Experimental Investigations of the Formation and Restriction of Abstract Grammatical

Experimental Investigations of the Formation and Restriction of Abstract Grammatical Constructions in Young Children

A thesis submitted to the University of Manchester for the degree of Doctor of Philosophy in the Faculty of Science and Engineering

2004

Ben Ambridge

Department of Psychology

CONTENTS

Title page1

Contents2

List of Tables9

List of Figures11

Abstract12

Declaration and Copyright Statement13

Acknowledgements14

Chapter 1: Generativist Approaches to Language Acquisition15

1.0 Thesis Introduction and Outline15

1.1 The generativist approach to language: “Words and rules”18

1.1.1 The X-bar theory of phrase structure: General principles20

1.1.2 Functional phrases23

1.1.3 Minimalism26

1.1.4 X-bar theory: Conclusion29

2.0 Generativist Theories of Language Acquisition29

2.1 Radford’s structure building theory30

2.2 Full competence accounts36

2.2.1 Valian’s general performance limitations account36

2.2.2 Parameter setting accounts44

2.2.2.1 Wexler’s (1998) agreement/tense omission45

model (ATOM) and very early parameter setting (VEPS)

2.2.2.2 Hyams’ (1999) pragmatic principle account53

2.2.3 Semantic bootstrapping and semantic constraints55

Pinker (1984, 1989)

2.2.3.1 Semantic bootstrapping (Pinker, 1984)56

2.2.3.2 Semantic constraints on verb57

argument structure privileges(Pinker 1984, 1989)

3.0 The Generativist Approach: Summary and Conclusion62

Chapter 2: A Constructivist Approach to Language Acquisition66

1.0 Background to the Account: Construction Grammar67

1.1 General principles of and evidence for construction grammar67

1.2 Radical Construction Grammar72

1.3 Construction grammar and Tomasello’s (2003) constructivist75

theory of language acquisition

2.0 Intention Reading and the Acquisition of Early Words and77

Utterance Wholes

2.1 Acquiring word meanings using skills of intention reading77

2.2 Segmenting the input stream into words79

2.3 Acquiring utterance wholes82

3.0 Schematization: The acquisition of partially productive, lexically83

specific construction schemas

3.1 The nature of the schemas: Verb islands or verb + other islands84

3.1.1 Evidence from an experimental study87

3.1.2 Evidence from a naturalistic data study88

3.1.3 Evidence from a computer modelling study91

3.2 Evidence for the lexically specific nature of early construction95

schemas

3.2.1 Novel verb studies95

3.2.2 Weird word order studies99

3.2.3 A syntactic priming study101

3.2.4 Comprehension studies102

3.2.5 Cross-linguistic studies106

3.3 The process of schematization107

3.3.1 Factors in the process of schematization: Token109

frequency of the frame in the input

3.3.2 Factors in the process of schematization: Type112

frequency of the variable item(s) and frame in the input

3.4 Schematization: conclusion113

4.0 The Process of Analogy and the Formation of Abstract113

Construction Schemas

4.1 Factors in the formation of abstract construction schemas118

4.1.1 Token and type frequency of the construction and its118

variable elements

4.1.2 Semantic generality of the verb and its interaction119

with token and type frequency of the construction and its

variable elements

4.1.3 Construction conspiracies124

4.2 The formation of abstract construction schemas: Conclusion125

5.0 Functionally Based Distributional Analysis and the Formation of126

6.0 The Appropriate Restriction of Linguistic Generalisations130

7.0 Challenges for the Constructivist Account and Conclusion132

7.1 Factors influencing children’s learning of productive132

construction schemas

7.2 Defining chunks, construction schemas, slots and frames135

7.3 Methodological factors and age differences137

7.4 Conclusion138

Chapter 3: Experiment 1. Children’s Acquisition of Non-subject140

Wh- Questions as a Test of Movement-Based (Generativist)

and Construction-Based (Constructivist) Accounts of

Language Acquisition

1.0 Introduction: Generativist and Constructivist Approaches to the141 Acquisition of Non-subject Wh- Questions

1.1 Testing generativist and constructivist accounts143

1.2 Generativist accounts of non-subject wh- question formation and144

their predictions

1.2.1 Wh- operator-specific approaches (DeVilliers, 1991;144

Valian et al., 1992)

1.2.1.1 De Villiers’ (1991) adjunct analysis144

1.2.1.2 Valian et al.’s (1992) optional inversion rule.145

1.2.2 Auxiliary-specific approaches (Stromswold, 1990;146

Santelmann et al. 2002)

