To be or not to be adultlike in syntax: An experimental study of language acquisition and processing in children

Thesis

Reference

To be or not to be adultlike in syntax: An experimental study of language acquisition and processing in children

LASSOTTA, Romy

Abstract

Children manifest difficulties in sentence comprehension, which may stem from an immature competence system and/or an immature performance system. This thesis examines children's syntactic knowledge (competence system), as well as processing abilities (performance system) involved in sentence comprehension. A series of six experimental studies with French-speaking children and adults were conducted. The investigations evidenced adultlike, rule-based syntactic knowledge of word order and movement before age 2. However, the syntactic processing mechanisms of incremental parsing and revision appeared to develop beyond age 5–6, reaching adultlike performance only around age 7–8. Importantly, domain-general cognitive control abilities, namely inhibition and working memory updating, were also found to be linked to comprehension performance in both children and adults. In conclusion, the results are interpreted as evidence for a child-adult continuity in terms of syntactic competence and comprehension difficulties are attributed to children's linguistic and non-linguistic processing abilities under development.

LASSOTTA, Romy. To be or not to be adultlike in syntax: An experimental study of language acquisition and processing in children. Thèse de doctorat : Univ. Genève, 2021, no. FPSE 756

DOI : 10.13097/archive-ouverte/unige:150402 URN : urn:nbn:ch:unige-1504025

Available at:

http://archive-ouverte.unige.ch/unige:150402

Disclaimer: layout of this document may differ from the published version.

1 / 1

Section de Psychologie

Sous la direction de Dr. Julie Franck

TO BE OR NOT TO BE ADULTLIKE IN SYNTAX:

AN EXPERIMENTAL STUDY OF LANGUAGE ACQUISITION AND PROCESSING IN CHILDREN

THESE

Présentée à la

Faculté de Psychologie et des Sciences de l’Éducation de l’Université de Genève

pour obtenir le grade de Docteure en Psychologie par

Romy LASSOTTA de

Allemagne

Thèse N^o 756 N^o étudiant 04-302-147

GENEVE Février 2021

JURY

Dr. Julie Franck (Université de Genève)

Prof. Anna Gavarró (Universitat Autònoma de Barcelona) Prof. Edouard Gentaz (Université de Genève)

Prof. Pascal Zesiger (Université de Genève)

SUMMARY

Children acquire their native language grammar with impressive speed and apparent ease and yet, on their developmental pathway towards an adultlike language proficiency level, they manifest difficulties in interpreting certain complex sentences. The central aim of the present dissertation is to understand the origins of these difficulties. One possible explanation of children’s misinterpretations is an immature competence system, suggesting that children lack adultlike syntactic representations (grammar). Another possibility is that misinterpretations result from an immature performance system (use of grammar), implying that the observed difficulties are related to processing limitations. Two types of factors are assumed to be at play in the process of sentence comprehension: factors within the language domain, like syntactic processing abilities, and domain-general factors, like cognitive control abilities. Cognitive skills develop until adolescence and could thus account for differences between child and adult comprehension performance.

This dissertation examines the nature of children’s syntactic representations underlying two key grammatical phenomena: word order and movement. A series of six experimental studies with French-speaking children and adults were conducted. Our investigations evidenced adultlike abstract representations of word order before age 2 and adultlike incremental parsing of ambiguous wh-questions involving movement and thus requiring to parse long-distance dependencies in 5–8-year-old children. Nevertheless, it also uncovered divergencies in their processing of syntactic revision as compared to the adults’, linked to their ability to integrate language’s distributional information in their parsing mechanisms, and in the maturity of some cognitive control functions from outside the language domain.

In conclusion, the results are interpreted as evidence of child-adult continuity in terms of syntactic competence (grammar) and child-adult divergencies are attributed to children’s processing abilities under development.

DEDICATION

In Gedenken an meinen Vater.

CREDITS

DIRECTOR Julie Franck SCRIPT SUPERVISORS Pascal Zesiger

Anna Gavarró Edouard Gentaz FOCUS-PULLER Akira Omaki

Severine Millotte Luigi Rizzi Jeffrey Lidz Colin Phillips

Catherine Walter-Laager PRODUCER Fonds National Suisse

Fondation Ernst et Lucie Schmidheiny Société Académique de Genève

CAST Bachelor students in Psychology & Speech Therapy Children and their parents

Teachers and educators SWING GANG Sandra Villata

Irene Rotondi Natalia Kartushina Anamaria Bentea Andrea Valente Jennifer Martin ANIMATOR Daniele Panizza

CUT Atelier Multimédia FPSE SPECIAL EFFECTS Svantje Titulski

Andres Posada Tom Fritsche Aaron Apple Equipe MAD Anke Naumann MAKE-UP Karin Pogade

Myriam Schallert THANK YOU ALL FOR INSPIRING, GUIDING,

CHALLENGING, ENCOURAGING, AND SUPPORTING ME ON MY WAY TOWARDS THIS PREMIERE!

BEST BOY Philipp Höhne BEST GIRL Wilma Lassotta BOOM OPERATORS Caroline Dombre

Laura Vuillequez Anne Ruef Diane Jousson Jeannette Molnar Jane Jöhr

Annett Kruschel Robert Kruschel Susanne Furchner Gretel Fuchs Vanessa Gallitz Markus Kuhn Bianca Lindner Katrin Drache Katrin Bünger Oliver Riemer Andreas Balke Tina Welz Ina Kuhnert Vicky Leightiger Jana Teuscher Evelyne Lamanque Marina Jahn Rebecca Nattefort Jasmin Luthard AUTHORS Martina Lassotta

Helmut Lassotta Gerda Schubert Andrea Lassotta Familie Berthold Familie Schink Familie Leuteritz Familie Kessinger Chantal Drapeau

“The greatest power of life is gratitude.”

—Hermann Bezzel

PUBLICATIONS AND PERSONAL CONTRIBUTION

During my PhD, I conducted a critical analysis of four Weird Word Order studies (Section 2.1), a corpus analysis (Section 2.2), and the six experiments of this dissertation, which comprised designing the paradigms, creating the material, recruiting the participants and running the experiments themselves, as well as coding and analyzing the data. I also participated in the realization of four other experiments, which go beyond the scope of this dissertation: a preferential looking paradigm testing 19–32-month-old Japanese-learners’ word order knowledge and a corpus study of Japanese child-directed speech (in collaboration with Julie Franck, Tessei Kobayashi, Akira Omaki, and Luigi Rizzi); a study examining the knowledge of the abstract grammatical principle of C-command in French-speaking 30-month-olds (in collaboration with Anamaria Bentea, Julie Franck, Akira Omaki, and Sandra Villata), and an experiment testing complement ambiguity resolution in 5–8-year-old English-speakers (in collaboration with Julie Gerard and Jeffrey Lidz). My publications that are relevant to the focus of this dissertation are listed below.

