converted it into some kind of abstract representational format” (p. 213). The two questions are related in that causality is a quality that is inferred, but not directly observed, from an event. Thus, evidence that babies have knowledge of causality could support the view that they have at least rudimentary conceptual knowledge (as opposed to merely perception) of animacy. A study by Woodward (1998) suggests that 5-month-old babies attribute goal-directedness to humans but not inanimate objects. Woodward showed babies a scene in which a hand moves across a stage to pick up one of two objects on the stage. After the babies had habituated to this scene, the scene was changed so that the position of the objects was reversed. Now the hand either moved in a different path but picked up the same object, or the hand moved in the same path as before but now picked up the other object. The babies looked longer (they were surprised) when the hand moved in the same path but picked up the different object. The important comparison is the further manipulation in which Woodward habituated another set of 5-month-olds to the same scene except that this time, instead of a human hand doing the reaching, it was a stick with a “multifingered sponge” on the end of it. In this case, the babies showed the opposite response, looking longer when the stick/sponge reached for the same object but took a different path to do so. The conclusion is that babies at this age attribute goal-directedness to human actors, but not to non-human actors. There is additional evidence that babies’ inference of causation arises by age 6 months. Leslie and Keeble (1987) showed 6-month-olds scenes in which

162

Animacy and children’s language

a red square entered from the left, touched a green square in the middle, and then the green square moved to the right. Half of the babies saw this scene, displayed such that the green square moved immediately upon being contacted by the red square (they called this the “direct launching” condition; the red square appeared to “cause” the green square to move). The other half of the babies saw the same scene, except that there was a half-second pause before the green square moved (they called this the “delayed reaction” condition; the green square moved as in the first scene but not apparently by direct causation from the red square). These scenes were presented to the infants until the infants habituated to them. Following habituation, the babies were then shown a test scene which involved the film of the habituation scene being played backwards. That is, the green square enters from the right, touches the red square, and then the red square moves to the left (with or without a pause, according to the condition). The result was that babies in the “direct launching” condition looked significantly longer at the test scene compared to the habituation scene, while the babies in the “delayed reaction” condition did not. What does this mean? The habituation scene and the test scene were identical in their spatiotemporal features: the test scene was merely the habituation scene run in reverse. Thus, they were perceptually identical. And the two conditions were exactly alike except for the slight delay before the second block moved. Leslie and Keeble reasoned that babies should look longer at the reverse of the “direct launching” scene if they had encoded causality as a component of the event. In other words, the difference between habituation and test movies in the “direct launching” condition was that in the habituation scene the red square was the causer, while in the test scene the green square was the causer. Since this causative role was absent in the “delayed reaction” condition, no inferential shift was required in apprehending the reversed scene. These findings suggest an ability to infer causality within the first year. However, care should be taken in drawing conclusions about 6-month-olds’ concept of animacy based on this result. Although Woodward’s (1998) study directly compared human and non-human agents, the stimuli employed in Leslie and Keeble’s study involved two moving inanimate objects (squares), not a contrast between people and inanimate objects. The stimuli were modeled on a series of experiments done by Albert Michotte on adults, in which people viewed scenes in which blocks or circles moved in different ways (Michotte, 1963). Adult participants ascribed animate properties to those objects that appeared to “chase” or “push” the other object, or to those that appeared to “hide” from another object and so forth. So the results of the Leslie and Keeble experiment

Development of the animacy concept

163

suggest that 6-month-olds make a Michottian inference about one dimension that is attributable to animates (causality), but does it mean that these babies conceptualize the “causer” block as animate? This remains an open question. Let us turn now to the question of when babies develop a concept of animacy, as opposed to distinguishing animates from inanimates on a perceptual basis. To see the distinction using one of the previously cited experiments, the explanation I offered for the Woodward et al. (1993) study described above was that 7-month-olds looked longer at the contact scene than the non-contact scene for people, but longer at the non-contact scene for objects, because these babies already know that people are a “kind” of animate entity, and that one of the properties of animate entities is self-propelled motion. On the other hand, inanimate shapes are a “kind” of inanimate entity, and such kinds do not have the capacity for self-propulsion. This kind of “rich” interpretation is espoused by Rochel Gelman, Elizabeth Spelke, Susan Carey and others (Gelman and Spelke, 1981; Gelman, 1990; Gelman et al., 1995; Carey, 2009). This view contrasts with what Carey (2009) calls the “lean” interpretation, according to which babies respond differently because of perceptual differences between the observed scenes. We will return below to Gelman’s and Carey’s specific arguments, but first let us look at one experiment with infants under one year of age which attempted to address directly the conceptual/perceptual distinction. Mandler and McDonough (1993) investigated this issue with 7-, 9-, and 11-month-olds. Using a manual habituation task, Mandler and McDonough showed babies a series of objects belonging to either the animal category (dog, fish) or vehicle category (car, airplane). Children were allowed to touch, play with, or otherwise explore the objects. After eight rounds of showing the child four different objects of the same category (e.g. four different types of dogs– the authors aimed for maximally perceptually different exemplars), one at a time, the child was then given a fifth exemplar of the same category, followed by an exemplar of a different category (e.g. a car). The amount of time the child spent attending to or exploring the test items was measured. The authors reasoned that if these babies had a conceptual category for the items seen, then if the new test item fit into the habituated category for the baby, she should spend as much time attending to the novel item as to the items in the familiarization phase (as measured at the end of the eight-round habituation period). In contrast, if the new object did not fit into the habituated category, babies should spend more time attending to it. According to this measure of conceptualization, Mandler and McDonough found that 9- and 11-month-olds conceptually distinguished animals from vehicles, cars from airplanes, and cars from motorcycles, but not dogs from

164

Animacy and children’s language

fish (or dogs from rabbits, in another version of the task). Seven-month-olds demonstrated a weaker tendency in the same direction as the older babies. The authors concluded that by the end of the first year (and beginning sometime around the second half of that year) babies can distinguish conceptually between animals and vehicles, though they do not form conceptual categories within the larger domain of animals until later. Importantly, Mandler distinguishes her findings from those showing a very early ability to distinguish animals perceptually. Eimas and Quinn (1994) showed that even 3-month-olds can perceptually distinguish certain animals: horses as opposed to zebras or giraffes, and cats as opposed to tigers and male lions. But the methodology used by Mandler and McDonough, the objectexamination task, does not measure perceptual similarity but rather conceptual category membership. This methodology had been used previously with the same age group to assess their categorization of objects where the exemplars in each category bore no perceptual similarity (“kitchen things” vs. “bathroom things”). Thus, Mandler argues that babies can perceptually distinguish different kinds of animals early in the first year, but they do not categorize them into different concepts until towards the end of that year. Furthermore, Mandler and McDonough included a trial with 9- and 11-month-olds that measured their conceptual categorization of birds and airplanes, the exemplars of which were highly similar in their perceptual features. The babies nevertheless treated these two types as belonging to different categories. Mandler concludes that at age 11 months babies have a concept of “animal” as distinct from vehicles, and a perceptual but not a conceptual distinction of kinds within the category of animals.3 A further argument for young children’s conceptual category of animate entities comes from Gelman, Durgin and Kaufman (1995). The in-principle part of Gelman et al.’s argument is that properties of object motion are ambiguous indicators of whether the moving object is animate or inanimate, and so perceptual properties of object motion cannot be the primary source of criterial information about whether something is animate or not. Similar to Michotte’s experiments, Gelman et al. showed adults computer-generated scenes of filled circles (“sprites”) moving along different kinds of paths, changing or not changing direction, and contacting or not contacting obstacles in the scene. Participants were then asked to describe what they saw and interpret the scene in terms of the actor(s) and event(s), and then to rate how “alive” the moving 3 See Rakison and Poulin-Dubois (2001) for criticisms of the view that babies have a concept of

animals within the first year of life.

Development of the animacy concept

165

object or objects were. Gelman et al.’s interpretation of the result was that while adults were influenced by certain aspects of the sprites’ motions (for example, they generally judged objects that moved along a straight or V-shaped trajectory to be “less animate” than objects that moved along a curved trajectory), their judgments of animacy were not unambiguously determined by these motion properties. For instance, one participant described an S-shaped trajectory (which overall triggered significantly more judgments of animacy than straight trajectories did) as “a smoke particle as seen from above in a wind tunnel.”4 Gelman et al. also support their argument with evidence that 3-year-olds make reliable designations of novel entities as animate or inanimate, even when the basis for their designation is only a static photograph. Massey and Gelman (1988) showed 3- and 4-year-olds pictures of unfamiliar animals, like echidnas and iguanas, as well as statues and unusual vehicles and asked the children whether each thing could go up or down a hill by itself. The 4-year-olds were at ceiling in responding that mammals and non-mammalian animals could go up or down a hill by themselves, that vehicles could go down but not up a hill, and that machines and statues could do neither, even if the statues had “feet” or otherwise resembled a human form. Interestingly, the 3-year-olds in their study deviated from this response pattern only in that they were less likely than 4-year-olds to say that an animal (mammalian or non-mammalian) could go up or down a hill by itself. However, the reason for their response was typically that they deemed the animal “too small” to manage a big hill by itself, or that its feet were not visible in the picture. Gelman et al. conclude that 3- and 4-year-olds have a concept of “animate thing” that does not require evidence of (self-propelled) motion, though self-propelled motion does appear to be entailed by the children’s designation of something as animate. Of course, 3- and 4-year-olds might be a different cognitive kettle of fish from 7-month-olds, and so it is unclear whether the Massey and Gelman findings bear directly on the question of whether a conceptual or a merely perceptual mechanism is at play in the earliest discriminations of animacy. Nevertheless, I find myself persuaded by the research pointing to children’s concept of animacy being in place by the beginning of the second year of life. 4 The authors argue that their findings are in agreement with the findings of Stewart (1982, 1984),

whose work is generally cited as favoring the view that adults make their judgments of animacy on the basis of object motion properties. However, Gelman et al. point to additional unpublished studies by Stewart that, they claim, illustrate the ambiguity of motion cues, as well as a statistical reanalysis of Stewart’s published data. On their reanalysis Stewart’s results appear to support the view that motion cues are ambiguous.

166

Animacy and children’s language

That animate entities hold a special place in children’s cognitive apprehension of the world is perhaps also supported by evidence that babies’ relationship to animate things is multifaceted. Not only do young babies attend differentially to scenes in which inanimate (as opposed to animate) entities move without an external causal force or in response to a command, but babies also respond differently to people vs. objects. In a study by Legerstee (1994) 4month-old infants were presented with either an object or an unfamiliar female person, which then went behind an occluding screen. When it was the object that was hidden, babies reached for it; when it was a person who was hidden, babies vocalized to the person. In another study, Legerstee et al. (1987) studied infants’ responses to their mother, an unfamiliar female adult, and a doll, over a period of several months starting when the babies were only 3 weeks old. The authors found that by about 7 to 9 weeks of age, the babies showed significant differences in their responses to the humans vs. the doll. They gazed at and reached for the doll, whether it was moving or not, significantly more than the people, and they vocalized to the people, whether moving or familiar or not, but never vocalized to the doll. A further issue to consider is whether, and at what age, children include in their concept of “animate thing” the full range of characteristics attributed to animates by adults. For instance, adults generally agree that animate entities not only move by themselves and have causation, but also have biological properties such as the ability to grow and reproduce, to communicate, and to have some intentional states. Thus, babies might have a concept of “animate thing” that includes some of these characteristics but not all of them. A broader discussion of this question is taken up in Section 5.1.4, but first let us examine children’s knowledge of entities’ capacity for intentional states in particular. In order to assess children’s understanding that animate entities have intentional states, we must look at studies with slightly older children. 5.1.3 Intentional properties of animates What does it mean for a child to understand that other people (and animals) have intentional states? It means understanding that they have desires, thoughts, beliefs, and emotions. How can we measure whether children understand this? One measure that has been used is the false belief task (Wimmer and Perner, 1983; Wellman, 1985). In a well-known version of this task the child observes a scenario in which two characters enter a room, and character A places an object in a closed location and then leaves. While character A is absent character B moves the hidden object to a new, hidden location. Character A returns. The child is asked where character A will look for the object.

Development of the animacy concept

167

If the child responds that A will look in the original location they are said to pass the task, because presumably they are able to represent A’s mental contents (beliefs) as being different from the actual state of affairs. If the child responds that A will look in the new location they are said to fail the task, because A could not have known about the change of the object’s location. Presumably, in the child’s mind character A’s beliefs are veridical with respect to the actual state of affairs. In traditional versions of this task children typically fail until about age 4. While this finding is robust (Wellman et al., 2001; Liu et al., 2008), certain manipulations of the task have produced passing results at earlier ages, and some researchers have questioned the use of false belief as the criterion for understanding intentional states more broadly. While I remain agnostic on the matter of when a representational theory of mind truly emerges, in what follows I will review some studies showing that well before age 4 children have some knowledge of others’ intentional states, and that very young children attribute intentional states to humans. Three studies are particularly informative. One study by Clements and Perner (1994) found that although most children before age 4;5 (4 years, 5 months) gave an incorrect verbal response in the traditional false belief task, a large majority (88%) of them between age 2;11 and 4;5 looked at the correct location before giving their answer. That is, when prompted with a leading comment, “I wonder where Sam is going to look (for the cheese)?” these children looked toward the location where Sam had placed the cheese. But when asked “Which box will he open (first)?” most children (71%) responded with the new location, the place to which the cheese had been moved. Children younger than 2;11, with one exception, both looked to the new location and responded verbally that Sam would look in the new location. Clements and Perner argue that this pattern demonstrates an implicit understanding of false belief by age 2;11, though not an explicit understanding of it until age 4 or so. A study by Onishi and Baillargeon (2005) points to an even earlier implicit understanding of false belief. In their experiment, 15-month-olds viewed a scene in which a woman plays with an object and then places it inside a green box. The infants then saw one of four scenes: (a) the woman watches while the adjacent yellow box moves back and forth, but the object remains in the green box (True Belief-green); (b) the woman watches while the object moves (by itself) from the green to the yellow box (True Belief-yellow); (c) a screen goes up blocking the woman’s view, and the object moves from the green to the yellow box (False Belief-green); (d) the woman watches while the object moves

168

Animacy and children’s language

from the green to the yellow box, but then the screen goes up and the object moves back to the green box without the woman seeing (False Belief-yellow). Following these scenes (babies were randomly assigned to a condition) each baby then saw a scene in which the woman reaches for the object in one of the boxes (either the green or the yellow one). Onishi and Baillargeon found that babies uniformly attended longer when the woman searched for the object in a location that was incompatible with her belief about where the object should be. That is, if the woman held a true belief about the object being in the green box, the babies looked longer if she then searched in the yellow box than if she searched in the green box. And if the woman held a false belief about the object being in the green box, the babies still looked longer if she searched in the yellow box, even though that is where the babies knew the object was. These findings are compatible with the possibility that infants have learned, perhaps through observation, that people generally look for things where they last saw them, or that babies know that people “know” things about the world, but they do not yet understand that people’s knowledge is their own interpretation (representation) of how things are. But even this kind of explanation attributes fairly sophisticated reasoning to 15-month-olds about other people’s perceptions and knowledge of the world; that is, to their mental contents. In a very different kind of study with slightly older children, Meltzoff (1995) tested whether 18-month-olds would infer the intentions of a person or a machine, even if they had not seen the intended action carried out. Children were divided into four test groups. One group saw a person demonstrate some action using each of five novel toys (e.g. pulling apart a dumbbell-like object with blocks on the ends of the rod); another group saw the person attempt to do this action but fail (e.g. pulling on the ends of the dumbbell but failing to pull it apart because one hand slipped off the end). The other two were control groups, who saw the novel toys but neither the target action nor an attempt at the target action. After demonstration (or attempted-but-failed demonstration), each object was placed within reach of the child, and children’s actions with the object for a 20-second period were recorded (they were not specifically instructed to do anything with the object). The result was that children in both test conditions (demonstration of the target action and attempt at the target action) responded by imitating the person’s target action, even if the target was never actually witnessed (as in the attempted-but-failed condition). Children in the control conditions were significantly less likely to perform the target action with the objects. The control groups established that children did not spontaneously perform the target

Development of the animacy concept

169

actions when presented with these particular objects. Rather, 18-month-olds performed these actions when they saw a person do it, or try to do it. In the latter condition the children had to infer the intention of the person to complete a particular action. A second experiment replicated the first in two conditions: one which replicated the first experiment by having a person demonstrate a failed attempt at an action (pulling apart the dumbbell), and the other which involved a mechanical device that performed the same action. That is, the mechanical device had two vertical “arms” that had horizontally-oriented pincers at the ends which could grasp an object. One end of the dumbbell was held by one set of pincers, while the other arm moved toward the dumbbell with open pincers and appeared to “grasp” it. This arm then pulled away and the pincers slipped off without pulling the dumbbell apart, the same way the human fingers did in the human condition. Thus, the two events were functionally the same, but differed by the animacy of the manipulator. In this experiment the two groups differed substantially in their rate of imitation: 60% of the children in the person condition imitated the target event (pulling the dumbbell apart), while only 10% of the children in the machine condition did, in spite of the fact that neither group of children actually saw the target action. This result provides quite striking evidence that 18-montholds infer intentions of people (that is, they infer that the human intended to pull the dumbbell apart), but not of inanimate objects. Although Meltzoff’s experiments do not necessarily point to 18-month-olds having a representational theory of mind (he does not claim they do), they certainly suggest the understanding that people have intentions, and what those intentions are, in contrast to inanimate objects. 5.1.4 Further conceptual change To say that children by the age of 2 years have a conceptual distinction between animate and inanimate things is not equivalent to saying that 2-year-olds have a completely adult-like understanding of biology, or even that they have an adult-like ontological category of “living thing.” That is, their early success in discriminating the categories of animate and inanimate does not preclude further conceptual change during childhood. As Rakison and Poulin-Dubois (2001) point out, children do not have an understanding of some of the internal biological properties of animates (such as laying eggs, or otherwise reproducing) until at least preschool age, and Carey (1985) argues that even 4- and 6-year-olds reason about biological properties of animals quite differently than older children and adults do.

170

Animacy and children’s language

Carey (1985) illustrates examples of children between ages 4 and 10 years imputing intentions to plants or the sun, claiming that cars and other objects are alive, and so forth, all of which appear to be strongly at odds with the sophisticated distinctions we saw that even small infants are capable of. These verbal reports from children in middle childhood were, of course, the basis for Piaget’s contention that children were animistic (Piaget, 1929; Laurendeau and Pinard, 1962). Carey offers a number of explanations for these responses that reconcile them with the facts about younger children. For example, the Piagetian method for extracting these judgments was an interview process in which children were asked whether something was “alive.” But a child’s ability to answer this question hinges on their understanding of the word alive, and it is not established whether children’s lexical entry for alive contains the same semantic content as the adult’s. In fact, the word alive is actually ambiguous as to the precise dimension of an entity it picks out. That is, the quality of being alive can be contrasted with a number of not-alive qualities: alive vs. dead (a living horse is alive but a dead horse is not, though it used to be), animate vs. inanimate (a horse is alive but a table is not, and it never was), real vs. imaginary (a horse is alive because it is real, but a ghost, for example, is not), and real vs. representational (a real horse is alive but a picture of a horse is not). Moreover, plants would be described by adults as “alive” but they bear none of the prototypical characteristics of animate things: they do not move on their own (save for subtly orienting toward the sun, which small children are unlikely to have witnessed), they lack prototypical animate features like faces, and they lack intentions as far as we know. In short, plants are alive but inanimate, so the label “alive” does not equate with “animate”. The problematic word alive trips young children up and makes them appear animistic in their beliefs, obscuring their quite accurate understanding of how animates and inanimates differ in their behavior and essential properties. Nevertheless, there are real differences between preschool-age children and adults in their knowledge of animacy. Extensive interview studies by Susan Carey and Frank Keil have demonstrated that children undergo enormous change between ages 4 and 10 in their concept of what it means for something to be a living thing, what the internal organs of people and animals are like, and about the “essence” of natural kinds like skunks and raccoons (Keil, 1979, 1989). Four- and five-year-olds typically deny that people are a kind of animal (their category of “animal” excludes humans), and they lack an umbrella ontological category of “living thing” so that plants are deemed unable to “be sick” or “grow” in contrast to animals and people. By age 10 children have

Development of the animacy concept

171

overcome these misconceptions and have considerably more adult-like knowledge of what it means for something to be alive. (See also cross-cultural studies by Hatano and Inagaki, some of them summarized in Hatano and Inagaki (1999), on children’s developing knowledge of biological properties of people and animals.) One final note about young children’s apparent “confusion” over the animate–inanimate distinction is that children will often act as if toys have certain features of living things, for example, making a doll eat, sleep, or go to the bathroom. However, this kind of pretend play should not be taken as an indication that children really think the doll is alive. Children are capable of pretend play by the age of 18 months (Fenson, 1984; Leslie, 1987), if not earlier. Moreover, the “animation” of inanimate things, as well as the “humanization” of animals (for instance, by making them speak, wear clothing, etc.) is ubiquitous in children’s literature. In one classic example, the story “The Little Engine that Could” (version by Watty Piper, 1930), the toys and dolls that are the train’s cargo are depicted throughout the book as being alive, and they converse with the equally animized train engines. Not only do the toys and engines speak, but they are explicitly described with emotive predicates (“She was a happy little train . . . ,” “How sad the little train and all the dolls and toys felt!” “Then indeed the little train was very, very sad, and the dolls and toys were ready to cry”) and given otherwise human-appropriate descriptions (“[The little blue engine] thought of the good little boys and girls . . . ,” “I think I can”). This nonveridical animization is perhaps even more prominent in children’s television programs and movies aimed at children. I make this observation for two reasons: first, depictions of animized inanimates, and humanized animals, are hardly rare in children’s experience, so if children in turn animize or humanize their own toys, this behavior should not (necessarily) be blamed on immature conceptions of animacy. Secondly, the fact that in spite of robust exposure to depictions of talking animals and objects children nevertheless have reasonably accurate distinctions between animate and inanimate things (to wit, children’s responses to what could go up a hill on its own in Massey and Gelman’s study) suggests a powerful native expectation on the part of children that real humans (and perhaps animals) are special in their ability to move on their own and have causative and intentional properties, while outside of books and television, real-world rocks and chairs do not have these properties. Children’s typical response to puppets (animized inanimates) is also revealing: my personal experience is that children are either fascinated or terrified by them, and sometimes both, suggesting that they represent something unexpected.

172

Animacy and children’s language

5.1.5 Agency In Chapter 3 we talked about the grammatical notion of agency: agent is the name given to the thematic role for the NP in the sentence that acts upon others, initiates some event, or causes or effects a change of state. It is no accident that the same term is used in the field of psychology to talk about similar abilities, but here predicated of actual entities as opposed to NPs. On what basis do young children attribute agency to an entity? Carey (2009) suggests two sources of evidence for agency. One source is featural: agents, being living things, generally have a face, while non-agents, being inanimate, generally do not. As noted above in Section 5.1.1 babies attend to faces from an extraordinarily young age, and they use faces or other human features (hands) in attributing goal-directed motions to people. Recall that in the Woodward (1998) study, 5-month-olds attributed goal-directedness to human hands, but not to a stick with a fingered sponge on the end of it. The other source according to Carey is motion. Although Massey and Gelman (1988) argue that motion by itself is neither necessary nor sufficient as a cue to animacy, there is substantial evidence that, when present, children and adults use motion cues in drawing inferences about the relative agency of an entity. This can be seen in the Michotte experiments with adults, in which adults attribute human-like properties of chasing, fleeing, and so forth to moving circles, and in the Leslie and Keeble (1987) study, which was a Michotte-style presentation with infants. That both facial features and motions support attribution of agency is shown by a study by Johnson, Slaughter, and Carey (1998). In this study, researchers introduced 12-month-olds to a robot that either had facial features (eyes that flashed) or did not (a light on top of the robot’s head flashed instead), and either responded contingently to the infant’s vocalizations or movements (i.e. flashing in response to vocalizations and moving its “arms” in response to infant movements) or responded noncontingently (flashed and moved in an apparently random pattern with respect to the infant’s vocalizations and movements). After a one-minute exposure period, the robot “turned” to orient itself at a 45-degree angle. Babies who had interacted with a robot that had a face or one that responded contingently turned to look in the direction the robot turned. Babies who had interacted with a faceless and noncontingently moving robot did not do this. Carey notes that “[a]gent detection is likely to be a problem of such significance for humans that evolution has built more than one redundant mechanism to contribute to its solution” (Carey, 2009, p. 187). I will continue to single out ‘animacy’ rather than ‘agency’ as the relevant feature for acquiring displacing predicates, for the reasons given in

Development of the animacy concept

173

Section 3.1.2, but of course there is significant overlap between the two terms and concepts. 5.1.6 Summary The studies described in this section all point to the animate–inanimate distinction developing extremely early in life. While the precise origin, characterization, and developmental trajectory of children’s concept of animacy is still debated, for our purposes it will be sufficient to have established that by the time syntactic acquisition begins in earnest (let’s say, 18 months) children have at their disposal a reliable means of differentiating animate from inanimate entities perceptually, and categorizing entities according to the concepts that accord with an animate/inanimate distinction. Crucially, before the end of the second year of life, children distinguish animate from inanimate entities on the basis of not only featural, but also behavioral and intentional properties, and they ascribe the properties of agency (and the related properties of intention and causation) to humans but not inanimate objects. Illustrative of young children’s keen awareness of what is alive is the following conversation with a girl who is not yet 4 years old, as reported to me by her mother.5 (1)

Mother: River, what are your ponies doing? Are they talking to each other? Are they playing? Child: (incredulous look) They’re not doing anything. They’re not alive. They’re play toys.

There was a second purpose of this section. We will see some evidence in the following sections (in particular Section 5.3.1) that children sometimes judge sentences with an inanimate subject and an intentional kind of predicate (e.g. want, try) to be well-formed. It has been objected (e.g. Hirsch and Wexler (2007)) that such a judgment stems from children’s animistic tendencies. However, my contention is that given young children’s sophisticated understanding of animacy, their linguistic deviances should not be taken as resulting from any confusion over animacy, or the mistaken belief that inanimate things are alive. Rather, I hope the details of this section have sufficed to allay this prejudice, and instead render more plausible the idea that children’s acceptance of sentences like The flower wants to be pink results from a more flexible lexical entry for the verb, as I will argue. 5 M. Fee, personal communication (Facebook post) 5/23/12.

174 5.2

Animacy and children’s language Children’s use of animacy in learning argument structure

The previous section established that by the time children are in the business of developing syntax, they distinguish conceptually between animate and inanimate entities, and they attribute intentional properties to humans but not to inanimate objects, including machines. The purpose of the present section is to look at how this cognitive distinction plays a role in the acquisition of sentence structure. It is one thing to show that 18-month-olds possess the cognitive distinction between animate, intentional entities and inanimate, non-intentional entities. It is another to show that this distinction manifests in a linguistic distinction between animate and inanimate NPs and the sorts of predicates that can apply to them. Gagliardi (2012) has demonstrated that children do not always rely on the semantic feature of animacy in noun class categorization, even when this cue is statistically more reliable than other cues, such as phonological ones. With respect to our question of predicate categorization, children might well understand that rocks and robots are incapable of having desires, and yet they might find it acceptable to apply predicates such as want or like to them. A study by Schwartz (1980) purported to show that children fail to restrict animate or intentional types of predicates only to animate NPs, and instead allow such predicates to apply to inanimate NPs. Schwartz’s study included both a Piagetian questionnaire (after Laurendeau and Pinard (1962)) to assess children’s metacognitive awareness of which entities were “alive,” and a sentence judgment task in which anomalous sentences contained inanimate subjects combined with different adjectives or verb phrases. The inanimate nouns were classified into three groups depending upon whether the object was capable of autonomous movement (he included sun and cloud in this group), whether the object was not capable of movement (termed “static,” e.g. lamp), or whether it could move but only via an externally caused force (car).6 Sentences were of the form The Adj N V-s PP and were made anomalous either by virtue of the adjective-noun combination within the subject NP (e.g. The hungry bed is in the room), or by virtue of the combination of the subject NP and main verb (e.g. The hot sun runs across the sky). A set of well-formed sentences contained animate subjects (e.g. The big fish swims in the water). Children had to judge whether each sentence sounded “OK” or “silly.” Children ranged in age from 4;2 to 8;10 with a mean age of 6;5. Schwartz categorized the children into groups according to their responses to the 6 It seems to me unlikely that children would consider the sun more capable of autonomous

movement than a car.

Children’s use of animacy in learning argument structure

175

Piagetian questionnaire, with Group 0 children responding that all objects were alive, Group 1 attributing life to those objects that move either autonomously or by external force, Group 2 attributing life only to those things that move autonomously (e.g. the sun), and Group 3 children denying that any inanimate objects were alive. Comparing their performance on the questionnaire and the sentence judgment task, Schwartz found that, generally speaking, children who judged more inanimate things to be alive in the questionnaire also judged more anomalous sentences to be “OK” as compared to children who were more adult-like in their assessment of which things were alive. The problematic nature of the Piagetian questionnaire notwithstanding (see Section 5.1.4), I find Schwartz’s results on the sentence judgment task questionable. First, children in Group 1, who judged both autonomously and non-autonomously moving entities (like the sun and cars) to be alive, were significantly better than Group 0 children (who judged everything to be alive) in rejecting anomalous sentences with non-autonomously moving subjects. That is, in spite of their determination that cars are alive, they rejected anomalous sentences with these NPs as subjects. Group 1 children were also significantly better in rejecting anomalous sentences with non-autonomously moving subjects (car) than autonomously moving subjects (sun), even though their metacognitive judgment was that both types of entities were equally alive. This shows that their performance on the sentence judgment task is more adult-like than their responses on the questionnaire with respect to these inanimate subjects. Secondly, even the Group 3 children, who were judged “non-animistic” by their questionnaire responses, were somewhat (though not significantly) worse in rejecting anomalous sentences with autonomously moving subjects than the other kinds of anomalous sentences. Thus, in spite of their judgment that the sun and clouds were “not alive,” they permitted about 30% of anomalous sentences with one of these subjects. Schwartz notes the possibility that children may have been giving these sentences a “metaphoric” interpretation, and I think this is a strong possibility. But if this is a possiblity for the Group 3 children, it is also a possible influence on the responses of the supposedly “animistic” children. Children may also have been thrown off by the use of both adjectives and verbs, across sentences, in contributing to the anomaly. That is, they may have been unsure exactly which aspect of the sentence they were supposed to be attending to since it varied across items. Finally, Schwartz notes that some children’s sentence judgments of anomaly were justified by appealing to other properties of the inanimate objects, like the fact that clouds “don’t have mouths.” Schwartz draws a distinction between the child’s judgment of clouds being alive and their having mouths, but isn’t this

176

Animacy and children’s language

the child’s way of marking what is alive and what isn’t? Given the problems in interpreting children’s determination of what the word alive applies to and what its semantic dimensions are, it seems to me that judgments of properties like “having a mouth” or “having feet,” or “having eyes” give a more accurate window into what the child actually thinks is alive. In contrast to Schwartz, my own interviews with 3- and 4-year-old children (note: much younger than Schwartz’s subjects) revealed a highly accurate rate of rejecting anomalous sentences with inanimate subjects (and accurate acceptance of sentences with animate subjects), according to the type of predicate applied to the subject (Becker, 2007). I showed twenty-four 3-year-olds and twenty-four 4-year-olds a series of normal cartoon pictures. “Normal” means that objects in the pictures were not “animized,” that is, inanimate objects were drawn as normal objects and not with added eyes or other typically animate features. A puppet described each picture with a single sentence, and the child’s task was to judge the puppet’s description as “OK” or “silly.” Four kinds of sentences were constructed by manipulating the animacy of the subject (e.g. Bert or The door) and whether the predicate typically applied to animate things (be friendly) or to inanimate things (be purple). Sentences in which both subject and predicate were animate or inanimate should have been judged as OK (Bert is friendly; The door is purple), and sentences in which the subject and predicate did not match in animacy should have been judged silly (Bert is purple; The door is friendly). Children were uniformly and significantly adult-like in their judgments; that is, they were significantly above chance for all types of sentences. There was no main effect of age, nor of sentence type, and there were no interactions. The percentage of correct responses, according to sentence type, are given in Table 5.1. Interestingly, the 3-year-olds had the lowest rate of correct responses in sentences with animate subjects and typically animate predicates, but even this

Table 5.1: Mean percent correct (with predicate type) Subject–Predicate Animate–Animate Animate–Inanimate Inanimate–Animate Inanimate–Inanimate

3-year-olds

4-year-olds

83.3% 87.0% 87.5% 87.5%

100% 87.0% 91.7% 83.3%

Children’s use of animacy in learning argument structure

177

rate (83.3%) was significantly above chance and not significantly different from their rate of accepting or rejecting sentences with inanimate subjects (87.5%). For neither age group was the lowest rate of correct responses found in sentences containing inanimate subjects paired with animate predicates; this would be predicted to be the most difficult condition if children think inanimate things are alive. In addition to the sentence judgment results, there is a vast and long-standing literature linking NP animacy to various aspects of children’s sentence structure. Early seminal work by Brown (1973), Bloom (1970), and Bowerman (1973, 1976) noted a strong tendency in the earliest multi-word utterances of English for preverbal nouns to indicate an animate entity (typically an agent) and postverbal nouns to indicate an inanimate entity (typically a patient). These patterns are reflected in children’s spontaneous productions, as well as their elicited judgments about sentence wellformedness (De Villiers and De Villiers, 1972). As noted by Scott and Fisher (2009, p. 781), “children keep track of the animacy of nouns in various grammatical positions.” A sizable body of work has explored children’s (as well as adults’) preference for animate agents and inanimate patients (and this normally translates into animate subjects and inanimate objects), much of it couched in terms of prototype effects (e.g. De Villiers, 1980; Slobin, 1981; Corrigan, 1988; Childers and Echols, 2004). That is, adults judge sentences with an animate agent and animate or inanimate patient to be “more prototypical” than sentences with an inanimate agent (causer) (Corrigan, 1986). In one study with children, Corrigan (1988) trained 2- and 4-year-olds to identify the agent (“do-er”) and patient of an action and tested their ability to correctly identify the agent and patient in sentences using different kinds of transitive verbs. Although all verbs chosen permitted both animate and inanimate subjects, and animate and inanimate objects (e.g. cover), some transitive verbs were considered “high prototypicality” verbs (e.g. kiss, pick up) since they have a stronger preference for an animate subject, while others were considered “low prototypicality” verbs (e.g. touch, blow), since they do not have this preference.7 Corrigan’s sentences were constructed by crossing animacy with semantic role, making for animate agent–inanimate patient (AI), animate agent–animate patient (AA), inanimate agent–animate patient (IA), and inanimate agent– inanimate patient (II) sentences. She found that children were both more 7 Corrigan notes that judgments of verb prototypicality were made on the basis of previous

norming studies she had conducted with adults, but some of these categorizations were counterintuitive to me, such as watch as a low-prototypicality verb, or move as high-prototypicality. Especially regarding move, see the following discussion of Gelman and Koenig (2001).

178

Animacy and children’s language

accurate and faster in identifying animate agents and less accurate and slower in identifying inanimate agents, for both high and low prototypicality verbs. There was also an effect of verb type, as the children were least accurate in identifying inanimate agents for the high-prototypicality verbs. Interestingly, the children in her study were also slightly better at the AA sentences than the AI sentences. This result is somewhat at odds with the other literature on prototypicality, which reports objects being more prototypically inanimate than animate (e.g. Childers and Echols, 2004; Slobin, 1981). There is also evidence from the developmental psychology literature for a preferential mapping of animacy to agenthood in preschoolers’ language. Studies by Gelman and Gottfried (1996) and Gelman and Koenig (2001) indicated that 3- and 5-year-olds interpreted the verb move differently according to whether the subject was animate or inanimate. In Gelman and Koenig’s study, children and adult controls watched video clips of animals or toys engaging in either spontaneous or caused motion events. In the spontaneous motion events, the (live) animal walked (or slithered, in the case of the snake) across a table, while the toy (e.g. a toy car or a wind-up toy) was shown moving across the table by itself. In the caused motion condition, each animal or toy was carried in the palm of a person’s hand from one side of the table to the other. In order to ensure that the animals did not produce any of their own movements in this condition realistic replicas were used. After each video, participants were asked “Was the [animal/toy] moving?” The main result was that while all subjects were at ceiling in answering that both the animal and the toy moved in the spontaneous movement condition, they differentiated the animal from the toy in the caused motion event, though to different degrees across the different age groups. Both 5-year-olds and adults gave significantly more “yes” answers to “Is the toy moving?” than to “Is the animal moving?” in this condition; the 3-year-olds showed a trend in the same direction but the contrast did not reach significance. In other words, participants were more likely to say that a motionless toy moved when it was carried by a person, than they were to say that a motionless animal moved under the same circumstances. The results imply a strong linkage between animacy and thematic role in children. More specifically, when move is used in reference to spontaneous motion, or what Levin and Rappaport Hovav (1995) call “internally caused” motion, such that the entity moving would plausibly have the role of agent or causer (hence an external argument), children easily map this NP to subject position. Put another way, when the NP is animate children easily construe the NP as an external argument, and thus the question Is NP moving? as having

Children’s use of animacy in learning argument structure

179

an unergative structure. On the other hand, when move is used in reference to (externally) caused motion, where the motion is a global movement of the whole object but without internal movement of the object, the question is asking about an internal argument. As would be expected, both children and adults are inclined to map an inanimate NP to this internal argument, but disinclined to map an animate NP to that position. That is, they take the question Is NP moving? to have an unaccusative structure, and unaccusatives allow an inanimate subject (see Section 2.3). What is interesting is that apparently there is a preference for this argument to be inanimate, as opposed to animate, due to its thematic role of patient. As Gelman and Koenig express it, inanimate patients can be subjects (as follows from an unaccusative structure), but speakers “reject animate patients as subjects” (p. 697). 5.2.1 The power and limitations of Semantic Bootstrapping Building upon some of the early empirical observations of a connection between animacy and sentence structure, but adopting a formal approach influenced by Grimshaw (1981), Pinker (1982, 1984) proposed an account of how children begin to figure out the structure of sentences of their language by using certain semantic primitives to “pull themselves up by their bootstraps.” The proposal rests on the assumption that children are able to identify a fair number of nouns as labels for objects and people in their environment, as well as some simple verbs as labels for actions, and thereby learn some word meanings. In turn, this assumes that although linguists do not define lexical categories in terms of semantic properties, strong semantic correlations will be found: objects and people will be labeled with nouns, events and states will be labeled with verbs, qualities and properties with adjectives, and so forth. In addition, it is assumed that the child knows certain structural correspondences, so that the entity perceived to be the instigator of an event will be mapped to the subject position, the event label will be mapped to the head of the predicate (verb), and any entity perceived to undergo the verb’s event, or otherwise a change of state, will be mapped to the object position of the sentence. In this way, “[the] categorization of words can be inferred from their semantic properties, and their grammatical relations can be inferred from the semantic relations in the event witnessed” (Pinker, 1984, p. 40). Using these assumptions, the child begins to build a limited set of grammatical rules that can then be applied with greater degrees of abstraction to sentences whose nouns do not refer to concrete objects, and whose verbs denote non-observable states. This later step is what Pinker called “structuredependent distributional learning.” In other words, after inferring that Mom

180

Animacy and children’s language

is the subject in the sentence Mom is throwing a ball because Mom denotes the agent of the action, and doing similarly across many other instances in which the syntactic subject is also the “do-er” of the verb’s action, the child can then compile a set of non-semantic properties that belong to the structural category of “subject.” For example, the subject is the thing that triggers Nominative or zero case marking (in some languages) and/or verb agreement (in some languages); it controls the referent of the implicit subject of an embedded infinitive clause; it is the thing that becomes oblique under passivization; and so forth (notice the resemblance to Keenan’s Promotion to Subject Hierarchy; see Section 3.4). Then, when another NP (identified as such by distributional properties, such as being preceded by a determiner) also has (some of) these subject properties, it can be analyzed as a syntactic subject, even if it does not denote a concrete entity, or an agent. Pinker gives the example, The situation justified extreme measures. The child is able to use this inductive process because the cluster of properties the child associates with semantically canonical subjects are exactly the same properties used by linguists in identifying subjects in languages under study. That is, these properties are valid criteria for establishing this grammatical category, even if they are not necessary conditions for category membership in the adult grammar. Is there evidence that children employ these assumptions about syntaxsemantics correspondence, apart from the fact that the target language contains them and children come to acquire the target language? There is evidence from results of experiments in which children were taught novel predicates that have different kinds of argument structure relations with their arguments (Lebeaux and Pinker, 1981; Marantz, 1982; Pinker et al., 1987). Marantz (1982) wondered whether children could be taught novel verbs that had different sorts of thematic grids (recall the brief description of this experiment in Section 3.4.1). He taught preschoolers a series of novel transitive verbs by viewing a scene involving some action, paired with a transitive sentence. The sentences had one of the following frames (Marantz’s experiment 2): (2)

a. Agent subject, Patient object (AP) The lion is puming the bear ‘The lion is pounding its head against the bear’ b. Patient subject, Agent object (PA) The frog is moaking the zebra ‘The zebra is hitting the frog with the back of its knee’ c. Agent subject, Location object (AL) The cow is nading the dog ‘The cow is rolling its head while standing on the dog’

Children’s use of animacy in learning argument structure

181

d. Location subject, Agent object (LA) The horse is dabing the cat ‘The cat is twirling its arm while standing on the horse’

The children watched up to six videos of the same event (different agents and different patients across the six videos, only the action itself remained the same), and after each trial had to say what was happening in the video using the novel verb (moak, etc.). For example, in the puming example above, the child would have to say something like “The lion is puming the bear.” What was measured was how many trials it took before they produced a spontaneous use of the novel verb to describe the event. Furthermore, the frame they produced was recorded, i.e. either a correct frame (AP for an AP verb) or a reversal (e.g. AP for a PA verb). The result was that 3- and 4-year-olds learned the “canonical” types of verbs, those that took an agent subject and either a patient or location object, significantly faster than the “anti-canonical” verbs, those with an agent as object. The 5-year-olds showed a trend in this direction but the difference was not significant; lack of significance resulted from the better performance of the older children on learning the anti-canonical verbs. The younger children also produced significantly more reversals for PA and LA verbs than for AP or AL verbs. The 5-year-olds did not produce significantly more reversals of anticanonical than canonical verbs, but a look at Marantz’s data shows that while the 5-year-olds in fact produced equal numbers of AL and LA reversals (8 of each), they produced more than twice as many PA reversals than AP reversals (12 vs. 5). Thus, even in the older children there was a strong tendency towards reversing PA verbs. There are a couple of important observations to make about Marantz’s results. One is that although the younger children appeared to rely more heavily on semantic role mapping than word order mapping (in that they were biased to express the agent as the subject in their own sentences), a number of these younger children did learn the anti-canonical verbs. That is, they eventually produced spontaneous utterances for the PA and LA verbs, and some did so without producing reversals. This shows that although children may have a bias to map the agent to the subject position, the bias can be overturned, even at age 3. This result underscores my contention that animacy cues are probabilistic, rather than deterministic.8 8 One can ask whether these “spontaneous” productions were truly spontaneous, since these sen-

tences had been given to the children as descriptions of the video scenes. Thus, responses would have been judged correct if children were simply imitating the sentences. Although there is

182

Animacy and children’s language

The second observation concerns an interesting difference between Marantz’s two experiments. The results I reported above come from his second experiment. The first experiment differed from the second, among other things, in having only animate (human) agents and only inanimate patients. For example, the sentence corresponding to (a) above in experiment 1 was Larry is puming the book, while the sentence corresponding to (b) was The book is moaking Larry. The results of the two experiments differed as well. Most significant for our purposes is that experiment 2 (where both NPs were animate) yielded significantly more reversals than experiment 1 (human agent, inanimate patient). The less frequent use of reversals in experiment 1 could be due to resistance on the part of children to put an inanimate NP in subject position. A novel word study by Pinker et al. (1987) showed a similar kind of effect in the passive construction. We will return to children’s learning of passives below in Section 5.3.4. At the moment what is relevant is that the children in Pinker et al.’s study quite willingly formed novel passives for novel verbs that took a patient object, like Marantz’s canonical verbs, but did so far less willingly for verbs that took an agent or location object (and a theme subject; i.e. anti-canonical verbs). In addition to results from these novel word-learning experiments involving English-speaking children, further evidence for children’s underlying expectations about argument structure comes from cross-linguistic studies. As mentioned above in Section 3.4.1 Pye (1990) found that children acquiring K’iche’ (a language with ergative case marking but accusative argument structure) have little difficulty acquiring the case marking system of their language. However, the one construction which involves something like an ergative syntactic structure (i.e. the agent argument shows object agreement), the Focus Antipassive construction, presents difficulties for children until after age 3. This is expected, since the argument structure of an ergative syntax system involves a fundamentally different mapping of thematic roles to syntactic positions. The structures are repeated here: (3) and (4) both have the same meaning (“Mary hugged John”), but the agent is mapped to the object position in the ergative syntactic structure.

evidence that even children’s “imitative” utterances are legitimate productions by children’s grammar (Slobin and Welsh, 1973), the possibility of rote production is always a concern when the exact target sentences have been modeled for the child.

