Frequency Effects in Language Acquisition: Defining the Limits of Frequency as an Explanatory Concept 9783110977905, 9783110196719

Citation preview

Frequency Effects in Language Acquisition

W DE

G

Studies on Language Acquisition 32

Editor Peter Jordens

Mouton de Gruyter Berlin · New York

Frequency Effects in Language Acquisition Defining the Limits of Frequency as an Explanatory Concept Edited by Insa Gülzow Natalia Gagarina

Mouton de Gruyter Berlin · New York

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter GmbH & Co. KG, Berlin.

Printed on acid-free paper which falls within the guidelines of the ANSI to ensure permanence and durability.

Library of Congress Cataloging-in-Publication Data

Frequency effects in language acquisition : defining the limits of frequency as an explanatory concept / edited by Insa Giilzow, Natalia Gagarina. p. cm. — (Studies on language acquisition ; 32) Includes bibliographical references and index. ISBN 978-3-11-019671-9 (cloth : alk. paper) 1. Language acquisition. 2. Frequency (Linguistics) I. Giilzow, Insa, 1967— II. Gagarina, Natalia. P118.F755 2007 401'.93-dc22 2007039594

Bibliographie information published by the Deutsche Nationalbibliothek The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.

ISBN 978-3-11-019671-9 ISSN 1861-4248 © Copyright 2007 by Walter de Gruyter GmbH & Co. KG, D-10785 Berlin. All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Cover design: Sigurd Wendland, Berlin. Printed in Germany.

Contents

Introducing the frequency debate in studies of language acquisition Insa Giilzow and Natalia Gagarina

1

Opening statement What frequency can do and what it can't Tom Roeper

23

The acquisition of determiners The role of input frequency in article acquisition in early child Swedish Ute Bohnacker

51

Testing the effects of frequency on the rate of learning: Determiner use in early French, German and Italian Tanja Kupisch

83

Input frequencies across development The acquisition of non-agent subjects in child Hebrew: The role of input Sigal Uziel-Karl and Nancy Budwig

117

The role of input frequency in early language production: Children's usage of Serbian prepositions Maja Savic andDarinka Andelkovic

145

Characteristics of maternal input in relation to vocabulary development in children learning German Christina Kauschke and Gisela Klann-Delius

181

What happens when adults often use infinitives? Natalia Gagarina

205

vi

Contents

Frequency mismatches between caregiver input and child language Structural versus frequency effects in LI acquisition of the passive and impersonal in Serbian Milja Djurkovic

237

The (non-) effect of input frequency on the acquisition of word order in Norwegian embedded clauses Marit Westergaard and Kristine Bentzen

271

Factors determining the acquisition of animacy in Czech Denisa Bordag

307

(Non-)Frequentist perspectives within UG A note on acquisition in frequency-based accounts of binding phenomena Jason Mattausch and Insa Gülzow Principles, parameters and probabilities Rosalind Thornton, Stephen Grain and Graciela Tesan

331 359

Comments The role of frequency in language acquisition Katherine Demuth

383

Counting grammars Charles Yang

389

List of contributors

407

Index

410

Introducing the frequency debate in studies of language acquisition Insa Giilzow and Natalia Gagarina

1. Introductory remarks The present volume contains selected papers from a workshop on frequency effects in language acquisition at the DGfS' annual conference held in Cologne in February 2005 and the IASCL Congress,1 which took place in Berlin in July 2005. The main aim of this volume is to bring together researchers with an interest in the impact of a linguistic item's frequency in the input on the process of (first) language acquisition. What is the nature of the relationship between the distribution of grammatical or lexical items in the input and their distribution in early child language? Can frequency effects serve as an explanatory concept for empirical findings? In what way can the mechanisms of a highly frequent item's acquisition be related to that of an infrequent item? The papers collected in this volume represent the two leading approaches to language acquisition: UG-based and usage-based accounts. As the frequency debate tends to divide researchers into those with a more formalistic and those with a more functional theoretical approach, the papers explicitly relate to the tension created between these two positions, which is debated in the three parts of the volume: an opening statement relating to the current frequency debate, an empirical section discussing findings in a variety of languages, and concluding comments. The first contribution is a mainly theoretical account in which the author makes the somewhat categorical statement that 'frequency is not an explanatory concept' (Roeper, this volume). The papers that follow Roeper's chapter discuss findings in a variety of languages and mirror the frequency debate in that they relate to Roeper's statement by identifying the areas in which frequency effects can and cannot contribute to a theory of language learning (Bohnacker; Kupisch; Uziel-Karl & Budwig; Savic & Andelkovic; Kauschke & Klann-Delius; Gagarina; Djurkovic; Westergaard & Bentzen; Bordag; Mattausch & Gülzow; Thornton, Crain & Tesan). At the end of the volume,

2

Insa Gülzow and Natalia Gagarina

two researchers (Demuth; Yang) formulate their comments with regard to the frequency debate. In the process of language acquisition, a child moves from an initial state of limited language comprehension and production to a state of adultlike language use. During this process, innate structures and their interaction with the input are assigned a different status by the two leading approaches. For theories that assume the existence of a rich inventory of innate structures, the question of how frequency of exposure leads to acquisition is basically irrelevant. Theoretically, single exposure may be as successful as many, or, when arguments of an impoverished stimulus are considered, even in cases in which exposure is absent or ambiguous, acquisition is believed to take place. Functionalist accounts, on the other hand, are based on the assumption that language structure more directly relates to language use. In the acquisition process, children operate with formfunction pairings and develop these in their continuous experience of the linguistic environment. Among other factors that make linguistic elements identifiable for the language-learning child, frequency of exposure is believed to add to the saliency of a linguistic element. Theories with a focus on how continuous experience with language structure moves the child to a target-like language use implicitly or explicitly have to relate to the question of a critical mass of exposure. How many exposures of a linguistic item are necessary in the learning process and how are the items represented before, during and after acquisition? The empirical findings presented and discussed in this volume mirror the current state of the frequency debate in that they show that no clear-cut situation exists in which a theoretical approach can be neatly associated with a clearly defined position regarding frequency effects. On the one hand, it is impossible for approaches with a more universalist orientation to ignore the fact that on some level, frequency can or even must affect the acquisition of linguistic elements, as children do not acquire language in complete absence of exposure. This is most clearly reflected by statements such as those recently formulated by Yang (2002), who argues for a leading role of frequency in grammar competition. On the other hand, it is acknowledged by researchers with a more usage-based orientation that frequency effects must be interpreted from an integrative perspective, which departs from the statement that the frequency of linguistic elements in child language is merely a reflection of the input. In a recent publication, Bybee (2006) has argued that the relation between language use and grammatical structure is reciprocal. If grammar is viewed as "the cognitive organization of one's experience

Introducing the frequency debate in studies of language acquisition

3

with language", it follows that grammar will determine the way in which language is used just as much as language use affects the emergence of grammar (Bybee 2006: 711). Being part of language structure, the impact that the frequency of a linguistic item may have on its representation can never be assessed in isolation. It is demonstrated in the contributions of the volume that frequency is always in interaction with other factors and relates, for instance, to an element's prosodic prominence, complexity, conceptual transparency, or general accessibility. This is a position that is currently also defended in second language acquisition theory (e.g. Ellis 2002; Gass & Mackey 2002; Hulstijn 2002).

2. Approaches to the frequency debate in linguistic theory Heated debates regarding the concept of frequency and its status in linguistic theory go back as far as the 1930s. In studies exploring evolutionary processes, such as the emergence of linguistic units, patterns or constructions, it has been argued that diachronic changes may be rooted in the frequency of use of certain patterns. Zipf (1935), for instance, demonstrated that there is a connection between the complexity of a linguistic structure and an either low or high frequency of its use in language. This connection, which Zipf calls "the principle of least effort", is still regarded as a basic mechanism in frequency-based explanations and in more recent linguistic theories such as bidirectional optimality theory. In current approaches to diachronic linguistics and typology, proponents of the idea that (morphosyntactic) change and (morphosyntactic) typologies can be explained from a frequentist perspective are Joan Bybee (1998, 2001, 2006) and Martin Haspelmath (2003, 2006). Bybee (2001) argues for a prominent role of frequency in the entrenchment and productivity of constructions in French. She shows, for instance, that for the French liaison type and token frequencies facilitate the resistance to modification and productivity in constructions consisting of more than one word. Addressing a more general issue, Bybee & Hopper (2001: 3) argue that the frequency of a given linguistic element (construction and/or pattern) "has a profound influence on the way language is broken into chunks in memory storage, the way such chunks are related to other stored material and the ease with which they are accessed". In his appeal against markedness, Haspelmath relates to Zipf s "principle of least effort" by replicating its essence in the statement that "the un-

4

Insa Gülzow and Natalia Gagarina

marked term is more frequent than the marked term" (Haspelmath 2006: 55, cf. also Greenberg 1966). Haspelmath argues that "complexity and difficulty typically lead to low frequency" (Haspelmath 2006: 63) and he suggests that the notion of markedness can be discarded in favor of the low or high frequency of a linguistic item. In this sense, frequency may replace markedness in cases where markedness is traditionally regarded "as a morphological difficulty/unnaturalness" or "as a multidirectional correlation" (Haspelmath 2006: 64). In his later work, Haspelmath (in press) assigns frequency an even more prominent role when accounting for morphosyntactic asymmetries. Since "frequent patterns are encoded with less material", "[a] 11 morphosyntactic asymmetries can be explained on the basis of frequency asymmetries; i.e. in a sense they are all instances of economic motivation" (Haspelmath in press: 1). He thus epitomizes the slogan that "grammars code best what speakers do most" (Du Bois 1985: 363). With regard to the relation between usage and grammar, Bybee (2006) assigns frequency a major role in determining the way in which exemplars are represented. Based on findings that memory of non-linguistic phenomena can be highly detailed and extensive (e.g. Goldinger 1996), Bybee argues that there is reason to believe that this also applies to the memory of linguistic items. Linguistic structures with a low frequency have a low impact on the accumulated representation effected by previously encountered exemplars, which also reduces the probability of low frequency items being relevant for the representation of larger units: "[OJnly when a sequence is repeated will access to it as a unit rather than by its parts become more efficient" (Bybee 2006: 717). A further pro-frequency position is formulated by the representatives of bidirectional optimality theory (OT). Jäger & Zeevat (2002) argue for a frequency-based approach to the development of differential case marking paradigm patterns. Using his bidirectional gradual learning algorithm as a formal tool, Jäger (2003) demonstrates how a frequency-based analysis can account for diachronic change and typology. Developing Jäger's line of research, Mattausch (2005) investigates the evolution of anaphoric binding phenomena from a frequentist perspective. In the Iterated Learning Model (Kirby & Hurford 1997) bidirectional learning is combined with bias constraints. Mattausch argues for "the interaction of bias constraints with the two other types of constrains, markedness constraints and faithfulness constraints in a context of bidirectional learning" (Mattausch 2005: 164). In Mattausch's model, a learner may replicate a grammar on the basis of distributional patterns of linguistic elements in a given language.

Introducing the frequency debate in studies of language acquisition 5 3. Positions in the frequency debate in first language acquisition

As already mentioned above, approaches differ with regard to the realm in which they assign frequency a status as an explanatory concept. While there is a tendency of non-nativist approaches to assign distributional characteristics of the input an effect on language acquisition, which is generally denied by nativist approaches, the interpretations of empirical findings are far from homogenous. Accordingly, the impact of input frequencies on the order of acquisition and distributional patterns in children's productions has been a debated issue among acquisitionists since the seventies. Brown (1973) for instance, in his classical study investigating the acquisition of fourteen grammatical morphemes in English, found a high degree of constancy2 regarding the order of acquisition within the three children investigated (see also Roeper 1999; Yang 2002; Rowland et al. 2003). While Brown could demonstrate that there is a high degree of stability in the order across the children and in frequencies across the parents, no relation could be demonstrated between the two. Brown also compared individual orders with individual frequencies and the effect of the parent's use of particular prepositional phrases. In none of the cases could he show frequency to be a significant variable and thus he concluded that the distribution of linguistic elements in the input does not affect the children's path of acquisition. Proponents of the strong innateness hypothesis like Farwell (1973) not only denied frequency any role in the acquisition process, but the whole input as such: "it remains to be shown that adult language, simplified or not, has any direct effect on the progress of acquisition" (Farwell 1973: 49, cited in Forner 1979). In order to gainsay or to confirm these strong assumptions contra frequency and contra input in general, a number of studies have been published since which investigate the precise nature of the relation between parental input and child speech. Demuth (1989), for instance, relates the early use of passive structures in Sesotho to their high frequency in the input and a similar finding is reported by Gil (2006) for the use of passive suffixes in Jakarta Indonesian. According to Gil (2006), passive suffixes in Jakarta Indonesian are acquired early due to their high frequency in parental input and due to their formal simplicity. Likewise, Tomasello (2003) in reviewing studies on the early acquisition of passives in languages such as Sesotho and Inuktitut observes that "in these languages passives are used more frequently and saliently in adult speech to young children" (Tomasello 2003: 174). In order to test the role of frequency in the acquisition of a given linguistic

6

Insa Gülzow and Natalia Gaganna

structure, he performed experiments with English-speaking children between three and three and a half years of age in which the children were exposed to passive constructions more frequently than in the regular input. The children were then able to produce full passive sentences with a nonsense verb (Brooks & Tomasello 1999). Theakston et al. (2004) discuss to what extent input frequency, together with semantic generality, influences the acquisition of certain lexical items in the early stages of language development. They show that there is no difference in the use of semantically general verbs in children's speech and in the input. This finding suggests "that the children are not using semantically general verbs in place of more specific verbs that they have not yet acquired. Instead, their relative frequency of use mirrors the relative frequency of use in the input" (Theakston et al. 2004: 90). In contrast to these findings, it has been demonstrated for the acquisition of English datives that there is no correlation between the order in which children acquire the two types of datives and their representation in the input (Snyder & Stromswold 1997). In contrast to to-datives (John gave the book to Mary), double object datives for the verb give (John gave Mary the book) were represented with an average of about 73% in the adult's speech. Despite this fact, there was no significant correlation between the age of acquisition of /o-datives, which generally occurred later in the children's data, and their relative frequency in the input. While there is little consensus with regard to the direct and exclusive impact of input frequencies on the process of language acquisition, many studies indicate that it may not be frequency alone that explains why certain forms/constructions appear early while others do not. These studies address the issue of frequency interacting with various input components affecting the course of language acquisition, like cognition or linguistic complexity (cf. MacWhinney 2004; Gathercole 2005). Savic (1975), for instance, compared the order of appearance and frequency of use of w//-questions in the input and the order of emergence and frequency of use of these questions in child Serbo-Croatian. Neither with regard to the order nor with regard to the frequencies of questions in the speech of adults could Savic see a correspondence. Four years later, Savic's interpretation of her data was challenged by Forner (1979), who also replicated the study using the German data of her bilingual son. In this detailed analysis, Forner applies a "more accurate and more revealing" (Forner 1979: 25) statistical measure, the Pearson product-moment correlation coefficient, in order to statistically prove significant relationships between the order and frequency of wh-

Introducing the frequency debate in studies of language acquisition

1

questions in mother and child speech. In both Forner's and Savic's corpus it is demonstrated that "the mother's frequency highly correlates with the child's frequency and nearly always correlates with his order of production" (Forner 1979: 39). On the basis of her findings, Fomer introduces a model of acquisition which relates children's cognitive development and the semantic complexity of language to the input offered by the children's caretakers. While semantic (cognitive) complexity could not be shown to have a direct influence on the children's order or frequency of productions, semantic complexity could be shown to surface indirectly in the children's productions mirroring a given mother's adjustments in her language addressed to the child. In a much later study, Rowland et al. (2003) along similar lines show for the acquisition order of English w/z-questions that in contrast to input frequency, the syntactic and semantic complexity of whquestions could not reliably predict the order of their acquisition. Rather, an interaction between complexity and frequency seems to explain the late acquisition of relatively frequent w/z-constructions. Naigles & HoffGinsberg (1998) investigated the influence of the verb input on the order of verb acquisition for English-speaking children. They showed that input frequency of verb use, together with the utterance-final position and diversity of syntactic frames influences the children's production of the same verbs. The last group of theoreticians to be noted here work on the simulation of language acquisition as computational mechanisms and discuss the impact of frequency in terms of modelling statistical learning versus rulegoverned behavior (cf. Elman et al. 1996; Seidenberg & MacDonald 1999; Marcus 1998). Recurrent issues in connectionist models are how one can account for children's acquisition of linguistic elements with a low frequency in the input and mismatches between the distribution of elements in natural databases and the number of elements that are needed for the modelling of, for instance, inflectional learning. Marcus (2000) has argued that human language learning depends both on a statistical mechanism and a rule-based mechanism. In contrast to earlier models, such as Plunkett and Marchman (1993) or Rumelhart & McClelland (1986) in which the frequency with which a linguistic item appears in the training has a major impact on qualitative changes such as the appearance of overgeneralizations in past tense forms, Marcus argues that the production of regular and overgeneralized forms is dependent on a rule-based mechanism and thus independent of highly regular type frequency. For the acquisition of irregular English past inflections on the other hand, a statistical mechanism that is

8

Insa Gülzow and Natalia Gagarina

sensitive to type frequency is needed. According to Marcus, irregular past inflections are produced on the basis of resemblance to stored exemplars in a statistically sensitive memory. This makes the overgeneralization of irregular past tense forms with many similar neighbors more unlikely than the overgeneralization of irregular past tense forms with few such neighbors. Finally, Charles Yang (2002) proposes an approach to language acquisition in which statistical regularities in the input and a universal innate grammar both contribute to the process of language acquisition. The linguistic experience of a language-learning child interacts with a finite number of possible human grammars that are accessible to the learner from the start. Borrowing from Darwinian evolutionary biology (cf. Lewontin 1983) variant grammar types are expected to coexist in the learning process. A grammar is established once the target type persists while others have disappeared. Taking a general purpose learning model from the behaviorist tradition as a basis, Yang argues that during the learning process each grammar G is associated with a weight p that will be either rewarded or punished by the linguistic experience a child makes. Grammar competition can thus account for both the non-uniformity and the gradualness in individual samples of language acquisition. If more than one grammar succeeds in analyzing a specific linguistic structure in a child's input, frequency of exposure is crucial in that the relative frequency with which the (target) structure is reinforced guides the child in language acquisition. All in all, a continuum of opinions is presented in previous research. While representatives of the one extreme emphasize the strength and power of frequentist explanations, representatives of the other insist on the weakness of such an approach. As already mentioned, it is remarkable that the positions formulated above cannot necessarily be correlated with either a UG or a usage-based approach to language acquisition. Scholars of both groups investigate frequency on various linguistic levels and assign it a different status. It seems that the advocates of either theory adjoin, overlap, coincide, concur and struggle in their formulation of what kind of impact frequency may have in the acquisition process. This makes the frequency debate as presented in the chapters even more fascinating. In the present volume an attempt is made both at resolving the divergent positions and at defining the realm of frequency as an explanatory concept.

Introducing the frequency debate in studies of language acquisition

9

4. Frequency as an explanatory concept The opening statement of the present volume is formulated by Thomas Roeper, who argues against frequency as an explanatory concept. Roeper assigns frequency a status in parametric variation, assuming that children entertain multiple grammars while acquiring a given language. He strongly differentiates between frequency as the counting of instances of information and the developmental process of adding new information. Only the latter is assigned an explanatory status. As learning equals the change of mental representations and frequency is a merely cumulative effect, it is not necessary or relevant in a majority of cases to relate frequency to learning. Rather than contributing to the constitution of mental representations in the process of constructing and acquiring a grammar, frequency can be assigned a minor role when it comes to the choice of mental representations. One of Roeper's main arguments against frequency as an explanatory concept is the difficulty or even impossibility of defining the precise nature of the relationship between a linguistic element that is learned and its embeddedness in the grammar of a given language. As there is no clear understanding of how and which aspects of a mental representation change when learning takes place, frequency cannot explain or predict how and which exposure to a linguistic element contributes to the change of a mental representation. Instead of asking how many exposures may be needed for an element or one of its components to be learned, one might as well ask why it would take more than one exposure. This is not to say that linguistic elements that have a high frequency in the input do not increase the probability of a child paying attention to it. But as elements with a high frequency may nevertheless appear relatively late in acquisition while there are non frequent elements that are acquired relatively early, distribution cannot be regarded as being equally relevant in all cases. It is a well-known fact, for instance, that in some languages the distribution of linguistic elements in the input, such as articles, is mirrored from early on, while in others they remain absent for a relatively long time (cf. Kupisch, this volume). Roeper suggests that frequency is relevant when it comes to categorical grammatical choices in the acquisition process. Assuming that a child who produces and accepts non target linguistic elements and structures is entertaining more than one grammar, frequency may play a role in deciding which of the alternatives is to be chosen. Frequency metric may either directly contribute to the choice of a target grammar or it may affect the learning process indirectly when the child keeps track of the number of contexts in which a certain grammar may apply.

10 Insa Giilzow and Natalia Gaganna 5.

Discussing empirical data

The papers collected in the present volume discuss empirical data from a variety of languages such as French, German, Swedish, Italian, Czech, Hebrew, Serbian, Russian and Norwegian. They address topics such as the acquisition of grammatical elements, the relation between caregiver input and early productions across development, children's ability to encode concepts such as animacy and (non-)agency and the acquisition of syntactical structures as reflected by word order or passives. All papers relate to Roeper's opening statement and thereby create a strong link between their individual findings and the contribution they make to the frequency debate. The discussion of the empirical findings mirrors the current debate in that the authors of the present volume critically identify those aspects in the process of language acquisition that help to define frequency as a theoretically relevant concept. Departing from the observation that the frequency of a linguistic element in child language simply reflects the input, the contributions are united by realizing that the impact that frequency has on the way in which a child learns his or her language is complex and cannot be addressed in isolation. An attempt is made both to define the areas in which frequency interacts with acquisition and to pin down the limitations of frequency of an explanatory concept.

5.1. The acquisition of determiners The two papers following Roeper's contribution deal with children's acquisition of determiners. Bohnacker investigates two monolingual Swedish children's acquisition of definite and indefinite articles and Kupisch discusses the possible impact of frequency on determiner omission from a cross-linguistic perspective in monolingual children acquiring French, German and Italian. Although taking a non-functionalist approach, both contributions advocate a position that assigns frequency a role as a descriptive tool in the explanation of empirical findings. At the same time, they raise the issue of how correlations between input and early child language can be incorporated into a theory of language learning. Bohnacker adopts a fairly radical perspective and argues that frequency effects are little more than reflexes of the target language's structure and must therefore be regarded as an epiphenomenon. This is not to say that elements that are highly frequent in the input may show a tendency to emerge early as long

Introducing the frequency debate in studies of language acquisition

11

as they match the child's communicative goals. Conversely, infrequent elements may appear late in acquisition. However, Bohnacker's data also demonstrate that indefinite pronominal articles, which are relatively frequent in caregiver speech, as well as the comparatively infrequent enclitic definite articles are used target-like at about the same time. In this respect, the inconsistencies that Kupisch finds between adult and child language are especially relevant in cases in which the relative frequency of determiners in the input of children acquiring different languages would make different predictions than what findings actually show. Kupisch's findings suggest that neither can the language-specific properties of languages as closely related in their use of determiners as Italian and French be ignored when interpreting children's early language use. Nor is it possible to reduce frequency to a factor bearing on the rate of determiner acquisition across different languages. Rather, it must be acknowledged that the quantity of a linguistic element in the input may interact with different parameters such as the syntax/prosody mapping and the form/function mapping at different points in development. 5.2. Input frequencies across development A dynamic relation between input and the children's productions across development is documented in Uziel-Karl & Budwig's and Savic & Andelkovic's studies. Uziel-Karl & Budwig show that the frequency with which mothers used particular change of state verbs (COS) was mirrored in the children's data. This distribution is stable across development, that is, as the mothers increased their use of COS verbs and subject types, so did their daughters. A similar effect was also noted by Bohnacker for the use of bare nouns by the two Swedish children and their caretakers. Uziel-Karl & Budwig argue that although the frequency with which mothers offer linguistic elements to their children plays a role in the choice and order of acquisition, frequency cannot induce learning. Rather, the mother's choice of different distributional patterns at different points in development will promote those elements that are acquired early. Savic & Andelkovic investigate the distribution of prepositions in child language in comparison to different registers of adult language: child directed speech (CDS), written and conversational samples. The closest match is found between CDS and the children's productions, which is interpreted by Savic & Andelkovic as evidence for a reciprocal dynamic between children's and adults' language

12 Insa Gülzow and Natalia Gagarina

use. Rather than reacting to the structure of the target language in general, children specifically interact with the language structure offered to them in CDS. Both studies point out that while frequency seems to affect the choice children make at a given stage of development with regard to which linguistic element they focus their attention on, there are factors interacting with frequency. This will result in a mismatch between the distribution of a linguistic element in the input and the children's data. In the Uziel-Karl & Budwig's study of Hebrew children's language development, the choice of a particular morphological pattern for COS verbs may be attributed to a pure frequency effect, as this verb pattern happens to be associated with frequently used verbs although most COS verbs are part of a different morphological paradigm. Focusing on maternal input, Kauschke & Klann-Delius investigate the relationship of maternal input and child language with respect to the children's lexical development. Their data demonstrate that there is a strong correspondence between the distribution of elements in the input and the children's productions. Kauschke & Klann-Delius show that child-directed speech may somewhat change during the course of development and that the children's productions vary accordingly. This is especially evident in type and token correspondences between maternal input and the children's productions which appear at different points of development. For nouns, it could be shown that both mothers and children increase the number of types during the course of development, while verbs types are used with increasing frequency. Gagarina, in her investigation of children's non target uses of Russian infinitives, focuses on the important role of the input as a source of these errors. She shows that the frequency of analytic verb constructions in the input may run counter to the children's struggle with the structural complexity of these elements. The relative frequency of infinitives increases their salience for the language-learning children and facilitates their easy uptake and use, while the structural variety and the complexity of analytic constructions with infinitives in Russian impedes their acquisition, thus promoting persisting errors. 5.3. Frequency mismatches between caregiver input and child language In her study of children's comprehension and production of the Serbian passive and impersonal construction, Djurkovic finds no evidence of simi-

Introducing the frequency debate in studies of language acquisition

13

lar distribution patterns in caregiver input and the children's productions. On the contrary, while in Serbian adult language passives and impersonals appeared in amounts that displayed no statistical differences, children's productions of passives and impersonals could be shown to differ statistically. Moreover, in a comprehension study children reacted correctly to actives and impersonals in about the same amount of instances while they performed much worse on passives. Djurkovic adopts a very critical position regarding an innate universal grammar, but at the same time acknowledges that a 'critical mass' of exposures is theoretically not plausible, just as the distributional patterns and the acquisition of infrequent linguistic structures are problematic for accounts that rest on the assumption that children construct their language from linguistic experience. Rather, Djurkovic argues that if a level of representation for abstract syntactic structure is assumed, the relative difficulty with which Serbian children learn the passive may not only have to be attributed to mapping difficulties between form and function. Westergaard & Bentzen find a similar effect in the distribution of embedded clauses and questions in caregiver speech and the children's productions in their study of Norwegian children's acquisition of word order. Although the frequencies with which clauses and questions with a non-V2 structure and declaratives with a V2 structure correspond to the children's errors in production, embedded questions with a non-V2 structure were mastered by the children from early on, although they are comparatively infrequent in the input. Westergaard & Bentzen argue that although frequency may be assigned a role in the children's choice of a structure, their results show that there is also reason to believe that other aspects of language acquisition must be interacting with such effects. Comparing their results to findings in German child language, Westergaard & Bentzen adopt a position in which frequency or lack of frequency cannot directly cause a child to move down a certain path in language acquisition. Rather, they assume that the children's productions can be analyzed by assuming an economy principle of movement which hinders children from moving elements higher than corresponds to the evidence in the input. In these cases, it may be the rarity of embedded clauses and question with a non-V2 structure in the input that leads to a persistence of children's errors with these structures at least up to an age of six. Bordag investigates Czech children's acquisition of different masculine plural inflectional endings that mark animate and inanimate nouns. In an experiment with nonsense words, she was able to demonstrate that children's

14

Insa Gülzow and Natalia Gagarina

choice of an inflectional ending was not in accordance with the target morphology, but rather followed the predominant pattern in the input. Without paying much attention to the animacy distinction, Czech children tend to choose the plural inflection most frequently represented in their input. Bordag's contribution calls attention to the fact that frequency effects are affected by the well-known difference between children's performance in experimental conditions versus naturalistic settings. As shown in Berko's famous study (Berko 1958), children produce non-target inflections more readily with unknown words than familiar ones.

5.4. (Non-)Frequentist perspectives within UG The last two papers discuss the possible role of frequency within two different generative approaches to grammar. The first paper shows how the particular distributions of elements available to children in the input may serve as a basis for replicating the grammatical structure of the target language. The second paper concentrates on a number of well-known phenomena in child language that are interpreted within the Principles and Parameters approach. Mattausch & Gülzow offer a model for the acquisition of English personal pronouns and reflexive pronouns which is based on a frequentist theory of grammar. Somewhat similar to the proposal of Yang (2002) in that the frequency of elements is relevant in competitive processes, Mattausch (2006) developed a model for the evolution of binding phenomena within bidirectional optimality theory. Taking this as a basis, Mattausch & Gülzow discuss the so-called 'pronoun interpretation problem' or the 'delay of Principle B effect', which is a well-known phenomenon in English child language. While English children produce personal pronouns and reflexive pronouns virtually error free and reflexive pronouns are comprehended early, they allow a coreferential interpretation for personal pronouns up to six or seven years of age. The proposed model offers an explanation for these empirical facts by formulating a number of straightforward constraints that are combined with a markedness constraint reflecting the fact that self-forms in English are to be avoided whenever possible. Feeding the distributions of personal pronouns and reflexive pronouns of adult input into the model, it can be demonstrated that the constraints eventually acquire the values of the target system in English. Before they do so, however, their values account for the children's prolonged difficulties when interpreting personal pronouns versus reflexives. Thornton, Crain & Tesan

Introducing the frequency debate in studies of language acquisition

15

discuss three kinds of empirical data that challenge a mechanism of language learning that is based on continuous experience. Regarding, for instance, children's production of yes/no-questions, Thornton, Grain & Tesan draw attention to the fact that crucial information for the acquisition of this structure is not offered in abundance but represented by low frequencies. Moreover, children never produce utterances that are in accordance with a solely linear analysis of the input. Rather, they can be shown to observe structural rules from the onset of language production. In the case of whquestions with why, English children can be shown to produce noninversed structures on the basis of not only low frequencies of input but with no available input at all. As Thornton, Grain & Tesan argue, this finding is compatible with the Continuity Hypothesis in Universal Grammar, which allows structures of a possible human language to surface in children's productions. With regard to negation, Thornton, Grain & Tesan show that the children's productions document a path of development that is more characteristic of an abrupt and categorical change than of a gradual adaptation as proposed by Yang (2002). At the end of the book, Katherine Demuth and Charles Yang give a short comment and express their view on the debate of the role of frequency in the acquisitional process.

Notes 1. DGfS: Deutsche Gesellschaft für Sprachwissenschaft; IASCL: International Association of the Study of Child Language 2. The rank-order correlation coefficients for order of acquisition of the fourteen morphemes are 0.88 for Adam and Sarah, 0.86 for Adam and Eve and 0.87 for Sarah and Eve (Brown 1973: 272).

References Brooks, Patricia & Michael Tomasello 1999 How children constrain their argument structure constructions. Language 75(4): 720-738. Brown, Roger 1973 A first language: The early stages. Cambridge: Harvard University Press.

16

Insa Gülzow and Natalia Gagarina

Bybee, Joan 1998 The emergent lexicon. Chicago Linguistics Society 34: 421^435. 2001 Frequency effects in French liaison. In Frequency and the emergence of linguistic structure, Joan Bybee & Paul Hopper (eds.), 337-359. Amsterdam: Benjamins. 2006 From usage to grammar: The mind's response to repetition. Language, 82(4): 529-551. Bybee, Joan & Paul Hopper (eds.) 2001 Frequency and the emergence of linguistic structure. Amsterdam: Benjamins. Demuth, Katherine 1989 Argument structure and the acquisition of Sesotho applicatives. Linguistics, 36(4)781-806. Du Bois, John 1985 Competing motivations. In Iconicity in syntax, John Haiman (ed.), 343-365. Amsterdam: Benjamins. Ellis, Nick 2002 Reflections on frequency effects in language processing. Studies in Second Language Acquisition 24(2): 297-339. Elman, Jeffrey L., Elizabeth Bates, Mark H. Johnson, Annette Karmiloff-Smith, Domenico Parisi & Kim Plunkett (eds.) 1996 Rethinking Innateness: a connectionist perspective on development. Cambridge, MA: MIT Press. Farwell, Carol B. 1973 The language spoken to children. Papers and reports in child language development, Vol. 5. Stanford University. Forner, Monika 1979 The mother as LAD: interaction between order and frequency of parental input and child production. In Studies in first and second language acquisition, Fred R. Eckman & Ashley J. Hastings (eds.), 17—44. Rowley, MA: Newbury House Publishers. Gass, Susan & Alison Mackey 2002 Frequency Effects and Second Language Acquisition. Studies in Second language Acquisition 24(2): 249-260. Gathercole, Virginia Mueller 2005 Input, Cognicion, y Estructura Lingüistica: Influencias Conjuntas en el Desarrollo del Lenguaje en los Ninos. [Input, cognition, and linguistic structure: Concerted influences on language development in children.] In Estudios sobre la adquisicion del lenguaje (Studies of Language Acquisition), Maria Angeles Mayor Cinca, Begona Zubiauz & Emiliano Diez-Villoria (eds.), 27-57. Dialnet: Universidad de la Rioja.

Introducing the frequency debate in studies of language acquisition Gil, David 2006

17

The acquisition of voice morphology in Jakarta Indonesian. In The Acquisition of Verbs and their Grammar: The Effect of Particular Languages, Natalia Gagarina & Insa Giilzow (eds.), 201-227. Series: Studies in Theoretical Psycholinguistics, Vol. 33. Dordrecht: Springer. Goldinger, Stephen 1996 Words and Voices: Episodic traces in spoken word identification and recognition memory. Journal of Experimental Psychology: Learning, Memory and Cognition 22: 1166-1183. Greenberg, Joseph H. 1966 Some universals of grammar with particular reference to the order of meaningful elements. In Universals of language, Joseph H. Greenberg (ed.), 73-113. Cambridge, MA: MIT Press. Haspelmath, Martin 2006 Against markedness (and what to replace it with). Journal of Linguistics 42(1): 25-70. in press Creating economical morphosyntactic patterns in language change. In Language universals and language change, Jeff Good (ed.). Oxford: Oxford University Press. Hulstijn, Jan 2002 What does the impact of frequency tell us about the language acquisition device? Studies in Second Language Acquisition 24 (2): 269273. Jäger, Gerhardt 2003 Learning constraint sub-hierarchies: the bidirectional gradual learning algorithm. In Pragmatics in OT, Reinhard Blutner & Henk Zeevat (eds.), 251-287. Palgrave: MacMillan. Jäger, Gerhardt & Henk Zeevat 2002 A reinterpretation of syntactic alignment. In Proceedings of the 3rd and 4th International Symposium on Language, Logic and Computation, Dick de Jongh, Henk Zeevat & Marie Nilsenova (eds.), Amsterdam: ILLC (Institute for Logic, Langugae and Computation). Kirby, Simon & Jim Hurford 1997 The evolution of incremental learning: language, development and critical periods. Technical report, Language Evolution and Computation Research Unit, University of Edinburgh. Kupisch, Tanja this vol. Testing the effects of frequency on the rate of learning: Determiner use in early French, German and Italian. Lewontin, Richard 1983 The Organism as Subject and Object of Evolution. Scientia 188: 6582.

18 Insa Gülzow and Natalia Gagarina MacWhinney, Brian 2004 A multiple process solution to the logical problem of language acquisition. Journal of Child Language 31: 883-914. Marcus, Gary 1998 Can connectionism save constructivism? Cognition 66: 153-182. 2000 Pa bi ku and ga ti ga: Two mechanisms children could use to learn about language and the world. Current Directions in Psychological Science*): 145-147. Mattausch, Jason 2003 On the Optimization and Grammaticalization of Anaphora. ZAS Papers in Linguistics, 38. Michaelis, Susanne & Haspelmath, Martin 2003 Ditransitive constructions: Creole languages in a cross-linguistic perspective. Creolica 2003-04-23 (www.creolica.net, open-access online journal). Naigles, Letitia & Erika Hoff-Ginsberg 1998 Why are some verbs learned before other verbs? Effects of input frequency and structure on children's early verb use. Journal of Child Language 25: 95-120. Nelson, Katherine 1973 Structure and strategy in learning to talk. Monographs of the Society for Research in Child Development, 38 (1-2, Serial No. 149). Plunkett, Kim & Virginia Marchman 1993 From rote learning to system building: acquiring verb morphology in children and connectionist nets. Cognition 61: 299-308. Roeper, Thomas 1999 Universal bilingualism. In Bilingualism 2(3): 169-208. this vol. What frequency can do and what it can't. Rowland, Caroline, Pine, Julian, Elena Lieven & Anna L. Theakston 2003 Determinants of the order of acquisition of wh- questions: Re-evaluating the role of caregiver speech. Journal of Child Language 30: 609-^635. Rumelhart, David E. & James L. McClelland 1986 Parallel distributed processing: Explorations in the microstructure of cognition. Vol. 1. Cambridge, MA: MIT Press. Sauerland, Uli, Jan Andersen & Kazuko Yatsushiro in press The plural is semantically unmarked. In Linguistic Evidence, Stephan Kepser & Marga Reis (eds.), 413^34. Berlin/New York: Mouton de Gruyter. Savic, Svenke 1975 Aspects of adult-child communication: The problem of question acquisition. Journal of Child Language 2: 251-260

Introducing the frequency debate in studies of language acquisition

19

Seidenberg, Mark S. & Maryellen C. MacDonald 1999 A probabilistic constraints approach to language acquisition and processing. Cognitive Science 23: 569-588. Snow, Catherine E. 1972 Mothers' speech to children learning language. Child Development 43: 549-65. Snyder, William & Karin Stromswold 1997 The structure and acquisition of English dative constructions. Linguistic Inquiry 28(2): 281 -317. Theakston, Anna, Elena Lieven, Julian Pine & Caroline Rowland 2004 Semantic generality, input frequency and the acquisition of syntax. Journal of Child Language 31:61 -99. Tomasello, Michael 2003 Constructing a language: A usage-based theory of language acquisition. Cambridge, MA: Harvard University Press. Yang, Charles 2002 Knowledge and Learning in Natural Language. Oxford: Oxford University Press Zipf, Gerge 1965 The psycho-biology of language: an introduction to dynamic philology. Cambridge, MA: MIT Press. Original edition, Houghton: Mifflin, 1935.

Opening statement

What frequency can do and what it can't Tom Roeper

1.

Introduction1

Frequency is often alluded to as an ingredient of the learning process. It is very difficult, as we shall show, to articulate what role it could possibly have in a deductive theory of learning built around mental representations. We shall argue that almost every form of an alleged frequency effect reduces to a non-frequency account where every learning event involves the addition of new information. Under a specific notion of parametric variation and Multiple Grammars (see Roeper 1999; Yang 2002), a definable role for frequency effects is possible. Otherwise, allusion to frequency as an explanatory concept makes the actual learning mechanism more difficult to see. In general, we argue that the notion of frequency in the domain of learning is not a coherent concept.

1.1. Overview We can of course make machines that count, like computers, and if the mind is an unconscious machine, then it is possible that it counts unconsciously. Therefore, it is logically possible that counting, resulting in frequency, could be connected to the unconscious activity behind the use of language. It is quite another question to ask whether it is involved in the construction and acquisition of the grammar behind language use. Since the topic has become popular, let us reason as carefully as we can about the relationship between frequency and mental representations. They do not use the same concepts or vocabulary, so it is not clear that one can or must forge a connection between them. Here is a brief synopsis of our argument: If learning changes the mental representation in some way, then we have to understand the mechanism of information change. Counting instances ofXdoes not obviously changed. Learning inherently involves adding information. But again, the addition of information is quite separate from counting information. Each time infor-

24

Tom Roeper

mation is added, the object is new, and therefore not countable as a member of the previous set. Like a word-search on a computer, one change and the word will not be found again. If "similar" things are being counted, then we need a mental representation of the superset defining similarity. What is that and how is it acquired?2 Notions like similarity are where the real mental talents are hidden. However, the effect of using words of that kind, instead of explicit models, is to obscure the object of inquiry. Instead, even if quite inadequate, it is better to have an explicit form of representation, otherwise progress and insight in how mental processes work itself cannot be stated. If we have a representational model, we can systematically modify it. If we have none, then we are always gathering data outside of our real interest. Linguistic theory has been built around always maintaining an actual mechanical model of what occurs, even though, looking back, the initial models were startlingly crude. Current models are far more subtle though still radically imperfect. We will argue that frequency can have a narrow role in the choice of mental representations, where no new information is added to either, but not in their original constitution. This formulation may sound cryptic now. We will go through this argument in great detail, because it is in the details where the core principles of learning are present.