Peer-reviewed journal articles integrated in this dissertation

Lassotta, R., Omaki, A., & Franck, J. (2021). Syntactic revision in wh-questions: Developmental trajectories and the role of cognitive control [Unpublished manuscript]. Faculté de Psychologie et des Sciences de l’Education, Université de Genève.

Lassotta, R., Omaki, A., & Franck, J. (2020). Abstract knowledge of canonical and non-canonical word order before age two: An eye-tracking study. [Unpublished manuscript]. Faculté de Psychologie et des Sciences de l’Education, Université de Genève.

Lassotta, R., Omaki, A., & Franck, J. (2016). Developmental changes in misinterpretation of garden-path wh-questions in French. Quarterly Journal of Experimental Psychology, 69(5), 829–854. https://doi.org/10.1080/17470218.2015.1054845

Franck, J., & Lassotta, R. (2012). Revisiting evidence for lexicalized word order in young children. Lingua, 122(1), 92–106. https://doi.org/10.1016/j.lingua.2011.11.010

Additional peer-reviewed journal article

Franck, J., Millotte, S., & Lassotta, R. (2011). Early word order representations: Novel arguments against old contradictions. Language Acquisition, 18(2), 121–135.

https://doi.org/10.1080/10489223.2011.530536

Conference talks

Lassotta, R., Omaki, A., & Franck, J. (2014, November 9). Abstract knowledge of non-canonical word order by age 2 [Talk]. Boston University Conference on Language Development, Boston, MA, United States.

Lassotta, R., Omaki, A., Panizza, D., Villata, S., & Franck, J. (2013, July 25). Garden-path recovery and cognitive control in children and adults: Evidence from French wh-questions with filled-gaps [Talk]. International Congress of Linguists, Geneva, Switzerland.

Lassotta, R., Omaki, A., Panizza, D., Villata, S., & Franck, J. (2012, September 6). Linking Cognitive Control to Revision of Garden-Path wh-Questions in Adults and Children [Talk]. Architectures and Mechanisms for Language Processing Conference, Riva del Garda, Italy.

Conference posters

Lassotta, R., Omaki, A., Panizza, D., Villata, S., & Franck, J. (2012, November 3). Sentence revision and executive control in French-speaking children and adults: Evidence from wh-questions with filled-gaps [Poster presentation]. Boston University Conference on Language Development, Boston, MA, United States.

Lassotta, R., Omaki, A., Panizza, D., Villata, S., & Franck, J. (2012, September 19). Linking garden-path recovery and cognitive control in adults and children: Evidence from French wh-questions [Poster presentation]. Cognitive Modules & Interfaces at SISSA, Trieste, Italy.

Omaki, A., Kobayashi, T., Lassotta, R., Rizzi, L., & Franck, J. (2012, September 6). Input Effects on Parser Development: Evidence from Japanese Word Order Development [Poster presentation]. AMLaP, Paris, France.

Omaki, A., Lassotta, R., Kobayashi, T., Rizzi, L., & Franck, J. (2012, August 2). Delay of word order development in Japanese? Evidence from a preferential looking study with 19 and 30-month-old children [Poster presentation]. Annual meeting of the Cognitive Science Society, Sapporo, Japan.

Omaki, A., Lassotta, R., & Franck, J. (2011). Sentence revision difficulties in French-speaking children and adults: Evidence from wh-question with filled-gaps [Poster presentation].

AMLaP, Paris, France.

Franck, J., Lassotta, R., Omaki, A., & Rizzi, L. (2011, April 15). Sentence processing in 19-month-olds: The role of abstract word order representations [Poster presentation].

10th Symposium of Psycholinguistics at BCBL, San Sebastian, Spain.

ETHICAL CONSIDERATIONS

Both the syntactic representations project (Chapter 2) and the syntactic processing project (Chapter 3) have been approved by the ethical committee of the Faculté de Psychologie et des Sciences de l’Éducation of the Université de Genève. For the word order experiments (Section 2.2), infants were recruited by contacting families on the basis of the listings from the

“Office Cantonal de la Population”, registering all births in Geneva. Informed consent had been obtained from each family prior to the experiment. For the wh-question experiments (Sections 3.1 and 3.2), the Département d’Instruction Publique de Genève authorized us to conduct our studies at primary schools of Geneva. Teachers and families were informed about the aim and the modalities of the experiment and parents provided informed consent prior to the testing of their children. Adult participants were recruited among bachelor students enrolled in the Psycholinguistics course at the Université de Genève and their informed consent has been obtained before the beginning of the experiment.