Children’s use of animacy in learning argument structure (3)

S

Ergative

NPabs

VP

(4)

183

NPerg

V

Mary A

hugged

John O

S

Accusative NPnom Mary A

VP V

NPacc

hugged

John O

If children are biased to map the agent to the subject position in syntax, a crucial ingredient in Semantic Bootstrapping, then structures like (3) should be difficult (if not impossible) to acquire. Pye’s finding suggests that they in fact are difficult for young children. Ochs (1982) also looked at the acquisition of ergativity in children, focusing on the acquisition of Samoan. She reports that unlike Schieffelin’s (1979) finding of early acquisition of ergative case marking in Kaluli (and unlike Pye’s (1990) later findings for K’iche’), correct production of ergative morphology emerges quite late in Samoan-speaking children. Out of six children studied, none but the oldest child (age 3;4) used ergative case marking at all in their speech (except for two counterexamples, one of which was a partial repetition of an adult utterance). Ergative case marking in Samoan is subject to a number of distributional restrictions. Subjects of transitive verbs are marked with the ergative e marker if they follow the verb (e.g. VSO, VOS), but not if they precede the verb. Furthermore, only subjects of “canonical” transitive verbs, verbs that denote actions (e.g. fasi ‘hit’, ’ai ‘eat’), can bear the ergative marker. Transitive verbs that denote desires, emotions and perceptions (Ochs calls them “middle” verbs, e.g. mana’o ‘want’, ita ‘hate’, va’ai ‘see’) do not have ergative-marked subjects, even if the subject follows the verb. The one exception Ochs notes is the verb iloa ‘know,’ whose subject does bear ergative case if it follows the verb. So the appearance of ergative case marking in Samoan is

184

Animacy and children’s language

subject to variation, according to word order as well as aspectual (or perhaps idiosyncratic) properties of the verb. More importantly, in Ochs’s view, there are strong sociolinguistic constraints on the use of ergative case marking. In particular, while formal speech (especially speech by men to non-family members) contains a high percentage of ergatively marked postverbal subjects (73%), informal speech (speech to family members) contains a much lower percentage (18.8% averaged across both men and women speakers; the rate for women is 20%). In Samoan culture children spend nearly all of their time in the family home, in the care of women and surrounded by family members. Therefore, they are exposed primarily to informal speech. Ochs concludes that it is the low rate of exposure in the linguistic input that results in the delayed acquisition of this morphology by children. As noted above, ergative case marking systems are not problematic for the Semantic Bootstrapping hypothesis, since in so far as a morphologically ergative language has (primarily) accusative syntactic structure, the canonical mapping of semantic to syntactic positions will proceed just as in languages with accusative case systems. And so there is no a priori prediction that ergative, as opposed to accusative, case marking systems should be relatively hard or easy to acquire. However, as in the case of K’iche’, Ochs notes that Samoan is a “split-ergative” language in that some grammatical constructions appear to follow an accusative type of syntax (e.g. raising-to-object), while others appear to follow an ergative type of syntax (e.g. nominalizations). Therefore, we might ask whether children acquiring Samoan display any preference for an accusative type of syntactic structure in their own utterances. Although Ochs does not report children’s productions of raisingto-object versus nominalizations, she does report a strong tendency in the children’s word orders to reflect an accusative kind of grouping, rather than an ergative one. Specifically, Ochs notes that Samoan children resist placing the agent immediately after the verb in their transitive utterances. That is, even though the “basic” word order of adult Samoan is VSO (or, VAO using A to indicate a transitive subject), this word order is extremely infrequent in children’s speech, outnumbered by a large margin by both VOS (52%) and SVO (30%) (compare with about 11% for VSO). Ochs cites the observation of Lehmann (1973) and Slobin (1975) that “the verb and patient form a ‘perceptual Gestalt which resists interruption”’ (Ochs, 1982, p. 663). This syntactic grouping of verb and patient, rather than verb and agent, is the hallmark of an accusative syntactic structure. Thus, this word order preference among Samoan-acquiring children

Children’s use of animacy in learning argument structure

185

provides further evidence to support a developmental bias for an accusative type of argument structure.9 The great strength of the Semantic Bootstrapping hypothesis lies in its intuitively correct supposition that children in the early stages of sentence structure representation will be biased to map the agent of an action to the syntactic subject position, and the patient or theme to the object position in the structure. This bias accords with the strong tendency to find this mapping in the world’s languages, with parental use of language, and with children’s own biases as demonstrated in experimental studies. But it is not without its limitations. The major limitation from my point of view is that because its coverage is limited to basic sentences, it cannot straightforwardly account for the opaque constructions under consideration here. Of course, Pinker intended Semantic Bootstrapping as an explanation of how children begin to build syntactic representations at the very earliest point in language acquisition: when a child hears an NVN sentence, how does she know whether the word order is really SVO or OVS? If she hears VNN, how does she know whether it is VSO or VOS? All are logical possibilities, but children home in on the right word order for their language at a remarkably early point, both in production (Brown, 1973) and comprehension (Hirsh-Pasek and Golinkoff, 1991; Fernald, 1992), suggesting that they do not go through an extended period of entertaining all of the logical possibilities. But it is important to recognize that this approach works for “basic” sentences and not for “non-basic” sentences, as Pinker acknowledges. My critique is not that basic sentences shouldn’t be the first ones a child represents, but rather that in order to develop a comprehensive theory of language acquisition we need to go beyond the basic sentences and develop an approach that will cover non-canonical sentences, including the opaque constructions that are of interest here. The problem is not with the identification of the subject as syntactic subject, since the child should be able to do this via structure-dependent distributional learning. The problem, instead, lies in discovering the semantic role played by the syntactic subject, and in discovering that there is no semantic selectional relationship between the subject and the main predicate. In other words, Semantic Bootstrapping will not lead to the discovery that a syntactic subject is not an external argument. The null 9 On the other hand, the Samoan-speaking children tend to produce the lone argument of intransi-

tive verbs postverbally (i.e. VS is the primary word order), which appears to reflect an ergative grouping (S and O to the exclusion of A). However, this is mainly reflective of the ergative case marking grouping. Thus, Samoan-speaking children do have traits of ergativity in their speech, reflected in their word order preferences. But they do not appear to have a preference for mapping the agent to the object position, which is the pattern required by an ergative syntax.

186

Animacy and children’s language

hypothesis should always be that the subject is the external argument. We saw in Chapter 4 that adults do have this bias. What matters, though, is that the bias can be derailed. What I have argued here is that it is encountering an inanimate subject that helps to derail the bias and cue the fact that the subject is not the external argument of the verb. Put another way, my proposed learning procedure is the flip-side of Semantic Bootstrapping: not only is an animate NP likely to be a subject, but an inanimate subject is likely to be a derived subject. A better known challenge to Semantic Bootstrapping came from Lila Gleitman and the Syntactic Bootstrapping literature. Beginning with Gleitman (1990), Gleitman argued against one of the key premises of Semantic Bootstrapping, namely that children can plausibly infer not only the meanings of nouns based on observation of the world, but also the meanings of verbs. Pinker’s approach required children to know something about the verb’s meaning in order to figure out the sentence structure, but Gleitman’s insight was that children learn verb meanings via the syntactic structure of a sentence, not the other way around. 5.2.2 The power and limitations of Syntactic Bootstrapping In 1985 Barbara Landau and Lila Gleitman published a study of language development in blind children. One of the aspects of their language development Landau and Gleitman wondered about was the acquisition of verbs relating to sight, look, and see. These were verbs whose denotations were completely outside of the realm of blind children’s experience, and so the authors supposed that blind children would have some difficulty acquiring them. To their surprise, they found that blind children had no problems understanding them (at the appropriate stage of grammatical development) and little difficulty using them correctly and appropriately in their own speech. Both in spontaneous usage and in experiments on comprehension, the blind child they studied most closely, Kelli, indicated that look meant to apprehend or explore something (for her in the haptic, rather than the visual modality, of course), and that seeing was the result of looking at something. In attempting to explain how this could be, Landau and Gleitman looked to the child’s mother’s speech: perhaps her mother used verbs like look and see when the object to be explored was near Kelli, or in her hand, but not when the object was far away. But in comparing Kelli’s mother’s use of look and see with other common verbs, such as hold, put, give, have, and so forth, they found that look and see did not have the highest proportional use when an object was close or in Kelli’s hand. Instead, the verbs hold, put, and give all

Children’s use of animacy in learning argument structure

187

had higher rates of use when the object was close by than either look or see. So maternal use of these verbs in the “right” context was not the answer. Looking deeper, Landau and Gleitman hit upon a solution that involved taking the sentence frames in which only verbs of perception were used, but not other kinds of verbs, and looking at what proportion of the time the object was close by when Kelli’s mother used the verbs look and see in just these frames. Now the picture changed substantially: after isolating the sentence frames that admitted verbs of perception (e.g. V [S]), but excluding other frames like simple transitive sentences, which admit a host of non-perceptual verbs, the mother’s use of these verbs when the object was close by increased dramatically. Landau and Gleitman concluded that it was the sentence frames in which a verb occurred, in conjunction with contextual support, which led blind children to discover the meanings of these verbs of whose denotation they had no direct experience. Further, Landau and Gleitman reasoned that if blind children use sentence frame information to restrict the meanings of verbs, then sighted children should do the same. Thus was born the Syntactic Bootstrapping hypothesis. The Syntactic Bootstrapping hypothesis allowed Gleitman to resolve an important flaw in Pinker’s Semantic Bootstrapping hypothesis (Gleitman, 1990). Recall that in Pinker’s approach, there is an explicit assumption that the child can infer the meanings of sentences given knowledge of the individual lexical items. Crucially, this includes knowledge of verbs. While Gleitman agreed that nouns could be learned by some fairly unsophisticated mapping between objects and labels, she denied that verbs could be learned this way. One objection she raised was that events are not normally labeled for children the way objects are (when mom comes home at the end of the day and opens the door, she says “Hi! How was your day?” not “I’m opening the door”). A second objection was that even if an event is labeled, how does the child know which event, out of the myriad things going on in the world around her, is supposed to be the particular event this word denotes? A third objection was that there are pairs of verbs that can be used to describe exactly the same event but have opposite meanings (give/receive, buy/sell, chase/flee). How, in principle, should the child know that a given sentence means “The dog is chasing the cat” instead of “The dog is fleeing the cat”? In the wake of these objections, a flurry of research on children’s verb learning ensued. Over the course of the following twenty-five years this work showed just how powerful sentence frame information can be in restricting the possible meanings of a verb by weeding out the unlikely meanings. A verb used with one argument might mean “sleep” or “laugh” but is very unlikely

188

Animacy and children’s language

to mean “hit” or “give”. A verb used with two arguments might mean “hit” but would be unlikely to mean “sleep,” and so on. The relevant studies are too numerous to recount in detail or in any exhaustive manner (see Gleitman et al. (2005) for a good overview), but the upshot is that children, from the age of about 18 months onward, rely heavily on the syntactic cue of sentence frame as a way to home in on verbs’ meanings. It should not be assumed, however, that semantics has no place in Syntactic Bootstrapping. Below in Section 5.3.3 we will look at experiments that specifically investigate the role of animacy in acquiring unaccusative and unergative verbs. For the moment, I wish simply to note that the Syntactic Bootstrapping hypothesis provides an account of how children can use simple argument structures, combined with some linguistic semantic and extralinguistic information, to infer something about the meanings of verbs they are learning. Nevertheless, there is a weakness in Syntactic Bootstrapping just as there is in Semantic Bootstrapping that renders it unable to account, straightforwardly, for how children acquire opaque constructions. This is the fact that in order for Syntactic Bootstrapping to work, the child must assume that the nouns adjacent to a verb are in fact the verb’s semantic arguments. The sentence frame is only informative in so far as it is the syntactic rendering of the semantic relations the verb requires. But in the case of raising verbs and tough-adjectives, the syntactic NP subject is not semantically related to the main predicate, yet they are adjacent all the same. So although children should be biased to assume that adjacent arguments are semantically related, in order to make use of Syntactic Bootstrapping for the canonical cases, we also need something that will break this bias so that opaque constructions can be acquired. Inanimate subjects provide this key. Recall that there are other constructions that disambiguate displacing from non-displacing predicates. As discussed in Section 2.5 most displacing predicates occur with expletives while non-displacing predicates do not. (5)

a. There seems to be a tack on the floor. b. * There claims to be a tack on the floor.

(6)

a. It is hard to solve this problem. b. * It is eager to solve this problem.

There is good evidence that children reason across occurrences of predicates, using the range of structures in which a verb occurs to home in on its lexical properties (Fisher et al., 1991; Lederer et al., 1995; Naigles, 1996; Bunger and Lidz, 2004, 2008). My focus on inanimate subjects as the relevant

Children’s use of animacy in learning argument structure

189

cue should not be mistaken for a disavowal of expletives or the important role they can play in informing the shape of syntactic configurations. In fact, expletives are immensely helpful: their presence indicates that the predicate does not select a semantic argument in that position. My reason for focusing on inanimate subjects as the learning cue comes from six observations listed above in Section 2.5. I summarize them briefly here: (a) expletives are not by themselves completely unambiguous in the case of raising and control verbs because of the existence of ambiguous verbs like begin which can occur with expletives but can also function as control verbs (Perlmutter, 1970); (b) expletive subjects do not exist in all languages (e.g. Spanish, Mandarin), while all languages have both animate and inanimate NPs; (c) expletive subjects are even rarer in speech to children than inanimate referential subjects (to be demonstrated in Chapter 6); (d) there is psycholinguistic evidence that inanimate subjects are preferentially interpreted by adults as displaced (Chapter 4); (e) inanimate subjects are as good a cue as expletive subjects for the lack of a semantic role in biclausal sentences (such as raising and tough-constructions), and nearly as good a cue in monoclausal sentences (unaccusatives); and (f) there are some predicates that allow an expletive subject but do not allow a referential NP to be raised to subject position (e.g. suck, be probable). On the other hand, expletives may serve as a more reliable cue to an adjective being in the narrow class of tough-adjectives (as opposed to adjectives that participate in other sorts of object-gap constructions). So there is something of a trade-off between expletives and inanimate NPs in terms of their usefulness and robustness as cues to displacing predicates, and it is difficult to determine which is really the better cue. Further complicating matters is the fact that they are, for the most part, the only diagnostic environments for displacing predicates. As we will see in the next section, my own approach to studying how children learn displacing predicates is to present the predicates (real or novel) with an animate or inanimate subject, and then to use expletive constructions as my assessment tool (if the child thinks the predicate sounds fine with an expletive subject, then they have categorized it as a displacing predicate). One could, instead, present predicates with expletive subjects, and then use inanimate subjects as the test environment. But this alone would not tell us which cue was more useful. If the results turned out the same, it might mean that both cues are equally useful. Or it could mean that one cue is more powerful than the other, but in one case the stronger cue is used in the presentation phase and in the other case the stronger cue is used in the assessment phase. Unless we have an independent environment that can be used to test each cue separately, it will be hard to answer this question.

190

Animacy and children’s language

As we will see in the next section, children are sensitive to subject inanimacy as a means of identifying (or categorizing) displacing predicates. That they can use expletive subjects in conjunction with inanimate subjects in real-life language learning is not disputed. The point is to show that inanimate subjects, by themselves, can be used as a learning tool by children. 5.3

Children’s acquisition of displacing predicates

In the studies discussed in the following sections I look at what children know about displacing predicates, and the role that inanimate subjects can play in children’s acquisition of these predicates. Given that children are shown to use a semantic cue (subject animacy) as an indicator of syntactic structure (displaced vs. non-displaced subject), which then informs the lexical properties of the main predicate (displacing vs. non-displacing predicate), the general story here involves a significant amount of interplay between Semantic and Syntactic Bootstrapping. Both approaches play a vital role in the acquisition of displacing predicates and their unusual structures, and both approaches are in turn extended and strengthened in their empirical coverage by applying them to these non-canonical forms. Before turning to the empirical studies, I want to address the idea that probes of syntactic competence using both animate and inanimate arguments within a clause are uninformative, since children might simply rely on “semantics” (meaning real world knowledge) to figure out the meaning of the sentence, and not actually project the sentence’s syntactic structure. In essence, the worry is that one could interpret John hit the chair correctly with John as agent and chair as patient simply because chairs make for unlikely hitters, but in order to interpret John hit Bill with John as agent and Bill as patient, one needs to know that John is in subject position. So, if I show that children correctly interpret The chair is easy to push but not John is easy to push, maybe this is “only” because they know that chairs don’t push other things. While I recognize that in order to talk about children’s syntax we want to be sure we’re talking about children’s syntax, and not some other mechanism, I reject the notion that probing children’s competence with sentences whose NPs vary in animacy is uninformative. I do so for three reasons. The first is that we know that NP animacy can affect the processing load of a sentence. Goodluck and Tavakolian (1982) showed that children committed significantly more errors in acting out sentences with three animate NPs (The dog kicks the horse that knocks over the sheep) than the same sentences with two animate NPs and one inanimate NP (The dog kicks the horse that knocks over

Children’s acquisition of displacing predicates

191

the table). In these sentences, one can’t explain the worse performance in the three animate NP condition by appealing to world or semantic knowledge. So, if children perform poorly on sentences with an animate displaced subject (as they are claimed to do for raising sentences by Hirsch and Wexler (2007)), how do we know this is because the child is trying to rely on semantics, and not because of a processing overload? Furthermore, since we know that inanimate displaced subjects are processed more easily by adults than animate displaced subjects (see the work on object relative clause constructions discussed in Section 4.1), why should we insist that children demonstrate ability to parse a given structure under the hardest conditions before acknowledging that they interpret the sentence via a syntactic representation rather than a semantic one? A second reason I reject the idea that one must test children’s knowledge of sentence structure with only animate NPs is that the main claim I am making is that children exploit patterns of NP (in)animacy in order to learn how to represent the structure of their language. I show (below) that children are in fact influenced by NP animacy in their categorization of novel predicates, and, presumably, this is indicative of how they categorize predicates in their natural target language. So if children exploit the cue of NP inanimacy as a means of acquiring structure, why should they not exploit these cues in interpreting sentences of their target language? The point is that children use these animacy cues as a means of projecting structure, so if they demonstrate that they correctly interpret a sentence where these cues are available, we don’t need to reject out of hand the idea that they are still arriving at the right syntactic representation. The preceding two arguments make the case that the “semantics only” argument is not the only conclusion we are driven to, if children show good comprehension of sentences with mixed NP animacy. My third argument goes a step further in claiming that it is the wrong conclusion. Pinker (1984) makes a detailed case for the Continuity Principle, namely, that if children display adult-like behavior regarding one or another linguistic form, then the most parsimonious account explains this behavior by saying that the child has adultlike knowledge of the same linguistic form. In the present case, if children give evidence of understanding sentences like The chair is easy to push or The rock seems to be heavy, the simplest explanation is that they assign these sentences an adult-like structure. If we say they understand these sentences by some alternative means, the theory is further complicated by the need to then explain how children “unlearn” this alternative mechanism. Moreover, the proposal of an alternative mechanism is not independently warranted, since to the extent that

192

Animacy and children’s language

children exhibit difficulty with The boy is easy to push or The man seems to be heavy (i.e. the same sentences but with an animate subject), this difficulty can be easily explained either by processing load (see above) or, more likely, by the string-overlap with control structures (The boy is eager to push, The man claims to be heavy). Therefore, it seems to me that testing children’s knowledge of these complex kinds of sentences by using both animate and inanimate NPs can be an informative means of discovering their ability to represent the syntactic structures of these sentences.

5.3.1 Acquisition of raising and control A number of researchers of early child language have noted the rarity of verbs in children’s speech that take sentential (tensed or untensed) complements, including raising and control verbs (Bloom et al., 1975; Brown, 1973). While some early utterances contain combinations of want or can’t with another verb (want go), Pinker (1984) takes the view that the structure of these utterances is not clearly biclausal. It is hard to know what structure generates these early V+V sequences at a stage when most utterances contain at most two words. On the other hand, at the point at which children begin building three- and four-word utterances, complement-taking verbs appear robustly. These verbs, including want/wanna, gonna, and hafta, are used correctly in the sense that the controller of the lower predicate is always the main clause subject. Limber (1973) observed the use of various types of complement-taking verbs: initially subject control (want/wanna) and raising-to-subject verbs (he classified need as raising, though actually it is ambiguous; see Section 2.5), then auxiliaries, which, in terms of their argument structure, are equivalent to raising-to-subject verbs (Bresnan, 2001). At a slightly later stage children begin producing object control verbs (tell) and raising-to-object (want used with an object and a complement clause). Finally, verbs taking tensed complements (think, know) are observed in productive use. It should be borne in mind, though, that by “later” we are still talking about a relatively early stage in development: according to Limber these verbs (and their complements) are used productively before age 3. Experimental work in children’s acquisition of complement-taking verbs has focused primarily on the acquisition of object control verbs, with the central question being how children determine which of the two NPs in the matrix clause is the correct controller for PRO (John told Bill to leave – who will leave?). Since object control verbs are not the focus of this book I will put this literature to the side (see, for example, Chomsky (1969); Goodluck and

Children’s acquisition of displacing predicates

193

Behne (1992); Wexler (1992); McDaniel et al. (1991); Cairns et al. (1994); Cohen Sherman and Lust (1995)). Rather less attention has been paid to the acquisition of subject control and raising-to-subject verbs. Pinker (1984) observed that with an MLU of only 2.65, Eve (Brown, 1973) used verbs he classified as raising-to-subject (can’t, auxiliary have, don’t, and future go (gonna)) as well as the subject control verbs forget and try. Following closely upon that, Eve used gotta, may, will, and shall, all deemed by Pinker to be raising-to-subject, and wanta (subject control). Raising-to-object want appeared in Eve’s production at a somewhat later point, around MLU 4.0–4.5 (Brown’s Stage V). In terms of argument structure (and in the Lexical-Functional Grammar framework in which Pinker was working) auxiliaries and raising-to-subject verbs are formally equivalent. Both types of verbs fail to semantically select a subject argument, and both select a verbal or clausal complement (i.e. not an NP complement). But there are also important differences between auxiliaries and raising verbs as I’ve been using this term (i.e. to pertain to verbs like seem, tend, etc.) such that it is unclear whether their structures should be seen as truly equivalent. One obvious difference is that auxiliaries take a bare verbal complement, while raising verbs take an infinitival complement (that is, with the overt to infinitive marker). This distinction holds not only for English, which does not have bound infinitival morphology, but also in German, which does have infinitival morphology on verbs, as well as in Chamorro (see examples in Section 3.3.1). (7)

Er scheint die Sache vergessen *(zu) haben He seem-3S the thing-f forget-inf to have-inf “He seems to have forgotten the matter.”

Since many of children’s earliest productions of these (generalized) raising constructions either involve true auxiliaries (can, will) and semi-auxiliaries (gonna, hafta), in which the to infinitive is incorporated phonetically into the verb itself, it is unclear whether children are representing the complement clause as an infinitive clause or as a VP. A further difference between auxiliaries and raising verbs has to do with their morphosyntactic status. In English, modal auxiliaries do not inflect to agree with the subject (though the nonmodal auxiliaries have, be, and do do), and auxiliaries uniformly invert with the subject in questions. Raising main verbs are dissimilar in both of these respects: they exhibit 3sg agreement with the subject, and they fail to invert in questions, just like canonical main verbs.

194

Animacy and children’s language

In addition, auxiliaries (to the extent that they are uniform in not selecting a subject argument) do not have a parallel structure like raising verbs do, in which the subject is potentially an argument of the verbal item. That is, there are no control verbs that behave morphosyntactically like auxiliaries.10 For all of these reasons I will put aside auxiliary verbs in the discussion to follow.11 My own search of the CHILDES database turned up vanishingly few uses of non-auxiliary raising verbs in children’s spontaneous utterances. Across all available files in the Bloom (1970), Brown (1973), Kuczaj (1977), MacWhinney (2000), Sachs (1983), Suppes (1974), and Valian (1991) corpora (the number of child utterances in these corpora totalling 248,940) I found three uses of the verb seem with an infinitive complement, listed here. (8)

a. It seems to be (Adam age 3;07 from file 34) b. Yeah because Marky seems to get hurt (Ross age 4;11 from MacWhinney file 57a2) c. for Marky and bucking bronco because he seems to get hurt (Ross age 4;11 from MacWhinney file 57a2)

In addition, there were two uses of seem with a non-infinitival complement (e.g. That seems fun). The semi-auxiliary used-to was used a respectable number of times, but still not what I would consider frequently. Searching the same corpora as above I found seventy-five spontaneous productions of usedto (meaning past, i.e. excluding the handful of uses of be used to, meaning ‘be accustomed to’), sixty of them with an animate subject, twelve with an inanimate referential subject, and three with an expletive subject. These productions are exemplified in (9). (9)

a. I be a little baby when I used to play with this? (Ross age 2;9 from MacWhinney file 27b1) b. I found one that used to go in there. (Adam age 3;08 from file 35) c. It used to be daytime now it’s getting nighttime. (Abe age 2;10 from Kuczaj file 049)

While some of the productions come from somewhat “older” data points, like Ross’ two uses of seem, notice that some of the utterances with used-to were produced before age 3.

10 An argument could be made that German wollen ‘want’ fits this bill, since it is a modal but

takes an experiencer subject. See Section 3.3.1. 11 The semi-modal going to/gonna will be included in the corpus searches discussed in Chapter 6.

Children’s acquisition of displacing predicates

195

Like the work on tough-constructions we will see in the next section, the experimental work on children’s acquisition of raising verbs has placed an emphasis on the question of what children know about raising verbs and the construction of raising-to-subject in the target language, and when they know it. I will describe some of this research before turning to my own work on raising and control, which focuses instead on the question of how children come to distinguish raising verbs and structures from those associated with control. A body of work by Ken Wexler and his colleagues (Wexler, 2004; Hirsch and Wexler, 2005, 2007; Hirsch et al., 2008) has shown that children have difficulty interpreting sentences with seem in which the subject of the lower clause is raised to the matrix clause, though not when that NP remains in the lower clause. That is, children incorrectly interpret (10a) but not (10b). (10)

a. Bert seems to Ernie [t to be wearing a hat]. b. It seems to Ernie [that Bert is wearing a hat].

The theoretical explanation offered for this asymmetry rests on the claim that children have trouble forming A-chains (argument-chains) until late in development. “Late” generally means around age 4 or 5 years (Babyonyshev et al., 2001; Hirsch and Wexler, 2005), though explicit estimates of what “late” means vary in the literature from sometime after age 2;6 (Borer and Wexler, 1992) to age 7 (Hirsch and Wexler, 2007). The original proposal for A-chain maturation was made by Borer and Wexler (1987, 1992) to account for the relatively late acquisition of passives. However, since it is relevant to the arguments of Hirsch and Wexler I will give a brief description of it here. The formal mechanism used to account for children’s problem forming Achains can be understood in terms of the A-Chain Deficit Hypothesis (ACDH; Borer and Wexler, 1987; Babyonyshev et al., 2001) or, as in later work, in terms of the Universal Phase Requirement (UPR; Wexler, 2004; Hirsch and Wexler, 2007). According to the ACDH, children are simply unable to form Achains in their grammar until this capacity matures along a timecourse of biological maturation. As revised by Borer and Wexler (1992) the ban on A-chains does not include argument movement of an external subject as generated in the Specifier of VP into the Specifier of IP (movement assumed under the VPInternal Subject hypothesis; Kitagawa, 1986; Koopman and Sportiche, 1991). However, it does include the sort of A-movement found in passives, raisingto-subject, and unaccusative constructions (recall that tough-constructions are, or were, standardly thought to involve A -movement). The idea of the A-chain

196

Animacy and children’s language

maturation approach is that due to the inability to form A-chains, utterances that appear to involve A-movement actually involve a non-movement-based structure. For example, if children produce or correctly interpret truncated passives, children represent these constructions as involving an adjectival passive structure, and unaccusatives would have an unergative representation. The prediction is that children should perform poorly on production or comprehension tasks of raising-to-subject constructions (as well as full passives). The Universal Phase Requirement (Wexler, 2004) recasts this constraint in terms of Phase Theory (Chomsky, 2001). Recall from Section 2.1.1 that v is considered defective if it does not project an external argument (i.e. for raising verbs, passives, and unaccusatives); as such, it does not constitute a phase, and therefore movement may take place from within this segment of the structure into the T domain (i.e. the main clause) prior to Spell-Out. Wexler’s idea is that in child grammar v always constitutes a phase, whether or not v is defective. Therefore, movement into the T domain is always blocked. What matures, then, is the distinction between defective and non-defective v such that only non-defective v is a phase. However the restriction on A-movement is formalized, the intuition behind this restriction in early grammar is very appealing in many respects. In order for a Pinker-style Semantic Bootstrapping mechanism to function in early grammar, it is necessary that the argument relations between NPs and predicates be projected locally, and according to the Thematic Hierarchy. The whole point of Semantic Bootstrapping, and Marantz’s (1984) canonical argument relations, is that agents are mapped to the external argument position and themes or patients are mapped to the internal argument position. A-movement disrupts this canonical alignment, and so from this perspective, we should expect children, at least at the earliest stages of language acquisition, to eschew structures that “mess up” the canonical alignment of arguments. On the other hand, it is debatable whether A-movement is really impossible for children. My own data support the view that when the displaced subject is inanimate, children have no trouble representing A-movement. To flesh this out, let us look in a bit more detail at Wexler’s evidence against A-chains in raising verb constructions, and then at my own evidence for children’s ability to represent these constructions. Hirsch and Wexler (2007) presented children ages 3–9 with a spoken utterance and two pictures, one which matched the utterance, and one which did not match the utterance. Apart from active transitive sentences used as an experimental control, the spoken utterance (the experimental prompt) was one of three types of constructions, illustrated in (11).

Children’s acquisition of displacing predicates (11)

197

a. Lisa thinks that Bart is playing an instrument. (tensed complement) b. It seems to Lisa that Bart is playing an instrument. (unraised seem sentence) c. Bart seems to Lisa to be playing an instrument. (raised seem sentence)

Both pictures that accompanied the verbal prompt contained two characters, one drawn in the foreground with a thought-bubble (as used in cartoons), and the other character inside the thought-bubble. This depiction was meant to indicate that one character was thinking about the other character. Thus, the “correct” picture for any of the three sentences in (11) would involve Lisa in the foreground with a thought-bubble, and Bart inside that thought-bubble playing a saxophone. The non-matching picture (the foil) deviated from the spoken prompt along one of three dimensions. For what they called Matrix-Reversal (MR) foils, the experiencer argument was incorrect. So, in the above examples this picture would have Bart playing the saxophone while thinking about Lisa. For Embedded-Reversal (ER) foils, the agent of the embedded clause would be incorrect. In this case, the picture would show Lisa playing the saxophone while thinking about Bart. Finally, Double-Reversals (DR) reversed both roles, so the DR foil for the above sentences would show Bart thinking about Lisa playing the saxophone. Children were shown two pictures at a time and asked to point to the picture matching the utterance they heard. The overall result, averaging across the different foil types, was that while children in every age group performed at greater than 90% accuracy for active controls, think sentences and unraised seem sentences, only the 9-year-olds performed at this level for the raised seem sentences. Performance on the raised seem sentences (i.e. (11c)) ranged from 43.9% (3-year-olds) to 75.6% (8-yearolds). There is a small “jump” in accuracy between the 6-year-olds (51.7%) and 7-year-olds (71.1%), and this jump is associated with a shift from at-chance performance to above-chance performance. Breaking down the results according to foil type, Hirsch and Wexler found that for the think sentences and unraised seem sentences, children younger than age 7 did the worst on the Matrix-Reversal (MR) foils, with 3-year-olds making up half of the children who performed below chance on these items. That is, younger children frequently selected the MR foil rather than the correct picture. This error entails selecting a picture with the meaning of the embedded clause correctly depicted (e.g. Bart playing the saxophone), but an incorrect experiencer (Bart having a thought-bubble instead of Lisa).

198

Animacy and children’s language

For the raised seem sentences, which overall had a much lower accuracy rate, children did worst on the Double-Reversal (DR) foils, actually preferring these pictures to the correct picture a significant portion of the time. (Children younger than 7 were at chance on both the MR and ER foils for these sentences.) Selecting a DR foil for a raised seem sentence (e.g. (11c)) means selecting a picture in which Bart has the thought-bubble, and in that thought-bubble it is Lisa who is playing the saxophone. Hirsch and Wexler reason that children apply what they call a “think analysis” to raised seem sentences, since their interpretation of these sentences appears to be equivalent to taking the matrix clause subject as the experiencer and the actual experiencer (in the to-phrase) to be the agent of the lower clause. (12)

DR-foil error a. Bart seems to Lisa to be playing an instrument. b. = Bart thinks Lisa is playing an instrument

While Hirsch and Wexler’s results indicate significant change over time in children’s comprehension of these complex raising sentences with an overt experiencer, I believe a number of criticisms can be leveled both against the design of this experiment and the conclusions drawn. For example, the fact that thought-bubbles were used in all of the drawings, and this depiction was explained to children as indicating what a character was thinking about, could easily have biased children to a seem=think kind of interpretation of the raised seem sentences. In fact, to my mind it is not clear that thought-bubbles give an entirely accurate depiction of “seeming,” since I think of “seeming” as typically involving some kind of perception of something, not simply imagination. Moreover, Hirsch and Wexler point out that children tended to make errors with the think sentences in which the experiencer argument was switched (MR foils), so to say that children interpreted Bart seems to Lisa to be playing an instrument as Bart thinks Lisa is playing an instrument is not quite consistent with the children’s own interpretation of the sentence Bart thinks Lisa is playing an instrument. A further criticism is that, just as various researchers criticized Carol Chomsky’s investigation of tough-constructions because she investigated only a single lexical item (see Section 5.3.2 below), it is noteworthy that Hirsch and Wexler limited their study to the single raising verb seem. They justify their focus on this lexical item by claiming that it is the most frequent raising verb in speech to children (according to their search of the CHILDES database),

Children’s acquisition of displacing predicates

199

but they undoubtedly overlooked the verb going-to/gonna which, according to my searches, is vastly more frequent than seem; the raising verb used (to) is also more frequent than seem (see Section 6.1). Moreover, the lexical meaning of going-to/gonna is a simple future tense meaning and does not include anything about appearance or perception which could, plausibly, introduce some difficulties for young learners (as is argued, for instance, by Orfitelli (2012); her work is discussed below). Finally, although Hirsch and Wexler defend their use of the overt experiencer phrase (to Lisa above), it is highly probable that the inclusion of this phrase made the task harder, thus potentially masking earlier ability to comprehend raising sentences. In fact, Hyams and Snyder (in preparation) maintain that while raising over an overt experiencer is problematic for children (they dub this restriction the Universal Freezing Hypothesis, after Müller (1998)), raising without an intervening experiencer is unproblematic. Hirsch and Wexler themselves found that in a follow-up task with adults which measured Reaction Time (RT), adults took over a full second longer, and made marginally more errors, on the raised seem sentences than on either the think or unraised seem sentences. Thus, it seems to me imperative to test children’s comprehension of raising sentences without an experiencer argument. (We will return below to a study by Orfitelli (2012) that addresses the issue of intervening experiencers.) My own studies of children’s acquisition of raising verbs differ from that of Hirsch and Wexler in a number of relevant respects: I tested a number of raising predicates, not only seem, and I used sentences without an overt experiencer argument. In addition, an important observation about Hirsch and Wexler’s stimuli is that both the experiencer and embedded clause subject were animate. The use of two animate NPs was meant to “ensure” that children are not simply invoking a non-syntactic, semantically-based means of interpreting the sentence (see the beginning of Section 5.2). However, I contend that this is a false measure. If children correctly interpret a string of words in the presence of a semantic cue, such as animacy, this does not necessarily mean that children did not also parse the sentence syntactically. In keeping with Pinker’s Continuity assumption I submit that if children give an adult-like interpretation of a sentence, we should not simply reject the assumption that the child had an adult-like means of attaining that interpretation. While it remains possible that the child did use a different (i.e. non-syntactic) means of interpreting the sentence, the burden would then be on the researcher to explain how the child gets from this non-adult-like mechanism to the adult-like mechanism for interpretation. Furthermore, what we are really after here is an account of how children come to understand these structures. If the use of a semantic cue such

200

Animacy and children’s language

Table 5.2: Examples of test items in Becker (2006b) Sentence

Main verb type

Semantic compatibility between lower predicate and matrix subject

The flower wants to be pink The flower wants to fly away The hay seems to be on the ground The hay seems to be excited

control

compatible

control

incompatible

raising

compatible

raising

incompatible

as animacy (specifically, encountering an inanimate NP subject) provides a means of deriving this complex structure, then the child’s good performance on just these items (i.e. sentences with this semantic cue) is informative about the child’s learning process. Therefore, in my own experiments I also tested children’s understanding of raising constructions with inanimate subjects. In one experiment (Becker, 2006b) I presented children with sentences containing an inanimate matrix subject, either a raising or control matrix verb, and a lower (infinitive) predicate that was either semantically compatible or semantically incompatible with the matrix predicate (see Table 5.2). The two raising verbs tested were seem and appear, and the two control verbs used were want and try. I was interested in seeing whether children would look for a semantic relationship locally, between the main verb and main clause subject, or between the embedded predicate and the main clause subject, or both. The task was a sentence well-formedness judgment task, so upon hearing the sentence (paired with a picture), the child had to say whether the sentence, uttered by a puppet, sounded “OK” or “silly.” I reasoned that if children looked for a local semantic relationship between the main verb and matrix subject (presumably taking both types of verbs to be control verbs and assigning a semantic role like agent or experiencer to the subject) they should reject all of the sentences as “silly,” since inanimate NPs cannot be agents or experiencers. If children looked for a non-local semantic relationship between the matrix subject and the embedded predicate (i.e. taking all verbs to be raising verbs), then they should accept or reject each sentence only on the basis of the semantic compatibility between the lower predicate and the subject. That is, they should reject sentences where

Children’s acquisition of displacing predicates

201

Table 5.3: Predicted responses depending on child’s assumptions

Type control/compat. control/incompat. raising/compat. raising/incompat.

Item

All control

All raising

Adult-like

OK

OK

OK

Silly

Silly

Silly

The flower wants to be pink The flower wants to fly away The hay seems to be on the ground The hay seems to be excited

Reprinted with permission from MIT Press.

the lower predicate expressed things like “to fly away” or “to be excited” but accept sentences where the lower predicate was “to be pink” or “to be on the ground,” since these are plausible properties of inanimate things. On the other hand, if children looked for a semantic relationship both locally and long-distance, correctly according to whether the main verb was raising or control, they should give adult-like responses: they should reject all of the control sentences (since control verbs generally require agent or experiencer subjects and thus cannot have inanimate subjects) but should accept or reject the raising sentences according to the compatibility of the lower predicate. In Table 5.3 I present the predicted outcomes. The children who participated ranged in age from 3 to 5 years. The result revealed an interesting progression across the three age groups. As depicted in Table 5.4, while all groups of children gave variable responses, 3-year-olds tended to behave as if they categorized the predicates as raising predicates, 5-year-olds resembled the adult-like pattern, and 4-year-olds were in the middle. That is, the greatest shift across the age groups was in their responses to the control verb items with a semantically compatible lower predicate. Younger children were more willing to accept these items (e.g. The flower wants to be pink), while the older children rejected these sentences as silly. In Becker (2006b) I interpreted these results as pointing to a bias on the part of children to interpret these structures as raising structures, since all groups of children were significantly better than chance in evaluating the “goodness” of the raising sentences, and the pattern of errors the younger children made with the control sentences was in the direction of the expected pattern of responses

202

Animacy and children’s language

Table 5.4: Results of experiment: relative proportion of OK/silly responses 3-year-olds Type control/compat.

control/incompat.

raising/compat.

raising/incompat.