1.2. Frequency from a broad perspective The value of counting throughout scientific enterprises is that many other interactions achieve a numerical reflection without having a counting mechanism built in. Shoes that have taken 10,000 steps will show more wear and tear than those that took 100 steps. So if we want to predict wear and tear, we can count steps. But the shoes themselves do not count. They experience a tiny degree of wear with each step. Nevertheless, counting is an available indirect method for first-guess assessment in many domains. It is crucial however, to ask each time when and whether counting is associated with the core mechanism or its derivative. We will argue that most uses of frequency in psychology are derivative. If you wore your shoes for three years, they are probably worn out. The countable years did not wear out your shoes, walking did it, via friction on the soles.

What frequency can do and what it can't

25

1.3. Frequency perception

Human overt perception of frequency is not reliable according to my informal sources. (I write as a linguist, not someone familiar with the extensive behaviorist and connectionist literature on this topic, but my inquiries among connectionists have not led me to believe that they have a challenge to my reasoning). Piatelli-Palmerini (p.c.) informs me that if asked how often one hears a vowel three letters from the end of a word, people say it is rare. But if asked how often they hear -ing, they have a judgement that it is frequent. But -ing involves precisely a vowel three letters from the end. Yang (2002) as well shows that frequency perception at the phonological level presupposes an analysis. Humans and Tamarinds are equally good at recognizing syllables, but humans are much better at recognizing phonemes (Piatelli-Palmerini p.c.). These differences suggest that the items whose frequency one is sensitive to are themselves pre-determined by pre-existing biases, in our terms, by preexisting templates for mental representations. This presupposes that the templates are present and not under construction. In linguistic terms, the properties of UG are presupposed before counting, they are not the result of counting (see Yang 2002; Chomsky 2005). We turn now to two forms of frequency effect that are external to mental representations, but often seem to be entailed in psychological studies. 2. The magnification problem It is possible that on those occasions where we think many exposures are necessary, in fact we learn a subpart of the expression with a single exposure. This is what can be called the magnification problem, where frequency is apparently operative because our mode of analysis is too gross. Frequency looks like it is operative because we are counting at the wrong level of abstraction, and therefore counting the wrong things. Suppose one were a Martian and observed from a far star the building of houses on earth, just barely visible. Looking at gross movement, one might observe: (1)

a. large houses involve many truck-visits b. small houses involve fewer truck-visits

26

Tom Roeper

Therefore, if you want to build a big house, then visit it more often with trucks. This conclusion completely misunderstands how a house is constructed. In fact, truck-visits reflect the amount of material needed, and the actual building involves all kinds of separate stages and activities that are not visible to the Martian. A frequency analysis is completely misguided. Consider the following linguistic situation: one word is learned only with 26 exposures and another is learned with exactly 5 exposures. It seems that frequency is a factor and operates in a differential way. However, if we now discover that the word that requires 26 exposures has 26 syllables and the word that takes 5 exposures has 5 syllables, then it is immediately obvious that the choice of "word" as the basis for counting was a mistake: one should have chosen syllable. And then it becomes obvious that the words are learned syllable by syllable, with one exposure for each syllable. The frequency effect sprang from the wrong level of magnification. Now if we accept that what is learned is a mental representation, let us see if we can reason as carefully as we can about such a representation. Suppose we have an object: (2)

a. QXPR,

and we put it in our representation: b. QXPR, why do we need another exposure? It is sometimes said that the representation is "reinforced", but what could that possibly mean in informational or representational terms? Any proponent who claims that frequency is relevant to learning owes us an explanation of what the connection between a mental representation and "reinforcement" means. Either the information is present or absent. There is nothing in between. What is usually meant by reinforcement is that the information is not really all there, so more information is necessary. Suppose we just got: c. QX

at the first exposure, maybe like the first consonant cluster, which is enough for a partial representation but not enough for one to actually pronounce anything. Now the next exposure reveals P: d. QXP,

What frequency can do and what it can't

27

and the next R: e. QXPR. Now there is no role for frequency whatsoever. Different pieces of information are added at each exposure, and possibly the hearer did not actually hear the entire expression with each exposure. This argument at the phonological level can be repeated at the phonetic level. We may just get one feature of a phoneme like /z/, not enough to distinguish it from Is/ or /sh/. Suppose one tried the opposite assumption: nothing is represented until one hears it five times. Note that this is really an impossible claim. Suppose one hears: (3)

a. XNJQ,

and it is erased in entirety. Now one hears it again: b. XNJQ. It is impossible to know that you even heard it before, because the entire representation was eliminated. Logically, one just has heard it for the first time each time, so there is no basis for even counting how often one has heard it. The only possible conception of a frequency effect requires that one get some piece of a representation with each exposure, which we have just shown is not a frequency effect at all. The goal for a representation theory must be to capture the level of representation where one can describe what information is added with each exposure. Whenever a frequency effect is claimed for a mental representation, the first question should be whether a more refined level of analysis is called for. This approach may well apply to the phonetic level. One might not even get an entire phoneme, one might mishear something, yet still know that it was a vowel or a consonant cluster. Upon second hearing, with a representation that says only: (4)

"initial consonant cluster",

the hearer might focus on what the properties of the cluster are: was it a stop or a fricative? In effect, the hearer has narrowed the hypothesis space and begun to look at a smaller range of options. The reason that it seems that frequency is involved is again that we are looking at the wrong level of

28

Tom Roeper

magnification. At the right level, every exposure will add information. If we cannot get to that level, then we really do not know what is being learned.

2.1. The discovery problem Suppose we arrive at the following probabilistic judgment: if you look for coins on a beach, you will find a penny with ten tries, a dime with fifty tries, and a quarter with a hundred. This reflects how much change gets dropped on the beach. Each time one searches without success, there is no additional information added to a mental representation, so our previous critique does not apply. However, this scenario refers to proportions in the world being investigated. It is frequency with respect to the learning environment and not the mental representation. It has no implications for a learning procedure. When one finds a quarter, one finds the whole thing at once, no frequency is involved. This is analogous, in language, to the question of whether a child hears or pays attention to a word in his environment. If a child ignores 75% of what he hears, then the child is more likely to hear it once if it is said more often. This uses frequency to describe how a child is exposed to language, but not how it is learned, i.e. represented in the mind. It is still learned with one exposure, or with a number of exposures with new information in each instance. 2.2. No new information criterion The conclusion to which one is led is that frequency is not the right diagnostic whenever new information is added to a representation. A higher level of magnification will eliminate the apparent frequency effect. We can summarize this discussion with the following principle: (5)

Frequency is relevant where no new information is added with each exposure.

Where would such a situation arise? It arises where there is phonological optionality and where UG provides parametric alternatives, a view developed by Yang (2002), to which we return. It is simply a fact, given exceptions, given performance errors, and given the potential for misanalysis,

What frequency can do and what it can't

29

that the child is exposed to optionality or to sentences which could justify opposite sides of a parameter. Here, frequency could be relevant, but there are many deductive properties to consider as well.3

3. Where frequency does not have an effect We begin by outlining where it is fairly evident that frequency has little or no effect. If children learn 14,000 words by the age of six, with roughly one new word an hour, then frequency is not great for each exposure, often just one. When we introduce ourselves to a new person, we tend to say a person's name just once. Therefore, it is obvious that one can learn something with a single exposure. If one fails to "get the name", then part but not all of the information has been represented. The question then arises why any word or morpheme, or even structure, should take more than one exposure. Once again, the natural hypothesis is: more than one thing is learned. The full expression of possible forms in English grammar is often rare. We know that in English (6a) is ungrammatical, but (6b) is grammatical, though it contains a five-part verb that a child may never have heard. (6)

a. * John must will come. b. John must have been being shot at, when the police arrived.

These facts follow from the grammar. English, not German, has only one modal node, so (5a) is excluded. The rule for the full set of auxiliaries, though very rare, allows (5b), so it is allowed, even though it is extremely rare. A Google search for must have been being reveals 900 examples, and for must have been it lists 22,000,000. Chances are very great that the proportions in the environment of 4-year-olds is substantially less than the 22,000/1 it is for adults. Although Pullum & Scholz (2001) argue that rare sentences do occur sufficiently for acquisition, it is not very likely that children have heard such a five-part expression, and it is not one they discuss. (Chomsky, in a lecture, said that he had heard only 15 examples in 20 years of listening for such cases.) Not all words are learned from one exposure. Why not, if one can sometimes suffice? Susan Carey Block (1978)4 showed that "fast mapping" occurred quickly with a gross meaning from a few exposures, while a more exact mapping required many exposures. Is this an example of a frequency effect? Why should one meaning take longer than the other? Presumably

30

Tom Roeper

the child is changing subtle features of a word, or filling in uncertain parts of the meaning with further exposures. In other words, there is an addition of information to the original representation. For instance, the child may be narrowing subtle features, in order to know that chromium refers not just to dark green but to olive green. What happens after fast mapping during refined mapping? Children must hear many more examples, but each helps give a subtle part of olive green. Again, the learning is incremental. The frequency itself does not clearly play a role. It is the additional information that plays a role. In fact, there is suggestive experimental evidence that pits pure frequency against subtle diversity, showing that diversity helps more than frequency.

3.1. Diversity rather than frequency Landau & Gleitman (1985) suggested that verbs were learned by hearing them in a variety of contexts. If we hear the word see in the forms (7a)(7d) then an abstract notion of see that is simultaneously visual and mental can be developed which allows many particular forms to all be realized. (7)

a. b. c. d.

see your hat see what you mean see that John is nice see you to the door

Hearing a word in just one context provides too little information to see the abstraction. Informal pilot experiments, which perhaps should have been elevated to a demonstration about frequency, showed there is a contrast between pure frequency and what we can call "diversity of input". Just as Landau & Gleitman (1985) suggest, we found that it was better to vary the context for novel words rather than to use exactly the same story several times over. Only non-identical cases led to improvement. The learner must keep some representation of each situation to allow it to coalesce into a single abstract meaning. An interesting question is exactly what range of meanings allow a child to confidently triangulate what a word means to the point where they will use it. How does a child decide that a tentative meaning is right on the basis of further evidence?

What frequency can do and what it can V 31

3.2. Pragmatic variation How much does apparent frequency variation reveal? Upon greater magnification, a great deal of pragmatic and semantic variation. In a very close analysis, Kupisch (2006) shows that the appearance of definite articles in Italian and German, like English, is gradual, but that distinct stages fit a variety of pragmatic distinctions. Here is her adaptation of a definiteness tree (from Roeper, to appear) with ages for each type of meaning in Italian: Table 1. First appearance of article functions in Italian5 Function

Associated category

PROPER NAME

D

DEMONSTR. DEICTIC

D

DEFINITE UNIQUE

D

PART-WHOLE

D

DEFINITE EXPLETIVE

D

IDENTIFYING (=SPECIFIC INDEF.)

D

NON-PARTICULAR

NP

GENERIC

NP

NAMING

NP

DEFAULT KIND

N

Example of first occurrence

Age

la mamma voglio Ί want (the) mum' 2;0 // cane 'the dog' 1;7 i l sole 'the sun' 2;0 // naso 'the nose' 2;1 fare la musica 'make music' 2;4 ehe prepari? - un polio 2;4 'what do you prepare? - 'a chicken' non e una chiocciola 'it's not a smail' 2;3 solo i cervi Only deer' 2;8 n gatto 'a cat' I un treno 'a train' 1 ;8/l ;9 cane 'dog' 1 ;8

Since each of these forms is subject to variation across grammars, some captured by bare nouns, some by different articles, the child, to avoid a forced retreat, should receive ample evidence of each usage before he extends the definite article to a new pragmatic domain in a particular language. Much of the change of representation for referential terms may involve pragmatic distinctions, many of which are not yet fully formulated in linguistic theory. Here, particularly, where our own theoretical awareness is dim, the temptation to believe "it is just a matter of frequency" of exposure may be strong. 3.3. The goed/went variation: more subtle semantics One of the most famous examples of child grammar is the presence ofgoed in the grammar of many children. It illustrates the power of categorical

32

Tom Roeper

rules: if -ed is unmarked, then it should apply to all verbs. More interesting, though, is the fact that the move to went has the trademarks of gradualism. How is this to be explained? We argue that the child, like the adult, is operating with far greater subtlety than our rules suggest. It is not true for adults that every form of go has the past tense of went. Among the many meanings forgo is "belong": If one keeps knives in a top drawer for years, and then shifts them to a bottom drawer, we do not so easily say (8a) but rather (8b): (8)

a. * The knives went in the top drawer. b. The knives used to go in the top drawer.

In order for adults to have this exception, they must be paying attention to subtle meaning differences, which would suggest that children do as well. In fact, in the first stage we find that children use only went, then they use goed and then there is an interesting middle stage where at least for one child (Anisfeld 1973), we find the four stages (9a)-(9d):6 (9)

a. b. c. a.

stage 1: went stage 2:goeJ stage 3: goed/went stage 4: went

Now what has not been investigated to my knowledge is what the semantics of stage 3 are like, nor actually how many uses are involved in stage 1 and stage 2. Consider all these usages of go involving "walk", "travel", "say", see (lOc) and "belong": (10) a. b. c. d. e.

combination: pepper does not go with sugar location: it goes in the drawer say: and John goes Ί will' last: the fan will go for two years abstract pattern: the line goes along the edge

Ray Jackendoff (2004), and elsewhere, has articulated these extensions and what they entail. Even within the "move" meaning of go we can use it for short distances or long distances. It is possible that the child uses goed for just "walk" and went for distant travel or for a period of time. The differences may be even more subtle than these, or even rather irrelevant. A child

What frequency can do and what it can't

33

might have / goed and we went, or one could entail intentionality more than the other, or many other possible pragmatic features. It feels, but perhaps mistakenly, that goed is more intentional than went. So we might have we went to town and then / goed to the toy store, where the first reflects a parental decision and the second the child's own volition. It is often difficult for adults to imagine that children function with greater semantic subtlety than they do, but the evidence is that children are on the whole very conservative in their generalizations. That conservatism must reflect sensitivity to subtle variables. Remember that the child is on the lookout for options that occur in other languages, but not English. In German, for instance, es geht 'it goes' can be used to mean "it is grammatical" but we do not have that extension in English, though we have many others. The gradual shift from 50:50 to almost 100% went could reflect the child's stepwise experience of the use of went in a variety of contexts. Could the child at some point, say if 7 out of 10 possibilities were present, simply decide to use went everywhere? This is logically possible, but it would not explain the used to go exception. It is not likely, but we will return to this issue in our discussion of how children learn V2.

3.4. Parameters and frequency What are the ingredients that set parameters? We have effectively outlined two possibilities. One is that the child has a small amount of evidence that sets up more than one alternative, then a pure frequency metric decides which is right. The other mechanism is that the child uses incremental knowledge, accruing a list of contexts, to decide that one side of a parameter is productive and the other is not. If the list is dropped in favor of a categorical choice just in terms of the length of the list, then it is a subtler kind of frequency effect. Let us examine two parameters from this perspective. The first is V2 (see Roeper 1999 for more extensive discussion), where we find that it is a chosen option in German, and a minor variant in English, found only in quotation and stylistic inversion: (11) a. "Nothing" said, muttered, screamed, hollered Jim. b. Into the room ran Bill. Notably, one can add that be is compatible with V2, which could influence a child:

34

Tom Roeper

(12) a. b. c. ά.

John is not here *John do not be her tall is this man *tall do this man be

Presumably the German child hears not only these cases, but thousands of others. One must now ask, does he pay attention to the raw amount, or to the properties of those verbs? The verbs of quotation are potentially a large class, but they are defined exactly as verbs of quotation. In order for the English child to avoid overgenerating, he must pay attention to the verb class (quotation), and the emphatic nature of stylistic inversion. He may also pay attention to the class to which be belongs. There is a small amount of evidence which I collected that suggests precisely that the child pays attention to the notion of class and generalizes within it. One child, for a week or so, said both (13a) and (13b): (13) a. what means that b. what calls that The examples notably involve meanings that are equative in nature, and therefore, like is in what is that?, this brief overgeneralization of V2 was presumably overruled when the child readily, perhaps within a day or two, heard cases like (14) where there is do-insertion instead of V2. (14) What does that mean ? The German child will presumably start the same way: V2 for particular verbs or classes. Within days or hours, however, he will hear verbs of movement, eating, thinking, and at some point he substitutes for all of the classes the simple concept verb. Thus he moves from a representation like (15) to a categorical definition that refers to the notion of verb, see (16): (15) NP IP say, think, move... (16) NP IP verb It is also possible that the diversity of verbs which leads to selecting the whole verb category has only to do with the expanse of meaning: if verbs of doing, thinking, intention, raising all allow V2, then V2 is productive, no more of a list is needed. This could occur with a small number of verbs if they were sufficiently diverse.

What frequency can do and what it can't

35

Notice that the two methods implied here are ultimately different. Both depend upon keeping track of a number of uses. One, however, keeps track of frequency of V2 purely syntactically by counting, while the other is expanding a set of verb-class or pragmatic meaning alternatives to a point where the list is too long and a categorical decision is made: all verbs undergo V2. The second form is also a kind of frequency count, but at a different level of abstraction where V+verb-class is registered. The notion "the list of particulars is too long, so let us allow all verbs to undergo V2" entails a notion of frequency, but the informational basis is crucially different.

3.5. Pro-drop Yang (2002) argues that pro-drop reflects frequency variation by looking at results obtained by Valian and others which show a steady decrease in the missing-subject phenomenon. Here again we can apply our "no new information" assumption, and argue that each side of the parameter represents exactly the same range of pragmatic and semantic variation. This is, however, yet to be determined. We need an analysis that is fine-grained, like Kupisch (2006) above, to be sure whether it is pure frequency which the child keeps track of, or a set of meaning alternatives that in turn might be sensitive to frequency. This is particularly relevant in the case of pro-drop where social register factors are relevant: pro-drop entails not only some form of given information, but also a notion of informality in English. Logically, one might imagine that the child first defines pro-drop as argued in Roeper & Rohrbacher (2000), as a Chinese grammar that has no tense, for which Yang shows evidence (namely that missing subjects are possible only with adjunct wh-, as it would be in Chinese). When the child realizes that the language has empty subjects, like there is a dog, then the Romance parameter is introduced. (17) -Pro-drop (+there) +pro-drop (-there) Then, however, the child experiences subjects in formal situations with a clear introduced subject. Therefore [-pro-drop] is supported for those cases. In addition he hears expletives for many of the expletive verbs: happen, matter, continue. He continues to hear: (18) a. Got a cigarette?

b. Would you like to have lunch? Seems nice to me.

36

Tom Roeper

The sentences are then support for +pro-drop in two environments: edges of sentences, and with a few expletive verbs seems. These are very limited, so this side of the parameter is essentially exceptional and the other becomes productive and defined in categorical terms: all sentences require subjects. Here is where pure frequency can play a role. If the child regards the subject as optional in informal situations, then what would tell him that +pro-drop is wrong? The answer would be that if he never hears a missing subject except for certain verbs, then it is clear that the -pro-drop option is favored. So he must count the number of informal environments where it does not happen. He must also decide in many ambiguous situations exactly what "informal" means. Then he concludes that English is -pro-drop and +pro-drop only for well-defined lexical exceptions and edge phenomena. The child in Spanish experiences the fact that the subject is dropped in formal environments as well as informal, but the English child does not. The Spanish child's grammar will then allow it everywhere where prior discourse has introduced a possible subject. However, the actual statistics run the other way, English children do not add pro-drop environments, but subtract them, so we must reverse the logic. English children initially allow pro-drop everywhere and gradually restrict it to certain verbs with expletives and informal dialogues. If the restrictions are systematic, which requires a very close analysis of the pragmatics of the data, then we would find that the child is gradually adding contexts where subjects are obligatory. In both grammars, a prior discourse reference is needed to justify the deletion. In English, this occurs only with informal contexts. With Spanish, it occurs with all contexts (although this is also a simplification). Presumably the child has a complex definition of "formal". Is formal just outside the family? Is formal just beyond peers? Is formal only with strangers? Is it only for the first five minutes with strangers? Is there a category of older strangers? One can imagine that the set of honorifics applied to socially differentiate people in Japanese could apply to the definition of formality that determines where subjects are deletable, as a UG option. In order to deal with this problem, that the child needs a lot of information to make a grammatical choice, the theory of Multiple Grammars provides an approach (Roeper 1999; Kroch & Taylor 1997; Yang 2002): the child maintains both a pro-drop and a non-pro-drop grammar until enough information arises to decide which is categorical and productive. If it is categorical, then the pragmatically varied representation is dropped and the presence of a prior discourse element is sufficient. This

What frequency can do and what it can't

37

argument is itself an idealization, since there are surely pragmatic distinctions still available to the child and adult. If the child determines that in many formal contexts, pro-drop occurs, then he knows that he is in Spanish. We are suggesting that the child does not simply count up the presence or absence of subjects, but keeps track of contexts as well. Then the frequency concept can be applied to the notion of +subject/+formal. Here one could make an absolute deductive claim: if one formal context occurred, Spanish is chosen. However, given the fact that these contexts are unclear, one can invoke frequency to say: if the subjects are often dropped and the context is not clearly informal, then Spanish is chosen.

4. Past tense frequency and optionality Optionality implies two alternatives that are both known and equally adequate. In other words, no further information is needed for the representation of either. Yang (2002) discusses the case of optional past tense: (19)

dived/dove

Both are used, but neither frequently, so we simply leave them as options in our vocabulary without making a choice. In effect, we have two mini-grammars for past tense and tolerate them both for this verb because they are found elsewhere in the language, but neither is frequent enough to drive out the other. Where there are two options with exactly the same meaning, then the child can pay attention to the one that occurs most frequently. Children, and quite a few adults, adapt the paradigm for sing to bring: (20) a. child: sing, sang, sung', bring, brang, brung b. adult: bring, brought, brought

Because the "mistake" is prevalent, children hear not only their own output, but other children's and adults as well. In fact, a grammar change might be underway. We find among adults quite frequently: (21) we brang three bottles, but much less frequently: (22) three bottles -were brung.

38

Tom Roeper

Children hear both and eventually tend to substitute brought for brung and the basis could be purely statistical since it is a fairly high frequency word.

4.1. -ed Is this always the case? It is far from clear. A famous case from the Connectionist literature purports to show that past tense -ed is acquired gradually. However, it is not clear that -ed is a single object. It has at least four dimensions: (23) a. b. c. d.

past: John walked participle: the house was painted adjectival: the three-cornered roof implicit agent: it was poured to drink

If children acquire these one-by-one, then it would be expected that simply counting occurrences of -ed would give the appearance of gradual acquisition. In a sense, there is gradual acquisition of the full range of forms of -ed, but it could easily be the case that each one is acquired separately and suddenly when a rule is realized. These four versions of -ed mean that there is no object -ed as such in the grammar. Moreover, even with past tense, there are variations absent in English that the child could use initially. For instance, African-American English distinguishes between past and remote past. Wagner (1998) reports that initial -eduse is linked to telic verbs. This would again produce the impression of gradualism. It is an error to lump -eds together, just as -ing in sling, no swimming, and is swimming each have a different status. We are far from the bottom of factors that can create gradualism in grammar, and we may not have identified all the relevant factors if subtle semantics, pragmatic, and speech register information are all operative. 4.2. Transition probabilities Classical transition probabilities are where the most arguments have been made on behalf of the effect of frequency. Suppose one tabulates transition probabilities between X and Υ, Χ and Ζ, Χ and Q. Now what have we found? The set of possibilities already exists: XY, XZ, XQ. Nothing about the inner content of those sets is open to revision.7

What frequency can do and what it can't

39

Now suppose those consonant clusters in language X are only chosen as part of the primary grammar if they exhibit frequent transition probabilities. The others are relegated to another system, in effect, a different grammar. If we hear an African name like Mboya, we create a very small new grammar with /mb/ as an initial cluster, unavailable in English. We do not throw it away. It will occur in English rarely, though some possible examples exist, like tomboy which might have a different syllabification if /mb/ is part of the system. The rarity of/mb/ will simply not allow it to be entered into the phonology of English. Now we have reduced the notion of transition probabilities to the process of creating and choosing among alternative grammars. The organizing of the set of consonant clusters is not simply a question of adding frequent ones to a list of acceptable ones. If a cluster is added, then an analytic system will determine how it fits in, whether the acceptable phonetic rules need to be altered. An act of analysis, not just addition, is entailed. If it does not fit into the phonological logic of the grammar, then, we argue, it is represented as a subgrammar of a different type. Thus, even if Mboya were to become an American whose name we heard frequently, it would remain in a separate phonology because it does not obey the English rules. A logical possibility is this: prior to establishing any rules, a child could list the consonant clusters that he hears and simply say the grammar must contain the frequent ones and then projects rules that will make sense of them. Does it work this way? That is an empirical question. Imagine a hypothetical language in which there were 5 consonant clusters drawn from 5 incompatible grammars. It is quite possible that it would be essentially unlearnable, at least in a productive system, because no rules are recognized. If true, then it is only the recognition of rule relations that would actually enter a consonant sequence into a grammar. This remains an empirical question which I do not have the technical knowledge to explore further.

4.3. The shift from pragmatic to deductive Does frequency alone guarantee obligatory usage? It is very possible that information from a different module can enforce obligatory usage. Schmilz (2004) found that in monolingual German and bilingual children the use of articles was very haphazard, just as was reported in Kupisch above. She observes, however, that when case assignment is realized, then the articles

40

Tom Roeper

are suddenly no longer optional, but obligatory in a sharp break. In other words, one would find that in a sentence like (24) the articles were either present or absent. Later, sentences like (25) occur. (24) Er gibt die Mann die Hund. he gives the man the dog 'He gives the dog to the man.' (25) Er gibt dem Mann den Hund. he gives the. DAT man the-ACC dog 'He gives the dog to the man.' Why was the article optional before case marking? We do not know, but we can guess that the article was sensitive to a range of semantic and pragmatic distinctions that are, in effect, suddenly blocked by the presence of the case module which requires visibility. The interaction of modular information could have the effect of eliminating a form of variable pragmatic sensitivity because the case module does not tolerate optionality. If one looks at the kind of pragmatic variation associated with intonation, which have been described as coercive, concessive, agreeing, demanding, and so forth, such variation might get associated with articles as well. Consider a case like (26): (26) This is THE restaurant to go to in New York on New Year 's Eve. It means something quite different from: (27) This is the best restaurant in New York to go to on New Year 's Eve. The stressed the is not only emphatic, but implies an "incrowd" meaning. The child might be searching for many such special meanings that we do not detect. They are all wiped out when a deductive factor overrules that kind of pragmatic variation. 5. Recursion and productivity Under the Multiple Grammars approach, one grammar or one option inside a module wins. What does "winning" mean? We suggest here an idea explored elsewhere, as one identifying feature of the grammar that dominates.

What frequency can do and what it can V 41

(28) Recursion operates in the primary grammar. Recursion can be found in a universal form in the operation of merge and in complementation, though much variation is involved in how it is expressed (see Roeper & Snyder (2005), Snyder & Roeper (2004) for extensive discussion of the role of recursion in grammar acquisition). Thus, English allows recursive possessives where German does not, see (29): (29) John 's friend's mother's car These forms are not frequent, so it cannot be the case that the child counts up how many he hears. Instead, once a categorical decision is made, then the grammar allows the formation of a rule which refers to itself, a recursive rule. It is notable that under recursion we have a completely compositional reading: only possession is implied. In a local, non-recursive domain, all sorts of meanings are possible. A single possessive remains local and can acquire lexical-like or construction specific readings. Consider a case like (30): (30) John gave Bill a friend 's advice. Now that is a special kind of advice, friend-style. Note that this reading is lost if we embed the possessive: (31) Bill 's friend 's advice In (31), friend"s just fixes a relation to Bill, but not to style of advice. One can rebracket the phrase and get (32): (32) Bill's "friend's advice" In (32) Bill is advising in that special way. Recursion, once again, operates on the categorical basis of the grammar, and therefore it cannot operate where numerous lexically-specific or construction-specific factors are registered. Therefore, recursion blocks the kinds of special, pragmatic readings that the child may explore where constructions are not recursive. In the spirit of Construction Grammar, one can argue that constructions themselves, as long as they are not recursive, are extended lexical items that can accrue quite unusual meanings and pragmatically specific uses. This line of reasoning goes beyond our topic, but it is instructive. Recursion, outside of

42

Tom Roeper

the universal property of merge, cannot be frequent. We do not have many noun phrases with three adjectives, so it is no wonder that three-adjective noun phrases do not have special readings, while one- adjective noun phrases often do. For instance, (33) can mean he tells big lies. But the expression in (34) means the person is big and strange not that he tells big, strange lies. (33) he's a big liar (34) he 's a big, strange liar Therefore, we have here a fundamental constraint on what is worth counting. The child who excludes recursive rules from associated pragmatic factors, says in effect, never count how often double possessives occur in formal or informal situations, for instance. Universal Grammar will no more allow that computation than to include a computation of how many feet apart speaker and listener are when relative clauses are used.

5.1. Reinforcement: Language-external role of frequency If frequency is a poor diagnostic within grammar, does it help in the realm of language use, where all the other properties of mind become relevant? Once again, the notion that we reinforce knowledge or strengthen connections may seem attractive. Once again, the same arguments apply. Does reinforcement not change grammar, but somehow change its relevance to use? Do we use words more easily if we have used them frequently before? The notion of reinforcement seems to imply not that a representation is changed, but that its status is changed with respect to other mental systems. Now we are really addressing ourselves to a totally different system, but the core arguments apply. Imagine a pragmatic or social status, or even a neurological matrix whose requirements must be satisfied. The neurological map is again a representation of information where frequency will translate into something else. Without at least a minimal statement of the representation involved, the observation leads to no new insights and if accepted as a form of explanation, distorts our view of the mind. I have found that defenders will say that some physiological does object actually change, like a muscle growing fatter with use. In such an instance of course, we do not say "muscle plus frequent use" we would measure the muscle (show fat biceps) or find some other mode which would imply that

What frequency can do and what it can't

43

frequency is a misleading abstraction. Alternatively, one finds that researchers will admit that every experience is unique and that it is its added qualities that somehow add to the object. Again, quite obviously, a real explanation would focus on the added qualities, not the fact of repetition. This elucidation is much more than an inconsequential clarification. Many statistical models are equipped only to manipulate a given object, not to change its character. Therefore, the research agenda is built around how to increase frequency, study their effects, rather than to seek a conceptualization of what micro-changes are occurring in the course of repetition. The same argument holds for the concept of memory. Memory, a concept that is not well-understood at all, is supposed to be a possible place where frequency would be present.8 We really do count how often a word (or a knife) is used and put it in a convenient spot, near to the top of a memory bank so it is easy to retrieve. The properties of memory are so different in different domains (vision, language, social) that one can challenge the idea that the concept represents something coherent; that there are really common principles for visual and verbal memory. Nevertheless, the organizing factors are most probably linked to the principles in the domain itself, not to any simple version of frequency. We remember words that are linked to important events, even if they are rare, so word-memory is going to be a very complex object where frequency, again, may obscure rather than highlight the true ingredients. One could imagine that there are 50 distinct kinds of memory, each linked to different mental abilities or representations. One could discuss changes in various representations without every invoking the notion of memory quite profitably, I would imagine. Instead, we have only a few distinctions, like long-term and short-term memory, or visual and verbal memory, leaving out most of mental life. Those changes could involve biological growth with no new information, a kind of addition over time and subtraction over time, if we could state how a particular representation naturally allows subtraction and constancy over time, which is true of much of our knowledge. Consider the diverse kinds of memory that are involved in our concept of a person: size, voice, smell, social status, sexual attitude, style of movement, and a host of personality variables which might each require separate kinds of memory: social shyness and intellectual persistence are such different qualities that the array of experiences called upon to build up a representation probably call for distinct kinds of memory.

44

Tom Roeper

5.2. The moral dimension Like it or not, the ideas advanced in cognitive science have an instant impact in the real world of parenting and education, even if we label them a "theory" or confess to their incomplete status. The acceptance of frequency as a legitimate concept leads to the use of repetition as a method of changing verbal and other kinds of behavior. Speech pathology uses "repetition" as a major technique of intervention. It follows naturally from the "frequency" doctrine. It is much easier to repeat something than to develop the indirect methods of explanation that a real understanding of the relevant mental representation involves. For instance, if one simply repeats the demonstration of how to tie your shoes morning, noon, and night, then the pattern of instruction follows a very different path than if one breaks down tying one's shoes into a series of steps a child can learn one by one, even if the early steps are not actually sufficient to tie a shoe without help. In the realm of language, the path to learning that who? may have a multiple answer could lie in helping the child understand what every means. If one denies that a word like who could have an inner representation that contains several parts which one can identify one by one, then this path of instruction is excluded philosophically from the outset. Reliance on the notion that frequency is playing a fundamental role inhibits parents from looking for a deeper scaffolding upon which to advance the child's mental knowledge. 6. Conclusion The less we know, the more attractive the frequency diagnostic may be. And in fact, it is a first pass guide to where one might focus one's attention. It should not be treated as "explanatory". We have defined one domain, parametric choice, where frequency can have a clearly describable role in choosing among mental representations. That approach says that frequency only applies where no new information is added to a representation. We argued that the language acquisition device, as well as adult grammars, project sets of incompatible grammars. The simplicity of grammar is maintained if we establish separate grammars, rather than attempt to write exceptions into those grammars. For instance, we treated quotation as a Germanic subgrammar which will keep a simple definition of V2, rather than attempting a more complex rule (see Collins & Branagin 1997) where

What frequency can do and what it can't

45

we define the English rule of residual V2 in a way that captures both quotation and non-quotation. The acquisition device, having both V2 and residual V2, must decide which rule is productive and which is limited based on keeping track of what verb classes are involved. German, unlike English, decides to define V2 categorically, generating it everywhere. That decision is dictated by the fact that so many verbs undergo V2. The term "so many" implies a frequency representation defined over verb classes. We have argued that frequency is inherently unable to change a mental representation that exists. Adding information to an existing mental representation is a different mental act from counting instances of a mental representation. Counting presupposes that the items counted are the same. Adding information has nothing to do with counting information. Unless it has a well-defined role in a choice domain, which does not change the knowledge being preferred, it is not clear that frequency is ever an explanatory concept.

Notes 1. Thanks to the audience at the Deutsche Gesellschaft für Sprachwissenschaft symposium on input, Markus Bader, Greg Carlson, Johannes Johannsen, Lila Gleitman, and Charles Yang for helpful discussions. 2. Occasionally, proponents retreat with a cautious assertion that they make no claims about mental representation. However, then no claims should be made about learning, since the mind clearly has mental representations that have been learned, and to which these proponents allude. The phrase "is explained by frequency" is often used, but nothing that fulfills the concept of explanation is offered, nor do we hear that the concept of learning is distorted without this admission. It is important to note that this style of description has great consequences in the applied world that consumes psychological studies. 3. Chomsky (2004) argues analogously that once merge has occurred, the internal information is not open to change, but it is open to "movement", a further operation which does not inherently change the variables being moved. "A natural requirement is the "No Tampering Condition". ... Merge cannot break up or add new features to them." This means that the Mental Representation is fixed - nothing on the inside of these pairs is being learned. We can, however, keep track of how often follows X, etc. That information must fit into a .different system at a higher level. 4. See V. Johnson (2005) for a more recent discussion of fast mapping and references therein for literature on the topic. 5. Table adapted from Kupisch (2006).

46

Tom Roeper

6. There is a great deal of statistical work on this question, e.g. Marcus (1995, 2001). Our focus is upon how individual children progress with an eye on the subtle pragmatics that may be involved. 7. This is again analogous to the No Tampering Constraint in Note 3. 8. Ken Wexler (pc) recently observed that work on how children learned tensemarking now has an explanatory path which reaches from genetics to linguistic theory. He noted that it bypassed psychological concepts like "working memory".

References Anisfeld, Mosche 1984 Language Development from birth to three. Hillsdale, NJ: Erlbaum. Carey, Susan 1978 The child as Word-learner. In Linguistic Theory and Psychological Reality, Morris Halle, Joan Bresnan & George Miller (eds.), 264293. Cambridge, MA: MIT Press. Chomsky, Noam 2004 On Phases. MS, Massachusetts Institute of Technology. 2005 Three Factors in Language Design. Linguistic Inquiry 36 (1): 1-22. Collins, Christopher & Phil Branigan 1997 Quotative inversion. Natural Language and Linguistic Theory 15 (1): 1-41. Jackendoff, Ray 2002 Foundations of Language. Oxford: Oxford University Press. Johnson, Valerie E. 2005 Comprehension of the third person singular /s/ in AAE-speaking children. Language Speech and hearing services in Schools 36 (2): 116-124. Kroch, Anthony & Ann Taylor 1997 Verb Movement in Old and Middle English: dialect variation and Language Contact. In Parameters of Morphosyntactic Change, Ans van Kemenade & Nigel Vincent (eds.), 297-325. Cambridge: Cambridge University Press. Kupisch, Tanja 2005 The emergence of article functions in a German-Italian bilingual child: Steps towards morpho-syntactic explicitness. MS, University of Hamburg. 2006 The emergence of article forms and functions in a German-Italian bilingual child. In Interfaces in Multilingualism: Acquisition, Representation and Processing, Conxita Lleo (ed.), 45-109. Amsterdam: Benjamins.

What frequency can do and what it can't

47

Landau, Barbara & Lila Gleitman 1985 Language and experience: Evidence from the blind child. Cambridge, MA: Harvard University Press. Marcus, Gary F. 1995 Children's overregularization of English plurals: a quantitative analysis. Journal of Child Language 22: 447-459. 2001 The Algebraic Mind: Integrating Connectionism and Cognitive Science. Cambridge, MA: MIT Press. Roeper, Tom 1999 Universal Bilingualism. Bilingualism 2(3): 169-186. 2006 Watching noun phrases emerge. Seeking compositionality. In Acquisition meets Semantics, Veerle van Geenhoven (ed.). Dordrecht: Kluwer. Roeper, Tom & Bernhard Rohrbacher 2000 Null subjects in Early Child English and the theory of Economy of Projection. In Acquisition of Scrambling and Cliticization, Susan Powers & Cornelia Hamann (eds.), 345-396. Dordrecht: Kluwer. Roeper, Tom & William Snyder 2005 Language Learnability and the Forms of Recursion. In UG and External Systems: Language, Brain and Computation, Anna Maria Di Sciullo (ed.), 155-170. Amsterdam: Benjamins. Snyder, William & Tom Roeper 2004 Learnability and Recursion across Categories. In Proceedings of the 28th Annual Boston Uniersity Conference on Language Development, Alenja Brugos, Linnea Micciulla & Christine E. Smith (eds.), 543552. Sommerville, MA: Cascadilla Press. Schmitz, Katrin 2004 Indirect objects and Dative case in monolingual German and bilingual German/Romance language acquisition. MS, University of Wuppertal. Yang, Charles 2002 Knowledge and Learning in Natural Language. Oxford: Oxford University Press. 2004 Universal Grammar, Statistics, or Both. Trends in Cognitive Sciences 8(10): 451-456. Wagner, Laura M. 1998 The semantics of Time in Language. PhD diss., University of Pennsylvania.