1.3 A constructivist account of wh- question formation and149

its predictions: A lexical learning (wh- operator + lexical auxiliary

subtype- specific) approach (Rowland & Pine, 2000)

1.4 Summary of the predictions of the different accounts151

2.0 Method152

2.1 Participants152

2.2 Materials154

2.3 Design154

2.4 Procedure155

2.5 Scoring159

3.0 Results and Discussion.160

3.1 Uninversion errors162

3.1.1 Uninversion errors by wh- operator163

3.1.2 Uninversion errors by auxiliary164

3.1.3 Uninversion errors by wh- operator + auxiliarycombination165

3.1.4 Uninversion errors by lexical auxiliary subtype166

3.2 Correct questions168

3.2.1 Correct questions by wh- operator168

3.2.2 Correct questions by auxiliary169

3.2.3 Correct questions by number170

3.2.4 Correct questions by wh- operator + auxiliary combination171

3.2.5 Correct questions by wh- operator + lexical auxiliary172

subtype combination (i.e., by wh- operator x auxiliary x number)

3.2.6 Correct questions and the role of input frequency174

4.0 General Discussion179

4.1 Generativist accounts179

4.2 Constructivist accounts180

4.3 Comparing experimental and naturalistic data181

4.3.1 Wh- operator-specific approaches182

4.3.2 Auxiliary-specific approaches183

4.3.3 The constructivist approach186

4.3.4 Comparing naturalistic and experimental data: Conclusion186

5.0 Conclusion187

Chapter 4: Experiments 2 & 3. The Formation of Abstract189

Syntactic Construction Schemas: An experimental investigation

of the effects of temporally distributed input and verb type

frequency

1.0 Experiment 2: Formation of Partially Abstract Construction:189

Massed vs distributed pairs

1.1 Introduction: The distributed learning effect189

1.1.1 Temporally distributed presentations of instantiations193

of a grammatical construction: Help or hindrance?

1.2. Method196

1.2.1 Participants196

1.2.2 Materials197

1.2.3 Design and procedure198

1.2.4 Scoring201

1.2.5 Inter-rater reliability203

1.3 Results204

1.3.1 Analysis of target responses204

1.3.2 Analysis of non-target utterances206

1.3.3 Analysis of training schedule207

2.0 Experiment 3: Formation of a Partially Abstract Construction:208

Massed vs distributed pairs vs distributed x type frequency

2.1 Introduction208

2.1.1 The role of type frequency in the formation of abstract209

constructions

2.2 Method213

2.2.1 Participants213

2.2.2 Design213

2.2.3 Procedure214

2.3 Results214 2.3.1 Analysis of target responses 215

2.3.2 Analysis of non-target utterances217

2.3.3 Analysis of training schedule218

3.0 Experiments 2 & 3: Discussion218

Chapter 5: Experiments 4-6. Restricting Linguistic223

Generalizations: An Experimental Investigation of the

Entrenchment Hypothesis

1.0 Introduction: The Formation and Restriction of Linguistic223

Generalizations

1.1 The no-negative-evidence problem225

1.2 Early attempted solutions and their limitations226

1.2.1 Implicit negative evidence226

1.2.2 Innate constraints228

1.2.3 Principles of UG229

1.3 More successful proposals232

1.3.1 Entrenchment232

1.3.2 Pre-emption239

1.3.3 Problems for a pre-emption account242

1.3.4 The formation of semantic verb classes245

1.3.5 Entrenchment, pre-emption and the formation of247

(semantic) verb classes: ‘Three sides of the same coin’?

2.0 Experiment 4: Investigating the Entrenchment Hypothesis252

2.1 Introduction252

2.2 Method253

2.2.1 Participants253

2.2.2 Materials253

2.2.3 Design and procedure254

2.2.4 Scoring256

2.3 Results258

2.4 Discussion258

3.0 Experiment 5259

3.1 Method259

3.1.1 Participants259

3.1.2 Materials, design and procedure259

3.2 Results260

3.3 Discussion261

4.0 Experiment 6262

4.1 Method262

4.1.1 Participants262

4.1.2 Materials263

4.1.3 Design and procedure263

4.2Results267

4.3 Discussion268

5.0 Conclusion271

Chapter 6: Discussion273

1.0 Experiment 1: Wh- Questions274

1.1 Conclusions and theoretical implications274

1.2 Potential problems and refinements276

1.3 Additional future work283

2.0 Experiments 2 & 3: Distributed Learning and the Formation285

of an Abstract Construction Schema

2.1 Conclusions and theoretical implications285

2.1.1 Implications for maturational and other 286

generativist accounts

2.1.2 Implications for a construction-based account of 288

language acquisition and the effect of construction

token frequency

2.1.3 Implications for theories of word-learning290

2.1.4 Implications for a construction conspiracy account 291

of language acquisition (Abbot-Smith & Behrens, submitted):