Item

OK

The flower wants to be pink The flower wants to fly away The hay seems to be on the ground The hay seems to be excited

Silly

4-year-olds OK

Silly

5-year-olds OK

Silly **

*

**

*

**

*

**

*

**

**

*p ≤ 0.05, **p ≤ 0.01 Reprinted with permission from MIT Press.

for raising items – they accepted or rejected sentences on the basis of the semantic compatibility of the lower predicate with the matrix subject. This conclusion was called into question by Hirsch and Wexler (2007) and Hirsch et al. (2008), who correctly pointed out that the prior literature indicates quite early acquisition of subject control (although, as noted above, virtually all of this literature focuses on object control, or subject control into adjuncts, and not subject control with want and try). But while this point is well taken, and I would not claim that 3-year-olds incorrectly believe that want means ‘seem,’ I would maintain that when a verb like want is used with an inanimate subject, as it was in my experiment, young children are more willing to allow a bending of its lexical meaning, such that it takes on more of the semantic properties of raising-like verbs. This “bending” of the lexical meaning of a verb is known as the Verb Mutability Effect (Gentner and France, 1988) (see Section 3.3.1.1). In the present case, I take children’s “bent” meaning of want to be something like a future or modal meaning (The flower wants to be pink = the flower might/will be pink), akin to the “raising-like” meanings of want adults allow in some cases (see the discussion of the raising–control continuum in Section 3.3.1). This interpretation

Children’s acquisition of displacing predicates

203

of the above data is supported by two further experiments I conducted with the same age group, and lends further support for the primary learning story offered here. Namely, an inanimate subject cues a raising (or more generally displacing) structure, which in turn limits the plausible lexical meanings of the predicate to those related to modality, tense, and a few others. In the first follow-up experiment, also reported in Becker (2006b), children were given a Truth-Value Judgment (TVJ) task (Crain and Nakayama, 1987) in which a puppet used a raising or a control sentence to comment on what had just happened in a scene acted out for the child. The child then had to say whether the puppet’s comment was “true” or “false” given the scenario just witnessed. The puppet’s comments contained only animate subjects and were constructed so that the child had to parse both the main clause verb and the lower clause predicate, separately, in order to answer correctly (the original purpose of this follow-up experiment was to ensure that children in the first task were not responding simply by ignoring the main clause predicate and answering only on the basis of the matrix subject and embedded predicate). For example, in one story a small horse expresses a wish to be bigger. The horse is offered a “magic” apple, which he eats, and then goes to sleep. Upon waking, the horse is now big (having been replaced by a larger toy behind the experimenter’s back). Following the scenario the puppet said, “The horse used to be small” (true condition) or “The horse used to be big” (false condition). In both cases, the child had to correctly interpret the raising verb used to since taking into account only the matrix subject and embedded predicate (“The horse . . . be small/big”) should have led to wrong answers. Instead, both 3- and 4-year-olds were significantly above chance in responding correctly (although, admittedly, 3-year-olds did not perform at ceiling: they were correct 64% of the time). For comparison, another set of scenarios were followed by a control-verb sentence by the puppet. In one example, a pig wanted to eat a doughnut but in fact ended up eating a banana. The follow-up sentence was “The pig wanted to eat the doughnut” (true) or “The pig wanted to eat the banana” (false). Again, children were significantly above chance in responding correctly. As both Hirsch and Wexler (2007) and Becker (2009) note, this result is in fact surprising if children “cannot understand” subject control verbs. However, in light of the fact that in this experiment all subjects were animate, this result is entirely unsurprising. I believe it was the inanimacy of the subjects in the initial experiment that led to the apparent raising bias, and so we would not expect that bias to extend to cases where control verbs are used with animate subjects.

204

Animacy and children’s language

Another aspect of this particular task is also revealing about the role of animacy. In a description of this task in Becker (2006a) I noted that one raising verb in particular posed problems for children: the verb tend. Unlike the other raising verbs tested (seem, happen-to, and used-to) it appeared that children simply had no idea what tend meant. One child even stopped the experiment to ask me explicitly what it meant. At the time the best course of action seemed to be to simply remove that item from analysis, and in fact, once tend was removed the results improved dramatically for both 3- and 4-year-olds (only one child out of nine responded correctly to the tend question, and it was the child who had asked for a definition). But viewed in another light, children’s terrible performance on items with tend suggests that since this verb was used with an animate subject, children were construing it as a control verb. The story that accompanied this item was as follows: (13)

This is a story about a horse who always ate hay. Every day, the horse ate hay for breakfast and lunch and dinner. Then one day, the farmer came along with a piece of chocolate cake. The farmer said to the horse, “You are such a good horse. Today I’m going to give you a treat. Instead of giving you hay to eat, I’m going to give you a piece of chocolate cake to eat!” The horse said, “Oh boy! I’ve never had chocolate cake before! I think I’ll eat some!” So the horse ate the chocolate cake.

Following the story, the child was offered either “The horse tends to eat hay” (true) or “The horse tends to eat chocolate cake” (false). In the story the horse in fact ate both hay and chocolate cake, so if children were simply ignoring the unknown verb tend, they should have answered randomly. But children almost uniformly answered “false” to the first question, and “true” to the second. Given the horse’s excitement over the offer of chocolate cake, a plausible interpretation of the test sentence is something like “The horse wanted to eat hay/chocolate cake” or “The horse was excited to eat hay/chocolate cake,” either of which would involve a control structure. The second follow-up experiment that points to a role for inanimate subjects in leading children to construe control verbs as having a raising-type interpretation comes from a sentence well-formedness judgment task reported in Becker (2009). In that experiment, children were shown pictures of weather events and were given sentences with a weather-it subject and either a raising or a control main verb, e.g. It seems to be snowing or It’s trying to be sunny. Children had to say whether the sentence was “OK” or “silly.” The purpose of this experiment had been to see whether children’s acceptance of control verbs with inanimate subjects in the first experiment in Becker (2006b) might be due to the fact that

Children’s acquisition of displacing predicates

205

they allow control verbs to select an inanimate subject as external argument, as opposed to allowing these verbs to function as raising verbs (which do not select any kind of external argument). The results of this study showed that both 3- and 4-year-olds were significantly above chance in accepting raising verbs in these weather contexts, and that 3-year-olds were also significantly above chance in accepting control verbs in these contexts. The 4-year-olds in the study were almost, but not quite, significantly above chance in accepting control verbs in weather contexts. What does it mean to say that a sentence like It’s trying to be sunny or It wants to be raining is grammatical? First, recall from Section 3.3.1.1 that want appears to exhibit some variable behavior as both a control and a raising verb, so to the extent that children permit want in raising contexts, they are not engaging in something completely outlawed by UG. Secondly, informal conversations with these children about what these sentences mean revealed that children in fact had quite sharp intuitions about their meanings, and therefore were not simply answering randomly. For example, one child explained that you could not say “It wants to be raining” if it was already raining, or if it was sunny, but you could if the sky had dark clouds in it. Thus, It wants to be raining means something like “It might rain” or “It’s likely to rain,” i.e. the sentence has a future or modal kind of interpretation. Future and modal meanings are precisely the kinds of meanings associated with raising predicates (cf. English gonna), and therefore this is the kind of meaning we would expect children to assign to these verbs if they were construing them as raising verbs. The upshot of these studies is that children at ages 3 and 4 years can interpret raising structures (without an overt experiencer argument), and when the subject of the sentence is inanimate they have a bias to interpret the construction as a raising construction, even if the main verb is a control verb in the adult language. By age 5, however, children are not so swayed by subject animacy. These older children stick to a control interpretation of sentences with control verbs, rejecting such sentences as “silly” if the subject is inanimate. To use the terminology of Lidz et al. (2004), these older children have become “verb compliant” rather than “frame compliant.” A slightly different perspective is taken by Orfitelli (2012). She claims that what is difficult for children is particular raising predicates such as seem and appear, which select an experiencer argument, whether or not that argument is expressed overtly. In contrast, raising predicates that do not select an experiencer (John is about/tends/is going (*to me) to be in the kitchen) present no difficulties for children. Thus, her claim is that A-movement per se is not problematic, contra Wexler (2004); Hirsch and Wexler (2007), but rather

206

Animacy and children’s language

A-movement over an (implicit or explicit) experiencer that is problematic. Her evidence for this claim comes from a number of studies showing that young children (a) make mistakes interpreting sentences with seem where the subject is referential (i.e. raised), whether or not the experiencer is overt, and (b) do not make mistakes on raising constructions whose predicates do not select an experiencer argument (e.g. going-to/gonna), even if that predicate is somewhat uncommon and perhaps semantically complex (tend). Furthermore, she shows that children’s good performance with these non-experiencer raising predicates does not stem from children misconstruing these predicates as control predicates, since they perform equally well when the subject is inanimate (and thus not a good subject of a control construction). While I would not be surprised if the semantics of seem and appear with their implicit experiencer involved some added semantic complexity (as compared to, for example gonna), Orfitelli’s study lacks a crucial comparison needed to dissociate the effect of the implied experiencer from the effect of animacy. In particular, she did not test seem (with or without an overt experiencer) with inanimate subjects. If such sentences were shown to be as difficult as those with an animate subject, then this would support Orfitelli’s claim that the crucial factor is the presence of an (overt or covert) experiencer. On the other hand, if seem with an inanimate subject is easier for children, then this would demonstrate a clear effect of animacy. The next step in this line of research is to test children with novel verbs in these opaque constructions, with animate or inanimate subjects, and see how children categorize the novel verbs. A small pilot study conducted in my lab with five 5-year-olds indicated a very promising effect of exactly this sort. Children were shown a video that depicted a short conversational scenario acted out with toys and props, in which the characters in the scenario used a novel verb five times either with inanimate subjects or with animate subjects (see Table 5.5). Following the video children were asked a simple yes/no question that had a syntactic form which either admitted a raising verb but not a control verb (i.e. a there-construction) or admitted a control verb but not a raising verb (i.e. a transitive construction). For example, children heard the questions in (14). (14)

a. Did there meb to be a banana in the soup? b. Did Mrs. Farmer meb the soup?

If the verb meb is a raising verb it should sound grammatical in (14a) but ungrammatical in (14b), but if meb is a control verb it should sound grammat-

Children’s acquisition of displacing predicates

207

Table 5.5: Response time in msec to transitive

vs. there-construction questions (5-year-olds) Subject used in video

Transitive question There question

Animate

Inanimate

1556 948

4084 1196

ical in (14b) but ungrammatical in (14a). How should children know whether meb is a raising or a control verb? Our hypothesis was that if meb was used exclusively with inanimate subjects in the video, it should be interpreted as a raising verb, but if it was used exclusively with animate subjects in the video, it should be interpreted instead as a control verb (or, it should remain unclassified, since an animate subject is actually uninformative about the category of the verb). In order to assess children’s categorizations we measured how long it took children to answer each question, on the premise that it will take children longer to respond to an ungrammatical prompt than a grammatical one (modeled after Naigles et al. (1995); more on this methodology below). It turned out that children indeed took measurably longer to answer questions that had a transitive structure than those with a there-construction if the novel verb had been used with inanimate subjects in the video. The difference was so stark and uniform across the children that the effect was highly significant even with data from only five children. While children were generally faster in answering the there-question than the transitive question, an analysis of variance on the mean estimate of the reaction time (under a log transform to normalize extreme values) revealed a significant interaction between question type and subject animacy. The lag in responding to the transitive question for verbs seen with inanimate subjects was significantly larger than that for verbs seen with animate subjects (μ = 3.50, χ 2 = 10.57, p = .001). Although a full study remains to be carried out, these preliminary results seem quite promising. To wrap up this section on raising verbs, I want to add a word about raisingto-object. As I noted earlier, the raising-to-object construction is not a major focus of this book. Recall from Section 2.1.3 that it is unlike the other opaque constructions in that it involves a derived object rather than a derived subject, and objects are canonically inanimate (or, less animate than subjects). Yet, we

208

Animacy and children’s language

find the same pattern here, where an inanimate NP signals a derived NP, even though it is in object position. This is because the “constructional homonym” of raising-to-object, namely object control, requires an animate NP in object position. (15)

a.

John expected the book to be boring.

b. # John persuaded the book to be boring.

Given this, it is actually object control that is somewhat unusual (or less prototypical) in that it requires an animate object. This might suggest that raising-to-object constructions might be easier for children to acquire, or acquired earlier, as compared to object control constructions. Very little work has examined children’s acquisition of raising-to-object. One study, however, directly compared children’s interpretations of raisingto-object and object control constructions. Kirby (2009) found that 4- and 5-year-olds were at ceiling in correctly interpreting both types of constructions. But she did find an advantage for raising-to-object constructions when the embedded clause was passivized. That is, both age groups were significantly above chance on the raising-to-object embedded passives (e.g. He wanted Tigger to be called by Elmo), but only 5-year-olds were above chance on the object control embedded passives (e.g. She told the policeman to be sniffed by the dog), and, even so, they performed far worse on these than the raising-to-object embedded passives. In fact, Kirby’s 4-year-olds performed marginally better on embedded passives under raising-to-object verbs than on monoclausal passives (The policeman was sniffed by the dog), and her 5-year-olds performed equally well on both kinds of passives. The suggestion is that there is something about the raising-to-object construction that facilitates the interpretation of a derived argument in its object position. 5.3.2 Acquisition of tough-constructions When Carol Chomsky set out to uncover how children learn the “true grammatical relations” between words when those relations “are not expressed directly on the surface structure,” one of the constructions she focused on was the tough-construction (TC). As described above in Section 2.2 the grammatical analysis of the TC in adult grammar has been quite contentious. But the clear hallmark of these constructions is that the syntactic subject of the matrix clause is the semantic object of the embedded clause, not the semantic subject of that clause (see (16)). In this, TCs contrast with superficially similar constructions involving control adjectives (see (17)).

Children’s acquisition of displacing predicates (16)

Johni is tough [PROarb to please t i ]

(17)

Johni is eager [PROi to please e]

209

Chomsky wanted to know if children would correctly interpret the subject of a TC as the embedded object, or incorrectly interpret the TC as if it were a control adjective construction. She presented children ages 5 to 10 years with a blindfolded doll, placed on a table, and asked the children, “Is the doll easy to see or hard to see?” Children gave two types of answers. One group of children (ages 5;0 to 8;5) answered that the doll was “hard to see,” and when asked, “Would you make her easy to see?” these children removed the doll’s blindfold. Such verbal and behavioral responses indicate that these children had interpreted the doll being “hard to see” to mean that it was hard for the doll to see; that is, they assigned the sentence a subject-gap interpretation. The other group of children (ages 5;2 to 10;0) answered that the doll was “easy to see,” and when asked, “Would you make her hard to see?” placed the doll under the table. These children had interpreted the initial question in an adult-like way (hence an object-gap interpretation). Although there was considerable overlap in the ages of the two groups of children, the average age of the children giving incorrect responses was lower (6;0) than that of the group that responded correctly (8;1). Chomsky’s interpretation of these results was that until the age of 8 to 10 years children could not reliably interpret TCs with the correct adult structure. Critics of Chomsky’s methodology and interpretation pointed out that the presentation of a blindfolded doll (while necessary for teasing apart the two potential interpretations of the TC) may have been misleading for children, causing them to give a non-adult-like response when such would not have been their normal interpretation of the question otherwise (Morsbach and Steel, 1975), and that it is hard to draw solid generalizations based on only a single test item (Cromer, 1970). In my own view, one could also object to the use of the verb see in the test sentences, since young children are known to sometimes make errors in mapping the “seer” and “see-ee” roles onto subject and object, respectively (Armbuster, 1981). An entertaining example of this error comes from one of my own children: my daughter had covered a toy elephant’s eyes with her fingers and announced, “Nobody see elephant!” Then, uncovering the elephant’s eyes, she declared, “Yeah-body see elephant!” The cuteness of yeah-body as the opposite of nobody aside, this example clearly demonstrates an error in mapping the roles of stimulus and perceiver to the arguments of see. In fact, the difficulty in establishing the perceiver as the subject and the

210

Animacy and children’s language

stimulus as object should not be surprising when one considers the circumstances under which young children typically hear the verb see used. In particular, parents of young children frequently use it in games like “peek-a-boo,” uttering “I see you” when the child’s eyes become visible; this usage provides conflicting evidence for the mapping of the thematic roles of see. Another frequent use of see is in the sense of “find out” (Let’s see if Grandma’s home), which does not necessarily imply a visual state (Landau and Gleitman, 1985). Apart from a few studies suggesting somewhat earlier comprehension of TCs (Cambon and Sinclair, 1974; Perry and Shwedel, 1979), subsequent experimental tests of children’s interpretation of TCs have largely replicated Chomsky’s initial finding that many children younger than 6 give inconsistent interpretations of TCs. Cromer (1970) used an act-out task and tested four exemplars of each of three types of adjectives: control adjectives, which admit only a subject-gap reading (e.g. glad; Cromer called these S-type adjectives), tough-adjectives, which admit only an object-gap reading (Cromer’s O-type), and a set of adjectives that, as noted above in Section 2.2.2, admit either a subject- or an object-gap reading (e.g. ready; Cromer’s A-type). After a pretest in which children had to demonstrate with a duck and a wolf puppet actions corresponding to “The wolf bites the duck” and vice versa, children were asked to act out sentences like “The wolf/duck is happy to bite” (S-adjective), “The wolf/duck is easy to bite” (O-adjective), and “The wolf/duck is nice to bite” (A-adjective). While it is not clear what it means to enact one character being “easy” or “nice” to bite, Cromer was interested in whether children would take the syntactic subject of the sentence to be the agent of biting (indicating a subject-gap interpretation) or the patient (indicating an object-gap interpretation). He found that children with a mental age (MA) of 5;9 to 6;6 (according to scores on the Peabody Picture Vocabulary Test) consistently allowed both subject-gap and object-gap interpretations of TCs (he classified these children as “Intermediates”). These children also gave inconsistent responses to the Aadjectives, and many of them, when retested the very next day, gave responses that conflicted with their responses of the day before (sometimes correcting a previously incorrect response, but sometimes going the other way). Younger children (MA 2;11–5;7) consistently gave subject-gap interpretations, incorrectly to tough-adjectives, correctly to control adjectives, and almost uniformly to the A-adjectives (Cromer termed these children “Primitive Rule users”). Only the most advanced children (MA 6;8–10;8) consistently gave object-gap interpretations for tough-adjectives and subject-gap interpretations for control adjectives, and did not uniformly assume a subject-gap with A-adjectives (he called these children “Passers”). Like Chomsky, Cromer concluded that only

Children’s acquisition of displacing predicates

211

the Passers could represent the adult-like syntax of TCs; Primitive-Rule users adopted a simple rule of subject equals agent, and Intermediates were guessing. A recent study by Anderson (2005) likewise found that children between the ages of 3;4 and 6;3 gave inconsistent responses to TCs. In a truth-value judgment task, Anderson acted out scenarios for children with toys and then had a puppet comment on what had happened. For example, in one scenario a dog teaches a pig how to play soccer, and he does so quite successfully. But when the pig then tries to teach the dog how to slide down a slide the teaching is unsuccessful: the dog does it wrong and then is distracted by a cat, and so never manages to learn how to slide properly. So the post-scenario sentence The dog was difficult to teach is true on the adult reading. But if that sentence were interpreted to mean “It was difficult for the dog to teach” (i.e. the subject gap interpretation), the sentence should be judged false. Anderson tested a variety of tough-adjectives, easy, hard, and difficult. Although there was some variation among the different predicates in terms of children’s rate of correct (adult-like) responses, it was far less than the variation between the age groups: 3-year-olds’ performance ranged from 35% to 45% correct (mean 39.7%), 4-year-olds from 41% to 56% (mean 45.0%), and 5-year-olds from 43% to 61% (mean 54.6%), with worse performance associated with difficult for the youngest children, and with easy for the 4- and 5-year-olds. Six-yearolds’ performance, however, ranged from 77% to 88% correct (mean 82.3%; worst performance on hard). Statistical analysis revealed that the three younger age groups behaved alike, in contrast to the single oldest age group, who were statistically indistinguishable from adult controls. Anderson’s account of children’s behavior differs from previous accounts in that she does not link children’s difficulty with problems representing the syntax of tough-constructions. The fact that children in all the age groups she studied gave evidence of object-gap interpretations for some items led her to believe that children could in fact represent the adult structure of TCs. In addition, children who were asked to justify their responses to particular questions (e.g. “Why was the dog difficult/not difficult to teach?”) gave appropriate justifications; that is, their justifications were largely consistent with the true/false answer they had given. Thus, unlike Cromer (1970), she did not conclude that children’s inconsistent responses were indicative of a guessing strategy. Rather, she argued that children are capable of representing both the adult-like TC structure and a subject-gap structure (on analogy to the correct structure of control adjectives like eager); they sometimes project the subject-gap structure for tough-adjectives because they have a preference to interpret the matrix subject as the same as the embedded subject. Anderson points out that in an earlier

212

Animacy and children’s language

stage of English (until sometime in the nineteenth century), and in several other languages today, such an interpretation is in fact grammatical (Finnish, Mandarin; see Section 3.3.1). Thus, English-speaking children’s acceptance of subject-gap interpretations of TCs is unsurprising from a UG perspective: the type of ambiguity in TCs displayed by children acquiring English is sanctioned by Universal Grammar. What is interesting, as Anderson points out, is that no language appears to allow only the subject-gap interpretation of tough-constructions. This fact would appear to undermine the view that English-speaking children’s apparent preference for a subject-gap reading of English TCs arises from markedness pressures. At least, the subject-gap interpretation of TCs does not appear to be the unmarked option. Additional studies of childrens’ tough-constructions were conducted by Cambon and Sinclair (1974); Solan (1978); Macaruso et al. (1993); McKee (1997a,b), among others, and studies of related object-gap constructions were conducted by Goodluck and Behne (1992) (e.g. purpose constructions). (Please see Anderson (2005) for a comprehensive overview of the literature.) The general consensus is that these constructions present difficulties for young children, though the explanation for why this is so differs across the literature. A number of researchers point to the particular syntactic structure, which, as discussed in Section 2.2.1, is generally agreed to be unusual in a number of respects (Chomsky, 1969; Cromer, 1970; Goodluck and Behne, 1992). Others have suggested difficulties due to processing load (Macaruso et al., 1993), or lexical knowledge of the adjectives involved (McKee, 1997a,b). Still others have suggested nonlinguistic cognitive challenges, such as taking another’s point of view (Cambon and Sinclair, 1974). Consistent with the experimentally demonstrated difficulty with TCs (regardless of the underlying cause) is the fact that true tough-constructions are quite rare, both in children’s spontaneous speech (Bloom, 1970; Nishigauchi and Roeper, 1987) and in adult speech. Mair (1990, p. 67) reports finding only eighty-five occurrences in a corpus of 875,000 utterances of English. In my own search of 248,555 maternal utterances in several corpora of child-directed speech (Bloom, 1970; Brown, 1973; Suppes, 1974; Kuczaj, 1977; Sachs, 1983; Clark, 1987; Valian, 1991; MacWhinney, 2000) I found 116 occurrences of tough-constructions with a referential NP subject; tough-adjectives occurred an additional seventy-five times with an expletive subject in this dataset (see discussion of these frequences in Section 6.1). In their own speech children will sometimes produce elided forms like “hard to read” (Bloom et al., 1984), or what Anderson calls degree or object purpose constructions (“It’s too big

Children’s acquisition of displacing predicates

213

for you to eat,” “Grandma has a present for me to blow on”; Nishigauchi and Roeper (1987)), but few tough-constructions proper, with an unambiguous tough-adjective and a derived subject. When I searched the corpora of Bloom (1970), Brown (1973), Suppes (1974), Kuczaj (1977), Sachs (1983), Valian (1991), and MacWhinney (2000) (total child utterances: 248,940) I found thirty-four spontaneous productions of tough-constructions with an overt derived subject. An additional eighteen utterances had a tough-adjective with an expletive subject (so, underived). A further eight utterances had a tough-adjective but either no overt subject or a subject whose status could not be determined. Some of these utterances are exemplified in (18). (18)

a. The corner’s hard to get down, right? (referential derived subject; Sarah age 4;4 from file 103) b. No, it’s hard to open it. (expletive subject; Peter age 2;8 from Bloom, 1970 file 17) c. Hard to be out. (null subject; Sarah age 4;0 from file 089)

Interestingly, none of the children’s spontaneous utterances contained animate subjects. The only utterance that might have had an animate subject but was difficult to parse was Abe’s You are starting hard to get off and sit on the couch (Kuczaj file 086). As we will see in Section 6.1 this asymmetry in subject animacy is also true of adult productions of tough-constructions. What all of these approaches focus on is the what side of the acquisition question (What do children know about these constructions in their target language, and when do they know it?), but not the how side of things (How do children eventually come to learn them?). Only one study, Cromer (1970), attempted to teach children a novel tough-adjective in a particular context and measure how children categorized that adjective. Since Cromer’s novel adjective study bears some similarities to my own study, described below, it is worth looking at in detail. In addition to his puppet act-out task, in which children had to demonstrate the meaning of sentences like The wolf is easy/happy to bite, Cromer also presented children with two novel adjectives, larsp and risp, each in a context that favored either a tough-adjective or a control adjective interpretation based on the sentence frame in which it was introduced. For the novel control adjective, Cromer showed children a picture of a dog holding a bone and gave the following input: “See? Someone gave this dog a bone. And so he’s feeling very risp. He’s feeling very risp.” And for the novel tough-adjective, Cromer showed them a picture of a cat with a rose and gave this narration: “This cat climbed up and picked a rose. He found that chewing the rose was larsp. Chewing the

214

Animacy and children’s language

rose was larsp.” Crucially, the frame in which each adjective was used allows only one type of adjective and not the other (He’s feeling very happy/*easy; Chewing the rose was easy/*happy). After only these two exemplars of each adjective, the children had to act out “The wolf is risp/larsp to bite.” Cromer found that although his “Primitive Rule” and “Intermediate” children largely interpreted the subject as the semantic subject of the embedded verb, all but one of his “Passers,” those children who gave uniformly adult-like interpretations of English tough-adjectives, correctly distinguished the two novel adjectives. That is, they interpreted risp as a control adjective, upon hearing it twice in the frame “NP is feeling very risp,” and larsp as a tough-adjective, upon hearing it twice in the frame “VP-ing was larsp.” Cromer’s results show that children use sentence frames as a cue to adjective categorization, and that this categorization process can happen very quickly, once the category is clearly in place. Bruno Estigarribia, Duna Gylfadottír, and I wanted to know whether children would categorize novel adjectives as either tough or control adjectives upon hearing them in the frame “NP is Adjective to VP” according to whether the NP subject was animate or inanimate. We also wondered whether children could perform this categorization at a younger age than Cromer’s children, namely at an age at which children are reported to not yet be able to represent tough-constructions.12 We presented forty children ages 4 to 7 years with novel adjectives used in conversational contexts, and we systematically varied the animacy of the sentence subject. Children viewed videos in which toy characters had conversations about events taking place in the video. During the course of their conversation, which lasted about one minute, a novel adjective was used five times (a portion of this study is published in Becker et al. (2012), and a fuller description is given in Becker (to appear)). To maximize coherence in the conversations, the novel adjectives were used with a consistent meaning. Our two novel tough-adjectives were used consistently with a meaning akin to “easy” (that is, one could have substituted the word easy into the conversations and everything would have made sense), and our two novel control adjectives were used consistently with a meaning akin to “happy” or “willing” or “excited.” For example, in one video Mrs. Farmer wants to buy Mr. Farmer a surprise birthday present and asks the nurse for 12 It is difficult to compare our subject samples directly, since Cromer reports only Mental Age as

measured by scores on the Peabody Picture Vocabulary Test (PPVT), whereas we report only chronological age.

Children’s acquisition of displacing predicates

215

advice on what to get him. The limitation that Mrs. Farmer imposes is that the gift must fit inside a box (shown in the video) so that it will be a surprise. The nurse and Mrs. Farmer consider both a motorcycle and a ladder, neither of which are deemed “stroppy [=easy] to hide” because they do not fit in the box, before settling on a wristwatch, which is “stroppy to hide” because it is small, and does fit in the box. In another video, a mother and her two children talk about the children’s first day at school, and one child reports that her teacher was “narpy [=excited] to teach” because she gave the kids lots of fun projects to do, while the other child reports that his teacher was “not narpy to teach” because she made them sit quietly all day doing nothing. Thus, there were contextual cues in the videos to the meanings of the predicates, but the predicates themselves were not defined for the children explicitly. After viewing each video, the child participant was asked two questions by a puppet, the order counterbalanced across participants. One question was always of the form “Is it Adjective to VP?” (e.g. Is it stroppy to hide a motorcycle in the box?), which should sound grammatical if the adjective has been categorized as a tough-adjective (cf. Is it easy to hide a motorcycle. . . ?) but ungrammatical if it is construed as a control adjective (cf. *Is it eager to hide a motorcycle. . . ?). The other question was of the form “Is NP Adjective?” (e.g. Is the motorcycle stroppy?) which should sound grammatical if the adjective is taken to be a control adjective but ungrammatical if taken to be a toughadjective. Notice that in this case, a question like Is the motorcycle eager? is also impossible for adults, but only because of the meaning of the control adjective. That is, if children took The watch is stroppy to hide to be grammatical with a control interpretation, then they should also allow Is the motorcycle stroppy? In contrast, if they took a sentence like The watch is stroppy to hide to indicate that stroppy must be a tough-adjective, since it allows an inanimate subject, then Is the motorcycle stroppy? should sound ungrammatical. Our prediction was that children would indeed use the animacy of the subject as a means of categorizing the novel adjectives, and to preview our results we found a very strong effect in the case of adjectives used with an inanimate subject. That is, children who heard adjectives used only with inanimate subjects robustly categorized these adjectives as tough-adjectives. Before presenting the results of our study, I must note some further details about the design. Adjective type was manipulated within participants, meaning that all children were exposed to our two novel tough-adjectives, and our two novel control adjectives. In addition, we manipulated the subject animacy variable for novel tough-adjectives between subjects. Recall that the adjectives were always used with a particular contextual meaning, either “easy”

216

Animacy and children’s language

or “happy/excited.” In order to isolate the effects of subject animacy versus contextual meaning we divided our participants into two groups (twenty children in each, with roughly equal numbers of children in each age group from 4 to 7 years). Group 1 heard our novel tough-adjectives only with inanimate subjects and our novel control adjectives only with animate subjects, as in the examples above. Group 2 heard all adjectives, both tough and control, only with animate subjects. Thus, the videos for the control adjectives were the same between the two groups, but the videos for the tough-adjectives differed. The tough-adjective videos for Group 2 held the contextual meaning of the adjectives the same as in Group 1 (stroppy still meant something like “easy”), but the sentence subjects were all animate. In the dialogue for stroppy for Group 2 Mrs. Farmer wanted to throw a surprise birthday party for Mr. Farmer, but this time instead of hiding a gift she wanted to hide her friends, and then have her friends jump out and surprise Mr. Farmer. So, for example, the nurse was said to be “stroppy to hide” because she was very quiet, while the policeman was “not stroppy to hide” because he was making a lot of noise (see Table 5.6). We wanted to know how children categorized these novel adjectives (if they did at all) after hearing them only five times in an enacted conversation, but how could we determine children’s categorization? Since we wanted to Table 5.6: Design of tough-adjective study Novel Adj Type Group 1: tough

control

Group 2: tough

control

Adjective

Approx. meaning

Example sentence

daxy stroppy

easy easy

greppy

willing/ happy

narpy

excited/ happy

Apples are very daxy to draw. A watch is small and very stroppy to hide. I’m sure Mr. Farmer would be greppy to help. If my teacher isn’t narpy to teach I won’t learn anything!

daxy

easy

stroppy

easy

greppy narpy

willing/ happy excited/ happy

The policeman is not daxy to draw. She [the nurse] is stroppy to hide because she is very quiet. (same as Group 1) (same as Group 1)

Children’s acquisition of displacing predicates

217

know whether children had categorized a given adjective as belonging to the tough or control class, we were not necessarily interested in knowing exactly what these adjectives meant for the children. Certainly, the general meaning of the adjective would be pertinent to its categorization, but the precise lexical meaning was unimportant. Therefore we could not employ a preferential looking type of task, which assesses children’s interpretation of a given word or sentence (i.e. they look towards the scene that matches the meaning of the word or sentence). Moreover, meanings such as ‘easy’ and ‘excited’ are sufficiently abstract that I am not sure how such meanings could be easily and unambiguously depicted. Rather we were after a grammaticality judgment. But while children can be trained to give metalinguistic grammaticality judgments for certain types of ungrammaticality (McDaniel and Cairns, 1996), my own experience is that it is virtually impossible to get children under the age of about 7 years to give reliable metalinguistic grammaticality judgments on constructions involving expletives, which was one of our test constructions. Instead, we adopted a methodology used by Naigles et al. (1995), in which we measured how long it took children to answer a yes/no question. The premise of this methodology is that it will take children longer to respond to an ungrammatical prompt or question than a grammatical one, which is what Naigles et al. (1995) found. (This was the methodology used to assess children’s categorization of novel raising and control verbs described in Section 5.3.1; a similar methodology was used by Corrigan (1988) who found longer response times to less prototypical sentences.)13 As a check of the methodology, we exposed our participants to two warmup videos which involved only real English words, two filler videos which included a novel transitive or intransitive verb, and two videos in which they heard real English adjectives, one tough-adjective (hard) and one control adjective (afraid).14 In all of these cases, the post-video questions included one that should be grammatical (e.g. Did the farmer play with the car or Did the nurse borrow a basket?) and one that should sound ungrammatical based on an argument structure violation (Did the farmer play the car to his friend? or 13 Goodluck et al. (1995) also measured reaction time in older children’s processing of object-gap

purpose constructions using a self-paced reading task. Interestingly, these authors found that children (age 10) did not seem to show a preference for the gap being in object position when the preceding context favored that kind of interpretation, although the adults in their study did. 14 Initially we tried using the English control adjective eager, but children indicated confusion about what it meant. (Anderson (2005) and Macaruso et al. (1993) also report lack of comprehension of eager.) Afraid appeared to be familiar to the children we interviewed; at least, they did not ask us what it meant or otherwise indicate confusion about it.

218

Animacy and children’s language

3.4 3.3 3.2

RT (log10)

3.1

3 2.9 grammatical 2.8

ungrammatical

2.7 2.6 2.5

pl

ay

ow rr bo b

lo al

p(

i

r nt

an

s)

s) an tr ( rp zo

Figure 5.1. Mean RT in msec (log10) to warm-ups and fillers

Did the nurse borrow?). As shown in Figures 5.1 and 5.2 below, children uniformly and significantly responded more slowly to the ungrammatical than the grammatical questions, even for hard and afraid.15 Now let us turn to our target items, the novel tough- and control adjectives. As noted above, we constructed our two post-video questions so that one question would sound grammatical if the adjective had been categorized as a tough-adjective, but ungrammatical if it had been categorized as a control adjective (19), and the other question would sound grammatical if the adjective had been categorized as a control adjective, but ungrammatical if it were taken to be a tough-adjective (20). Thus, if our hypothesis was correct and children used the inanimate subject to categorize the novel adjective as a toughadjective, they should answer question (19) faster than question (20), for those adjectives heard only with inanimate subjects. (19)

Is it stroppy to hide a motorcycle in the box?

(20)

Is the motorcycle stroppy?

15 In all of the following figures the y-axis measures the log10 of the RT duration as measured in

milliseconds (msec). The reason for using the log10 transformation on these measurements is to smooth over variance due to extremely long or extremely short response times.

Children’s acquisition of displacing predicates

219

3.5 3.4 3.3

RT (log10)

3.2 3.1 grammatical 3

ungrammatical

2.9 2.8 2.7 2.6 hard

afraid

Figure 5.2. Mean RT in msec (log10) to hard and afraid

In fact, this is precisely what we found. Children in Group 1 were significantly faster in answering question (19) than question (20) for the adjectives used with inanimate subjects, suggesting that hearing the adjective used with an inanimate subject led the children to categorize the novel adjective as a tough-adjective. See Figure 5.3. In contrast to the novel tough-adjectives, though, these children were not significantly faster in answering one question or the other for the novel control adjectives (those that meant roughly ‘excited’ and were used with animate subjects), suggesting that in fact they had not uniformly categorized adjectives heard with animate subjects as control adjectives. Initially we were surprised by this result. Recall that in our experiment with adults on categorization of novel raising and control verbs (see Section 4.2.2), adults had a strong bias to construe an animate subject as the external argument of the main predicate, i.e. they showed a bias toward a control categorization. We had therefore expected children to show this bias with respect to the adjectives used with animate subjects. However, recall that since animate subjects are actually uninformative as to the category of a predicate (both tough and control adjectives are perfectly grammatical with them), children should have had no basis for categorization. Thus, it appears that children may not have quite the same

220

Animacy and children’s language 3.3 3.2

RT (log10)

3.1 3 grammatical

2.9

ungrammatical 2.8 2.7 2.6 2.5 tough-adj

control-adj

Figure 5.3. Mean RT in msec (log10) for Group 1

bias that adults do for interpreting an animate subject as the semantic subject in these biclausal constructions. The source of adults’ control bias may stem from frequency effects (see Section 6.1). Another possibility is that the difference stems from a difference between the two constructions we were testing: raising vs. control verbs with adults, and tough vs. control adjectives with children. In our current on-going work we are testing this categorization process in raising and control verbs in children (see end of Section 5.3.1), and so a fuller picture must await those results. The children who heard novel tough-adjectives used with inanimate subjects categorized these novel predicates as tough-adjectives, but recall that there were also contextual cues to the adjectives’ meanings. As described above our novel tough-adjectives were used consistently with a contextual meaning of ‘easy,’ and perhaps it was this contextual meaning that allowed children to categorize the adjectives rather than hearing them used with inanimate subjects. So let’s now turn to the children in Group 2, who heard all adjectives used with only animate subjects (see Figure 5.4). Here the results were quite different. The log10 data in the graph are slightly misleading, because the contrast between the responses to grammatical and ungrammatical questions for novel tough-adjectives appears to be very large

Children’s acquisition of displacing predicates

221

3.4

3.3

RT (log10)

3.2

grammatical

3.1

ungrammatical 3

2.9

2.8 tough-adj

control-adj

Figure 5.4. Mean RT in msec (log10) for Group 2

when in fact it is not significant. In Table 5.7 I provide the raw response times in milliseconds. In the table we can see that the response times to grammatical and ungrammatical questions for novel tough-adjectives are very similar and in fact go slightly in the wrong direction (slightly longer RT for the grammatical question). However, the contrast between the two types of questions for the novel control adjectives was significant, suggesting that children were able to use the situational context of the scenarios to home in on the adjectives’ general meanings. At least, children appeared to be able to glean enough cues from the context of the scenario to guess that the novel control adjectives meant something relating to emotion or other mental states, such that they could be used in the frame Is the NP Adj? (e.g. Is the teacher excited/happy/ afraid?). A rather surprising aspect of these results is that we had expected the categorization of control adjectives to look the same in the two groups of children, because the stimuli were identical. And yet, while the children in Group 1 showed no difference in response time to the grammatical and ungrammatical questions, the Group 2 children did manage to answer the grammatical questions more quickly.

222

Animacy and children’s language

Table 5.7: Response Time (msec) for novel toughadjectives, group 2 Adjective type novel tough-adjective novel control adjective

Grammatical

Ungrammatical

1964.5 1732.5

1723.8 2841.3

I think there are two important implications of these results. One is that context did play a significant role in children’s categorization of certain adjectives. The fact that the Group 2 children successfully categorized the novel control adjectives without any syntactic cues shows that there are important cues from the situational context (i.e. the context of the story) which can be informative about predicate meanings. Such an effect of situational context might be similar to that found by Papafragou et al. (2007), where the context of a false belief situation (i.e. one character was tricked by another into thinking the wrong thing) contributed to children’s ability to draw inferences about the meanings of some mental or credal verbs (think, believe, know). While such a claim might appear to run counter to the Syntactic Bootstrapping hypothesis (since there the claim is that abstract predicate meanings cannot be acquired primarily on the basis of observing the world), Papafragou et al. maintained, as I do, that sentence-level cues still play a more informative role than context-level cues. Specific evidence that supports this view is that Group 2 participants were not able to reliably categorize novel tough-adjectives in the absence of the subject animacy cue (see Table 5.7). The other implication of our child tough-construction results is that inanimate subjects are actually an extremely powerful cue when they are present. There are two aspects of the data above that indicate this, plus additional evidence from a further study, described below. The first thing to note is that when we compare Group 1 and Group 2 directly, we find that the children who heard novel tough-adjectives only with inanimate subjects (Group 1) were significantly faster in answering the grammatical questions (Is it stroppy to hide a motorcycle in the box?) than the children who heard these adjectives only with animate subjects (Group 2) (see Figure 5.5). Secondly, I would argue that the reason for the Group 1 children’s failure to categorize novel control adjectives was actually the presence of inanimate subjects in the other stimuli they saw (i.e. the tough-adjective videos). Recall that adjective type was manipulated within participants, so each child

Children’s acquisition of displacing predicates

223

3.3 3.2

RT (log10)

3.1 3 tough-grammatical

2.9

tough-ungrammatical 2.8 2.7 2.6 2.5 Group 1

Group 2

Figure 5.5. Mean RT in msec (log10) for Groups 1 and 2, novel tough-adjectives

in Group 1 heard both tough-adjectives (with inanimate subjects) and control adjectives (with animate subjects). In other words, hearing inanimate subjects with tough-adjectives in those video conversations interfered with their ability to categorize the control adjectives encountered in different video stimuli. This interference is reminiscent of an effect that turned up in the adult experiment described in Section 4.2.2 (Becker and Estigarribia, 2013). In that experiment we found that when participants received the cue of an inanimate subject for novel raising verbs (i.e. the inanimate subject condition), not only were they better at categorizing raising verbs, but they were worse at categorizing novel control verbs (these were fillers) as compared to the definition-only (i.e. animate subject) condition. Although this interference can be seen in the combined data shown in Figure 4.1 in Chapter 4, we see the effect most clearly in the 1-exemplar condition, shown here in Figure 5.6.16 The source of this interference effect is unknown, but it could involve a kind of structural priming: encountering an inanimate subject in some stimuli 16 Another possible explanation for the worse performance on control-like fillers in the adult

experiment is the lack of lexical definitions in the Animacy condition.

224

Animacy and children’s language 100 90 80

Percent Correct

70 60 50 Raising (target) 40

Control (iller)

30 20 10 0 Deinition

Animacy 1 Exemplar

Both

Figure 5.6. Percentage of correct categorizations of novel raising (target) and control (filler) verbs in adult experiment, 1-exemplar condition (Becker and Estigarribia, 2013)

primes participants to project a structure with a displaced subject, and this primed structure influences their responses to other stimuli in the set. Thus, the contamination effect in Group 1 points to the fact that inanimate subjects are such a powerful cue to sentence structure (that is, to the subject being a displaced subject), that in its presence all other cues to predicate categorization, from biases to contextual cues, are diminished. Thirdly, an additional study (Becker, to appear) extended the novel toughadjective study to even younger children. In that study I taught 3- and 4-year-olds novel adjectives with animate or inanimate subjects but without any context at all – stimuli are semantically pared down conversations modeled after the conversations used in Yuan and Fisher (2009) (see (25) below in Section 5.3.3). An example of the conversations I used is given in (21–22). (21)

animate subjects A: The baker is stroppy to push! B: Really? The baker is stroppy to push? A: Yeah, the baker is stroppy to push.