The acquisition of determiners

The role of input frequency in article acquisition in early child Swedish Ute Bohnacker

This paper investigates patterns of article use in monolingual early child Swedish and in child-directed adult speech. Article omissions in the adult data are found to be more widespread than previously assumed, especially articleless, "bare" singular count nouns (e.g. sten 'stone' instead of en sten 'a stone') and article omissions in doubly determined nominals (e.g. lilla tummen (little thumb-the, 'the little thumb') instead of den lilla tummen (the little thumb-the; 'the little thumb'). Such omissions in the input may arguably influence the course of acquisition. In the two children studied, an initial determinerless stage (1;3-1;7) is followed by a stage of optional articles (1;8-1;11). Targetlike article provision is reached at 2;0, which is early compared to most other Germanic languages. Definite enclitic articles (e.g. -en 'the' as in s ten-en 'the stone') emerge at an earlier age and are produced at higher frequencies than indefinite prenominal articles (e.g. en 'a' as in en sten 'a stone') and at an earlier age and at much higher frequencies than definite prenominal articles (e.g. den 'the' as in den lilla tummen 'the little thumb'). These child frequency patterns appear to replicate those of the adult caregivers. However, input frequency is argued to be an insufficient explanation for Swedish article acquisition, because of striking mismatches in child and adult article use in other areas, especially bare nouns. Investigations of child-directed adult speech are nevertheless important because they tell us what the immediate target looks like for the young child, which may be different from what linguists and reference grammars tend to assume. 1. Introduction and background For a wide variety of languages, empirical studies have shown that articles tend to be omitted in very early child speech. When articles do emerge, acquisition is typically not instantaneous but gradual: children produce both determined and bare forms of the same noun (e.g. a fish ana fish) in similar

52

Ute Bohnacker

syntactic contexts and with similar semantic and pragmatic functions. Acquisitionists have tried to account for the initial absence of articles and for their variable provision in a variety of ways. Within the nativist generative paradigm - the paradigm that I have been working in - universal grammar (an innate faculty within general cognition but domain-specific to language) is assumed to generate mental representations of language structure. Minimally, language learners need to (a) set parameters, which will trigger discrete properties of grammar, and (b) learn lexical elements (free and bound morphemes) and the syntactic, semantic and phonological properties associated with them. Though some researchers have tried to attribute article omissions to parameter missettings, it is more widespread to explain article omissions via lexical learning, where the child may have a targetlike syntax, but is still learning the lexical elements and, importantly, the properties associated with them, a process that will be gradual and item-by-item. There may also be performance or processing deficits that need to be overcome in order to converge on the target (though often little is said about how exactly this is achieved). Alternatively, Roeper (1999, this volume) and Yang (2002) have proposed that children have coexisting and competing innate UG-constrained grammars. This would mean, for the nominal domain, that an overt-article grammar and a nullarticle grammar, or several more refined grammars, compete with each other, leading to variable production. In other syntactic domains, other sets of grammars will compete with each other. Depending on how well these grammars can parse the input, they get strengthened or weakened, and finally eliminated, until children converge on the adult distribution. Here, frequency may play a role when the child makes a choice between different mental representations (such as different parameter settings or weighting multiple grammars), but does not explain the construction of the actual mental representations, which are seen as UG-generated (cf. Roeper, this volume). On any of the above generative approaches, input frequencies of linguistic forms can influence the course, speed and order of acquisition, and in the present paper I will investigate some aspects of this influence for the acquisition of articles in early child Swedish. Articles and other determiners (demonstratives, possessives, quantifiers, etc.) and the grammatical features that linguists may attribute to them (defmiteness, specificity, number, person, gender, etc.) are part of nominal phrases. A determined nominal phrase is generally thought to minimally have the syntactic structure in (1), where the determiner heads its own func-

The role of input frequency in article acquisition in early child Swedish

53

tional projection (DP, determiner phrase) on top of the lexical noun phrase (NP). More elaborate structures have been proposed as well, but will not be discussed here. (1) [DP ...[DP D [ N P . . . [ N , N ]]]]

There is, however, little consensus on whether bare, i.e. determinerless, nominals in adults (e.g. milk, fish) project to the DP level (with an empty or covert D), or whether they are simply NPs. Children's bare nominals (which are often instances of article omission) could likewise be characterized as NPs only, and variability in children's article provision as competition between a DP-grammar and an NP-grammar (see Kupisch 2004 for this line of reasoning). During the 1980s and 1990s, children were often said to lack functional categories and projections (e.g. DP) in their mental representations until the age of ca. 2;0, at which this functional structure would mature in a genetically pre-determined manner, much in the same way as nonlinguistic processes like teething (e.g. Radford 1990a,b). However, other researchers were quick to point out that such theories of maturation made the false prediction that children should acquire articles roughly at the same time (2;0) across languages, while empirical studies showed just the opposite: individual children often differ in their rate of article acquisition, and crosslinguistically, articles have been found to emerge at very different ages and reach acquisition criterion at different ages for different languages. There are too many studies to be cited here, but to name a few, Lleo (1998, 2001) showed that both monolingual and bilingual infants start to produce articles earlier in Spanish than in German; Kupisch (2004, this volume) showed that both monolingual and bilingual children acquire articles earlier in French and Italian than they do in German, and in earlier work (Bohnacker 1994, 1997, 2003) I showed that articles are acquired earlier in Swedish than what has been reported for child English and German. Moreover, different types of determiners (e.g. definite articles, indefinite articles, demonstratives) are often found to emerge at different points in time within a language for the individual child, and the same goes for the point at which such morphological markers reach acquisition criterion. For these empirical reasons, theories of maturation of functional structure have largely been abandoned. Alternative explanations try to relate cross-linguistic differences in article acquisition to individual and/or language-specific properties of the in-

54

Ute Bohnacker

put. Researchers of various theoretical persuasions increasingly realize the necessity of investigating the input that children actually receive. Recordings show that child-directed speech may diverge from adult-to-adult speech and from the ultimate target language, and sometimes there are surprising similarities (e.g. distributional patterns and frequency matches) between child-directed speech and early child production. Such findings are highly important, and as linguists we may have to revise previous assumptions about what the input and immediate target is for the child, and sometimes even revise our assumptions about the grammar of the language (variety) in question (cf. Bohnacker 1999 for Icelandic; Kupisch 2004 for Italian). I believe this renewed interest in input issues is a healthy and promising development. If certain morphological markers, syntactic constructions and semantic/pragmatic functions are frequent in the input, will these markers, these constructions, these functions be the ones to emerge early in the child? Will they become productive early on? Will they also reach acquisition criterion early (whichever way that may be defined)? And will low-frequency items emerge late, become productive late, be acquired late? Existing studies tell us that such correlations are not always corroborated empirically - not all high-input frequency forms are acquired early for instance, which suggests that there must also be other factors at play. Apart from determining the effects of statistical structure of the language on the order of acquisition, we may also ask what input frequencies actually tell us about the process of acquisition. In non-generative frameworks, closely corresponding frequency patterns between input and child production have been taken as evidence for stochastic learning, for an acquisition device that takes the statistical properties of the input to be the primary determinant of development, basically computing and matching the frequency of various elements in the input. What is frequent will become "entrenched" (cf. Langacker 1987), what is infrequent will not. The child is assumed to make lexically-specific associations, build a repertoire of learned strings (formulae) and use these to construct schemata and more general schemata (slot-and-frame), ultimately a type of pattern learning that is not domain-specific to language (usage-based, frequency-based, item-byitem acquisition models, cf. e.g. Lieven et al. 2003; Tomasello 2003; Theakston et al. 2004). Constructivists thus assume that the child constructs mental representations from the statistical structure of the input, in the absence of any innate language structure. Nativists, on the other hand, assume that mental representations of language structure are in place from the out-

The role of input frequency in article acquisition in early child Swedish

55

set (UG), but that their specifics are triggered by the input to which the child is exposed. (The input is also the source for lexical learning of morphemes and the properties associated with them.) On this approach, the target of acquisition is not knowledge of frequency; in fact, the statistical structure of the input is only an epiphenomenon and with little explanatory force (cf. Bley-Vroman 2002: 210, 212). In generative frameworks, frequency correlations between input and child production have sometimes been interpreted as evidence for very early parameter setting (e.g. Wexler 1994; Hyams 1996: 94-97), or for early convergence on the adultlike weighting of multiple grammars. Mostly however, researchers have focused on how and why children go beyond what they hear in the input. The indisputable existence of frequency mismatches and children's creative, novel and at times non-targetlike utterances point to a grammatical system that is underdetermined by the input available, serving as arguments against learning models that are purely frequency-based. Of course, lexical learning, learned strings that turn into schemata and frequency pattern learning are vital components of language acquisition, they can influence acquisition orders and thereby affect the development of linguistic knowledge, and I would venture to claim that most generative acquisitionists think so (e.g. Hyams 1992, 2005). However, I do not believe that it logically or empirically follows from this that linguistic knowledge can be reduced to knowledge of the statistical properties of language, and that children do not have access to a generative, i.e. a creative, combinatorial language component and to innate principles of language (cf. also Jackendoff 2005). In this paper I will investigate the course of acquisition of articles in Swedish, by analyzing very early spontaneous production data from two monolingual Swedish-speaking children, Embla and Markus. In Sections 2 and 3, I will summarize relevant aspects of my earlier research (Bohnacker 1997, 2003) and for reasons of space must refer the reader to these works for details. Then I will take a slightly different look at Embla's and Markus's data and home in on the distribution patterns and frequencies of different kinds of articles and articleless nominals in child production vis-ä-vis caregiver speech. I will focus on four potential input frequency effects and discuss whether these can help explain the course of acquisition: definite vs. indefinite article frequency, definite enclitic vs. definite prenominal article frequency, bare noun frequency and bare noun subtype frequency (Sections 4-7). I will argue that even though there are matches between input and early child production, we also find major discrepancies in fre-

56

Ute Bohnacker

quency, early evidence of rule productivity, and thus evidence for article acquisition being more than frequency-based statistical learning.

2. Articles in adult Swedish Swedish has overt morphological definite and indefinite markers. The indefinite article used with singular count nouns is a free prenominal monosyllable, en 'a' or ett 'a' (e.g. (2a) ett far 'a sheep'). There is no indefinite article for mass nouns (substances and collectives), abstract nouns, indefinite plural count nouns, and these therefore occur bare (e.g. _ mjölk 'milk', _ enighet 'unity', _ ko-r 'cow-s'). The definite article is postnominal and enclitic and takes the form of a syllabic or non-syllabic suffix (e.g. (2b)faret 'sheep-the', ko-n 'cow-the'). Swedish also allows multiple morphological definiteness marking, so-called "double determination", in nominal phrases that contain a demonstrative or adjective: here, a prenominal definite article and the definite enclitic article co-occur (cf. (2c—e)). Elements in the nominal phrase agree in number (singular, plural) and grammatical gender (common, neuter).

(2)

a. ett far a.SG.NEU sheep.SG.NEU 'a sheep/one sheep' b. far-et sheep.SG.NEU-the.SG.NEU 'the sheep'

indefinite article

definite article

c. det här far-et the.SG.NEU here sheep-the.SG.NEU 'this sheep'

double determination

d. det far-et the.SG.NEUiirejierf sheep-the.SG.NEU 'THIS sheep'

double determination

e. det svartafar-et the.SG.NEU black sheep-the.SG.NEU 'the black sheep' double determination, adjective e! svarta far-et black sheep-the.SG.NEU 'the black sheep'

single determination, adjective

The role of input frequency in article acquisition in early child Swedish

57

It is less well known that Swedish definite nominal phrases with an adjective need not be doubly determined, but can carry only one definite marker: for certain classes of adjectives, there never is any prenominal definite article, and for other adjective classes, the prenominal definite article of the double determination construction is optional in many varieties, cf. (det) svarta far-et (2e'). This fact is often ignored in theoretical accounts of Swedish and also, sadly enough, in studies of Swedish child language, which erroneously classify singly definite phrases such as svarta far-et across the board as non-targetlike prenominal article omissions.1 Another often-ignored fact about Swedish articles is the widespread occurrence of "ungrammatically bare" singular count nouns without the prenominal indefinite article. Such bare singulars have a non-individuated reading, the type-properties of the noun take precedence over the individual (or token) properties (Bohnacker 2003: 201-202,243-247), as illustrated in (3a). This use is highly productive, and not - as some textbooks and reference grammars of Swedish would have you believe - restricted to a handful of idioms or fixed constructions (verb + bare singular, preposition + bare singular). (3) gives a selection of the bare singulars I noticed when listening to the radio and leafing through a newspaper magazine for only a few minutes. (3)

a. nu tar vi _far och _ varg. now take we sheep and wolf 'Now let's talk about/take {the/a/0} sheep and {the/a/0} wolf/ wolves.' cf. nu tar vi ett far och en varg. now take we a sheep and a wolf 'Now let's take a/one sheep and a/one wolf.' b. finns det _ varg däruppe? exists it wolfthere-up 'Are there wolves up there?/Can the wolf be found up there?' c. vill du ha _päse? will you have bag 'Would you like a bag?' (cashier to customer) d. jag har köpt _ ny tekanna. vill du se den? I have bought new teapot want you see it 'I've bought a new teapot. Do you want to see it (i.e. the teapot)?' e. iford _jacka med fuskpäls och _ kapuschong wearing coat with fake-fur and hood 'wearing a coat with fake fur and a hood'

58

Ute Bohnacker

f. det är inte sä lätt att vara _ svensk mamma. it is not so easy to be Swedish mum själv har jag __ kurdisk mamma. self have I Kurdish mum 'It's not that easy to be a Swedish mother. I have a Kurdish mother myself.' g. det är _ solig eftermiddag utanför Kungsholmens kyrka, it is sunny afternoon outside Kungsholmen church, det doftar av vär. it smells of spring 'It's a sunny afternoon outside Kungsholmen church and spring is in the air.' h. _ cykel gär bra, men inte _ bil. bike goes well but not car a sen sä ßnns detju _ taxi pä on. and then so exists it well taxi on island-the bike works well/is fine, but not a car. And then there's a taxi on the island.' i. _ buss avgär klockan 9:00 frän Stationen. bus leaves clock 9:00 from station-the _ guide möter vid fägeltornet kl 10. guide meets at birdtower-the clock 10 bus leaves at 9 a.m. (i.e. There is a bus at 9 a.m.) from the station. There will be a guide meeting you at the bird-watchtower at 10a.m.' As (3d-g) show, bare singular count nouns can easily take modifiers such as postnominal prepositional phrases or prenominal adjectives (e.g. ny tekanna 'new teapot'); "bare" refers to the fact that the noun occurs without an article. Bare singulars occur as predicates (e.g. vara _ svensk mamma 'be (a) Swedish mum'; jag är _ lingvist am (a) linguist') and as arguments (e.g. ha _pase 'have (a) bag'). Bare singulars often occur with copula verbs, as the object of thematic verbs and as the object of prepositions, but they also occur as the subject of existential, presentational and locational clauses, see (3b) and (3f-i).2 Swedish native speakers (and linguists) vary considerably in their acceptance of bare nouns, and when presented with bare singular count nouns they just produced themselves, some reject these in judgment tasks, indicating a mismatch in actual and perceived usage.

The role of input frequency in article acquisition in early child Swedish

59

Many bare singulars occur in contexts where a determined nominal with an indefinite article could be used as well, e.g. (3c) vill du ha _ päse? ('want you have 0 bag') is just as acceptable as vill du ha en päse? ('want you have a bag'), or (3d), where jag har köpt _ ny tekanna ( have bought 0 new teapot') is equally acceptable as jag har köpt en ny tekanna ( have bought a new teapot'). The semantic difference between a determined and a bare noun is often very subtle; the bare noun emphasizes type properties, but does not rule out token reference, as shown in (3d), where a bare singular (ny tekanna 'new teapot') is the antecedent of a token-anaphor (den 'it', i.e. a specific teapot). Theorists are in disagreement as to whether bare singulars (and bare nouns in general) should be analyzed as DPs with an empty D, or simply as NPs, in which case Swedish could be said to have both a DP- and an NP-grammar. Even though there are no large-scale corpus studies of bare nouns as yet, Swedish appears to allow substantially more bare singular count nouns than closely-related languages such as German and English, and this may influence acquisition. 3. Recent insights about articles in child Swedish Existing acquisition studies of Swedish articles have investigated taperecorded naturalistic production data from the Stockholm-based Child Language Syntax Project of the early 1970s (Söderbergh 1973) and Strömqvist & Richthoff s Gothenburg-based corpora of the 1980s/1990s (e.g. Strömqvist, Richthoff & Andersson 1993, partially available on CHILDES). Besides hard-to-come-by internal research reports, there have been few publications on Swedish article acquisition until recently. In Bohnacker (1997, 2003), I hand-searched the corpora of two of the Stockholm and Gothenburg children, Embla (1;8-2;1, 10 recordings) and Markus (1;3,19-2;0,25, 15 recordings), with additional data from Tor (2;3-2;5, 5 recordings), and investigated their nominal phrase development. Interestingly, I found an initial no-article (no-determiner) stage, which, however, is short and can only be documented if data collection starts early enough. It is attested for Markus (1;3,19-1;7,25), but not for the other children for whom recording started at age 1 ;8 (Embla) or even later (Tor). It should be noted that the children's speech in general is affected by processes of radical phonetic and phonological simplification, e.g. simplified syllable structure (CV, VC, CVC), consonant cluster reduction, vowel harmony, consonant harmony, stopping, final consonant deletion resulting in open CV syllables. These processes also affect nominals and may render

60

Ute Bohnacker

them as rudimentary forms that look like non-targetlike, determinerless bare nouns. The no-article stage is followed by an optional-article stage (also attested for other children, cf. e.g. Lange & Larsson 1975; Svartholm 1978; Santelmann 1998). For Embla and Markus there is a rapid increase in article provision, the development is relatively fast and is reminiscent of how articles emerge at an early age in many children acquiring Romance languages (e.g. Spanish, Italian, French), but different from Germanic languages studied to date, such as English or German, where articles typically emerge late and reach adult levels late (cf. e.g. for German, Clahsen, Eisenbeiß & Vainikka 1994: 99; Eisenbeiß 2000; Kupisch 2004; Lleo & Demuth 1999; for Italian, Bottari et al. 1993/94; Crisma & Tomasutti 2000; Kupisch 2004, this volume; for Spanish, Lleo 1997; Lleo & Demuth 1999). Recent findings from Troms0 Norwegian (Anderssen 2006) suggest, however, that child Norwegian patterns with child Swedish. Both Markus and Embla produce nouns with overt articles already at age 1;8 (corresponding to a MLU of 1.01 for Markus and 1.64 for Embla).3 Soon after, they produce sizable percentages of articles and by 2;0, both have reached adult provision levels (>90%). Markus' article provision in obligatory contexts is shown in Figure 1 sample by sample in percent. The rightmost column shows a comparison with his adult caregivers. 100 -

20 -

Figure 1. Markus' provision of articles in obligatory contexts, percent, vs. adult caregivers (based on Bohnacker 2003).

The role of input frequency in article acquisition in early child Swedish 61 Figure 2 displays the same information for Embla and her caregivers. 100 i-

1;8,2

1;9,0

1;9,2 1;10,0 1;10,2 1;11,0 1;11,2 2;0,0

2; 1,0

2; 1,2

Adults

Figure 2. Embla's provision of articles in obligatory contexts, percent, vs. adult caregivers (based on Bohnacker 1997).

Figure 3 illustrates Markus' articles with common nouns, raw figures aggregated by stage (for sample-by-sample figures and a detailed analysis, see Bohnacker 2003). During his initial no-article stage (age 1;3,19-1;7,25) all nouns are bare, i.e. articleless (white columns). Bare nouns predominate also for the next few months (1;8,5-1;9,7), even though enclitic definite articles begin to appear at this point (checkered columns). At age 1;ΙΟΙ;! 1,12, definite enclitic articles become more frequent, and prenominal indefinite articles (black columns) and prenominal definite article (grey columns) are emerging as well. Though not shown in the above chart (Bohnacker 2003), Markus produces high rates of definite articles before producing high rates of indefinite ones (67% (14/21) definite articles vs. 12% (10/85) indefinite ones at age 1;10,14; and 88% (60/68) defmites vs. 62% (16/26) indefinites at age 1;11,12). Soon after, at 1;11,25-2;0,25, adult provision levels are reached for both types of articles and the distribution of articles in Markus' productions closely matches that of the adults in the recordings, as shown by the rightmost column conglomerate in Figure 3.

62

Ute Bohnacker

300 i-

Bare Noun N w/ definite enclitic N W / indef. prenominal article w/ definite prenominal

250 200 150

100

50 0

1;3,19-1;7,25

1;8,5-1;9,7

1;10-1;11,12

1;11,25-2;0,25

Adults, 1/4 of total

Figure 3. Markus' articles and (common) nouns by stage, vs. input, raw figures (based on Bohnacker 2003).

Some examples of Markus' and Embla's use of articles are given in (4)-(6). (4)

Playing with cars and little toy people. *ADU: ja all t ing fmns kvar #. yes everything is still-there *CHI: andagubbe-n. other man-the 'the other man' *ADU: den andra gubben ja de e bara en gubbe i bilen #. the other man-the well it is only one man in car-the *ADU: de ska vara en gubbe till ja. it shall be a man more yes 'There should be one more man.' *CHI: gubb ekaj #. man.STEM [??] (stem + unclear; perhaps imitation of ADU (en) gubb-e till) *CHI: hu. *CHI: hm. *ADU: sä du att du e blötdar ja du ha-r nog kissat+... said you that you are wet there yes you have well peed

The role of input frequency in article acquisition in early child Swedish

*CHI: *ADU: *CHI:

hmm. ha#. engubbe. [...] a man

*ADU: ja. *CHI: en gubbe. a man *ADU: defattas en gubbe #. it lacks a man 'There's a man missing.' (5)

(6)

63

a. *CHI: gubbe-n ärglad man-the is happy 'The man is happy.' b. *ADU: vad är del da? what is this then 'What's this then?' *CHI: en gubbe. a man 'a man' a. *CHI: en bil. a car b. *CHI: bil-en stä. car-the stand c. *CHI: äh den bil-en. no the car-the 'no, THIS car' d. *CHI: bil-ar-nana. all- bil-ar-nanana. car-PL-the.PL all-PL car-PL-the.PL 'the cars, all the cars'

Markus (1;10,14)

(Emblal;8,2)

(Emblal;9,2)

(Emblal;8,2) (Embla 1 ;9,0)

(Emblal;9,0)

(Embla 1;9,0)

In Bohnacker (1997: 68-69, 2003: 217-219), I argued that most children's nouns with articles are productive uses of determiners (N+Art, Art+N, etc.) and not just frozen, unanalyzed wholes. I still take this view because Embla and Markus often produce the same noun both as a bare form and as a determined form with different articles, e.g. with a prenominal article in some

64

Ute Bohnacker

instances, with an enclitic article or a plural marker and an enclitic article in others, as exemplified in (4)-(6). This suggests that the children have separate entries for the noun and the determiners in their lexicons, and that they are appropriately assigning a complex structure to the nominals with overt determiners. What the abstract categories involved in this complex structure exactly are is up for linguistic discussion, and acquisitionists may disagree about whether these abstract categories are given (innate) or constructed by the child; the point I want to make is simply that the children in their productions in the corpora at hand appear to appropriately assign a complex structure to the nominals that feature overt determiners. Moreover, the children's articles do not point to any semantic or pragmatic deficiencies. Most of Embla's and Markus' early nominals with definite articles refer deictically or anaphorically to entities in the discourse, and many of their nominals with indefinite articles are used in the function of naming, as is often the case in early child language. In choosing between definite and indefinite articles in these functions, Embla and Markus appear to make hardly any errors, as far as one can tell from the transcripts.4 However, this need not mean that the children have learned all the semantic functions articles have (cf. Kupisch 2004). Spontaneous production data should of course be complemented by experimental data designed to test semantic and pragmatic competence, e.g. Emslie & Stevenson (1981). Interestingly, different morphophonological and semantic types of articles have been found to emerge at different times for Markus (Bohnacker 2003: 224-243), as had been hinted at earlier for other Swedish children (e.g. Lange & Larsson 1975: 119, 1977: 6, 36; Svartholm 1978). As shown in Figure 3, enclitics emerge in Markus' speech (1;8) before prenominals do (1;10), definite articles emerge (1;8) before indefinite articles do (1;10). But they reach adult provision levels at the same time by 2;0. This staggered development is interesting from an input frequency perspective (see next section), and from a cross-linguistic perspective: is there something about the semantics and/or function of a particular article that makes it easier to acquire? Or is there something about the morphophonological form of a particular article that makes it easier to acquire (e.g. enclitic articles, metrical structure)? I also found that different types of articles were produced with very different frequencies (Figures 4 and 5). For both Embla and Markus, most of their articles are definite enclitics in singly definite nominals (Embla 78%, Markus 51%). Indefinite prenominal articles are less common (Embla 10%,

The role of input frequency in article acquisition in early child Swedish

65

Markus 35%). Double determination (definite prenominal + enclitic) is least common (Embla 10%, Markus 14%). Embla

Embla's parents definite article suffix indefinite prenominal article doubly definite definite prenominal only

Figure 4. Kinds of articles in Embla's productions (1;8-2;1) vs. Embla's caregivers (percent).

As the pie charts in Figures 4 and 5 show, the distribution of the children's article types is very similar to that in the recorded input. It thus appears as if the children are indeed matching the frequency patterns of the input in their output. In the following sections, I will investigate this match in greater detail. Markus

Markus' caregivers definite enclitic article indefinite prenominal article doubly definite definite prenominal only

Figure 5. Kinds of articles in Markus' productions (1;3-2;0) vs. Markus' caregivers (percent).

66

Ute Bohnacker

4. Potential input frequency effect 1: Definite vs. indefinite article frequency Are definite articles the first articles to be produced in Swedish (Figure 3) because they are the most frequent article forms in the input? This has not been explicitly suggested but is implied in Bohnacker (1997).5 As the pie charts in Figure 4 and Table 1 show, Embla's caregivers indeed produce many more definite articles (84%) than indefinite articles (16%).6 Table 1. Definite vs. indefinite articles: Embla

Definite articles Indefinite articles

Embla's caregivers (first 5 samples)

Embla (1 ;8,2-2;l ,2)

84% (67/80)7 16% (13/80)

90%(161/178)7 10% (17/1 78)

Embla produces both types of articles already in her first sample, though raw figures are low and indefinite articles are rare (Bohnacker 1997: 66-67). We must therefore turn to Markus, whose corpus is much larger than Embla's, for determining a developmental order. I hand-searched Markus' samples and found that his caregivers produce definite articles slightly more frequently (55%) than indefinite ones (45%). This distribution appears to hold across the observation period. Caregiver and child frequencies roughly match, as can be seen from the pie charts in Figure 5 and Table 2. Table 2. Definite vs. indefinite articles: Markus

Definite articles Indefinite articles

Markus' caregivers (all samples)

Markus (1;3,19-2;0,2, all samples)

55% (615/1111) 7 45% (496/1111)

63% (235/S73)7 37% (138/373)

Markus starts to produce definite articles before indefinite ones (Figure 3), but reaches adult provision levels for both at the same time. This suggests that even though the frequencies of definite articles and indefinite articles as compared to each other match in child and parental speech, this by itself does not explain the acquisitional path (for a discussion of other, phonological, situational/interactional etc. explanations, see Bohnacker 2003: 232-247).

The role of input frequency in article acquisition in early child Swedish

67

There is another issue that disturbs the otherwise impressively similar distribution of articles in child and caregiver speech (Figures 4 and 5), namely the issue of article omissions. So far, I have concentrated on nominal phrases with overt articles. However, when we compare all the types of nominals that involve common nouns for Markus and his caregivers (i.e. all nominals, except for pronouns, proper nouns and name-like kinship terms), we find a strikingly dissimilar distribution. Markus produces predominantly bare nouns, with an average of 58% (567/927), whereas the adults produce 88% overtly determined nouns and only 22% (327/1487) bare nouns, as shown in Figure 6. From the breakdown by age (Figure 3), we know that during the first months of recordings, Markus' common nouns at first (1;3-1;7) are exclusively bare (100%) and then predominantly so (92% bare at 1;8-1;9, 63% bare at 1;10-1;11), in striking contrast to the adults. Markus

Markus' parents

definite enclitic article indefinite prenominal article doubly definite ;/ definite preI nominal only I bare noun possessive prenominal

Figure 6. Kinds of nominals with common nouns in Markus' productions (1;3-2;0) vs. caregivers (percent).

It is unclear how the heavy preponderance of bare nouns, i.e. the omission of articles, in early child speech could be captured by a frequency-based approach to language acquisition. Unless one can explain why children seem to pay attention to some input frequencies (e.g. definite vs. indefinite articles), but are willing to ignore input frequency patterns in others (e.g. determined nominals being so much more frequent than bare ones), I doubt whether speaking of an input frequency "effect" in the former case explains very much. One might, of course, say that the child only produces bare nouns at first because s/he has to start somewhere and will start with bare

68

Ute Bohnacker

nouns before going on to learn morphological article markers. Note, however, that such a view has little to do with constructivist frequency-driven approaches (e.g. Tomasello 2003; Theakston et al. 2004), according to which the child starts to produce unanalyzed strings that are frequent in the input. As determined nouns are much more frequent in the recorded input than bare nouns (Figure 6), Markus should start producing unanalyzed [N+Art] and [Art+N] chunks. The fact that he exclusively produces bare nouns in the early recordings suggests that the unanalyzed learned string approach is on the wrong track.

5. Potential frequency effect 2: Definite enclitic vs. definite pronominal Are definite prenominal articles produced rarely and late by Embla and Markus because they are the least frequent article forms in the input? There is indeed a close frequency match between children and adults (recall Figures 4, 5, 6): only 9% of the definite nominals produced by Embla's caregivers have a prenominal definite article, and 11% of Embla's, cf. Table 3. 15% of Markus' caregivers' definite nominals involve a prenominal article, and 22% of Markus', cf. Table 4. Thus, the overwhelming majority of definite articles are not prenominal but enclitic. Table 3. Definite articles, enclitic vs. pronominal: Embla

Definite enclitic only Definite pronominal + enclitic Definite pronominal only

Embla's caregivers (first 5 samples)

Embla (1;8,2-2;1,2)

91% (61/67) 9 % (6/67) 0% (0/67)

89% (143/161) 10 % (16/161) 1% (2/161)

Table 4. Definite articles, enclitic vs. pronominal: Markus

Definite enclitic only Definite pronominal + enclitic Definite pronominal only

Markus' caregivers (all samples)

Markus (1;3,19-2;0,2, all samples)

86% (540/631) 12% (75/631) 3 % (16/631)

78% (200/255) 14% (35/255) 8 % (20/255)

The role of input frequency in article acquisition in early child Swedish

69

Low numbers of prenominal articles in Swedish child speech have sometimes been interpreted as non-targetlike article omissions (e.g. Santelmann 1998: 653-657). As I have argued in Bohnacker (2003: 232-236) and will do so again here, such an interpretation cannot be upheld in light of the input these children receive. Firstly, the input frequency of prenominal definite articles (as documented in the recordings) is very low (see Table 3 and Table 4). This is not very surprising because prenominal definite articles run a much lower chance to occur than enclitic articles in Swedish, simply because prenominal definite articles can only occur in nominal phrases that involve a prenominal demonstrative or an attributive adjective, and attributive adjectives are rare in adult-to-child speech cross-linguistically. (In Markus' input samples, only 7% of all nominals with a common noun contain an attributive adjective.) Secondly, as was discussed in Section 2, the input — and colloquial Swedish varieties in general - are characterized by variability, i.e. "omissions" of prenominal definite articles from nominals with adjectives. Let's take a closer look at Markus' caregivers. Their 75 doubly determined nominals (definite prenominal + enclitic) break down into three types: 23 with a stressed demonstrative prenominal followed by N+Art as in (7a); 30 with an unstressed prenominal followed by a stressed demonstrative här/där plus N+Art as in (7b), and 22 cases with an unstressed prenominal followed by an adjective plus N+Art as in (7c). In addition, the caregivers produce 17 singly determined nominals with adjectives (Adj+N+Art), e.g. (8).8 (7) a. nä inte DEN stol-en Markus. no not thestressed chair-the Markus 'No, not THAT chair, Markus.'

(8)

(caregiver, 1 ;3,19)

b. mamma har den HÄR troja-n. mum has the herestressed jumper-the 'Mum's got THIS jumper.'

(caregiver, 1 ;8,05)

c. var e den lilla flicka-n? where is the little girl-the 'Where's the little girl?'

(caregiver, 1;3;19)

a sä kommer _ lilla hund-en. and so comes little dog-the 'And then comes the little dog.'

(caregiver, 1;4,27)

70

Ute Bohnacker

This means that 44% (17/39) of the caregivers' definite nominals with an attributive adjective are singly determined, i.e. in about half of the cases the prenominal article is "omitted", whereas definite nominals with a demonstrative are doubly determined all the time (prenominal + enclitic). This is what the immediate target for the child looks like, and the child productions are quite similar: Markus omits definite prenominal articles from nominals with an adjective, but rarely from nominals with a demonstrative, as shown in Table 5 (raw figures, article omissions are indicated by strike-through). Table 5. Definite prenominal article: Markus

demonstrative Det + N+Art demonstrative Bet + N+Art Art + demonstr. {harldar} + N+Art Art + demonstr. {härldär} + N+Art Art + Adj + N+Art Art + Adj + N+Art Art + nominalized Adj, no N Aft + nominalized Adj, no N

Markus' caregivers

Markus (1;3,19-2;0,2)

23 0 30

35 0 1

0

2

22 17 16 0

1 9 19 6

What Markus and Embla9 appear to be doing, then, is presumably to pay attention to input frequencies of definite prenominal and definite enclitic articles, and most certainly to pay attention to input characteristics, that is to the grammar of the immediate target language. 6. Potential input frequency effect 3: Differences in bare noun frequency across languages Let us now return to articles and article omissions in general (Sections 2 and 3). Could the extent to which children omit articles be correlated with the relative frequency of articleless (bare) nouns in the input? Kupisch (2003, 2004, this volume), for instance, sets out to demonstrate that children start to produce articles later and omit articles for a longer period if they are exposed to a Germanic language with a high frequency of bare nouns. Kupisch (2004: 22-30) convincingly shows that bare nouns in child-

The role of input frequency in article acquisition in early child Swedish

71

directed adult German are substantially more frequent (18% bare nouns) than in child-directed adult French and Italian, but also that there are frequency differences between the Romance languages (French (6%), Italian (11%)). Kupisch argues that these adult bare noun frequencies affect the children's rate of acquisition, as most of her German children start producing articles later than the French and Italian children do (Kupisch 2003). Her proposed bare noun input frequency effect receives support from a cross-sectional elicitation study: German children that are older than their French and Italian peers nevertheless show a higher rate of omission of determiners, with significant differences between the three languages for the youngest children: 59% bare nouns in child German, 41% in child Italian, 19% in child French. (Kupisch 2004: 94-95).10 Such a frequency account of bare vs. determined nouns rests on the assumption that children indeed count grammatical categories and have at least a rough idea of what a determiner is. Can Kupisch's bare noun input frequency account be extended to Swedish? Like German, Swedish is a Germanic language and has bare nouns. And due to the fact that Swedish allows bare singular count nouns in a variety of contexts (cf. Section 2), the input frequency of bare nouns may be quite high, even though exact figures have not been established for any corpus yet. Kupisch would thus predict that Swedish articles emerge crosslinguistically late. I will show that Swedish child-directed speech has indeed a high ratio of bare nouns, at least as far as the recordings of Embla's and Markus' caregivers are concerned, which I hand-searched. To make the database comparable to Kupisch's (2004), I counted as bare nouns (BNs) bare singular count nouns, bare mass nouns, bare abstract nouns and bare plurals. Proper nouns, (namelike) kinship nouns and pronouns were excluded. Bare noun ratio was established as the number of bare nouns divided by the sum of bare and determined nouns. As shown in Table 6 and Table 7, there are 20% bare nouns in Embla's caregivers' and 22% bare nouns in Markus' caregivers' speech. Table 6. Bare nouns: Embla Embla's parents/caregivers (first 5 samples)

Embla (first 5 samples, 1 ;8,2-l; 10,2)

20% bare nouns (23/116)

52% bare nouns (62/119)

72

Ute Bohnacker

Table 7. Bare nouns: Markus Markus' caregivers (all samples)

Markus (all samples, 1;3,19-2;0,25)

22% bare nouns (327/1487)

58% bare nouns (567/972)

These Swedish BN percentages should advantageously be verified for a larger caregiver corpus, but they are higher than what Kupisch found for adult-to-child speech in French (6%), Italian (11%) and German (18%). Yet despite the 20%-22% bare nouns in the sampled input, Embla and Markus do not produce articles as late as Kupisch's German children. Rather, as we have seen, Embla and Markus produce articles early on and reach adult provision levels by 2;0 (Figures 1 and 2). Hence, the input frequency of bare nouns (in Swedish) does not appear to affect the course of acquisition as directly as Kupisch (2004) suggests.

7. Potential input frequency effect 4: Bare noun (subtype) frequency over time There could be another, language-internal, input frequency effect to do with bare nouns, for instance, a substantial shift in bare noun frequency over time, correlating with child production. There might also be a shift in bare noun composition over time. For instance, a breakdown into four semantic types - bare singular count (e.g. päse 'bag'), mass (e.g. ost 'cheese'), abstract (e.g. sömn 'sleep') and plural (e.g. hästar 'horses') - in Table 8 reveals that a full 30% of the adults' BNs are bare singulars. Even though linguists disagree about the acceptability of bare nouns and tend to judge such bare singulars as ungrammatical, we are interested here in what Swedish-speaking parents actually say to their children. Table 8. Bare nouns by semantic type: Markus' caregivers (all samples) Semantic type

Percentages

Bare singular count noun Bare mass noun Bare abstract noun Bare plural noun

30% 31 % 5% 34%

(98/327) (102/327) (17/327) (110/327)

The role of input frequency in article acquisition in early child Swedish

73

Table 9 provides a rough breakdown into five distributional syntactic contexts, BNs in isolation (i.e. BNs that constitute utterances of their own, not being part of a sentence), BNs with a copula verb, BNs as the subject or object of a thematic verb, BNs as the object of a preposition, and "other". Table 9. Bare nouns by syntactic context: Markus' caregivers (all samples) Syntactic context

Percentages

In isolation With copula With thematic verb With preposition Other

22% 33% 33% 11 % 1%

(71/327) (109/327) (107/327) (37/327) (3/327)

A breakdown of Markus' caregivers' BNs over time shows that there is a decrease in the overall frequency of bare nouns (Figure 7). For ease of exposition, I have grouped the adult productions into four periods, the same periods that Markus' productions were grouped into earlier (1;3,19-1;7,25; 1;8,5-1;9,7; 1;10-1;11,12; 1;11,25-2;0,25). o/ /o

Caregivers' bare nouns

35 rΠ All BNs | BNs in isolation -Φ- bare singulars

30 — 25 — 20 15 — 10 —

X. r-

1

1;8,5-1;9,7

*ϋ c G Ν^ O N

prepositions

Ό Ο Ν Ρ*

Μφ

Ο "^ cO O N Q Ό Ό w W Ν^ ,^ωθ

ί,'" a ^ °.|t-

Ν

0Ρ

Ο '3p T3 'ηΐ1 Η 4>Η

a3

Figure 6. Frequency of various prepositions in child speech at five different ages

5.2. Do children follow the pattern evident in CDS? An important finding in this study (presented in Section 3.1) is that the distribution of prepositions in CDS differs from the distributions in texts and conversational language. The difference is observed in the prepositions kod and sa. It means that adults adjust their language to a child's age, and this adjustment changes as the child develops (presented in Section 5.1). The results of previous studies revealed that the frequency of prepositions in adult language is a relevant determinant of their order of acquisition (Johnston & Slobin 1979; Tomasello 1987; Hallan 2001; Sinha 1999; Rice 1999; Andelkovic 2000a, 2000b). On the basis of these findings, we expected patterning in child productions to follow that of CDS so that the prepositions kod and sa would appear early and be used frequently. Results with kod supported our intuitions, whereas results with sa were less consistent. This is evident in the scattergrams presented in Figure 7 that plot prepositional usage as a function of frequency in CDS at three ages. Points located to the right of the trendline indicate early acquisition and rather frequent use. Those located to the left of the trendline indicate delayed acquisition.

Children 's usage of Serbian prepositions

0,2 0,4 children 20

0.2 children 26

163

0,2 children 32

Figure 7. Predicted and observed frequency of kod and sa at three ages (20, 26, and 32 months)

The preposition kod occurs early in children production and, at the age of 26 months, it is used more frequently than in CDS. Later it evidently loses its privileged status. At 32 months it is used as often as in CDS and in a proportion consistent with conversational language. Inspection of the samples reveals that phrases like dodi kod mame 'come to mommy', sedi kod mene 'sit by me', etc., become less frequent as age increases. Unlike the preposition kod, patterning for the preposition sa does not comply with our predictions. It is not present at the earliest age at all, and it has a lower frequency than predicted at the ages of 26 and 32. A possible explanation is that the extensive variation in the meanings for sa exceeds the child's capacity. Inspection of the samples in our study offered support for this interpretation. 5.3. The early input for the homonymous preposition sa and how it is reflected in child language As described in Section 3.1.2, the meanings of the preposition sa are diverse: comitative, ablative, attributive, object, instrumental, etc. This may set a more complex task for young children and cause delayed production of the preposition sa. A one-to-one mapping is easier to learn than one-tomany or many-to-one mapping (Pinker 1984; Slobin 1985; Clark 1993) and our results seem to reflect this developmental principle. Inspection of CDS reveals that multiple meanings of sa were evident in CDS addressed to children at the youngest age (20 months). Examples are presented below.

164 Maja Savic and Darinka Anäelkovic (2)

a. Comitative Jesi se igrao sä njima? be.PRES.2SG REFL play.PERF.2SG with.PREP they.INS.PL 'Did you play with them?' b. Object Igraj se s um. play.IMP.2SG REFL with.PREP it.INS.SG 'You play with that.' c. Place Ma, to se sa Meseca vidi well it.ACC REFL from.PREP moon.GEN see.PRES.3SG 'Well, it could be seen from the moon.' d. Instrument Ne mozes tetu dirat s grabljama not may.PRES.2SG aunty.ACC touch.lNF with.PREP rake.lNS 'You may not touch the aunt with the rake.' e. Attributive Ovu sa cvjeticem. this.ACC with.PREP flower.INS 'This one with the flower.' f. Ablative Skloni prst sa usta. take.IMP.2SG fmger.ACC from.PREP mouth.GEN 'Take your finger away from the mouth.'

In the children's production, the earliest uses of the preposition sa appeared at the age of 26 months. Only three out of eight children used this preposition at that age. Five children were even later with the first usage of sa (32, 38 or 44 months of age). Examples of the earliest uses of particular meanings are presented below. (3)

a. Comitative (Ana, 26 months) Ocu ja sa mamom. want.PRES.lSG I.NOM with.PREP mummy.INS want (to go) with mommy.' b. Object (Nik, 26 months) Sa loptama. with.PREP ball.INS.PL 'With the balls.'