TABLE OF CONTENTS

LIST OF TABLES ... 19

LIST OF FIGURES ... 21

CHAPTER 1: SEEKING GROUND ... 23

1.1 INTRODUCTION ... 25

1.2 RESEARCH AIMS ... 26

1.3 ROADMAP ... 28

CHAPTER 2: SYNTACTIC REPRESENTATIONS OF WORD ORDER ... 31

2.1 Revisiting evidence for lexicalized word order in young children ... 33

2.1.1 Introduction ... 33

2.1.2 The debate ... 33

2.1.3 From theory to data: critical issues with empirical predictions ... 36

2.1.4 The problem of missing data: Critical issues with data analysis ... 40

2.1.5 From data to theory: Critical issues with data interpretation ... 45

2.1.6 General discussion ... 53

2.1.7 Conclusion ... 60

2.2 Abstract knowledge of canonical and non-canonical word order before age two: An eye-tracking study ... 63

2.2.1 Introduction ... 63

2.2.2 Study 1: Canonical word order ... 67

2.2.3 Study 2: Non-canonical word order ... 73

2.2.4 Study 3: Corpus analysis ... 79

2.2.5 General discussion ... 84

CHAPTER 3: SYNTACTIC PROCESSING OF WH-QUESTIONS ... 93

3.1 Developmental changes in misinterpretation of garden-path wh-questions in French ... 95

3.1.1 Introduction ... 95

3.1.2 Experiment 1: Adjunct wh-questions ... 100

3.1.3 Experiment 2: Argument wh-questions ... 108

3.1.4 Experiment 3: Acceptability judgment ... 117

3.1.5 General discussion ... 123

3.2 Syntactic revision in wh-questions: Developmental pathways and the role of cognitive control ... 129

3.2.1 Introduction ... 129

3.2.2 Current study ... 135

3.2.3 Experiment ... 136

3.2.4 Discussion ... 150

3.2.5 Conclusion ... 157

CHAPTER 4: FINDING CLOSURE ... 159

4.1 Summary of findings and opening perspectives ... 161

4.1.1 Novel findings about syntactic representations ... 161

4.1.2 Novel findings about syntactic processing ... 169

4.2 FINAL CONCLUSION ... 176

REFERENCES ... 182

APPENDIX A ... 204

APPENDIX B ... 205

APPENDIX C ... 207

APPENDIX D ... 208

APPENDIX E ... 219

LIST OF ABBREVIATIONS

DLD Developmental Language Disorder

NP Noun Phrase

OSV Object - Subject - Verb OVS Object - Verb - Subject

PLA Proportion of Looks to Agent-first scene PP Prepositional Phrase

S1S2V first Subject - second Subject - Verb S1VS2 first Subject - Verb - second Subject SOV Subject - Object - Verb

SVO Subject - Verb - Object

VP Verb Phrase

VSO Verb - Subject - Object WO Word Order

WWO Weird Word Order

LIST OF TABLES

Table 1 Summary of the methods used in the four studies under review 37 Table 2 Predictions of the lexical hypothesis in the two age groups 38 Table 3 Distribution of Match responses in the ungrammatical conditions

(based on data reported in Matthews et al., 2007) 45 Table 4 Distribution of Match and Mismatch responses in ungrammatical

conditions involving novel verbs across the various studies 46 Table 5 Mean number of produced words by participants of Study 1 and 2 68

Table 6 Trial procedure in Study 1 70

Table 7 Proportion of looks towards the agent-first video with canonical NVN

sentences in Study 1 71

Table 8 Test trial procedure in Study 2 76

Table 9 Proportion of looks towards the agent-first video with canonical NVN

and non-canonical NNproV sentences in Study 2 77 Table 10 Distribution of the offline responses in Experiment 1 106 Table 11 Model of embedded-verb response in Experiment 1 106 Table 12 Distribution of the offline responses in Experiment 2 112 Table 13 Model of empirical logit transformed embedded-verb responses in

Experiment 2 113

Table 14 Response times in Experiment 2 114

Table 15 Models of log-transformed response times in Experiment 2 114 Table 16 Examples of questions in the eight experimental conditions of

Experiment 3 119

Table 17 Models of z-transformed acceptability judgments in Experiment 3 121 Table 18 Means of cognitive control indexes of DCCS, N-back and Corsi tasks

for each age group 148

Table 19 Partial correlations between EV increase and N-back, DCCS, and

Corsi task indexes (controlled for age) 150

Table 20 Overview of rates of active gap filling, misparsing and revision (in %) in studies on wh-attachment in French bi-clausal wh-questions

(Lassotta et al., 2016, and current study), depending on main clause

argument structure and main verb semantics 152

Table A1 Test sentences in Study 1 204

Table A2 Test sentences in Study 2 204

Table E1 Correlations between EV increase and N-back, DCCS and Corsi

task indexes 219

LIST OF FIGURES

Figure 1 Distribution of the raw frequencies of Match and Mismatch responses in the low frequency condition in Matthews et al. (2005) 43 Figure 2 Distribution of the mean frequencies of Match and Mismatch responses

in the low frequency condition in Matthews et al. (2005), calculated on the subset of responses containing a test verb 43 Figure 3 Baseline-corrected proportion of looks to the agent-first video with

canonical NVN sentences in Study 1 72

Figure 4 Baseline-corrected proportion of looks to the agent-first video with

canonical NVN and non-canonical NNproV sentences in Study 2 78 Figure 5 Distribution of sentence types in French child-directed speech 82 Figure 6 A sample scene presented during a story stimulus in Experiment 1 104 Figure 7 Mean z-transformed acceptability judgments in Experiment 3 120 Figure 8 Example of story pictures in the wh-question task 139 Figure 9 Example of fruit sequence in the N-back task 140 Figure 10 Example of 2-back detection in the familiarization step V of the

N-back task 142

Figure 11 Example of (A) congruent and (B) incongruent trials in the DCCS task 143 Figure 12 Proportions of EV responses in the wh-question task by question type

(ambiguous vs. filled-gap) and age group (5–6-year-olds vs. 7–8-year-olds

vs. adults) 147

C ^HAPTER 1:

S EEKING GROUND

“An idea that is not dangerous is not worthy of being called an idea at all.”

—Oscar Wilde

1.1 I

For a long time, babies have been thought to be born as “blank pages” to write on and were perceived as being rather “ignorant” compared to adults’ outstanding cognitive skills, perceiving the world “as one great blooming, buzzing confusion” (James, 1890, p. 462). Luckily and importantly though, this conception has been challenged by an accumulation of empirical evidence from psychological and psycholinguistic research. One astonishing observation is the fact that infants can perfectly acquire a new language within only a few years without any apparent effort. Therefore, infants must dispose of an incredibly powerful faculty allowing this impressive development. Rich empirical evidence is available in the field of psycholinguistics concerning the acquisition of prosodic, phonological, lexical, and syntactic knowledge and processing. However, contradicting views about the acquisition of syntax give rise to two major debates.

First, the nature of children’s early syntactic knowledge is under debate. Do young children around age 2 already develop adultlike abstract symbolic representations independent of lexical units (grammatical approach; e.g., Franck et al., 2011; Gertner et al., 2006)? Or do they rather accumulate exemplar-based, lexically-stored, item-specific structures that they have encountered in the input and that gradually generalize into adultlike abstract representations only later in development, around age 4 (constructivist, usage-based approach; e.g., Abbot-Smith et al., 2001; Ambridge, 2020; Tomasello, 2000)?