Children’s acquisition of displacing predicates B: A: (22)

225

What about the bus driver? Is the bus driver stroppy to push? No, the bus driver is not stroppy to push.

inanimate subjects A: The chair is stroppy to push! B: Really? The chair is stroppy to push? A: Yeah, the chair is stroppy to push. B: What about the book? Is the book stroppy to push? A: No, the book is not stroppy to push.

The results of this study showed that 3-year-olds behaved just like the older children in Group 1 of the earlier experiment: given an inanimate subject, children categorized a novel adjective as a tough-adjective, but given an animate subject children did not categorize the novel adjective either way. Note that in this experiment subject animacy was manipulated between participants, so there is no contamination effect here, but there was also no contextual evidence children could have used to help figure out what the novel adjectives meant. Four-year-olds, on the other hand, behaved quite differently: when the subject was inanimate these children did not appear to categorize the novel adjectives one way or another. These findings contribute to both of the claims made above: for children of all ages, an inanimate subject is strongly indicative of a displacing structure; younger children can use this cue in the absence of situational cues, while older children (4 and up) use this cue but only in conjunction with situational cues. A potential objection to my interpretation of these results is that children may have been incorrectly interpreting the it in the questions of the form Is it Adj to VP? (Is it stroppy to hide a motorcycle?) as referential it rather than expletive it, and therefore their faster responses to this type of question for novel tough-adjectives is not necessarily indicative of a tough-adjective categorization. That is, if children interpreted it as referential, the question Is it eager to hide a motorcycle in the box? would be grammatical. I have two responses to this objection. One is that since it is inanimate, even referential it should still not really be a possible subject for a control adjective. My other response is that if children interpreted both the Is it Adjective to VP? and the Is NP Adjective? questions as control structures, we should not have seen such a dramatic asymmetry between their responses to the two types of questions. Especially, we should not have seen this asymmetry for one type of novel adjective and not the other type of novel adjective. In the context of previous work on children’s acquisition of TCs an important observation about our results is that in the first study described above (Becker et al., 2012) we found no age effects among our child participants.

226

Animacy and children’s language

That is, the 4-year-olds and 7-year-olds behaved alike. Moreover, in the later study even 3-year-olds were able to represent TCs when the stimuli were made simple enough. This is somewhat surprising, given the prior literature’s virtually unanimous finding that TCs present difficulties for children until at least age 6. And recall that in Cromer’s novel adjective study, only his “Passers” correctly categorized novel tough-adjectives. But in our study even 3- and 4-year-olds had no difficulty mapping novel adjectives, heard only five times, to a tough-adjective category. Why were our participants able to do this? Echoing Anderson (2005), I submit that the structure of tough-constructions is not inherently difficult to represent, but because of the surface ambiguity of a sentence with an animate subject in the string NPanim is Adjective to VP children would plausibly have difficulty determining that the string could be mapped to the TC structure (with object gap) as opposed to one of the other possible structures, some of which have subject gaps (e.g. subject control, subject purpose constructions). In all of the previous literature, animate or animate-like NPs (e.g. the doll) were used. Given an animate subject, children may have been led to misanalyze known tough-adjectives either as control adjectives, or as tough-adjectives that allow a subject gap (as is permitted by UG; see Section 3.3.2). We showed that an inanimate subject is an unambiguous cue and therefore leads children to the correct interpretation of tough-adjectives.17 One could still object that the cue of animacy allows children to use a purely semantic rule for interpreting these structures, so that they are not in fact representing the syntax of tough-constructions. How do we know that children do not simply uniformly interpret inanimate NPs as patients, so that NPinanim is Adjective to VP implies an object gap without any syntactic machinery at all? Such a suggestion was made by Carol Chomsky, who conceded that children could correctly interpret sentences like The book is hard to read earlier than The doll is easy to see by means of a purely semantic mechanism. As noted at the beginning of Section 5.3 I appeal to Pinker’s (1984) argument for the Continuity Principle. That is, if children get the same interpretation as adults but via alternative means, we must then explain how they come to abandon those alternative means and take on the adult mechanism. If, instead, children’s apparent likeness to adults is indicative of a deep likeness, i.e. they are using the same mechanism to reach the same interpretation, our developmental explanation

17 Cromer’s novel adjective task did not have this problem, since his novel tough-adjectives were

used with a VP subject (Chewing the rose was larsp). However, it is possible that he simply gave his participants too few exposures (two, as opposed to our five), or not enough contextual support to promote categorization in his younger subjects.

Children’s acquisition of displacing predicates

227

is simpler: we don’t have to explain how children change their fundamental mechanism for deriving sentence structures. The foregoing discussion and evidence have focused on the utility of inanimate subjects in cueing the underlying structure of tough-constructions. But more likely it serves as a more general cue that the surface subject is displaced, and in these adjectival constructions inanimate subjects seem to be quite strongly linked to an underlying object position. That is, it should work equally for other kinds of so-called Null Operator Structures, such as degree constructions (e.g. These rocks are too heavy [PRO to lift t]), ready constructions (The sandwich is ready [PRO to eat t]), and others noted above, even if they are not, strictly speaking, tough-constructions. Crucially for my purposes, in each such construction in which the subject is linked to an object gap, the subject can be inanimate. When it is obligatorily linked to a subject gap, as with control adjectives, it must be animate. The question that remains, then, is how children eventually come to distinguish among these different kinds of Null Operator Structures with object gaps. As mentioned in Section 2.5 expletives are an important clue here: they occur with true tough-adjectives but not other predicates that occur in these other types of object gap constructions, like ready. Thus, children will be able to use the different distributional properties of these constructions to further iron out their differences. 5.3.3 Acquisition of unaccusatives The opaque constructions discussed thus far contain a biclausal structure: the main clause predicate selects a clausal complement, and the matrix subject is derived from within that infinitive complement clause, either from the subject (raising-to-subject) or object (tough-construction) position. We also saw in Chapter 2 that there are monoclausal constructions in which the subject is derived; namely, unaccusative constructions and the passive. Can inanimate subjects cue derived subjects in these monoclausal constructions as well? Let’s consider why the cue of subject animacy might work differently in the case of monoclausal constructions. Verbs that select both an external argument and a clausal complement, as opposed to an NP complement, tend to have meanings associated with mental states, desires, and otherwise properties of sentient beings (think, know, believe, say, want, decide, claim, try, etc.). Inanimate entities contrast rather sharply with the set of entities that can be plausible subjects for these verbs, and so encountering an inanimate subject in the main clause of a biclausal structure is unexpected if the subject is an external argument. This is why inanimate subjects serve as a good cue to a derived subject in these cases.

228

Animacy and children’s language

The situation is a bit different for verbs that select an NP complement. Transitive and intransitive verbs include a wide array of lexical meanings, denoting all manner of motions, events, states, and causative relations. Psycholinguistic work has shown that verbs that occur in transitive or intransitive frames are relatively unconstrained in their possible meanings, as compared to verbs that take clausal complements (Kako, 1998). So we can imagine that encountering an inanimate subject with a transitive or intransitive verb might be relatively uninformative. And yet, we saw in Section 2.3.1 that there is a semantic component to the unaccusative/unergative distinction: unaccusative verbs easily permit inanimate subjects and generally have meanings that do not denote self-controlled actions, while unergative verbs generally require animate, volitional subjects (see exceptions in Sections 2.5 and 3.1.2). As for transitive verbs, although there are languages that disallow inanimate subjects of transitives (such as Jacaltec and Japanese; see Section 3.1.1), English and many other languages do allow them (cf. The rock broke the window), and so we do not want children hearing such sentences to interpret the subject as an internal argument. On the other hand, even in English there are transitive verbs that deviate from the Agent–Verb–Patient prototype in that they take two internal arguments, one a theme and one an experiencer. Psychological verbs, such as please or frighten, are such verbs (Belletti and Rizzi, 1988). Without complicating the picture too much by bringing psych-verbs into the discussion, I mention these verbs to suggest the possibility that inanimate subjects may still provide a cue to displaced subjects even in certain transitive sentences. That is, in a sentence like Horror movies frighten John, the inanimate theme subject is underlyingly an internal argument (as is the experiencer object, John). So it is worth asking whether inanimate subjects might indicate derived subjects in the case of monoclausal frames. We will see in this section that there is an effect of subject animacy in cueing unaccusative versus unergative verbs. In the next section, however, we will see that this effect does not extend straightforwardly to the passive. 5.3.3.1 Novel verb studies As in the case of raising and tough-constructions, much of the literature on children’s acquisition of unaccusatives has focused on what children know: do children in fact analyze unaccusative structures as involving a raised internal argument, as opposed to an underlying external argument (i.e. as unergatives)? However, unlike raising and tough-constructions, a bit more attention has been

Children’s acquisition of displacing predicates

229

paid in the Syntactic Bootstrapping literature to the question of how children distinguish unaccusatives from unergative verbs. Some of this focus began with an important study by Letti Naigles (1990; 1996), in which she asked the question of how a child might learn the meaning of a verb that alternates between a transitive and an intransitive sentence frame. There are two distinct ways in which a verb could enter into this alternation. In one way, the verb is underlyingly transitive but permits the object to be optionally omitted (Omitted Object, or OO verbs; as in (23)). The other way, the verb is one of those that alternate between a transitive frame and an unaccusative intransitive frame (Causative Alternation, or CA verbs18 ) as in (24) (Levin, 1993). (23)

Omitted Object a. The girl ate a cookie. b. The girl ate.

(24)

Causative Alternation a. The girl opened the door. b. The door opened.

Naigles’ work did not focus specifically on animacy as a cue to differentiating these alternating verbs. Instead she probed a preference for CA verbs to have a meaning related to physical causation, with OO-type verbs tending to denote actions with non-causative physical contact. But more recent work has highlighted a role for animacy in cueing one verb type vs. the other. This role for animacy is evident in the contrast between (23b) and (24b). In (23b) the intransitive subject is animate, as is typical for both transitive and unergative verbs. In contrast, the subject of (24b) is inanimate, as is typical for transitive objects, and easily permitted for unaccusative verbs (see Section 2.3 above). The study by Gelman and Koenig (2001), discussed above in Section 5.2, found that when move had an unergative type of use (Is NP moving? was a question about an event of spontaneous, internally driven movement), children permitted it with both animate and inanimate subjects, but when it had an unaccusative type of use (Is NP moving? was a question about an entity being moved by an external causer), children had a strong preference for it to have an inanimate subject. Gelman and Koenig’s experiment was not specifically designed to tease apart children’s acquisition of unaccusative verbs, but its results shed light on children’s restrictions on argument structure of unac18 This is also called the causative-inchoative alternation.

230

Animacy and children’s language

cusatives with respect to animacy. Specifically, children strongly preferred the derived subject to be inanimate. While it is slightly surprising that children also allowed inanimate subjects of an unergatively used verb, given unergative verbs’ strong preference for animate subjects, recall that the inanimate entities used in the experiment were capable of self-propelled motion (by batteries or other mechanical means) and thus were, in the relevant sense, “animized.” Two experiments by Ann Bunger and Jeffrey Lidz (2004; 2008) suggested an effect of animacy on children’s encoding of novel predicates as unaccusative or unergative, when heard in the kinds of sentences above in (23–24). In their first study, Bunger and Lidz wanted to know whether children understand the different components of causative events. Since CA verbs denote a change of state in the internal argument, the meaning of a CA verb includes something about the result state of the verb’s action. That is, in the pair in (24) the door undergoes a change of state such that it results in being open after the event is finished. Bunger and Lidz describe this as the “result subevent,” as distinct from the “cause subevent” which involves an actor performing the opening action. The 2004 study presented 2-year-olds with videos of causative events, such as a girl bouncing a ball, accompanied, across conditions, with verbal commentary that included either an intransitive sentence (unaccusative, by virtue of the subject being the patient, in this case ball; e.g. Look! The ball is pimming!), a transitive sentence (Look! The girl is pimming the ball!), or both (plus a control condition in which no verb was used, and the verbal commentary served only to draw the child’s attention to the video: Wow, look at that!). Following the initial exposure to the complete event, children were then shown two separate subevents, each displayed on two different screens. One depicted only the cause subevent, in this case the girl patting the ball as before, but this time the ball remains stationary, or the result subevent, in this case the ball bouncing without the girl touching it. Bunger and Lidz expected children to look longer at the scene that “matched” the verbal prompt. This would amount to children looking longer at the result subevent if they initially heard the unaccusative, or unaccusative plus transitive sentences, but not if they had initially heard only the transitive sentence when they saw the complete event. Although there was a slight overall preference for the result subevent (perhaps due to the novelty of the inanimate object appearing to move on its own), this preference was significant only for the children who had indeed heard the unaccusative, or unaccusative plus transitive sentences during the familiarization phase (i.e. when they saw the whole event). The implication is that children knew that the intransitive sentence was unaccusative and not

Children’s acquisition of displacing predicates

231

unergative because the patient subject of the intransitive sentence was both perceptually the entity affected by the action witnessed and it was always inanimate. But how do we know that children in fact took these intransitive sentences to be unaccusative, and not unergative (with an inanimate subject)? In a follow-up study, Bunger and Lidz (2008) again showed 2-year-olds videos of people acting upon inanimate subjects, such as a boy pressing a bicycle pump which makes a flower spin. But this time the intransitive sentence used during the familiarization phase was unergative, e.g. Look! The boy is blicking. That is, the subject was animate and referred unambiguously to the agent in the scene. Children heard either an intransitive (unergative) or a transitive sentence while viewing the complete event (all events, as before, were causative). But now during the test phase, children saw two different non-causative events. One depicted the same agent performing the same type of action but without an object being affected (for example, the boy presses a bicycle pump but the flower does not spin), and the other depicted the same agent doing a different (equally non-causative) activity (e.g. hitting the flower, flower does not spin or do anything else). The results of this study showed that children preferred the scene where the agent performs the same action as before, even though it has no causative effect this time, but only if they had heard the intransitive (unergative) sentence (The boy is blicking!) during familiarization, and not if they had heard a transitive sentence (Look! The boy is blicking the flower) during that phase. Bunger and Lidz conclude that the children had mapped the novel verb onto the “means” subevent (i.e. what the agent is doing in the event) when they were presented with the verb in an unergative sentence. Taken together, these two studies support the idea that children interpret an intransitive sentence with an inanimate patient subject as unaccusative (and referring to the result subevent), but they interpret an intransitive sentence with an animate agent subject as unergative (and referring to the means subevent). These studies are suggestive of a role for subject animacy in encoding novel intransitive verbs as unaccusative or unergative, but they are not conclusive. Bunger and Lidz did not specifically pit unaccusative versus unergative interpretations against one another while controlling for NP animacy. Moreover, there is apparently contradictory evidence from work by Cynthia Fisher (1996; 2002). In Fisher’s studies children ages 2–5 and adult controls watched causative scenes taking place between two animate entities. For example, one girl pulls another girl in a wagon. Half of the children hear an accompanying transitive sentence (She’s pilking her over there) and half hear an intransitive sentence (She’s pilking over there). Since both entities are animate and

232

Animacy and children’s language

have the same gender, and since they are referred to with pronouns rather than proper names, the reference of the intransitive subject she is ambiguous. Children were then asked to point to “Which one pilked the other one?” or “Which one pilked?” according to the condition. Fisher found that children who had heard the transitive sentence (and question) pointed to the agent significantly more frequently than those who had heard the intransitive sentence (and question). That children in the transitive condition pointed to the agent is unsurprising. But children in the intransitive condition appeared to have an unaccusative bias, pointing to the agent as the one who “did the pilking” only a quarter of the time or less (actually, the 3.5-year-olds did this only 5% of the time). While there is nothing wrong with mapping an animate NP to the subject (theme/patient argument) of an unaccusative verb, this result seems to be at odds with both Bunger and Lidz’s and Gelman and Koenig’s results, according to which children appear to disprefer animate patients as subjects of unaccusative verbs. On the other hand, Scott and Fisher (2009) specifically manipulated animacy in their sentences containing novel transitive and intransitive verbs and found a significant effect of animacy on children’s interpretation of novel verbs. In their study, 2-year-olds first watched a video of two women having a short conversation using a novel verb that alternated between a transitive and an intransitive use. The nature of this alternation followed that illustrated above in (23–24). Namely, for the OO-type verb, the lone argument in the intransitive version was animate, and for the CA-type verb, the lone argument in the intransitive version was inanimate. (25)

Sample Dialogues from Scott and Fisher (2009) Causative-Alternation Omitted Object A: Matt dacked the pillow. A: Matt dacked the pillow. B: Really? He dacked the pillow? B: Really? He dacked the pillow? A: Yeah. The pillow dacked. A: Yeah. He dacked. B: Right. It dacked. B: Right. He dacked.

Following the dialogue, the children then viewed each of two videos, on different screens, one depicting a causative event (a girl pushing a boy to make him bend forward) and the other depicting a non-causative (but still transitive) event (a girl dusting a boy’s shoulder with a feather duster). Upon seeing these new scenes the children were told “The girl is dacking the boy. Find dacking.” On average children looked significantly longer toward the causative action if they had initially heard the Causative-Alternation dialogue (61%) than if they had heard the Omitted Object dialogue (39%). Crucially, the only difference

Children’s acquisition of displacing predicates

233

between the two dialogues was the animacy of the subject in the intransitive sentences. Scott and Fisher determined that the ability of the intransitive sentence to have an inanimate subject signaled an unaccusative structure, which in turn signaled a verb whose transitive use has a causative meaning, in line with the set of verbs that undergo the causative-unaccusative alternation. 5.3.3.2 Knowledge of the unaccusative-unergative distinction in the target language The experiments just discussed looked at one side of the acquisition of unaccusatives: whether children take account of the number of arguments a verb occurs with, and the animacy of those arguments, to draw inferences about the meaning of the verb; more specifically, whether the verb has a causative meaning. But the causative aspect of unaccusative verbs’ lexical meaning is only one part of the picture. For one thing, it applies only to those verbs that actually undergo the Causative Alternation. There are many unaccusative verbs that do not alternate with a transitive use (e.g. fall, arrive, come, etc.). In addition, there are morphosyntactic reflexes of the unaccusative-unergative split in many languages, and some work has explored how much of this distinction children know about, and at what age. One well-known study by Babyonyshev et al. (2001) looked at Russianspeaking children’s production of the Genitive of Negation construction. In this construction, an indefinite object in the scope of negation receives Genitive case, and since it is restricted to objects of transitives, and postverbal themes of passive and unaccusative verbs, the construction is considered unaccusative (Pesetsky, 1982) (but see arguments by Potsdam and Polinsky (2011) against the A-movement analysis of this construction). Babyonyshev and her colleagues elicited these forms from children ages 3 to 6.5 and found that while both younger and older children they tested correctly produced the Genitive case with nonspecific objects of transitive verbs, and correctly avoided the Genitive (they produced Nominative instead) on both specific objects of transitives and on subjects of unergative verbs, the older children (mean age 5;4) produced the Genitive case with arguments of unaccusative verbs more frequently than the younger children did (mean age 4;0). The authors conclude that younger children have difficulty representing the covert A-movement associated with the underlying object in these constructions, and therefore performed poorly in the task. Babyonyshev’s study was taken as evidence that children have difficulty with A-chains (after Borer and Wexler 1987, 1992), and that true unac-

234

Animacy and children’s language

cusative structures with a raised internal argument are therefore difficult, if not impossible, for children to produce or comprehend until the ability to form A-chains has matured. Instead, where children appear to use unaccusative verbs with a raised internal argument, they are actually representing these forms as unergative. On the other hand, other work has been used to argue for an earlier facility with unaccusative constructions. Of note, Snyder et al. (1994) and Hyams and Snyder (in preparation) report that children acquiring French and Italian use the correct form of the auxiliary (be and not have) with reflexive clitic constructions, indicating an adult-like representation of these constructions as unaccusative. In a study of children acquiring Hebrew, Friedmann (2007) showed that children as young as 2 years correctly distinguish unaccusatives from unergatives in terms of word order: unaccusatives allow both SV and VS orders, while unergatives allow only SV order. In Friedmann’s study, her children allowed both orders with unaccusatives but limited unergatives to SV order. These studies provide some mixed evidence on children’s knowledge of unaccusatives in terms of the morphosyntactic properties of these verbs: children appear to have difficulty with them when the raising of the theme argument occurs covertly (Babyonyshev et al., 2001), but not when the raising occurs overtly (Snyder et al., 1994; Hyams and Snyder, in preparation; Friedmann, 2007; see also Lorusso et al., 2005). None of these studies examined whether children allow unaccusatives to take both animate and inanimate subjects, and whether they limit unergative verbs to taking animate subjects. In collaboration with Jeannette Schaeffer I looked at this question in the spontaneous speech of children acquiring English (Becker and Schaeffer, 2013). We found that children indeed make a distinction between unaccusatives and unergatives with respect to subject animacy. Using data from Adam, Eve (Brown, 1973), and Nina (Suppes, 1974) (total child utterances in these corpora: 117,086), we found that these children produced over 93% of their unergative verbs with an animate subject, but subject animacy was split evenly among their unaccusatives: only 51.5% of their unaccusative verbs had animate subjects; a full 48.5% had inanimate subjects (see Table 5.8; occurrences of verbs without overt subjects were not counted). Unaccusatives were also vastly more frequent in the children’s production than unergatives, mainly due to a few highly frequent verbs: go and come, though Adam also used fall with very high frequency. The fact that the children permit unaccusative verbs to occur with inanimate subjects at such a high rate (significantly higher than with unergatives) suggests that the children correctly distinguish the classes of unaccusative and

Children’s acquisition of displacing predicates

235

Table 5.8: Children’s use of animate subjects with unaccusative

vs. unergative verbs in spontaneous speech

% Animate Subject Total N

Unergative

Unaccusative

93.1 413

51.5 2822

unergative verbs in terms of their argument structure representations, and the extreme frequency of their unaccusatives (we tallied 2,822 uses, compared to 413 unergatives) suggests little difficulty producing unaccusatives correctly. Moreover, we found this asymmetry to be present both in earlier as well as later stages of development. The data we examined ranged from age 1;6 (Eve’s earliest file) to 5;2 (Adam’s latest file). When we compared the rate of animate subject use with each verb type in the younger files (below age 3) with the rate in the older files (age 3 and above) we found no difference for the unergative verbs (95% animate subjects before age 3; 92% animate subjects after age 3). For the unaccusative verbs there was a significant effect of age, but this turned out to be due entirely to Eve’s data, who used vastly more animate subjects with both verb types than Nina or Adam, and whose files spanned only the “early” period. That is, when we compared Adam’s rate of using animate subjects with unaccusatives before age 3 with his use after age 3, he used virtually the same rate of animate subjects with unaccusatives in both time periods (46% before age 3 and 41% after age 3). The same was true of Nina’s data (76% before age 3 and 75% after age 3). Eve, on the other hand, produced significantly more animate subjects overall compared to the other children, almost 86% with unaccusatives. While this could be taken as evidence that Eve incorrectly represented (some) unaccusatives as unergatives, it is worth noting that (a) Eve still produced significantly fewer animate subjects with unaccusatives (85.7%) than with unergatives (98.6%), and (b) the use of animate subjects with unaccusatives is not inappropriate, it just makes it harder to claim conclusively that she represented unaccusatives differently from unergatives. 5.3.4 Animacy and the acquisition of the passive Compared to acquisition studies of raising-to-subject, tough-constructions, and unaccusatives, the literature on children’s acquisition of the passive is vast. Here I will restrict my attention to studies that specifically manipulated or took

236

Animacy and children’s language

into account NP animacy. Most recent work on children’s acquisition of the passive has limited its experimental inquiry to so-called reversible passives, sentences with two animate NPs, such that either NP could plausibly correspond to the agent or the patient (e.g. Crain et al., 1987; Fox and Grodzinsky, 1998; Hirsch and Wexler, 2007). I think the reasons for this limitation have to do with concerns I spelled out at the beginning of Section 5.3, namely the worry that if children can correctly interpret passives when the patient (subject) is inanimate and the agent (oblique) is animate, they might be doing so “only” by means of semantics and not representing the syntax of the construction (but see my objections to this view, listed there). However, older work on the passive did not exclude this useful comparison. One of the earliest experiments on children’s knowledge of the passive was conducted by Dan Slobin (Slobin (1966); see also Bever (1970)). Slobin presented children with a series of pictures accompanied by sentences of one of four syntactic forms, and children had to indicate whether the sentence was “right” in relation to the picture, or “wrong.” The four syntactic forms were affirmative active (e.g. The dog is chasing the cat; he called them “kernel” after Harris (1951); Chomsky (1957)), affirmative passive (The cat is being chased by the dog), negative active (The dog is not chasing the cat), and negative passive (The cat is not being chased by the dog). Crucially, half of the stimuli were “reversible” (as in the examples just given), meaning that both of the NPs in the sentence were animate and therefore candidates for either logical role in the sentence, and the other half of the stimuli were “non-reversible,” meaning that the patient was inanimate (for the true affirmative sentences, e.g. The girl is watering the flowers). The children in Slobin’s study ranged in age from 6 (kindergarteners) to 12 years, and even the youngest children had good accuracy on the passive sentences (over 80% correct on average). But Slobin measured not just accuracy of responses, but also how long it took subjects to respond: their Reaction Time (RT). Like the adult controls, children were uniformly faster in responding to the non-reversible passives than the reversible ones. In fact, the effect of reversibility was mildly facilitative for all types of sentences and for all age groups, with minor exceptions.19 However, Slobin concluded that “nonreversibility seems to be especially facilitative in regard to passive and passive negative sentences, while not especially speeding comprehension of kernel and 19 For example, the youngest participants were equally fast in responding to reversible and non-

reversible “false kernel” sentences, e.g. The cat is chasing the dog/The flowers are watering the girl, and they were slower on non-reversible true negative sentences (The flowers are not watering the girl) than on reversible true negative sentences (The cat is not chasing the dog).

Children’s acquisition of displacing predicates

237

negative sentences” (Slobin, 1966, p. 226). While the younger children were quite a bit slower in responding overall, their ratio between RTs to reversible and non-reversible passives was on the order of the adults’ ratio. In general, reversible passives were slower than non-reversible passives by a factor of 1.36 for 6-year-olds (ratio range: 1.32–1.41, across affirmative, negative, true, and false passives). For adults, reversible passives were slower than non-reversible passives by a factor of 1.51. Slobin’s data show that non-reversibility speeds comprehension of passives (i.e. inanimate patient subject and animate demoted agent), and this effect holds for speakers at all ages studied, kindergarteners through adults. Another study that found a pattern reminiscent of Slobin’s non-reversibility effect was a novel verb study by Pinker, Lebeaux, and Frost (1987). Pinker and colleagues taught children ages 5 to 8 transitive verbs that either took an agent subject and theme object (in the active form), or a location subject and theme object (e.g. one of these verbs meant “to be hanging from”). These verbs were considered “canonical” since their active voice subjects were higher on the Thematic Hierarchy than their active voice objects (see Section 3.2). But they also taught the children “anti-canonical” verbs; these were verbs with the thematic roles reversed: active voice subjects were themes, and direct objects were either agents or locations (recall Marantz’s study; see Section 5.2.1). During the training phase verbs were modeled either in the active or passive voice, and then children were asked to both produce on their own, and act out (to indicate comprehension), the same verbs in both voices. While children rather easily passivized the agent–theme verbs, they were somewhat less willing to passivize the (more spatially connoting) location– theme verbs, and far less willing to passivize the anti-canonical verbs of either type, compared to their canonical counterparts. But we have to be careful in drawing connections between these results and any effect of animacy. Although Pinker et al.’s stimuli contained a mixture of animate and inanimate NPs (e.g. The monkey is floosing the giraffe; The fork is being floosed by the pencil), and although the researchers manipulated the thematic roles assigned to subject and object position, animacy was not aligned with thematic role as far as I can tell. That is, it appears from the examples provided in the text that both NPs in a given training sentence were either animate or inanimate. Thus, agents were not uniformly animate and themes were not uniformly inanimate. So the general preference shown by Pinker’s children for creating passives in which the subject was a theme and the byphrase NP an agent does not necessarily imply that subjects of passives should be inanimate and by-phrase obliques animate.

238

Animacy and children’s language

Further complications emerge when we look more closely at Slobin’s results and at additional experimental data. Hayhurst (1967) pointed out that Slobin’s youngest subjects made as many errors on reversible as non-reversible passives, while the older children made fewer errors on non-reversible passives. Therefore, she suggested that non-reversibility is not entirely helpful for the youngest children. In her own study (not a true-false task, like Slobin’s, but one in which children had to produce passives modeled after a prompt), Hayhurst’s youngest subjects (average age 5 years) produced no correct non-reversible passives and did produce some correct reversible passives. She argued that only her oldest subjects (mean age 9;6) were aided by non-reversibility. Similarly, Horgan (1978) denied that children uniformly begin by producing non-reversible passives and then later graduate to using reversible passives. Horgan elicited picture descriptions from children ages 2 to 13 (and from adult controls) and analyzed the descriptions that had a passive structure. She found that some of her youngest subjects used reversible, but not nonreversible full passives, while others used only non-reversible full passives. In fact, in her study, no child younger than 11 years produced both reversible and non-reversible passives. Averaging the responses within each age group (2–4, 5–7, and 9–13), she found that while all age groups produced both reversible and non-reversible passives in about equal proportions, the younger children (ages 2–7) only produced non-reversible passives with an instrument by-phrase, that is, a by-phrase with an inanimate NP (e.g. The window was broken by/with the ball),20 and never with an agent by-phrase (The window was broken by the boy). The older children, ages 9–13, did produce nonreversible passives with both types of by-phrases. Horgan suggested that children interpret the passive specifically as a means of expressing non-agentive causation.21 In light of these results, it does not seem to be the case that the first passives to be acquired are those with an inanimate patient subject and an animate agent by-phrase. Horgan does not give a breakdown of the children’s production of animate versus inanimate patients in full passives, but she notes that in the youngest group of children (ages 2–4) truncated passives almost exclusively contained inanimate subjects (patients) while their full passives generally contained animate subjects. 20 These sentences are considered non-reversible, but they are different from Slobin’s non-

reversible sentences in that here both NPs are inanimate. 21 But see Pinker et al. (1987) for counterarguments to Horgan’s analysis.

Children’s acquisition of displacing predicates

239

Lempert (1989) also elicited passives from children in response to depicted events. These children, ages 2;6 to 5 years, were pre-tested on their knowledge of the passive and were only included in the study if they scored less than 67% correct on that pre-test. Children were trained using an imitation task to produce passive sentences with either animate or inanimate patients (all agents were animate).22 Lempert measured the proportion of the time children correctly imitated the passive sentences. Additionally, in a post-training generalization phase (which took place 1–3 days after training) children were asked to describe pictures, some requiring verbs already encountered during training and some requiring verbs not encountered during training. In both the training and generalization phases, children who were trained on pictures/sentences with animate patients produced significantly more passives than children who were trained with inanimate patients. That is, children trained on sentences like The bunny is being touched by the boy produced more passives than did children who were trained on sentences like The plant is being touched by the boy. Not only did the animate patient-trained children produce more passives overall, they produced the most passives with animate patients in the generalization phase (55% of their responses, as compared to between 24% and 42% with inanimate patients at this stage of the study). Children trained on inanimate patients, instead, produced equal proportions of passives when tested with animate or inanimate patients (26% across the board). This result appears to imply that inanimate patient subjects are not facilitative of learning the passive. Lempert explained these results by proposing that children’s category of subject comprises a number of semantic features including the feature ‘animate.’ This general bias is consistent with what we have seen so far for nondisplaced subjects, but it appears to run counter to my claim about the association between animacy and displaced subjects. However, recall from Section 2.4.2 that there is an important difference between the passive and other displacing constructions. Namely, in the opaque contexts like NP verbs to VP (raising or control verb), NP is Adjective to VP (tough or control adjective), or NP verbs (unaccusative or unergative) the meaning of the predicate in question is fundamentally different, depending on the underlying structure of the string. Easy is not simply a different version of the lexical item eager, but (be) eaten is a different version of eat. Active and passive voice denote the same set of events, 22 See also De Villiers (1980). However, De Villiers did not train the children in her study on

passives with inanimate patients; she only trained them on passives with two animate NPs, some with an active kind of event (hit) and others with a less “active” kind of event (wear).

240

Animacy and children’s language

that is, they are in most cases truth-functionally equivalent, but they differ in the perspective taken. But pairs like John is easy to please and John is eager to please are not in any way truth-functionally equivalent. In thinking about how inanimate subjects could help children learn opaque predicates like seem and easy, I do not think we want to extend such a strategy to passives.23 What motivates the use of a raising or TC structure as opposed to a control structure is the choice of the main predicate. But what motivates the use of passive over active voice is the need to foreground or promote the patient and to background or demote the agent or instrument. These pressures come from discourse features such as what is given (vs. new) in the discourse, what is known to both participants, and so forth (Haviland and Clark, 1974; Tomlin, 1983). At this point recall that we have seen evidence in adult languages for a very different pattern in passives than in other displacing structures. In particular, we saw in Section 3.4 that Jacaltec requires subjects of passives to be animate, although it allows (certain) other derived subjects to be inanimate (and it generally requires subjects of transitives to be animate). According to Ransom (1977), Japanese which, like Jacaltec, requires transitive subjects to be animate, likewise requires passive subjects to be animate. Moreover, I mentioned in Chapter 3 (Section 3.1.1) that Ransom (1977) argued for a preference in English passives for the subject (i.e. the promoted argument) to be more animate than the demoted NP. Although inanimate passive subjects are certainly grammatical in English, she notes that the relative position of the subject and oblique on the animacy hierarchy (human > non-human animate > inanimate) accounts for the subtle differences in acceptability of sentences like these (the judgments are Ransom’s, but I agree with them; pp. 422–423):24 (26)

a. ?? The fence was jumped by the man. b. ? The fence was jumped by the horse.

23 On the other hand, Crawford (2012) found that English-speaking children ages 4 to 6 years

were quite good in judging passives with a purpose clause to be grammatical (Carrots are being grown to make soup), where all of her passive sentences had inanimate subjects. Since Crawford did not compare animate vs. inanimate subjects of passives it is impossible to be sure about any effect of animacy, but perhaps this feature facilitated their performance. 24 This is actually at odds with Jackendoff’s (1972) Thematic Hierarchy Condition (see Section 3.5). I’m not sure what to make of this contradiction, but it strikes me as indicative of how the passive construction entails a conflict between two broader tendencies in language: for subjects of all kinds to be animate, and for the argument reversal that is the hallmark of the passive to flag patient/theme promotion and agent/experiencer promotion.

Children’s acquisition of displacing predicates

241

c. ? The cat was attacked by the man. d. The cat was attacked by the dog. e. The fence was struck by lightning.

In contrast, when the promoted patient subject denotes a human, the oblique argument can be anywhere on the animacy hierarchy (subject to lexical restrictions of the verb) and well-formedness is not compromised. (27)

a. The linguist was attacked by the informant. b. The linguist was attacked by the dog. c. The linguist was struck by lightning.

Consistent with these sentence judgments, recall that the sentence construction task by Ferreira (1994) discussed in Section 4.3 showed that adults produced significantly more passives when provided with a theme–experiencer verb (e.g. worry or challenge), i.e. verbs which in the active voice take a theme subject and experiencer object, than they did when provided with an agent– theme or experiencer–theme verb. In other words, adults exhibited a drive to produce sentences with animate subjects, and so passives were more often produced that resulted in an animate subject. Another way to think about the split between passives on the one hand, and subject-raising, tough-constructions, and unaccusatives on the other, is in terms of Hyams et al.’s (2006) Canonical Alignment Hypothesis (CAH). The CAH is defined as follows (Hyams et al., 2006, p. 1064): (28)

Canonical Alignment Hypothesis Children cannot form A-chains that derive the misalignment of thematic and grammatical hierarchies; i.e., an external argument (agent, experiencer), if there is one, maps onto the subject (SpecIP or TP)

The crucial point is the phrase “an external argument, if there is one,” because an external argument is projected by verbs that can be passivized, but not by raising verbs, tough-adjectives, or unaccusatives. In other words, if the lexical entry for a given verb projects an external argument (which, following the discussion in Chapter 3, is likely to be higher on the Animacy Hierarchy than internal arguments), children will be biased to represent that argument in subject position. An inanimate subject can still provide a cue to learners that the subject is derived, even in the case of passives, but there is a further disruption in the projection of the argument structure with passives that one does not encounter with the other displacing predicates discussed here. This seems to me to be consistent with the observation in the

242

Animacy and children’s language

experimental data described above, that older children and adults appear to be facilitated in their construction and interpretation of passives by inanimate patients, while younger children experience more facilitation by animate patients.25 5.4

Summary

The first part of this chapter established that soon after birth babies are drawn to certain properties of animate entities, in particular human faces, and that within the first year of life they discriminate animate from inanimate entities on the basis of perceptual properties (for example, movement behaviors) and begin to form conceptual categories of these different entities in the world around them. By midway through the second year of life toddlers attribute intentions to people but not inanimate objects. Around this time, they are beginning to put words together to form sentences. In the second part of this chapter we saw that animacy plays an important role in children’s representations of sentence structure. We see this in children’s own productions (their own earliest sentences tend to have animate subjects and inanimate objects), and in their comprehension of sentences (they interpret animate NPs as agents and inanimate NPs as patients). This bias is robust, as children have a much harder time learning novel verbs that violate the canonical argument structure pattern (Marantz’s and Pinker et al.’s anti-canonical verbs). Two well-known approaches to accounting for children’s early acquisition of sentence structure, Pinker’s Semantic Bootstrapping hypothesis and Gleitman’s Syntactic Bootstrapping hypothesis, explain a great deal about how children can exploit cues from NP animacy in the construction of their earliest sentences, all of which are assumed to be “basic” or canonical. However, neither Semantic nor Syntactic Bootstrapping extends straightforwardly to opaque constructions such as raising-to-subject or tough-constructions, in which the matrix subject is not an argument of its adjacent predicate. We saw that in just these cases, encountering an inanimate subject can provide the necessary cue 25 See also a recent study by Messenger et al. (2012), which found that 6-year-olds produced more

passives following passive primes than they did after active primes, but that they also produced many “passive reversals,” where the sentence has a passive structure but the arguments are reversed. Messenger and her colleagues argue that children by this age can represent passive syntax correctly but have difficulty with the non-canonical mapping of thematic roles in the passive construction.

Summary

243

that the subject is derived. In the experiments reported above, it was demonstrated that children can use this cue both in interpreting known predicates of their language (as in the experiments on subject raising, and in children’s adultlike responses to hard in the tough-adjective study) and in the categorization of novel predicates. There is strongly suggestive evidence that learners can use inanimate subjects as a cue to unaccusative as opposed to unergative verbs (Bunger and Lidz, 2004, 2008), as well as evidence for a stark asymmetry in children’s production of inanimate subjects with unaccusative but not unergative verbs. However, experiments on novel verb categorization of intransitives specifically manipulating subject animacy have not yet been carried out. So it seems that the usefulness of the inanimacy cue for derived subjects may extend beyond biclausal structures to certain monoclausal structures as well. However, there are limits on the utility of this cue. In particular, we saw that it does not work quite the same way in children’s learning of the passive construction. While older children and adults appear to receive some benefit from inanimate patient subjects of passives (they are faster and/or more accurate in interpreting them), younger children appear to receive the most benefit from animate patient subjects of passives. This may be linked to Gelman and Koenig’s (2001) finding that younger children (3-year-olds) resisted allowing animate patient subjects of move. That is, perhaps it is the violation of canonical argument alignment in this direction (animate patients rather than inanimate subjects) that is most useful in acquiring the passive. In Section 5.3.4 I attributed the asymmetry between raising verbs, toughadjectives, and unaccusatives on the one hand, and passives on the other, to the fact that the different voices of a verb in a passive–active alternation do not correspond to completely different lexical meanings, in contrast to the raising– control, tough–control, or unaccusative–unergative distinctions. Instead, the active–passive distinction turns on discourse factors involving givenness and the pragmatic need to promote or demote certain arguments. It seems that inanimate subjects are a useful cue for distinguishing different predicates that differ in their lexical meanings, but not necessarily for distinguishing different voices of the same basic predicate. Up to this point we have talked about which cues in language can lead children to discriminate displacing predicates and the constructions they occur in, from non-displacing predicates and their syntactic structures. This is essentially a categorization process. But we have not considered what the actual

244

Animacy and children’s language

input to children looks like in this respect, and whether the cues I have argued to be useful in principle for discovering displacing predicates, and used by children in experimental settings, are available to learners in everyday exposure to language. In the next chapter we will consider the facts about the distribution of displacing predicates with respect to subject animacy in child-directed speech, and we will look at some computational models of learning to distinguish these predicates on the basis of this input.