Children 's usage of Serbian prepositions

165

c. Place (Luk, 26 months) Sa dnige strane. from.PREP other.GEN.SG side.GEN.SG 'From the other side.' d. Attributive (Nik, 32 months) Imam konja sa prikolicom. have.PRES.lSG horse.ACC with.PREP carriage.INS 'I've got a horse with a carriage.' e. Ablative (Laz, 38 months) Ja had dodem s posla. I.NOM when.CONJ come.PRES.lSG from.PREP work.GEN 'When I come back from work.' f. Instrument (Ana, 32 months) Hocu s drugim bojicama want.PRES.lSG by.PREP other.INS.PL pen.lNS.PL Ί want (to draw) with the other colored pens.' The variety of meanings of the preposition sa that children have to master is presumed to be the source of its prominent departure from the prediction made on the basis of CDS. It should be noted, however, that deviations from prediction for some other homonymous prepositions were smaller. A preposition of medium frequency za 'for, in favor of, in direction of, behind' occurred at the youngest age and at a proportion consistent with CDS. This also holds for the low frequent prepositions po 'at, by, from, to', and ο 'upon, about'. However, because of their low frequency, deviations of po and ο are difficult to estimate in a sample of such limited size. In general, the pattern of usage for children corresponded closely to patterns for adults. In order to make a definite conclusion about whether the late onset and rare usage of sa is caused by conceptual complexity, or something else, further analyses need to be undertaken. However, the intralexemic frequency of particular meanings of Serbian prepositions is not known, so additional linguistic investigations are necessary first. In sum, deviation of the preposition sa, was statistically detected. It is an important outcome insofar as it indicates that input frequency alone cannot predict children's production, and that accounts of acquisition must consider semantic aspects of prepositions.

166 Maja Savic and Darinka A näelkovic 5.4. Relevance of other factors in children's production of prepositions In the introductory section, we noted that the strong effects of input frequency may diminish the influence of other relevant properties of language so that they become difficult to detect. Although it proved to be relevant in many previous studies (Johnston & Slobin 1979; Kuzcaj & Maratsos 1975; Parisi & Antinucci 1970; Sinha et al. 1999; Meints et al. 2002), the semantic and structural complexity was not emphasized in the present study. The relevance of semantic and structural aspects is more perceptible in qualitative analyses of use, and this holds for our study as well. In spite of their fluent production, children do not always use prepositions the same way as adults do. The following examples point to some non-standard prepositional phrases produced by a child named Ana. (4)

a. Child production (Ana, 26 m.) Ja i medo biti (veliki) kod tavana I and teddy-bear be (high) near.PREP attic.GEN 'Me and teddy bear will be high up near the attic' Target Ja i medo biti veliki do tavana. I and teddy-bear be high up to.PREP attic.GEN 'Me and teddy-bear will be high up to the attic.' b. Child production (Ana, 26 m.) Na flapone. on.PREP ceiling.ACC.PL '(High) on the ceiling' Target Do plafona. Up to.PREP ceiling.GEN.SG '(High) up to the ceiling' c. Child production (Ana, 38 m.) Krav od dijete. cow.NOM of.PREP baby.NOM 'Cow of the baby' Target Dijete od krave. baby.NOM of.PREP cow.GEN 'Baby of the cow.'

Children 's usage of Serbian prepositions

167

d. Child production (Nik, 44 m.) Ονο je iz mora u Krasicima. thisbe out of.PREP seaside.GEN at.PREP Krasici.DAT 'This one is out of the sea at Krasici.' Target Ονο je s mora u Krasicima. this be from.PREP seaside.GEN at.PREP Krasici.DAT 'This one is from the seaside at Krasici.' e. Child production (Nik, 44 m.) Gde je led on se odseljava od sume wherever be ice it REFL move off.PREP wood.GEN 'Wherever there is ice (in the woods), it moves of the wood' Target Gde je led on se odseljava iz sume. wherever be ice it REFL move out of.PREP wood.GEN 'Wherever is ice (in the wood), it moves out of the wood.' f. Child production (Ane, 36 m.) Krv u prstic blood.NOM in.PREP fmger.NOM. 'Blood in the finger.' Target Krv iz prstica blood.NOM from.PREP fmger.GEN 'Blood from the finger.' g. Child production (Ane, 44 m.) Ono od na+mora. that.NOM from.PREP on+seaside.PREP+GEN 'That from on+the+seaside' Target Ono sa mora. that.NOM from.PREP seaside.GEN 'That from the seaside.' It should be emphasized that the uses cited above were not the first instances of particular prepositions in a given child. Although not very numerous, similar examples were present in the productions of other children at different ages. They illustrate that a gradual and complex learning process

168

Maja Savic and Darinka Andelkovic

goes on long after the first usage of particular prepositions. Since semantic and structural factors have been investigated in previous studies, some of which are described in the introduction, we will not elaborate here. Most important to note is that both frequency and semantic factors are observable in the same sample: the long-lasting learning of prepositional forms and functions and the concurrent fluent production by means of successful approximation of input frequency. This highlights the lag between early spontaneous production of prepositions, and their late full comprehension noted in other studies (Leikin 1998; Sinha et al. 1999; Andelkovic 2000a, 2000b).

5.5. Summary of results Statistical exploration of the usage of prepositions in a variety of adult and child samples of language showed how patterning of frequencies in children's production follows their patterning in the input language. Analyses reveal that prepositional use in the adult samples is very stable, showing that the relational meanings which prepositions depict appear in the diverse registers of adult language. Nevertheless, certain differences between adult samples emerged: when addressing children, adults use the prepositions kod and sa more frequently than in texts and conversation. This is pragmatically constrained (as explained in Sections 3.1.1 and 3.1.2), since both are frequently used in utterances typically directed to children when searching for proximity (dodi kod mene 'come to me'; hajde sa mamom 'go with mommy'). Interestingly, this property of CDS declines as the child develops, and this was statistically documented. On the side of children, the findings show that the relative proportion of lexemes is very stable across age samples and is strongly guided by the corresponding proportions in adult language. Child usage of prepositions can be predicted from their proportion in adult language at the amount of 84% of variation. Thus, input frequency provides children with unique experience about the adult's usage of prepositions and prevails in shaping child language over purely conceptual and structural complexity. This enables children to replicate the distributional pattern of adult language very early, producing an early image of apparently successful performance. However, discrepancies associated with the prepositions sa and kod in child usage suggest that adult input frequency might not be the only factor that influences production. Although pragmatic constraints of both prepositions are reflected in the distributional patterning in the input, sa does not

Children 's usage of Serbian prepositions

169

follow the same pattern as kod and as other prepositions, which was predicted from CDS. Notably sa is homonymous, and, according to previous studies (Johnston & Slobin 1979; Pinker 1984; Slobin 1985; Clark 1993), every departure from one-to-one mapping between concepts and linguistic units may be a source of difficulty in acquisition. Thus, the late onset and rare incidence of sa at the early ages in the present study support claims for the complexities inherent in homonymy. Finally, qualitative examples of usage illustrate that meanings and formal properties of particular prepositions are not fully mastered until long after their first incidence. After early appearance, children continue to use lemmas even well before final refinement of their meaning and grammatical properties. A question that still remains to be answered is whether the major effect of input frequency identified in this study mirrors a universal regularity in acquisition, or whether it is an idiosyncrasy of Serbian. On the one hand, strong effects of input frequency argue for a universal regularity. On the other hand, structural differences between languages and particular features of the Serbian language may prove critical. These include: (a) a relatively homogenous grammatical position of relational terms in Serbian; (b) the morphological simplicity of the majority of Serbian prepositions; (c) the need for congruency with the noun case, where one preposition can appear with two or three cases, and Parezanovic's finding (1989) that this is irrelevant in comprehension; (d) high functional overlap amongst prepositions (homonymy, polysemy, synonymy). The first three features facilitate acquisition, whereas the last one may pose some challenges. While in Serbian it seems easy for children to establish the grammatical status of prepositions, it may be difficult for them to choose the proper lemma. One way to cope with such a system could be to rely on frequency in language input. Conclusions as to whether this is a language-specific or more universal regularity depend on the availability of converging cross-linguistic evidence on comprehension, production and conceptual development.

6. Conclusion: Factors in the acquisition of prepositions In recent years, studies have offered an emergentist account of language competence, and our results are in tune with this approach. Instead of being unified under a set of abstract rules, language emerges primarily from idio-

170

Maja Savic and Darinka Andelkovic

syncratic and item-based units of production (Tomasello 2003; Lieven et al. 1997; Lieven et al. 2003; Behrens 2003, 2005). The approach applied here reveals that quantitative properties of language input play a major role in shaping children's productions. Child usage of prepositions proved to be strongly guided by frequency distribution of prepositions in child-directed speech. The pattern helps children to produce language successfully long before full mastery of prepositions. While approximating input frequency and successfully performing the communication tasks (comprehension and production), children explore the form and function of prepositions concurrently with and by means of their usage. It is important that similarity of child language to CDS is larger than its similarity to the conversational and written adult samples. This demonstrates the relevance of the immediate linguistic environment as contrasted with an ideal set of abstract rules. Also, partial adjustment of CDS to a child's age enlightens the dynamic interdependence between child language and child-directed speech. Furthermore, it should be noted that highly frequent prepositions tend to capture conceptually and formally simple relations, and the interdependence can account for the early usage of particular prepositions. Thus, the impact of frequency invites an examination of the influence of other factors. Tokens with the highest frequency mark the simplest relations, so linguistic experience strongly supports prelingual conceptual development, representing its outcomes in the language and preserving them from oblivion. When frequent tokens are conceptually and/or structurally complex (as, for example, the Serbian preposition so), production cannot benefit fully from input frequency. It is important to note that the interdependence between high frequency and conceptual/structural simplicity is statistically testable. Moreover, the contribution of semantic and structural complexity can be statistically contrasted with the contribution of input frequency. The relevance of the semantic and formal factors traditionally was advocated by means of qualitative descriptions of semantic and formal properties of the utterances produced by children or presented to them. However, a quantitative investigation of instances of different meanings and of structural properties of prepositions in spontaneous child language is possible. In order to achieve such an aim, further quantitative exploration of semantic and formal properties of prepositions in adult language is necessary. In conclusion, the findings reveal that children fluently produce prepositions in a pattern that approximates the adult pattern of usage, without achieving full refinement of their semantic and formal properties. Theoreti-

Children 's usage of Serbian prepositions

171

cally, the words are 'empty' at the start, or to be more precise, they encompass only the contents available at the moment of interaction. This explains why early words are so idiosyncratic, context-dependent and non-consistent. Interestingly, it seems that an old discussion that Vygotsky started with Piaget (Vygotsky 1985) regains its relevance: words are being acquired even before they are understood - syntax of words precedes syntax of thoughts. Early productions provide reliable evidence for acquisition of form, not acquisition of meaning. The outcome calls for a thorough semantic and quantitative analysis of prepositional forms and functions in children and adults, as well as for cross-linguistic comparisons of input frequency.

7. Discussion: Input frequency as a factor vs. mechanism in the theory of L A In our opinion, much of the current disagreement about the role of frequency in language acquisition derives primarily from two sources. On the one hand, it is due to a somewhat unclear definition of input frequency per se, while on the other, the insight that the role of frequency depends primarily on the learning abilities of a child at a particular developmental level is often neglected. With respect to the first issue, we consider frequency as an environmental modulating determinant that makes it a factor of acquisition, not a mechanism. Being external to a developing mind, frequency cannot be a mechanism, as a mechanism is a neural and/or psychological process that underlies developmental change and determines the nature and content of representations. Mechanisms are means by which factors like frequency realize their influence. For instance, mechanisms of associative learning are related to frequency in that highly frequent items are learned before less frequent ones. Thus, factors enabled via mechanisms accelerate or impede developmental change. However, factors (e.g. frequency, semantic complexity, etc.) do not form mental representations, and do not generate developmental change. Therefore, they cannot be considered explanatory concepts. When input frequency is considered to be a factor, rather than a mechanism, its contribution can be empirically evaluated. Although rarely declared so, this is what most of the studies that test input frequency do. The problem, however, starts when interpreting the results and forming conclusions where representations are inferred on the basis of effects of input

172

Maja Savic and Darinka A näelkovic

frequency. In our view, effects recorded in linguistic and statistical investigations (like the study presented here) may well contribute to the body of knowledge on language acquisition by describing formal, semantic or distributional properties of language samples, and by revealing statistical correlations between the variables of interest. They do not, however, reveal anything about the nature and content of representations. Conclusions about the nature of representation could not be achieved by means of linguistic and statistical analysis, (even when they are well supported by logical argumentation,) without appealing to learning/neural mechanisms. Unlike Roeper (this volume) who suggested that One could discuss changes in various representation without even invoking the notion of memory quite profitably', we believe that inferring about representations on the basis of structural properties of language misguide linguistic and psychological research into isomorphisms between language and mind, which fail to receive any support from developmental cognitive and neural science. Second, it is important to emphasize that the impact of particular factors is very much dependent on the learning abilities that a child has at his or her disposal at a particular developmental level (Piaget 1959; Vygotsky 1985; Smith 2000; Bloom 2000). Thus, input frequency has a major role in what will be learned (most frequent items first), not how it will be learned. How language will be learned is determined by the constraints of a child's developing mind. At the very early stages, when the sensory-motor system and basic reflexes are essentially the only tools at the child's disposal, only dumb mechanisms of statistical learning are applicable (Elman et al. 1996; MacWhinney 1998; Smith 2000). Children's cognitive architecture changes over time and, later on, new and more complex learning mechanisms emerge during development. Likewise, long-term semantic representations change with usage and exposure to linguistic constructions in communicative events (Seidenberg & MacDonald 1999). As pointed out by Seidenberg et al. (2003), the limits of statistical learning are not yet known, and the same holds for grammar-based learning. Language is an assemblage of different abilities that may be activated at different times and in different contexts. Thus, the research agenda is not how to explain representations on the basis of frequency, but how to explain the integration of fundamental mechanisms and the emergence of complex dynamical system of abilities and concepts on that basis.

Children 's usage of Serbian prepositions

173

Appendix The table below shows the proportion of prepositions at different age samples of children's language and samples of written language, Child-Directed Speech and conversational language.

u na s(a) od za kod po iz do pored iza preko bez oko ispred ispod kroz 0

pokraj usred kraj

Σ

20

26

32

38

44

Written

CDS

Convers

0.46 0.19 0 0 0.12 0.15 0 0 0 0 0 0.04* 0 0.04* 0 0 0 0 0 0 0 1

0.36 018 0,06 0.05 0.04 0.18 0.04 0.01 0.02 0.03 0.01 0.01 0.01 0.01 0.01 0 0 0 0.01 0 0 1

0.36 0.21 0.09 0.05 0.13 0.09 0.04 0.01 0.01 0 0 0.01 0.01 0 0 0 0 0 0 0 0 1

0.33 0.15 0.14 0.13 0.09 0.07 0.02 0.02 0.03 0.01 0 0.01 0 0 0 0 0 0 0 0 0 1

0.25 0.20 0.13 0.13 0.10 0.04 0.09 0.02 0 0.01 0.01 0 0.01 0 0 0.01 0 0 0 0 0 1

0.33 0.16 0.09 0.08 0.1 0.01 0.02 0.01 0.03 0 0 0.01 0.02 0.01 0 0 0.02 0.03 0 0 0.01 1

0.26 0.20 0.16 0.07 0.1 0.09 0.03 0.02 0.01 0.01 0 0 0.01 0 0 0.01 0.01 0.01 0 0 0 1

0.33 0.17 0.09 0.06 0.08 0.06 0.02 0.04 0.04 0 0 0.01 0.01 0.01 0 0 0.01 0.02 0 0 0 1

"unproductive usage only (songs and rhymes)

174

Maja Savic and Darinka Andelkovic

Notes *

1.

2.

3. 4.

The data for this research were retrieved from the Serbian Corpus of Early Child Language, which was supported by RSS Grant No. 256/1998. We would like to express our appreciation to Laurie Beth Feldman, Draga Zee, Aleksandar Kostic, and Insa Giilzow for helpfull comments on this manuscript. Address for correspondence: Laboratory of Experimental Psychology, Faculty of Philosophy, University of Belgrade, Cika Ljubina 18-20 11000 Belgrade, Serbia. Andelkovic, Seva and Moskovljevic (2001). The corpus is compiled into electronic form according to the CHILDES system (MacWhinney 1989; MacWhinney and Snow 1985) and contains longitudinal recordings of spontaneous interactions of eight children (4 boys and 4 girls) with people in their family environment. Recordings cover 16 age samples from 18 to 48 months at 2months intervals and each session lasted approximately 90 minutes. In lexical decision experiments, it was demonstrated that frequency estimates derived from daily press are better predictors of processing latency than estimates derived from other registers, like prose and poetry (A. Kostic 1996). www.serbian-corpus.edu.yu Each point in Figures 4 and 5 is a value of coefficient of determination r2 in a regression analysis, in this study. Each is treated as a measure of similarity between samples. Following the changes in values of r2 over ages can give us insight into increasing similarity between CDS and other adult samples (Figure 4), or children's increasing ability to replicate the pattern of adults' usage of prepositions (Figure 5).

References Andelkovic, Darinka 2000a Razvoj imenicke morfologije srpskog jezika na predskolskom uzrastu. (Development of noun morphology in Serbian preschool children). Psihologija XXXIII (1-2): 145-169. 2000b Razvoj razumevanja imenickih sintagmi sa prostornim znacenjem na predskolskom uzrastu. (Comprehension of spatial meaning nounphrases by preschool children). Psihologija XXXIII (1-2): 145-169. Andelkovic, Darinka, Nada Seva & Jasmina Moskovljevic 2001 Srpski elektronski korpus ranog dec/eg govora (Serbian Electronic Corpus of Early Child Language), Laboratory of Experimental Psychology, Faculty of Philosophy, and the Department of General Linguistics, Faculty of Philosophy, University of Belgrade. Barrett, Martyn 1995 Early lexical development. In The Handbook of Child Language, Paul Fletcher & Brian MacWhinney (eds.), 362-375. Oxford: Blackwell.

Children 's usage of Serbian prepositions

175

Behrens, Heike 2003 Verbal prefixation in German child and adult language. Act Linguistica Hungarica, Vol. 50 (1-2): 23-55. 2005 The input-output relationship in first language acquisition. Language and cognitive processes 21 (23): 2-24. Bloom, Lois 2000 The intentionality model in word learning: How to learn a word, any word. In Becoming a word learner: A debate on lexical acquisition, Roberta Golinkoff, Kathryn Hirsh-Pasek, Lois Bloom, Linda Smith, Amanda Woodward, Nameera Akhtar, Michael Tomasello & George Hollich (eds.), 19-50. Oxford: Oxford University Press. Bohnacker, Ute this vol. The role of input frequency on article acquisition in early child Swedish. Bowerman, Melissa & Soonja Choi 2003 Space under construction: Language-specific spatial categorization in first language acquisition. In Language in Mind: Advances in the Study of Language and Thought, Dedre Gentner & Susan GoldinMeadow (eds.), 387-428. Cambridge, MA: MIT Press. Brown, Roger 1973 A first language: The early stages. Cambridge, MA: Harvard University Press. Clark, Eve 1973 Non-linguistic strategies and the acquisition of word meanings. Cognition 2: 161-182. 1977 Strategies and the mapping problem in first language acquisition. In Language learning and thought, John Macnamara (ed.), 147-168. New York: Academic Press. 1993 The lexicon in acquisition. Cambridge: Cambridge University Press. Cromer, Richard 1974 The development of language and cognition: The cognition hypothesis. In New Perspectives in Child Development, Brian Foss (ed.), 184-252. Harmondsworth: Penguin. Dromi, Esther 1979 More on the acquisition of locative prepositions: An analysis of Hebrew data. Journal of Child Language 6 (3): 547-562. 1987 Early lexical development. New York: Cambridge University Press. Ellis, Nick 1998 Emergentism, Connectionism, and Language Learning. Language Learning 14(4): 631-664. Elman, Jeffrey, Elizabeth Bates, Mark Johnson, Annette Karmiloff-Smith, Domenico Parisi & Kim Plunkett (eds.) 1996 Rethinking innateness: A connectionist perspective on development. Cambridge, MA: MIT Press/Bradford Books.

176

Maja Savic and Darinka Andelkovic

Hallan, Naomi 2001 The path to prepositions? A corpus-based study of the development of a lexicogrammatical category. In Frequency and the emergence of linguistic structure, Joan Bybee & Paul Hopper (eds.), 91-120, Amsterdam/Philadelphia: Benjamins. Halpern, Esther, Roberta Corrigan & Ora Aviezer 1983 In, on, and under: Examining the relationship between cognitive and language skills. International Journal of Behavioral Development 6(2): 153-166. Ivic, Milka 1957 Jedno poglavlje iz gramatike naseg modernog jezika - sistem mesnih padeza. Godisnjak Filozofskog fakulteta u Novom Sadu. Knjiga II: 145-158. Jocic, Mirjana 1981 Jedan aspekt usvajanja padeznog sistema na ranom uzrastu: padezi za obelezavanje prostornih odnosa. Godisnjak SDPLJ 4—5. Zagreb, 87-91. Johnston, R. Judith & Dan Slobin 1979 The development of locative expressions in English, Italian, SerboCroatian and Turkish. Journal of Child Language 6: 529-545. Klikovac, Duska 2006 Semantika predloga: Studija iz kognitivne lingvistke (Semantics of prepositions: A study in cognitive linguistics). Beograd: Filoloski fakultet. Kostic, Aleksandar 1996 Reprezentativnost jezickog korpus i mentalni leksikon (Representativity of language corpus and mental lexicon). LEP saopstenja 39: 132. Kostic, Dorde 1999 Frekvencijski recnik savremenog srpskog jezika {Frequency Dictionary of Contemporary Serbian language). Institute of Experimental Phonetics and Speech Pathology and Laboratory of Experimental Psychology, Belgrade. 2001 Kvantitavni opis srpskog jezika. Korpus srpskog jezika (Quantitative Description of the Serbian Language. Corpus of Serbian Language), Institute of Experimental Phonetics and Speech Pathology and Laboratory of Experimental Psychology, Belgrade. Kostic, Dorde & Spasenija Vladisavljevic 1995 Govor ijezik deteta u razvoju. Zavod za udzbenike i nastavna sredstva. Beograd. Kupisch, Tanja this vol. Testing the effects of frequency on the rate of learning: Determiner use in early French, German and Italian

Children 's usage of Serbian prepositions 2005

111

Frequency effects of in the evolution of determiners in monolingual and bilingual children. [In]determinismus in der Sprache. 27. Jahrestagung DFGS 2005. Köln. Landau, Barbara & Ray Jackendoff 1993 'What' and 'where' in spatial language and spatial cognition. Behavioral and Brain Sciences 16: 217-265. Lange, Joke de, Sergey Avrutin & Maria Teresa Guasti 2005 Cross-linguistic differences in Child and Adult Speech Optional Omissions: A comparison of Dutch and Italian. Proceedings of GALA 2005. Leikin, Mark 1998 Acquisition of locative prepositions in Russian. Journal of Psycholinguistic Research 27: 91-108. Lieven, Elena V. M., Heike Behrens, Jennifer Speares & Michael Tomasello 2003 Early syntactic creativity: a usage-based approach. Journal of Child Language 30: 333-370. Lieven, Elena V. M., Julian M. Pine & Gillian Baldwin 1997 Lexically based learning and early grammatical development. Journal of Child Language 24: 187-219. Majid, Asifa, Melissa Bowerman, Sotaro Kita, Daniel B. M. Haun. & Stephen C. Levinson 2004 Can language restructure cognition? The case for space. TRENDS in Cognitive Sciences 8 (3): 108-114. MacWhinney, Brian 1989 The CHILDES Project. Computational Tools for Analyzing Talk. Hillsdale, NJ: LEA. MacDonald, Maryellen C. 1999 Distributional information in language comprehension, production and acquisition: Three puzzles and a Moral. In The Emergence of Language, Brian MacWhinney (ed.), 177-196. Mahwah, NJ: Erlbaum. MacWhinney, Brian 1998a Models of the emergence of language. Annual Review of Psychology 49: 199-227. 1998b The emergence of language. Hillsdale, NJ: Erlbaum. MacWhinney, Brian & Catherine Snow 1985 The Child Language Data Exchange System. Journal of Child Language 12:271-296. Maye, Jessica, Janet F. Werker & Lou Ann Gerken 2002 Infant sensitivity to distributional information can affect phonetic discrimination. Cognition 82: 101-111. Meints, Kerstin, Kim Plunkettt, Paul Harris & Debbie Dimmock 2002 What is On' and 'under' for 15-18- and 24-months-olds? Typicality effects in early comprehension of spatial prepositions. British Journal of Developmental Psychology 20: 113-130.

178 Maja Savic and Darinka A näelkovic Munnich, Edward & Barbara Landau 2003 The effects of Spatial Language on Spatial Representation: Setting Some Boundaries. In Language in Mind: Advances in the Study of Language and Thought, Derdre Gentner & Susan Goldin-Meadow (eds.). Cambridge, MA: MIT Press/Bradford Books Opacic, Gordana 1964 A comparative study of general functions of certain Serbo-Croatian and English prepositions. Institute for experimental phonetics and speech pathology. Belgrade. Parezanovic, Sanda 1989 Uticaj predloga i pragmatike na razvoj razumevanja recenica sa predlosko-padeskom sintagmom u srpskohrvatskom jeziku. (Prepositions and pragmatics in comprehension of sentences with prepositional-noun phrase in Serbian). Unpublished MA diss., Faculty of Philosophy. University in Belgrade. Parisi, Domenico & Francesco Antinucci 1970 Lexical competence. In Advances in Psycholinguistics, Giovanni Flores DeArcais & Willem J. M. Levelt (eds.), 197-210. Amsterdam: Noth-Holland. Piaget, Jean 1959 Language and thought of the child. London: Routledge & Kegan Paul. Piaget, Jean & Bärbel Inhelder 1956 The child's conception of space. London: Routledge & Kegan Paul. Piaget, Jean, Bärbel Inhelder & Alina Szeminska 1970 The child's conception of geometry. London: Routledge & Kegan Paul. Pinker, Stephen 1984 Language learnability and language development. Cambridge: Harvard University Press. Rice, Sally 1999 Patterns of acquisition in the emerging mental lexicon: The case of to and for in English. Brain and Language 68: 268-276. Roeper, Tom 1999 Universal Bilingualism. In Bilingualism 2(3): 169-186. 2005 What frequency can do and what it can't. [Indeterminismus in der Sprache. 27. Jahrestagung DFGS 2005. Köln this vol. What frequency can do and what it can't. Rumelhart, David E. & James L. McClelland 1986 Parallel distributed processing: Explorations in the microstructure of cognition. Vol. 1: Foundations. Cambridge, MA: MIT Press. Saffran, Jenny, Richard Aslin & Elissa Newport 1996 Statistical learning by 8-month-old infants. Cognition 81: 149-169. Saffran, Jenny, Elizabeth K. Johnson, Richard Aslin & Elissa Newport 1999 Statistical learning of tone sequences by human infants and adults. Cognition 70: 27-52.

Children 's usage of Serbian prepositions

179

Savic, Svenka & Melanija Mikes 1974 Noun phrase expansion in child language. Journal of Child Language 1. Savic, Svenka & Vesna Polovina 1989 Razgovorni srpskohrvatski jezik (Conversational Serbo-Croatian Language), Institute of South Slavic Languages, Faculty of Philosophy, Novi Sad. Seidenberg, Mark & Maryellen C. MacDonald 1999 A probabilistic constraints approach to language acquisition and processing. In Connectionist models of human language processing: Progress and prospects, Morten H. Christiansen, Nick Chater & Mark S. Seidenberg (eds.). Special issue of Cognitive Science 23(4): 569-588. Seidenberg, Mark, Maryellen C. MacDonald & Jenny Saffran 2002 Does grammar start where statistics stop? Science 18: 553-554. 2003 Response to the Marcus and Berent's letter 'Are there limits to statistical learning?' Letters. Science 300. Sinha, Chris, Lisa A. Thorseng, Mariko Hayashi & Kim Plunkett 1999 Spatial language acquisition in Danish, English and Japanese. In Language and thought in development, Peter Breeder & Jaap Murre (eds.). Tübingen: Narr. Slobin, Dan 1985 Cross-linguistic evidence for the language-making capacity. In The cross-linguistic study of language acquisition. Volume 2: Theoretical issues, Dan Slobin (ed.), 1157-1249. Hillsdale: Erlbaum. Smith, Linda 2000 Learning how to learn words: An associative crane. In Becoming a word learner: A debate on lexical acquisitioni, Roberta Golinkoff, Kathryn Hirsh-Pasek, Lois Bloom, Linda Smith, Amanda Woodward, Nameera Akhtar, Michael Tomasello & George Hollich (eds.), 51-80. Oxford, New York: Oxford University Press. Stevanovic, Mihailo 1964 Savremeni srpskohrvatski jezik. Belgrad: Naucno delo. Tomasello, Michael 1987 Learning to use prepositions: A case study. Journal of Child Language 14: 79-98. 2003 Construing a Language: A Usage-Based Theory of Language Acquisition. Cambridge, MA: Harvard University Press. Weist, Richard M. 1991 Spatial and temporal location in child language. First language 11 (32): 253-267. Wilcox, Sthephen & David Palermo 1975 'In', On', and 'under' revisited. Cognition 3: 245-254.

180

Maja Savic and Darinka A ndelkovic

Vygotsky, Lev S. 1985 Thought and Language. Cambridge, MA: MIT Press. Yang, Charles D. 2004 Universal Grammar, statistics, or both? TRENDS in Cognitive Sciences 8(10): 451-456.

Characteristics of maternal input in relation to vocabulary development in children learning German* Christina Kauschke and Gisela Klann-Delius

Despite the divergent perspectives on child language acquisition, there is consensus that linguistic input is a necessary condition for language learning. In the present study, we describe the characteristics of child-directed speech (CDS) in German with respect to word frequency and word categories, and we analyze the relations between maternal input and child language in German. The study longitudinally investigates the lexical properties of CDS to 32 children at age 1;1, 1;3, 1;9 and 3;0 years. The results show that mothers and children increase the number of word types and tokens. Correlations between the word production of mothers and that of their children were found at ages 1;9 and 3;0 for types and at age 3;0 for tokens. With regard to the distribution of word categories, the results indicate that mothers offer all adult-like parts of speech from the earliest stages of lexical acquisition. The overall composition of the mothers' vocabulary remains quite constant, but also shows some age-related changes. With respect to the use of personal-social words, we found a decline in CDS and in children's word production. In addition, the number of noun types and verb tokens in CDS increased with the children's age. The results illustrate the data-providing and supportive functions of the input for specific aspects of lexical development.

1. Introduction There is no controversy that children need linguistic input in order to learn their mother tongue. However, with respect to grammatical development it is still controversial whether linguistic knowledge is genetically endowed and only triggered by the input data (Chomsky & Lasnik 1993), or learned from scratch, firstly in an item-wise and usage-based fashion, and later by abstraction and analogy (Tomasello 2000). Accordingly, the functions as-

182

Christina Kauschke and Gisela Klann-Delius

signed to the input differ: for nativists following the theory of Principles and Parameters, the input can only function as a trigger of the innate linguistic knowledge of the learner, because it is quantitatively as well as qualitatively underdetermined. The recent contributions of Yang (2004) and Roeper (this volume) refine this position. They suggest that statistical information contained in the input helps the child in parameter setting, and that innate linguistic knowledge is not triggered but expressed via probabilistic learning. In this perspective, linguistic input not only matters in regard to its quantity, but even more so in regard to its quality: "... there is suggestive experimental evidence that pits pure frequency against subtle diversity, showing that diversity helps more than frequency" (Roeper, this volume). Proponents of the usage-based approach place much more explanatory weight on the role of the input and stress its "highly repetitive item-based frames" (Tomasello 2006: 13). This corresponds to the finding that children first imitate concrete and specific constructions: "children reproduce what they hear and form their own abstractions later and only very gradually" (Tomasello 2000: 10). This usage-based approach (Tomasello 2003) highlights the data-providing function of the input, i.e. the regularity, specificity, frequency, and distributional properties of the data. As to lexical acquisition, one can observe a similar line of controversy: for constraint theorists, word learning is guided by biologically-shaped hypotheses about how to relate sound and meaning (Markman 1994; Woodward & Markman 1997; Waxman & Senghas 1992), whereas others focus on the experiential and interactive cues a child exploits to correctly pair sound and meaning (Nelson 1996; Baldwin & Tomasello 1998). In addition, the acquisition of a word's full meaning is seen to depend on rich exposure to the target word in variable contexts, as the experiments on fast mapping and slow mapping demonstrate (Carey 1978; Rice & Woodsmall 1988; Schafer 2005; Werker et al. 1998; Schafer & Plunkett 1998; HoustonPrice, Plunkett & Harris 2005; Woodward, Markman & Fitzsimmons 1994). With respect to the acquisition of word categories ("parts of speech"), a main component of lexical acquisition, a growing number of studies investigating the development of word categories in different languages document an increasing awareness of the specific role that the linguistic input plays. In this area of research there is a long-standing debate on the socalled 'noun bias' in lexical development (Gentner 1982). Proponents of the noun-bias hypothesis assume that nouns are universally privileged for early acquisition, because verbs are conceptually more complex than nouns. In contrast, it has been suggested that the developmental sequence and the relative proportions of word categories are strongly influenced by input

Characteristics of maternal input in relation to vocabulary development

183

characteristics, i.e. by features such as frequency and positional saliency. Empirical findings in favor of or against either one of these claims are heterogeneous and sometimes conflicting (see Kim, McGregor & Thompson 2000). In light of a wealth of cross-linguistic findings on the acquisition of word categories like nouns and verbs, Gentner & Boroditsky (2000) and Gentner (2006) maintain the proposal that there is an early conceptuallybased noun advantage across languages. At the same time, the authors acknowledge clear effects of the input language modulating the timing of the acquisition of word categories, their relative proportions during lexical development, and the degree of the noun bias. Recent cross-linguistic empirical research on the acquisition of word categories in relation to input characteristics has shown that input properties modify the noun bias and influence the course of the acquisition of nouns and verbs. At present, only a few studies investigate the influence of input properties on the acquisition of word categories other than nouns and verbs (e.g., see Akoyunoglou Blackwell 2005 on the acquisition of adjectives). Apart from language-specific influences, various studies of lexical acquisition show that lexical learning seems to be widely influenced by social and environmental factors; the studies of Hart & Risley (1995) and Hoff (2002), for instance, have documented that the socioeconomic status of parents, and the quantity as well as the quality of the linguistic input they provide correlate with the rate and level of children's lexical acquisition. Pan et al. (2005) found that growth patterns in children's vocabulary production are correlated with the diversity of maternal lexical input, and with maternal language and literacy skills. Although there is evidence that lexical acquisition is influenced by environmental factors, which show up in the frequency and the quality of input characteristics, one has to concur: "pointing to the environment, however, does not explain how the environment exerts its influence" (Hoff 2003: 1368). With respect to the question addressed in this volume, one could rephrase Hoff s statement as follows: "pointing to the frequency of the data provided in the input does not explain how frequency exerts its influence". More detailed research into which aspects of the environment are crucial for which developmental areas has yet to be carried out. This should include the investigation of frequency effects. Hoff (2003) favors the "principle of environmental specificity", according to which specific aspects of the environment influence specific aspects of development. Environmental factors influencing language acquisition may be social-pragmatic aspects or data-providing features. Social-pragmatic factors include joint attention, maternal attentiveness, responsiveness and contingency, language use in

184

Christina Kauschke and Gisela Klann-Delius

routines, positive affect, and attachment. Children benefit from these social features particularly in the early stages of language acquisition. These features are also relevant for the development of pragmatic and conversational skills. But, according to the results of Hoff & Naigles (2002), social support has little influence on children's lexical development; in contrast, dataproviding factors account for the variance in children's productive vocabularies. These aspects comprise the total amount of words, lexical richness and diversity, MLU, and the diversity of syntactic frames in the input. For lexical development, it seems to be important how much data is available and how informative the data is. The issue of how the frequency and diversity of the input matters is addressed in the present study with regard to words and word categories in German CDS. In addition, we investigate the open question of whether the data-providing and the social-supportive functions of the input are both important in the course of lexical acquisition, particularly with regard to the acquisition of word categories. We hypothesize that maternal input to young children shows characteristics of data provision as well as support in the form of an attunement to the child's mastery of word categories. With respect to the data-providing function, we assume that an input displaying diversity as well as repetition (mere frequency) should be helpful for language learning, since varied repetitiveness supports "one of the central tendencies" of the child's mind, i.e. "... the tendency to order the world by seeking invariants. A format in which each successive variation is both familiar (the part that is repeated) and novel (the part that is new) is ideally suited to teach infants to identify [...] invariants" (Stern 1985: 74). These assumptions are addressed in the present study. The study presented here focuses on the characteristics of maternal input to children learning German and its relation to characteristics of the child's developing lexicon. In particular, we focus on the frequency and distribution of word categories in German CDS. German children's use of word categories has been investigated in a longitudinal study of the same sample (Kauschke and Hofmeister 2002), which illustrated the dynamics involved in the development of word categories. In the earliest stages of lexical development, children used mostly relational words (like up or gone), personal-social words (like hello or yes) and some onomatopoeic terms (like moo). The categories which were initially dominant underwent a substantial decline; they were gradually complemented first with nouns, then with verbs, function words and other words. Thus, different word categories followed characteristic developmental curves during particular periods in the children's development.

Characteristics of maternal input in relation to vocabulary development

185

In the present longitudinal study, we analyze overall word frequency and the quantitative distribution of word categories in spontaneous speech that mothers addressed to their children, and we seek possible correlates with the distribution of word categories in the children's verbal productions. We address the following questions: Which characteristics does maternal input show with respect to the frequency and diversity of the vocabulary employed, and to the use of word categories? In particular, does maternal input show characteristics of fine-tuning to the child's level of lexical development and/or of mere data provision? Are there any relations between maternal input and child language? Hereby, we aim to identify which aspects of the linguistic input that might contribute to lexical acquisition can be observed during the early stages of lexical development. In addition to this explanatory aim, we pursue a descriptive aim: With our study on CDS in German, we hope to broaden the knowledge about input characteristics in languages other than English.

2. Method 32 mother-child dyads were studied during the second and third years of the children's lives.1 The study group of children comprised 16 girls and 16 boys. Information about the subjects' socioeconomic status was gained through a questionnaire assessing the education and professional occupations of the parents. Three-quarters of the dyads came from middle-class or upper middle-class families; one quarter came from lower middle-class or working-class families. Most children were first-borns. All children had experienced normal birth and were developing normally during their first year of life. The longitudinal study presented here comprises four recording samples, three of which were taken during the second year, at age 1;1 (13 months), 1;3 (15 months), and 1;9 (21 months). The final recording was taken at the age of 3;0 (36 months). During each 30-minute recording session, the mother interacted with her child as she would normally do in a room furnished with appropriate toys. The last ten minutes of these videotaped sessions were transcribed and analyzed. All verbal utterances, as well as vocal expressions (vocalizations, and babbling), other sounds (e.g. crying, and sniffing), and gestures of the mother and child were transcribed (see Klann-Delius & Hofmeister 1997 for details on this method). These transcripts formed the basis for the subsequent analyses.

186

Christina Kauschke and Gisela Klann-Delius

The data set is partially incomplete either because a mother-child dyad missed a session, or because of missing data due to technical problems. Data are lacking for one dyad at age 1 ;3 (boy), two dyads at age 1 ;9 (girl, boy) and one dyad at age 3;0 (girl). 2.1.

Procedure

In the present study, we focus on the mothers' use of words. The children's data were previously analyzed by Kauschke & Hofmeister (2002). Given the ambiguities in early word use, the determination of word status and the assignment of words to word classes were more difficult in the case of the children than in that of the mothers. Kauschke & Hofmeister (2002) describe in detail the procedure they used for the analyses of the child data. The procedure used for the analyses of the mothers' speech is described below. 2.1.1. Counting of word types and word tokens The first step was to determine any incidence of word production in the mothers' utterances. Paralinguistic expressions (sounds like coughing) and expressive utterances (laughing and sighs) were excluded. Onomatopoeic forms, conventional expressive interjections (e.g. oh), and short utterances (discourse particles, and assertions like yes) were counted as words. Every word determined in this fashion was then noted down in citation form (not in the form of actual realization); morphological markers were not taken into account. Every word type and every word token was stored in a database which had been developed especially for this purpose. Within this database, the words were sorted into lexical categories (see below). The complete database consisted of 2,128 different words that occurred in the sample of all the 32 mothers' CDS. In sum, a frequency of 16,036 types and 47,849 tokens was calculated. The number of total words (tokens) was calculated as a measure of word frequency; the number of different words (types) served as an indicator of lexical diversity. 2.1.2. Coding of word categories The next step was the assignment of words to lexical categories. While the classification of words produced by children requires some adaptations to

Characteristics of maternal input in relation to vocabulary development

187

the unstable and ambiguous nature of early words, the categories for adult word use may rely on the well-known categories of open and closed class vocabulary (see Evans 2000; Knobloch & Schaeder 2000). The classification system used in this study consists of nine main categories: nouns, verbs, adjectives, adverbs, pronouns, function words, particles, personalsocial words, and other words. Some of the main categories were divided into separate subcategories (see Table 1). Table 1. Classification of word categories in CDS Main category

j Main category includes

: Subcategories

Nouns

i proper names, persons, other ani- | proper names ' mates, inanimate objects, abstract \ count nouns \ nouns, internal-state nouns

Verbs

! actions and activities, processes, intransitive verbs i events, movements, states, mental transitive verbs j verbs, internal-state verbs ; ditransitive verbs I modal verbs

Adjectives

I adjectives denoting attributes or qualities, modifying elements, internal-state adjectives

Adverbs

heterogeneous class of uninflected words that modify a range of non-nominal classes

Personal-social words

interactive and expressive words

Pronouns

personal pronouns, demonstrative pronouns, possessive pronouns

interactive terms: I assertions and negations (yes/no), greetings, discourse signals, attention-getting devices expressive terms: personal, expressive utterances, interjections onomatopoeic terms

Function words j prepositions, conjunctions, articles, question words, auxiliaries, copulas Particles Other words

manner particles, intensifiers, particles that denote the speaker's attitude j ambiguous, non-classifiable words

188

Christina Kauschke and Gisela Klann-Delius

Semantic and distributional criteria helped to determine the word-class membership of every word. Unlike traditional classifications of parts of speech, we decided to treat personal-social words as an independent category. This category contains words which are supposed to be frequent in CDS. The category includes interactive words (attention-getting devices like look, short forms like yes and no, greetings like hello), and discourse markers (like hm), expressive utterances (expressive interjections like oh), and onomatopoeia. Words like these are sometimes excluded from classification systems, or they are included in other categories, e.g. particles or interjections. In order to capture the characteristics of the personal-social words, we counted interactive and expressive words, on the one hand, and onomatopoeia, on the other hand, as distinct categories in the trend analyses. The classification system used here is intended as a means of structuring and organizing the data in order to code words in a manner which allows comparability across the mothers' and the children's word production. 2.1.3. Statistical analysis The statistical analysis first involved the computation of descriptive statistics for the overall amount of types and tokens calculated for all participants who were present in the sessions. In order to measure changes in the mothers' vocabulary size over time, one-factorial analyses of variance (ANOVA) with repeated measures were conducted for types and tokens. Estimates of linear, quadratic, and cubic trends were derived from a polynomial decomposition of the data measured at the four sampling points. A second trend analysis was based on the logarithms of the data. Since the logarithm of an exponential growth is a straight line, a significant linear trend was interpreted as an indicator of exponential change over time. Descriptive statistics of lexical categories were derived from the mean number of words in each word category at each sampling point. For the purposes of further statistical analysis of the lexical categories, the data were converted into relative frequencies by dividing the number of types and tokens in every category by the respective totals (the sum of all types or tokens for every mother). As a result of the conversion to proportions, the data did not fit the assumptions of parametric methods. Accordingly, analyses of the course of development were conducted by applying non-parametric methods: firstly, the Friedman ranking test was used in order to measure significant differences in the mean proportions of the lexical categories at the four sampling points; secondly, based on the mean ranks computed by the Fried-

Characteristics of maternal input in relation to vocabulary development

189

man test (see Friedman 1937), a trend analysis (see Page 1963) was carried out in order to investigate whether the composition of the mothers' vocabularies followed particular patterns during the period of observation. In a final analysis, the mothers' word production was related to their children's vocabulary growth by computing Pearson's correlations between the amount of types and tokens produced by the mothers and by the children at each sampling point. For the analyses of variance and the Friedman tests, we excluded four dyads which had missed one of the sessions. Thus, these results are based on data from 28 mother-child dyads who participated in all four sessions. Missing data were not replaced by averages, because this procedure may seriously affect or even distort the results. 3.