Children’s production of non-target constructions

2.1.5 Implications of the null effect for verb type frequency293

2.1.6 Implications for the wider distributed learning literature295

2.2 Practical and methodological Implications298

2.3 Potential problems and refinements300

2.4 Additional future work304

3.0 Experiments 4-6: Restricting Linguistic Generalizations: 306

The entrenchment hypothesis

3.1 Conclusions and theoretical implications306

3.2 Refinements and future work307

4.0 Conclusion310

References315

Appendices331

A. Experimenter prompts for Experiment 1 (Wh- questions)332

B. Verbs used in Experiments 2 and 3 (Distributed learning) and 333

frequencies in the British National Corpus (spoken texts section)

C. Scoring criteria for Experiments 2 and 3334

D. Sample parent’s letter (Experiment 1)335

E. Sample parent information sheet (Experiment 1)336

LIST OF TABLES

Table 3.1 Predictions Made by Different Theories of Non-subject153

Wh- Question Acquisition

Table 3.2. Sample Experimenter Prompts Illustrating the Use157

of the Appropriate Pronominal Subject, Wh- Operator and

Lexical Auxiliary form in Uninverted Order

Table 3.3 Proportion of Correct and Erroneous Questions by161

Wh- Operator + Auxiliary + Number Combination, and

Corresponding Standard Deviations

Table 3.4 Significantly Different Uninversion Rates for Particular165

Wh- Operator + Auxiliary Combinations

Table 3.5Significantly Different Uninversion Rates for Particular167

Lexical Auxiliary Forms (auxiliary + number combinations)

Table 3.6 Significantly Different Rates of Correct Question172

Production for Particular Wh- Operator + Auxiliary Combinations

Table 3.7 Significantly Different Rates of Correct Question 173

Production for Individual Wh- Operator + Lexical Auxiliary

Forms (i.e., wh- operator + auxiliary + number combinations)

Table 3.8 Frequency of Each Wh- Operator + Lexical Auxiliary177

Combination in the Sample Input Corpus, and Number of

Correct Questions and Uninversion Errors for Each Combination

in the Experimental Study

Table 3.9 Proportion of Utterances Using Contracted Form of184

Auxiliary is

Table 4.1Experiment 2, Number of Children in Each Experimental 206

Group Producing at Least one Novel Utterance Using a Target

Object Cleft Construction

Table 4.2 Experiment 2. Mean Proportions of Non-Target Utterances207

as a Function of Each Child’s Total Number of Utterances

Table 4.3 Experiment 3. Mean Proportion of Target Object Cleft 215

Utterances as a Function of Each Child’s Total Number of Utterances

Table 4.4Experiment 3. Number of Children in Each Training217

Schedule Group Producing at Least One Novel Utterance

Using a Target Object Cleft Construction (collapsed across verb types)

Table 4.5Experiment 3. Mean Proportions of Non-Target 218

Utterances as a Function of Each Child’s Total Number of Utterances

Table 5.1 Experiment 4. Mean Number of Productive Transitive258

Utterances Using the Novel Verb

Table 5.2 Experiment 5. Mean Number of Productive Transitive 261

Utterances Using the Novel Verb

Table 5.3Training procedure for Experiment 6266

Table 5.4 Experiment 6. Mean Number of Productive Transitive 267

Utterances Using the Novel Verb (max=2)

Table 5.5 Mean Number of Utterances for Each of the Non-Target268

Categories (max = 2). Standard deviations are shown in brackets

LIST OF FIGURES

Figure 3.1 Uninversion rates (as a proportion of all Reponses) 163

by wh- operator

Figure 3.2 Uninversion rates (as a proportion of all responses) 164

by auxiliary

Figure 3.3 Uninversion rates (as a proportion of all responses) 165

for each wh- operator + auxiliary combination

Figure 3.4. Uninversion rates (as a proportion of all responses)167

for each lexical auxiliary form (auxiliary + number combination)