Second, the interpretation of children’s occasional non adultlike performance in sentence processing is a strongly debated topic. Do these differences reflect fundamental discontinuities with the adult syntactic parser? Or can they be attributed to divergencies in the weighting of parsing strategies and/or in cognitive functions external to the linguistic functions, but which are nevertheless required during language processing (e.g., working memory and cognitive control;

Mazuka et al., 2009)?

Both debates hinge on the more general question of the relationship between children’s and adults’ linguistic competence. The working hypothesis underlying this dissertation is the idea that children’s syntax is fundamentally in continuity with the adults’.

1.2 R

The present work aims to provide new theoretical and empirical arguments on the debates about (1) the nature of early syntactic representations and (2) the interpretation of children’s difficulties in sentence processing. For this purpose, we conduct a critical analysis of existing findings as well as novel experimental studies of the acquisition of two fundamental and intrinsically related properties of natural languages: word order and movement.

The word order of a language determines the specific position of syntactic constituents relative to each other, and more particularly of heads relative to their complements. Although languages vary, some placing complements before their heads and others placing them after, head-complement order is strikingly stable throughout the paradigm of constituents within a language. This property of the language is one of the earliest aspects of syntax that children acquire since, as soon as they start combining words around 20 months of age, their productions respect the canonical word order of the language (e.g., Bloom, 1991). And accurate comprehension of sentences displaying the language’s canonical word order has been observed even earlier, at as early as 14 months (Hirsh-Pasek & Golinkoff, 1996). Yet, it is debated if children’s knowledge underlying word order representation is of an abstract format as in adults (e.g., The cow is kissing the donkey would be represented by way of abstract variables like subject, verb, and object) or whether it is represented as item-specific information (e.g., word order would be tied to lexical knowledge of the verb kissing) during the first years of life.

Movement is a grammatical operation enabling to move constituents to non-canonical positions in the sentence, a property of the language that has been argued to be mastered later in development. Indeed, difficulties have been widely reported when children are required to process sentences involving movement, leading to misinterpretations sometimes until age 8 (with non-canonical sentences like object relative clauses, e.g., Show me the donkey that the cow is kissing). However, children’s difficulties with such sentences disappear under some conditions (e.g., Friedmann et al., 2009) and adults are also subject to misinterpretations when processing certain complex sentences (e.g., Christianson et al., 2001). So, the question arises as to why syntactic competence sometimes fails to manifest itself in performance.

We investigate whether children’s syntactic representations are in an abstract format early on (see Chapter 2) and whether their sentence processing relies on similar mechanisms as compared to adults (see Chapter 3). In line with the hypothesis of child-adult continuity in syntax, the observed divergencies in performance could be related to non-linguistic factors, i.e., domain-general cognitive functions. Cognitive control abilities maturing until adolescence

(Davidson et al., 2006) may play a role in children’s non-adultlike linguistic processing. Hence, even if children dispose of adultlike syntactic representations of word order and movement, and their parsing processes are adultlike as well, parsing difficulties may still occur due to factors outside the language module. The ultimate goal of our approach is to reach a more complete and adequate understanding of what appears like a discontinuity with adult syntax and to add new informative pieces to the extraordinary puzzle of language acquisition.

1.3 R

In CHAPTER 2, we first discuss the old but still prevailing controversy of the nature of children’s early syntactic knowledge of word order in a critical review of four of the foundational studies using the Weird Word Order paradigm (Franck & Lassotta, 2012; see Section 2.1). We uncover various problems in terms of predictions, data analysis and interpretation of the results that invalidate the authors’ theoretical conclusion that young children encode item-specific word order knowledge. Moreover, revisiting their methodological and inferential grounds brings evidence, from these studies themselves, for abstract word order representations before age 3.

Next, we report two experimental studies using the intermodal preferential looking paradigm to examine word order representations of 19–22-month-old children as well as a corpus study of child directed speech in French (Lassotta et al., 2020; see Section 2.2). Children correctly interpreted canonical but also non-canonical sentences with pseudo-verbs, even though lexical (verb-specific) information was not available and the non-canonical structures tested (object topicalization with a full nominal subject) were virtually absent from the input. These findings support the hypothesis of children representing word order abstractly in early stages of development, that is before age 2.

In CHAPTER 3, we start by exploring the processing of bi-clausal wh-questions involving movement and requiring syntactic revision, which often leads to misinterpretations, in French-speaking children aged 5–8 years and in adults (Lassotta et al., 2016; see Section 3.1).

The results from two comprehension experiments and an acceptability judgment experiment reveal three key findings. First, we provide evidence of an incremental sentence processing mechanism in children and adults alike. Second, revision difficulties were found in both groups as well. Third, the observed instances of successful revision suggest that adults and children are guided by different types of constraints: while the child parser appears to prioritize syntactic information that comes directly from the error signal in the sentence, the adult parser prioritizes distributional information to resolve the syntactic error and hence to prevent misinterpretations.

Then, we examine the role of non-linguistic cognitive functions in resolving this type of syntactic error in 5–8-year-olds and adults using the same comprehension experiment with bi-clausal wh-questions and a new N-back task measuring cognitive control (Lassotta et al., 2021; see Section 3.2). Although misinterpretations occur in all groups, 7–8-year-olds and adults succeed in revision to a large, similar extent, whereas 5–6-year-olds more often fail to revise. Crucially, revision abilities are linked to performance to the N-back task: participants with stronger cognitive control revise more often than those with weaker cognitive control. Thus, difficulties in sentence

1.3ROADMAP 29 processing are at least partly due to non-linguistic cognitive abilities, hindering the full syntactic competence to display in performance. We conclude that children’s (mis-)interpretations involve the same incremental parsing mechanisms, revision process and non-linguistic skills as in adults.

In CHAPTER 4, we conclude that from early on, child syntax comprises the relevant representations and constraints or “background knowledge” for meaningful interpretations and generalizations to be possible. Such constraints are necessary for language acquisition and processing in that they delimit the infinite set of hypotheses a speaker (or listener) could consider.

And only when taking into account both representations and processing mechanisms, a theory can fully explain human language.

1.00

0.75

0.50

0.25

0.00 0.75

0.50

0.25

0.00

C HAPTER 2:

S YNTACTIC REPRESENTATIONS OF WORD ORDER

“You cannot create experience. You must undergo it.”