Results

The number of words produced by the mothers increases with their children's age. Table 2 shows the mean numbers of types and tokens for all mothers who participated in the sessions (n=32 at 1 ;1, n=31 at 1;3, n=30 at l;9,andn=31 at3;0). Table 2. Descriptive values for types and tokens

Age

1;1 (13m.) n=32

l;3(15m.) n=31

1;9(21 m.) n=30

3;0(36m.)

n=31

106.22 14-190 43.86

113.13 49-192 36.43

140.87 52-208 37.22

158.19 83-249 40.51

309.84 19-713 163.40

321.23 76-674 140.43

458.90 126-776 163.20

458.35 186-825 168.56

Number of types Mean types Range types S.D. types Number of tokens Mean tokens Range tokens S.D. tokens

3.1. Growth pattern of word types The growth rate of the number of types produced increases from an average of 106.2 to 158.2. This growing amount of different words produced by the

190

Christina Kauschke and Gisela Klann-Delius

28 mothers who participated in all four sessions constitutes a significant increase as a factor of age (F(3,81) = 25.536, p < .001). The variable 'age' is responsible for the variance in types to a degree of 49%. Figure 1 illustrates the growth patterns of types. An analysis of the first three sampling points results in a linear (F(l,27) = 22.017, p < .001) and a quadratic (F(l,27) = 4.894, p < .05) trend. An analysis of the logarithms of the type values during this period suggests exponential growth of word types (linear trend: F(l,27) = 15.098, p < .01; no significant quadratic trend). This observation does not continue into the children's third year of life. A trend analysis of all sampling points shows a significant linear (F(l,27) = 68.138, p < .001) but no significant quadratic or cubic growth. The mothers' pattern of lexical diversity - an exponential increase in the second year followed by a further linear growth - is comparable to the development of their children's word-production skills (cf. Kauschke & Hofmeister 2002). 3.2. Growth pattern of word tokens As could be expected from the analysis of types, the overall amount of words produced also increases during the period of observation (see Figure 1). Table 2 shows how mean values increase from 309.84 at the beginning to 458.35 words at the end of data collection. Again the factor 'age' is significant (F(3,81) = 17.999, p < .001); age explains 40% of the token variance. An analysis of the data recorded during the children's second year of life shows that the overall amount of words produced by their mothers grows in an exponential fashion. From age 1;1 to 1;9 a linear (F(l,27) = 32.191, p < .001) and a quadratic (F(l,27) = 9.344, p < .01) trend can be observed. An analysis of the logarithms of the token values for this period shows a linear (F(l,27) = 18.507, p < .001) but no quadratic trend. In contrast to the types, the growth pattern of the tokens also contains a cubic component. A trend analysis encompassing all sampling points results in significant linear (F(l,27) = 39.200, p < .001) and cubic (F(l,27) - 10.941, p < .01) trends. After a moderate increase between the ages of 1;1 and 1;3, the amount of total words produced reaches a peak at age 1;9. Between the ages of 1;9 and 3;0, the overall production of words seems to reach a plateau.

Characteristics of maternal input in relation to vocabulary development

191

500 450 400 350 υ .ο

c a u

ε

300 250 200 150 100 • Types

50 -

• Tokens

Ο 13 months

15 months

21 months

36 months

Figure 1. Growth patterns of types and tokens

3.2.1. Correlations between maternal input and children's vocabulary production The computation of Pearson correlations between the number of words produced by the mothers and children (all dyads who participated at each sampling point) reveals different patterns for types and tokens. The amount of types uttered by the children correlates with type frequency in the input at age 1;9 (r = .367, ρ < .05) and at 3;0 (r - .491, ρ < .01). In contrast, no significant correlations were found with respect to the amount of tokens produced by the mothers and children before age 3;0. At age 3;0, the amount of tokens uttered by the mothers and children correlates significantly (r=.491, ρ *bezat (begut) run-3SG —> *3P (target)

Agreement

Number

Tense

Gender

lib. Two stems celuju —> *celut' (celovat') kiss-1S —» *INF (iNF-target)

Influence of an infl. class, i.e. j-addition igraf -igraj'u (1st class) celovat' - *celovaju

Influence of a .partner stem' celuju —> *celut' kiss-lS:INF

Number-gender

Notes I will consequently use the term optional infinitives below, since infinitives in Russian have the inflectional suffix -/', \.e.pisa-t' 'to write', where pis- is a root and -a is a thematic vowel. Cf. however, Farwell (1973) who questions the effect of a 'restricted' parental language on the child's acquisition of this language. I stress at this point that since the goal of this study is not to account for the existence of OI in Russian but to explain the durative use of infinitives along with the productive use of correct finite verb forms, I will not review in detail the large amount of literature on this topic. Rather, I will only point to the fact that nowadays OIs were documented for Danish, Dutch, English, French, German, Greek, Hebrew, Icelandic, Norwegian, Russian and Swedish. On the contrary, languages such as Catalan and Spanish, Inuktitut (typically developing children do not exhibit OIs in their speech production, however the production of OIs is documented for one child with SLI, see Crago & Allen 2001), Italian, Japanese, Polish, Portuguese and Tamil do not exhibit an OI stage at all or show

230

Natalia Gagarina

a very low (insignificant) percentage of subject-verb agreement errors (for the overview, see Gagarina 2002; Legate & Yang in press). Modern Greek does not have an infinitive construction, but a construction resembling OIs (an -z form which corresponds to participle) was found in early child Greek by Varlokosta, Vainikka & Rohrbacher (1998), see also Hyams (2002). 4. Only the relevant children's utterances are glossed. 5. On the Null Modal Hypothesis, see Hoekstra & Hyams 1998. 6. For the definition of salience, see Koepcke (1993) "Salienz ist die Bestimmung des Ausmaßes, mit dem eine morphologische Markierung vom Hörer identifizierbar ist, also ihre akustische Prominenz."

References Babby, Leonard H. 2000 Infinitival Existential Sentences in Russian. Paper presented at Formal Approaches to Slavic Linguistics 8, 1999, University of Pennsylvania, Philadelphia. Bates, Elizabeth & Brian MacWhinney 1978 Sentential devices for conveying givenness and newness: A crosscultural developmental study. Journal of Verbal Learning and Verbal Behavior 17: 539-558. Bittner, Dagmar, Wolfgang U. Dressler & Marianne Kilani-Schoch (eds.) 2003 Development of Verb Inflection in First Language Acquisition. A Cross-Linguistic Perspective. Berlin: Mouton de Gruyter. Brun, Dina, Sergej Avrutin & Marina Babyonyshev 1999 Aspect and its Temporal Interpretation during the optional Infinitive Stage in Russian. Paper presented at BUCLD 23: Proceedings of the 23rd annual y Conference on Language Development, Boston University. Chomsky, Noam 1986 Knowledge of language: Its Nature, Origin, and Use. New York: Praeger. Crago, B. Martha & Shanley E. M. Allen 2001 Early Finiteness in Inuktitut: The Role of Language Structure and Input. Language Acquisition 9: 59-111. Demuth, Katherine 1990 Subject, Topic and Sesotho Passive. Journal of Child Language 17: 67-84. Farwell, C. 1973 The language spoken to children. Papers and reports in child language development, 5: Stanford University Press.

What happens when adults often use infinitives?

231

Forner, Monika 1979 The mother as LAD: interaction between order and frequency of parental input and child production. In Studies in first and second language acquisition, F. R. Eckman & A. J. Hastings (eds.), 17^4. Rowley, MA: Newbury House Publishers. Fräser, Colin & Naomi Roberts 1975 Mothers' speech to children of four different ages. Journal of Psycholinguistic Research 4: 9-16. Gagarina, Natalia 2002 Thoughts on optional infinitives (in Russian). In Linguistics by Heart, Webfest for Horst-Dieter Gasde. Daniel Hole, Paul Law & Niina Zhang (eds.), 1-23. Berlin: ZAS. www.zas.gwz-berlin.de/mitarb/ homepage/ webfest. 2003 The early verb development and demarcation of stages in three Russian-speaking children. In Development of Verb Inflection in First Language Acquisition. A Cross-Linguistic Perspective, Dagmar Bittner, Wolfgang U. Dressler, Marianne Kilani-Schoch (eds.), 131169. Berlin/New York: Mouton de Gruyter. Gagarina, Natalia, Maria Voeikova & Sergej Gruzincev 2003 New version of morphological coding for the speech production of Russian children (in the framework of CHILDES). In Investigations into Formal Slavic Linguistics, Peter Kosta, Joanna Blaszczak, Jens Frasek, Ljudmila Geist & Marzena Zygis (eds.), 243-258. Frankfurt am Main: Peter Lang. Gagarina, Natalia & Insa Giilzow (eds.) 2006 Acquisition of Verbs and their Grammar: The Effect of Particular Languages. Series: Studies in Theoretical Psycholinguistics,Vol. 33: 229-259. Gierut, Judith A., Michele L. Morrisette & Annette Hust Champion 1999 Lexical constrains in phonological acquisition. Journal of Child Language 26:261-294. Gil, David 2006 The acquisition of voice morphology in Jakarta Indonesian. In The Acquisition of Verbs and their Grammar: The Effect of Particular Languages, Gagarina, Natalia, Insa Gülzow (eds.), 201-227. Series: Studies in Theoretical Psycholinguistics, Vol. 33: 201-227. Ginneken, Jac van 1917 De Roman van een Kleuter. Hertogenbosch/Antwerpen: Malmberg. Gvozdev, Aleksandr N. 1949 Formirovanie u rebenka grammaticheskogo stroja russkogo jazyka [The formation of grammar of Russian in child language]. Moskau: APN RSFSR.

232

Natalia Gaganna

Hirsh-Pasek, Kathy & Roberta Golinkoff 1993 Skeletal supports for grammatical learning: What the infant brings to the language learning task. In Advances in infancy research, Vol. 8 C. K. Rovee-Collier & L. P. Lipsitt (eds.), 299-338). Norwood, NJ: Ablex. Hoekstra, Teun & Nina Hyams 1998 Aspects of Root Infinitives. Lingua 106: 81-112. Hyams, Nina 2002 Clausal Structure in Child Greek. The Linguistic Review 19: 225269. Kiebzak-Mandera, Dorota 2000 Formation of the verb system in Russian children. Psychology of Language and Communication 4: 27—46. Koepke, Klaus-Michael 1993 Schemata bei der Pluralbildung im Deutschen. Versuch einer kognitiven Morphologie. Studien zur deutschen Grammatik 47. Tübingen: Narr. Legate, Julie Anna & Charles Yang in press Morphosyntactic Learning and the Development of Tense. Language Acquisition. MacWhinney, Brian 2000 The CHILDES Project: Tools for Analyzing Talk. Mahwah, NJ: Erlbaum. Mattausch, Jason & Insa Gülzow this vol. A note on acquisition in frequency-based accounts of binding phenomena. Naigles, Letitia & Erika Hoff-Ginsberg 1998 Why are some verbs learned before other verbs? Effects of input frequency and structure on children's early verb use. Journal of Child Language 25: 95-120. Nelson, Katherine 1973 Structure and strategy in learning to talk. Monographs of the Society for Research in Child Development 38(1/2): 1-135. Rhee, Jaemin & Kenneth Wexler 1995 Optional infinitives in Hebrew. In Papers on Language Acquisition and Processing, Carson T. Schuetze, Jennifer B. Ganger & Kevin Broihier (eds.), 383^402. Cambridge: MIT Working Papers in Linguistics. Rizzi, Luis 1994 Some notes on linguistic theory and language development: the case of root infinitives. Language Acquisition 3: 371-393. Roeper, Thomas 1999 Universal bilingualism. Bilingualism 2: 169-186.

What happens when adults often use infinitives?

233

Roeper, Thomas this vol. What frequency can do and what it can't. Savic, Svenke 1975 Aspects of adult-child communication: The problem of question acquisition. Journal of Child Language 2: 251-260. Slobin, Dan Isaack 1997 From "Thought and Language" to "Thinking for speaking". In Rethinking Linguistic Relativity, John J. Gumperz & Steven C. Levinson (eds.), 70-96. Cambridge: Cambridge University Press. Snow, Catherine E. 1972 Mothers' speech to children learning language. Child Development 43: 549-65. Snyder, Willyam & Eva Bar-Shalom 1998 Word Order, Finiteness, and Negation in Early Child Russian. Paper presented at BUCLD 22: Proceedings of the 22nd annual Conference on Language Development, Boston University. Storkel, Holly L. 2004 Do children acquire dense neighborhoods? An investigation of similarity neighborhoods in lexical acquisition. Applied Psycholinguistics 25:201-221. Thornton, Rosalind, Stephen Grain & Graciela Tesan this vol. Principles, Parameters and Probabilities. Tomasello, Michael 2003 Constructing a language: A usage-based theory of language acquisition. Cambridge, MA: Harvard University Press. Tomasello, Michael & William E. Merriman (eds.) 1995 Beyond Names for Things: Young Children's Acquisition of Verbs. Hillsdale, NJ: Erlbaum. Varlokosta, Sryridoula, Anne Vainikka & Bernhard Rohrbacher 1998 Functional Projections, markedness, and "root infinitives" in early child Greek. The Linguistic Review 15: 187-207. Voeikova, Maria D. 2000 Russian Existential Sentences: A Functional Approach. (LINCOM Studies in Slavic Linguistics). München: LINCOM Europa. Wexler, Kenneth 1994 Optional infinitives, verb movement and the economy of derivation in child grammar. In Verb movement, D. Lightfoot & N. Hornstein (eds.). Cambridge: Cambridge University Press. Wijnen, Frank, Masja Kempen & Steven Gillis 2001 Root infinitives in Dutch early child language: an effect or an input? Journal of Child Language 28: 629-660. Yang, Charles 2002 Knowledge and learning in natural language. New York: Oxford University Press.

Frequency mismatches between caregiver input and child language

Structural vs. frequency effects in LI acquisition of the passive and impersonal in Serbian Mil/a Djurkovic

Why is it that in languages that have sets of "passive-like" constructions, some of these seem to be easier for children to acquire than others? In this paper, I argue that this is so because some of these passive-like constructions are only functionally and not morphosyntactically passive, and that it is easier for children to acquire structurally simpler (non-passive) constructions. Comparing a set of Serbian and Russian data, I also question the potential of the frequency patterns of these constructions in the input to explain their LI acquisition and argue that it is not justified to treat constructions solely as "meaningful linguistic symbols" disregarding their underlying structural properties as significant both in theoretical and language acquisition terms.

1. Theoretical background A satisfactory theory of language acquisition should not only correctly explain the ability of a child to acquire knowledge of language on the basis of linguistic experience and the course of language acquisition, but also her ability to arrive at an adequate representation of linguistic knowledge (Culicover & Nowak 2003: 24, original emphasis; see also Roeper, this volume). This is where the relationship between the language acquisition theory and linguistic theory becomes apparent, as it is the linguistic theory that is supposed to account for the properties of this final linguistic representation (i.e. end-state grammar) and explain which of its properties appear to be universal cross-linguistically and why, and why some of the logically possible options are never found in human languages (Culicover & Nowak 2003: 23). Hence, a satisfactory language acquisition theory should be able to account for the child's route to this final grammatical system that is sufficiently general to be productive and sufficiently constrained not to represent an instantiation of a "wild grammar". There are two basic theoretical approaches to first language (LI) acquisition. ' One is the nativist approach, which presupposes an innate universal

238

Milja Djurkovic

grammar (UG) that accounts for the apparently rapid acquisition of certain deeply regular "core" grammatical principles, whereas vocabulary, various kinds of idioms, fixed expressions and other phenomena on the "periphery" of grammar are considered to be acquired by "normal", inductive learning processes (Tomasello 2003: 181). Thus, this approach postulates a kind of dual process model of LI acquisition. The second approach is generally referred to as usage-based. This approach does not assume that language represents an innate and domain-specific cognitive faculty. It posits a single process model of LI acquisition, i.e. a single set of cognitive and learning processes to account for the way both the most abstract and the most concrete linguistic phenomena are acquired. The classic representative of the former approach is the Principles and Parameters (P&P) approach, based on various versions of the Chomskyan syntactic theory and modified in accordance with the modifications of that syntactic framework. On the strongest versions of this approach, since the general principles are already innate, a limited amount of input (which could theoretically be a single exemplar of a linguistic structure) should be enough to "trigger" setting or resetting of parameters, which are taken to represent specific instantiations of the core UG principles in individual languages. However, apart from positing the dual process acquisition model, the above-mentioned approach has been criticized for not being able to account in sufficiently general and plausible terms even for the way the core of language is acquired, encountering a number of logical, theoretical and empirical problems.2 For instance, it has been criticized for its biological implausibility and its inability to explain how children map from an abstract parameter setting to its overt linguistic realization in a particular language, so as to be able to use this overt form as a "trigger" for setting the parameter (Tomasello 2003: 185-186). Importantly also, this approach faces problems with defining which phenomena in languages represent the actual parameters and how many of those there are in languages - to the extent that virtually any superficial difference between languages could be seen as corresponding to a parameter (Culicover & Nowak 2003: 31; Newmeyer 2004). Importantly, for the discussion that is to follow in this paper, this testifies to the difficulty this approach has in dealing with cross-linguistic variation (and even the variability within a single language) theoretically, and by extension in language acquisition. One of the most prominent approaches that arose as a reaction to the P&P account of the language acquisition process is the usage-based approach. As summarized by Tomasello (2003: 327), it proposes that language

Structural vs. frequency effects in LI acquisition

239

structure emerges from language use, both historically and ontogenetically. In ontogeny, it is hypothesized that a child hears and stores concrete utterances and then finds patterns in them. This process is gradual and uneven and depends crucially on the type and token frequency3 with which certain structures appear in the input. Together with general cognitive principles and mechanisms (e.g. categorization, intention-reading, pattern-finding), this enables LI acquisition. Linguistic constructions are taken to be the most fundamental units of language acquisition. They are treated theoretically (e.g. Croft 2001; Goldberg 1995, 2005) and proposed to be learned by children as "meaningful linguistic symbols'V'linguistic gestalts" (i.e. pairings of form and meaning/function) generalized by means of "inheritance hierarchy" from the more concrete to the more abstract ones (e.g. Tomasello 2003, 2004; Goldberg 2005). Thus, as opposed to the P&P's syntax-centric orientation, the usage-based approach assumes as a suitable framework for describing the system of adult grammar, a theoretical framework - (some version of) Construction Grammar - that denies the autonomy of syntax altogether. Since there is by now ample evidence that the above-mentioned cognitive mechanisms are employed in the LI acquisition of a big portion of linguistic phenomena, a reasonable question posed within the usage-based approach is why it is at all necessary to invoke an innate universal grammar to account for the acquisition of certain special aspects of linguistic structure. However, this question is asked from a theoretical standpoint which does not assume any "invisible" levels of syntactic structure to be existent in the representation of language. There is reason to believe that the theoretical framework assumed within the P&P acquisition theory (its latest version being the Minimalist Program) is overly syntax-centric and is founded on empirically insufficiently supported theoretical assumptions (interface uniformity, derivational formal technology, etc.).4 Once these assumptions are abandoned, it becomes possible to provide a different account of the adult grammatical system, which involves postulating much less abstract syntactic structure with a constraint-based formalism (cf. the "Simpler Syntax Hypothesis" of Culicover & Jackendoff 2005), integrating the core and the periphery of grammar into a representational continuum, and potentially providing a unified account of LI acquisition for both of these aspects of language. In other words, reducing the amount of abstractness of the linguistic system on independent grounds reduces the amount of what needs to be acquired by means of the specialized linguistic cognitive processes in the first place

240

Mil/a Djurkovic

(cf. Culicover & Jackendoff 2005: 12). Thus, although the proponents of the P&P approach might be right in suggesting that it would probably be impossible for children to reach the final level of generalization as quickly and effortlessly as they apparently do without the help of certain specifically linguistic, innate and genetically coded "guidelines" (e.g. Chomsky 1965, 1995, 2000, 2005; Wexler 1999; Borer 2004; Wunderlich 2004; etc.), these guidelines might turn out to be much less abstract, or at least less numerous then originally proposed - and hence more plausible as genetically coded elements. However, the justifiability of completely ousting the syntactic representation as one of the levels necessary for explaining how language works is still questionable, although perhaps desirable considering the learnability issues. But, in the words of Culicover & Jackendoff (2005: 43), "[o]ne of the most important goals of linguistic theory should be to establish the proper balance between purely structural principles, interface principles, and functional principles," while detecting the influence of each of these in language is an empirical issue. Hence, if evidence can be found for the necessity of positing an independent structural level of representation, the potential influence of this "hidden syntactic level" on the acquisition process would have to be accounted for in a language acquisition theory as well, perhaps even by proposing its innateness. What I will try to show in this paper is that there are linguistic phenomena that seem to call for positing this syntactic level theoretically, as their properties cannot be accounted for by assuming direct symbolic mapping from conceptual structure to overt linguistic form (as proposed, for instance, in Construction Grammar). If it can also be shown that the language acquisition process is somehow sensitive to the properties of this level, this could provide evidence for its psychological reality as well. Specifically, I will discuss the morphosyntactic properties of the passive and the impersonal constructions in Serbian, arguing that a specialized syntactic level of argument structure needs to be posited in order to provide a satisfactory account of their syntactic behavior. I also discuss their acquisition pattern and review a set of data related to the acquisition of similar constructions in Russian. I argue that the emerging frequency mismatches do not explain the Serbian and Russian data satisfactorily and they undermine the usual frequency related explanations of the acquisition of the similar phenomena in other languages. I will suggest that the structural complexity that arises in one of the constructions by means of disrupting the canonical setup of the above-mentioned level of representation might be

Structural vs. frequency effects in LI acquisition

241

one of the reasons for its peculiar acquisition pattern and might be added to the list of potential conspiring factors that influence LI acquisition. In Section 2,1 briefly present previous research related to LI acquisition of the passive. In Section 3, I discuss the morphosyntactic properties of Serbian "passive-like" constructions, suggesting a reanalysis that should enable a clearer comparison of the acquisition patterns of the individual constructions within the passive-like set. Section 4 sets out the predictions for the acquisition study. In Sections 4.1 and 4.2, I describe the methodology. The results are presented in Sections 4.3 and 4.4. Section 5 contains discussion and conclusions.

2. Why is the passive late in LI acquisition? The following facts about the acquisition pattern of the passive have been observed in a number of studies of children's spontaneous production and their comprehension in experimental setting. Although the passive is quite infrequent in the input, children start producing certain instances of passives spontaneously quite early in development (before the age of 3). However, they perform poorly in comprehension experiments with it - in most cases until the age of 4 or 5 (Babyonyshev & Brun 2004; Maratsos et al. 1985; Fox & Grozdinsky 1998; Sudhalter & Braine 1985; etc.). Apart from this discrepancy, it has been observed in comprehension experiments that children are better at understanding passives that clearly denote a resultant state. In addition, consistent with the previous observation, it has also been found that the stative/resultative uses of the passive precede the eventive5 ones even in spontaneous production (see Israel et al. 2000 for English; Abbot-Smith & Behrens 2006 for German). There are a number of questions that arise from this state of affairs. What causes the time gap between children's ability to produce and comprehend the passive? What does this gap reveal about the process of LI acquisition of the passive? To what extent does the small input frequency of the passive cause its late acquisition? The answers to these questions differ substantially as a function of the data patterns observed and researchers' general views of the nature of the LI acquisition process. Within the classic nativist approach, the most influential account of the cross-linguistically attested delay in LI acquisition of the passive is the following: the structural complexity of the passive (creation of Α-chains in the passive derivation through movement of the object to subject position)

242

Mil/a Djurkovic

presents a problem for children and until certain maturational conditions are met (maturation of Α-chains), the children will not be able to acquire the passive. They will only be able to interpret/produce it on an incorrect, adjectival analysis which does not involve movement and the presence of Α-chains (Borer & Wexler 1987; Babyonyshev & Brun 2004). This is supposed to account for the observed discrepancy in the production and comprehension of the resultative vs. eventive passives. On the usage-based approach, the obvious explanation for the late acquisition of the passive in general is that it is very infrequent in the input and hence children take longer to arrive at the adult-like representation (Tomasello et al. 1998; Brooks & Tomasello 1999). The prevalence of adjectival/resultative passives in child language and their earlier acquisition (at least in English) is due to the greater frequency of the so-called truncated passives - the passives without "by-phrases" expressing the original subject argument, which are thus homophonous with the adjectival ones. With respect to the former approach, I argue below that both its theoretical account of the passive (the derivational, movement account) and the maturational account of its acquisition are problematic, both theoretically and empirically. With respect to the latter approach, I will suggest that the frequency mismatches observed in the data call for an alternative explanation, especially in view of the fact that both constructions under examination are infrequent in the input, yet one of them has a clearly facilitated acquisition pattern.

3. Serbian "passive-like" constructions: Towards a unified characterization of passivization Numerous typological studies of the passive and the descriptive accounts in grammars have assumed its communicative function of foregrounding the "patient" and backgrounding the "agent" as one of the basic criteria for identifying this construction in languages, thus extending the notion of the passive to various constructions that in one way or another despecify the agent and inhibit its syntactic realization in a construction. This results in establishing families of "passive-like" constructions,6 with similar communicative use, but also with a number of divergent morphosyntactic properties. These properties (on which I elaborate in more detail below) have to be accounted for by positing certain additional constraints on passivization and by increasing the number of its defining properties. This in turn calls

Structural vs. frequency effects in LI acquisition

243

into question the justifiability of such an approach to identifying construction inventories. This approach has also been adopted within the previously mentioned framework of Construction Grammar (cf. for instance Croft's (2001) Voice Continuum and its application in the analysis of the Czech passive-like constructions in Fried (to appear)). The following constructions in Serbian have been subject to a similar misanalysis in both theoretical and descriptive literature (Stanojcic & Popovic 1992; Stevanovic 1986; I vie 1963; Milosevic 1980; Baric et al. 1995; Kucanda 1992, 1999; Vukojevic 1992; Belaj 2001; etc.): (1)

The passive (auxiliary verb biti 'be' + passive participle) Uliks je napisan od strane Dzojsa. Ulysses.NOM be.AUX.3SG write.PASS.PART.MASC by Joyce.GEN 'Ulysses was written by Joyce.' 7 (Djurkovic 2004: 74)

(2)

The active impersonal (transitive) / "reflexive passive" (personal active verb + originally reflexive marker se) Kuca se gradi (*od strane Marka) house.NOM.SG REFL build.PRES.3SG (*byMark) One is building the house.' / 'People are building the house.' / The house is being built.'

(3)

The active impersonal (intransitive) Opet se mirno spava. again REFL peacefully sleep.PRES.3SG One sleeps peacefully again.' / 'People sleep peacefully again.'

The misanalysis has to do with classifying the reflexive impersonal construction, formed of transitive verbs (example (2)), as the "reflexive passive" based on its similarity in function and meaning with the canonical participial passive (1) and disregarding its morphosyntactic similarity with the construction in (3), formed of intransitive verbs, that is unanimously recognized in the above-mentioned literature as an impersonal construction. Following Blevins (2003) and Kibort (2004), in the following sections I will show that the reflexive passive in Serbian only functionally resembles the passive, while its selectional and morphosyntactic properties testify that it belongs to a discrete diathesis type: the active impersonal.8 Hence, the set of passive-like constructions in Serbian can be shown to actually comprise two discrete sets of constructions - the passive and the active impersonal.

244

Mil/a Djurkovic

Each of these constructions has a set of its own defining properties, restrictions, morphosyntactic reflexes and frequency patterns and might be expected to exhibit different patterns of acquisition as well. In the next three subsections, I elaborate on the respective properties of the two constructions, suggesting the sort of architecture that a theoretical framework that accounts for them should have.

3.1. The passive: Defining properties In the Serbian participial passive, as in other prototypical passives crosslinguistically, the major syntactic reflexes of passivization are subject demotion and object promotion. Crucial support for the view that it is subject demotion that is primary in passivization comes from recognizing the existence of subjectless passives of intransitive verbs in a number of languages like German, Dutch, Turkish, etc. (cf. Comrie 1977; Keenan 1985). However, it has also been observed that there are certain systematic restrictions on passivization of intransitive verbs, i.e. only a subclass of these verbs can actually passivize. Thus, although all active intransitive predicates have subjects, and although what the passive does is demote the subject, some of the passives of intransitive verbs turn out to be ungrammatical. This suggests that a more precise characterization of the passive rule, or rather of the level of representation at which it applies, is needed. The basic insight that helps determine the level of representation at which the passive applies comes from the "Unaccusative Hypothesis" of Perlmutter (1978), which refers to the phenomenon of split intransitivity, i.e. to the fact that the class of intransitive verbs is not homogenous but consists of two classes that exhibit distinct syntactic behavior and semantic properties. As is generally assumed in the literature, the crucial difference between the unaccusative and unergative verbs is syntactically encoded in their initial argument structure, with unergatives specifying a subject and unaccusatives presumably an object argument.9 As this distinction is revealed also in their ability to passivize, and as passivization seems to be singling out only the unergative predicates as its potential input, it is plausible to assume that passivization is actually sensitive to the properties of predicate argument structure and applies at this level of representation. The following examples from Polish10 illustrate the pattern with an unergative (4a) and an unaccusative (4b) verb:

Structural vs. frequency effects in LI acquisition (4)

245

a. Tutaj bylo tanczone. here was.3SG.NEUT dance.PART.SG.NEUT 'There was dancing here.' / 'The dancing was done here.' (Kibort 2004: 73) b. *Wkuchni bylo zostawane (przez ludzi). in kitchen was.3 SG. NEUT remain. PART.SG.NEUT (by people) 'There has been remaining in the kitchen (by people).' (Kibort 2004: 73)

Furthermore, if unaccusatives are treated as lacking initial subjects, and if passivization is defined as demoting initial subjects, it straightforwardly follows that '[n]o impersonal passive clause in any language can be based on an unaccusative predicate' (Perlmutter & Postal 1984: 107). This observation is largely cross-linguistically confirmed." A passive rule that demotes the initial subject fails to apply to verbs without initial subjects. As the lack of the initial subject is what defines unaccusative verbs as a class, it is expected that they should not passivize (cf. Blevins 2003). Thus, passivization can be seen as a valence-changing operation which alters the verb argument structure by deleting the underlying subject argument from the argument structure of a verb, allowing it to be optionally expressed in an oblique agent phrase.

3.2. The active impersonal or the "reflexive passive" - a reanalysis Different morphological types (generally evolved from historically passive participle or reflexive forms) of impersonal constructions are attested in many languages, notably Slavic (Polish, Czech, Ukrainian, Lithuanian, Slovenian, Serbian, Croatian), Romance (Italian, European Portuguese, Spanish), Balto-Finnic (Estonian, Finnish), and Celtic (Breton, Welsh, Irish). They can be divided into intransitive and transitive, depending on the valence of the input verb. While the intransitive impersonals have been at least descriptively correctly classified as such (cf, e.g., Siewierska 1984; Stanojcic & Popovic 1992; Stevanovic 1986; Baric et al. 1995), the status of the transitive impersonals has not been clear. Focusing on Serbian, there are presumably several reasons why its transitive impersonal construction (cf. example (2) above) has been treated as a kind of passive. For one, this has to do with its passive-like function of backgrounding the agent. Related to that is the extensive use of the dynamic

246

Mil/a Djurkovic

"reflexive passive" in place of the participial passive, where the latter cannot be used in Serbian (namely, in the present tense to denote current activity). Another argument for such an analysis has been the apparent object promotion in this construction as evidenced in the nominative case marking on the remaining argument and the consequent agreement of the verb and this argument - which are characteristic of the Serbian participial passive (and of passives cross-linguistically). However, one of the most striking characteristics of the impersonal as opposed to the passive has been disregarded: its apparent insensitivity to the properties of argument structure. This property is evidenced in the possibility of forming the impersonal not only from unergative transitive and intransitive verbs (as with the passive), but also from unaccusative verbs (including copulas) and from already passivized predicates. The following examples from Serbian illustrate this pattern: (5)

unaccusative 'go' Leti se ide na more. in summer REFL go.PRES.3SG to the seaside 'People go to the seaside in the summer.'

(6)

copula 'be', unergative 'dance', 'sing' Kada se je veselo, igra

when REFL be.COPULA.PRES.3SG cheerful dance.PRES.3SG se i peva. REFL and sing.PRES.3SG 'When people are cheerful, they dance and sing.' (7)

transitive unaccusative 'have' Ima se svega dovoljno. have.PRES.3SG REFL everything.GEN.SG enough One has everything in abundance.'

(8)

passive/copula 'be' Bilo

se

hapseno

be.PAST.3 SG REFL arrested.PASSIVE.PART.IMPERF.NEUT.3SG / maltretirano od strane policije.

and molested.PASSlVE.PART.lMPERF.NEUT.3SG by police 'People were being arrested and molested by the police.'

Structural vs. frequency effects in LI acquisition

247

While the participial passive has been shown to apply at the initial argument structure and to crucially depend on its properties, the insensitivity of the impersonal to the properties of the verb's argument structure suggests that it applies at a different level of representation, i.e. at the level of surface syntactic functions. Hence, while the passive demotes the initial subject argument from the argument structure of the verb, the impersonal appears to just inhibit the syntactic realization of the surface subject, without affecting the initial argument structure of the verb. However, although impersonals do not exhibit sensitivity to the argument structure of the predicate, they are sensitive to the nature of the surface subject referent - they can only be formed of the predicates whose surface subjects can be construed as human. Weather verbs, non-animate verbs and verbs denoting, for instance, actions typical of animals cannot be impersonalized felicitously or they have to be interpreted metaphorically and as if referring to humans (as in (9) and (10) below).12 Passivization is generally not constrained in this way. (9)

# Laje se u dvoristu. bark.PRES.3SG REFL in the backyard One is barking in the backyard.'

(10) #Gnezda se grade visoko nadrvecu. nests.NOM.PL REFL build.PRES.3SG high-up in the trees One builds nests high up in the trees.' This is compatible with the fact that impersonal constructions usually have an active interpretation which is associated with an indefinite, human agent.13 This ties in with the fact that although the syntactic subjects are not expressed in impersonals, whether transitive or intransitive, they are still active in the construction, which is confirmed by the fact that they can, for instance, serve as antecedents for reflexive pronouns as in (11) (this is not the case with the demoted subject of the passive, see (12)). (11) Susedi sej voli u njegovoj, kuci neighbor.NOM.SG REFL love.PRES.3SG in his.POSS.3SG.MASC house a ogovara u svojojj. but gossip-about.PRES.3SG in his.POSS.REFL 'People love the neighbor in his house but gossip about him in their own.' (adapted from Vidakovic 2004: 291)

248

Milja Djurkovic

(12) Marko i

je

ogovaran

Marko.NOM.MASC.SG be.AUX.3SG gossip-about.PASS.PART.3SG.MASC

öd

strane

susedaj

u svojoji

a

ne u

from side.GEN neighbor.GEN in his.POSS.REFL and not in

njegovojj kuci. his.POSS.3SG.MASC house 'Marko was gossiped about in his own house and not in the neighbor's house by the neighbor.' The above-mentioned properties are also compatible with the fact that impersonals virtually uniformly14 resist the expression of an oblique agent phrase, unlike the prototypical passive (cf. example (2) above). Furthermore, the argument that the oblique agent phrases may not even be an integral part of the passive construction since they are optional (e.g., Keenan 1985: 261263) has actually been supported by the observation that there are many languages that present "passives" which do not permit agent phrases. However, while these languages virtually strictly forbid the expression of this phrase in their impersonal constructions, they optionally allow it in the passives (which they also present alongside the impersonals). Therefore, as pointed out by Blevins (2003: 506), while there may indeed be some passives where available strategies for expressing agent phrases must be disallowed by some sort of constraint, it is likely that the number of such cases will turn out to be much smaller than usually believed, once the impersonals are correctly classified. Unlike in passives, the remaining argument in transitive impersonals is retained as the structural object and does not undergo promotion. As noted by Blevins (2003: 475), if a language permits object cases to be expressed in constructions without an expressed nominative, the object may occur in the accusative (as in Polish or Ukrainian), or in the partitive, or it can alternate between the partitive and nominative (as in Estonian or Finnish). In other cases, the object may occur in the nominative (as in Serbian) or alternate between the nominative and accusative, as in Lithuanian, Czech or in some varieties of Slovene and Croatian. As for the case pattern of Serbian transitive impersonal constructions, the object status of the remaining syntactic argument is obscured by the fact that it surfaces in the nominative form, generally associated with subjects, and consequently triggers agreement. However, as pointed out by Blevins (2003: 506), the fact that these arguments trigger subject agreement is inconclusive since it has been shown (e.g. in Icelandic) that nominative com-

Structural vs. frequency effects in LI acquisition

249

plements can control agreement, i.e. that the agreement actually seems to obtain with the nominative argument, not with the subject of a clause. In addition, the fact that the nominative argument in the transitive impersonal cannot serve as an antecedent of the reflexive (unlike the nominative argument of the true passive) suggests that it is not a true subject. Also, although the impersonal constructions in the above-mentioned languages exhibit the alternating case patterns, their other morphosyntactic properties and sense (as presented previously) are exactly the same and are distinct from the properties of the canonical passives in these languages.15 Summing up, the properties of impersonalization include its insensitivity to the properties of the predicate argument structure, its resistance to the expression of the oblique agent phrase, necessarily indefinite human interpretation and retention of structural objects. All these properties identify impersonalization as a valence-preserving operation which inhibits the realization of the surface subjects, without affecting the argument structure of the input predicate. While the indefinite human interpretation (also characteristic of subjectless passives) and the resistance to the expression of the oblique agent phrase (some impersonals seem to allow it, although very infrequently) could be said to be strongly characteristic of impersonals, though perhaps not defining, its insensitivity to argument structure and its valence-preserving effect seem to be truly defining of this alternation. Thus, the basic (and defining) contrast between passive and impersonal constructions is that the former is a valence-changing operation, which reduces the number of arguments in the argument structure of the predicate by deleting its logical subject, while the latter is a valence-preserving operation, which preserves transitivity of the predicate and only inhibits the realization of the surface subject. On the whole, the properties of the two constructions shown above suggest that the passive is a more complex alternation which, unlike the impersonal, disrupts and changes the canonical, active, alignment of verbs' semantic arguments with syntactic functions. 3.3. A note on theoretical implications The previous discussion leads to a conclusion that the theoretical framework that could potentially successfully encode the properties of these constructions needs to distinguish between three independent levels of representation: thematic structure, argument structure and syntactic/grammatical functions, which are then mapped onto the surface morphological representation according to the phrase structure conventions of individual languages (e.g.