Figure 3.5. Correct questions (as a proportion of all responses)169

by wh- operator

Figure 3.6. Correct questions (as a proportion of all responses) 170

by auxiliary

Figure 3.7 Correct questions (as a proportion of all responses) 171

by wh- operator + auxiliary combination

Figure 4.1. Experiment 2. Mean proportion of target object cleft 205

utterances as a function of each child’s total number of utterances

by age-group and condition

Figure 4.2. Experiment 3. Mean proportion of target object cleft 216

utterances as a function of each child’s total number of utterances

by condition

ABSTRACT

Under traditional generativist accounts, children acquire language (a system of formal rules acting on variables such as NOUN, VERB and TENSE) with the help of some innate knowledge of syntax. Recently, these generativist accounts (e.g., Pinker, 1989; Radford, 1990; Wexler, 1998) have been challenged by functionalist accounts (e.g., Pine, Lieven & Rowland, 1998; Bybee, 1995, Bates & Goodman, 2001) under which children acquire an inventory of meaningful chunks of linguistic material of various sizes, that become increasingly abstract as development proceeds (e.g., I want X [SUBJECT] [VERB] [OBJECT]). Tomasello (2003) draws together many different strands of research to present a relatively complete constructivist account of language acquisition. The goals of this thesis are (1) to test the predictions of this account, and competing generativist accounts; and (2) to investigate aspects of this account that currently remain somewhat underspecified.

Chapters 1 and 2 outline generativist and constructivist accounts of language acquisition respectively, and present evidence in support of the claim that only constructivist approaches can potentially explain the pattern of child language acquisition observed.

Experiment 1 (Chapter 3) tested the predictions of these two approaches with respect to children’s acquisition of non-subject wh- questions (e.g., Who is Mickey hitting?). Questions using each of 4 wh- operators (what, who, how and why), and 4 auxiliaries (copula BE, auxiliary BE, DO and CAN) in 3sg and 3pl form were elicited from 28 children aged 3;6-4;6. Generativist theories claim that uninversion errors (e.g., Who Mickey is hitting?) will pattern by wh- operator (De Villiers, 1991; Valian et al., 1992) or auxiliary (Stromswold, 1990; Santelmann et al., 2002). Although errors did show some tendency to pattern by auxiliary, interactions between the variables of wh- operator, auxiliary and number suggest that Rowland and Pine’s (2000) constructivist model, under which children acquire frequent wh- operator+lexical auxiliary combinations from the input, can potentially provide the best fit for the data.

Experiments 2 and 3 (Chapter 4) investigated two factors thought to influence the process by which children form abstract grammatical constructions: (1) temporal distribution of instantiations of the construction and (2) type frequency of the variable element in the construction. 48 children aged 3;6-5;10 and 72 children aged 4;0-5;0 were given 10 exposures to theconstruction it was the [OBJECT] that the [SUBJECT] [VERB]ed all in one session (massed), or on a schedule of 2 trials per day for 5 days(distributed pairs), or 1 trial per day for 10 days (distributed). Children in both the distributed conditions learned the construction better than children inthe massed condition, as evidenced by productive use of this construction with a verb that had not been presented during training, though a VERB type frequency manipulation was found to have no effect.

Experiments 4-6 investigated a specific aspect of Tomsello’s account: the hypothesis that repeated presentation of a particular verb (e.g., kick) in a particular argument structure construction (e.g., John kicked the ball) leads to the inference that the use of that verb in non-attested constructions (e.g., *the ball kicked) is not permitted (the entrenchment hypothesis). These studies did not demonstrate an entrenchment effect, but remain a work in progress.

In Chapter 6, I conclude that the findings of Experiments 1-6 are broadly consistent with Tomasello’s (2003) account, but argue that specific aspects of the constructivist account require much more detailed investigation, and present several suggestions as to how this might be accomplished.

DECLARATION

I declare that no portion of the work referred to in this thesis has been submitted in support of an application for another degree or qualification of this or any other university or other institute of learning

COPYRIGHT STATEMENT

(1)Copyright in text of this thesis rests with the Author. Copies (by any process) either in full, or of extracts, may be made only in accordance with instructions given by the Author and lodged in the John Rylands University Library of Manchester. Details may be obtained from the Librarian. This page must form part of any such copies made. Further copies (by any process) of copies made in accordance with such instructions may not be made without the permission (in writing) of the Author.

(2)The ownership of any intellectual property rights which may be described in this thesis is vested in the University of Manchester, subject to any prior agreement to the contrary, and may not be made available for use by third parties without the written permission of the University, which will prescribe the terms and conditions of any such agreement.