—Albert Camus

2.1 R

Franck, J., & Lassotta, R. (2012). Revisiting evidence for lexicalized word order in young children. Lingua, 122(1), 92–106. https://doi.org/10.1016/j.lingua.2011.11.010

2.1.1 Introduction

Language acquisition is not as much of a challenge for the child who is learning to speak without any apparent effort, as it is for the researcher who is trying to discover the nature of this learning process. Despite the growing body of both empirical and theoretical evidence concerning the processes involved in the acquisition of phonological and lexical representations, the acquisition of grammar remains highly controversial. Controversies concern the format of the early grammatical representations that the child is acquiring (whether it is a set of abstract symbolic rules or an accumulation of individual exemplars of particular structures), the role of innate constraints in grammar induction from the input (whether induction is filtered out by internal constraints or operates solely on the basis of the input), and, if such constraints are acknowledged, the nature of these constraints (whether they are specific to language or domain-general, in the form of cognitive or perceptual constraints). This paper focuses on the controversy about the format by which young children represent one major property of the grammar of their language: word order.

2.1.2 The debate

The basic structure of syntactic constituents is made of a head, which constitutes the principal element (e.g., the verb in the verb phrase), its complements (e.g., the object) and specifiers (the subject; Chomsky, 1981). The order of the head and the complement is generally fixed for a given category in a language (i.e., a language has Verb–Object or Object–Verb order).

Moreover, a language tends to choose a single order across categories (Greenberg, 1963), even though coherence holds only statistically, not absolutely across languages. In the Principles and Parameters model, this aspect of variation in word order is captured by the head-direction parameter (e.g., Chomsky, 1995).

When do children develop knowledge of the word order of their language? Acquisition research has shown that word order is already represented when children start to combine words into sentences. Analyses of early sentences, arising around 18–24 months, show that word order mistakes are hardly ever made (e.g., Bloom, 1991; Brown, 1973; Lust & Wakayama, 1981;

Tomasello, 1992). But evidence suggests that this knowledge is actually encoded much earlier.

Head-direction correlates with cues in the speech signal that infants are sensitive to and which were found to play a critical role in the bootstrap of word order. First, head-direction correlates with rhythmic properties of the language due to the fact that within phonological phrases, prominence always falls on the complement. Hence, while prominence falls on the right edge of constituents in head-complement languages, it falls on the left in complement-head languages (Nespor & Vogel, 1986). Babies as early as 3 months old were found to rely on this cue to discriminate between languages with different head-direction (Christophe et al., 2003). Preference for the head-direction’s prosodic patterns of the input language was argued to arise around 14 months, as German infants at that age were found to prefer NV phrases, following German’s canonical word order (SOV), to VN phrases (Hofmann et al., 2003). Second, head-direction correlates with statistical regularities in the distribution of grammatical morphemes in the sentence. Whereas head-complement languages typically have their grammatical morphemes in initial position in the sentence, complement-head languages have them in final position (Gervain et al., 2008). Gervain and colleagues found that 8-month-old Italian infants prefer an artificial language with frequent words (i.e., typically grammatical morphemes) situated in utterance-initial position in the sentences while Japanese infants prefer one with frequent words in final position.

Hence, as early as 8-month-old infants appear to develop knowledge of at least some of the surface properties that correlate with the word order of their mother tongue. All these observations were argued to support the view that acquisition of word order relies on a mechanism of parameter setting (Gibson & Wexler, 1994; Pinker, 1994) by which the infant, on the basis of prosodic and statistical cues present in the input, rapidly develops an abstract representation of head-direction (e.g., Christophe et al., 2003; Gervain et al., 2008). In this view, children develop abstract knowledge of surface properties of the word order of their language before developing a lexicon.

But when do children start to know that syntactic positions map onto thematic roles at the semantic level? The experimental work by Hirsh-Pasek and Golinkoff using the intermodal preferential looking paradigm showed that as early as 14 months, English children interpret the noun phrase following the verb as its object (e.g., Hirsh-Pasek & Golinkoff, 1996). Although this report clearly attests that 14-month-olds already use word order to comprehend sentences, it is insufficient to tease apart two theoretical interpretations as to the format in which word order is

represented. According to what we will refer to as the grammatical hypothesis, children develop abstract knowledge of the mapping between syntactic positions and thematic roles from very early on. This hypothesis is grounded in the account of language acquisition in terms of parameter setting according to which language acquisition amounts to fix the value of a number of universal properties that are subject to parametric variation across languages on the basis of limited input (e.g., Radford, 1990; Clahsen et al., 1996). In this view, child language may only diverge from adult language within the possibilities offered by Universal Grammar. Critical evidence was recently provided by Gertner and colleagues who tested children’s comprehension of sentences involving pseudo-verbs, i.e., verbs that children had never encountered before the experiment (Gertner et al., 2006). In line with the results obtained by Hirsh-Pasek and Golinkoff with known verbs, English children aged 21 months interpreted the argument preceding the pseudo-verb as the agent of the action (e.g., “the duck” in The duck is gorping the bunny) while they interpreted the one following the pseudo-verb as the patient (“the bunny”). Similar results were obtained with French infants aged 19 months (Franck et al., 2013).

According to the lexical hypothesis, children learn the appropriate use of grammatical relations on a verb-by-verb basis. In this view, grounded in the constructivist, usage-based account of grammatical development, word order involves a slow and gradual process of generalization across an important sample of lexically specific examples, on the basis of general inferential mechanisms (e.g., Tomasello, 2000, 2003). An important piece of the empirical foundation of this hypothesis comes from the analysis of children’s use of word order in elicited productions in experiments in which they are initially confronted to ungrammatical sentences (an avenue paved by early research within the competition model; MacWhinney & Bates, 1989). In this set of studies, children are taught novel (or low frequent) verbs presented in ungrammatical structures.

Some ungrammatical structures consist of isolated verbs or verbs lacking one argument (the subject or the object, e.g., Akhtar & Tomasello, 1997; Olguin & Tomasello, 1993; Tomasello et al., 1997). Other structures involve ungrammatical word orders as in the “Weird Word Order”

experimental paradigm (Abbot-Smith et al., 2001; Akhtar, 1999; Matthews et al., 2005, 2007).

The bulk of these studies is that young children, in contrast to older children, tend to re-use the ungrammatical structure modelled by the experimenter in their own productions, a finding which is interpreted as attesting to children’s lack of productivity with word order. The constructivist framework proposes that word order develops from lexically specific schemas formed around frequent verbs, i.e., lexical units that are relatively consistently ordered (Pine et al., 1998), and only slowly generalizes, around 3 or 4 years old, to an abstract representation independent of the verbs.