250

Mil/a Djurkovic

as in the constraint-based frameworks like LFG, HPSG, or (with some modifications) the "Simpler Syntax" framework of Culicover & Jackendoff 2005).16 It does not seem possible to account for the morphosyntactic differences between the two constructions only in terms of the symbolic mapping from semantics to overt representation (as in Construction Grammar, e.g. Croft 2001; Fried in press) as the difference between the selectional properties of the constructions is not only semantic or functional. Therefore, it seems that a specialized syntactic level of representation (the level of grammatical functions) needs to be posited (cf. Culicover & Jackendoff 2005).17 These two authors, however, propose direct mapping from thematic structure to grammatical functions. But the properties of the passive and the impersonal, and their different sensitivity to split intransitivity suggest that a level of argument structure (at which split intransitivity is encoded) also needs to be posited, in order to provide the mapping between the thematic structure and grammatical functions (as is the case, for instance, within the LFG framework, see e.g. Alsina (1996), Kibort (2004), etc. for discussion). Hence, the passive operates at the level of argument structure by demoting the first argument of the unergative verb, whereas the impersonal inhibits the overt syntactic realization of the subject function, without disrupting the mapping from the argument structure to grammatical functions.18 As discussed in Section 1, although it is desirable on the basis of learnability considerations to minimize recourse to the hidden and specifically syntactic levels of grammatical representation, certain theoretical considerations suggest that positing a version of such a level may be necessary if a satisfactory account of certain grammatical phenomena is to be provided. The acquisition-related question then is how is this hidden level learnable if it cannot be straightforwardly inferred from the input? I leave the speculation regarding this issue for the discussion section and turn to examining language acquisition data with a view of potentially detecting some reflexes of the existence of this level which could be considered to influence acquisition over and above frequency. As mentioned previously, this should then lend support to the psychological and not only theoretical reality of this level of representation. 4. The present study: Predictions Having established clear morphosyntactic distinctions between the passive and the impersonal constructions in Serbian, we are now in a better position

Structural vs. frequency effects in LI acquisition

251

to give a unified and clear prediction and explanation for their acquisition patterns and potentially establish what might be the locus of the problem that children have when acquiring the passive. Assuming that positing an abstract syntactic level of representation at which the two constructions crucially differ is justified theoretically, we could look for the evidence of this difference in acquisition as well, and propose the following prediction: (a) since, unlike the passive, the impersonal does not involve the disruption of the canonical mapping between argument structure and grammatical functions and essentially has the active mapping, one would expect a facilitated acquisition of the impersonal and would expect to see it patterning with the active transitive, as they involve the same, active mapping. However, assuming the usage-based approach, one would not be in a position to refer to the underlying structural properties of these constructions as these are not considered as theoretically relevant. So the prediction would be that (b) the development should proceed in keeping with the nature of the input (e.g. the more frequent construction should be acquired earlier), with perhaps the influence of the respective communicative functions of the two constructions. In what follows, I present part of the data from my own two acquisition studies (an experimental comprehension study and a corpus study) of the Serbian constructions and review an acquisition study of a pair of similar constructions in Russian (Babyonyshev & Brun 2004) (Section 4.5) and then compare which of the above-mentioned approaches explains the observed patterns better and how the above-mentioned predictions fare with respect to the data at hand. 4.1. Methodology: The comprehension study As has been demonstrated in a number acquisition studies (cf. Section 2), there is often a discrepancy between children's production in a naturalistic setting and their performance in experiments. Many researchers have pointed out that even if children are considerably productive with a certain construction, their interpretation of this construction may be quite different from that of the adults. Therefore, an experimental study was conducted where possible interpretations that children might attribute to the constructions could be controlled (at least to some extent), i.e. where children could be pressed to either interpret the items in the adult-like way, or not be able to interpret them at all if they did not have the adult-like representation of the items.

252

Milja Djurkovic

As already mentioned in Section 2, one possibility when it comes to children's production of the passive at an early age is that what they are producing are actually resultative adjectives and not verbal passives. Therefore, in this comprehension experiment I used examples of the passive that could only be interpreted as verbal (imperfective verbs denoting a past action) to prevent the possibility of children interpreting them as adjectives. The task was a comprehension picture-matching task. There were three conditions: passive, impersonal, and active with four test sentence-picture pairs in each of them. The same verbs (kupati 'bathe', sisati 'cut-hair', nositi 'carry', hraniti 'feed') and arguments (all human or human-like: Dusko Dugousko 'Buggs Bunny', Snezana 'Snow White\patuljci 'dwarfs', beba 'baby', Crvenkapa 'Little Red Riding Hood', mama 'mother') were used in each of the conditions. The arguments of the verbs were all human or human-like and thus semantically reversible, so that the children would not simply resort to semantics to help them determine their respective roles in the situation. The test sentences were of the kind: (13) a. Dusko Dugousko je kupan.19 'Buggs Bunny was bathed.'

passive

20

b. Dusko Dugousko se kupa. 'Buggs Bunny is being bathed.' c. Dusko Dugousko kupa patuljka^ 'Buggs Bunny is bathing the dwarf.'

impersonal active

The procedure consisted of showing the child two pictures, asking the question Na kojoj slid...? ('In which picture...?'), followed by one of the experimental sentences which was true for one of the pictures (the other picture always showed a reversed situation) and asking the child to point to the correct picture. The pictures were presented in pairs, varying the side (left/ right) where the correct picture would appear so that the child would not get used to pointing only in one direction and thus accidentally getting all the answers right/wrong. Ninety-six monolingual Serbian-speaking children took part in the experiment. They were divided in five age groups (number of subjects per group in brackets), the mean ages of the groups being: 2;08 (18), 3;03 (19), 4;02 (19), 5;01 (20), and 6;01 (20).

Structural vs. frequency effects in LI acquisition

253

4.2. Methodology: The corpus study

In order to examine the potential influence of the frequency patterns in the input on the acquisition of the constructions, I conducted an analysis of the Serbian Corpus of Early Child Language (Andelkovic et al. 2001). This corpus contains the recordings of eight children, four boys and four girls, recorded every two months for two hours from the age of 1 ;6 to 4;0. For the present study, the following age groups were selected for examination: 1;6 1;8, 2;2, 2;8, 3;2, 3;8, 4;0. During the recording sessions in these age groups the children produced around 42,867 utterances and the adults around 85,184 utterances. The count included all tokens of the participial passive forms used as complements of the verb be (with potentially both verbal and adjectival interpretation). The cases where the participial form was used attributively were not counted. As for the impersonal, the count included the tokens of impersonal constructions with the transitive verb, as these are the ones generally assumed to be a kind of passive (cf. Section 3).

4.3. Results: The comprehension study Figure 1 below shows the percentage of correct responses for the passive, impersonal and active for all children against age. 100 "A

H PAS S • IMP Π ACT

AGE Figure 1. Percentage of correct responses for the passive, impersonal and active

254

Mil/a Djurkovic

The one sample t-test of the levels of correct performance per age shows that for the active and the impersonal, the children perform either at 100% correct or significantly above chance already from the age of 2;08. For the passive, they are at chance until 4;02 and only after that start performing significantly above chance level. The mixed ANO VA results show a significant effect of age for all constructions (F (4, 91) = 11.77, p < .001). They also show a significant difference between the response patterns for different constructions (F (2 182) = 78.49, p < .001). There was no age and construction interaction, suggesting that the difference remains the same over age. In order to see how exactly the patterns for the constructions differed, a pair-wise analysis was conducted. This showed overall and for each age group individually that there was no significant difference between the response patterns for the impersonal and active, whereas there was a significant difference between the passive and active F (1, 91) = 106.57, p < .001 and between the passive and impersonal F (1, 91) = 89.17, p < .001. Thus, it appears that it is the active and the impersonal that pattern together in comprehension, whereas the passive exhibits a different pattern. The question is, what is causing these differences in children's comprehension of the three constructions? Is it something about their structural properties (which would be in keeping with the hypothesis that the structural properties of the constructions might in some cases have an effect on acquisition over and above other factors), or perhaps the way they are distributed in the input with respect to frequency? 4.4. Results: The corpus study The corpus study was intended to shed some light on the latter possibility. Table 1 and Table 2 (as well as Table 3) below suggest that although both children and adults produce more impersonals than passives, there might be some discrepancy between the frequency ratios of the constructions in child language vs. those in the input. Table 1. Distribution of the passive and the impersonal in child production Passives and transitive impersonals

Passives

Transitive impersonals

178(100%)

44(24.72%)

134(75.28%)

Structural vs. frequency effects in LI acquisition

255

Table 2. Distribution of the passive and the impersonal in the input Passives and transitive impersonals

Passives

Transitive impersonals

512(100%)

206(40.23%)

306(59.77%)

Table 3 shows that both the passive and the impersonal are quite infrequent in both adult and child language. A statistical analysis (mixed ANOVA) of the ratios in Table 3 confirms this impression. Namely, it turns out that although the impersonal is somewhat more frequent than the passive in the input, the difference between the frequency of the passive and the impersonal is not significant F (1,7) = 1.08, whereas in child production it is significant F (1, 7) = 8.10, p < .03. It is as if the children are not producing the amount of the participial passive that would be expected on the basis of the input pattern. Table 3. Frequency of the passive (verbal and adjectival) and transitive impersonal in child and adult production (*child/Agroup etc. = child and the group of adults (mainly family and the experimenters) that interacted with the child during the recording sessions) Children

Pass.

Imp.

Adults

Pass.

Imp.

Child 1

0.051% 3/5834

0.120% 7/5834

Agroup 1 *

0.276% 26/9429

0.276% 26/9429

Child 2

0.175% 10/5726

0.297% 17/5726

Agroup2

0.614% 42/6841

0.249% 17/6841

Child 3

0.088% 4/4539

0.419% 19/4539

Agroup 3

0.266% 26/9758

0.359% 35/9758

Child 4

0.038% 2/5248

0.343% 18/5248

Agroup 4

0.065% 10/15446

0.434% 67/15446

Child 5

0.024% 1/4147

0.096% 4/4147

Agroup 5

0.155% 22/14158

0.240% 34/14158

Child 6

0.148% 11/7428

0.269% 20/7428

Agroup 6

0.245% 26/10598

0.538% 57/10598

Child 7

0.102% 6/4888

0.123% 8/4888

Agroup 7

0.283% 29/10234

0.371% 38/10234

Child 8

0.158% 8/5057

0.811% 41/5057

Agroup 8

0.287% 25/8720

0.367% 32/8720

256

Milja Djurkovic

The observed pattern is unexpected on the frequency-based account. If LI acquisition of these constructions is crucially influenced by their frequency in the input and if their structural properties do not play a role, there should be no such contrast, i.e. the child production should not show a significant difference in frequency between these constructions. Hence, the discrepancies observed in the corpus data suggest that the frequency patterns may not be able to account for the data straightforwardly. 4.5. Russian: Perfective (participial) passive vs. imperfective (sja-) passive However, before reaching any conclusions about which approach best accounts for the Serbian data, I review a corpus study of Russian child and child-directed speech which deals with constructions grammatically and historically related to the Serbian constructions discussed here. Russian, like Serbian, has a set of passive constructions - the participial and the so-called "s/a-passive" (sja being the Russian counterpart of the reflexive pronoun se in Serbian, except that it is affixed to the verb in Russian). However, unlike in Serbian, both of these constructions are recognized in the literature to be real, morphosyntactic passives, and both involve the same kind of structural complexity. The division between them is aspectual. Namely, only the perfective verbs can be used with the participial passive while the imperfective verbs are used with the s/a-passive. Babyonyshev & Brun (2004) conducted an analysis of the spontaneous speech of eight monolingual Russian speaking children, aged 2;06 to 3;09, looking at the patterns of production of the passive constructions in Russian. The data obtained in this study show an interesting asymmetry. As can be seen in Table 4 below, the children produced significantly more perfective passives (91%) than imperfective ones (9%). Table 4. Distribution of the two types of passives in Russian child language Total number of passives

Perfective (participial passive)

Imperfective (sja-passive)

212(100%)

193(91%)

19(9%)

One explanation for this could be that a similar pattern is present in the input as well. However, the authors report the data from an analysis of informal child-directed speech of four adults that they conducted, showing that there is no preference for perfective verbs in the adult use of the passive con-

Structural vs. frequency effects in LI acquisition

257

structions. On the contrary, the perfective forms were used by the adults only 44.2% of the time, as opposed to 55.8% of the imperfective. They also report the data from an analysis of formal adult-directed speech, which gave similar results: 43.1% of perfective passives and 56.9% imperfective. Neither of the differences was found to be significant. Thus, the child production and the input patterns seem to be opposite - a fact not expected on the usage-based account. The explanation that the authors offer is that the perfective passives that children produce are actually misanalyzed resultative adjectives. As in the authors' view, the children exhibit an Α-chain deficit early on in development, they misanalyze the verbal for adjectival passive until maturationally ready to use the verbal passive in an adult-like way. The children have this possibility in Russian, as the resultative adjectives and passive participles are syntactic homophones (s-homophones).22 On the other hand, as the imperfective passive does not have an s-homophone in Russian,23 the children do not produce this construction until a certain point in development when maturationally ready to do so. As mentioned in Section 3.3 and Note 15, the derivational movementbased account of the passive is theoretically problematic in a number of ways and it seems justified to question the validity of positing such a representation of the passive as a valid foundation for providing accounts of its language acquisition pattern. In addition, the accounts of L l acquisition in terms of maturation have proven to be problematic in many ways as well, notably with respect to their biological (im)plausibility and numerous mismatches in terms of acquisitional order in monolingual and especially bilingual language development.24 These mismatches would be unexpected if the development of certain linguistic structures was genetically and maturationally preconditioned. However, although I argue that the theoretical and acquisition account of the passive proposed by Babyonyshev & Brun is flawed, the fact remains that an account invoking the decisive influence of the frequency patterns in the input cannot explain the patterns emerging from the Russian data and that an alternative account in terms of structural differences should be sought. 5. Discussion and conclusions To summarize, the following patterns emerge from the Serbian and Russian data. The comprehension study showed that the impersonal and the active pattern together from the earliest age tested. Already at that age 2;08 the

258 Milj a Djurko vie

children are above chance in their comprehension of these two constructions. The passive, on the other hand, shows a different pattern, the correct responses not reaching the level above chance until the age of 5;01. The corpus study of Serbian child and child-directed speech revealed that the impersonal is to some extent more frequent in the input than the passive. However, although the frequency ratio between the two constructions in child language goes in the same direction, the discrepancy between the levels of production of the two constructions is significantly larger than in adult language, in favor of the impersonal. The Russian corpus data show that children are significantly more productive with the perfective than with the imperfective passive, although both constructions are used in adult speech roughly to the same extent, with the imperfective passive being somewhat more frequent. The prediction of the usage-based approach that the development should proceed in keeping with the nature of the input (that is, that the more frequent construction should be acquired earlier) seems not to be supported by these patterns. Namely, as the impersonal is more frequent in adult language in Serbian, one might be tempted to resort to this fact as a straightforward explanation for the apparent facilitation of the impersonal both in production and comprehension in Serbian. However, as there is a larger discrepancy in child production of the passive and impersonal in favor of the impersonal than in adult production, the frequency-related explanation turns out not to be that straightforward. This is more so if one takes into consideration the patterns in the Russian data that are difficult to account for by resorting to the frequency patterns in the input as an explanation. The children's production seems to be the opposite of what they get in the input. Furthermore, as can be seen from Table 3, both the passive and the impersonal in Serbian are quite infrequent in the input on the whole. Neither of them reaches even 1% of the total number of the recorded utterances. Yet, in comprehension, already from the age of 2;08, the impersonal is equally well understood by the children as the personal active transitive construction, which certainly forms the bulk of the utterances children hear and is far more frequent than the impersonal. The question is, how is this possible in spite of such a small input frequency of the impersonal (albeit somewhat bigger than that of the passive) if a similarity between the active and the impersonal at the level of structural representation is not helping children here? Or is this less than 1 % just about enough for children to acquire the impersonal to the extent that the

Structural vs. frequency effects in LI acquisition

259

patterns emerging in comprehension are practically indistinguishable from those of the much more frequent active transitive? Here we encounter the problematic issue of the "critical mass" and the number and nature of the exemplars required to trigger generalizations. Although, obviously, a child needs to be exposed to language and starts learning item by item in order to form certain syntactic generalizations, it is not clear at this point what amount of exemplars a child needs to be exposed to in order to arrive at the appropriate generalization (Tomasello 2003: 166-167). What is the nature of the "critical mass" (Marchman & Bates 1994) in the case of different constructions, and can learning be triggered perhaps by a single exemplar at the right moment? (cf. similar learning outcomes in Childers & Tomasello (2001) with four times more exemplars than in Abbot-Smith et al. (submitted), cited in Tomasello 2003). Also, as shown in the Abbot-Smith & Behrens' (2006) study of LI acquisition of the passive and future in German, the input frequency of a particular construction in isolation may not have a deterministic influence on acquisition. In their view, input frequency should be examined in relation to a network of related constructions, as the influence of prior acquisition of related constructions might counterbalance the influence of input frequency. Furthermore, there is abundant evidence of frequency mismatches reported in the literature (cf, e.g., Bohnacker, this volume; Kupisch, this volume; Westergaard & Bentzen, this volume; Abbot-Smith & Behrens 2006; etc.) and here. This suggests, by now perhaps uncontroversially, that there are other influences that have to be taken into account (e.g. morphophonological structure, syntagmatic complexity, processing complexity, pragmatic properties), not only to explain why a linguistic structure is acquired in a particular way, but also, crucially, why it is frequent or infrequent in a language in the first place. Ultimately, it is acknowledged by, for instance, Tomasello (2003) and Diessel (2004) that input frequency and structural complexity interact in complex ways in the developmental process. Although their notion of "structural complexity" refers to syntagmatic complexity (constructions containing anything from a single element to multiple components) and the scale of schematicity/abstractness (Diessel 2004: 18), it is perhaps warranted to consider the representational complexity as well as a possible influence on LI acquisition process in some cases. All these empirical problems related to frequency bring into sharper relief the basic theoretical problem with using the notion of frequency to explain the actual process of language acquisition and the formation of linguistic representations. As argued by Roeper (this volume), while frequency

260

Milja Djurkovic

can have a narrow role in the choice of mental representations where no new information is added to either of the representations, it is not clear how it can explain the way these representations are constructed in the first place. He points out that as the process of learning consists of creating new representations, each exposure to a phenomenon should add new information to the previous representation until a final generalization is reached. Hence, even if it is the same phenomenon being repeated with certain frequency, since the object (i.e. representation) to which information is added is changed with each addition, the effect of each exposure can actually be considered to be unique. Hence, the learning process is not facilitated by a small or large frequency of a phenomenon, but by the new information that is added with each exposure. The abovementioned frequency mismatches testify to the fact that it is the nature and perhaps complexity of the information contained in a phenomenon being learned, rather than its frequency (which is derivative), which conditions the learning pattern. Therefore, while it is certainly valid to look at the way frequency affects language use and at how a child is exposed to language, it is dubious to what extent frequency actually explains the process of acquisition, apart from being a kind of "explanatory wild card" used when the nature of the phenomenon being learned or the nature of the learning process is insufficiently understood. Returning to the data patterns presented in this paper, the frequency mismatches observed further suggest that there might be more to the constructions under discussion judging by their acquisition pattern. Focusing on their structural properties (perhaps as a window into the kind of representational "information" that these constructions might contain), the first prediction in Section 4 was related to the issue of whether it is justified to posit an abstract syntactic level of representation to account for the difference between the impersonal and the passive theoretically, and whether the effects of this level of representation being affected in some way could be noticed in acquisition. The comparison between the Serbian and Russian data seems to be particularly relevant here. As both passive-like constructions in Russian are passives and involve the same disruption of argument structure-function mapping, the same acquisition pattern can be expected for both constructions. The fact that the perfective one is facilitated in production can be accounted for by proposing that what the children are using are the resultative participles (morphologically and semantically related to the passive ones, but lacking the argument structure-function mapping change - cf. Note 22). Hence, the actual problem that Russian children have with the passive is revealed in

Structural vs. frequency effects in LI acquisition

261

their difficulty with using the imperfect!ve passive, which does not have a semantically related s-homophone. On the other hand, although Serbian has the same inventory of constructions homophonous to the passive and impersonal, Serbian children do not show the same asymmetry in the usage of the passive and the impersonal. Although the Serbian impersonal (morphologically similar to the Russian imperfective passive) also does not have a proper semantic and s-homophone, Serbian children are able to use it and comprehend it from early on in development. While this may seem surprising, it is actually explained on the present account of the differences between the passive and the impersonal in terms of the nature of argument structure-function mapping. As, unlike the Russian imperfective passive, the Serbian impersonal construction does not involve any change in this mapping and has essentially the same representation as the active construction, the children are able to use it and comprehend it in an adult-like way from early on. On the other hand, as the passive does involve this change, this seems to present problems for children, and it takes them longer to form correct generalizations with respect to this construction. Furthermore, it seems plausible to suggest that even the early "passives" used by the Serbian children are actually misanalyzed resultative participles, as in the Russian case. This is supported by the comprehension pattern where children seem to be unable to comprehend the actual verbal passives reliably until the age of 5;01. Hence, the asymmetry in the Serbian data which goes in the opposite direction to Russian is actually predicted. In view of this, it seems problematic to dismiss the underlying syntactic structure as an additional factor potentially influencing late LI acquisition of the passive. More generally, the findings presented here support the view that it is not justified to treat constructions as just symbolic entities encoding certain meaning and communicative function without taking into consideration their underlying structural complexity as theoretically relevant even if this involves positing a "hidden" syntactic level of representation that poses additional learnability problems. This may get in the way both of correct cross-linguistic description of constructions and of understanding the patterns of their acquisition. As has been shown here, there seems to be evidence of this level being somehow represented in the brain, although it is far from clear at this point in which particular way. One way of looking at it (purely speculatively) could be along the lines of Culicover & Jackendoff (2005: 539), who suggest that this level of representation (the GF-tier in their terminology) '[mjight be a late evolutionary add-on to the syntax-

262

Mil/a Djurkovic

semantics interface, hardly inevitable but affording adaptive advantages in efficiency and reliability of communication'. Its existence - a kind of representational template perhaps (cf. Roeper, this volume) - might facilitate early acquisition of the constructions that involve its canonical representation, whereas the ones where this representation is somehow disrupted lag behind. On the whole, it seems that it is the structural complexity of the passive, after all, that has the crucial influence on its late acquisition. Constructions functionally similar to it but structurally simpler seem to be facilitated in acquisition. This finding supports the view that although the properties of the input in terms of frequencies cannot be disregarded, they alone could not help children to get to the appropriate level of generalization in acquisition if they did not interact with certain specifically linguistic mechanisms and structural properties of the constructions in a language.

Acknowledgements I would like to thank my supervisors, Teresa Parodi and Jim Blevins, for their guidance and advice. I am also grateful to Darinka Andelkovic for giving me access to the corpus data used in this study and for very helpful discussions. Big thanks are also due to the children and teachers of the nurseries 'Nase dete', Sabac, for taking part in the comprehension study and for making them such fun to do. Notes 1. See, e.g., Tomasello (2003), Culicover & Nowak (2003), etc. for more detailed discussions. 2. See, e.g., Culicover & Nowak 2003, Tomasello 2003 and Culicover & Jackendoff 2005 for more detailed overviews of the problems and criticisms. 3. The potential of these notions to explain the process of language acquisition will be discussed in the final section of the paper - i.e. to what extent the frequency effects in language use are involved in the construction and acquisition of the grammar which underlies language use (cf. Roeper, this volume). 4. See Culicover & Jackendoff (2005) for a detailed overview and a critical exposition of the development of argumentation and assumptions behind different versions of the Chomskyan syntactic theory.

Structural vs. frequency effects in LI acquisition

263

5. The resultative/stative passive denotes an ongoing state resulting from a previously occurring event as in, 'The window is broken'. Hence it is often referred to as the "adjectival passive". The eventive passive denotes a dynamic event as in, 'The window was broken by the children from the neighborhood.' 6. See the analyses and overviews in e.g.: Comrie (1977), Siewierska (1984, 1988), Babby (1989), Avgustinova et al. (1999), Blevins (2003), Kibort (2004), etc. 7. Serbian periphrastic passive constructions cannot refer to the present when expressing an activity (except when the auxiliary verb be is used iteratively). This use is highly marked and very rare, however. Hence the past tense gloss in the translation. 8. Parts of this analysis were presented in Djurkovic (2004). See also Djurkovic (2007) for a more extensive analysis along the same lines. 9. Blevins (2003: 5) suggests that it might suffice if unaccusatives are defined as merely failing to specify a subject, i.e. without reference to the exact nature of their existing argument. 10. I give examples from Polish because the subjectless passive, although morphologically possible, is rarely used in Serbian (except with sentential complements as subjects). This is presumably because the functionally related intransitive impersonal construction has taken over this use from it. 11. However, this prediction also has a number of exceptions, related to the socalled unaccusativity mismatches, both within a single language and cross-linguistically (cf, e.g., Rosen 1984; McClure 1990; Levin & Rappaport-Hovav 1995; Sorace 2004; Keller & Sorace 2003; etc.). 12. Similar data from Polish, Turkish, Latin, Spanish, Italian, Arabic, Dutch and German are provided by Frajzyngier (1982) confirming the same generalization, although in a functional and not morphosyntactic analysis of the impersonal. See also Kibort (2004), Blevins (2003), Torn (2002) for additional examples. 13. As observed by Blevins (2003: 489), an indefinite human interpretation is not restricted to impersonals, as evidenced by a similar interpretation of subjectless passives in languages that allow them (e.g. in German). This suggests that this interpretation is associated with subjectless forms of personal verbs, irrespective of the source of their subjectlessness. This interpretation also seems to be a matter of degree and while being the only possible one in some languages (Estonian, Finnish, Polish, Lithuanian, Breton), the impersonals in other languages appear to have more of a 'passive' interpretation, independent of human agency, and consequently allow agentive phrases (Ukrainian, Welsh). 14. There are examples of impersonal constructions with the agentive pooltphrases in Estonian, for instance. However, they are characterized as 'not often used' (Tuldava 1994: 273, cited in Blevins 2003: 485), 'intrusive' (Matthews 1955: 370, cited in Blevins 2003: 485), or 'questionable' (Torn 2002: 93) by the authors who cite these examples. 15. For a more detailed discussion and more examples of the case pattern, see Djurkovic (2007).

264

Mil/a Djurkovic

16. For reasons of space, I do not discuss the movement-based, derivational account of the passive in the various versions of the Chomskyan framework here. Suffice it to say that there are numerous passive-related as well as independent arguments against positing various derivational principles and mechanisms such as movement (and consequently Α-chains), abstract case (assignment/checking/ absorption), EPP, etc. In addition, giving up other empirically and theoretically unmotivated assumptions of the derivational approach to grammar such as, for instance, structural and interface uniformity leads to assuming a constraintbased formalism (as proposed above) and encoding the active-passive relations in terms of alternative mapping of argument structure to syntactic functions, (see, e.g., Bresnan 2001; Dalrymple 2001; Culicover & Jackendoff 2005 for extensive argumentation along these lines). 17. See Culicover & Jackendoff (2005) for a review of other independent arguments in favor of positing this level of representation (the GF tier in their terminology). 18. For a more detailed argumentation along these lines and for explicit formalization of these levels within a modified LFG framework, see Djurkovic (2007). 19. The correct picture shows Buggs Bunny in the bath and a dwarf (from Snow White) standing next to the bath with a sponge in his hand. The agent phrase was used in two of the four sentences (no significant difference was observed in children's comprehension in either condition). 20. The correct picture shows Buggs Bunny in the bath and three dwarfs standing around him with sponges/towel/shampoo bottle. The plurality of the agent was supposed to suggest part of the usual interpretation of the impersonal. 21. The correct picture shows a dwarf in the bath and Buggs Bunny standing next to the bath with the sponge in his hand. 22. S(yntactic) Homophone: A phrase α is an s-homophone of a phrase β if α and β have distinct structure but common pronunciation (Babyonyshev et al. 2001: 7). 23. This observation needs some qualification. Namely, both the perfective and imperfective passive in Russian actually have structural homophones (i.e. the resultative participle for the perfective passive, and the reflexive, reciprocal, inherent, etc. use of the sja form attached to verbs for the imperfective passive, cf. Wade 2000; Forbes 1964). However, only the former is also semantically related to the passive and the children only seem to be helped by this form. Therefore, it seems that it is not only important that the two forms are s-homophones they need to be similar in their semantic orientation (as the passive and resultative participles are). Furthermore, as proposed by Kibort (2005), the similarity between the two participles is indeed only semantic, as the resultative participle formation can be argued to involve a morphological change (verb-adjective conversion) and a change in semantic orientation, while the passive participle involves the change in argument structure-function mapping and also the change in semantic orientation. The same situation obtains in Serbian.

Structural vs. frequency effects in LI acquisition

265

24. See, for instance, Bohnacker (this volume) for a short review of the problems with the maturational account with respect to the acquisition of determiners.

References Abbot-Smith, Kirsten & Heike Behrens 2006 How known constructions influence the acquisition of new constructions: the German periphrastic passive and future constructions. Cognitive Science 30(6): 995-1026. Alsina, Alex 1996 The Role of Argument Structure in Grammar: Evidence from Romance. Stanford: CSLI Publications. Andelkovic, Darinka, Nada Ilic Seva & Jasmina Moskovljevic 2001 Serbian Corpus of Early Child Language. Laboratory of Experimental Psychology, Faculty of Philosophy and Department of Linguistics, Faculty of Philology. University of Belgrade. Avgustinova, Tania, Wojciech Skut & Hans W. Uszkoreit 1999 Typological Similarities in HPSG: A Case Study of Slavic Verb Diathesis. In Slavic in Head-Driven Phrase Structure Grammar, Robert D. Borsley & Adam Przepiorkowski (eds.), 1-28. Stanford, CA: CSLI Publications. Babby, Leonard H. 1989 Subjectlessnes, External Subcategorization, and the Projection Principle. Zbornikzafilologiju i lingvistiku XXXII12: 7^0. Babyonyshev, Maria, Jennifer Ganger, David Pesetsky & Kenneth Wexler 2001 The maturation of grammatical principles: Evidence from Russian unaccusatives. In Linguistic Inquiry 32: 1—44. Babyonyshev, Maria & Dina Brun 2004 The acquisition of perfective and imperfective passive constructions in Russian. Penn Working Papers in Linguistics: Proceedings of the 27th Annual Penn Linguistics Colloquium 10(1): 16-31. Baric, Eugenija, Mijo Loncaric, Dragica Malic, Slavko Pavesic, Mirko Peti, Vesna Zecevic & Marija Zinka 1995 Hrvatska gramatika. Zagreb: Skolska knjiga. Belaj, B. 2001 Gramaticki i leksicki pasiv u hrvatskom standardnom jeziku. Ph.D. diss., University of Zagreb. Blevins, James P. 2003 Passives and Impersonals. In Journal of Linguistics 39(3): 1-^8. Bohnacker, Ute this vol. The role of input frequency on article acquisition in early child Swedish.

266

Milja Djurkovic

Borer, Hagit & Kenneth Wexler 1987 The maturation of syntax. In Parameter Setting, Thomas Roeper & Edwin Williams (eds.), 123-172. Dordrecht: Reidel. Borer, Hagit 2004 The Grammar Machine. In The Unaccusativity Puzzle: Explorations of the Syntax-Lexicon Interface, Artemis Alexiadou, Elena Anagnostopoulou & Martin Everaert (eds.), 288-331. Oxford: Oxford University Press. Bresnan, Joan 2001 Lexical-Functional Syntax. Oxford: Blackwell. Brooks, Patricia & Michael Tomasello 1999 Learning the English Passive Construction. In Cognition and Function in Language, Barbara A. Fox, Dan Jurafsky & Laura A. Michaelis, (eds.), 84-98. Stanford: CSLI Publications. Chomsky, Noam 1965 Aspects of the Theory of Syntax. Cambridge, MA: MIT Press 1995 The Minimalist Program. Cambridge, MA; London: MIT Press 2000 The Architecture of Language. Oxford: Oxford University Press. 2005 Three Factors in Language Design. In Linguistic Inquiry 36(1): 1-22. Comrie, Bernard 1977 In defense of spontaneous demotion: The impersonal passive. In Syntax and Semantics 8: Grammatical Relations, Peter Cole & Jerrald M. Sadock (eds.), 47-58. New York: Academic Press. Croft, William 2001 Radical Construction Grammar. Syntactic Theory in Typological Perspective. Oxford: Oxford University Press. Culicover, Peter W. & Andrzej Nowak 2003 Dynamical Grammar. Oxford: Oxford University Press. Culicover, Peter W. & Ray Jackendoff 2005 Simpler Syntax. Oxford: Oxford University Press. Dalrymple, Mary 2001 Lexical Functional grammar. San Diego: Academic Press. Diessel, Holger 2004 The Acquisition of Complex Sentences. Cambridge: Cambridge University Press. Djurkovic, Milja 2004 Passive and Impersonal in English and Serbian. In Working papers in English and applied linguistics 10, Henriette Hendriks (ed.), 49101. Cambridge: Research Centre for English and Applied Linguistics. 2007 Theoretical Status and First Language Acquisition of the Passive and Impersonal Constructions in Serbian. Unpublished Ph.D. diss., University of Cambridge.

Structural vs. frequency effects in LI acquisition

267

Forbes, Nevill 1964 Russian Grammar. Oxford: Oxford University Press. Fox, Danny & Yosef Grodzinsky 1998 Children's Passive: A View from the By-Phrase. Linguistic Inquiry 29:311-332. Frajzyngier, Zygmunt 1982 Indefinite agent, passive and impersonal passive: A functional study. Lingua 58 (3): 267-290. Fried, Miriam in press Agent back-grounding as a functional domain: reflexivization and passivization in Czech and Russian. In Agent demotion, Torgrim Solstad & Benjamin Lyngfelt (eds). Amsterdam: Benjamins. Goldberg, Adele E. 1995 Constructions: A Construction Grammar Approach to Argument Structure. Chicago: University of Chicago Press. 2005 Constructions at work. The Nature of Generalization in Language. Oxford: Oxford University Press. Israel, Michael, Christopher Jonson & Patricia J. Brooks 2004 From states to events: The acquisition of English passive participles. Cognitive Linguistics 11(1/2): 103-129. Ivic, Milka 1963 Kategorija "man-Satze" u slovenskim jezicima. In Godisnjak Filozofskogfakulteta u Novom Sadu, Vol. VII: 93-98. Kibort, Anna 2004 Passive and passive-like constructions in English and Polish. Ph.D. diss., University of Cambridge. 2005 The ins and outs of the participle-adjective conversion rule. In Proceedings of the LFG05 Conference, Miriam Butt & Tracy HollowayKing (eds.). Stanford: CSLI publications. Keenan, Edward. L. 1985 Passive in the world's languages. In: Language typology and syntactic description, Vol.. I: Clause structure. Timothy Shopen (ed.), 243281. Cambridge: Cambridge University Press Keller, Frank & Antonella Sorace 2002 Gradient auxiliary selection in German: an experimental investigation. Journal of Linguistics 39(1): 57-108. Kucanda, Dubravko 1992 Funkcionalni pristup analizi pasiva u hrvatskom. Savremena lingvistika34: 175-184. 1999 Pasivizacija kao strategija subjektivizacije/topikalizacije. JezikoslovIje2(19): 17-33. Kupisch, Tanja this vol. Frequency effects in the acquisition of determiners: Correlation or causality?

268

Milja Djurkovic

Levin, Beth & Malka Rappaport Hovav 1995 Unaccusativity: at the syntax-lexical semantics interface. Cambridge, MA: MIT Press. McClure, William 1990 A lexical semantic explanation for unaccusative mismatches. In Grammatical Relations: A Cross-Theoretical Perspective, Katarzyna Dziwirek, Patrick Farrell & Errapel Mejias-Bikandi (eds.), 45-57. Stanford: CSLI Publications. Maratsos, Michael, Dana E. C. Fox, Judith A. Becker & Mary A. Chalkley 1985 Semantic restrictions on children's passives. Cognition 19: 167-191. Marchman, Virginia & Elisabeth Bates 1994 Continuity in lexical and morphological development: A test of the critical mass hypothesis. Journal of Child Language 21: 339-366 Milosevic, Ksenija 1980 Jedan negramatican recenicni model u srpskohrvatskom jeziku. Juznoslovenskifuolog XXXVI: 47-61. Newmeyer, Frederick 2004 Typological evidence and Universal Grammar. Studies in Language 28 (3): 527-548. Perlmutter, David M. 1978 Impersonal Passives and the Unaccusative Hypothesis. In Proceedings of the fourth annual meeting of the Berkeley Linguistic society, 157-89. University of California, Berkeley. Perlmutter, David M. & Paul M. Postal 1984 The 1-Advancement Exclusiveness Law. In Studies in Relational Grammar 2, David M. Perlmutter & Carol Rosen (eds.), 81-125. Chicago: University of Chicago Press. Rosen, Carol 1984 The interface between semantic roles and initial grammatical relations. In Studies in Relational Grammar 2, David M. Perlmutter & Carol Rosen (eds.), 38-77. Chicago: University of Chicago Press. Roeper, Thomas this vol. What frequency can do and what it can't. Siewierska, Anna 1984 The Passive: A comparative linguistic analysis. London: Croom Helm. Sorace, Antonella 2004 Gradience at the Lexicon-Syntax Interface: Evidence from Auxiliary Selection and Implications for Unaccusativity. In The Unaccusativity Puzzle: Explorations of the Syntax-Lexicon Interface, Artemis Alexiadou, Elena Anagnostopoulou & Martin Everaert (eds.), 243-269. Oxford: Oxford University Press.

Structural vs. frequency effects in LI acquisition

269

Stanqjiic, Zivojin & Ljubomir Popovic 1992 Gramatika srpskogajezika. Beograd: Zavod za udzbenike i nastavna sredstva. Stevanovic, Mihailo 1986 Savremeni srpskohrvatskijezik I i II. Beograd: Naucna knjiga. Sudhalter, Vicki & Martin D. S. Braine 1985 How does comprehension of passives develop? A comparison of actional and experiential verbs. Journal of Child Language 12: 455—47. Tomasello, Michael, Patricia J. Brooks & Elissa Stern 1998 Learning to produce passive utterances through discourse. First Language 18: 223-37. Tomasello, Michael 2003 Constructing a Language: A Usage-Based Theory of Language Acquisition. Cambridge, MA: Harvard University Press. 2004 What kind of evidence could refute the UG hypothesis? Commentary on Wunderlich. Studies in Language 28 (3): 642-664. Torn, Reeli 2002 The status of the passive in English and Estonian. In Working papers in English and applied linguistics 7, Henriette Hendriks (ed.), 81-106. Cambridge: Research Centre for English and Applied Linguistics. Vidakovic, Ivana 2004 Case and Grammatical Relations in English and Serbian. In Working papers in English and applied linguistics 10, Henriette Hendriks (ed.), 245-301. Cambridge: Research Centre for English and Applied Linguistics. Vukojevic, Luka 1992 Sintaksa pasiva. In Rasprave zavoda za hrvatski jezik 18: 235-258. Wade, Terence L. B. 2000 A Comprehensive Russian Grammar. Oxford: Blackwell. Westergaard, Marit & Kristine Bentzen this vol. The (Non-) Effect of Input Frequency on the Acquisition of Word Order in Norwegian Embedded Clauses. Wexler, Kenneth 1999 Maturation and Growth of Grammar. In Handbook of Child Language Acquisition, Ritchie, William C. & Tej K. Bhatia (eds.), 1155. San Diego: Academic Press. Wunderlich, Dieter 2004 Why assume UG? Studies in Language 28 (3): 615-664.

The (non-)effect of input frequency on the acquisition of word order in Norwegian embedded clauses Marit Westergaard and Kristine Bentzen

This paper argues that frequency may have an effect in language acquisition, but only in connection with other factors such as complexity or economy. The study focuses on the acquisition of word order in two types of embedded clauses in Norwegian, using data from a corpus of young children (below the age of three) and some diary notes and experimental data from somewhat older children. Samples of child-directed speech are also investigated. Norwegian is a verb second language in main clauses, while embedded clauses generally display no verb movement. An investigation of the child data reveals that children produce non-target-consistent word order in embedded declaratives, moving the verb across negation or an adverb, while there is no verb movement across the subject in embedded questions. The two clause types are found to have comparable frequencies in the input, both relatively low. Children's target-consistent production in embedded questions is accounted for by a syntactic model where main and embedded questions have different functional structure in the highest domain of the clause, reflecting different illocutionary force. Thus, there is no overgeneralization between the two. Children's errors in embedded declaratives are accounted for by a principle of economy affecting a lower domain in main clauses, and overgeneralization to embedded clauses is therefore possible. Thus, low input frequency is not the cause of the children's errors, but it is argued that once an error pattern appears in the child grammar, low input frequency may make this persist for an extended period of time.