6 Modeling the acquisition of displacing predicates

The central thesis of this book is that an inanimate subject signals to language learners that it is likely to be derived, and that the main predicate it is adjacent to is what I have called a displacing predicate (one that does not underlyingly select an external argument, or “semantic subject”). There are four ingredients needed to support this thesis, and in the preceding chapters I have provided three of these ingredients: (1) languages universally prefer subjects to be animate and allow inanimate subjects if they are derived or displaced (Chapter 3); (2) adults are sensitive to NP animacy in their parsing of sentences and preferentially treat inanimate subjects as derived (Chapter 4); (3) the animate–inanimate distinction is available to children’s conceptual apparatus and children, like adults, treat inanimate subjects as derived in experimental settings (Chapter 5). The fourth ingredient is to extend the experimental demonstrations of Chapter 5 to language acquisition “in the wild.” That is, we must show that the subject animacy cues that appear to guide children’s guesses about sentence structure (and predicate categorization) in experimental settings are plausibly used by children in the natural course of language acquisition. We will do this in two ways. The first is to demonstrate that these cues are available to children in the language they hear around them. To do this I will present data on patterns of subject animacy in input to children with the different kinds of predicates we have focused on here, showing that subjects are inanimate asymmetrically more often with displacing predicates than with non-displacing predicates. One could think of this as a computational-level analysis: it shows that, in principle, the cues are present in the input data for children to use. The second way I’ll extend my hypothesis to naturalistic language acquisition is to apply computational modeling techniques in order to model the kind of probabilistic inference procedure I believe children must employ in order to learn these predicates. So far in this book I have been careful to state the learning cues I have proposed in terms of likelihood: a sentence with an inanimate 245

246

Modeling the acquisition of displacing predicates

subject is likely to indicate a non-canonical underlying structure with a displaced subject; an animate subject is likely to be underived (see the Derived Subject Condition in Section 3.5). The reason it was necessary to couch the cue information in such terms is that the cue is not deterministic in all sentence frames: not all derived subjects are inanimate, and not all inanimate subjects are derived. A sentence such as The hammer broke the window should not trigger a structural representation of the sentence with a derived subject. Even in the sentence frames under consideration there are cases where this cue is not uniquely informative. Most blatantly, displacing predicates surely occur with animate subjects – they simply do not require animate subjects the way non-displacing predicates generally do. Relatedly, recall that certain verbs are ambiguous between being raising or control verbs. Although these verbs generally function as raising verbs when the subject is inanimate and as control verbs when the subject is animate, thus adhering to the picture presented here, these verbs are always free to have a raising interpretation when the subject is animate (John started to feel sick can mean – in fact, most naturally means – that John came to start feeling sick without it being under his own volition). More problematically, a small number of control verbs allow inanimate subjects (serve, help), and a subgroup of unergative verbs allow or require inanimate subjects (see Section 3.1.2). As we will see below, raising verbs are actually used with animate subjects a very high proportion of the time in speech to children, and yet children should not consider them to be non-displacing predicates. Unaccusative verbs are used more frequently with inanimate subjects than raising verbs are, but still less than 50% of the time. Of the displacing predicates only tough-adjectives are used almost exclusively with inanimate subjects. In turn, these distributions imply the need for a probabilistic approach to learning. So one purpose of this chapter is to specify some ways we could model children’s learning of displacing predicates using subject inanimacy as a probabilistic cue. One could think of this as an algorithmiclevel analysis: it shows there is a way for children to extract the cue of subject animacy to use in distinguishing among predicate types.1 To start out, let us establish that a feature like animacy could plausibly be used for learning to categorize predicates. A study by Merlo and Stevenson (2001) points to a very helpful role for subject animacy in an automated 1 I am not making the direct claim that children use the exact procedure implemented in

Section 6.3, namely Markov Chain Monte Carlo estimation. However, there is some evidence that natural neural networks represent probabilities and perform Bayesian inference (Ma et al., 2006), and much of the work by Perfors and colleagues (see Section 6.2.2) strongly suggests that children reason via Bayesian inference.

Modeling the acquisition of displacing predicates

247

machine learning experiment. The target of learning in their study was three verb classes that overlap to a large degree in their surface syntactic distributions: unergative (jump), unaccusative (break) and “object-drop” transitive verbs (eat). Each of these three classes of verbs allows both an intransitive (NP V) and a transitive (NP V NP) syntactic frame. The unaccusative/transitive alternation and transitives that allow an omitted object were discussed above in Section 5.3.3; unergative verbs are generally described as being only intransitive, but some of them allow a transitive use (with causative meaning) in some languages (e.g. The trainer jumped the horse over the fence = the trainer caused the horse to jump over the fence). (Merlo and Stevenson note that transitive uses of unergative verbs are rare in their cross-linguistic distribution, and, by my judgment, their use in English is somewhat marked.) Merlo and Stevenson identified five features that can be probabilistically associated with the different argument structures that these verbs participate in: animacy, transitivity, causativity (whether transitive uses of the verb have a causative meaning, e.g. He broke the vase but not He played soccer), and two morphological features having to do with participial verbal forms (these will not be relevant for our discussion). For example, as just mentioned, unergative verbs are quite marked in a transitive structure: they are rare crosslinguistically, and probably marked even within languages that allow them. Thus, the feature of transitivity would have a low probability for unergative verbs but a higher one for both unaccusative and transitive verbs. With respect to animacy, Merlo and Stevenson reason that since unergative and transitive verbs select an agent subject in both their transitive and intransitive sentences (and agents are animate), and since unaccusatives select an agent (= animate) subject only when they occur in transitive structures (but not when they are intransitive), then as long as unaccusatives are not extremely uncommon in intransitive sentences they will, on balance, occur less frequently with animate subjects. This assumes, of course, that their intransitive theme subjects will not denote animate entities as often as agents denote animate entities. Although this is not logically necessary (nor necessary by linguistic constraints) it appears to be true in the data considered by Merlo and Stevenson, and it is true in the child-directed speech we’ll look at in Section 6.3.3. (Recall that it was true of children’s own productions of intransitive unaccusative verbs – see Section 5.3.3.) Thus, the feature of animacy will have a low probability with unaccusative verbs but a higher one with unergative and transitive verbs. Twenty verbs in each class were searched in the Wall Street Journal corpus, and the frequency with which each of the verbs occurred with each of the five features was calculated using an automated procedure and checked manually.

248

Modeling the acquisition of displacing predicates

Then, a vector template for each verb consisting of the frequency of each feature for that verb, plus the label of that verb’s class, was used as the input to a machine learning algorithm (C5.0; Quinlan (1992)). This algorithm uses decision trees to generate rules that determine category membership for novel items. For example, to determine which category a new verb belongs to, one could ask whether it has a high frequency for the animacy feature. If yes, it is probably either in the unergative class or the object-drop transitive class, but not the unaccusative class. One can then ask if it also has a high frequency for the causativity feature. If yes, it is probably in the unergative class but not the transitive class, and so forth.2 The algorithm was trained using two supervised methods, and tested on a portion of the dataset not seen during training. The two methods were tenfold cross-validation, where the algorithm is trained on a randomly chosen 90% of the data and tested on the remaining 10%, repeated fifty times, and a “single hold-out training and testing” method, where the algorithm is trained on all but one of the verbs and tested on the one held out, repeated to hold out different verbs. The algorithm’s task was to classify these as yet unseen verbs as unergative, unaccusative, or (object-drop) transitive. When all five features were used together, the model correctly categorized verbs almost 70% of the time, significantly better than chance (33.9%) and nearly as good as expert human categorizations (86.5%), where “correct” categorizations are established by Levin’s (1993) determination of verb class membership. By removing individual features from the vector inputs and comparing results to the full feature set, Merlo and Stevenson were able to compare the relative usefulness of the individual features. They found that while the animacy feature in isolation was not terribly useful (it generated performance only slightly above the chance rate of 33.9% correct categorization),3 when used in concert with the other features it was highly informative. More specifically, the animacy feature greatly improved the ability of the model to distinguish unaccusatives (which freely allow inanimate subjects) from either unergative or object-drop transitive verbs, relative to chance, but not between unergatives 2 Decision trees can operate on either discrete or continuous feature values. Here we have contin-

uous values (frequencies), so we might imagine that the answer to the question at each node in the decision tree is “yes” if the value is .5 or greater but “no” if the value is less than .5. Merlo and Stevenson do not state what the criterion was in their model. 3 It might be worth observing that the automated coding of animacy actually underestimated the distribution of animate subjects with unergative and transitive verbs, as determined by comparison to manual coding of a portion of the data. Thus, it is possible that the relatively weak performance of the animacy feature when used alone could be attributed in part to artificially low frequency estimates of animate subjects.

Modeling the acquisition of displacing predicates

249

and object-drop transitives. The degree of usefulness of the animacy cue itself can be seen in the fact that when this feature was omitted (and only the other four features were used), the success of the model was significantly impaired. Merlo and Stevenson’s conclusion is that, when used in conjunction with other features such as transitivity and causativity, the semantic feature of animacy is demonstrably useful in categorizing certain classes of verbs using an automated categorization procedure. Scott and Fisher (2009) performed a similar analysis on child-directed speech in CHILDES, though their study focused only on causative– unaccusative alternating verbs (The vase broke; John broke the vase) and object-drop transitive verbs (John ate; John ate a sandwich). Instead of using a supervised learning algorithm, Scott and Fisher used an unsupervised learning algorithm known as k-means clustering (MacQueen, 1967). Since children are not given explicit feedback about their classification of predicates unsupervised learners like this are seen as somewhat more realistic than supervised learners. The k-means clustering algorithm works by taking the values of a set of variables, such as transitivity and animacy, for a set of objects, in this case verbs. The values for these variables are determined by frequency, so a verb that has many occurrences of animate subjects will have a high value for the animacy variable, just as in the Merlo and Stevenson study. And like Merlo and Stevenson, Scott and Fisher determined subject animacy by means of an automated procedure that counted pronominal subjects as animate. The algorithm then groups the verbs into a predetermined number of clusters in such a way that it minimizes the within-cluster distances between the verbs, where distance is measured according to similarity of the variable values. For example, two verbs that both have similarly high values for the animacy variable (or both have similarly low values) will be clustered closer together than verbs that have very different values for that variable. The verbs are regrouped iteratively, and when the within-cluster distances cannot be further reduced the algorithm stops. Scott and Fisher evaluated the results of the clustering algorithm according to how accurately the verbs were categorized (i.e. matched against linguists’ determination of verbs’ categories), as well as by a second measure that rates the overall quality of clustering, called the Adjusted Rand Index. They compared each of these accuracy measures to a clustering of a random selection of values for each of the variables. Despite the slight differences between the algorithms used by Merlo and Stevenson, and Scott and Fisher, similar results were obtained: Scott and Fisher found that including the variable of subject animacy produced highly accurate

250

Modeling the acquisition of displacing predicates

verb categorizations, but only with respect to intransitive uses of verbs. That is, while transitive sentences generally had animate subjects (99%) for both types of verbs (i.e. causative verbs like John broke the vase, and transitive verbs used with an object, like John ate a sandwich), unaccusative verbs (The vase broke) occurred with animate subjects only 44% of the time; transitive verbs used without an object (John ate) occurred with animate subjects 97% of the time. Using the distribution of subject animacy in objectless sentences led to a very robust distinction of the two verb classes. 6.1

Displacing predicates in the input to children

The Scott and Fisher study reveals an asymmetry in the rate of animate subjects among different types of intransitive verbs in parents’ speech to children. As in adult-directed language (cf. Merlo and Stevenson), unaccusative verbs occur with animate subjects at a much lower rate than transitive verbs used with or without an object. Scott and Fisher also demonstrated how an unsupervised learning algorithm can use this asymmetry in order to correctly categorize different types of objectless verbs. As discussed in Chapter 2, unaccusative verbs are a type of displacing predicate, and so these results support the notion that an asymmetrically high rate of inanimate subjects could inform learners that the predicates with this distribution are displacing predicates. In this section I want to look at whether this distribution is true of displacing predicates generally, not just unaccusative verbs. It turns out that displacing predicates do generally occur with asymmetrically more inanimate subjects than non-displacing predicates do, though the steepness of the asymmetry varies greatly across the pairs of constructions. The fact that we find some variation in this respect suggests that inanimate subjects may be a stronger cue for some (types of) displacing predicates than for others. But the fact that asymmetries are statistically detectable means that a learning procedure sensitive to the animacy feature will be able to group predicates into subclasses, especially if we relax the assumption that there are two strict subclasses (see discussion of ambiguous verbs in Section 2.5 and variable behavior of raising and control verbs in 3.3.1.1).4 In the following tables, I give the distributions of particular displacing and non-displacing predicates in child-directed speech, starting with raising and 4 Some caution is needed, however: we should not conclude that raising verbs are somehow

“less displacing” than, say, tough-adjectives because raising verbs occur with animate subjects proportionally more often than tough-adjectives do.

Displacing predicates in the input to children

251

Table 6.1: Mothers’ use of animate/inanimate subjects with raising

and control verbs Raising verbs

Animate subj.

Inanimate subj.

seem used (to) going (to)

4 45 1197

6 5 58

total

1246

69

Control Verbs want like try love hate

405 210 86 10 1

2 0 0 0 0

total

712

2

% Inanimate subj.

5.2%

0.3%

control verbs. In all of these tables, individual verbs were searched and their subjects were coded for animacy by hand. The data in Table 6.1 come from all files from Adam, Eve, and Sarah in the Brown (1973) corpus in CHILDES (MacWhinney, 2000), and the total maternal utterances is 60,250. Here the rate of use of inanimate subjects with both raising and control verbs appears to be similarly low, 5.2% and 0.3%, respectively. However, three things point to inanimate subjects still being a plausible means of distinguishing the two verb classes. One is that the distribution in this table is actually statistically significant (χ 2 (1) = 33.8, p ≤ 0.001). Moreover, if we think in terms of the proportion of sentences with inanimate subjects that contain raising vs. control verbs, the overwhelming majority are raising verbs (69/71). To anticipate the discussion in the next section somewhat, this is good for a Bayesian learner since the probability of a raising verb given an inanimate subject (P (raising|inanimate)) will be quite high. The third fact is that expletive subjects are even rarer in speech to children than inanimate referential subjects are. For comparison, raising verbs occurred sixty-nine times with inanimate subjects in the dataset above, but only eleven times with expletive subjects (not shown in the table). Thus, although the rates of both inanimate and expletive subjects are dwarfed by the large number of animate subjects, inanimate referential subjects occur much more frequently than expletives (Hirsch and Wexler (2007) likewise report a very low rate of expletive subjects occurring with raising verbs). Therefore it is quite likely that children will encounter a raising verb with an inanimate referential subject before they

252

Modeling the acquisition of displacing predicates

Table 6.2: Adults’ use of animate and inanimate subjects with

tough/control adjectives Tough-adj.

Animate subj.

Inanimate subj.

hard difficult easy tough

2 0 1 0

93 3 17 0

total

3

113

Control Adj. happy afraid anxious willing glad eager

8 7 2 1 14 0

0 0 0 0 0 0

total

32

0

% Inanimate subj.

97.4%

0%

encounter it with an expletive subject, so it behooves the learner to use the cue offered by the inanimate subject as a means of categorization. Turning now to tough and control adjectives, the distribution with respect to subject animacy is given in Table 6.2. In contrast to raising and control verbs, (some of) which have reasonably high token frequencies, tough- and control adjectives are vanishingly infrequent in child-directed speech. Unlike the counts in Table 6.1, which come from only the Brown (1973) corpus, the data in Table 6.2 come from a much larger set of corpora, necessitated by the extremely low token frequencies of both types of adjectives. The corpora searched here include Bloom (1970), Brown (1973), Clark (1987), Kuczaj (1977), MacWhinney (2000), Sachs (1983), Suppes (1974), and Valian (1991), and the total number of adult utterances out of which these examples were found is 248,555.5 The asymmetry between the two types of adjectives in their occurrence with animate vs. inanimate subjects is quite stark. Although tough-adjectives are perfectly grammatical with animate subjects, they occurred with such subjects only three times in nearly a quarter of a million utterances. Those three examples of tough-constructions with animate subjects are given in (1). 5 Mostly these are maternal utterances, but in some corpora in which other adults contributed

significantly to the conversations, those adult utterances were included as well. This includes the father’s speech in Kuczaj and MacWhinney, and investigators in Bloom and Clark.

Displacing predicates in the input to children (1)

253

a. They’re [=chimpanzees] still hard to train. . . (Kuczaj file 129) b. He’s [=deflated blow-up horse] awful hard to ride now. (Valian file 05a) c. Well I’m a girl mommy and I’m easy to hold. (MacWhinney file 50b2)

It is interesting to observe that one of these utterances (1b) involves the use of an animate pronoun to refer to an inanimate object; thus, the referent of the subject is inanimate, but the instance was classified as having an animate subject because the pronoun is grammatically animate. So tough-adjectives receive a robust input cue from inanimate subjects. Interestingly, they also occur a reasonably high proportion of the time with expletive subjects. In addition to the 113 occurrences of tough-adjectives with inanimate referential subjects, they were used with expletive subjects in this dataset seventy-five times. This is a much smaller ratio of inanimate to expletive subjects than we saw with raising verbs (recall: 69 vs. 11) though inanimate referential subjects are still more frequent. This is perhaps a good thing since, as discussed in Section 2.5, expletives are potentially a more important cue for distinguishing tough-adjectives from other sorts of adjectives that occur in object-gap constructions, than they are for distinguishing raising verbs from other verbs that take an infinitive complement. Finally, let us look at unaccusative and unergative verbs. Recall that Scott and Fisher (2009) found that unaccusative verbs occurred with inanimate subjects about 56% of the time (they reported 44% animate subjects), while transitive verbs used without an object occurred with inanimate subjects only 3% of the time. But Scott and Fisher had not searched for unergative verbs. Looking at a smaller set of corpora in CHILDES (Adam and Eve from Brown (1973) and Nina from Suppes (1974); total maternal utterances is 65,345), I hand-coded a set of unaccusative and unergative verbs according to the animacy of their subjects (thus, I counted only intransitive uses of verbs like open and close). Similar to the asymmetry Scott and Fisher found for unaccusative and object-drop transitive verbs, I found that unaccusative verbs occur with inanimate subjects much more frequently than unergative verbs do (see Table 6.3). The proportions are not exactly the same as Scott and Fisher’s, possibly due to hand coding rather than automated coding of animacy, or to the difference in sample size, but the rates I found are roughly in the same neighborhood as theirs. Like the previous tables, this distribution is significant by a chi-square test (χ 2 (1) = 85, p ≤ .001). The rate of inanimate subject use with unaccusative verbs (37.8%) is higher than that with raising verbs (recall: 5.2%) but lower than that with tough-adjectives (97.4%). However, unergative verbs are used

254

Modeling the acquisition of displacing predicates

Table 6.3: Mothers’ use of animate and inanimate subjects with unaccusatives and unergatives Unaccusatives

Animate subj.

Inanimate subj.

close come fall open

0 375 176 1

6 169 139 21

total

552

335

Unergatives cry dance laugh

122 60 25

5 6 1

total

207

12

% Inanimate subj.

37.8%

5.5%

much more frequently with inanimate subjects than control predicates are (control verbs: 0.3%, control adjectives: 0%). The twelve occurrences of unergative verbs with inanimate subjects are shown here in (2). (2)

a. How can a cracker cry? (Adam 18) b. Can bread cry? (Adam 18) c. The teapot’s crying? (Nina 18) d. Because the teddy bear’s crying? (Nina 23) e. I think the teddy bear’s crying because she’s getting mad. (Nina 23) f. It wasn’t dancing. (Adam 17) g. A house can’t dance. (Adam 21) h. What kind of egg dances around? (Adam 21) i. It’s dancing, yes. (Eve 07) j. You making the dolly dance? (Nina 05) k. And now do you want to make the clown dance with the magnet? (Nina 34) l. The sunshine is laughing. (Adam 12)

Four of these examples contain predicates like cry and dance used with toys (2d,e,j,k). So although the subject is coded as inanimate, it may be “animized” in play by the child, parent, or both. Recall a similar “error” in the data above on tough-adjectives, where a deflated horse is referred to with the animate pronoun he. If we remove these utterances the rate of unergatives with inanimate subjects drops to 3.7%, still quite a bit higher than the rate with control verbs

Displacing predicates in the input to children

255

and adjectives. Two additional utterances in this list (2f,g) involve sentential negation, which should perhaps be considered part of the predicate, and therefore these examples are not really semantically ill-formed. In addition, four of the interrogatives (2a,b,c, and h) have the flavor of incredulity, suggesting that the predicate might not truly be applicable to the subject. So, depending on our criteria the number of counterexamples could range between two and twelve utterances. These kinds of utterances raise a difficult question about how to consider animized toys and objects in conversations with children, both in adult speech to children and in children’s own productions. For comparison, in their own speech Adam, Eve, and Nina produced 7% of their unergative verbs with inanimate subjects (e.g. Is it gonna cry? Adam 36; see Section 5.3.3). In Chapter 5 I went to great lengths to argue that children, from a very early age, are cognizant of which entities in their world are animate (agentive, sentient) and which are inanimate (non-agentive, non-sentient). Should these apparently paradoxical expressions of animacy be chalked up to playfulness and imagination, or should they be considered errors? It seems unfair to count the occurrences in adult speech as errors, but will they impede a child’s use of the inanimacy cue in identifying displacing predicates? And what about the occurrence of these expressions in the children’s own speech? As far as I know we are not in a position to give a solid answer to these questions. My own subjective experience with young children tells me that children are able to entertain simultaneously a fantasy-based conception of objects and a reality-based one, so that their allowance of toys, dolls, trains, etc. to speak, sleep, have emotions, and so forth in play, in children’s literature, and in animated cartoons does not warp their sense of what “real” inanimate objects are capable of. My 2-year-old can aver that a plain, ordinary rock is incapable of talking or eating, because, in her words, “it’s not have a mouth” (see the discussion about pretend play and children’s literature in Section 5.1.4). Those utterances in (2) that do not appear to involve obvious cases of animized toys, such as the first two examples, may be cases in which parents are playfully probing children’s awareness of what things can perform which actions. The higher proportion of unergative verbs with inanimate subjects in both adults’ and children’s speech (higher compared to control verbs and adjectives) may arise from the fact that these are verbs that denote physically observable actions and may, for that reason, be more likely to be probed in playful conversation than the more abstract control predicates. That is, parents might be more likely to ask a child, “Can a tree dance?” than “Can a rock try to be heavy?” or “Can a ball be eager to roll?”

256

Modeling the acquisition of displacing predicates

The main pattern we can observe in the tables above is that inanimate subjects are used asymmetrically more frequently with displacing predicates than non-displacing predicates in child-directed speech. Thus, this cue is at least in principle available to children as a means of identifying displacing predicates. We can also observe that inanimate referential subjects are arguably a more prevalent cue than expletive subjects in the case of raising verbs (this is less clear in the case of tough-adjectives). A third observation is that the precise rate of occurrence of inanimate subjects with each type of displacing predicate is vastly different, ranging from 5.2% (raising verbs) to 37.8% (unaccusatives) to 97.4% (tough-adjectives). The fact that none of these predicates are used exclusively with inanimate subjects will become important in the development of probabilistic models in the following sections. 6.2

Computational modeling of language acquisition

Computational modeling via learning simulations is a natural outgrowth of modern linguistic theory. The child’s task as language learner has long been recognized as a computational one: the child must compute the correct grammar (a finite system capable of generating the infinite set of possible structures in the language and none of the impossible ones) based on partial evidence in the input. The input is said to be partial or incomplete (or “impoverished”) in the sense that (a) it cannot contain all of the possible sentences of a language, since that set is infinite, and (b) it does not contain explicit negative evidence – children are not told which sentences are not possible in the language.6 As theorists, we would like to test our hypotheses about how children learn language by engineering a learner (one that makes particular guesses and not others), or by manipulating the input in a particular way and observing the result. As in the Merlo and Stevenson study described above, we can construct a learning algorithm, a procedure or set of guesses we think a real child might employ, and “expose” it to engineered input of the sort children are likely to hear (taken, for example, from actual corpora of child-directed 6 Not all researchers agree that the input to children is truly impoverished; for instance, one

could argue that even though the input does not contain the full set of sentences allowed by the language, it includes enough exemplars of the types of sentences allowed that the child could converge on a close enough approximation of the grammar to communicate effectively. Likewise the argument that children lack negative evidence is sometimes disputed on the grounds that not hearing some form in the input could serve as indirect negative evidence. See, for example, the debate in The Linguistic Review 19, issues 1–2 (2002).

Computational modeling of language acquisition

257

speech), and measure the results. Does the right grammar or categorization emerge? How does manipulation of the input to favor or disfavor certain cues or features affect the outcome? What constraints are necessary to build in to the model in order to achieve successful learning? Computational models can be (and have been) used to explore the learning of various aspects of language: speech segmentation (Saffran et al., 1996; Thiessen and Saffran, 2003; Newport and Aslin, 2004, i.a.), word order (Gomez and Gerken, 1999), word categorization (Mintz, 2002, 2003), regular and irregular morphology (Rumelhart and McClelland, 1986), and inflectional paradigms (Legate and Yang, 2002, 2007), as well as various aspects of phonological acquisition (Pearl and Goldwater (in press) give a good overview of this literature). Here my focus will be on modeling the acquisition of a specific kind of argument structure, namely, the kind in which the subject is not selected by its adjacent predicate. As just mentioned, one of the great strengths of modeling work is that it can help shed light on what (kinds of) assumptions must be in place in order for learning to succeed, and which aspects of the input are most useful for learning certain components of grammar. Much of the work using computational models is aimed at demonstrating that it is possible to learn complex structures in language based on the type of input children hear spoken around them, and therefore that it is unnecessary to postulate innate language structures (Rumelhart and McClelland, 1986; Elman, 1990, 1993). This is certainly not true of all such approaches (e.g. Yang (2002); Pearl and Lidz (2009); Mitchener and Becker (2011); Gagliardi (2012)), and in current modeling work there seems to be agreement that some constraints are innately given and some information is learned from input (Elman et al., 1996). As we will see, Bayesian approaches inherently require some prior assumptions, though these may be specified as domain-specific assumptions (i.e. specific to linguistic cognition) or as domain-general. Much of the disagreement among researchers centers around questions of exactly what is innate and what is learned, and whether the innate constraints are specific to a particular cognitive domain (like language) or can be applied generally to various domains of cognition (see Pearl and Lidz 2009 for a discussion of domain-specific and domain-general learning mechanisms with regard to language acquisition). Before looking at concrete examples of modeling the distinction between displacing and non-displacing predicates, I want to flesh out the problem of generalization, and point out how this problem leads to the need to restrict the hypothesis space. Then I’ll introduce Bayesian models, which can model probabilistic learning but require hypothesis restrictions, and we’ll look at an

258

Modeling the acquisition of displacing predicates

example of how a Bayesian learner can model the acquisition of argument structure. Following that I will show how such models can be applied to the use of animacy features in acquiring the distinction between each of the pairs of predicates discussed in the preceding chapters: raising vs. control verbs, tough vs. control adjectives, and unaccusative vs. unergative verbs. The approach I take in this chapter is to present the information about computational models in general enough terms for non-computationalists to understand, rather than in specific enough terms for computationalists to employ and replicate these models. I do this partly for reasons of accessibility, and partly because much of the computational work on opaque constructions and displacing predicates is in its infancy, and it is more important for the conceptual groundwork to be laid than for precise implementations to be carried out at this stage. Certainly, I see this as a budding area of research, and I hope that collaborations between modeling experts and language acquisition experts will take this as a starting point for more thorough explorations of this learning problem. 6.2.1 Learning as generalization Perhaps the central puzzle to be explained by theories of language acquisition is how language learners manage to generalize patterns without grossly over-generalizing. And, more puzzling still, it must be explained how children retreat from the overgeneralizations they do make without explicit negative evidence (Chomsky, 1959; Marcus, 1993; Pinker, 1989). The importance of correct generalization can be illustrated in a number of ways. One illustration is what Pinker (1987) calls “Baker’s paradox”: certain verbs occur in both the double object and prepositional dative constructions, while other quite similar verbs allow one construction but not the other. (3)

(4)

a.

Mary gave a painting to the museum.

b.

Mary gave the museum a painting.

a.

Mary donated a painting to the museum.

b. * Mary donated the museum a painting. (5)

a. * Mary asked a question to me. b.

Mary asked me a question.

How do children figure out that donate and ask can occur in only one of the two constructions? While they will never hear the offending utterances (4b/5a), they will also never be told that these forms are impossible. Children

Computational modeling of language acquisition

259

frequency

0.1

5 · 10−2

0 0

20

40 60 rank

80

100

Figure 6.1. A power law distribution, illustrating Zipf’s law

must generalize the relevant properties that determine which verbs allow both constructions vs. those that allow only one. Another illustration of the problem of generalization, specifically relating to the poverty of the stimulus, comes from Zipf’s law (Zipf, 1949/1965). Zipf noticed a number of statistical correlates in language, the most well-known of which is that a small number of words occur very, very frequently in language, and a vast number of words (essentially, all the rest) occur with rapidly decreasing frequency. The relationship between the rank and frequency of words is that of a power law distribution (see Figure 6.1), which, when plotted on a log–log scale results in a nearly perfect line with a slope of −1 (see Figure 6.2). The same turns out to be true of syntactic constructions when formalized as phrase structure rules (e.g. PP → P NP) (Yang, 2011). That is, a small number of phrase structure rules are extremely frequent in language, while the rest of these rules occur very rarely. Thus, the problem for the learner is that the vast majority of words and constructions in the language will be encountered only a small number of times, if at all. And yet, children will successfully generalize the rarely encountered words and syntactic rules to new contexts. This outcome implies a large degree of generalization on the part of the learner. With respect to displacing predicates the generalization to be made is that certain predicates that occur in the opaque string [NP verbs [to VP]] or [NP is adjective [to VP]] will also be compatible with an expletive subject (or a

260

Modeling the acquisition of displacing predicates

log(frequency)

10−1

10−2

10−3 100

101 log(rank)

102

Figure 6.2. A power law distribution plotted on a log–log scale

null one, in non-expletive languages), will allow a Quantifier Lowering interpretation (in the case of verbs) or an object-gap interpretation (in the case of adjectives), and so forth, while other predicates that occur in these very same strings will not have those properties. The distinction between the two sets of properties follows from the fact that predicates that have them (can occur with expletives, etc.) do not select an external argument and therefore participate in a very different sort of argument structure than those predicates that lack these properties (i.e. cannot occur with expletives, do not have a QL reading, disallow an object gap interpretation, etc.). The reason this constitutes a generalization problem is that learners will not be given negative evidence concerning the inability of certain predicates to have these properties (e.g. *There claims to be a problem; The girl is eager to see  someone is eager to see the girl). For reasons discussed in previous chapters, it is not sufficient to say that learners assume as a default that predicates are non-displacing, and let the categorization of a predicate as displacing be triggered by hearing the predicate with an expletive. Although expletives surely form part of the informative input children can use in figuring out that a predicate is displacing, they do not by themselves solve the generalization puzzle. Moreover, I think it is implausible, particularly in the case of these highly unobservable predicate meanings, that children primarily acquire these abstract and subtle meanings from their observations of and interactions with the world, in the absence of linguistic structure as a guide. Recall from Section 5.3.2 that there was

Computational modeling of language acquisition

261

evidence for some role of context in children’s categorization of novel adjectives as belonging to the tough or the control class, but that subject animacy played a more decisive role in categorization. Although the lexical semantic properties of displacing predicates are unquestionably bound up with these predicates’ argument structure properties (displacing predicates have meanings related to aspect, appearance, difficulty, modality, or negation, or a few other “bleached” meanings – see Section 3.3.1), as is true for all predicates (Fisher et al., 1991; Lederer et al., 1995), I submit that children do not discover that a verb means “seem” and then conclude that it does not select its subject; rather, they figure out that the verb does not select its subject (most likely by hearing it with an inanimate subject) and then narrow down the verb’s possible meanings: maybe it relates to appearance or happenstance, but not to emotion. 6.2.2 Restricting the hypothesis space One solution to the generalization problem is to build in biases or expectations on the part of the learner in order to restrict the hypothesis space. Children cannot entertain every possible grammar on their way to finding the right one; instead, there must be some assumptions, or inductive biases, in place (Quine, 1969; Gallistel, 1990; Mitchell, 1990, among others). Although somewhat of an oversimplification, this was at the heart of Chomsky’s idea of Universal Grammar: it provides a means of delimiting the possible hypotheses a child can entertain by asserting that certain aspects of grammar are innate.7 In fact, I do take certain elements of the learning of displacing predicates to be very likely innate. For instance, the notion that the sentence subject is prototypically animate is so widespread and so useful as to be a good candidate for an innate expectation, and it is the violation of this prototypical state of affairs that leads to learning the non-prototypical displacing predicates. We will come back to questions about innateness, and will further refine the notion of subject, in Chapter 7; for now what is important is that we can restrict our hypothesis space by including a bias for subjects to be animate, and for predicates to select their subjects. Before refining our proposal for learning displacing predicates, let us look at how the hypothesis space can be restricted within certain Bayesian learning models. In these models, the hypothesis space is restricted by specifying the prior probability of some hypothesis, for example, the hypothesis that a 7 According to some modelers, inductive biases can be “learned” as well (e.g. Kemp et al., 2007;

Perfors et al., 2010). We will come back to these kinds of models in Section 6.3 below.

262

Modeling the acquisition of displacing predicates

string of words is generated by a particular syntactic structure. Much of my discussion here relies on a very useful tutorial provided in Perfors et al. (2011), and I refer readers to that article for a more articulated tutorial in Bayesian modeling. Following this general discussion of Bayesian models we’ll look at one example of how this approach has been used to model the acquisition of argument structure for transitive and intransitive verbs. Then, a slightly more complex type of Bayesian model, a Hierarchical Bayesian Model, will be used in modeling the acquisition of raising verbs, tough-adjectives and unaccusative verbs. Bayes’ rule is a theorem in probability theory that can be applied to calculating the probability of a hypothesis given some set of data, in terms of the prior probability of that hypothesis (the probability that the hypothesis is right in the absence of any data), and what is called the likelihood of the data (i.e. the likelihood of encountering this data if the hypothesis is right). The rule is given in (6) (P = probability; h = hypothesis; d = data). (6)

likelihood prior

      P (d|h) · P (h) P (h|d) = P (d)

The above rule reads as follows: the posterior probability of the hypothesis given the data is equal to the probability of the data given the hypothesis (this is the likelihood) times the probability of the hypothesis (this is the prior probability), divided by the probability of the data alone. One way to describe the rule intuitively is to say that one’s hypothesis is more likely to be correct if either the hypothesis is more likely to be true in the absence of the observed data, or if the data is less likely to be explained by an alternative hypothesis (more on alternative hypotheses in a moment). As Perfors and colleagues put it, “it balances between a sense of plausibility based on background knowledge on one hand and the data-driven sense of a ‘suspicious coincidence’ on the other” (Perfors et al., 2011, p. 303). What they mean by “suspicious coincidence” is that if some data is observed more frequently than would be otherwise expected, the hypothesis predicting that data is likely to be correct; if it were not correct, the high frequency of the data would be a suspicious coincidence. In practice, we generally have multiple hypotheses about a given set of data, and we are trying to decide which hypothesis is right (which one has the highest probability). The term P (d) in (6) is equivalent to the sum of all

Computational modeling of language acquisition

263

probabilities P (d|hi ) · P (hi ) for all hi ∈ H (H symbolizing the set of hypotheses under consideration), so sometimes the simple term in the denominator is spelled out with a “normalizing term” that sums across all of the likelihoods and priors for each of the hypotheses under consideration, as in (7). So, (6) and (7) are equivalent. (7)

P (hi |d) = 

P (d|hi ) · P (hi ) hj ∈H P (d|hj )P (hj )

Bayes’ rule helps us evaluate and compare different hypotheses about some data. Perfors et al. (2011) give a transparent example, which newcomers to modeling might find instructive, in which they use Bayes’ rule to evaluate hypotheses for explaining why someone is coughing, e.g. a cold, lung cancer, heartburn. They show how assigning values to the terms in (7) according to our intuitions about how frequent each of these ailments is in the general population (colds and heartburn are common but lung cancer is not; this is the prior probability) and how likely each one is to cause coughing (colds and lung cancer are likely to cause coughing, heartburn is not) can yield one of these hypotheses winning out over the others. The “winning” hypothesis is the one with the highest posterior probability, and this comes about when the product of the prior probability of one hypothesis and the likelihood of that hypothesis is the highest value (closest to 1) compared to the competing hypotheses (in their example, it turns out that a cold is the most probable cause of coughing). Their example also demonstrates how the posterior probabilities of each of the hypotheses under consideration sum to 1, which reflects our intuition that as one hypothesis becomes more likely to be true (closer to 1), others become less likely to be true (closer to 0). How does Bayesian reasoning apply to restricting the hypothesis space in learning language? This procedure can be easily adapted for categorization problems (e.g. learning the shape bias, as in Kemp et al. (2007), or words for object categories, as in Xu and Tenenbaum (2007)), and categorization is a component in many of the learning tasks within language acquisition. To see an example of linguistic categorization based on Bayesian inference, suppose you are learning a new language and encounter a verb a number of times in an intransitive frame. One hypothesis is that this verb is unergative and will only ever occur in intransitive frames (e.g. laugh; as discussed above, some languages will allow a few of these verbs to also occur in transitive frames). Another possibility is that the verb is unaccusative and allows alternations

264

Modeling the acquisition of displacing predicates

between intransitive and transitive frames (break), but the transitive occurrences have not been encountered yet. A further possibility is that the verb is transitive but can be used optionally without a direct object (eat). As more intransitive occurrences show up in the data, the more likely becomes the narrower hypothesis that the verb can only be intransitive and thus is unergative (Niyogi, 2002). It would be, as Perfors and colleagues call it, a “suspicious coincidence” if the verb were actually an alternating verb, but no transitive uses are found in the data. Alishahi and Stevenson (2008) developed a Bayesian model for acquiring these classes of transitive and intransitive verbs (see also Alishahi and Stevenson, 2005a,b). They wanted to show that a Bayesian learner could generalize over pairings of constructions (e.g. transitive, intransitive) and semantic features associated with different constructions (e.g. the transitive use of a transitive-intransitive alternating verb has a “causative” meaning; this is similar to the Merlo and Stevenson (2001) study described previously, but it employs a Bayesian learner instead of a decision tree algorithm). For their model, Alishahi and Stevenson make the assumption that the learner has access to the argument structure frame of the sentence each verb occurs in (neighboring nouns are assumed to be the verb’s arguments) and enough semantic information about the sentence and the verb to be able to extract semantic “primitive” features (e.g. cause and move for the verb put) and the thematic roles of the arguments in the sentence (agent, theme, etc.). For reasons given in Chapter 5, and elaborated below, neither of these assumptions will be valid for learning raising and control verbs. But let’s allow them as plausible assumptions for transitive and intransitive verbs for the moment. Each syntactic frame is then associated probabilistically with semantic features, and groups of frames that overlap to a large degree in their semantic features are called a “construction.” For example, frames that have two arguments, semantic primitives such as act and participant roles in which the first argument is an agent and the second is a theme will be grouped together as a generalized transitive construction. As new input sentences are encountered, the model determines the argument structure frame for each sentence and associates it with a construction with some probability. Predicates are then associated with constructions, and the strength of those associations depends on how frequently a predicate is encountered in each construction. In Bayesian terms, the posterior probability of a construction (k) given a particular frame (F ) is proportional to the

Hierarchical Bayesian Models

265

likelihood of the frame times the prior probability of that construction (how frequent that construction is overall). (8)

P (k|F ) =

P (F |k) · P (k) P (F )

Alishahi and Stevenson evaluate their learning model by having it “predict” the syntactic frame of a novel verb, given only semantic features (which they liken to the task of production in real language acquisition) and also by having it “predict” the semantic features of a novel verb, given only argument structure input (akin to comprehension in children). The results of their simulation show that the model performed fairly well, though better in nonnoisy than in noisy data. For some verbs the model appeared to go through an initial stage of “imitating” the frames or features given (i.e. not generalizing), followed by a stage of (over-)generalizing based on the frames or semantic features. In cases of overgeneralization, subsequent positive input was sufficient to retreat from the overgeneralization enough to reach about 80% accuracy. 6.3

Hierarchical Bayesian Models

I noted in the previous section that some of the assumptions used in Alishahi and Stevenson’s Bayesian learner are not feasibly applied to the problem of learning displacing predicates. Let’s review why this is so. Briefly, the construction of a “surface parse” of a sentence based on a sentence string (NP V NP → [NP [V NP]], etc.) requires the assumption that adjacent NPs are the arguments of the verb. But in the case of displacing predicates, this assumption leads to the wrong conclusion, since their adjacent subject is precisely not their argument (at least, not their external argument). Moreover, I believe it is implausible that children might figure out the lexical semantics of these verbs prior to having cues from argument structure, given their rather extreme abstractness (see Gleitman et al. (2005) and Section 5.2.2 above). In addition, because the verbs we are interested in are few in number and some quite rare in speech, we cannot make use of more traditional training and testing methods as we’ve seen used with, for example, transitive and intransitive verbs. In fact, when my collaborator Garrett Mitchener and I applied Alishahi and Stevenson’s model (which requires strict categorization of verbs) to our input data on raising and control verbs (see Section 6.1, Table 6.1) it

266

Modeling the acquisition of displacing predicates

was unsuccessful. Instead of generalizing two categories of verbs (raising and control), the model distinguished individual verbs exactly according to their proportion of occurrence with animate vs. inanimate subjects (see Mitchener and Becker (2011) for details).8 Thus, the learner failed to generalize from the data. However, we had better success with two other types of learning models which, rather than requiring strict categorization of predicates, place predicates along a continuum from “less displacing” (more control-like) to “more displacing” (more raising-like). An advantage of having our model place the verbs along a continuum is that we can better model the actual behavior of these predicates, which we saw in Section 3.3.1.1 is more of a continuum than a true dichotomy. A possible disadvantage is that we lack a means of “testing” our model on novel data. However, we can still evaluate how well our model ranks the verbs with respect to each other by inspecting the continuum post hoc. If the model ranks seem, for example, as being “more displacing” than try, then it has achieved an accurate placement of these verbs; if not, it has failed. Mitchener and I have employed two different kinds of models with similar results. These were a Hierarchical Bayesian Model (HBM) and an accumulator model developed by Mitchener. Here I will focus on the Hierarchical Bayesian Model. HBMs employ not only a prior probability of some hypothesis, but also prior hypotheses about the hypotheses themselves (Kemp et al., 2007; Perfors et al., 2010, 2011). As explained in Perfors et al. (2011), the intuition behind HBMs is that there are often higher-order generalizations about the lower-order generalizations, also called inductive constraints. They give the example, attributed to Goodman (1955), of selecting colored marbles from different bags. Imagine that on each draw you select white marbles from one bag, black marbles from another bag, and red marbles from yet another bag, but in every case you draw the same color marble from a given bag. In other words, the data tell you what color marbles are in each bag, but also that the bags are uniform in color. Then, upon encountering a green marble drawn from a new bag you can develop a hypothesis about the color of the next marble to be picked from the same bag, namely, that it will also be green. This is 8 A possible complicating factor in evaluating that model is that we also used data about the

eventivity of the embedded predicates, and these additional features may have led to problems. The rationale for including this data came from the result in Becker (2005), discussed in Section 4.2.1, that adults were influenced in their choice of raising vs. control verbs by the eventivity of the lower predicate, but the source of that influence is unclear. We exclude these features from the model discussed below.

Hierarchical Bayesian Models

267

because you already have the broader hypothesis that the marbles in each bag are uniform in color. This is called an “overhypothesis,” or “hyperparameter” (Goodman, 1955; Perfors et al., 2010). 6.3.1 A model of learning raising and control By employing overhypotheses, or hyperparameters, we can model the intuition that subjects generally tend to be animate in language (see extensive discussion in Chapter 3), and that predicates generally tend to select their subjects (this is what makes displacing predicates unusual). Our model has three layers, illustrated in Figure 6.3 below. Starting from the bottom of our model, the data for our learner is the known value y i for each verb, which represents the counts of how many sentences contain each verb occurring with an animate or an inanimate subject (taken from naturalistic speech, to be described below). Since we are interested in the counts of occurrences with both animate and inanimate subjects, y i is a set of pairs ((seem–animate, seem–inanimate), (try–animate, try–inanimate), . . . ). We also include ni , the total number of sentences for each verb (ni = y i,animate + y i,inanimate ) (ni is not shown in the picture). The next layer of our model, the hypotheses or parameters, are represented by β, the baseline tendency of subjects to be animate in language, and λ, the

φ1

γ1

φ2

βseem

γ2

λseem

yseem

λtry

ytry

λneed

yneed

Figure 6.3. Graphical representation of the Hierarchical Bayesian Model; the highest layer is the hyperparameters φ 1 , φ 2 , γ 1 and γ 2 ; the middle layer includes the parameter β, the baseline rate of subjects being animate (its value is the same across the data set), and the parameter λ, a verb’s selectivity (its value varies for each verb); the lowest layer includes the counts of animate and inanimate subjects for each verb based on the data set.