To reconcile an apparent contradiction in the empirical findings, we undertook a detailed analysis of the methods, results and argumentation proposed in four studies based on the Weird Word Order (WWO) methodology and often referred to in support of the constructivist hypothesis (Abbot-Smith et al., 2001; Akhtar, 1999; Matthews et al., 2005, 2007).¹ The paper is organized as follows. The Section “From theory to data” analyses the logical, deductive pathways from theory to data underlying the lexical hypothesis. We present the predictions formulated by the authors, based on the lexical hypothesis and then argue that these predictions are not borne out under a fine-grained functional analysis of the task and considerations from the adult psycholinguistics literature. In the Section “The problem of missing data”, we address the recurrent problem of data analysis in these papers with regard to the issue of missing data and then proceed to the Section “From data to theory” raising critical issues in data interpretation. The punch line is that it shows not only that most of the predictions of the authors are, in contrast to what they claim, not borne out by their own data, but, more crucially, that the data provide evidence for the alternative, grammatical hypothesis.

2.1.3 From theory to data: critical issues with empirical predictions Different variables are manipulated in WWO studies. The first variable is age:

performance of children from two (sometimes three) age groups is typically compared.² Under the assumption of the lexical hypothesis, the young group (between 2;2 and 3;6 years old, depending on the study) lacks grammatical knowledge of word order (i.e., has lexical knowledge that gradually turns into grammatical knowledge) while the older group (from 4 years old onwards) has developed abstract word order representations. Hence, the young group constitutes the critical test case opposing predictions from the lexical hypothesis to predictions from the grammatical hypothesis. The second variable is the lexical status of the verb. Two conditions are usually contrasted: one involving high frequency (familiar) verbs, the other one involving low frequency or novel verbs. A third variable is also manipulated in some papers, namely the grammaticality of the modelled sentences (grammatical vs. ungrammatical model, i.e., SOV or VSO order). A summary of the experimental conditions illustrated with examples across the various papers is provided in Table 1.

1 The paper by Abbot-Smith et al. (2008), also based on the WWO methodology, is addressed in the General discussion since it reports data (and interpretation) radically different from those reported in the WWO papers reviewed here.

2 Although age is treated as a binary variable in these studies, one should keep in mind that the underlying assumption of the lexical hypothesis is that of a gradual process of language development.

Children’s performance in the task is scored according to three categories: (1) Matches, i.e., reproduction of the word order modelled, (2) Mismatches, i.e., modifications of the word order modelled, or corrections constituting a subset of these responses,³ and (3) miscellaneous responses involving mostly non-responses or responses including lexical units that were not in the model.

Three major predictions differentiating the younger and the older group are examined across papers. The first prediction bears on the reproduction effect which is based on the comparison between responses that “Match” the input word order and responses that “Mismatch”

it. The lexical hypothesis predicts that when confronted with novel verbs,

the youngest children will be willing to spontaneously use the non-SVO orders. They are not expected to switch these orders to SVO. Older children, on the other hand, will either ignore these orders or will actively switch them to SVO, as they will have already formed the generalization that English is an SVO language. (Akhtar, 1999, p. 343)

Table 1

Summary of the methods used in the four studies under review

Note. *Indicates that the word order is ungrammatical in the language tested.

The second prediction bears on the frequency effect which is based on the comparison between high frequency verbs (familiar) and low frequency verbs (unfamiliar or pseudo verbs).

The lexical hypothesis predicts that young children should match ungrammatical orders less if the verb is frequent than if it is infrequent, since they have lexical knowledge of the word order for

3 As we will discuss it in the following section, virtually all Mismatch responses consisted in corrections of ungrammatical orders since children nearly never changed an ungrammatical word order into another ungrammatical order. We nevertheless decided to keep the generic category of Mismatch responses rather than corrections since this category was used in the papers discussed.

the former. For the same reason, they will tend to correct ungrammatical orders with high frequency verbs more than with low frequency verbs. In contrast, no frequency effect, or at least a weaker effect is expected in the older group which is assumed to perform verb-generally, independently of lexical items. As Matthews and colleagues remark:

We anticipated that the number of occasions on which the younger children matched the SVO word order would be roughly inversely proportional to the frequency of the verbs they heard and that this effect of verb frequency would gradually diminish with age.

(Matthews et al., 2005, p. 124)

The third prediction bears on the grammaticality effect which is based on the comparison between grammatical and ungrammatical input models. Akhtar (1999) and Abbot-Smith and colleagues (2001) introduced grammaticality as part of their experimental designs and predicted that in the presence of a novel verb, young children would behave similarly whether the model sentence was grammatical or ungrammatical, and that they would tend to match the word order presented in both cases. In contrast, older children were expected to match grammatical orders more than ungrammatical orders and to mismatch-to-correct ungrammatical orders more than the grammatical one. A summary of the predictions is presented in Table 2.

Table 2

Predictions of the lexical hypothesis in the two age groups

Note. The lexical hypothesis assumes that only the younger group has a lexical representation of word order. To anticipate the results discussed in the Section “The problem of missing data”, predictions borne out by the data are indicated by , whereas predictions not validated by the data are indicated by .

Table 2 provides a preliminary view over the empirical evidence reported in the papers.

While some of the predictions are indeed validated by the data, as indicated by a in the table, others are not, as indicated by a (see the following two sections for an in-depth analysis of the data).

More problematically, some of the predictions are clearly incorrect, either because they fail to take into account relevant factors that may play a role in performance, or because they

ignore important findings from adults’ psycholinguistics. When confronted with grammatical sentences, the parser of languages like English or French assigns the agent role to the preverbal noun phrase in a declarative sentence, and assigns the patient role to the postverbal noun phrase.

But if the noun phrases do not occupy their canonical grammatical positions, the parser cannot assign thematic roles: it has to guess. One source of information that may guide this inferential process is the visual context: the video illustrates who is the agent and who is the patient. If the verb is familiar, knowledge of its argument structure may also influence the inferential process by telling the child whether the two NPs constitute the two arguments of the verb (for transitive verbs) or a single argument in some sort of conjunct phrase (for intransitive verbs, see, e.g., Naigles, 1996, for evidence that argument structure is already finely processed by 30-month-olds).