1. Introduction In this paper, we investigate the acquisition of word order in Norwegian embedded clauses. More specifically, we look at how Norwegian children acquire verb placement in embedded w/z-questions and all types of embedded

272

Marit Westergaard and Kristine Bentzen

clauses containing negation or an adverb. We also consider some childdirected speech data, as we believe that it is important for studies in first language acquisition to take into account the role of input in language development. Whether or not one assumes that children are innately endowed with Universal Grammar (UG), it is obvious that certain parts of language, such as vocabulary and phonetic inventory, have to be learned from the primary linguistic data (PLD). Lately, the effect of input on the acquisition process has received considerable attention. In much recent work on language acquisition within the constructivist framework (e.g. Tomasello 2003; Theakston et al. 2004), it is argued that input frequency is vital to understanding both the order of acquisition of particular constructions and children's non-target-consistent production. In fact, it is often argued that children's early multi-word utterances are not the result of rule-governed behavior at all, but that they simply follow from a functionally-based distributional analysis of the input. Thus, in children's early production, there is claimed to be little or no syntactic structure underlying their utterances. This stands in stark contrast to the generative approach to language acquisition, where it is commonly assumed that UG provides the child with the necessary functional structure and constraints, and that all the child needs to do is to learn lexical items and the setting of certain language-specific parameters. Here we argue that input frequency plays a role in the acquisition of word order, but only in combination with other factors. Thus, our approach is in line with several of the contributions to the present volume, e.g. Roeper, Kupisch, and Bohnacker. The children in our study are acquiring a Northern dialect of Norwegian spoken in the city of Troms0. Two constructions with similar input frequencies are investigated: embedded questions on the one hand, and (all) embedded clauses containing negation or an adverb on the other. Both constructions are very infrequent in the input. We show that children make mistakes in embedded clauses with negation or an adverb, overgeneralizing the word order from main clauses (producing structures with verb movement across negation or an adverb). In contrast, they do not overgeneralize main question word order into embedded questions (producing structures with verb movement across the subject). This is accounted for within a Split-CP model of clause structure and a structurebuilding approach to language acquisition where input and economy principles interact in the development of word order. Thus, we argue that the lack of input cues for the target word order in itself is not the reason for children's non-target-consistent production. However, low input frequency

The (non-)effect of input frequency on the acquisition of word order

273

may be one of the contributing factors causing the target word order in embedded clauses with negation or an adverb to be acquired relatively late. While children have to rely on input to acquire the word order in lower domains of the clause, UG provides them with the information that embedded questions are different from main clause questions with respect to illocutionary force. Consequently, children do not project the same functional architecture for the two constructions, and overgeneralization of features from main to embedded questions should therefore be impossible. The paper is organized as follows. In the next section we outline the relevant word order facts of Norwegian. In Section 3 we present the acquisition data from the children in this study, while Section 4 contains an investigation of some of the adult data in the acquisition corpus. Then, in Section 5 we analyze the child data within an economy-based account of language acquisition. Here we also discuss the role of input frequency in the acquisition process. Section 6 is a summary with concluding remarks.

2. The word order of Norwegian Norwegian is a VO language with a rule of verb second (V2), which means that the finite verb has to appear in second position in all main clauses. This is standardly analyzed as verb movement to the topmost head position of the clause, C (see, e.g., Vikner 1995). This can be seen in both subjectinitial and non-subject-initial clauses, as illustrated in (1) and (2), respectively. Norwegian also shows V2 effects in main w/z-questions, as in (3). (1)

John liker ikke tog. John likes not trains 'John does not like trains.'

(2)

Ijjor dro John til Peru to ganger. last-year went John to Peru two times 'Last year John went to Peru twice.'

(3)

Hvorfor liker John tog? why likes John trains 'Why does John like trains?'

In embedded clauses, the finite verb remains within the VP. This is illustrated in (4), where the verb has to follow negation. As we see in example

274

Marit Westergaard and Kristine Bentzen

(5), most embedded clauses (such as embedded w/z-questions) do not allow V2, as the verb must also follow the subject. (4)

Jeg kjenner en mann [som {*liker} ikke {liker} tog]. I know a man who likes not likes trains know a man who doesn't like trains.'

(5)

Har du h0rt [hvorfor {*liker} John {liker} tog]? have you heard why likes John likes trains 'Have you heard why John likes trains?'

There are some exceptions to the generalization that the verb does not move in embedded contexts. First of all, Norwegian in general optionally allows verb movement in that-c\auses that are complements of assertive and semi-factive predicates (say, believe, discover, etc.). In the subjectinitial embedded clause in (6) verb movement past negation is optional (cf. the subject-initial main clause in (1)). In the non-subject-initial embedded clause in (7) V2 is obligatory (cf. the non-subject-initial main clause in (2)). Although verb movement past negation is accepted in sentence (6), the preferred option in Norwegian is generally to leave the verb in the VP, according to Garbacz (2005).' However, as shown in (7), these constructions allow embedded topicalization, and then subject-verb inversion is obligatory. Thus, certain thai-clauses like those in (6) and (7) are arguably contexts where embedded V2 is available. (6)

Hun sier fat John {liker} ikke {liker} tog lenger]. she says that John likes not likes trains longer 'She says that John doesn't like trains any longer.'

(7)

John sa at [ißor dro han til Peru to ganger]. John said that last-year went he to Peru two times 'John said that he was in Peru twice last year.'

Secondly, Bentzen (2003, 2005, 2007) has shown that several Northern Norwegian dialects also allow verb movement past certain adverbs in nonV2 contexts, such as embedded w/z-questions, relative clauses, and adverbial embedded clauses. The Troms0 dialect, which is the target dialect of the children in the current study, allows finite auxiliaries preceding certain adverbs such as ofte Often' and allerede 'already,' as illustrated in (8) below. Again, verb movement is not the preferred option, and we may thus

The (non-)effect of input frequency on the acquisition of word order

275

assume that the word order in (8) is relatively infrequent in the input. Crucially, verb movement is never possible past negation and certain other adverbs such as heldigvis 'fortunately' and ogsä 'also', as we see in (9). (8)

Vi begynte ä bli spente na... we began to become excited now ... ettersom vi ville allerede kunne vite resultatet päfredag. ... as we would already could know result.the on Friday 'We started to get excited now as we would be able to know the result already on Friday.'

(9)

*... ettersom vi ville ikke kunne vite resultatet for päfredag. ... as we would not could know result.the before on Friday '... as we would not be able to know the result until Friday.'

There are also some exceptions to the V2 requirement in main clauses. In several Norwegian dialects, verb movement is optional in main whquestions (cf. Westergaard 2003; Westergaard & Vangsnes 2005; Vangsnes 2005). The Troms0 dialect has optional V2 in main wA-questions with the monosyllabic wA-words kern 'who,' ka 'what,' and kor 'where', as we see in (10)-(12). The example in (13) shows that in the non-V2 cases the verb also has to follow negation, in line with the restriction on moving the verb past negation in non-V2 contexts (cf. example (9) above): (10) Kern {like} han John {like} best? who likes DET John likes best 'Who does John like the best?' (11) Ka {er}favorittlandet ditt {er}? what is favorite-country.the yours is 'What is your favorite country?' (12) Kor {parkerte} han {parkerte} bilen henne? where parked he parked car. the LOC 'Where did he park the car?' (13) Kern han {*lante} ikke {länte} penga til? who he lent not lent money to 'Who didn't he lend money to?'

276

Marit Westergaard and Kristine Bentzen

In wA-questions with long w/z-phrases and the disyllabic question words koffer 'why', korsn 'how', and katti 'when', V2 is obligatory in main whquestions in this dialect, as illustrated in (14)-(16): (14) Koffer {gikk} han {*gikk} hjem sä tidlig? why went he went home so early 'Why did he go home so early?' (15) Korsn {visste} du {*visste} kem del var? how knew you knew who it was 'How did you know who it was?' (16) Katti {lande} flyet din {*lande} i Lima? when lands plane.the yours lands in Lima 'When does your plane arrive in Lima?' Summing up, Norwegian in general has obligatory verb movement to the second position in all kinds of main clauses, but no verb movement in embedded clauses. However, embedded V2 is possible in thai-clauses embedded under certain predicates. Furthermore, the Troms0 dialect optionally allows verb movement past certain adverbs, but not past negation in nonV2 contexts, such as embedded w/z-questions, relative clauses, and embedded adverbial clauses. In addition, this dialect has optional V2 in a subgroup of w/z-questions (those introduced by kem 'who,' ka 'what,' and kor 'where'). In this study, we focus on how Norwegian children growing up in Troms0 acquire verb placement in embedded clauses. More specifically, we investigate how they acquire verb placement with respect to the subject in embedded w/z-questions, as in (5), on the one hand, and how they acquire verb placement with respect to negation and adverbs in (all) embedded clauses and non-V2 main w/z-questions, as in (4) and (13), on the other. We consider data from three very young Norwegian children below the age of 3, as well as data from two older children, up to the age of 8. In order to consider the potential effect of input frequency, we also investigate a sample of some adult data in an acquisition corpus. 3. Previous studies and Norwegian child data Previous studies on the acquisition of word order in V2 languages suggest that verb placement in main clauses is in place from very early on. Wester-

The (non-)effect of input frequency on the acquisition of word order

277

gaard (2005) shows that this is also the case for Norwegian children. As soon as multi-word utterances appear in the child data, verb movement generally applies in non-subject-initial clauses, questions, and subject-initial clauses with negation or adverbs. Such early acquisition of V2 in main clauses is also attested in Swedish (Santelmann 1995; Platzack 1996), Dutch (Jordens 1990), German (Poeppel & Wexler 1993; Müller 1996), and Lucernese Swiss German (Schönenberger 2001). Findings concerning the acquisition of word order in embedded clauses are more varied. Clahsen & Smolka (1986) find that German-speaking children correctly place the verb clause-finally in their very first production of embedded clauses. Penner (1996) reports on data from a Bernese Swiss German child which indicate that there is correct verb placement in embedded contexts (clause-finally) until about the age of 3;2, but this stage is followed by a period of a few months when the child produces embedded clauses both with and without verb movement. Occasional non-targetconsistent verb movement in German embedded clauses is also reported for monolingual German children by Gawlitzek-Maiwald, Tracy & Fritzenschaft (1992), and for bilingual German-English children by Döpke (1998). Furthermore, Schönenberger (2001) found that her two Lucernese Swiss German subjects consistently moved the finite verb in a non-target-like manner in embedded clauses. This pattern occurred until the age of 4;11, when the verbfinal pattern gradually took over. Finally, Häkansson & Dooley Collberg (1994) have shown that Swedish-speaking children seem to move modal auxiliaries across negation and adverbs in Swedish embedded clauses. Again, this is a pattern not found in the adult language. Previous findings are thus inconclusive, some studies suggesting that verb placement in embedded clauses is unproblematic, whereas others report this to be an area where children make mistakes for an extended period of time. In the following sections, we present data from Norwegian-speaking children indicating that there is evidence of overgeneralization of verb movement past negation or an adverb into constructions that do not allow verb movement in the target language. However, non-target-like verb movement past subjects is not attested in the children's production. 3.1. Young children In this section, we provide some evidence from three very young Norwegian children. These data come from a relatively large corpus consisting of

278

Marit Westergaard and Kristine Bentzen

altogether 70 one-hour recordings of three children between the age of approximately 1;9 and 3;3. ^ Given the young age of these children, there are not many instances of embedded clauses in the data. However, there are a few relevant examples in the later files, altogether 108 embedded questions, 28 embedded clauses with negation or an adverb and one non-V2 main whquestion with negation. Let us start with the 28 embedded clauses with negation or an adverb (all declaratives). As many as 15 of these had to be excluded from our discussion because they are unclear with respect to the question of verb movement. In three of these examples, although they display the targetconsistent non-V2 word order, the verb involved seems to be non-finite, and judging from the context, there is a modal missing in the structure.3 The remaining 12 of the excluded examples display the most common word order in the children's embedded clauses, viz. the order where negation ikke/ikkje 'not' occurs immediately following the complementizer (if present), i.e. above the verb as well as the subject, as illustrated in (17). This Neg-S-Vfm word order is also possible in the adult grammar, although as mentioned in Section 2, it is much less frequent than S-Neg-Vfm (see Note 1). It is also unclear exactly what position negation is attached to in such sentences.4 Nevertheless, we must conclude that these sentences cannot reveal anything about possible verb movement. (17) neihoskal passe pä mce ikkje reven komme a ta mce. no she shall watch on me not fox.the comes to take me 'She is to watch out so the fox doesn't come and take me.' Preferred: Ho ska!passe pä mce sä reven ikkje kommer og tar mce. (Ina.l8,2;8,12) The 13 remaining embedded clauses with negation were relevant to the current study. Four of these 13 sentences can be said to be true examples of target-consistent word order without verb movement in embedded contexts. Two of these are illustrated in (18) and (19), where negation appears between the subject and the verb, indicating that no verb movement has taken place. (18) ikkeda[//]at detda ikkeblir stramt. not then that it then not becomes tight '... that it doesn't get (too) tight.'

(Ole.18,2;9,15)

The (non-)effect of input frequency on the acquisition of word order

(19) bare när dem ikke hold pa da dettedemxxx. only when they not hold on then fall they xxx Only when they are not holding on, then they fall.'

279

(Ina.27, 3;3,18)

Target form: Bare när dem ikke hold(er) (fast?), da dcett dem.

But nine of the examples in the child data do in fact display verb movement in embedded contexts. Five of these are f/za/-clauses in which the target grammar also allows (but disprefers) verb movement. One of these is illustrated in (20). The four remaining embedded clauses in the corpus exhibit verb movement in non-V2 contexts where it is clearly ungrammatical in the target language. An example is given in (21). (20) hansa hanville ikke spise [?]. he said he would not eat him 'He said that he wouldn't eat him.' Preferred: Han sa at han ikke ville spise han. (21) deter ho mamma som har ogsa tegna. it is she mommy who has also drawn 'It is mommy who has also been drawing.'

(Ann. 17, 2;8,4)

(Ina.26, 3;2,5)

Target form: Dei er ho mamma som ogsa har tegna.

Another construction where the target language does not show verb movement is non-V2 main w/z-questions, as we saw in example (13) above. In the data from the younger children there is one such question containing negation, and in this example the verb has indeed moved across negation. The word order in (22) indicates that there is overgeneralization of verb movement in these cases, not to a position in the clause structure above the subject, but presumably to a lower functional head. (22) kem somvil ikkje vcere Hag med han? who that will not be together with him 'Who doesn't want to be with him?'

(Ina.25, 3; 1,8)

Target form: Kem som ikkje vil vcere i lag med han?

Thus, 10 out of 14 relevant examples show that all three children seem to prefer verb movement over a word order without movement, both in cases where it is completely ungrammatical in the target grammar and in cases where it is only dispreferred (in certain //za/-clauses). This is interesting, as it goes against a minimalist account, where movement is always considered

280

Marit Westergaard andKristine Bentzen

to be a more costly operation than no movement. It could of course be the case that these examples are restarts - i.e. that they are biclausal structures with two main clauses. Such an explanation is also supported by the fact that none of these clauses are introduced by complementizers. However, within a minimalist account it is generally assumed that children start out with the least costly approach to word order, viz. a structure with no movement (cf. Platzack 1996; Roberts 1999), and that they only produce structures with movement if there are strong and consistent cues for this in the input. Given that the option not to move the verb should be available to the children (as they do produce some embedded clauses with targetconsistent word order), it is surprising from the point of view of a minimalist account that they seem to prefer the least economical structure in these cases. This indicates that economy interacts with other factors in the acquisition of word order. Next we consider embedded questions, of which there are a total of 108 examples in the corpus. What is striking about these clauses is that virtually all of them occur with target-consistent non-V2 word order; that is, with no verb movement across the subject. Examples of these embedded questions are provided in (23)-{25), both from relatively early and relatively late files in the corpus. (23) se her ka Inagjer. look here what Ina does 'Look here what Ina is doing!'

(Ina.04, 1;11,22)

(24) Ann vet ikkekor han erkenne. Ann knows not where he is LOG 'Ann doesn't know where he is.'

(Ann.09,2;2,19)

(25) skalcEvise # korsen man tr0kke pä knappen? (Öle.20, 2; 10,15) shall I show ... how one pushes on button.the 'Do you want me to show (you) how you push the button?' There is only one potential exception to the lack of overgeneralization of V2 word order in embedded w/z-questions in the child data, and this is illustrated in (26): (26) du, ser du ka er det der sann der der? (Ina.27, 3;3,18) you see you what is that there such there there 'You, do you see what that is/do you see: what is that?'

The (non-)effect of input frequency on the acquisition of word order

281

A possible explanation for the word order in (26) is again that the child is producing a biclausal structure, i.e. that there should be a restart between the question word ka 'what' and the preceding part of the sentence, as illustrated by the alternative translation. This analysis is not completely implausible, especially given the linguistic context in the file. It is clear that the adult responding to the child's question has interpreted this as a w/z-question and not as a yes/no-question, the reply to (26) being a specification of what "that" refers to.5 In any case, 1 out of 108 examples does not constitute evidence that there is a rule of verb movement across the subject in embedded questions in the child's internalized grammar, and we therefore conclude that in general there is no evidence of overgeneralization of V2 from main into embedded questions. Another possible type of verb movement in these embedded w/z-questions would be movement across negation or an adverb, as we saw in the embedded declarative clauses above. Unfortunately, none of the 108 embedded questions in this corpus include negation or an adverb. This section has investigated the occasional examples of embedded clauses in early child data. To summarize so far, virtually no cases of verb movement across the subject were attested in embedded questions, see (23)-(25). In embedded clauses with negation or an adverb, on the other hand, the majority of cases show overgeneralization of verb movement. This was the case in four clear examples, illustrated in (21), and five further examples of embedded thai-clauses, as shown in (20). In comparison, only four examples of target-consistent word order without verb movement were found in the child data, see (18)-(19). Furthermore, the only relevant example of a non-V2 main w/z-question also displayed verb movement across negation, as illustrated in (22). These findings are summarized in Table 1, and as we can see, the children moved the verb across negation in 10 of the (preferred) non-V2 contexts, whereas they left it in the target-like position following negation only 4 times. Table 1. Overview of word order in embedded clauses with negation or an adverb in the corpus of three Norwegian children, age approximately 1;9 to 3;3

Embedded thai-clauses Other embedded declaratives Non-V2 wA-questions Total

(S)-Vfin-Neg/Adv

S-Neg/Adv-Vfin

5 4 1 10

0 4 0 4

282

Marit Westergaard and Kristine Bentzen

3.2. Older children The investigations of the older children are based on sporadic recordings and diary notes from two boys, Henning (2;4,4-8;0,17) and Iver (1;8,105;9,15), as well as the results from a small experimental study with the same two children at the age of 8;0,20 and 5;9,18. The patterns reported for the very young children in the section above are generally confirmed in the data from the older children. In the recordings and diary notes several embedded clauses with negation are attested, especially from the age of around four. The data show that the children at this stage display both verb movement and V in-situ in a target-like manner in thai-clauses, as illustrated in (27) and (28). Both of these examples are acceptable, but the word order in (27) is dispreferred in the adult language. (27) ce vet at ce har ikke gjort det. I know that I have not done it Ί know that I haven't done it.' (28) cesa at han ikkesku ... I said that he not should Ί said that he shouldn't...'

(Henning 4;8,13)

(Henning4;8,0)

In Bentzen (2003) it was shown that these children also move the verb past negation in many other types of embedded non-V2 contexts. Such movement was attested in relative clauses, as illustrated in (29)-(30), in adverbial subordinate clauses, (31), and in embedded w/z-questions, as illustrated in (32)-(33). There is also one instance in the data of verb movement in a non-V2 main ^//-question, given in (34). The following examples are all ungrammatical in the adult language: (29) ce like alt som er ikke sterkt og alt som er sterkt. I like everything that is not hot and everything that is hot Ί like everything that isn't hot, and everything that is hot.' (Henning 4;2,7) (30) du ma fa da} enbiffkniv som er ikke sann. you must get you a steak.knife that is not like-that 'You need to get a steak knife that isn't like that.'

(Iver 5;8,16)

The (non-)effect of input frequency on the acquisition of word order

283

(31) ce ma ta pä ullcesta for at ce skal ikke bli sä kald. I must take on wool.socks for that I shall not get so cold need to put on wool socks in order to not get too cold.' (Iver4;ll,29) (32) nar han Iver er ikke her sa kan ce ta med when DET Iver is not here then can I take with den store skjeia. the big spoon.the 'When Iver isn't here, I can use the big spoon.'

(Henning 4;6,27)

(33) men ce lik 'ikke det när det er ikke sann. but I like-not it when it is not like-that 'But I don't like it when it isn't like that.'

(Henning 4;7,16)

(34) kem som var ikke helt i form? who that was not completely in shape 'Who wasn't feeling too well?'

(Henning 4;5,0)

Several embedded w/7-questions were also attested in the recordings and diary notes of the older children. None of these exhibit verb movement past the subject, as illustrated by the following examples. (35) vet du ka det her er, tante? know you what this here is auntie 'Do you know what this is, auntie?' (36) ce vet korsn dem läge et hus sann her. I know how they make a house like-that here know how to make a house in this way.'

(Henning 3; 11,12)

(Iver4;7,10)

Thus, evidence from the sporadic recordings and diary notes shows that the overgeneralization of main clause word order into embedded declaratives found in the very young children is also attested for the two older children. Furthermore, the lack of such overgeneralization into embedded w/iquestions in the corpus of the young children is confirmed in the data from the older children. However, the sporadic recordings and diary notes do not give any indications as to how frequently these two children display the non-target-like word order, nor for how long such patterns persist in the

284

Marit Westergaard and Kristine Bentzen

children's grammars. Therefore, a small experiment was conducted with the two children at the age of 5;9,18 and 8;0,20. The small experimental study was designed to elicit embedded whquestions with negation or an adverb. In the experiment, we introduced the children to the hippo Härek. The children were told that Härek was a very peculiar hippo who had three special features: (i) he claimed to have the best memory in the world, (ii) he did not talk to adults, and (iii) importantly, he would not respond to you unless you started your sentences with 'Do you remember...?'. The investigator (the second author) read a story with the children about a four-year-old boy, Karsten, who was ill and had to stay at home instead of going to kindergarten. The children were told that Härek also knew the story, and that they were now going to test how much he remembered of it, asking questions starting with 'Do you remember...?'. We attempted to elicit altogether 16 embedded questions, 12 of which were supposed to contain negation or an adverb. The remaining four questions where included as fillers. The elicitation setup is illustrated in (37): (37) INV: So, Karsten didn 't go to kindergarten today, and that was because he was ill. Therefore, he didn't go to kindergarten. We remember that that was why, but ask Härek whether he remembers why. CHILD: Do you remember why Karsten didn't go to kindergarten today? The older child, Henning, included negation or an adverb in 11 of the 12 designated questions, and in all cases, negation or the adverb preceded the verb in a target-like manner, as shown in (38)—(40): (38) huske du koffer han Karsten ikke var i barnehagen? remember you why DET Karsten not was in kindergarten.the 'Do you remember why Karsten wasn't in the kindergarten?' (39) huske du koffer ho ikkeville kjope den potta? remember you why she not wanted buy that pot.the 'Do you remember why she didn't want to buy that pot?' (40) huske du koffer en mann ikkeßkk kj0pe L0veungen? remember you why a man not got buy lion.baby.the 'Do you remember why a man didn't get to buy the lion baby?'

The (non-)effect of input frequency on the acquisition of word order

285

The younger child, Iver, included negation or an adverb in only 8 of the 12 designated questions, and in 7 of these 8, he produced the non-target-like word order with the verb preceding negation or the adverb, as illustrated in (41H43): (41) huske du koffer han Karsten vor ikke i barnehagen? remember you why DET Karsten was not in kindergarten.the 'Do you remember why Karsten wasn't in the kindergarten?' (42) huske du koffer dama ville ikke kjope en nattpotie? remember you why lady.the wanted not buy a night.pot 'Do you remember why the lady didn't want to buy a chamber pot?' (43) huske du koffer Loveungen var ikke til salgs? remember you why Lion.baby.the was not to sale 'Do you remember why the Lion baby wasn't for sale?' Neither of the children moved the verb past the subject in these embedded w/z-questions. Thus, the sporadic recordings, the diary notes, and the small experiment with older children constitute evidence that children up to the age of (at least) around 6, optionally move the verb past negation and adverbs in non-V2 contexts. Furthermore, the small experiment shows that they do not move verbs past subjects in embedded questions. One possible explanation for the children's word order patterns in embedded clauses could be that there is a word order change taking place in the language. This seems unlikely, however, given that most of the examples of verbs preceding negation or adverbs in the diary notes and sporadic recordings are from the older child Henning at the age of approximately 45. The fact that at the age of 8 he hardly uses this word order anymore suggests that this is a feature of a certain developmental stage in the acquisition process, rather than e.g. an indication of a syntactic change taking place in the dialect. Summing up the investigation of both very young and somewhat older Norwegian children, we found substantial evidence for overgeneralization of verb movement past negation and adverbs from main clauses into embedded clauses, at least up to the age of 6. Such verb movement is generally not accepted in the target language. In the few contexts where it is possible in the target language, viz. past certain adverbs in non-V2 contexts, and past both negation and adverbs in certain thai-clauses, verb movement

286

Marit Westergaard and Kristine Bentzen

is the dispreferred option. Within a minimalist account of language acquisition, it is at first sight unexpected that children should prefer verb movement where it is not allowed or dispreferred in the target language. Assuming that economy principles play an important role in language acquisition, one would expect children to avoid costly operations such as verb movement, unless there is strong and consistent evidence for such movement in the input. Furthermore, it does not seem to be the case that the children in the study are simply applying main clause V2 word order in embedded clauses in general. This is evident from the fact that they do not move the verb past the subject in embedded w/z-questions in analogy with main w/z-questions. Thus, what needs to be explained is why children overgeneralize verb movement past negation and adverbs but not past subjects. In the following sections, we discuss possible reasons for this asymmetry in the acquisition of verb placement. An important question is whether this is a result of asymmetries in the frequency of the relevant constructions in the input, or whether other factors may play a role.

4. Input frequencies As mentioned in the introduction, there has recently been an increased interest in the role of the input, within functional as well as formal approaches to language acquisition, and especially within the constructivist framework, where input is often argued to be the sole explanation for acquisition orders and children's error patterns. An example of a constructivist approach relevant to the constructions at hand is Tomasello (2003), who argues that children's early production of embedded clauses provides no evidence for a hierarchical structure in children's linguistic systems. Investigating examples from German child language, he finds that early embedded clauses always appear with the same matrix verbs, normally only two or three different ones. Therefore, these are better analyzed as linear constructions, he argues, where the matrix verb is simply stuck onto the beginning of the clause, which remains a main clause structure. Applying this line of reasoning to the Norwegian child data in the previous section, it could be argued that the embedded clauses with negation are not really embedded constructions, but rather main clauses with an initial chunk that looks like a matrix clause. Thus, we find main clause word order in these constructions with the verb preceding negation or an

The (non-)effect of input frequency on the acquisition of word order

287

adverb, as was illustrated in e.g. (20) and (21). Presumably the V-Neg/Adv combination in the main clause would on this approach not be the result of verb movement, but simply a linguistic chunk which is reproduced from main to embedded clauses (or rather, structures which look like embedded clauses). If children learn syntactic structure from input only, we would then expect to find the following frequencies of the relevant constructions in the input that the children in this study are exposed to: embedded clauses with negation or an adverb should be infrequent in the input, since this is the clause type where children make word order mistakes for an extended period of time. On the other hand, main clauses with negation or an adverb should be relatively frequent, since this is where the V-Neg/Adv pattern that the children are overgeneralizing is found. Furthermore, embedded questions should also be quite frequent, as the children were found not to overgeneralize the V-Subject word order found in main w/z-questions. Obviously, it is possible to argue that children's early utterances have more syntactic structure than what is normally assumed within a constructivist approach and still argue for a frequency effect. On such an approach within a generative framework, the word order of main clauses would be the result of verb movement, and because of a high frequency of main clauses with V2, this type of movement would then be overgeneralized to embedded clauses. This means that frequency would override economy in this case, since, as discussed above, a common idea within the minimalist framework is that syntactic movement is always more costly than the lack of movement. For a frequency effect to play a role here, one would expect to find exactly the same input frequencies as was sketched above for the construct!vist approach: embedded clauses with negation or an adverb should be considerably less frequent than the corresponding main clauses, while embedded questions are expected to be quite frequent. In order to get an indication of what child-directed speech may consist of in terms of frequency of syntactic constructions, some samples of the adult material from the Troms0 corpus were investigated in detail. First, one file (corresponding to approximately one hour of spontaneous speech) was hand-searched and all complete sentences of the investigator (INV) were counted. In this file, the investigator produced a total of 793 utterances, out of which there were 668 complete clauses, 554 matrix and 114 embedded clauses. There are altogether 123 subject-initial main clause declaratives in the sample, 43 of which contain negation or an adverb, see Table 2. This means that the evidence for verb movement across negation

288

Marit Westergaard and Kris tine Bentzen

or adverbs in main clauses makes up 6.4% of the total input. Furthermore, there are as many as 337 examples of questions and non-subject-initial declaratives (50.4%), providing the child with evidence for verb movement across the subject. This means that there is ample evidence in the input that Norwegian is a V2 language in main clauses, see Westergaard (2006) for a more detailed analysis. Similar findings have been attested for much larger samples of Swedish input data in Josefsson (2004), altogether 14,033 adult utterances, where V2 constructions such as yes/no-questions are attested in 22-28% of all utterances, and non-subject-initial declaratives 12-27%. Table 2. Overview of evidence for V2 and non-V2 in a sample of child-directed speech, the investigator in the file Ole.14 (age of child 2;6,21), with percentages calculated relative to the total number of complete (matrix and embedded) clauses (N=668) Evidence for V2

Evidence for non-V2

Subject-initial decl. 6.4% (43) Embedded clauses with Neg/Adv 0.9% (6) with Neg/Adv Non-V2 w/z-questions with Neg/Adv 0.1 % (1) Non-subject-initial 50.4% (337) Embedded questions 1.6% (11) decl. and questions

The evidence that the verb does not move past negation or adverbs in nonV2 contexts should be expressed in all embedded clauses as well as in nonV2 main w/z-questions with negation or adverbs. These constructions are indeed very infrequent in the input. As illustrated in the right-hand column of Table 2, the investigator produces only six embedded clauses with negation or an adverb in the file, corresponding to 0.9% of the input in the sample. One of these sentences is illustrated in (44) and could be compared to the non-target-consistent child utterances in (20)-(21) and (29)-(33) above. (44) pass pä at denikkje faller over. watch on that it not falls over 'Watch out so it doesn't fall over.'

(IN V, file Ole.14)

Furthermore, there is only one example of a non-V2 main w/z-question with negation, which increases the evidence for the lack of verb movement across negation, but only by 0.1%. This example is given in (45) and should be compared with the non-target-consistent child utterances in (22) and (34) above.

The (non-)effect of input frequency on the acquisition of word order

(45) kern som ikkjefär kjore? who that not gets drive 'Who doesn't get to drive?'

289

(INV, file Ole.14)

This means that the total evidence for the lack of verb movement across negation or an adverb in Norwegian embedded contexts and non-V2 main w/z-questions is attested in only 1.0% in the input sample. Moreover, there is also an example of an embedded {hat-clause with the word order V-Adv in this file, which is illustrated in (46). As discussed in Section 2 above, these are also grammatical in the target language, further complicating the structures that have to be acquired by the child. (46) ce trur hanmä bare sitte der. I think he must only sit there think he just has to sit there.'

(INV, file Ole.14)

So far, our predictions with respect to frequency seem to be borne out: the evidence for Neg/Adv-V word order in embedded contexts and non-V2 main w/z-questions is extremely infrequent in the input (1.0%), and compared to the 6.4% evidence for the opposite word order in subject-initial main clauses, it could be argued that the more frequent word order is overgeneralized to the less frequent one. The 50.4% evidence for V2 in nonsubject-initial declaratives and questions, i.e. a word order where the verb precedes the subject, could possibly be added to this, as these utterances provide the child with general evidence for verb movement in the language. But what about the embedded questions, which were also predicted to be frequent in the input? As illustrated in Table 2, it turns out that evidence that the verb does not move across the subject in embedded (non-subject) w/z-questions is also very infrequent in the sample of adult data. The investigator produced only eleven such examples, making up as little as 1.6% of the input sample. One of these is provided in (47). (47) vet du ka slags farge det er? know you what kind color that is 'Do you know what color that is?'

(INV, file Ole.14)

Since the input sample discussed here is quite small and also produced by only one person, a more focused search of larger samples of the corpus was made, in order to check whether more considerable differences in input

290

Marit Westergaard and Knstine Bentzen

frequencies could be attested between the two types of embedded constructions requiring non-V2 word order. More specifically, we searched for negation (no adverbs) and specific question words in the production of the investigator in files Öle. 15-22, and in the production of one of the parents in files Ann.01-21 (MOT). As shown in Table 3, the investigator produced altogether 6,351 utterances. Out of these, there were 32 (0.5%) embedded clauses with negation, and no non-V2 main w/z-question with negation. In addition, the investigator produced 66 (1.04%) embedded (non-subject) wA-questions. Ann's mother produced a total of 8,860 utterances, 39 (0.44%) of these were embedded clauses with negation, and 41 (0.46%) were non-V2 main w/z-questions with negation. Furthermore, she produced 224 (2.5%) embedded (non-subject) w/z-questions. Thus, the more focused search in the corpus also indicates that the evidence for not moving the verb in non-V2 contexts is relatively infrequent. For both adult speakers the number of embedded questions is somewhat higher than the total number of clauses providing evidence for the lack of verb movement across negation or an adverb, 32 (0.5%) vs. 66 (1.04%) for the investigator, and 80 (0.90%) vs. 224 (2.5%) for Ann's mother.6 Table 3. Overview of utterances providing evidence for non-V2 word order in samples of child-directed speech, the investigator (INV) in files Ole.1522 (N=6351, all utterances) and the mother (MOT) in files Ann.01-21 (N=8860, all utterances) Evidence for non-V2

INV MOT

Emb. clauses w/Neg

Non-V2 w/z-questions w/Neg

Total

Embedded w/z-questions

32(0.5%) 39 (0.44%)

0 41 (0.46%)

32(0.5%) 80 (0.9%)

66(1.04%) 224(2.5%)

Thus, in the small hand-searched adult sample, as well as in the focused search of the larger corpus samples, there is a slightly higher percentage of embedded questions than embedded contexts with negation or adverbs. However, we doubt that a difference between e.g. 0.5% and 1.04% could be the only explanation for children producing a considerable number of non-target-consistent constructions in the former case and displaying a virtually error-free production in the latter. Why would 1.04% be enough input to acquire a certain word order, while 0.5% - or the 0.9% produced by the mother - is not? And even if the children's production were the result

The (non-)effect

of input frequency on the acquisition of word order

291

of differences in input frequency, one would expect such a small difference to have an effect only for a short period of time. However, as shown in Section 3 above, the non-target-consistent word order produced by the children in embedded contexts is quite persistent, possibly lasting well into school age. We thus find it highly unlikely that frequency could be the sole explanation for this, and we therefore reject a purely constructivist approach to the child data. Moreover, we believe that such an approach would also have a problem explaining why the V-Subject word order of all main clauses with V2 does not overgeneralize to embedded questions. After all, in the small input sample investigated in Table 2, the difference in input frequency between main clauses with V2 (V-Subject) and embedded clauses (Subject-V) is as much as 50.4% vs. 1.6%, which is much higher than the difference between main and embedded clauses with respect to the position of the verb in relation to negation or an adverb (6.4% vs. 1.0%). Thus, if input frequencies were responsible for overgeneralization from the relatively frequent V-Neg/Adv word order of main clauses to embedded contexts, we see no reason why the extremely frequent V-Subject word order should not overgeneralize in the same way. We therefore want to argue that the results of our investigation of the input clearly reveal that other aspects of language acquisition such as complexity or economy must be invoked to explain the error patterns described in Section 3. This will therefore be the focus of the next section. 5. An economy-based account Having rejected an analysis which explains the children's performance solely by reference to input frequency, we will now turn to an account of the observed facts in terms of economy, complexity and to a certain extent, frequency. The framework we adopt is a Split-CP model of clause structure, and this is outlined briefly in the next section. In Section 5.2 we account for the appearance of the children's non-target-consistent word order in embedded contexts with negation by referring to an economy principle which causes them not to move elements higher up in a clausal structure than there is evidence for in the input. The reason why this does not apply in embedded questions is related to the syntactic model we adopt, where main and embedded clauses have different clausal architecture, reflecting their different illocutionary force. Finally, in Section 5.3 we discuss some reasons why the children's errors in embedded clauses are so persistent, and here frequency will be argued to play a role.

292

Marit Westergaard and Kristine Bentzen

5.1. Theoretical background

We adopt a Split-CP model of clause structure, which is inspired by Rizzi's (1997) original model and later work on Italian syntax, e.g. Rizzi (2001), Benincä & Poletto (2004), but which is in many ways different from these accounts. The model was originally developed in Westergaard & Vangsnes (2005) and somewhat revised in Westergaard (2005, 2006). The most important aspect of the model is that different clause types have different heads in the CP-domain, reflecting the illocutionary force of the clause type. For example, a w/z-question is an Int(errogative)P(hrase), a yes/noquestion a Pol(arity)P, and a declarative a Top(ic)P. The syntactic heads in the CP and IP domains of the clause that are relevant for the present discussion are provided in the bracketed structure in (48): (48)

C p[(Int°

Top 0 ...) ... (Wh°) Fin0 ,P[InTop° T°

Another crucial aspect of this model for our present purposes is that embedded clauses have a restricted CP domain. That is, embedded declaratives are assumed to be bare Fin(initeness)Ps, while embedded questions are bare WhPs. This reflects the different illocutionary force of main and embedded clauses. For example, embedded questions are not 'real' questions and lack interrrogative force, and thus there is no Int° head present in the clausal structure. The model was developed mainly to account for different types of V2 grammars in English and Norwegian dialects, many of which have no strict V2 requirement in w/z-questions, as mentioned in Section 2. The main parametric tool of the model is the presence of a specific EPP head feature on individual functional heads in the CP domain, called [X°Epp]·7 This feature must be lexicalized, a requirement which may be met by verb movement. Grammars differ with respect to whether a particular head is endowed with this EPP feature, which means that there are several sources for V2 word order. According to this model, Norwegian dialects which have no V2 requirement in w/z-questions, e.g. the Nordm0re dialect described by Äfarli (1986), are argued to have no EPP feature on the Int° head, but as they are strictly V2 in declaratives, this feature is present on the Top0 head. English, which has subject-auxiliary inversion in all questions but (generally) no inversion in declaratives, has the opposite requirements on these two heads. The Troms0 dialect, which the children in this study are acquiring, is argued to have the EPP head feature on Int° as well as the Top0 head. The

The (non-)effect of input frequency on the acquisition of word order

293

former is necessary to account for V2 word order in w/z-questions introduced by long w/z-phrases (see examples (14)-(16) in Section 2), while the EPP feature on the Top0 head accounts for verb movement in all declarative sentences, across the subject in non-subject-initial declaratives and across negation or an adverb in subject-initial declaratives.8 Finally, the distinction between main and embedded clauses with respect to the presence of C-heads accounts for the differences between main vs. embedded clause word order: the heads Fin0 and Wh° are not endowed with the EPP feature, and consequently, there is no verb movement to the CP domain in embedded clauses in Norwegian. For our present analysis, we also adopt a general view of language acquisition which could be described as a continuity approach which includes structure-building (see Westergaard 2005). The continuity aspect of this is taken care of by a universal "pool" of possible functional categories, where rules for their relative order (and a number of other rules and constraints) are provided by UG. In the process of language acquisition, children select categories from this universal set, based on principles of UG and cues in the input. Additionally, children need cues to know how the different functional projections are realized syntactically in their particular language, e.g. by verb movement triggered by the EPP feature. We also argue that in this process, children are guided by economy principles. One of these is the principle of structural economy proposed in the Lexical Learning Hypothesis of Clahsen, Eisenbeiss & Penke (1996), originally from Safir (1993). Another economy principle, which is crucial for our analysis of the child data at hand, is a principle of economy of movement (see also Westergaard 2005). These principles will ensure that children do not build more structure or move elements higher in the structure than there is evidence for in the input. This means that movement operations should initially target positions that are as low as possible in the clause structure. This corresponds to what is often found in early child language: to the extent that children produce non-target forms, they normally seem to be due to children producing less movement than in the adult language, see e.g. Schaeffer (2000) on the lack of scrambling in Dutch, or Radford (1994) on the lack of inversion in some English-speaking children's w/i-questions. Superficially, this is of course the opposite of what we see in the acquisition data presented in this paper, and in the next section we therefore consider this economy principle in more detail in relation to the Norwegian child data.

294

Marii Westergaard and Kristine Bentzen

5.2. Economy of movement As mentioned in Section 2, V2 word order is in place more or less immediately in those clause types that require it, and Norwegian children must therefore realize very early that their language requires some filled C° head. Still, the question could be asked whether early verb movement indeed targets the same head positions as in the adult grammar. In the syntactic model adopted here, it is assumed that the verb in all Norwegian main clauses moves to the highest head, Int° in w/z-questions, Pol0 in >>es/«o-questions, and Top0 in all declaratives, subject-initial as well as non-subject-initial clauses. According to the approach to language acquisition we are assuming, UG provides children with the knowledge that all main clauses have a CP domain and specifies the head that is necessary to produce different clause types, e.g. Int° for w/z-questions or Pol0 for yes/no-quesuons. Thus, it is not unlikely that early verb movement in questions and non-subjectinitial declaratives is in fact movement to the appropriate heads. In subjectinitial main clauses, on the other hand, this is not immediately obvious. Since subjects in the world's languages are not universally in the specifier position of the highest C-head (e.g. not in English), UG will not provide Norwegian children with the information that subjects are default topics in this language and move to SpecTopP. This must therefore be learned from input. Likewise, that the verb is attracted by the [X°EPP] head feature on Top0 and moves to the head of this functional projection must also be learned. Unfortunately, there is no clear evidence in subject-initial declarative main clauses that the verb (and accordingly also the subject) moves all the way to the TopP in Norwegian. Nevertheless, there should be ample evidence in the input that there is verb movement in these sentences, as illustrated by the relatively high frequency of main clauses with negation in the sample of child-directed speech investigated above, obviously displaying the target V-Neg word order (see Table 2). There were 43 examples in the speech sample, making up 6.4% of the total (43 out of 668 clauses). However, if we assume that children only focus on the relevant clause type when searching for cues, as the Split-CP model indicates (see Westergaard 2005, 2006), then the evidence for verb movement is much higher, in fact as much as 35%, as the relevant figure to relate this to is the total number of subject-initial declaratives in the sample, which is 123. In any case, children apparently realize very early that finite verbs move across negation and adverbs in subject-initial main clauses, and they produce target-

The (non-)effect

of input frequency on the acquisition of word order

295

consistent forms from the onset of the appearance of relevant constructions, as illustrated in (49). (49) cegjorikkje. I do not 'I'm not doing (it).'