(3)Further information on the conditions under which disclosures and exploitation may take place is available from the Head of the Department of Psychology

ACKNOWLEDGEMENTS

My greatest thanks are, of course, due to my supervisors, Professor Elena Lieven, Dr. Anna Theakston and Professor Mike Tomasello, who have made this thesis possible.

I would also like to thank my fellow students in Manchester, both past and present for their many helpful discussions. They are Rob Maslen, Danielle Matthews, Ceri Savage and Thea Cameron-Faulkner. Several members of the Max Planck Institute in Leipzig have also provided help with methodology and statistics, in particular Daniel Stahl, Franklin Chang and Kirsten Abbot-Smith.

Thanks must go also to my long-suffering co-experimenters for Experiments 2 and 3: primarily Victoria Hulks and Ellie O’ Malley but also (on occasions) Anna Roby and Evan Kidd. Thank you, and may you never have to hear that construction again in your lifetime!

I have enjoyed being a (somewhat occasional) member of the Chester reading group, and would like to single out for particular thanks Virginia Gathercole, who made a helpful methodological contribution to Experiments 4-6, and Bill Croft, who has radically changed the way I see language.

Special thanks go to Caroline Rowland, for help with the wh- questions literature, and Julian Pine, whose lectures at the University of Nottingham inspired me to undertake this doctorate. Thanks to you both, and I look forward to working with you in Liverpool.

Finally, I thank the University of Manchester and the Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany. This research was supported by a studentship from the Max Planck Institute for Evolutionary Anthropology.

Chapter 1: Generativist Approaches to Language Acquisition

1.0 Thesis Introduction and Outline

Do children learn their native language, or can we only say that they acquire it? One’s answer to this question seems to be inextricably linked to one’s view of precisely what language is. Under the generativist (nativist, UG, formalist) view, a language consists of a lexicon – analogous to a mental dictionary – and a grammar; an abstract, infinite, recursive, formal system, analogous to that of propositional logic. Those who favour this view of language generally believe that the grammar cannot be learned from a finite set of utterances generated by it, and that language can only be acquired with the help of some innate knowledge of syntax. Under the constructivist (socio-pragmatic, functionalist, usage-based) view, which draws no sharp distinction between the grammar and the lexicon, language is an inventory of meaningful chunks of linguistic material of various sizes, and various levels of abstraction, which serve some communicative or socio-pragmatic function. Those who favour this view generally believe that humans’ general cognitive abilities can be used to learn units of meaningful material, and to abstract across and generate links between them.

In the present thesis, having argued that the generativist position is untenable, I present a series of six experiments designed to test and extend an alternative constructivist account. The structure of this thesis is as follows.

In Chapter 1, five leading generativist theories of language acquisition[1] (Radford, 1990,1996; Valian, 1991; Wexler, 1998, Hyams, 1999; Pinker, 1984, 1989) are outlined and evaluated. It is argued that none of these theories can explain the pattern of child language acquisition data found in either naturalistic or experimental studies.

Chapter 2 presents a recent constructivist account (Tomasello, 2003) and argues that such an approach is, at least potentially, compatible with the available data. The remainder of the thesis then endeavours to test and extend this theory.

In Chapter 3 I report the results of an experiment (Experiment 1) designed to mediate between generativist and constructivist accounts of language acquisition with respect to children’s acquisition of non-subject wh- questions. These structures represent a particularly good test case for both constructivist and (especially) generativist accounts, as both make specific predictions with regard to the pattern of acquisition and errors. It is argued that the pattern of results observed is compatible with only the constructivist approach.

Having provided a further demonstration that constructivist approaches are most compatible with the data, I then return to Tomasello’s (2003) specific account and seek to extend this account by investigating, in some detail, the nature of the processes by which the abstract grammatical constructions that are taken to underlie adult linguistic competence may be (1) abstracted from the input (Chapter 4) and (2) appropriately restricted, in order to avoid overgeneralization errors (Chapter 5).

Experiments 2 and 3 (Chapter 4) investigate two factors which are hypothesised to influence the formation of abstract syntactic construction schemas: (1) the temporal distribution of substantive instantiations of the construction in the input and (2) the type frequency of the variable element in the construction, in this case the verb. This study shows a distributed learning effect such that acquisition of the construction is facilitated by temporally distributed (over several days) as opposed to massed (in one session) exposure. It is argued that this pattern of results is compatible with Tomasello’s (2003) approach, under which grammatical constructions, words and non-linguistic stimuli are all acquired via domain-general cognitive processes, but not with domain-specific generativist accounts under which the grammar, but not the lexicon, is acquired with the help of processes such as parameter setting and biologically determined maturation.