But the child’s performance to the task is also influenced by another type of guessing which concerns what the experimenter expects from her: correct the ungrammatical word order or adjust to the experimenter’s weird way of speaking? The situation is particularly puzzling since the child hears the experimenter speak correctly during all the familiarization phase before the test sentences are presented. The role of the socio-pragmatic context in which the experiment takes place was demonstrated by Chang and colleagues who found that Japanese children matched WWO more often when speaking to a robot than when speaking to the experimenter (Chang et al., 2009). In a recent experiment replicating Matthews et al.’s (2007), we obtained significantly more WWO corrections with a slightly different procedure in which test sentences were presented via a computer voice (rather than the experimenter) and children were asked to describe the scene to a zebra whose eyes were hidden (Franck et al., 2011).

Thus, it is incorrect to predict that if children have grammatical representations, they will correct ungrammatical word orders with novel verbs. The reason is that the presence of abstract grammatical representations of word order is not a sufficient condition to trigger WWO’s correction with novel verbs: in order to correct a WWO, the child has to be able (a) to infer the argument structure of the novel verbs involved in the WWO input sentences and (b) to infer that the experimenter expects her to correct the incorrect structure from the input. Rather, the implication goes the other way around: if children do correct ungrammatical orders (above chance level) with novel verbs, they must have grammatical representations. It is also incorrect to predict that if children do not have grammatical representations, they will reproduce (i.e., match) the ungrammatical word order. Indeed, in order to be able to reproduce it, they must be able to infer (a) and (b) as well. Rather, the prediction is that if children lack grammatical representations, they should not correct ungrammatical word orders.

As for predictions related to the role of lexical frequency in children’s performance, they fail to take into account the literature in adults’ psycholinguistics showing massive influence of lexical frequency in language processing, including syntactic processing (e.g., MacDonald et al., 1994). We argued that when confronted with sentences with ungrammatical word orders, the parser is insufficient to build a semantic representation for these sentences. Lexical information on the verb provides important cues to this process since argument structure allows determining whether and which NPs may be part of the argument structure of the verb. Hence, lexical knowledge is expected to facilitate the processing of WWO sentences. However, this does not tell us what children will produce.

Thus, it is incorrect to predict that children will reproduce WWO less with high frequency (known) verbs: frequent verbs are expected to boost the general response rate, but the actual responses that children will produce (reproduction or correction of the WWO) depend on the inferences they will make, as discussed previously. It is also incorrect to predict that younger children will be more sensitive to frequency than older children, given that the adult’s mature parser, assumed to involve abstract grammatical knowledge, is significantly influenced by frequency (see Fisher, 2002, for an extended argumentation on the interpretation of lexical frequency effects in the literature on language development).

In sum, erroneous predictions have been formulated in the papers. Predictions related to the reproduction effect fail to take into account cognitive factors recruited in performing the task and which are independent of grammar. Predictions related to the frequency effect fail to take into account evidence about language processes in adults showing across-the-board sensitivity to frequency. In the General discussion, we discuss the consequences of these errors in the interpretation of the data.

2.1.4 The problem of missing data: Critical issues with data analysis One systematic aspect of the data sets reported in these studies is the very high rate of missing data of two types: (1) missing children: children who were excluded from the analyses because they never produced a single test verb throughout the whole experiment, and (2) non-responses: responses from children who were included in the analyses but which failed to contain the target verb (either because nothing was produced or because another verb was used).

We examine these two sorts of missing data in turn.

With respect to missing children, Abbot-Smith and colleagues (2001) report that 10 out of 42 children tested (i.e., 24 %) never produced any of the novel verbs that were presented to them.

Matthews and colleagues (2007) report that out of the 112 children tested, 23 failed “to use a test verb at any point and instead preferred to use alternative, often higher frequency, verbs in their place” (p. 392; an additional set of 28 children was also discarded because of experimenter error or because they failed to use a multiword utterance using any verb). Similarly, in Matthews and colleagues (2005) 43 children out of the 139 children tested were not included in the study “due to experimenter error or because they failed to complete the testing session or to produce any multi-word utterance using any test verb”⁴ (Matthews et al., 2005, p. 124). Hence, it seems that between 20 % and 30 % of the children in those studies were excluded from the analyses.

The high rate of Missing children would not be a problem if it was not asymmetrically distributed across experimental conditions. Table 1 in Matthews and colleagues (2007) shows that whereas only 2 children out of 56 (3.6 %) failed to produce a test verb in the high frequency condition, 26 failed in the low frequency condition (53.6 %). Similarly, in Matthews and colleagues (2005) only 1 child out of 32 failed to respond in the high frequency condition (3.1 %) while 12 failed in the low frequency condition (37.5 %). It is likely that this strong imbalance also characterizes the other studies, although they do not provide the distribution of Missing children.

Hence, Missing children appear to be distributed according to frequency, a variable manipulated in the experiment.

Out of the children who were included in the analyses, Non-responses constitute the vast majority of children’s performance across all experiments, ranging from 46 % to 98.8 %. Like Missing children, Non-responses are not distributed equally across conditions. Akhtar (1999) reports 97 % and 97.5 % Non-responses in the two ungrammatical conditions against 93.2 % in the grammatical condition at age 2, a difference which is also found at ages 3 and 4. Abbot-Smith and colleagues (2001) report that some children did not use the target verb at all in one or two of the three experimental conditions. On average, 98.8 % Non-responses are reported in the ungrammatical VS condition against 95.4 % in the corresponding SV grammatical condition.

Response rates calculated over the numbers provided in Table 2 of Matthews and colleagues (2005) show that children aged 2;9 years provided 92.8 % Non-responses in the low frequency condition against 77.5 % in the high frequency condition, while a difference of 83.7 % vs. 46.9 % was found in children aged 3;10 years. Matthews and colleagues (2007) report that two-year-old

4 The authors did not break down experimenter errors from problems arising from the children (failing to complete the session or to produce a test verb); nevertheless, this is not an issue for the main point developed here which is that missing children are not missing at random.

children failed to respond 82 % of the time in the low frequency condition and 63 % of the time in the high frequency condition, while the 3-year-olds responded 55 % and 52 % of the time, respectively. Again, Non-responses appear to be systematically distributed according to the variables manipulated, i.e., grammaticality, frequency, and age.