268

Modeling the acquisition of displacing predicates

Figure 6.4. Graphical representation of β and λ in Hierarchical Bayesian Model; β represents the baseline rate of occurrence of animate subjects in the language. λi represents a verb’s “choosiness” about occurring with animate subjects relative to that baseline. Both are beta distributions ranging between 0 and 1. q i is the probability that a given verb will occur with an animate subject.

degree to which a verb is selective, or “choosy,” about having an animate subject. The value of β will be constant across a data set, since it represents a baseline value for the language, while the value of λ will be potentially different for each verb, so for each verb i there will be a λi . The values of λi and β are the values we are most interested in inferring. We will use the inferred values of λi for each verb in order to rank the verbs along a continuum. The values of both β and λ are represented as beta distributions. A beta distribution is a commonly used distribution for random numbers between 0 and 1.9 The value of β is unknown (it will be estimated by our model), so in Figure 6.4 it falls at some unknown point along the continuum between 0 and 1. The value of λi will also be between 0 and 1, and λ = 0 means that the preference for animate subjects will correspond to the baseline rate of animate subjects in the language. We can also combine λi and β to calculate q i , the probability that a given verb will occur with an animate subject (q i = λi + (1 − λi )β). A beta distribution is associated with two parameters, whose values determine the shape of the beta distribution’s density curve. The ratio between the two parameters determines how centered (or skewed to the left or right) the distribution is (if their ratio is close to 1 it will be centered), and their absolute magnitudes determine how “pointy” or “flat” the distribution is: a large value will be associated with more pointiness (one is more certain about the distribution because there is more data), while smaller values will be associated with 9 A note about terminology: the words beta in beta distribution and the parameter β in our model

bear the same label accidentally; they are not related.

Hierarchical Bayesian Models

269

a flatter curve (less certainty). Since these are parameters about parameters, moving up to the next level of our model we have two hyperparameters, or overhypotheses. Let’s call the hyperparameter about subject animacy φ (so, the parameters for β are φ 1 and φ 2 ), and the hyperparameter about predicate selectivity γ (so, the parameters for λ are γ 1 and γ 2 ). Just like β and λ, we don’t know exactly what the values of φ and γ are, in fact we want the learner to infer their values from the data. As the parameters of a beta distribution their value needs to be positive but is otherwise unconstrained, and so we represent them as being exponentially distributed with a mean of 1. Setting the mean to the low value of 1 is a standard assumption made about unknown priors and simply means that the model does not have strong prior expectations – this allows results to be more heavily influenced by the data rather than prior expectations.10 To summarize so far: at the top level of the model the hyperparameters φ 1 and φ 2 control how often predicates generally occur with animate subjects (which is β), and the hyperparameters γ 1 and γ 2 control how choosy predicates are about having an animate subject (λ). Both β (whose value holds for a language) and λ (whose value will differ across individual verb types) determine the distribution of animate and inanimate subjects seen for a given verb y i . The actual counts for y i come from two types of corpora, one consisting of child-directed speech, taken from the Brown (1973) corpus within the CHILDES database (MacWhinney, 2000), and the other consisting of adultdirected speech, taken from an annotated version of the Switchboard corpus (Bresnan et al., 2002; Taylor et al., 2003). The child-directed speech data are essentially those reported above in Table 6.1, but with ambiguous verbs included. We searched for specific raising, control, and ambiguous verbs and counted (by hand) the number of times each verb occurred with an animate or inanimate subject. The distribution of these verbs with respect to subject animacy is given in Tables 6.4 (child-directed speech) and 6.5 (adult-directed speech). In the case of raising and control verbs, we can apply the HBM in the following way. Particular verbs are observed in the dataset to occur with animate subjects with a certain frequency (y i ), and based on this frequency and the total 10 We based our model on Perfors et al.’s model L3, in which two layers of overhypotheses are

learned, rather than having the highest level be prespecified. This contrasts with their model L2, in which the highest-level parameters were predetermined. Their L2 model was less flexible in learning than the L3 model.

270

Modeling the acquisition of displacing predicates

Table 6.4: Distribution of raising and control verbs with animate/inanimate subjects, from CHILDES Raising

Animate

Inanimate

seem used (to) going (to)

4 45 1197

6 5 58

total

1246

69

Control want like try love hate

405 210 86 10 1

2 0 0 0 0

total

712

2

Ambig. start begin need

4 1 38

0 0 4

total

43

4

% Inanimate subj.

5.2

0.3

8.5

number of sentences containing that verb (ni ) we can calculate the proportion y ). Then, based on of the time the verb occurs with animate subjects ( i,animate ni this proportion relative to a baseline (β), we can infer how choosy that verb is about having an animate subject (λi ). The idea is that verbs that occur with animate subjects a high proportion of the time relative to β are inferred to have a high value for λi (approaching 1); these verbs are very selective and require an animate subject. Verbs that occur with animate subjects a low proportion of the time relative to β are inferred to have a low value for λi (closer to 0); these verbs are not selective and do not require an animate subject. Given that the lower bound of λ is linked to the value of β, a low or near-0 value for λi essentially means that the verb allows inanimate subjects about as often as they would ever occur in language, not that inanimate subjects are required. As we have seen in Chapters 2 and 3 some predicates are more selective about having animate subjects than others. Thus, rather than perform a strict categorization task our model places verbs along a continuum ranging from very choosy (essentially requiring animate subjects) to not choosy (allowing

Hierarchical Bayesian Models

271

Table 6.5: Distribution of raising and control verbs with

animate/inanimate subjects, from switchboard Raising

Animate

Inanimate

seem used going tend happen

81 252 55 73 33

94 37 9 19 8

total

494

167

Control want try like love hate choose

465 161 214 18 27 6

3 1 0 0 0 0

total

891

4

Ambig. need have start begin continue

270 547 15 4 10

26 11 9 3 1

total

846

50

% Inanimate subj.

25.3

0.4

5.6

inanimate subjects). Representing verbs in a continuum in this way allows us to mirror the actual behavior of these predicates (see Sections 2.5 and 3.3.1). So we might conclude that a verb like try is very selective about having an animate subject, while a verb like begin is somewhat less selective, and a verb like seem is completely flexible. These relative preferences are best modeled as probability densities, rather than discrete probabilities. Fortunately, Bayes’ rule works for inferring probability densities as well as discrete probabilities (e.g. probabilities of events), and so we use the symbol p instead of P to denote this difference. We let y i be the number of sentences with verb i with an animate or inanimate subject, and ni be the total number of sentences for each verb. Using Bayes’ rule we can then estimate the value of the parameters λi and β for each verb:

272

Modeling the acquisition of displacing predicates

(9)

      p(yi |λi , β, ni ) · p(λi , β) p(λi , β|yi , ni ) = p(yi |ni )

likelihood

prior

In the equation in (9) the first term in the numerator on the right side of the equals sign is the likelihood term (the probability of the data conditioned on the parameters – so, the probability of observing this distribution of y i given λi , β, and ni ), the second term is the prior (the expectation that predicates will be selective and that subjects will be animate), and the denominator is a normalizing term. Recall that there is a different λi value for each verb, but only one β value for the whole dataset. Thus, more generally we have the following equation, across verbs: (10) , β| p(λ y , n , φ i , φ 2 , γ 1 , γ 2 ) =

β, n , φ 1 , φ 2 , γ 1 , γ 2 ) · p(λ , β|φ 1 , φ 2 , γ 1 , γ 2 ) p( y |λ, p( y | n)

The vector notation in (10) means that it is including data from the whole dataset, not only from individual verbs ( y = ((seem–animate, seem–inanimate), (try–animate, try–inanimate), . . . )). Mitchener used a Markov Chain Monte Carlo (MCMC) method to estimate the posterior values for each λi . We can represent the outcome of our model by plotting the distribution of posterior values for λi . Figures 6.5 and 6.6 plot this distribution, i.e. the relative selectivity of each verb, on a scale from 0 to 1 (as before, 0 meaning it is flexible – it allows inanimate subjects – and 1 meaning it is completely selective – it requires animate subjects), first for the CHILDES data and then for the Switchboard data. The median value of λi is indicated by the narrow black line within each bar plot, and differently shaded segments of each plot indicate deciles. More data causes the posterior values to be more densely clustered around the median (less spread out). Verbs with fewer data points (e.g. love, start, hate, begin in CHILDES) have less densely clustered values (more spread out).11 The result of applying the HBM to both data sets was quite good. In both cases the plot yields a progression of verbs from “very selective” (e.g. like, want, try) to non-selective (seem). Nearly all of the control verbs have their 11 These figures do not appear in Mitchener and Becker (2011); they were compiled by Mitchener

after that paper was published.

Hierarchical Bayesian Models

273

like want try love start begin hate going need used seem 0.0

0.2

0.4

0.6

0.8

1.0

Figure 6.5. Posterior distribution plots for λ; CHILDES data. Values range from 0 (flexible; allowing inanimate subjects) to 1 (selective; requiring animate subjects).

like want try hate have love choose continue need used going happen tend begin start seem 0.0

0.2

0.4

0.6

0.8

1.0

Figure 6.6. Posterior distribution plots for λ; Switchboard data. Values range from 0 (flexible; allowing inanimate subjects) to 1 (selective; requiring animate subjects).

median λi values directly at or very close to 1.0, and raising and most ambiguous verbs have a median λi much lower than 1. Only two verbs are placed out of order: start and begin (ambiguous) are placed below several raising verbs. In the CHILDES data all raising verbs other than going have median λi values below .8, and only hate (used only once in the corpus) is placed

274

Modeling the acquisition of displacing predicates

out of order, below start and begin. In both data sets the verb with the lowest median λi is seem. Given these results, can we conclude that raising and control verbs can be distinguished on the basis of subject animacy? Although some verbs were placed slightly out of order with respect to our categorization of these verbs (see Tables 6.4 and 6.5), the model performed quite well in estimating that the verbs that select their subjects (control verbs) are pickier about having an animate subject than verbs that don’t select a subject (raising verbs). It is also important to recognize that the input to the model involved quite sparse data. Except for a couple of verbs with very high frequencies (going and want in CHILDES, want and have in Switchboard), most verbs occur so few times it could be difficult to draw strong conclusions about their subject selectivity (this is seen in the very spread out bar plots for many verbs). An additional indication of the sparseness of the data comes from the inferred values of β. Recall that this term represents the baseline (or minimum) rate of animate subjects in the language, and our learner inferred this value from the data. As expected, the inferred value of β was quite low: .43 for CHILDES and .38 for the Switchboard data (i.e. these values are “low” compared to the actual mean proportion of animate subjects in language, which ranges from .91, in Switchboard, to .96, in CHILDES). As baseline values these seem appropriate; however, the standard deviations for β were quite large: .156 for CHILDES and .140 for Switchboard, suggesting uncertainty about its value. Uncertainty can arise when there are not enough observations (input data) to be certain about something. We can derive three implications of these results. One is that although the HBM worked well using only information about verbs’ distribution with animate and inanimate subjects, perhaps the model would work even better if this information were used in conjunction with other cues, such as expletives. In Section 2.5 I gave several reasons why expletives could not solve this learning problem by themselves (their rarity, their non-universality, among others), but surely they provide some important information about a verb’s selectional properties. It would be interesting to examine the distribution of inanimate subjects with raising verbs in languages that do not have expletives, to see if there are stronger empirical cues from subject animacy in these languages. To my knowledge this has not yet been studied. A second thing to notice is that the model performed similarly on both adultdirected and child-directed speech. Both forms of speech serve as input for language learning. Child-directed speech certainly has direct and measurable effects on children’s language development (Snow and Ferguson, 1977; Hart

Hierarchical Bayesian Models

275

and Risley, 1995), and there is evidence that overheard speech contributes to children’s learning of vocabulary (at least in experimental settings; OshimaTakane (1988); Akhtar (2005)) (see also Blum-Kulka and Snow (2002)). Moreover, in cultures in which children generally receive very little directed speech input, children do not show delays in their own language development, suggesting that speech directed at other individuals around them (including adults) forms their primary input data (de Leon, 1998) (on the other hand, see Schneidman et al. (2013) for caveats about the effectiveness of overheard speech in naturalistic settings). A third implication is that given the somewhat weak asymmetry between raising and control verbs in terms of their occurrence with inanimate subjects (i.e. both types of verbs mostly occur with animate subjects), particularly in child-directed speech, learners will be at an advantage if they have strong and/or specific prior expectations about both the tendency of subjects to be animate and the tendency of predicates to select their subjects. In building our model, Mitchener and I set the mean value of the hyperparameters to be low so as to allow the posterior densities to be driven more by the data than prior expectations. But it is important to remember that these prior expectations played some role in the model. With these expectations in place, counterevidence can be informative of a non-canonical pattern. It is worth noting that these expectations about subject animacy and predicate selectivity are a domain-specific type of bias, which is different from the sorts of biases other researchers typically build into these models. For instance, Perfors et al. (2010) argue that an advantage of HBMs is that although the highest-level priors have to be specified in some way, they are sufficiently abstract that they might not need to be considered domain-specific. That is, if the built-in (innate) assumptions are things like “there are categories,” this would be an assumption that is not specific to language; rather, it applies to many different domains of inference. They see this as an advantage because it allows for a reduction in the amount of language-specific knowledge that has to be assumed to be priorly known. But when viewed more closely, the learning models that purport to contain no language-specific prior knowledge do contain assumptions that are suspiciously linguistic. For example, Perfors et al. (2010) note that their work “assumes a pre-existing knowledge of constructions” (p. 624, footnote 6). In their discussion they wonder whether their model worked so well because it was supplied with “clean features that already picked out precisely the constructions of interest,” (p. 636) and they speculate that a more realistic learner would require a “richer semantic representation,” but this strikes me as something that would be language-specific. Similarly,

276

Modeling the acquisition of displacing predicates

Alishahi and Stevenson’s (2008) model assumed knowledge of thematic roles and the mapping between thematic roles and syntactic positions. We will come back to a discussion of prior biases and expectations in Chapter 7, but it is important to keep these issues in mind when we evaluate the success (or lack of success) of learning models, because it is not clear that successful models necessarily learn these grammatical patterns and categorizations without language-specific expectations. It is also important to consider what the ability to categorize predicates buys us in the larger picture of language acquisition: what does it mean for two (or more) categories of predicates to be distinguished? Specific to the present case, what does it mean for a learner to realize that some predicates are very choosy about having an animate subject while others are more flexible? Will it follow automatically that the flexible predicates allow expletive subjects, Quantifier Lowering interpretations, and be able to split idioms, and that the truth value of a complement clause under a flexible verb is preserved under passivization (recall examples (5) and (6) from Chapter 2)? I submit that these additional properties do not automatically follow unless these categories are interpreted within a framework of grammatical knowledge in which it can be understood that a predicate that is flexible has a particular argument structure – it does not select an external argument. Thus, the ability to perform a Bayesian inference about predicate categories based on distributional data is useful for language acquisition in so far as linguistic knowledge is assumed as a prior hypothesis.

6.3.2 A model of learning tough-constructions As we saw in Section 6.1 there is only a small (though statistically significant) asymmetry in the use of raising and control verbs with inanimate subjects in parental speech, but a much starker asymmetry for tough-adjectives and control adjectives. Because of this starker asymmetry, the categorization problem is potentially less grave for tough-adjectives. One could imagine a deterministic procedure to distinguish these two classes (i.e. Subjinanimate → tough; Subjanimate → control), though this rule would potentially group non-tough-adjectives that permit object gaps, like ready, together with tough-adjectives. So let us see how our Hierarchical Bayesian Model fares in discriminating tough from control adjectives, plus the odd adjective ready. The input to our model in this case was the distribution of adjectival predicates with animate vs. inanimate subjects in Table 6.2, repeated here, plus the adjective ready.

Hierarchical Bayesian Models

277

Table 6.6: Adults’ use of animate and inanimate subjects with

tough/control adjectives Tough-Adj.

Animate subj.

Inanimate subj.

% Inanimate subj.

hard difficult easy tough

2 0 1 0

93 3 17 0

total

3

113

Control Adj. happy afraid anxious willing glad eager

8 7 2 1 14 0

0 0 0 0 0 0

total

32

0

0%

Other ready

138

16

10.4%

97.4%

The plot in Figure 6.7 shows the posterior values for λi , which, as in the raising and control model above, indicate the degree to which a predicate is selective about having an animate subject. A value of 1 means the predicate requires an animate subject; a value of 0 means the predicate has no animacy requirement. The solid black line is the median value, and more data causes the posterior values to be more densely clustered around that median. As with certain verbs in the section above, adjectives with fewer data points (e.g. anxious, willing) have more spread out values. What we see is that the three tough-adjectives have λi s at or very close to 0; the control adjectives, instead, have median λi values approaching 1, with more or less spreading around the median depending on the number of data points for each adjective. Thus, inanimate subjects appear to provide a very good cue to displacing predicates in these adjectival constructions. However, it is important to note that inanimate subjects are still not completely predictive, either of tough-adjectives or of object gap constructions more generally. The adjective ready, discussed above in Section 2.2.2, occurs more frequently than any of the control adjectives, it occurs with both animate and inanimate subjects, and its inanimate subjects are both semantic subjects

278

Modeling the acquisition of displacing predicates

Figure 6.7. Posterior distribution plots for λ for each adjective. Values range from 0 (flexible; allows inanimate subjects) to 1 (selective; requires animate subjects).

(indicating subject gaps; but see footnote 12) and semantic objects in the lower clause (indicating object gaps). Its distribution is shown in Table 6.7. Ready is an interesting adjective because unlike the straightforward control adjectives (happy, glad, anxious, etc.) it allows both subject and object gaps (it was one of Cromer’s “A-adjectives”). This is partly due to the fact that it permits inanimate subjects, unlike happy and other emotive adjectives, but notice that even when the subject is inanimate it is still a subject-gap construction a majority of the time (81%) (these were utterances like I think this train is getting ready to leave (Nina file 54)12 ). On the other hand, ready is not a tough-adjective since it does not allow an expletive subject (*It’s ready to read that book; cf. It’s easy to read that book). When we include the data from ready in our model, as shown in the figure above, ready is grouped solidly with the control adjectives. This is an interesting result, because it shows that occurrence with inanimate subjects at a relatively low rate does not automatically

12 All but three of the inanimate subject utterances with a subject gap contained an unaccusative

or passive embedded verb, and thus in a sense are underlyingly object gaps. Of the three that contained an unergative or transitive embedded verb, one involved an “animized toy” subject (the parent was reading The Little Engine that Could, whose pervasive animization was discussed in Section 5.1.4; this occurrence was in Valian file 04a); one involved an unergative verb of emission which easily takes an inanimate subject (I think it’s ready to pop; Kuczaj file 27), and the third was The reason that babies need formula is that . . . they’re so young that their tummies are not ready to really eat the milk yet (Sachs file 81).

Hierarchical Bayesian Models

279

Table 6.7: Distribution of ready in speech to

children Subject type Animate

Inanimate

Subject gap

Object gap

Subject gap

Object gap

138

0

13

3

indicate a displacing predicate. In my view, this is the advantage of employing a probabilistic learning procedure rather than a deterministic one: the learner will not be thrown off track by a few occurrences of the unexpected pattern. On the other hand, it potentially raises problems for displacing predicates that just happen to be used very frequently with animate subjects, as we will see with certain unaccusative verbs in the next section.

6.3.3 A model of learning unaccusatives and unergatives Previous research has demonstrated that animacy can be used as a probabilistic cue for distinguishing the classes of unaccusative and unergative verbs (Merlo and Stevenson, 2001; Alishahi and Stevenson, 2008, discussed above). Scott and Fisher (2009) extended Merlo and Stevenson’s analysis of the Wall Street Journal corpus by using child-directed speech in the CHILDES database. As discussed above, in their study they wanted to see which cues available in parental speech could support the distinction between unaccusative-transitive alternating verbs (what they call “causal” verbs; e.g. open or break) from transitive verbs that permit an omitted object (what they call “unspecifiedobject” verbs; e.g. eat or read). They note that unlike the WSJ data, in the child-directed speech data they examined the two types of verbs were used transitively at equivalent rates, so that transitivity itself was not an informative cue for distinguishing the two classes. However, subject animacy in intransitive uses of these verbs was a highly informative cue: unspecified-object verbs, when used intransitively, had significantly more animate subjects (97%) than causal verbs used intransitively (44%). Recall that their unsupervised learning algorithm robustly distinguished the two types of intransitive verbs based on the distribution of animate subjects (see discussion at the beginning of this chapter).

280

Modeling the acquisition of displacing predicates

Table 6.8: Mothers’ use of animate and inanimate subjects with unaccusatives and unergatives Unaccusatives

Animate Subj.

Inanimate Subj.

close come fall open

0 375 176 1

6 169 139 21

total

552

335

Unergatives cry dance laugh

122 60 25

5 6 1

total

207

12

% Inanimate Subj.

37.8%

5.5%

These studies point to subject animacy being a useful cue for discriminating among different classes of verbs that may or may not have a direct object. But the more relevant question to my work is whether subject animacy can help children discriminate intransitive unaccusative verbs from intransitive unergative verbs, in the sense of categorizing the first class as taking only an internal argument and the second class as taking only an external argument. We saw in Section 6.1 that unaccusative verbs are used asymmetrically more often with inanimate subjects than unergative verbs are, in speech to children. This asymmetry was better than what we found for raising vs. control verbs, but weaker than that with tough vs. control adjectives. For consistency with the other constructions investigated here, let us see whether our HBM can correctly place the two classes of intransitive verbs on the continuum from “more selective” (unergative) to “less selective” (unaccusative). Using as input the distribution data in Table 6.3, repeated here, our Hierarchical Bayesian Model does rank unaccusative verbs as generally more displacing than unergative verbs, though the distinction between the two types of predicates is less stark than it was for tough- and control adjectives. As seen in Figure 6.8, open and close have posterior λi values close to 0, and cry, laugh, and dance have λi values close to 1, as expected. However, come and fall wind up near the middle. This corresponds to their relatively higher rate of use with animate subjects (see Table 6.8). Nevertheless, we can see a distinction between the two classes of verbs around the value of .8, similar to what we found for raising vs. control verbs in the Switchboard data.

Summary of modeling results

281

Figure 6.8. Posterior distribution plots for λ; unaccusative and unergative verbs. Values range from 0 (flexible; allows inanimate subjects) to 1 (selective; requires animate subjects).

6.4

Summary of modeling results

In the preceeding sections we saw that a special type of Bayesian learner that incorporates “overhypotheses,” or generalizations about generalizations, is mostly able to place displacing and non-displacing predicates on different sides of a continuum on the basis of known distributions of these predicates with animate vs. inanimate subjects in naturalistic speech. Unsurprisingly, the model was most successful at differentiating tough-adjectives from control adjectives, as these predicates show the most asymmetric distribution with respect to subject animacy: control adjectives are used exclusively with animate subjects, and although tough-adjectives are grammatical with both animate and inanimate subjects, in fact they are used virtually only with inanimate subjects. The model differentiated unaccusative from unergative verbs on the basis of the frequency of occurring with (in)animate subjects, placing unergative verbs higher on the scale (“more selective”) than unaccusative verbs. However, we did not see as neat a split between unaccusatives and unergatives as we did between tough and control adjectives. Interestingly, it did not appear that the relatively more frequent use of unergative verbs with inanimate subjects caused problems (these verbs were still solidly judged to be non-displacing predicates), rather it was the relatively high frequency of certain unaccusative

282

Modeling the acquisition of displacing predicates

verbs with animate subjects (come and fall) that led to a weaker differentiation of these predicates. The model performed a little less cleanly with raising and control verbs. Control verbs were generally placed higher on the scale than raising verbs, and ambiguous verbs mostly clustered in between the control verbs at the top of the scale and raising verbs at the bottom. But a handful of verbs were “misplaced” with respect to each other. For example, in the CHILDES data hate (control) was placed below begin and start (ambiguous), and need (ambiguous) was placed below going (raising). In the Switchboard data begin and start (ambiguous) were placed below used, going, happen, and tend (all raising). Nevertheless, for both data sets the model placed more frequent control verbs solidly at the top of the scale and placed seem at the bottom of the scale. Moreover, it did so even though some of the verbs had very low numbers of occurrence in the corpus. The models described above relied heavily on distributional evidence in the input to children. But the models also incorporated some built-in assumptions about language in the form of overhypotheses. Specifically, our models have the prior expectation that some predicates will be very “choosy” about their subject, while other predicates will not be choosy. In a sense, then, we specify the knowledge that there might be displacing (i.e. “unchoosy”) predicates as prior knowledge (this was the parameter λ in our model). We also built in the expectation that subjects will be animate at least some portion of the time in the input (this is the other parameter, β in our model). It is important that our model worked with these prior assumptions in place. In the next chapter we will explore in some more depth what this implies about children’s innate knowledge.

7 Conclusions and origins

The end of this book is about beginnings. What is the nature of children’s prior knowledge that allows them to understand “true grammatical relations” that are “not expressed directly on [the] surface” form of a sentence – that is, to acquire the correct internal structure of opaque constructions? I have argued that hearing a main predicate used with an inanimate subject is one important clue to this puzzle. As spelled out in the Derived Subject Condition (Section 3.5), if the subject is the least animate NP in the sentence (it has the fewest Proto-agent entailments or the most Proto-patient entailments, in Dowty’s terms), it is likely to be derived. Therefore, this clue allows children to figure out that the predicate most likely permits a syntactic subject that is not a semantic subject, or external argument. Hearing an inanimate subject is evidence that comes from the input, but in Chapter 6 I argued that the learner required inductive biases in order to generalize, and the learning models we employed in Section 6.3 used languagespecific biases to solve the generalization problem: specifically, the assumption that subjects would be animate some proportion of the time, and that predicates could be more or less selective about the animacy of their subject. In this chapter I want to look a little more closely at the issue of innateness, and what arguments might be made for these biases being a type of innate knowledge in language learners. The first of these biases – that subjects will tend to be animate – presupposes a conceptual distinction between animate and inanimate referents. This type of knowledge, unlike the other two biases I have argued for, is not specific to language. So we will entertain separately the questions of whether there is domain-general innate knowledge of animacy, and whether there are domainspecific biases concerning animacy (and, more specifically, subject animacy) and predicate selectivity in language. Before turning to evidence bearing on these specific pieces of knowledge, I think it is important to lay out what is meant by the contentious term “innateness.” First, several things that are not meant are the following. 283

284

Conclusions and origins (a) Innate means inevitable in all circumstances: one obvious example illustrating that this cannot be what is meant is that an individual may be hard-wired to acquire language, but if no language input is presented (because of profound deafness or pathological language deprivation) language will not emerge. I take an innate ability to be “inevitable” only with the qualification that certain external factors must be present, i.e. in the normal course of events (see below). (b) Innate means inflexible: there is ample evidence of neurological plasticity especially early in life so that in the aftermath of localized trauma or lesioning, nearby cortical areas can in many cases be recruited to “take over” some of the tasks of the damaged areas. Such plasticity, which is not without limits, is not at odds with the fact that certain cortical regions are generally (in the normal course of events, in most individuals, etc.) associated with certain cognitive functions (vision, audition, language, etc.). The debate over modular vs. holistic cognitive architecture is actually orthogonal to the debate about innateness, though many proponents of innateness also lean in favor of modular views of mental organization. This is because if specific cognitive abilities are linked with certain cortical regions across the species, this organization may have developed along a trajectory dictated by a person’s genetic endowment, the same way humans develop arms but not wings or fins. My point is just that evidence for neural flexibility is not necessarily evidence against innate mental faculties. (c) Innate means 100% uniform across the species: nearly all humans are born with properly functioning hearts, livers, brains, and limbs. But some individuals are born with defects, to greater or lesser degrees, in some aspect of their physiology. Some defects are evident at (or before) birth, while others emerge only later. Some defects may be brought on by known environmental hazards while others do not have a clear environmental cause. But variation of this sort does not preclude a biological basis for certain physiological (including cognitive) structures. Rather, certain structures develop in a particular way in the normal course of events. In fact, we do see some variation across individuals in the timecourse of their language development. Not only is there a wide range of what is considered “normal,” but we also see deviations from “normal” language development that cannot be tied to defects in the linguistic input. For example, there are sibling pairs in which one child develops language exactly according to the “textbook” trajectory while the other sibling exhibits language

Conclusions and origins

285

delays or deficits, both children having been reared in virtually the same environment with respect to language input. Even if there is some environmental cause for the language delay/disorder, meaning, for example, that exposure to toxins in utero or after birth caused the disruption of some brain mechanism that subserves language, this is still an environmental disruption of a biological structure. I see this kind of environmental influence as being fundamentally different from the domain-specific environmental influence of hearing language. To my mind, these kinds of deviances that cannot be linked to any lack of linguistic input are one of the strongest arguments for language having a biological basis. (d) Innate means “we can find a gene for it”: although the argument that something is “innate” unquestionably comes down to an assertion that the source of this thing (ability, knowledge, etc.) is to be found in our DNA and not in our experience of the world, there is a large chasm between that assertion and the fantastical assumption that we could identify a single gene or gene sequence that is “responsible” for this very ability or bit of knowledge. While certain areas of certain chromosomes have been linked to particular disorders that involve language (e.g. chromosome 7), most linguists are quick to caution that this does not mean we have found a “language gene.” Just as we do not have to be reductionists in order to talk about mental structure, we do not have to reduce universal tendencies to specific genes in order to make interesting and useful claims about these tendencies. My own (admittedly superficial) understanding of genetics is that the human genome is sufficiently complex that we would expect an alsocomplex ability like human language to have its roots in a number of different parts of our genetic sequence. The qualification about the “normal course of events” and the allowance for some flexibility and variation are important. Building these assumptions into our definition of innateness is essential to ensure a realistic application to humans who are, undoubtedly, varied. But we are varied in ways not necessarily tied to our experiences (such as the example of divergent linguistic development in sibling pairs); and we are varied within certain parameters. That is, humans’ high degree of similarity in matters physiological, linguistic, and otherwise cognitive, is incredibly compelling and is what has led many philosophers over the centuries to posit innate knowledge or biases or constraints on our growth and development.

286

Conclusions and origins

Even if we take away (or weaken) the features of inevitability, inflexibility, and complete uniformity, I maintain that we still have a meaningful notion of innateness. There is agreement among nativists and non-nativists alike that something constrains our learning about the world – without any constraints nothing could be learned – but I would push for the stronger view that learners are constrained by expectations specific to language. As we discussed in Chapter 6, distributional information about subject animacy can help to distinguish displacing from non-displacing predicates, but in order for this distinction to be meaningful in the context of the larger grammar, the different types of predicates must be understood in terms of their grammatical properties. What makes something a displacing predicate is not that it can occur with inanimate subjects, but that it has no external argument. It is this deeper property, signalled by the occurrence of inanimate subjects, that the learner must figure out. If the property of not selecting an external argument is meaningful for the learner (i.e. it has grammatical significance), then once predicates of this sort are identified, the other properties of these predicates (splitting idioms and so forth) will follow. We will spell these ideas out in more detail in Section 7.2. First let us consider different ways of thinking about knowledge of the animate–inanimate distinction as a type of innate knowledge. Here the innate constraint may simply be an expectation to encounter beings like us (humans) and entities not like us (everything else), and finer-grained distinctions of humans vs. animals, and among higher and lower animals, can then be made on the basis of experience. Or, it might contain a richer set of expectations, a view which is explored in the next section.

7.1

Origins of knowledge of the animacy distinction

In Section 5.1 I cited evidence that babies have intuitions very early in life about animacy: their attention is drawn to objects that have animate features like faces, they expect humans but not geometric shapes to move on their own, they treat certain animals as being of the same “kind” as other similar animals, but not of the same kind as physically similar inanimate objects. I did not, in that section, make any claims about this knowledge being innate. Babies interact with humans from the time they are born (legends of children raised by wolves notwithstanding, babies who did not have human contact very early in life would not survive), so how can we tease apart knowledge that is innate from knowledge that is gained through very early experiences?

Origins of knowledge of the animacy distinction

287

We may not be able to distinguish once and for all between the possibility that children are truly hard-wired to expect an animate–inanimate distinction in the world and the possibility that they develop this expectation within the very earliest moments of experience out of the womb, but I want to suggest a few sources of evidence in favor of innateness.1 One source is the fact that a distinction of this sort is found in non-human species as well. One type of evidence for this comes from the ability to represent other individuals’ agency, an ability found in non-human primates (Hare et al. (2000); Flombaum and Santos (2005) and other work cited in Carey (2009)), and other types of abilities linked to social cognition, such as gaze following in dogs and food caching in birds (see Fitch (2010, pp. 159–160)). These abilities presuppose an ability to distinguish conspecifics, which itself presupposes an even more basic ability to distinguish animate from inanimate entities. If these abilities are found in non-human animals as well as humans, then the implication is that the ability to distinguish animate from inanimate (or, perhaps, the mechanism that underlies the learning of this distinction) evolved before humans split off from our common ancestor (which, in the case of birds, would have been a very long time ago indeed), and so it is encoded in the genetic material we share with these other species. A different kind of argument for a fundamental cognitive ability to distinguish animate from inanimate things could be taken from the literature on cognitive organization: there is strong evidence that objects and faces, for example, are recognized via distinct neural mechanisms (Farah, 1995). Although this tack does not translate directly into claims about innateness, there is at least an argument for (certain properties of) animate and inanimate things being represented distinctly at a very fundamental cognitive level and most likely developmentally prior to language being represented. At least, object and face recognition abilities develop earlier than language production and comprehension.

1 Even children with very different sources of information about the world, for example blind

children, will have experiences with both animate and inanimate entities and may distinguish them on the basis of sound or textural features as opposed to visual ones. Nevertheless, evidence that blind children seek divisions between animate and inanimate in largely the same way as sighted children could support the view that they are guided by the same innate constraints and expectations. Certainly, blind children come to develop normal social relationships with other humans (and normal language), but I do not know of studies of blind vs. sighted children’s early concepts (or perceptions) of animacy. Landau and Gleitman (1985) report normal conceptual development, as well as normal language development, in their study of blind children, but they do not comment specifically on the concept of animacy.

288

Conclusions and origins

Another angle on this question that I find useful is exploration of crosscultural similarities in people’s biological taxonomies. The idea is that if people living in very different environments, with different social structures and encountering very different kinds of animals (and different degrees of exposure to animals) make similar groupings and hierarchical rankings across the biota, this uniformity of conceptualization, paired with evidence of extremely early knowledge in infants, would make plausible the idea that (some aspect of) this knowledge is innate. Anthropological studies of “folk biology” or “folk taxonomy” have indicated a quite universal system according to which people around the world group animals and plants according to kind (Conklin, 1962; Bulmer, 1968; Berlin, 1973, 1992; Atran, 1990; Diamond and Bishop, 1999). These studies have shown strong consistency in the level which people take to be “basic” for animals and plants (e.g. cat and tree but not mammal or tabby) and people’s judgments about taxonomies of animals and plants (i.e. which animals and plants are likely to be more closely related). For instance, animals and plants are universally recognized as different at a fundamental level (the rank of “kingdom”), with nearly all living organisms being grouped into one of these kingdoms. Another pervasive trait of folk taxonomy is that humans are generally conceived of as having a separate ontological status from non-human animals. In addition, taxonomies within the kingdoms of animals and plants are discerned via kind, morphological similarity, or “biological essence” rather than, for example, dangerous vs. non-dangerous or edible vs. inedible things, though these are other logically possible (and evolutionarily justifiable) ways of grouping living things. Scott Atran notes that children and adults all over the world categorize living things in terms of their essential nature but nonliving things in terms of functional or physical properties. An example of the difference is that a three-legged dog is judged to be lacking a leg because it is four-legged by nature, while a three-legged chair – a stool – is not “lacking” its leg; having four legs is merely a typical physical property of chairs, not part of the chair’s essential nature. Not only is people’s categorization of their biological world strikingly uniform, but they reason about animals’ shared properties and likely relatedness in highly similar ways. For example, if told that the house sparrow shares with the russet sparrow the property of being susceptible to some disease, people may expect to find this property shared with other types of sparrows or perhaps related birds, such as finches, but not necessarily with spiders or whales. On the other hand, if told that an earthworm shares some property with a house sparrow, people expect this property to be shared by all (or most) animals, since

Origins of knowledge of linguistic animacy and displacing predicates

289

earthworms and house sparrows are only distantly related, and they are about as closely related to each other as they are to other living things like spiders and whales (Atran, 1998). The similarities in how people reason about animals’ taxonomic relations transcend their amount of “scientific” education and direct experience with animals. For instance, American college students have more formal schooling but less direct experience with animals than indigenous Itzaj Maya, while the Itzaj Maya have less schooling but more direct experience with animals. In one of Atran’s studies both of these groups of subjects judged bats to be more closely related to (perceptually similar) rats than to (perceptually dissimilar) cats, even though (a) bats (chiropterans) are not actually more closely related to rats (rodents) than to cats (felines), and (b) American college students judged bats to be mammals and Itzaj Maya judged bats to be a type of bird. Thus, even though cross-cultural differences do emerge in the details of taxonomic groupings, the basis upon which people judge genealogic relationships and kind adherence appears to be universal. Moreover, Atran (1998) comments on the stability of folk biological taxa across not only contemporaneous cultures but cultures through history (cf. Aristotelian divisions of animal “kinds”). Such global similarities, along with considerations of the plausibility of an innate folk biology being naturally selected for, led Atran to propose that the culturally influenced learning of identification of the “basic level” categories of animals (which presupposes an ontological animate–inanimate distinction) is “strongly motivated by cross-culturally shared cognitive mechanisms that do not depend primarily on experience” (Atran, 1998, p. 554). These shared cognitive mechanisms are, he argues, part of an innate core domain of human cognition.

7.2

Origins of knowledge of linguistic animacy and displacing predicates

Knowledge of the animate–inanimate distinction is not specific to language, and so even if this knowledge (or the capacity to make this distinction) is innate, it would constitute domain-general innate knowledge. There is a separate issue of how such knowledge relates to linguistic encodings of animacy and whether there is innate knowledge of animacy that is specific to language. More specifically, the question is whether there is innate knowledge of how animacy maps onto, and can serve as a cue for, grammatical relationships. I take this question to be the more contentious question in the field currently.

290

Conclusions and origins

In Chapter 3 we saw abundant examples of how the Animacy Hierarchy manifests in grammar through morphological marking, word order preferences and selectional restrictions. The presence of animacy distinctions is so pervasive in human language that only a fraction of the attested examples of linguistic marking of animacy could be included in that discussion. Does the pervasiveness of animacy marking in language mean that knowledge of the animate–inanimate distinction constitutes innate linguistic knowledge? Let’s think about what this would mean. The primary pattern we saw in Chapter 3 was that non-derived subjects are preferentially high on the Animacy Hierarchy, and that derived subjects can be lower on this hierarchy than nonderived subjects. The tendency is for canonical, non-derived subjects (external arguments) to be animate. So we could pose the question in terms of whether children innately expect subjects to be animate, and this would be a type of language-specific innate knowledge (this expectation is implicated in Pinker’s Semantic Bootstrapping (Pinker, 1984); see Section 5.2.1). But to even pose this question presupposes that children have a notion of subject a priori. One problem with supposing that the notion of “subject” is innate is that it has been argued that there are languages that do not have grammatical subjects, such as Lisu (Lolo-Burmese; Li and Thompson, 1976). Rather, these languages have a “topic–comment” sentence structure, and the topic is not necessarily linked to the agent or experiencer role. For example, in (1a – b) the topic may denote either the agent (biter) or patient (bite-ee), and the intended meaning is influenced by discourse or assumptions about the world. Thus, the most likely meaning of (1a) is “As for people, dogs bite them” and the most likely meaning for (1b) is “As for dogs, they bite people,” since dogs are more likely to bite people than people are to bite dogs. But both meanings are possible for both sentences (this datum is from Li and Thompson (1976, p. 472); these authors cite Hope (1974)). (1)

a. làthyu nya ánà kh`u”-a” people top dog bite-decl “As for people, they bite dogs/dogs bite them.” b. ánà nya làthyu kh`u”-a” dog top people bite-decl “As for dogs, they bite people/people bite them.”

Most languages that are considered “topic-prominent,” including Lisu, do allow subjects, even if they might be rare. Moreover, when we find such subjects they adhere to hierarchies and exhibit properties we have seen associated with subjects repeatedly. For example in Lisu when there is no candidate for

Origins of knowledge of linguistic animacy and displacing predicates

291

topic in a given sentence, the NP that is highest on the thematic hierarchy is marked with the topic marker nya and behaves, semantically, like a canonical subject. Schachter (1976) likewise provides some evidence that the category of topic in Tagalog and other Philippine languages has many of the properties prototypically associated with subjects: for example, the “topic” can be the argument of an intransitive predicate or an adjectival or nominal (“equative”) predicate, and it is the only NP in the sentence that can be relativized (in accordance with Keenan and Comrie’s (1972) Accessibility Hierarchy if only one position can be relativized, it is the subject).2 A true topic is not selected by any predicate the way a subject is; it is linked to a discourse topic and fails to exhibit (many of) the features prototypically linked with subjects, such as control of the reference of PRO or reflexives (see Keenan’s hierarchy of subject properties in Section 3.4). Since the topic is not selected by a predicate, it is considerably more flexible in what it can denote – unlike subjects, topics do not have an inherent preference for denoting agents or animate entities (Li and Thompson, 1976). But interestingly, even in topic-prominent languages the topic is preferentially associated with the agent role (and therefore with animate referents), and is more likely human than non-human (Givón, 1976). In other words, there is no necessary association between topic and animacy/agent, but in actual practice this association is widely attested. By some accounts, one of the characterizations of topic (as distinguished from comment) is identification with the persepctive of the speaker. That is, topics tend to take the speaker’s perspective, and since speakers are humans, topics will also tend to be prototypically human or agentive (certainly, this is not the only criterion for identifying topics; see discussion in Li and Thompson (1976); Bates and MacWhinney (1982)). The widespread preference for associating topics with animate entities or agents points again to animacy being a highly universal feature that can be linked to the “highest” or “most prominent” NP, whether it is syntactically a subject or a topic.3 Languages vary in terms of whether they have subjects vs. topics, and languages that have subjects exhibit variation in the morphosyntactic properties of their subjects (e.g. case marking, agreement patterns, restrictions on which 2 I should note, though, that Schachter concludes that Tagalog topics are not unambiguously

subjects. However, see also Schwartz (1976) for arguments that another Philippine language, Ilocano, has subjects. 3 There may yet be a stronger association between humanness or animacy and subjecthood than with topichood. Duranti and Ochs (1979) point out that left-dislocated direct objects in Italian discourse bear many of the traits of topics but denote humans at a much lower rate (26.8%) than true subjects (100%).