In sum, both Missing children and Non-responses are what is called in the statistics literature Missing Not At Random (MNAR): they are not evenly distributed across experimental conditions. Therefore, the mechanism underlying missing children and missing responses cannot be ignored. To obtain valid inferences, a model of the data that includes the mechanism responsible for missing children and missing responses is required. Unfortunately, in the MNAR setting it is very rare to know the appropriate model for the missingness mechanism (e.g., Allison, 2002).

How did the authors deal with missing data? All studies except Akhtar (1999) discarded Missing children, that is, calculated mean frequencies on the basis of the subset of children who produced at least one test verb during the experimental session. Although Non-responses were taken into account in Akhtar (1999) and Abbot-Smith and colleagues (2001), they were discarded in Matthews and colleagues (2005, 2007). In the latter two studies, statistical analyses were conducted over proportions, that is, mean frequencies of Match responses calculated over the subset of responses including a test verb (Match + Mismatch). By doing so, Matches and Mismatches are in complementary distribution. As a result, if young and old children produced a similar raw frequency of Match responses (as is the case in most of the studies) but older children produced more Mismatch than Match responses, the transformation into the mean frequency calculated on the subset of responses containing a test verb misleadingly suggests that young children matched more than older children. As an illustration, Figures 1 and 2 contrast the distribution of raw frequencies (Figure 1, based on Table 2 in Matthews et al., 2005) to that based on mean frequencies calculated on the subset of responses containing a test verb, that is, Match and Mismatch responses (Figure 2, based on Figure 1 in Matthews et al., 2005).⁵

5 The mean frequencies in the present figure are slightly different from those reported in Figure 1 in Matthews et al. (2005). The reason is that, for the sake of clarity, we calculated these proportions on the basis of the raw frequencies reported in our Figure 1.

In contrast, Matthews et al. (2005) calculated proportions for each child individually and then averaged them. Note that this does not change anything with respect to data distribution.

Figure 1

Distribution of the raw frequencies of Match and Mismatch responses in the low frequency condition in Matthews et al. (2005)

Figure 2

Distribution of the mean frequencies of Match and Mismatch responses in the low frequency condition in Matthews et al. (2005), calculated on the subset of responses containing a test verb

Excluding missing data also impacts on the effects of age and lexical frequency. Table 3 shows the distribution of Match responses in Matthews and colleagues (2007) averaged over the two ungrammatical word orders (SOV and VSO). The first column reports raw frequencies as reported in Table 1 in the paper. The second column reports mean frequencies calculated on the subset of responses involving a test verb, i.e., excluding missing responses and missing children

(data from Figure 1 in Matthews et al., 2007).⁶ These mean frequencies were those considered for the statistical analyses. The third column reports mean frequencies calculated on the total number of responses expected from the subset of children included in the study, i.e., excluding missing children but including missing responses. The denominator used to obtain these mean frequencies is the number of children included in each condition (N in Table 1 of their paper) 20 (the number of expected responses for each child). The fourth column reports mean frequencies calculated on the total number of responses expected from all children tested, including the children who failed to produce a test verb at any point in the experiment, i.e., including missing responses and missing children. The denominator is 28 (the number of children tested in each condition) 20 (number of expected responses for each child) = 560.

What appears from Table 3 is that both the effect of age and the effect of lexical frequency on Match responses are dramatically different depending on whether or not missing data are included. Whereas the occurrence of Matches is similar for the two age groups in the low frequency condition when missing data are excluded (left column), it appears higher in the older group when all missing data are included (right column) as well as when only missing children are excluded (middle column). In contrast, the effect of lexical frequency is similar across the two age groups when all missing data are excluded (left column), but appears stronger in the older than in the younger group when missing children are excluded (middle column) and it disappears in the older group and is even inverted in the younger group when all missing data are included (right column).

In sum, studies using the WWO paradigm report a considerable amount of missing data that are critically missing non at random. Missing data are distributed as a function of the experimental variables that were tested; hence, the probability of missingness appears to depend on the varying difficulty of the experimental conditions. This fact strongly suggests that they constitute an integrative part of children’s cognitive processing⁷ and that they can therefore not be discarded from the analyses. Facing missing values, and in the absence of an account for the reasons behind missing values, it is recommended that authors apply a replacement procedure or use a statistical technique that allows for the given incompleteness (see Lachaud & Renaud, 2011, for a tutorial applied to psycholinguistics data). As we will see in the next section, some of the

6 The mean frequencies reported here correspond to those reported by the authors in their Figure 1. They were calculated as the average of individual means of Match for each child (Matthews, personal communication). This explains why the numbers reported here do not exactly fit those obtained by dividing the total number of Match responses reported in Table 1 of the paper by the total number of responses (Match + Total reversion).

7 Missing responses consist in the failure to use a test verb, which means that this category of responses also involves cases where children replaced the verb presented with a familiar verb, which actually amounts to correcting the ungrammaticality of the model sentence.

observations reported are highly sensitive to how missing data are treated and therefore not interpretable as such. However, other observations appear insensitive to the way missing data are dealt with. These observations that are stable across analyses are precisely those one should capitalize on; we review them in the General discussion.

Table 3

Distribution of Match responses in the ungrammatical conditions (based on data reported in Matthews et al., 2007)

Note. The first column reports raw frequencies; the second column reports mean frequencies calculated on the subset of responses involving a test verb (data from Figure 1 in Matthews et al., 2007); the third column reports mean frequencies calculated on the total number of responses expected from the subset of children included in the study;

the fourth column reports mean frequencies calculated on the total number of responses expected from all children tested. The two ungrammatical orders manipulated in the study (SOV and VSO) were collapsed.

2.1.5 From data to theory: Critical issues with data interpretation The influence of the three variables on children’s behavior is now examined in turn: age, lexical frequency, and grammaticality. The analysis shows that the effects argued to provide empirical evidence for the lexical hypothesis are inexistent or at best compatible with it or, more problematically, they are incompatible with it and even provide clear evidence for the grammatical hypothesis.

Effect of age

The authors predicted that young children, who are assumed to lack grammar, would tend to re-use the ungrammatical word orders with novel verbs (Match) while, in contrast, older children would tend to correct them (Mismatch). Since no effect of age is expected on performance with familiar verbs (both age groups are expected to correct ungrammatical orders), we will concentrate on the condition involving novel verbs. We discuss in turn the effect of age on Match and on Mismatch responses.