292

Conclusions and origins

semantic roles can be associated with the position, and so forth). These facts suggest that the category of “subject” itself is unlikely to be entirely innate. We do not suppose, for example, that children are born expecting to find Nominative case marking, or to associate morphological marking with a particular structural position. But the more fundamental pattern we see is the robust association between animacy (or agency) and the most “prominent” NP in a sentence. It seems to me that an innate expectation to associate animate NPs with this structurally prominent category, with high probability at any rate, would be supremely helpful for language learners. It is in this sense that I want to suggest that children expect subjects to be animate (I believe this is also what Pinker had in mind). The learner will discover through language-specific experience whether their language has, for example, case marking on subjects, or whether their language has topics instead of subjects. A prior expectation for this “prominent” category to be associated with the feature of animacy is likely to be the first step in constructing the appropriate grammatical category. Asking the further question of why children might be predisposed to encode animacy in language and to associate animate NPs with a prominent structural position (subject or topic), I can imagine an answer that lies somewhere in considerations of language evolution. The ability to distinguish animate from inanimate entities is present in all humans (and, undoubtedly, even our pre-human ancestors), and distinctions of animacy (e.g. human vs. animal, or distinctions between species of animals) were important to early hominids for survival. When human language evolved people used it to talk about things that were important to them, such as humans and animals. Not only does vocabulary tend to be retained in direct proportion to its significance for speakers (Berlin et al., 1973), but highly frequent forms tend to become grammaticalized (Bybee and Hopper, 2001). Thus, animacy was something that was important to the earliest hominids who had language, and over time animacy itself evolved to take on a more formal role in language, ultimately becoming associated with the grammatical role of subject. Why animacy should be prototypically linked to the subject position may also have, at bottom, a functional explanation: humans and other animates are more likely to act upon their environment than inanimate entities are, and as grammar evolved to specify the subject, prototypically, as the actor or agent, the feature of animacy and the role of subjecthood became associated with one another. Certainly, arguments can be made for a learning mechanism that “learns” this connection on the basis of repeated input patterns (this is the assumption of constructivist approaches to language acquisition; Tomasello (2003); Ambridge and Lieven (2011)). However, in my view there is a trade-off

Origins of knowledge of linguistic animacy and displacing predicates

293

in efficiency: the disadvantage of not assuming innate constraints that could be universally helpful is that the learner is unnecessarily burdened with looking for relationships and patterns in the input. That is, if children do not innately expect the subject to be associated with greater animacy, then even if they knew animacy was represented somehow in language they would need to discard other possible ties between animacy and language (e.g. the object is preferentially animate, or the subject is preferentially inanimate, or animate NPs are preferentially oblique, and so forth). The cognitive resources required for considering and discarding all of these possibilities would be more efficiently used for learning things that could not reasonably be assumed innately, such as lexical meanings and language-specific rules or constraints, and by many accounts children are extremely efficient learners. In turn, efficiency in learning can be seen as an evolutionary advantage. In evolutionary biology the idea that some ability that was once learned through a general learning mechanism (and/or through social means) becomes innate via natural selection is known as the Baldwin Effect. Members of the species that do not have to learn this ability, since it is innate, will be able to acquire it more quickly and efficiently, and as long as the ability is advantageous for survival these individuals will be favored. If the expectation that subjects (or “the most prominent NP,” as per the above discussion) are prototypically animate gives children a leg up in acquiring language, in that it significantly constrains the hypothesis space, then learners who have this expectation innately will be more efficient language learners. We can imagine that early hominids who were good at language were better able to strategize with their fellow clanspeople, convey and understand information about the location of food, and resolve differences peacefully, and therefore more likely to survive and reproduce. The other piece of knowledge I have argued is central to the acquisition of opaque constructions is that language may contain displacing predicates. Clearly, this type of knowledge is specific to the domain of language. Let’s consider what it would mean for displacing predicates to be innate. One possible entailment is that displacing predicates constitute some kind of grammatical category, similar to the category of Noun or Verb, and this category is part of the innate linguistic apparatus. I think it is unlikely that such a specific type of predicate would be rightly considered to form a category on this level, since grammatical components that are plausible candidates for such a universal category are so much more general (e.g. Noun; and even this notion is disputed – see the constructivist literature cited above).

294

Conclusions and origins

It is also worth observing that apart from failing to select an external argument, these predicates are a fairly diverse group: they belong to different lexical categories (verbs and adjectives), they take different kinds of complements (raising verbs and tough-adjectives take a clausal complement, while unaccusatives take an NP complement), and raising verbs and tough-adjectives readily take an expletive subject, while unaccusatives are more constrained in this respect. In addition, recall that these predicates’ frequency of occurrence with inanimate subjects is widely variable, ranging from extremely rare (raising verbs) to almost categorical (tough-adjectives). So displacing predicates are highly specific compared to the kinds of general categories many people consider plausibly innate, and they have diverse syntactic and semantic properties, making it unlikely that they form a true natural class. So what are displacing predicates then? One way we might think about them in a cohesive way is as a hybridization of auxiliaries and main predicates. Pinker notes that auxiliaries “have the semantic property of designating sentence modality, tense, aspect, epistemic status, and modal concepts such as necessity, possibility, impossibility, and the like” and are a closed class (Pinker, 1984, p. 46). Displacing predicates have precisely these semantic properties and are also something of a closed class (note the relatively small set of raising verbs, and the extremely small set of true tough-adjectives). Indeed, some linguists have categorized the displacing predicates seem, going to, used to, and so on as “marginal modals,” “semi-auxiliaries,” or “emerging modals” (see Quirk et al. (1985); Krug (2001)), suggesting that these forms derived diachronically from main (non-displacing) predicates but became displacing predicates through the process of grammaticalization (note that these verbs do tend to be extremely frequent, and high frequency is associated with grammaticalized elements). On the other hand, it is important to recall that these predicates differ from auxiliaries in important morphosyntactic features: most of them are main predicates (see Section 5.3.1; though, admittedly, some of them may be losing this status as they undergo phonological reduction, e.g. gonna, usedta). So they could be characterized as main predicates that have the semantic features of auxiliaries, but it is not clear to me that such a characterization is sufficient to warrant designating them as an independent linguistic category. Moreover, these semantic features apply to raising verbs and tough-adjectives, but not necessarily to unaccusative verbs. Given that there does not appear to be evidence supporting a general category of displacing predicates, what does it mean to say that knowledge of these predicates is innate? One way of answering this question comes from the view that displacing predicates are, in some sense, the quintessence of

Origins of knowledge of linguistic animacy and displacing predicates

295

what makes human language more than simply strings of words. Chomsky’s argument against Bloomfieldian structuralism was based on the rejection of the notion that the grammar of a language could be extracted given data that consisted only of surface strings, an argument that is illustrated by “constructional homonyms” such as John is easy to please and John is eager to please (Lees, 1960). While my proposal for how these constructional homonyms are distinguished may appear to offer an input-driven solution to this basic conundrum, I submit that the cue of an inanimate subject would not do the work that it does if the learner did not expect that there can be non-local dependencies in syntax. With this expectation, plus the expectation that external arguments will be animate, an inanimate subject can then cue a non-local dependency, which is really what displacing predicates are about. Without this expectation, an inanimate subject could trigger (a) an ergative representation of syntax, (b) coercion of the inanimate subject to have animate properties (e.g. agency), or (c) determination that inanimate subjects are external theme arguments of predicates like seem and easy, among other possibilities. If we accept that children end up with the correct representation, the one where the subject is not the external argument of the main predicate and instead is displaced, then the most likely route to this conclusion is the expectation that such structures can exist in grammar. It is in this sense, then, that I claim that displacing predicates are a type of innate knowledge. An advantage of building in the expectation that language can have displacing predicates is that we can account for related phenomena exhibited by these predicates. Even though within each pair of predicates (raising and control verbs, etc.) the displacing predicate is statistically distinguishable from the non-displacing one by its higher rate of occurrence with inanimate subjects, it is not clear that the other crucial properties of each of these displacing predicates would necessarily follow from the distribution of displacing predicates with inanimate subjects. That is, simply distinguishing one predicate from another on the basis of its rate of occurrence with inanimate subjects does not lead to, for example, expecting one predicate to occur with expletive subjects, or to allow quantifier-lowering interpretations, or, in the case of tough-adjectives, an object-gap as opposed to a subject-gap reading (as we saw with the predicate ready). In order for these other properties to fall out from the observation that a given predicate occurs with inanimate subjects at a statistically higher rate than a superficially similar predicate, there must be some expectation about connections and correlations between a predicate’s ability to select an external argument and its ability to have these other properties. Thus, we can think of inanimate subjects as serving as a cue for a kind of

296

Conclusions and origins

macro-parameter: learners need to know that the parameter is part of language (predicates can select an external argument or not), and also that the value of this parameter relates to other properties of sentence structure, namely, those properties associated with NP-movement. To summarize, what I mean when I say that some aspect of language is innate is that children will expect to find this aspect or pattern in the language they are acquiring. Thus, children expect that in the normal case, the subject of a sentence will be animate and will be the external argument of the main predicate, and that a sentence with an inanimate subject must be represented differently from the normal case. Similarly, children will expect predicates to apply to arguments, and they expect language to allow for long-distance dependencies. The confluence of these expectations allows an inference to be made when an inanimate subject of a main verb is encountered: it is likely to be indicative of a long-distance dependency. 7.3

Further questions

This conclusion is about beginnings in another sense, as well. Research projects are like intellectual hydras, so that answering one question inevitably results in a dozen new ones springing up, begging to be investigated. The most pressing open questions, to my mind, concern the need for further experimental work on children’s inferences about novel predicates under different exposure conditions: with animate vs. inanimate subjects, but also with varying numbers of exemplars, with different types of embedded predicates, and with or without situational context cues which could potentially provide crucial semantic information. Moreover, these experiments, out of convenience, have been conducted only with English-speaking children so far. What inferences are drawn by children acquiring other languages with raising and tough predicates, like Tongan or Niuean? What inferences are drawn by children who speak a language that disallows inanimate transitive subjects, like Jacaltec, or languages that have topics instead of subjects, like Lisu? Are there languages that lack displacing predicates, and if so, how does subject animacy influence the inferences of children acquiring these languages? The question of what role situational context plays in the acquisition of these predicates is an important one. Certainly, cues from the non-linguistic context (whether observed visually or not) are necessary for filling in the full lexical entry for any predicate. While the syntactic frame in which a predicate occurs helps to narrow down the possible meanings of that predicate (for example, by excluding certain impossible or unlikely meanings) it does not unequivocally

Further questions

297

tell the learner its precise meaning. No one has claimed that it does. But how do children extract particular lexical meanings from experienced events in the world? Even if a child can use occurrence with an inanimate subject to figure out that seem or easy is a displacing predicate, how on earth does the child reach the further step of understanding just what these predicates actually mean? As yet, we do not have a good understanding of this process. The methodology I have used in my most recent studies with children (measuring reaction time in answering questions) is also new and needs to be tested more widely. Is it a reliable measure of all types of ungrammaticality, or only of argument structure violations? Can it be used to test other sorts of implicit judgments? In my view, if we are going to test children under the age of 6 or 7 on their categorizations of novel predicates we must develop reliable means of tapping into subconscious grammaticality judgments. Currently, experiments are being planned that would use measures of electrical responses in the brain (ERP) to measure children’s grammaticality judgments with these types of constructions (in collaboration with Masako Hirotani). A further open question is how children even determine that a biclausal sentence with an embedded infinitive involves a complement clause as opposed to adjunction. Although I have talked about these opaque sentences as if only two options were available for the underlying structure (e.g. raising or control), in fact multiple different structures are possible. A sentence like John gorps to stay in shape could have a raising structure (gorp means ‘seem’), a control structure (gorp means ‘try’), or it could have an intransitive verb with an adjoined purpose clause (gorp means ‘run’). This level of the learning problem, quite an important one to my mind, has been glossed over here. It may represent a further type of structural opacity that presents difficulties for the basic Syntactic Bootstrapping procedure, i.e. parsing a string in the absence of lexical clues about the predicate’s meaning. And, certainly, the computational models described in Chapter 6 are in the early stages of exploration and development. All of these issues represent avenues to be explored in future work.

Appendix

List of Abbreviations in Glosses 1 2 3 Abs Acc agr agt apass asp caus cj.pvb cl. CL comp compl Dat decl desid det dir dyn Erg ex f FA fut Gen incompl 298

first person second person third person Absolutive case Accusative case agreement (within NP; Chamorro) passive agent (Maori) antipassive aspect marker causative marker conjunctive preverb clitic classifier complementizer completive Dative case declarative marker desiderative determiner directional dynamic Ergative case exhaustive feminine gender focused antipassive (K’iche’) future Genitive case incompletive

Appendix ind inf instr Loc m n neg Nom nonpst obv p Part part pass poss pre pst refl s sub subj suf tam top tr

indicative infinitive instrumental (case or marker) Locative case masculine gender neuter gender negation Nominative case nonpast obviative plural Partitive case participle passive possession marker prefix past reflexive singular subject subjunctive mood suffix unspecified tense-aspect-mood marker topic marker transitive

299

Bibliography

Aissen, Judith. 2003. Differential object marking: Iconicity vs. economy. Natural Language and Linguistic Theory 21:435–483. Aissen, Judith, and David Perlmutter. 1983. Clause reduction in Spanish. In Studies in relational grammar 1, ed. David M. Perlmutter, 360–403. University of Chicago Press. Akhtar, Nameera. 2005. The robustness of learning through overhearing. Developmental Science 8:199–209. Alishahi, Afra, and Suzanne Stevenson. 2005a. The acquisition and use of argument structure constructions: A Bayesian model. In Proceedings of the ACL 2005 Workshop on Psychocomputational Models of Human Language Acquisition. 2005b. A probabilistic model of early argument structure acquisition. In Proceedings of the 27th Annual Meeting of the Cognitive Science Society. 2008. A computational model of early argument structure acquisition. Cognitive Science 32:789–834. Allen, Barbara J., and Donald G. Frantz. 1983. Advancements and verb agreement in Southern Tiwa. In Studies in relational grammar, ed. David M. Perlmutter, 303–314. University of Chicago Press. Ambridge, Ben, and Elena V.M. Lieven. 2011. Child language acquisition: Contrasting theoretical approaches. Cambridge University Press. Anderson, Deborah. 2005. The acquisition of tough movement in English. Doctoral Dissertation, University of Cambridge. Anderson, Stephen R. 1976. On the notion of subject in ergative languages. In Subject and topic, ed. Charles Li, 1–24. New York: Academic Press. Anderson, Stephen R., and Sandra Chung. 1977. On grammatical relations and clause structure in verb-initial languages. In Syntax and semantics 8: Grammatical relations, ed. Peter Cole and Jerrold M. Sadock, 1–26. New York: Academic Press. Armbuster, Tom. 1981. How to influence responses to “Is Kathy easy to see?”. Journal of Psycholinguistic Research 10:27–39. Asudeh, Ash. 2002. Richard III. In Papers from the 38th Regional Meeting of the Chicago Linguistic Society, ed. Mary Andronis, Erin Debenport, Anne Pycha, and Keiko Yoshimura, 31–46. Chicago Linguistic Society. 2005. Control and semantic resource sensitivity. Journal of Linguistics 41:465–511. Asudeh, Ash, and Ida Toivonen. 2012. Copy raising and perception. Natural Language and Linguistic Theory 30:321–380.

300

Bibliography

301

Atran, Scott. 1990. Cognitive foundations of natural history: Towards an anthropology of science. Cambridge University Press. 1998. Folk biology and the anthropology of science. Behavioral and Brain Sciences 21:597–611. Babyonyshev, Maria, Jennifer Ganger, David Pesetsky, and Kenneth Wexler. 2001. The maturation of grammatical principles: Evidence from Russian unaccusatives. Linguistic Inquiry 32:1–44. Baker, Mark. 1988. Incorporation: A theory of grammatical function changing. University of Chicago Press. Baker, Mark, Kyle Johnson, and Ian Roberts. 1989. Passive arguments raised. Linguistic Inquiry 20:219–251. Bates, Elizabeth, and Brian MacWhinney. 1982. Functionalist approaches to grammar. In Language acquisition: The state of the art, 173–218. Cambridge University Press. Becker, Misha. 2005. Learning verbs without arguments: The problem of raising verbs. Journal of Psycholinguistic Research 34:165–191. 2006a. Acquisition of A-movement. In Proceedings of the inaugural conference on Generative Approaches to Language Acquisition–North America, ed. K.U. Deen, J. Nomura, B. Schulz, and B.D. Schwartz, volume 4, 65–75. Storrs, CT: University of Connecticut Occasional Papers in Linguistics. 2006b. There began to be a learnability puzzle. Linguistic Inquiry 37:441–456. 2007. Animacy, expletives, and the learning of the raising–control distinction. In Proceedings of the 2nd conference on Generative Approaches to Language Acquisition–North America, ed. Alyona Belikova, Luisa Meroni, and Mari Umeda, 12–20. Somerville, MA: Cascadilla Proceedings Project. 2009. The role of animacy and expletives in verb learning. Language Acquisition 16: 283–296. Becker, Misha. To appear Animacy and the acquisition of tough-adjectives. Language Acquisition. Becker, Misha, and Bruno Estigarribia. 2011. Drawing inferences about novel raising and control verbs. In Selected Proceedings of the 4th Conference on Generative Approaches to Language Acquisition–North America, ed. Michaela Pirvulescu, Maria Cristina Cuervo, Ana T. Perez-Leroux, Jeffrey Steele, and Nelleke Strik. Somerville, MA: Cascadilla Proceedings Project. 2013. Harder words: How children learn abstract verbs with opaque syntax. Language Learning and Development 9: 211–244. Becker, Misha, Bruno Estigarribia, and Duna Gylfadottír. 2012. Tough-adjectives are easy to learn. In Proceedings of BUCLD 36. URL: Cascadilla Proceedings Project. Becker, Misha, and Jeannette Schaeffer. 2013. Animacy, argument structure and unaccusatives. In Generative linguistics and acquisition: Studies in honor of Nina M. Hyams, ed. Misha Becker, John Grinstead, and Jason Rothman, 13–33. Amsterdam: John Benjamins. Belletti, Adriana, and Luigi Rizzi. 1988. Psych-verbs and θ-theory. Natural Language and Linguistic Theory 6:291–352. Berlin, Brent. 1973. The relation of folk systematic to biological classification and nomenclature. Annual Review of Systematic and Ecology 4:259–271.

302

Bibliography

1992. Ethnobiological classification: Principles of categorization of plants and animals in traditional societies. Princeton University Press. Berlin, Brent, Dennis E. Breedlove, Robert M. Laughlin, and Peter H. Raven. 1973. Cultural significance and lexical retention in Tzeltal–Tzotzil ethnobotany. In Meaning in Mayan languages, ed. Munro S. Edmonson, 143–164. The Hague: Mouton. den Besten, Hans. 1977. On the interaction of root transformations and lexical deletive rules. University of Amsterdam. Bever, Thomas G. 1970. The cognitive basis for linguistic structures. In Cognition and the development of language, ed. John R. Hayes, 279–362. New York: John Wiley & Sons, Inc. Biggs, Bruce. 1969. Let’s learn Maori. Wellington, NZ: A.H. and A.W. Reed. Blake, Barry. 1977. Case marking in Australian languages. Linguistic Series 23. Canberra: Australian Institute of Aboriginal Studies. Blake, Frank R. 1930. A semantic analysis of case. In Curme volume of linguistic studies, ed. James T. Hatfield, Werner Leopold, and A.J. Friedrich Zieglschmid, 34–49. Baltimore: Linguistic Society of America. Bloom, Lois. 1970. Language development: Form and function in emerging grammars. Cambridge, MA: MIT Press. Bloom, Lois, Patsy Lightbown, and Lois Hood. 1975. Structure and variation in child language. Monographs of the Society for Research in Child Development . Bloom, Lois, J. Tackeff, and Margaret Lahey. 1984. Learning to in complement construction. In Language development from two to three, ed. Lois Bloom. Cambridge University Press. Blum-Kulka, S., and Catherine Snow. 2002. Talking to adults: The contribution of multiparty discourse to language acquisition. Mahwah, NJ: Lawrence Erlbaum Associates. Boeckx, Cédric. 1998. A minimalist view on the passive. Doctoral Dissertation, University of Connecticut. Bolinger, Dwight. 1961. Syntactic blends and other matters. Language 37:366–381. Borer, Hagit, and Kenneth Wexler. 1987. The maturation of syntax. In Parameter setting, ed. Thomas Roeper and Edwin Williams, 123–172. Dordrecht: Reidel. 1992. Bi-unique relations and the maturation of grammatical principles. Natural Language and Linguistic Theory 10:147–189. Bowerman, Melissa M. 1973. Early syntactic development: A cross-linguistic study with special reference to Finnish. Cambridge University Press. 1976. Semantic factors in the acquisition of rules for word use and sentence construction. In Normal and deficient child language, ed. D.M. Morehead and A.E. Morehead, 99–197. Baltimore, MD: University Park Press. Bresnan, Joan. 1971. Sentence stress and syntactic transformations. Language 47:257– 281. 1982a. Control and complementation. In The mental representation of grammatical relations, ed. Joan Bresnan, 282–390. Cambridge, MA: MIT Press. 1982b. The mental representation of grammatical relations. Cambridge, MA: MIT Press.

Bibliography

303

1982c. The passive in lexical theory. In The mental representation of grammatical relations, ed. Joan Bresnan, 3–86. Cambridge, MA: MIT Press. 2001. Lexical-functional syntax. Malden, MA: Blackwell. Bresnan, Joan, Jean Carletta, Richard Crouch, Malvina Nissim, Mark Steedman, Tom Wasow, and Annie Zaenen. 2002. Paraphrase analysis for improved generation, link project. Stanford, CA: HRCR Edinburgh-CSLI Stanford. Bresnan, Joan, Anna Cueni, Tatiana Nikitina, and R. Harald Baayen. 2007. Predicting the dative alternation. In Cognitive foundations of interpretation, ed. G. Bouma, I. Kraemer, and J. Zwarts, 69–94. Amsterdam: Royal Netherlands Academy of Science. Bresnan, Joan, and Jennifer Hay. 2008. Gradient grammar: An effect of animacy on the syntax of give in New Zealand and American English. Lingua 118:245–259. Brody, Michael. 1993. θ -theory and arguments. Linguistic Inquiry 24:1–23. Brown, Roger. 1973. A first language. Cambridge, MA: Harvard University Press. Bulmer, Ralph N.H. 1968. Worms that croak and other mysteries of Karam natural history. Mankind 6:621–639. Bunger, Anne, and Jeffrey Lidz. 2004. Syntactic bootstrapping and the internal structure of causative events. In Proceedings of the 28th Boston University Conference on Language Development, ed. Alejna Brugos, Linnea Micciulla, and Christine E. Smith, 74–85. Somerville, MA: Cascadilla Press. 2008. Thematic relations as a cue to verb class: 2-year-olds distinguish unaccusatives from unergatives. In Proceedings of the 31st Penn Linguistics Colloquium. Philadelphia: Penn Working Papers in Linguistics. Burgess, Curt. 1990. Interaction of semantic, syntactic and visual factors in syntactic ambiguity resolution. Doctoral Dissertation, University of Rochester. Burzio, Luigi. 1986. Italian syntax: A Government-Binding approach. Dordrecht: Reidel. Bybee, Joan, and Paul Hopper. 2001. Frequency and the emergence of linguistic structure. Amsterdam: John Benjamins. Cairns, Helen, Dana McDaniel, J.R. Hsu, and M. Rapp. 1994. A longitudinal study of principles of control and pronominal reference in child English. Language 70:260–288. Cambon, Jacqueline, and Hermine Sinclair. 1974. Relations between syntax and semantics: Are they “easy to see”? British Journal of Psychology 65:133–140. Carey, Susan. 1985. Conceptual change in childhood. Cambridge, MA: MIT Press. 2009. The origin of concepts. New York: Oxford University Press. Carpenter, Bob. 1997. Type-logical semantics. Cambridge, MA: MIT Press. Chen, Mel. 2012. Animacies: Biopolitics, racial mattering, and queer affect. Durham: Duke University Press. Childers, Jane B., and Catharine H. Echols. 2004. 2 1/2-year-old children use animacy and syntax to learn a new noun. Infancy 5:109–125. Chomsky, Carol. 1969. The acquisition of syntax in children from 5 to 10. Research Monograph 57. Cambridge, MA: MIT Press. Chomsky, Noam. 1957. Syntactic structures. The Hague: Mouton. 1959. Review of Skinner Verbal behavior. Language 35:26–58.

304

Bibliography

1964. Current issues in linguistic theory. The Hague: Mouton. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press. 1977. On wh-movement. In Formal syntax, ed. Peter Culicover, Thomas Wasow, and Adrian Akmajian. New York: Academic Press. 1980. On binding. Linguistic Inquiry 11:1–46. 1981. Lectures on government and binding: The Pisa lectures. New York: Mouton de Gruyter. 1995. The minimalist program. Cambridge, MA: MIT Press. 2001. Derivation by phase. In Ken Hale: A life in language, ed. Michael Kenstowicz, 1–52. Cambridge, MA: MIT Press. Chomsky, Noam, and Howard Lasnik. 1977. Filters and control. Linguistic Inquiry 8:425–504. Chung, Sandra. 1976. On the subject of two passives in Indonesian. In Subject and topic, ed. Charles N. Li, 57–98. New York: Academic Press. 1978. Case marking and grammatical relations in Polynesian. Austin, TX: University of Texas Press. 1983. ECP and government in Chamorro. Natural Language and Linguistic Theory 1:207–244. 1998. The design of agreement: Evidence from Chamorro. University of Chicago Press. Cinque, Guglielmo. 1995. Italian syntax and universal grammar. Cambridge University Press. 2000. “Restructuring” and functional structure. Unpublished ms., University of Venice. Clark, Eve. 1987. The principle of contrast: A constraint on language acquisition. In Mechanisms of language acquisition, ed. Brian MacWhinney, 171–188. Hillsdale, NJ: Lawrence Erlbaum Associates. Clark, Herbert H., and Jeffrey S. Begun. 1971. The semantics of sentence subjects. Language and Speech 14:34–46. Clements, Wendy, and Josef Perner. 1994. Implicit understanding of belief. Cognitive Development 9:377–395. Clifton, Charles, and Lynn Frazier. 1989. Comprehending sentences with long-distance dependencies. In Linguistic structure in language processing, ed. Greg Carlson and Michael Tanenhaus. Dordrecht: Kluwer. Clifton, Charles Jr. 1993. Thematic roles in sentence processing. Canadian Journal of Psychology 47:222–246. Cohen Sherman, Janet, and Barbara Lust. 1995. Children are in control. Cognition 45:225–255. Collins, Chris. 2005a. A smuggling approach to raising in English. Linguistic Inquiry 36:289–298. 2005b. A smuggling approach to the passive in English. Syntax 8:81–120. Comrie, Bernard. 1989. Language universals and linguistic typology. University of Chicago Press. Conklin, Harold C. 1962. The lexicographical treatment of folk taxonomies. International Journal of American Linguistics 28:119–141.

Bibliography

305

Corbett, Greville. 1991. Gender. Cambridge University Press. Corrigan, Roberta. 1986. The internal structure of English transitive sentences. Memory and Cognition 14:420–431. 1988. Children’s identification of actors and patients in prototypical and nonprototypical sentence types. Cognitive Development 3:285–297. Couper-Kuhlen, Elizabeth. 1979. The prepositional passive in English: A semanticsyntactic analysis, with a lexicon of prepositional verbs. Tübingen: Max Niemeyer Verlag. Craig, Colette. 1976. Properties of basic and derived subjects in Jacaltec. In Subject and topic, ed. Charles Li, 99–124. New York: Academic Press. 1977. The structure of Jacaltec. Austin, TX: University of Texas Press. Crain, Stephen, and Mineharu Nakayama. 1987. Structure dependence in grammar formation. Language 63:522–543. Crain, Stephen, Rosalind Thornton, and Keiko Murasugi. 1987. Capturing the evasive passive. Paper presented at the Boston University Conference on Language Development; Reprinted in Language Acquisition, 2009, v.16:123–133. Crawford, Jean L. 2012. Developmental perspectives on the acquisition of the passive. Doctoral Dissertation, University of Connecticut. Croft, William. 2003. Typology and universals. Cambridge University Press, 2nd edition. Cromer, Richard. 1970. “Children are nice to understand”: Surface structure clues for the recovery of deep structure. British Journal of Psychology 61:395–408. Cruse, D.A. 1973. Some thoughts on agentivity. Journal of Linguistics 9:11–23. Culicover, Peter, and Ray Jackendoff. 2001. Control is not movement. Linguistic Inquiry 32:493–512. Dahl, Östen. 2000. Egophoricity in discourse and syntax. Functions of Language 7:33–77. 2008. Animacy and egophoricity: Grammar, ontology and phylogeny. Lingua 118:141–150. Dahl, Östen, and Kari Fraurud. 1996. Animacy in grammar and discourse. In Referent and referent accessibility, ed. Thorstein Fretheim and Jeanette K. Gundel, 47–64. Amsterdam: Benjamins. Davies, William D., and Stanley Dubinsky. 2004. The grammar of raising and control. Malden, MA: Blackwell. De Villiers, Jill G. 1980. The process of rule learning in child speech: A new look. In Children’s language, ed. Keith Nelson, volume 2, 1–44. New York: Gardner Press. De Villiers, Peter A., and Jill G. De Villiers. 1972. Early judgments of semantic and syntactic acceptability by children. Journal of Psycholinguistic Research 1:299–310. DeLancey, Scott. 1981. An interpretation of split ergativity and related patterns. Language 57:626–657. Demuth, Katherine. 1989. Maturation and the acquisition of Sesotho passive. Language 65:56–80. Diamond, Jared, and K. David Bishop. 1999. Ethno-ornithology of the Ketengban people, Indonesian New Guinea. In Folkbiology, ed. Douglas Medin and Scott Atran, 17–45. Cambridge, MA: MIT Press.

306

Bibliography

DiSciullo, Anna Maria, and Edwin Williams. 1987. On the definition of word. Cambridge, MA: MIT Press. Dixon, R.M.W. 1979. Ergativity. Language 55:59–138. Dowty, David. 1991. Thematic proto-roles and argument selection. Language 67:547–619. Duranti, Alessandro, and Elinor Ochs. 1979. Left-dislocation in Italian conversation. In Syntax and semantics: Discourse and syntax, ed. Talmy Givón, volume 12, 377–416. New York: Academic Press. Eimas, Peter D., and Paul C. Quinn. 1994. Studies on the formation of perceptually based basic-level categories in young infants. Child Development 65:903–917. Elman, Jeff. 1990. Finding structure in time. Cognitive Science 14:179–211. 1993. Learning and development in neural networks: The importance of starting small. Cognition 48:71–99. Elman, Jeff L., Elizabeth A. Bates, Mark H. Johnson, Annette Karmiloff-Smith, Domenico Parisi, and Kim Plunkett. 1996. Rethinking innateness: A connectionist perspective on development. Cambridge, MA: MIT Press. Evers, A. 1975. The transformational cycle in Dutch and German. Doctoral Dissertation, University of Utrecht. Fantz, Robert L., Joseph F. Fagan III, and Simon B. Miranda. 1975. Early visual selectivity as a function of pattern variables, previous exposure, age from birth and conception, and expected cognitive deficit. In Infant perception: From sensation to cognition, volume 1: Basic visual processes, ed. Leslie B. Cohen and Philip Salapatek, 249–345. New York: Academic Press. Farah, Martha. 1995. Dissociable systems for visual recognition: A cognitive neuropsychology approach. In An invitation to cognitive science, vol. 2: Visual cognition, ed. Stephen M. Kosslyn and Daniel N. Osherson, 101–119. Cambridge, MA: MIT Press. Fenson, Larry. 1984. Developmental trends for action and speech in pretend play. In Symbolic play: The development of social understanding, ed. Inge Bretherton, 249–270. New York: Academic Press. Fernald, Ann. 1992. Human maternal vocalizations to infants as biologically relevant signals: An evolutionary perspective. In The adapted mind: Evolutionary psychology and the generation of culture, ed. Jerome H. Barkow, Leda Cosmides, and John Tooby. Oxford University Press. Ferreira, Fernanda. 1994. Choice of passive voice is affected by verb type and animacy. Journal of Memory and Language 33:715–736. Ferreira, Fernanda, and Charles Clifton. 1986. The independence of syntactic processing. Journal of Memory and Language 25:348–368. Fillmore, Charles. 1968. The case for case. In Universals in linguistic theory, ed. Emmon Bach and Robert T. Harms, 1–88. New York: Holt, Rinehart & Winston. 1977. The case for case reopened. In Syntax and semantics 8: Grammatical relations, ed. Peter Cole and Jerrold Sadock, 59–82. New York: Academic Press. Fisher, Cynthia. 1996. Structural limits on verb mapping: The role of analogy in children’s interpretation of sentences. Cognitive Psychology 31:41–81. 2002. Structural limits on verb mapping: The role of abstract structure in 2.5-yearolds’ interpretations of novel verbs. Developmental Science 5:55–64.

Bibliography

307

Fisher, Cynthia, Henry Gleitman, and Lila R. Gleitman. 1991. On the semantic content of subcategorization frames. Cognitive Psychology 23:331–392. Fitch, Tecumseh. 2010. The evolution of language. Cambridge University Press. Flombaum, Jonathan, and Laurie Santos. 2005. Rhesus monkeys attribute perceptions to others. Current Biology 15:447–452. Folli, Raffaella, and Heidi Harley. 2008. Teleology and animacy in external arguments. Lingua 118:190–202. Ford, M. 1983. A method for obtaining measures of local parsing complexity throughout sentences. Journal of Verbal Learning and Verbal Behavior 22:203–218. Fortescue, Michael. 1984. Learning to speak Greenlandic: A case study of a two-yearold’s morphology in a polysynthetic language. First Language 5:101–114. Foster, J.F. 1979. Agents, accessories and owners: The cultural base and the rise of ergative structures, with particular reference to Ozark English. In Ergativity. New York: Academic Press. Fox, Danny, and Yusef Grodzinsky. 1998. Children’s passive: A view from the byphrase. Linguistic Inquiry 29:311–332. Fraser, Colin, Ursula Bellugi, and Roger Brown. 1963. Control of grammar in imitation, comprehension, and production. Journal of Verbal Learning and Verbal Behavior 2:121–135. Frauenfelder, U., J. Segui, and J. Mehler. 1980. Monitoring around the relative clause. Journal of Verbal Learning and Verbal Behavior 19:328–337. Frazier, Lyn. 1978. On comprehending sentences: Syntactic parsing strategies. Bloomington: Indiana University Linguistics Club. 1987. Syntactic processing: Evidence from Dutch. Natural Language and Linguistic Theory 5:519–559. Frazier, Lyn, and Janet D. Fodor. 1978. The sausage machine: A new two-stage parsing model. Cognition 6:291–325. Frazier, Lyn, and Keith Rayner. 1982. Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology 14:178–210. Friedmann, Na’ama. 2007. Young children and A-chains: The acquisition of Hebrew unaccusatives. Language Acquisition 14:377–422. Gagliardi, Ann C. 2012. Input and intake in language acquisition. Doctoral Dissertation, University of Maryland. Gallistel, C.R. 1990. The organization of learning. Cambridge, MA: MIT Press. Gary, Judith Olmsted, and Edward Louis Keenan. 1977. On collapsing grammatical relations in universal grammar. In Syntax and semantics, vol. 8: Grammatical relations, ed. Peter Cole and Jerrold M. Sadock, 83–120. New York: Academic Press. Gelman, Rochel. 1990. First principles organize attention to and learning about relevant data: Number and the animate-inanimate distinction as examples. Cognitive Science 14:79–106. Gelman, Rochel, Frank Durgin, and Lisa Kaufman. 1995. Distinguishing between animates and inanimates: Not by motion alone. In Causal cognition: A multidisciplinary debate, ed. Dan Sperber, David Premack, and Ann James Premack, 150–184. Oxford: Clarendon Press.

308

Bibliography

Gelman, Rochel, and Elizabeth Spelke. 1981. The development of thoughts about animate and inanimate objects: Implications for research on social cognition. In Social cognition, ed. J.H. Flavell and L. Ross, 43–66. New York: Academic Press. Gelman, Susan, and Melissa Koenig. 2001. The role of animacy in children’s understanding of ‘move’. Journal of Child Language 28:683–701. Gelman, Susan, and John Opfer. 2002. Development of the animate–inanimate distinction. In Blackwell handbook of childhood cognitive development, ed. Usha Goswami, 151–166. Malden, MA: Blackwell. Gelman, Susan A., and Gail M. Gottfried. 1996. Children’s causal explanations of animate and inanimate motion. Child Development 67:1970–1987. Gentner, Dedre, and Ilene M. France. 1988. The verb mutability effect: Studies of the combinatorial semantics of nouns and verbs. In Lexical ambiguity resolution: Perspectives from psycholinguistics, neuropsychology and artificial intelligence, ed. Steven L. Small, Garrison W. Cottrell, and Michael K. Tanenhaus, 343–382. San Mateo, CA: Morgan Kaufmann Publishers, Inc. Gibson, Edward. 1998. Linguistic complexity: Locality of syntactic dependencies. Cognition 68:1–76. Gillette, Jane, Henry Gleitman, Lila Gleitman, and Anne Lederer. 1999. Human simulations of vocabulary learning. Cognition 73:135–176. Givón, Talmy. 1976. Topic, pronoun, and grammatical agreement. In Subject and topic, ed. Charles Li, 267–280. New York: Academic Press. Gleason, Jean Berko, Brenda Caldwell Phillips, Richard Ely, and Elena Zaretsky. 2009. Alligators all around: The acquisition of animal terms in English and Russian. In Crosslinguistic approaches to the psychology of language: Research in the tradition of Dan Isaac Slobin, ed. Jiansheng Guo, Elena Lieven, Nancy Budwig, Susan Ervin-Tripp, Keiko Nakamura, and Seyda ¸ Ózçali¸skan, 17–26. New York: Psychology Press. Gleitman, Lila. 1990. The structural sources of verb meanings. Language Acquisition 1:3–55. Gleitman, Lila, Kimberly Cassidy, Rebecca Nappa, Anna Papafragou, and John C. Trueswell. 2005. Hard words. Language Learning and Development 1:23–64. Golinkoff, Roberta, and Joyce Kerr. 1978. Infants’ perception of semantically defined action roles in filmed events. Merrill-Palmer Quarterly 24:53–61. Gomez, Rebecca, and LouAnn Gerken. 1999. Artificial grammar learning in one-yearolds leads to specific abstract knowledge. Cognition 70:109–135. Goodluck, Helen, and Dawn Behne. 1992. Development in control and extraction. In Theoretical issues in language acquisition, ed. Jürgen Weissenborn, Helen Goodluck, and Tom Roeper. Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Goodluck, Helen, M. Finney, and M. Ling. 1995. Gap accessibility and children’s sentence processing. Cahiers Linguistiques d’Ottawa. Dept. of Linguistics, University of Ottawa. Goodluck, Helen, and Susan Tavakolian. 1982. Competence and processing in children’s grammar of relative clauses. Cognition 11:1–27. Goodman, Nelson. 1955. Fact, fiction, and forecast. Cambridge, MA: Harvard University Press.

Bibliography

309

Goren, C.C., M. Sarty, and P.Y.K. Wu. 1975. Visual following and pattern discrimination of face-like stimuli by newborn infants. Pediatrics 56:544–549. Grimshaw, Jane. 1981. Form, function, and the language acquisition device. In The logical problem of language acquisition, ed. C.L. Baker and J.J. McCarthy, 165– 182. Cambridge, MA: MIT Press. Gruber, Jeffrey S. 1965. Studies in lexical relations. Doctoral Dissertation, MIT. Bloomington: Indiana University Linguistics Club. 1976. Lexical structures in syntax and semantics. New York: North-Holland. Haegeman, Liliane. 1991. Introduction to Government and Binding theory. Cambridge, MA: Blackwell. Hale, Kenneth. 1972. A note on subject–object inversion in Navajo. In Issues in linguistics: Papers in honor of Henry and Renee Kahane, ed. B. Kachru et al., 300–309. Champaign: University of Illinois Press. Hall, Barbara. 1965. Subject and object in modern English. Doctoral Dissertation, MIT. Hare, Brian, Josep Call, Bryan Agnetta, and Michael Tomasello. 2000. Chimpanzees know what conspecifics do and do not see. Animal Behaviour 59:771–786. Harris, Zellig. 1951. Methods in structural linguistics. University of Chicago Press. Hart, Betty, and Todd Risley. 1995. Meaningful differences in the everyday experience of young American children. Baltimore, MD: Paul H. Brookes Publishing. Hatano, Giyoo, and Kayoko Inagaki. 1999. A developmental perspective on informal biology. In Folkbiology, ed. Douglas Medin and Scott Atran, 322–354. Cambridge, MA: MIT Press. Haviland, Susan E., and Herbert H. Clark. 1974. What’s new? Acquiring new information as a process in comprehension. Journal of Verbal Learning and Verbal Behavior 13:512–521. Hayhurst, Hazel. 1967. Some errors of young children in producing passive sentences. Journal of Verbal Learning and Verbal Behavior 6:634–639. Heath, Jeffrey. 1976. Substantival hierarchies: Addendum to Silverstein. In Grammatical categories in Australian languages, ed. R.M.W. Dixon, Linguistic series 22, 172–190. Canberra: Australian Institute of Aboriginal Studies. Hicks, Glyn. 2009. Tough-constructions and their derivation. Linguistic Inquiry 40:535–566. Hirsch, Christopher, Robyn Orfitelli, and Ken Wexler. 2008. The acquisition of raising reconsidered. In Language Acquisition and Development: Proceedings of GALA 2007, ed. Anna Gavarró and Maria João Freitas. Newcastle: Cambridge Scholars. Hirsch, Christopher, and Kenneth Wexler. 2005. Children’s passives and their resulting interpretation. In Proceedings of the inaugural conference on Generative Approaches to Language Acquisition–North America, ed. K. U. Deen, J. Nomura, B. Schulz, and B.D. Schwartz, 125–136. Cambridge, MA: University of Connecticut Occasional Papers in Linguistics. 2007. The late development of raising: What children seem to think about seem. In New horizons in the analysis of control and raising, ed. William D. Davies and Stanley Dubinsky, 35–70. Dordrecht: Springer. Hirsh-Pasek, Kathy, and Roberta Golinkoff. 1991. Language comprehension: A new look at some old themes. In Biological and behavioral aspects of language

310

Bibliography

acquisition, ed. N. Krasnegor, D. Rumbaugh, M. Studdert-Kennedy, and R. Shiefelbusch, 301–320. Hillsdale, NJ: Lawrence Erlbaum Associates. Holisky, Dee Ann. 1987. The case of the intransitive subject in Tsova-Tush (Batsbi). In Studies in ergativity, ed. R.M.W. Dixon, 103–132. Amsterdam: North-Holland. Holmberg, Anders. 2000. Am I unscientific? A reply to Lappin, Levine, and Johnson. Natural Language and Linguistic Theory 18:837–842. Hope, Edward. 1974. The deep syntax of Lisu sentences. Doctoral Dissertation, Australian National University, Pacific Linguistics. Series B, no. 34. Hopper, Paul, and Sandra A. Thompson. 1980. Transitivity in grammar and discourse. Language 56:251–299. 1982. Syntax and semantics vol. 15: Studies in transitivity. New York: Academic Press. Horgan, Diane. 1978. The development of the full passive. Journal of Child Language 5:65–80. Horn, Larry. 1981. A pragmatic approach to certain ambiguities. Linguistics and Philosophy 4:321–358. Hornstein, Norbert. 1999. Movement and control. Linguistic Inquiry 30:69–96. Hsiao, Fanny, and Edward Gibson. 2003. Processing relative clauses in Chinese. Cognition 90:3–27. Huddleston, Rodney. 1970. Some remarks on case grammar. Linguistic Inquiry 1:501–511. Hyams, Nina, Dimitris Ntelitheos, and Cecile Manorohanta. 2006. Acquisition of the Malagasy voicing system: Implications for the adult grammar. Natural Language and Linguistic Theory 24:1049–1092. Hyams, Nina, and William Snyder. in preparation. Young children are frozen: Reflexive clitics and the universal freezing hypothesis. Unpublished ms., UCLA and University of Connecticut. Jackendoff, Ray. 1972. Semantic interpretation in generative grammar. Cambridge, MA: MIT Press. 1983. Semantics and cognition. Cambridge, MA: MIT Press. 1990. Semantic structures. Cambridge, MA: MIT Press. Jaeggli, Osvaldo. 1986. Passive. Linguistic Inquiry 17:587–622. Johansson, Sara, and Julie Brittain. 2012. The lexical semantics of (Northern East) Cree verbs of emission: A unified analysis of -piyi. Paper presented at the Linguistic Society of America annual meeting, Portland, January 2012. Johnson, Mark H., Suzanne Dziurawiec, Hadyn Ellis, and John Morton. 1991. Newborns’ prefential tracking of face-like stimuli and its subsequent decline. Cognition 40:1–19. Johnson, Susan, Virginia Slaughter, and Susan Carey. 1998. Whose gaze will infants follow? The elicitation of gaze following in 12-month-olds. Developmental Science 1:233–238. Kako, Edward. 1998. The event semantics of syntactic structures. Doctoral Dissertation, University of Pennsylvania. Katz, Jerrold, and Jerry Fodor. 1963. The structure of a semantic theory. Language 39:170–210.

Bibliography

311

Katz, Jerrold, and Paul Postal. 1964. An integrated theory of linguistic descriptions. Cambridge, MA: MIT Press. Kayne, Richard. 1984. Connectedness and binary branching. Dordrecht: Foris. Keenan, Edward. 1976. Toward a universal definition of “subject.” In Subject and topic, ed. Charles Li, 303–334. New York: Academic Press. Keenan, Edward, and Bernard Comrie. 1972. Noun phrase accessibility and universal grammar. Paper presented at the 47th Annual Meeting of the LSA, Atlanta, GA, December 27–29. Keil, Frank C. 1979. Semantic and conceptual development: An ontological perspective. Cambridge, MA: Harvard University Press. 1989. Concepts, kinds and cognitive development. Cambridge, MA: MIT Press. Kemp, Charles, Amy Perfors, and Joshua B. Tenenbaum. 2007. Learning overhypotheses with hierarchical Bayesian models. Developmental Science 10:307–321. King, J., and Marcel Just. 1991. Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language 30:580–602. Kirby, Susannah. 2009. Semantic scaffolding in first language acquisition: The acquisition of raising-to-object and object control. Doctoral Dissertation, University of North Carolina at Chapel Hill. Kirby, Susannah, and Misha Becker. 2007. Which it is it? The acquisition of referential and expletive it. Journal of Child Language 34:571–599. Kitagawa, Yoshihisa. 1986. Subjects in Japanese and English. Doctoral Dissertation, UMass Amherst. Koopman, Hilda, and Dominique Sportiche. 1991. The position of subjects. Lingua 85:211–258. Koster, Jan. 1978. Locality principles in syntax. Dordrecht: Foris. Krug, Manfred. 2001. Frequency, iconicity, categorization: Evidence from emerging modals. In Frequency and the emergence of linguistic structure, ed. Joan Bybee and Paul Hopper, 309–335. Amsterdam: John Benjamins. Kuczaj, Stan. 1977. The acquisition of regular and irregular past tense forms. Journal of Verbal Learning and Verbal Behavior 16:589–600. Kuno, Susumu. 1972. Subject raising in Japanese. Papers in Japanese Linguistics 1. 1973. The structure of the Japanese language. Cambridge, MA: MIT Press. Lakoff, George. 1966. Stative adjectives and verbs in English. Technical Report NSF16, The Computation Laboratory of Harvard University Mathematical Linguistics and Automatic Translation, Cambridge, MA. 1971. On generative semantics. In Semantics, ed. Danny Steinberg and Leon Jakobovits, 232–296. Cambridge University Press. Landau, Barbara, and Lila R. Gleitman. 1985. Language and experience: Evidence from the blind child. Cambridge, MA: Harvard University Press. Landau, Idan. 2003. Movement out of control. Linguistic Inquiry 34:471–498. 2010. The locative syntax of experiencers. Cambridge, MA: MIT Press. Lasnik, Howard, and Robert Fiengo. 1974. Complement object deletion. Linguistic Inquiry 5:535–570. Lasnik, Howard, and Juan Uriagereka. 1994. A course in GB syntax. Cambridge, MA: MIT Press.

312

Bibliography

Laurendeau, M., and A. Pinard. 1962. Causal thinking in the child: A genetic and experimental approach. New York: International Universities Press. Lebeaux, David, and Steven Pinker. 1981. The acquisition of the passive. Paper presented at the Boston University Conference on Language Development. Lederer, Anne, Henry Gleitman, and Lila Gleitman. 1995. Verbs of a feather flock together: Semantic information in the structure of maternal speech. In Beyond names for things: Young children’s acquisition of verbs, ed. Michael Tomasello and William E. Merriman, 277–297. Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Lees, Robert B. 1960. A multiply ambiguous adjectival construction in English. Language 36:207–221. Legate, Julie Anne, and Charles Yang. 2002. Empirical re-assessment of stimulus poverty arguments. The Linguistic Review 19:151–162. 2007. Morphosyntactic learning and the development of tense. Language Acquisition 14:315–344. Legerstee, Maria. 1994. Patterns of 4-month-old infant responses to hidden silent and sounding people and objects. Early Development and Parenting 3:71–80. 2001. Domain specificity and the epistemic triangle: The development of the concept of animacy in infancy. In Emerging cognitive abilities in early infancy, ed. Francisco Lacerda, Claes von Hofsten, and Mikael Heimann, 193–212. Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Legerstee, Maria, Andrée Pomerleau, Gérard Malcuit, and Helga Feider. 1987. The development of infants’ responses to people and a doll: Implications for research in communication. Infant Behavior and Development 10:81–95. Lehmann, Winfred P. 1973. A structural principle of language and its implications. Language 49:47–66. Lempert, Henrietta. 1989. Animacy constraints on preschool children’s acquisition of syntax. Child Development 60:237–245. de Leon, Lourdes. 1998. The emergent participant: Interactive patterns in the socialization of Tzotzil (Mayan) infants. Journal of Linguistic Anthropology 8:131–161. The Linguistic Review (2002) 19: issues 1–2. Leslie, Alan. 1987. Pretense and representation: The origins of theory of mind. Psychological Review 94:412–426. Leslie, Alan M., and Stephanie Keeble. 1987. Do six-month-old infants perceive causality? Cognition 25:265–288. Levin, Beth. 1993. English verb classes and alternations: A preliminary investigation. University of Chicago Press. Levin, Beth, and Malka Rappaport Hovav. 1995. Unaccusativity. Cambridge, MA: MIT Press. Li, Charles, and Sandra Thompson. 1976. Subject and topic: A new typology of language. In Subject and topic, ed. Charles Li. New York: Academic Press. Lidz, Jeffrey, Henry Gleitman, and Lila R. Gleitman. 2004. Kidz in the ’hood: Syntactic bootstrapping and the mental lexicon. In Weaving a lexicon, ed. D. Geoffrey Hall and Sandra Waxman, 603–636. Cambridge, MA: MIT Press. Limber, J. 1973. The genesis of complex sentences. In Cognitive development and the acquisition of language, ed. Timothy E. Moore. New York: Academic Press.

Bibliography

313

Lin, Chien-Jer Charles, and Thomas G. Bever. 2006. Subject preference in the processing of relative clauses in Chinese. In Proceedings of WCCFL 25, 254–260. Somerville, MA: Cascadilla Press. Liu, David, Henry M. Wellman, Twila Tardif, and Mark A. Sabbagh. 2008. Theory of mind development in Chinese children: A meta-analysis of false-belief understanding across cultures and languages. Developmental Psychology 44:523–531. Lorusso, Paolo, Claudia Caprin, and Maria Teresa Guasti. 2005. Overt subject distribution in early Italian children. In Supplemental Proceeding of the Boston University Conference on Language Development, at www.bu.edu/linguistics/ APPLIED/BUCLD. Lowder, Matthew, and Peter C. Gordon. 2012. The pistol that injured the cowboy: Difficulty with inanimate subject-verb integration is reduced by structural separation. Journal of Memory and Language 66:819–832. Lyons, John. 1968. Introduction to theoretical linguistics. Cambridge University Press. Ma, Wei Ji, Jeffrey M. Beck, Peter E. Latham, and Alexandre Pouget. 2006. Bayesian inference with probabilistic population codes. Nature Neuroscience 9:1432–1438. Macaruso, Paul, Donald Shankweiler, Brian Byrne, and Stephen Crain. 1993. Poor readers are not easy to fool: Comprehension of adjectives with exceptional control properties. Applied Psycholinguistics 14:285–298. MacQueen, J.B. 1967. Some methods for classification and analysis of multivariate observations. In Proceedings of the 5th symposium on math, statistics, and probability, ed. Lucien Marie Le Cam and Jerzy Neyman, 281–298. Berkeley, CA: University of California Press. MacWhinney, Brian. 1977. Starting points. Language 53:152–168. 2000. The child language data exchange system. Mahwah, NJ: Lawrence Erlbaum Associates, Inc. MacWhinney, Brian, and Csaba Pléh. 1988. The processing of restrictive relative clauses in Hungarian. Cognition 29:95–141. Mair, Christian. 1990. Infinitival complement clauses in English: A study of syntax in discourse. Cambridge University Press. Mak, Willem M., Wietske Vonk, and Herbert Schriefers. 2002. The influence of animacy on relative clause processing. Journal of Memory and Language 47:50–68. Mandler, Jean, and Laraine McDonough. 1993. Concept formation in infancy. Cognitive Development 8:291–318. Manzini, Rita. 1983. On control and control theory. Linguistic Inquiry 14:421–446. Marantz, Alec. 1982. On the acquisition of grammatical relations. Linguistische Berichte 80–82:32–69. 1984. On the nature of grammatical relations. Cambridge, MA: MIT Press. Maratsos, Michael, Danny Fox, Judith Becker, and Mary Anne Chalkley. 1985. Semantic restrictions on children’s passives. Cognition 19:167–191. Maratsos, Michael, Stan Kuczaj, D.E.C. Fox, and M. Chalkley. 1979. Some empirical studies in the acquisition of transformational relations: Passives, negatives, and the past tense. In Minnesota symposium on child psychology, ed. W.A. Collins. Hillsdale, NJ: Lawrence Erlbaum. Marcus, Gary. 1993. Negative evidence in language acquisition. Cognition 46:53–85.

314

Bibliography

Massey, Christine, and Rochel Gelman. 1988. Preschooler’s ability to decide whether a photographed unfamiliar object can move itself. Developmental Psychology 24:307–317. May, Robert. 1977. The grammar of quantification. Doctoral Dissertation, MIT. 1985. Logical form: Its structure and derivation. Cambridge, MA: MIT Press. McCawley, James. 1968. The role of semantics in a grammar. In Universals in linguistic theory, ed. Emmon Bach and Robert T. Harms, 125–170. New York: Holt, Rinehart and Winston. McDaniel, Dana, and Helen Cairns. 1996. Eliciting judgments of grammaticality and reference. In Methods for assessing children’s syntax, ed. Dana McDaniel, Cecile McKee, and Helen Smith Cairns, 233–254. Cambridge, MA: MIT Press. McDaniel, Dana, Helen Smith Cairns, and J.R. Hsu. 1991. Control principles in the grammars of young children. Language Acquisition 1:297–335. McKee, Cecile. 1997a. Lexical factors in language acquisition. Language and Cognitive Processes 12:853–858. 1997b. Some adjectives are “easy” and some are not. Lexicology 3:59–83. Mecklinger, A., Herbert Schriefers, K. Steinhauer, and A. Friederici. 1995. Processing relative clauses varying on syntactic and semantic dimensions: An analysis with event-related potentials. Memory & Cognition 23:477–494. Meltzoff, Andrew N. 1995. Understanding the intentions of others: Re-enactment of intended acts by 18-month-old children. Developmental Psychology 31:838–850. Merlo, Paola, and Suzanne Stevenson. 2001. Automatic verb classification based on statistical distributions of argument structure. Computational Linguistics 27:373–408. Messenger, Katherine, Holly P. Branigan, and Janet F. McLean. 2012. Is children’s acquisition of the passive a staged process? Evidence from six- and nine-year-olds’ production of passives. Journal of Child Language 39:991–1016. Michotte, Albert. 1963. The perception of causality. London: Methuen. Mintz, Toben. 2002. Category induction from distributional cues in an artificial language. Memory & Cognition 30:678–686. 2003. Frequent frames as a cue for grammatical categories in child directed speech. Cognition 90:91–117. Mitchell, Tom M. 1990. The need for biases in learning generalizations. In Readings in machine learning, ed. Jude W. Shavik and Thomas G. Dietterich, 184–191. San Mateo, CA: Morgan Kaufmann Publishers, Inc. Originally published as a Rutgers Technical Report, 1980. Mitchener, Garrett, and Misha Becker. 2011. A computational model of learning the raising-control distinction. Research on Language and Computation 8:169–207. Moro, Andrea. 1997. The raising of predicates: Predicative noun phrases and the theory of clause structure. Cambridge University Press. Morsbach, Gisela, and Pamela Steel. 1975. “John is easy to see” re-investigated. Journal of Child Language 3:443–447. Müller, Gereon. 1998. Incomplete category fronting. Dordrecht: Kluwer. Naigles, Letitia. 1990. Children use syntax to learn verb meanings. Journal of Child Language 17:357–374.

Bibliography

315

1996. The use of multiple frames in verb learning via syntactic bootstrapping. Cognition 58:221–251. Naigles, Letitia, Anne Fowler, and Atessa Helm. 1995. Syntactic bootstrapping from start to finish with special reference to Down Syndrome. In Beyond names for things: Young children’s acquisition of verbs, ed. Michael Tomasello and William E. Merriman, 299–330. Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Napoli, Donna Jo. 1981. Semantic interpretation vs. lexical governance: Clitic climbing in Italian. Language 57:841–887. Newport, Elissa, and Richard Aslin. 2004. Learning at a distance: Statistical learning of non-adjacent dependencies. Cognitive Psychology 48:127–162. Nilsen, Don Lee Fred. 1973. The instrumental case in English. The Hague: Mouton. Nishigauchi, Taisuke, and Thomas Roeper. 1987. Deductive parameters and the growth of empty categories. In Parameter setting, ed. Thomas Roeper and Edwin Williams, 91–122. Dordrecht: D. Reidel. Niyogi, Sourabh. 2002. Bayesian learning at the syntax-semantics interface. In Proceedings of the 24th Annual Meeting of the Cognitive Science Society, ed. W. Gray and C. Schunn, 697–702. Mahwah, NJ: Lawrence Erlbaum Associates. O’Brien, Karen, Elaine Grolla, and Diane Lillo-Martin. 2006. Long passives are understood by young children. In Proceedings of the 30th Boston University Conference on Language Development, ed. David Bamman, Tatiana Magnitskaia, and Colleen Zaller, 441–451. Ochs, Elinor. 1982. Ergativity and word order in Samoan child language. Language 58:646–671. Olsen, Susan. 1981. Problems of seem/scheinen constructions and their implications for the theory of predicate sentential complementation. Tübingen: Max Niemeyer Verlag. Onishi, Kristine, and Renée Baillargeon. 2005. Do 15-month-old infants understand false beliefs? Science 308:255–258. Orfitelli, Robyn. 2012. Argument intervention in the acquisition of A-movement. Doctoral Dissertation, UCLA. Oshima-Takane, Yuriko. 1988. Children learn from speech not addressed to them: The case of personal pronouns. Journal of Child Language 15:95–108. Øvrelid, Lilja. 2004. Disambiguation of grammatical functions in Norwegian: Modeling variation in word order interpretations conditioned by animacy and definiteness. In Proceedings of the 20th Scandinavian Conference of Linguistics. University of Helsinki Publications. Helsinki, January 7–9, 2004, Department of General Linguistics. Papafragou, Anna, Kimberly Cassidy, and Lila Gleitman. 2007. When we think about thinking: The acquisition of belief verbs. Cognition 105:125–165. Partee, Barbara. 1977. John is easy to please. In Linguistic structures processing, ed. A. Zampolli. Amsterdam: North-Holland. Pearl, Lisa, and Sharon Goldwater. in press. Statistical learning, inductive bias, and Bayesian inference in language acquisition. In Oxford handbook of developmental linguistics, ed. Jeff Lidz, Joe Pater, and William Snyder. Oxford University Press.

316

Bibliography

Pearl, Lisa, and Jeffrey Lidz. 2009. When domain-general learning fails and when it succeeds: Identifying the contribution of domain specificity. Language Learning and Development 5:235–265. Perfors, Amy, Joshua B. Tenenbaum, Thomas L. Griffiths, and Fei Xu. 2011. A tutorial introduction to Bayesian models of cognitive development. Cognition 120:302–321. Perfors, Amy, Joshua B. Tenenbaum, and Elizabeth Wonnacott. 2010. Variability, negative evidence, and the acquisition of verb argument constructions. Journal of Child Language 37:607–642. Perlmutter, David M. 1970. The two verbs begin. In Readings in English transformational grammar, ed. Roderick A. Jacobs and Peter S. Rosenbaum, 107–119. Waltham, MA: Ginn. 1978. Impersonal passives and the unaccusative hypothesis. In Proceedings of the Berkeley Linguistic Society 4, 157–189. Berkeley, CA. Perry, Fred L., and Allan Shwedel. 1979. Interaction of visual information, verbal information, and linguistic competence in the preschool-aged child. Journal of Psycholinguistic Research 8:559–566. Pesetsky, David. 1982. Paths and categories. Doctoral Dissertation, MIT. 1995. Zero syntax: Experiencers and cascades. Cambridge, MA: MIT Press. Piaget, Jean. 1929. The child’s conception of the world. London: Routledge and Kegan Paul. Pinker, Steven. 1982. A theory of the acquisition of lexical interpretive grammars. In The mental representation of grammatical relations, ed. Joan Bresnan. Cambridge, MA: MIT Press. 1984. Language learnability and language development. Cambridge, MA: Harvard University Press. 1987. Resolving a learnability paradox in the acquisition of the verb lexicon. In Lexicon project working papers 17, 1–100. Cambridge, MA: MIT Center for Cognitive Science. 1989. Learnability and cognition: The acquisition of argument structure. Cambridge, MA: MIT Press. Pinker, Steven, David S. Lebeaux, and Loren Ann Frost. 1987. Productivity and constraints in the acquisition of the passive. Cognition 26:195–267. Polinsky, Maria, and Eric Potsdam. 2008. The syntax and semantics of wanting in Indonesian. Lingua 118:1617–1639. Postal, Paul. 1962. Some syntactic rules of Mohawk. Doctoral Dissertation, Yale University. Published by Garland Press, New York, 1979. 1971. Cross-over phenomena. New York: Holt, Rinehart and Winston. 1974. On raising: One rule of English grammar and its theoretical implications. Cambridge, MA: MIT Press. Postal, Paul, and Geoffrey Pullum. 1978. Traces and the description of English complementizer contraction. Linguistic Inquiry 9:1–29. Potsdam, Eric. 2006. Backward object control in Malagasy: Against an empty category analysis. In Proceedings of the 25th West Coast Conference on Formal Linguistics, ed. D. Baumer, D. Montero, and M. Scanlon, 328–336. Somerville, MA: Cascadilla Proceedings Project.

Bibliography

317

Potsdam, Eric, and Maria Polinsky. 2011. Against covert A-movement in Russian unaccusatives. Linguistic Inquiry 42:345–355. Potsdam, Eric, and Jeffrey Runner. 2001. Richard returns: Copy raising and its implications. In Papers from the 37th Regional Meeting of the Chicago Linguistic Society, ed. Mary Andronis, Christopher Ball, Heidi Elston, and Sylvain Neuvel, 453–468. Chicago Linguistic Society. Poulin-Dubois, Diane, Anouk Lepage, and Doreen Ferland. 1996. Infants’ concept of animacy. Cognitive Development 11:19–36. Pullum, Geoffrey. 1997. The morpholexical nature of English to-contraction. Language 73:79–102. Pye, Clifton. 1990. The acquisition of ergative languages. Linguistics 28:1291–1330. Pylkkänen, Liina. 2008. Introducing arguments. Cambridge, MA: MIT Press. Quine, W.V.O. 1969. Natural kinds. In Ontological relativity and other essays, ed. W.V.O. Quine. New York: Columbia University Press. Quinlan, J. Ross. 1992. C4.5: Programs for machine learning. Series in Machine Learning. San Mateo, CA: Morgan Kaufmann. Quirk, Randolph, Sidney Greenbaum, Geoffrey Leech, and Jan Svartvik. 1985. A comprehensive grammar of the English language. London: Longman. Rakison, David H., and Diane Poulin-Dubois. 2001. Developmental origin of the animate–inanimate distinction. Psychological Bulletin 127:209–228. Ransom, Evelyn N. 1977. Definiteness, animacy, and NP ordering. In Proceedings of the 3rd Annual Meeting of the Berkeley Linguistics Society, ed. Kenneth Whistler, Robert van Valin, Chris Chiarello, Jeri Jaeger, Miriam Petruck, Henry Thompson, Ronya Javkin, and Anthony Woodbury, 418–429. Berkeley Linguistics Society. Rayner, Keith, Marcia Carlson, and Lyn Frazier. 1983. The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences. Journal of Verbal Learning and Verbal Behavior 22:358–374. Rett, Jessica, Nina Hyams, and Lauren Winans. 2013. The effects of syntax on the acquisition of evidentiality. In Proceedings of the Boston University Conference on Language Development 37. Somerville, MA: Cascadilla Press. Richards, Norvin. 2001. Interactions and architectures. Oxford University Press. Ritter, Elizabeth, and Sara Thomas Rosen. 2010. Animacy in Blackfoot: Implications for event structure and clause structure. In Syntax, lexical semantics and event structure, ed. Malka Rappaport Hovav, Edit Doron, and Ivy Sichel, 124–152. New York: Oxford University Press. Rizzi, Luigi. 2001. Relativized minimality effects. In The handbook of contemporary syntactic theory, ed. Mark Baltin and Chris Collins, 89–110. Oxford: Blackwell. Rochette, Anne. 1988. Semantic and syntactic aspects of Romance sentential complementation. Doctoral Dissertation, MIT. Rogers, Andy. 1974. Physical perception verbs in English: A study in lexical relatedness. Doctoral Dissertation, UCLA. Rosen, Sara. 1989. Argument structure and complex predicates. Doctoral Dissertation, Brandeis University.

318

Bibliography

Rosenbaum, Peter. 1967. The grammar of English predicate complement constructions. Cambridge, MA: MIT Press. Rothstein, Susan. 1991a. Heads, projections, and category determination. In Views on phrase structure, ed. K. Leffel and D. Bouchard. Dordrecht: Kluwer. 1991b. Syntactic licensing and subcategorization. In Syntax and semantics: Perspectives on phrase structure: Heads and licensing, ed. Susan Rothstein, volume 25. San Diego, CA: Academic Press. Rudanko, Juhani. 1989. Complementation and case grammar: A syntactic and semantic study of selected patterns of complementation in present-day English. State University of New York Press. Rumelhart, David, and J.L. McClelland. 1986. On learning the past tenses of English verbs. In Parallel distributed processing, volume 2. Cambridge, MA: MIT Press. Sachs, Jacqueline. 1983. Talking about the there and then: The emergence of displaced reference in parent-child discourse. In Children’s language, ed. K. E. Nelson. Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Saffran, Jenny, Richard Aslin, and Elissa Newport. 1996. Statistical learning by 8-month-old infants. Science 274:1926–1928. Schachter, Paul. 1976. The subject in Philippine languages: Topic, actor, actor-topic, or none of the above? In Subject and topic, ed. Charles Li, 491–518. New York: Academic Press. Schieffelin, Bambi. 1979. How Kaluli children learn what to say, what to do, and how to feel: An ethnographic study of the development of communicative competence. Doctoral Dissertation, Columbia University. 1985. The acquisition of Kaluli. In The crosslinguistic study of language acquisition, ed. Dan Slobin, volume 1: The Data. Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Schneidman, Laura A., Michelle E. Arroyo, Susan C. Levine, and Susan GoldinMeadow. 2013. What counts as effective input for word learning? Journal of Child Language 40:672–686. Schwartz, Arthur. 1976. On the universality of subject: The Ilocano case. In Subject and topic, ed. Charles Li, 519–543. New York: Academic Press. Schwartz, Richard. 1980. Presuppositions and children’s metalinguistic judgments: Concepts of life and the awareness of animacy restrictions. Child Development 51:364–371. Scott, Rose M., and Cynthia Fisher. 2009. Two-year-olds use distributional cues to interpret transitivity-alternating verbs. Language and Cognitive Processes 24:777–803. Seiter, William J. 1980. Studies in Niuean syntax. New York: Garland Publishing, Inc. Selkirk, Elisabeth. 1980. The role of prosodic categories in English word stress. Linguistic Inquiry 11:563–605. Sheldon, A. 1974. The role of parallel function in the acquisition of relative clauses in English. Journal of Verbal Learning and Verbal Behavior 13:272–281. Silverstein, M. 1976. Hierarchy of features and ergativity. In Grammatical categories in Australian languages, ed. R.M.W. Dixon. Canberra: Australian Institute for Aboriginal Studies.

Bibliography

319

Slobin, Dan. 1966. Grammatical transformations and sentence comprehension in childhood and adulthood. Journal of Verbal Learning and Verbal Behavior 5:219–227. 1975. The more it changes . . . : On understanding language by watching it move through time. In Papers and reports on child language development, volume 10, 1–30. Stanford University. 1981. The origin of grammatical encoding of events. In The child’s construction of language, ed. W. Deutsch. New York: Academic Press. Slobin, Dan I., and C. Welsh. 1973. Elicited imitation as a research tool in developmental psycho-linguistics. In Studies of child language development, ed. C.A. Ferguson and D.I. Slobin. New York: Holt, Rinehart and Winston. Snow, Catherine E., and Charles A. Ferguson. 1977. Talking to children: Language input and acquisition. Cambridge University Press. Snyder, William, Nina Hyams, and Paola Crisma. 1994. Romance auxiliary selection with reflexive clitics: Evidence for early knowledge of unaccusativity. In Proceedings of the 26th Annual Child Language Research Forum, 127–136. Stanford, CA: CSLI. Solan, Lawrence. 1978. The acquisition of tough-movement. In Papers on the structure of child development, ed. Helen Goodluck and Lawrence Solan, volume 4. Amherst: University of Massachusetts Occasional Papers in Linguistics. Spelke, Elizabeth. 1991. Physical knowledge in infancy: Reflections on Piaget’s theory. In The epigenesis of mind: Essays on biology and cognition, ed. Susan Carey and Rochel Gelman, 133–170. Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Sproat, Richard. 1983. VSO languages and Welsh configurationality. In MIT working papers in linguistics, vol. 5: Papers in grammatical theory, ed. I. Haik and D. Massam. Cambridge, MA: MIT Department of Linguistics and Philosophy. Stewart, Judith A. 1982. Perception of animacy. Doctoral Dissertation, University of Pennsylvania. 1984. Object motion and the perception of animacy. Paper presented at the Psychonomic Society Meeting, San Antonio, TX. Stowell, Timothy. 1981. Origins of phrase structure. Doctoral Dissertation, MIT. Suppes, Patrick. 1974. The semantics of children’s language. American Psychologist 29:103–114. de Swart, Peter, Monique Lamers, and Sander Lestrade. 2008. Animacy, argument structure, and argument encoding. Lingua 118:131–140. Tabossi, Patrizia, Michael J. Spivey-Knowlton, Ken McRae, and Michael Tanenhaus. 1994. Semantic effects on syntactic ambiguity resolution: Evidence for a constraint-based resolution process. In Attention & performance XV, ed. Carlo Umiltà and Morris Moscovitch, 589–615. Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Talmy, Leonard. 1985. Lexicalization patterns: Semantic structure in lexical forms. In Language typology and syntactic description, vol 3: Grammatical categories and the lexicon, ed. Timothy Shopen, 57–149. Cambridge University Press. Taylor, Ann, Mitchell Marcus, and Beatrice Santorini. 2003. The PENN treebank: An overview. In Treebanks: The state of the art on syntactically annotated corpora, ed. Anne Abeillé. Dordrecht: Kluwer.

320

Bibliography

Thiessen, Erik, and Jenny Saffran. 2003. When cues collide: Use of stress and statistical cues to word boundaries by 7- to 9-month-old infants. Developmental Psychology 39:706–716. Thorndike, E.L., and I. Lorge. 1944. The teacher’s word book of 30,000 words. New York: Columbia University Teachers College. Tomasello, Michael. 2000. Do young children have adult syntactic competence? Cognition 74:209–253. 2003. Constructing a language: A usage-based theory of language acquisition. Cambridge, MA: Harvard University Press. Tomasello, Michael, and Patricia Brooks. 1998. Young children’s earliest transitive and intransitive constructions. Cognitive Linguistics 9:379–395. Tomlin, Russell S. 1983. On the interaction of syntactic subject, thematic information, and agent in English. Journal of Pragmatics 7:411–432. 1986. Basic word order: Functional principles. London: Croom Helm. Traxler, Matthew J., Robin K. Morris, and Rachel E. Seely. 2002. Processing subject and object relative clauses: Evidence from eye movements. Journal of Memory and Language 47:69–90. Traxler, Matthew J., Rihana S. Williams, Shelley A. Blozis, and Robin K. Morris. 2005. Working memory, animacy, and verb class in the processing of relative clauses. Journal of Memory and Language 53:204–224. Trueswell, John C. 1996. The role of lexical frequency in syntactic ambiguity resolution. Journal of Memory and Language 35:566–585. Trueswell, John C., Irina Sekerina, Nicole M. Hill, and Marian L. Logrip. 1999. The kindergarten-path effect: Studying on-line sentence processing in young children. Cognition 73:89–134. Trueswell, John C., and Michael K. Tanenhaus. 1994. Toward a lexicalist framework for constraint-based syntactic ambiguity resolution. In Perspectives on sentence processing, ed. C. Clifton, L. Frazier, and K. Rayner, 155–179. Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Trueswell, John C., Michael K. Tanenhaus, and Susan M. Garnsey. 1994. Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language 33:285–318. Tyler, Lorraine K., and William D. Marslen-Wilson. 1977. The on-line effects of semantic context on syntactic processing. Journal of Verbal Learning and Verbal Behavior 16:683–692. Valian, Virginia. 1991. Syntactic subjects in the early speech of American and Italian children. Cognition 40:21–81. Van Valin, Jr., Robert D. 1990. Semantic parameters of split intransitivity. Language 66:221–260. Wanner, Eric, and Michael Maratsos. 1978. An ATN approach to comprehension. In Linguistic theory and psychological reality, ed. Morris Halle, Joan Bresnan, and George A. Miller, 119–161. Cambridge, MA: MIT Press. Weckerly, Jill, and Marta Kutas. 1998. The interaction of noun animacy and grammatical role in the processing of object relative sequences. Technical Report 98.01, UCSD Cognitive Science Tech Report.

Bibliography

321

1999. An electrophysiological analysis of animacy effects in the processing of object relative sentences. Psychophysiology 36:559–570. Wellman, Henry. 1985. The child’s theory of mind. Cambridge, MA: MIT Press. Wellman, Henry, David Cross, and Julanne Watson. 2001. Meta-analysis of theory-ofmind development: The truth about false belief. Child Development 72:655–684. Wexler, Kenneth. 1992. Some issues in the growth of control. In Control and grammar, ed. Richard K. Larson, Sabine Iatridou, Utpal Lahiri, and James Higginbotham. Dordrecht: Kluwer. 2004. Theory of phasal development: Perfection in child grammar. In MITWPL 48: Plato’s problems: Papers on language acquisition, ed. Aniko Csirmaz, Andrea Gualmini, and Andrew Nevins, 159–209. Cambridge, MA: MIT Working Papers in Linguistics. Williams, Edwin. 1981. Argument structure and morphology. The Linguistic Review 1:81–114. 1984. Grammatical relations. Linguistic Inquiry 15:639–673. Wimmer, Heinz, and Josef Perner. 1983. Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception. Cognition 13:103–128. Woodward, Amanda. 1998. Infants selectively encode the goal object of an actor’s reach. Cognition 69:1–34. Woodward, Amanda, Ann Phillips, and Elizabeth Spelke. 1993. Infants’ expectations about the motion of animate versus inanimate objects. In Proceedings of the Fifteenth Annual Meeting of the Cognitive Science Society, Boulder, CO, 1087–1091. Hillsdale, NJ: Lawrence Erlbaum Associates Inc. Wurmbrand, Susanne. 2001. Infinitives: Restructuring and clause structure. New York: Mouton de Gruyter. Xu, Fei, and Joshua Tenenbaum. 2007. Word learning as Bayesian inference. Psychological Review 114:245–272. Yamamoto, Mutsumi. 1999. Animacy and reference: A cognitive approach to corpus linguistics. Amsterdam/Philadelphia: John Benjamins. Yang, Charles. 2002. Knowledge and learning in natural language. New York: Oxford University Press. 2011. A statistical test for grammar. In Proceedings of the 2nd Workshop on Cognitive Modeling and Computational Linguistics, 30–38. Association for Computational Linguistics. Yuan, Sylvia, and Cynthia Fisher. 2009. “Really? She blicked the baby?” Two-year-olds learn combinatorial facts about verbs by listening. Psychological Science 20:619– 626. Zaenen, Annie. 1988. Unaccusative verbs in Dutch and the syntax-semantics interface. Report No. CSLI-88-123, Stanford, CA. Zipf, George K. 1949/1965. Human behavior and the principle of least effort: An introduction to human ecology. New York: Hafner Publishing Co. Zubizarreta, Maria-Luisa. 1985. The relation between morphophonology and morphosyntax: The case of Romance causatives. Linguistic Inquiry 16:247–289.

Index

A-chain maturation, 195–199, 233–234 A-movement (NP-movement), 19, 20, 22, 26, 34, 46, 80, 85, 88, 95, 100, 104, 106, 195, 205, 233, 296 A -movement, 20, 22, 33, 34, 104 agency, 62, 70–72, 75, 78, 110, 172–173 ambiguity, 7, 51, 52, 54, 89, 123, 170 lexical, 91, 97, 102, 123, 140, 143 structural, 9, 130 Ancient Greek, 67 Animacy Hierarchy, 11, 57, 63, 74, 85, 109, 114, 117, 128, 155, 240, 290 anti-canonical verbs, 181–182, 237 Arabana, 64, 65 Arctic Eskimo, 115 argument structure, 10, 47, 50, 64, 76, 86, 180, 192, 193, 217, 257, 261, 265 auxiliary verbs, 6, 88, 98, 192–194, 294 Basque, 65 Bayesian learning, 251, 257, 261–265 Blackfoot, 66 Canonical Alignment Hypothesis, 241 Cantonese, 104 case, 33, 44, 65–66, 77, 95, 105, 180 Case, 20, 21, 33–35, 39, 44, 46, 47, 80, 101 Case Grammar, 40, 77–78, 83 causation, 158, 161–163 Causative Alternation, 229–233, 249 Chamorro, 66, 96 computational models, 12, 128 concept of animacy, 62, 71 conceptual/cognitive development, 11, 12 constructional homonyms, 3, 4, 15, 29, 39, 85, 208, 295 Continuity Principle, 191, 199, 226

322

control adjectives, 3, 208 object control verbs, 124, 192, 208 subject control verbs, 30, 39, 53–54, 58–59, 140–147, 149–153 copy raising, 19, 21, 53, 101–102 Cree (Northern East), 73 degree adjectives/constructions, 38, 212, 227 Derived Subject Condition, 85, 122, 124, 283 domain specificity, 257, 275–276, 283 Dutch, 41–42 Dyirbal, 65, 115 ergative case/morphology, 65–66, 72, 110, 117, 119, 182–184 ergative syntax, 115–119, 182 expletive subjects, 5, 16–17, 19, 31, 33, 37–39, 47, 54–58, 90, 141, 188–189, 217, 227, 251, 259, 274, 276, 295 external argument, 2, 5, 15, 20, 22, 30, 37, 45, 47–48, 50, 53, 56, 61, 73, 81, 84, 86, 117, 121, 155, 178, 185, 219, 227, 241, 260, 276, 283, 286 false belief, 166–168 Finnish, 105–106 folk taxonomy, 288–289 frequency, 132, 146, 150–153, 212–213, 234, 247, 250–256, 294 Genitive of Negation, 233 Georgian, 67 German, 86–90, 103–104, 193 Government-Binding Theory, 20, 26, 34, 42, 46, 80 grammaticality judgment task – see sentence (grammaticality) judgment task

Index Guaraní, 67 Gumbainggir, 65 Hindi, 64 idioms, 16–17, 19, 29, 34, 276 Indonesian, 100 Indonesian (Bahasa), 106 innateness / innate knowledge, 3, 13, 68, 84, 114, 119, 171, 257, 261, 275, 276, 283–286 intention, 166–169 Italian, 43–44, 86, 90–92, 103 Jacaltec, 65, 66, 110–112, 116 Japanese, 66 Kaluli, 79, 117, 183 K’iche’, 116–117, 182, 184 Labrador Inuttut, 107 language evolution, 292–293 language processing, 11, 126, 128 Lexical Functional Grammar (LFG), 19, 27, 47, 50 Lisu, 107, 290–291 long-distance (non-local) dependency, 1, 2, 4, 201, 296 Malagasy, 100 Mandarin, 104, 189 Maori, 92–94 Minimalism, 10, 21–25, 34, 45, 48, 84 modality/modals, 98, 100, 202, 205 Mohawk, 113 motion, 159–161, 164–165, 172, 174, 178–179 Navajo, 66, 69 Niuean, 97, 106 Norwegian, 68 novel word learning, 148–153, 180–182, 237 NP-movement, see A-movement Onondaga, 113 opaque constructions, 6, 8, 10, 12, 13, 15, 29, 50, 51, 61, 76, 86, 147, 156, 157, 185, 188, 206, 227, 239, 259, 283 Ozark English, 67

323

passive, 5, 8–9, 16, 88, 99, 100, 109, 112, 154–155, 182, 208 Persian, 67 phase, 22, 45, 196 Piaget, 156, 170, 174 PRO, 26, 32, 39, 101, 145, 192 probabilistic learning, 54, 62, 86 123, 181, 246 proto-roles (Agent, Patient), 82–84, 99, 122, 123, 283 psychological (psych-) verbs, 57, 58, 83, 138, 228 Quantifier Lowering, see scope effects raising adjectives, 19, 56, 92 verbs (raising-to-object), 4, 124, 192, 207–208 verbs (raising-to-subject), 5, 9, 15, 30, 52–54, 71, 80, 85, 140–147, 149–153, 188 ready, 36–38, 210, 227 relative clauses object vs. subject, 136–139 reduced vs. unreduced, 131–136 Response Time, 217–221, 236–237 restructuring, 87–88 Ritharngu, 65, 66 Russian, 65, 66 Samoan, 92, 95, 117, 183–185 scope effects (Quantifier Lowering), 18–19, 59, 99, 102 Semantic Bootstrapping, 12, 68, 116, 117, 187, 190, 196 sentence completion, 140–147 sentence (grammaticality) judgment task, 127–128, 148–153, 175–177, 200, 204 smuggling, 10, 15, 22–25, 34, 35, 48 Southern Tiwa, 67 Spanish, 65, 189 Swedish, 68 Syntactic Bootstrapping, 8, 12, 61, 190, 222, 229 Thargari, 65 Thematic Hierarchy, 63, 76, 79, 83, 109, 114, 117, 196, 291

324

Index

thematic roles, 10, 15, 21, 26, 28, 29, 33, 35, 39, 40, 44–47, 49, 53, 62–63, 76, 89, 101, 145, 178 Tiwi, 67 Tlapanec, 66 Tongan, 94–96 topic-prominent languages, 290–291 tough-adjectives, 3, 5, 9, 15, 52, 56, 71, 71, 85, 188 tough-constructions, tough-movement: see tough-adjectives Truth-Value judgment task, 203, 211 Tsova-Tush, 72

unaccusative verbs, 5, 9, 16, 52, 57–58, 71, 73, 113, 133, 179, 247–250 unergative verbs, 5, 72, 73, 179, 247–250 Universal Grammar, 11, 84, 125, 205, 212, 261 UTAH, 84 Verb Mutability Effect, 98, 202 verbs of emission, 42, 52, 72–74, 121, 123 volition, 41, 53, 70, 78, 83 Welsh, 116 Zipf’s law, 259