(Ina.09, 2;2.12)

However, there are many other head positions than Top0 that could serve as the landing site for verb movement in such sentences. Given that the economy principle discussed above ensures that children do not move constituents any farther than they have to, we argue that they initially pick a lower head as the target for verb movement in these constructions. In this model, there are two functional projections in the IP-domain of the clause, the In(ner)Top(ic)P and T(ense)P, as was illustrated in (48) above. Sentence adverbs (including negation) normally occur adjoined to TP in the clause structure, i.e. in a position between the InTopP and TP. Thus, in order for the verb to appear in front of negation in sentence (49), it will minimally have to move to the head of the InTopP, the highest functional projection in the IP domain. Since the InTopP is the lowest possible projection in the clause structure that will ensure that the verb precedes negation, this is also in accordance with the children's shortest move approach. In other words, the children initially assume that there is verb movement to the head InTop0 in Norwegian. Note that this corresponds to V°-to-I° movement in traditional terminology, i.e. the children are in fact mis-setting a parameter, assuming that there is general verb movement to the IP domain in Norwegian. Verb movement to the InTop0 head will result in the surface word order V-Neg, which corresponds to what is found in the adult language. However, the syntactic representation of a sentence like (49) differs from that of the target grammar, in that the child version of the sentence is a bare FinP with verb movement to InTop0, as illustrated in the partial structure in (50). The corresponding adult structure is a full TopP, as illustrated in (51), where the [X°Epp] feature attracts the verb to the head position of this projection.

296

(50)

Marit Westergaard and Kristine Bentzen

FinP

(Child structure) InTopP

ce\ I

InTop'

TP

gJ0r-} do

TP

ikkje not tt

(51)

TopP

Top0 gj0r do

(Adult structure)

FinP

If this analysis of children's main clause declaratives is correct, we would expect to see verb movement to the InTopP also in embedded contexts. In fact, we expect to see the verb in front of negation and adverbs in all clauses where the verb does not move to the CP domain. This means that children should not only produce the non-target-consistent word order in all kinds of embedded clauses, but also in the non-V2 main w/z-questions. This is, of course, exactly what we saw above in the child data presented in Section 3, as illustrated in e.g. examples (27) and (34), repeated here for convenience. In all other clause types (questions and non-subject-initial declaratives), the verb moves to a head in the CP domain, and this will mask the V-to-I movement that the children seem to be assuming for their language.

The (non-)effect of input frequency on the acquisition of word order

(21')cevet at cehar ikke gjort det. I know that I have not done it know that I haven't done it.' Preferred: ^E1 vet at ce ikke har gjort det. (34') kem som var ikke belt i form? who that was not completely in shape 'Who wasn't feeling too well?' Target form: Kem som ikke var helt iform?

297

(Henning 4;8.13)

(Henning 4;5.0)

An additional example from the Tromse corpus is illustrated in (52). This sentence is a non-subject-initial declarative, where the verb has failed to move across the subject to Top0. Thus, this sentence displays non-targetconsistent word order, as the adult grammar requires V2 (see Westergaard 2004). As the children produce V2 constructions from the onset of multiword utterances, there are very few such cases attested in the child data, and only one which includes negation. Note that the verb in sentence (52) does occur to the left of negation, suggesting that verb movement has in fact taken place. However, note that the verb does not move to the position above the subject, which would be expected in the target grammar. The word order in this example indicates that the verb has moved, and in accordance with the argumentation presented here, it has moved to the head of the InTopP. (52) [/] og sä du kan ikke tegne mer sann. and s+.... and so you can not draw more such 'And then you can't draw more like that.' Target form: Og sä kan du ikke tegne mer sann.

(Ann. 17, 2;8.4)

Thus, the children's overgeneralization patterns in embedded clauses (and occasional examples from main clauses) provide some support for the analysis that initially in Norwegian child language there is verb movement to a lower head than Top0 in main clause declaratives. This means that the children's choice of an uneconomic word order pattern in embedded clauses (involving verb movement) is actually caused by an economy principle operative in main clauses, viz. the principle of economy of movement. But why don't we find any overgeneralization of V2 word order in embedded questions? This was illustrated by sentences such as (24) above, repeated here, where the verb correctly appears following the subject, unlike the word order in main clause questions.

298

Marit Westergaard and Kris tine Bentzen

(24') Ann vet ikkekor han erkenne. Ann knows not where he is LOG 'Ann doesn't know where he is.'

(Ann.09, 2;2.19)

The reason for this lack of word order overgeneralization, we argue, is due to the functional architecture of the Split-CP model provided by UG. Recall that main clause questions are IntPs, while embedded questions are bare WhPs, lacking interrogative force. Being endowed with this knowledge, children know that embedded questions are not real questions and consequently do not project an IntP in these cases. Overgeneralizing the [X°Epp] head feature on the Int° head to embedded clauses is therefore not possible, simply because that functional head is not present in this context.9 On the other hand, according to the account given for children's error patterns above, they will of course be expected to "transfer" verb movement to the InTop0 head also in embedded questions. The prediction is that although young children will not move the verb across the subject in an embedded question, they should in fact overgeneralize verb movement across negation. Thus, an ungrammatical sentence such as (53) should be unattested in child data, while non-target forms such as the hypothetical sentence illustrated in (54) are predicted to occur. (53) *^ vet ka vil han g)0re. I know what will he do (54) *yE vet ka han vil ikkje gj0re. I know what he will not do The first part of this prediction is generally borne out in both the Troms0 corpus of younger children, as well as in the diary notes and recordings of the older children. As for the second part of this prediction, the results from the small experiment described in Section 3 suggest that children at least up to the age of 6 overgeneralize verb movement to embedded w/z-questions as well, moving the verb past negation and adverbs, as illustrated in e.g. (41), repeated here. (4Γ) huske du koffer han Karsten var ikke i barnehagen? remember you why DET Karsten was not in kindergarten.the 'Do you remember why Karsten wasn't in the kindergarten?' Target form: Huske du koffer han Karsten ikke var i barnehagen?

The (non-)effect of input frequency on the acquisition of word order

299

5.3. The way to the target grammar But if children have mis-set a parameter, how can they reach the target grammar? We argue that in order for children to reset the V-to-I parameter and revise their initial hypothesis, they need to pay attention to the word order in sentences that do not display V2, i.e. all embedded contexts and non-V2 main w/z-questions. Note that this is different from the degree-0 learnability of e.g. Lightfoot (1999), which argues that children can only detect cues in unembedded contexts. Within the Split-CP model that we are assuming, where main and embedded clauses have different functional architecture, children must pay attention to the word order of relevant clause types separately, in order to acquire the status of the EPP head feature with respect to the individual syntactic heads in the CP domain. We also believe that in order for children to be able to distinguish between Norwegian and V2 languages which do display V-to-I movement, e.g. Icelandic, they will have to be sensitive to embedded word order. Icelandic in fact displays exactly the word order in embedded clauses that the children in our study produce in Norwegian, and as far as we can tell, there is no difference between Norwegian and Icelandic main clauses that will indicate to children which type of language they are learning. Thus, we argue that the cue that a V2 language also has V-to-I must be found in non-V2 contexts, i.e. generally in embedded clauses. Embedded clauses are naturally more complex structures than main clauses, and searching for cues in these contexts is arguably more difficult than finding cues in main clauses. This could be one reason why the nontarget-consistent word order is so persistent in children's production, possibly lasting beyond the age of six, as indicated by the results of our small experiment. Compared to the extremely early acquisition of word order in general, in Norwegian as well as in other languages, the target-consistent word order in embedded clauses indeed falls into place very late. Here frequency may also play a role. Recall that it is not sufficient for Norwegian children to pay attention to just any embedded clause; it must also contain negation or an adverb, otherwise the word order will be identical to that of main clauses. And as we saw in Tables 2 and 3 in Section 4, these clause types are extremely infrequent in the input, attested only between 0.5% and 1% in the samples of child-directed speech that we investigated. Thus, we would argue that the lack of input frequency does have an effect in this case, viz. the effect that it takes a considerable time for children to revise their initial hypothesis that Norwegian has V-to-I movement.

300

Marit Westergaard and Kristine Bentzen

However, their initial hypothesis is not directly caused by the lack of frequency, but rather the principle of economy of movement, as we argued above. One piece of evidence that may support the idea that there is a frequency effect here is the fact that this type of overgeneralization is generally not found in German child language, as we saw in Section 3.1. Being an SOV language with verb movement also across objects and adjuncts in main clauses, German will provide considerably more input evidence to children that embedded clauses are different from main clauses. That is, the difference will not only be visible in embedded clauses containing negation or sentence adverbs (which are generally infrequent in child-directed speech), but also in all embedded clauses containing an object or an adjunct. Without having performed a study on German child-directed speech to this end, we think we can safely assume that a German-speaking child will be exposed to the relevant contexts for non-V2 considerably more often than Norwegian-speaking children. We may also compare this to another non-target-consistent word order pattern produced by young Norwegian children in so-called 'subject-shift' constructions, where the target language requires pronominal subjects to appear preceding negation in questions and non-subject-initial declaratives. In Westergaard (2005) it is shown that the three children in the study (age approximately 1;9 to 3;3) all initially produce pronominal subjects in a lower position, following negation. In Westergaard (forthcoming) this is argued to be due to the same economy principle that is discussed in the present article, as well as the general complexity of the construction. However, in this case the children's error pattern is relatively short-lived, as the target-consistent word order falls into place between age 2;6 and 3;0. The difference between the subject-shift constructions and the embedded contexts discussed here may partly be due to different input frequencies. In the same sample of child-directed speech which was investigated in Table 2, evidence for word order in the subject-shift constructions is attested in 4.2% of the total input (28 out of 668) and in 8.3% of all relevant contexts, i.e. questions and non-subject-initial declaratives. This is, of course, considerably more than the 1.0% evidence for word order in non-V2 contexts, and input frequency may therefore be argued to play a certain role here.

The (non-)effect of input frequency on the acquisition of word order

301

6. Summary and concluding remarks In this paper, we have considered two similar constructions in Norwegian child language, embedded questions and (all) embedded clauses containing negation or an adverb. In the former clause type the children's word order is error-free from the beginning, in that they do not overgeneralize verb movement across the subject from main clause questions. In the latter clause type, on the other hand, children produce non-target-consistent word order for a considerable period of time, possibly beyond the age of six. That is, they move the verb across negation or an adverb, and this is a word order which also appears in other non-V2 constructions. An investigation of some samples of child-directed speech revealed that both constructions are extremely infrequent in the input, and that the difference between the two is too small to have such a considerable effect on the children's production. A possible account of the error pattern as a result of input frequencies only was therefore rejected. Instead, within a weak continuity/structure-building account to language acquisition, we explored an analysis which assumes an economy principle of movement, which generally says that children will not move elements higher in the clause structure than there is evidence for in the input. More specifically, we argued that Norwegian children's early subject-initial main clauses display verb movement to a lower functional head than in the target grammar, i.e. to a head in the IP domain. This will ensure target-consistent word order in main clauses (V-Neg/Adv), but result in non-targetconsistent word order in non-V2 contexts: embedded clauses and non-V2 main w/z-questions. On the other hand, the reason why there is no overgeneralization of V2 from main to embedded questions is related to the Split-CP model of clause structure that we assume, where different clause types have different functional heads in the CP domain. While main w/z-questions have an Int° head, embedded questions are bare WhPs, reflecting the fact that they have no interrogative force. Thus, the different functional architecture for the two clause types accounts for the lack of overgeneralization, as a feature value on the Int° head cannot be transferred to a clause type where this head is not present. However, input frequency is also argued to play a role in this analysis. Together with the general complexity of the relevant constructions, the lack of input frequency may be a reason why the non-target-consistent word order produced by the Norwegian children is so persistent, compared to word order in other constructions.

302

Marit Westergaard and Kristine Bentzen

Thus, we argue that there may certainly be effects of input frequencies in language acquisition, but we doubt that input frequency alone can account for acquisition orders and children's non-target-consistent production. Rather, we believe that explanations must be sought in a variety of areas. In the particular case discussed in the present paper, we have argued that economy as well as complexity interact with frequency to produce the particular error patterns found in the child data.

Notes 1. Garbacz (2005) searched the Big Brother corpus of spoken Norwegian (Big Brother-korpuset), and according to his findings, the order S-Neg-Vfin is by far the most frequent word order in thai-clauses, constituting 64% of embedded clauses with negation. The second most common word order is S-Vfin-Neg, occurring 29% of the time, whereas Neg-S-Vfm is the most infrequent pattern (7%). He also ran a similar search in the Oslo corpus of Standard (written) Norwegian, Bokmäl (Oslo-korpuset av taggede norske tekster). In this corpus, the preference for the S-Neg-Vf]n order was even more significant. As much as 96% of the sentences had this word order, whereas the other two orders were each used only about 2% of the time. 2. Apart from ten files that have been recorded and transcribed by the first author, the corpus has been collected by Merete Anderssen. See Anderssen (2006) and Westergaard (2005) for details about the corpus. 3. In the dialect the children in this study are acquiring, the present tense endig -er of the standard language has been reduced to -e, which means that the infinitive and the present tense verb forms of many verbs are identical. This is the case for the two classes of regular verbs, which make up approximately 96% of all verbs in the language, according to Endresen & Simonsen (2001). 4. A relatively standard assumption is that negation and adverbs may be adjoined to VP as well as to TP, but not AgrSP. It is, however, possible that certain light adverbs and negation may be adjoined even higher, as suggested in Holmberg (1993). In sentences such as (17) in child language, it could either be the case that negation is adjoined to this higher position, or alternatively, that the child has failed to move the subject (see Westergaard 2005). 5. The reply from the investigator is provided in (i). It should be noted, however, that pragmatically, also a ^es/no-question could call for such an answer. (i) det [/] det kalles for hyena. (INV, File Ina.27) it it is-called for hyena 'It is called a hyena.'

The (non-)effect of input frequency on the acquisition of word order

303

It should be noted that the number of embedded questions may be somewhat inflated as an effect of the recordings, especially in the speech of the parents. In an attempt to make the children speak as much as possible, they frequently produce utterances such as the following: (\)...har du fortalt ho Merete ka du gjorde i gär? (MOT, file Ann.03) have you told DET Merete what you did yesterday This abbreviation refers to the Extended Projection Principle (EPP) of earlier versions of generative theory (originally from Chomsky 1982), which ensured that all clauses have a subject. Within the Minimalist model, e.g. Chomsky (1995), an EPP feature on a syntactic head will require that this head projects a specifier in order for the uninterpretable EPP feature to be deleted. The optional word order in questions with the monosyllabic question words is accounted for by another C-head, the head of the Foc(us)P (see Westergaard & Vangsnes 2005; Westergaard 2005). Languages which do display V2 word order in embedded contexts, e.g. Belfast English, must then be assumed to have verb movement also to the head Wh°.

References Afarli, Tor 1986

Some syntactic structures in a dialect of Norwegian. University of Trondheim Working Papers in Linguistics 3: 93-111. Anderssen, Merete 2006 The Acquisition of Compositional Defmiteness in Norwegian. Ph.D. diss., University of Tromse. Benincä, Paola & Cecilia Poletto 2004 Topic, Focus and V2: Defining the CP sublayers. In The Structure of CP and IP: The Cartography of Syntactic Structures 2, Luigi Rizzi (ed.), 52-75. Oxford/New York: Oxford University Press. Bentzen, Kristine 2003 V-to-I movement in the absence of morphological cues: Evidence from Northern Norwegian. In Nordlyd 31.3: Proceedings from the 191 Scandinavian Conference of Linguistics, Anne Dahl, Peter Svenonius & Marit R. Westergaard (eds.), 573-588, University of Troms0. 2005 What's the better move? On verb placement in Standard and Northern Norwegian. Nordic Journal of Linguistics 28 (2): 153-188. 2007 The degree of verb movement in embedded clauses in three varieties of Norwegian. In Nordlyd 34: Scandinavian Dialect Syntax 2005, 127-146. University of Troms0. (Big Brother-korpuset. Available at http://www.tekstlab.uio.no/talespraak/bigbrother.)

304

Marit Westergaard and Kristine Bentzen

Chomsky, Noam 1982 Some Concepts and Consequences of the Theory of Government and Binding. Cambridge, MA and London: MIT Press. 1995 The Minimalist Program. Cambridge, MA and London: MIT Press. Clahsen, Harald & Klaus-Dirk Smolka 1986 Psycholinguistic evidence and the description of V2 phenomena in German. In Verb Second Phenomena in Germanic Languages, Hubert Haider & Martin Prinzhorn (eds.), 137-168. Dordrecht: Foris. Clahsen, Harald, Sonja Eisenbeiss & Martina Penke 1996 Lexical learning in early syntactic development. In Generative Perspectives on Language Acquisition: Empirical Findings, Theoretical Considerations and Crosslinguistic Comparison [Language Acquisition and Language Disorders 14], Harald Clahsen (ed.), 129-159. Amsterdam and Philadelphia: Benjamins. Döpke, Susanne 1998 Competing language structures: The acquisition of verb placement by bilingual German-English children. Journal of Child Language 25:555-584. Endresen, Rolf Theil & Hanne Gram Simonsen 2001 The Norwegian verb. In A Cognitive Approach to the Verb: Morphological and Constructional Perspectives, Hanne Gram Simonsen & Rolf Theil Endresen (eds.), 73-94. Berlin/New York: Mouton de Gruyter. Garbacz, Piotr 2005 Bisatsstrukturer som man kan inte bortse frän (om bisatsordfbljd i de fastlandsskandinaviska spraken) [Embedded clause structures which cannot be ignored (on embedded clause word order in the Mainland Scandinavian languages)]. MS, University of Lund. Gawlitzek-Maiwald, Ira, Rosemary Tracy & Agnes Fritzenschaft 1992 Language acquisition and competing linguistic representations: The child as arbiter. In The Acquisition of Verb Placement, Jürgen Meisel (ed.), 139-179. Dordrecht: Kluwer. Holmberg, Anders 1993 Two subject positions in Mainland Scandinavian. Working Papers in Scandinavian Syntax 52: 29-^41. Hakansson, Gisela & Sheila Dooley Collberg 1994 The preference for Modal + Neg. An L2 perspective applied to LI acquisition. Second Language Research 10(2): 95-124. Jordens, Peter 1990 The acquisition of verb placement in Dutch and German. Linguistics 28: 1407-1448. Josefsson, Gunlög 2004 Input and Output: Sentence patterns in Child and Adult Swedish. In The Acquisition of Swedish Grammar [Language Acquisition and

The (non-)effect of input frequency on the acquisition of word order

305

Language Disorders 33], Gunlög Josefsson, Christer Platzack & Gisela Häkansson (eds.), 95-133. Amsterdam and Philadelphia: Benjamins. Lightfoot, David 1999 The Development of Language: Acquisition, Change and Evolution. Maiden, MA and Oxford: Blackwell. Müller, Natascha 1996 V2 in first language acquisition. Early child grammars fall within the range of UG. Linguistics 34: 992-1028. (Oslo-korpuset av taggede norske tekster. Available at www.tekstlab.uio.no/norsk/bokmaal.) Penner, Zvi 1996 From empty to doubly-filled complementizers. A case study in the acquisition of subordination in Bernese Swiss German. Fachgruppe Sprachwissenschaft der Universität Konstanz. Arbeitspapier'Nr. 77. Platzack, Christer 1996 The initial hypothesis of syntax: A minimalist perspective on language acquisition and attrition. In Generative Perspectives on Language Acquisition: Empirical Findings, Theoretical Considerations and Crosslinguistic Comparison [Language Acquisition and Language Disorders 14], Harald Clahsen (ed.), 369^414. Amsterdam and Philadelphia: Benjamins. Poeppel, David & Kenneth Wexler 1993 The full competence hypothesis of clause structure in early German. Language 69: 365-424. Radford, Andrew 1994 The syntax of questions in child English. Journal of Child Language 21: 211-236. Rizzi, Luigi 1997 The fine structure of the left periphery. In Elements of Grammar: Handbook of Generative Syntax, Liliane Haegeman (ed.), 281-337. Dordrecht: Kluwer. 2001 On the position "Int(errogative)" in the left periphery of the clause. In Current Studies in Italian Syntax, Guglielmo Cinque & Giampolo Salvi (eds.), 287-296. Amsterdam: Elsevier. Roberts, Ian 1999 Verb movement and markedness. In Language Creation and Language Change: Creolization, Diachrony, and Development, Michel DeGraff (ed.), 287-327. Cambridge, MA: MIT Press. Safir, Ken 1993 Perception, selection, and structural economy. Natural Language Semantics 2: 47-70. Santelmann, Lynn 1995 The Acquisition of Verb Second Grammar in Child Swedish: Continuity of Universal Grammar in Wh-questions, Topicalizations and Verb Raising. Ph.D diss., Cornell University.

306

Marit Westergaard and Kristine Bentzen

Schaeffer, Jeanette 2000 The Acquisition of Direct Object Scrambling and Clitic Placement: Syntax and Pragmatics. Amsterdam and Philadelphia: Benjamins. Schönenberger, Manuela 2001 Embedded V-to-C in Child Grammar: The Acquisition of Verb Placement in Swiss German. (Studies in Theoretical Psycholinguistics.) Dordrecht: Kluwer. Theakston, Anna, Elena Lieven, Julian Pine & Caroline Rowland 2004 Semantic generality, input frequency and the acquisition of syntax. Journal of Child Language 31:61 -99. Tomasello, Michael 2003 Constructing a Language: A Usage-based Theory of Language Acquisition. Cambridge, MA: Harvard University Press. Vangsnes, 0ystein A. 2005 Microparameters for Norwegian w/z-grammars. Linguistic Variation Yearbook 5, Jeroen Van Craenenbroeck (ed.), 187-226. Amsterdam and Philadelphia: Benjamins. Vikner, Sten 1995 Verb Movement and Expletive Subjects in the Germanic Languages. New York: Oxford University Press. Westergaard, Marit R. 2003 Word Order in w/z-questions in a North Norwegian Dialect: Some Evidence from an acquisition study. Nordic Journal of Linguistics 26(1): 81-109. 2004 The Interaction of input and UG in the acquisition of verb movement in a dialect of Norwegian. Nordlyd 32.1: Troms0 Working Papers in Language Acquisition, 110-134. University of Troms0. 2005 The Development of Word Order in Norwegian Child Language: The Interaction of Input and Economy Principles in the Acquisition of V2. Ph.D diss., University of Troms0. 2006 Triggering V2: The amount of input needed for parameter setting in a Split-CP model of word order. In Language Acquisition and Development: Proceedings of GALA 2005, Adriana Belletti, Elisa Bennati, Cristiano Chesi, Elisa DiDomenico & Ida Ferrari (eds.), 658671. Cambridge: Cambridge Scholars Press. forthc. Verb movement and subject placement in the acquisition of word order: Pragmatics or Structural Economy? In First Language Acquisition of Morphology and Syntax: Perspectives across languages and learners. [Language Acquisition and Language Disorders], Pedro Guijarro-Fuentes, Pilar Larranaga & John Clibbens (eds.). Amsterdam and Philadelphia: Benjamins. Westergaard, Marit R. & Oystein A. Vangsnes 2005 WTz-questions, V2, and the left periphery of three Norwegian dialects. Journal of Comparative Germanic Linguistics 8: 117—158.

Factors determining the acquisition of animacy in Czech Denisa Bordag

The present study explores the role of the following factors in animacy acquisition by Czech children: structural aspects of plural formation, referentiality, and, in particular, frequency. Children and adults were asked to produce plural forms of masculine pseudo-nouns, which are distinctly marked for (in)animacy in Czech. The results reveal that in contrast to adults, who mark (in)animacy unambiguously (animacy over frequency), children are affected in their choice of inflectional endings by the frequent, unmarked [i]-plural irrespective of animacy (frequency over animacy). The data complies with the view that input frequency and distribution of particular formation patterns determines acquisition, interacting with other factors such as transparency of form-function mapping.

1. Introduction Animacy is an important category in natural language processing. It characterizes the tendency of human languages to expand the differentiation of grammatical and lexical forms associated with human or otherwise "animate" entities from their less animate counterparts. The animate vs. inanimate distinction plays a role in languages where it surfaces overtly as well as in those where it does not. Corpus studies clearly show that in transitive sentences, animate nouns occur more frequently in subject position and inanimate nouns in object position. Whereas in some languages the constraint that subjects should be higher than objects in animacy is grammaticalized, in other languages it presents itself as a statistical tendency (Dahl & Fraurud 1996, for Swedish). Consequently, at least two lines of research exploring animacy can be traced in the present literature. The first one focuses on animacy as a conceptual notion and its consequences for sentence construction, especially word order (the so-called "animate first" tendency). The other line of re-

308

Denisa Bordag

search is concerned with animacy as a grammatical category, sometimes referred to as a subgender (Corbett 1991; Stankiewicz 1968). In many languages,1 including most of those from the Slavic family, the animate/inanimate distinction is reflected in the declension of some noun classes. Assigning nouns to the animate or inanimate subgender is usually based on a straightforward semantic criterion: "animates" generally include humans, animals and immortal/imaginary beings that are human or animal-like, whereas "inanimates" denote non-living objects or plants. Historically, subgenders have developed more recently than the three primary genders (masculine, feminine, neuter), where semantic assignment accounts for only some nouns. Despite the fact that animacy has become a topical research question in recent years, its acquisition in Slavic languages has to date been largely neglected. This study attempts to characterize the difficulties children have to overcome in acquiring animacy as a grammatical category in the highly inflectional Czech language. It will be shown that although animacy in Czech is, to a large extent, a referential category (therefore enabling the successful application of the semantic criterion to distinguish between animate and inanimate nouns), other factors such as the formal expression of the animate/inanimate distinction and, in particular, frequency drive its acquisition. It appears that, in accordance with other recent data (Theakston, Lieven, Pine & Rowland 2004), input frequency and distribution of particular formation patterns determines acquisition and that children need to recognize these regularities to begin applying more abstract semantic generalization. Czech is a western Slavic language with a very rich inflectional system. It has fourteen declensional classes, seven cases and three grammatical genders: masculine, feminine and neuter. The masculine gender is subdivided into animate (Ma) and inanimate (Mi), based solely on the evidence of accusative and nominative plural. The two other genders, feminine and neuter, do not manifest a grammatical distinction between animate and inanimate.

Factors determining the acquisition ofanimacy in Czech

309

Table 1. Declensional classes (DC) in Czech after Smilauer (1971) Declensional class

Plural

Pl.End.

pan 'Mr.' muz 'man' pfedseda 'chairman' soudce 'judge' hrad 'castle' stroj 'machine' zena 'woman' nize 'rose' pisen 'song' kost 'bone' mesto 'town' more 'sea' kufe 'chick' staveni 'building'

pan-i/-ove muz-i/-ove pfedsed-ove soudc-i/-ove hrad-y stroj-e zen-y ruz-e pisn-e kost-i mest-a mofe-e kuf-at-a staven-i

-[i]+(A)/-ove -\\\l-ove -ove -\\\l-ove -[i] -e -[i] -e -iH-(A) -[i] -a -e -a -i

Gender Ma Ma Ma Ma Mi Mi F F F F N N N N

%*

8.92 3.87 1.1 0.39 25.13 2.96 21.92 6.21 2.24 7.02 5.61 0.73 0.36 11.62

*Proportion of nouns from a particular declensional class

Table 1 presents an overview of Czech declension. Each declensional class is represented by one traditional example word in the nominative that is prototypical for that declension. Masculine nouns, which are central to this study, fall into six main declensional classes. This paper focuses on the four most frequent classes: pan, muz, hrad and stroj. The first two are animate (Ma), the last two inanimate (Mi). They all have a typical masculine termination, which is a consonant in Czech. Animate nouns ending with a hard consonant belong to the declensional class pan, animate nouns ending with a soft (or palatalized) consonant to the declensional class muz. Inanimate nouns ending with a hard consonant are from the declensional class hrad, inanimate nouns ending with a soft consonant from the class stroj. Since animacy status is determined by the nature of the referent and the noun's termination is obvious from the word form, any Czech masculine pseudonoun can be reliably assigned to one of the four declensional classes based on these two criteria. The exceptions are masculine nouns ending with a neutral consonant. They cannot be assigned to a declensional class based on phonological criteria, but rather on the basis of morphological criteria.2 Only two such items were included in our experiments.

310 Denis a Bor dag

The second column in Table 1 lists nominative plural endings, which were the forms elicited in the experiments. They were chosen because they reliably indicate the animate/inanimate distinction and because previous research on child language production (see below) shows that these forms can be successfully elicited with very young children. Nominative plural of the animate classes can usually be formed in two ways: either with the ending -ove (or -e), which often signals "personality", or with the ending -/'. However, if the ending -/ is used with a noun from the declensional class pan, which ends with a hard consonant, a stem alternation must be performed to transform the hard consonant into its soft counterpart, e.g. hroch — hrosi ('hippo' - 'hippos'), kluk - kluci ('boy' - 'boys'), vrah vrazi ('murderer' - 'murderers'), etc. In Czech, as well as in other West Slavic languages, an alternation signals animacy (Belicovä 1998). A closer look at the table reveals that the ending -[i] has a special status in Czech plural declension. First, it is not restricted to a single declensional class or even gender. It appears in feminine as well as masculine animate and inanimate nouns. It must be stressed here that the difference between -i and -y in Czech, as opposed to e.g. Russian, is only orthographic. In speech, both letters are pronounced the same. Since the experiments were performed orally and it was assumed that children were unfamiliar with the different orthographic realizations of the sound -[i], the effects of orthography on the results can be excluded. Second, apart from being general in the sense described above, the ending -[i] is also very frequent. Not only does it appear in the two most frequent declensional classes hrad and zena, but it also covers more than 60% of Czech nominative plurals. The ending -[i] is thus the most frequent and probably the unmarked plural ending in Czech. This conclusion is further supported by data from acquisition of Czech as a second language, which shows that L2 Czech learners perceive this ending as typically plural (Bordag 2006). Table 2 summarizes the frequencies of Czech cases. It shows that at approximately 6%, nominative plural belongs to the less frequent Czech cases and that in plural cases genitive is more frequent. Thus, children are not highly exposed to nominative plural.

Factors determining the acquisition ofanimacy in Czech

311

Table 2. Frequencies of Czech cases (cf. Smilauer 1972)

Nominative Genitive Dative Accusative Locative Instrumental Vocative

Singular %*

Rank

Plural %*

Rank

22.52 16.92 3.43 15.01 8.52 7.66 1.76

1 2 9 3 4 5 12

5.94 7.56 1.03 4.87 2.37 2.22 0.19

7 6 13 8 10 11 14

Proportion of each case in written Czech (other data not available for Czech).

2. The study In this study, we applied Berko's design (Berko 1958) to research on plural formation by Czech children, focusing on their ability to produce animate and inanimate nominative plural forms and to distinguish between them appropriately. In her original experiments, Berko used drawings of Martian-like beings and objects to elicit plural forms of English pseudo-words. She argues that if a subject can supply the plural ending of a real word in a given language, he or she may simply have memorized the proper form. If, however, the subject can supply the correct plural of a pseudo-noun, he or she has an abstract representation of the category. Berko was later followed by other researchers who applied her design to elicit particular morphological markings in different languages (e.g. Kopeke 1987, 1988, 1998). The experiments revealed that (for example) German3 children use plural forms in plural contexts in 90% of cases at the beginning of their third year (Park 1978). However, they produce a greater number of incorrect plural endings than correct ones. Plural morphology is reliably acquired only by children as old as five years. Three- to four-year-old children still make errors when choosing a proper plural ending in 50% of cases (Schaner-Wolles 1988). Generally, the results for real words are always better than for pseudowords with subjects of all ages (Walter 1975). However, Ramge (1975) observed that children often use correct noun plurals in their spontaneous speech, while producing incorrect plurals for identical items in experimental situations. The most frequent type of error in experimental conditions is usage of a singular form in a plural context. This is true for real nouns (Schaner-Wolles 1988) and pseudo-words (Walter 1975). This type of error

312 Denisa Bordag

also appears in spontaneous speech (Park 1978). The number of zero plural forms that are produced with pseudo-nouns can even surpass 70% (Mugdan 1977), and such performance can be observed with children up to six years (Walter 1975). However, repeating the pseudo-noun form that subjects heard from the experimenter in the singular condition does not necessarily imply that children produce a singular form in a plural context. As argued by Kopeke (1998), it may simply mean that subjects interpret the intended singular form as a plural form, perhaps based on its phonological properties. Therefore, the term "repetition" will be used in cases where subjects repeated the stimulus in the present study.

2.1. Experimental design To prepare the materials for the experiments reported in this study, a number of pseudo-nouns were created, observing the phonological rules of Czech. Pictures representing the pseudo-words were computer-animated and printed on white sheets of paper. They depicted either inanimate objects or cartoonlike beings with typical features of living creatures (eyes, nose, mouth, hands, feet). Several real words were included as well. A text, omitting the target form, was typed on each sheet for the experimenter to read out: below the picture of a single item, the sentence Tohlejejeden/jedna/jedno... 'This is one ...' appeared. Note that the numeral One' in Czech is marked for gender (jeden - masculine,y'ei/A2ö - feminine, jedno - neuter), so that the gender of the item was predefined by this sentence. Below the item and the sentence, three exemplars of the same item were printed with an incomplete sentence: Tohlejsou t f i . . . 'These are three ...'; see Figure 1.

To je jeden barach. This is one barach'

To jsou tfi 'These are three...' Figure 1. Example of animate items

Factors determining the acquisition ofanimacy in Czech

313

Unlike in other similar experiments, the numeral 'three' instead of 'two' was used, because it has one invariant form for the three genders in Czech. The numeral 'two' has two different forms in Czech, one for masculine nouns (dva) and one for feminine and neuter nouns (dve). The adult subjects were tested on 63 pseudo-words and their experimental session took about 30 minutes. Children were tested on a reduced selection of items with 30 pseudo-words. Their session lasted approximately 1520 minutes. Pseudo-nouns of all three genders as well as several real nouns were included in the experiment. The focus of this study is on 18 pseudonouns that appeared both in the adult and in the child item set. They were all masculine pseudo-nouns; nine of them were supposed to elicit animate and nine inanimate plurals (see Table 3). Eight of them ended with a soft consonant, eight with a hard consonant, and two with a neutral consonant. Ten items were monosyllabic, eight disyllabic. The items from the animate and inanimate group differed only in the first letter. Table 3. The 18 target items of this study

Animate Inanimate

Hard consonant termination

Soft consonant termination

Ambg

barach tolek bak vouch varach molek mak douch

butoc rapez väj mutoc knapez kaj

bam lam

dos bos

All pseudo-nouns which were to elicit animate plurals were combined with pictures of non-existing living beings (Figure 1). The pseudo-nouns which were to elicit inanimate plurals were combined with pictures of nonexisting objects (Figure 2).

Figure 2. Example of an inanimate item

314 Denlsa Bordag

Remember that for Czech masculine nouns, the information about their animacy/inanimacy status and the type of their termination (hard vs. soft) is sufficient for their unambiguous assignment to one particular declensional class, which determines the plural ending.

2.2. Participants Fifteen adults and 40 children were tested in the experiment. Three children were excluded from the analyses because they failed to complete the experiment. The adult subjects, all of them university students, served as a control group. They received a booklet with all the experimental pictures and were asked to read the first sentence and then complete the second one. To draw their attention to the pictures, they were told that after this session they would receive a list of pictures and would have to mark those that appeared in the booklet. This task however did not take place. Subjects were paid for their participation. The children were tested at a preschool in Prague and ranged between 3;2 and 6;9 years in age. They were divided into two age groups: the younger children 3;2-4;10, the older children 5;l-6;9. Groups were approximately the same size, including 18 and 19 children, respectively, each group covering a period of 20 months with a gap of three months (4; 105;1) between them. Each child was tested individually. The sessions started with a short story designed to motivate the children to participate in the task. They were promised sweets and little toys if they did their best. If the attention of a child began to decline, he or she received a few sweets during the experiment. All children who participated were rewarded by the experimenter.

3.

Results

Table 4 presents results for all three groups of subjects for 16 masculine pseudo-nouns which ended with a hard or a soft consonant. The two items with a neutral termination have plural forms that are partially ambiguous with respect to animacy and will therefore be discussed later.

Factors determining the acquisition ofanimacy in Czech

315

Table 4. Results for masculine pseudo-nouns with hard and soft consonant endings

Non-living Adults (15x8= 120) 5;l-6;9(19x8=152) 3;2^;10 (18x8=144) Living Adults 5;l-6;9 3;2^;10

Inanimate plural Animate plural [i] or -e [i], [i]+A, -ove

Repetition

Other

56 .7% (68) 45 .1% (70) 43 .8% (63)

41 .7% (50) 48 .0% (73) 38 .9% (56)

0 4.6% (7) 11.8% (17)

1. 7% (2) 1. 3 % (2) 5.6 % (8)

13 .3% (16) 40 .8% (62) 42 .4% (61)

85 .0% (102) 57 .2% (87) 43 .8% (63)

0 1.3% (2) 10.4% (15)

1. 7 % (2) 0. 7 %(1) 3. 5 %(5)

The upper part of the table presents the results of the items depicting inanimate objects, the lower part the results of the items depicting animate beings. The vertical segmentation shows the number of actual inanimate, animate, repetition and other plurals the subjects produced. To have a better basis for comparisons, percentages give the proportion of a particular plural form used by each group of subjects. Repetitions are forms produced by the subjects as a completion of the plural sentence which were, however, identical with the forms that subjects heard in the singular (e.g. molek - molek). As can be observed, the number of such responses decreases rapidly with age and adults produce virtually no such forms in the plural context. This finding is also reported by other authors (e.g. Berko 1958). Surprisingly, the number of such responses in the present study is relatively low (only about 10% by the younger children) compared to e.g. Mugdan's (1977) study (over 70%). This might be due to the general complexity of the Czech inflectional system. The research of Mills (1986) on German gender acquisition demonstrates that the more ample the evidence for inflectional marking, the faster the acquisition (e.g. gender marking in German vs. English, where it is obvious only from personal pronouns). As "other" plurals we refer to forms whose stem was either substantially changed by the subjects (i.e., a different pseudo-noun than the target one was produced), or whose ending completely deviated from the expected one (very rare). These responses were produced mostly by the younger children, but it was generally a very infrequent type of error. If subjects are sensitive to animacy, we would predict that they produce predominantly inanimate plurals with non-living items and predominantly

316 Denisa Bordag

animate plurals with living items. This is precisely the pattern of results that was observed with the adult subjects. They clearly preferred animate plurals with living beings (85% to 13.3%), and the preference for inanimate plurals with non-living objects is also distinct, despite a less striking difference (56.7% to 41.7%). The results of the adult subjects show that the experimental paradigm was employed successfully and that it can reveal subjects' sensitivity to the factor of animacy. The high number of animate plurals with non-living items is a rather unexpected result, but it can be explained if the unmarked status of animate forms is taken into account. As pointed out by Stankiewicz (1968), within the category of animacy, the inanimate subgender opposes the animate as marked vs. unmarked, since the inanimate cannot designate animate beings, whereas the animate gender encompasses both animate and inanimate entities. Typically, these are the names of mushrooms, playing cards, dances, drinks or cars. Whereas a different pattern of results for living and non-living items was observed with adults, the child data showed a similar plural formation pattern for the items from both groups. In fact, the children demonstrated no clear preference for animate and inanimate plurals for items of corresponding groups, producing approximately identical numbers of animate and inanimate plurals irrespective of an item's animacy status. However, the adult-like pattern of results already starts to emerge in the production of older children: they seem to associate animate plurals with living beings (57.2% to 40.8%), though this preference is less strong than with the adult subjects (85.0% to 13.3%). Based on these data, we could conclude that although older children start to perform in a manner similar to adults, in most cases children randomly produce plural endings irrespective of the animacy factor. However, a detailed data analysis yields more insightful conclusions. 3.1. Results for items with a soft consonant termination In the next step, the data was analyzed according to the noun termination. Table 5 presents the results for the eight items with a soft consonant termination. Four of them, mutoc, knapez, kaj, bos, were designed to elicit inanimate plurals (declensional class stroj and plural ending -e), the other four, butoc, rapez, vaj, dos, animate plurals (declensional class muz and plural ending [i] (spelled -/) or -ove).

Factors determining the acquisition ofanimacy in Czech

317

Table 5. Results for masculine pseudo-nouns ending with a soft consonant Inanimate plural Animate plural Repetition Non-living Adults (15x4=60) 5; l-6;9( 19x4=76) 3 \1-4; 10 (18x4=72) Living Adults 5;l-6;9 3;2^;10

Other

60 .0% (36) 28 .9% (22) 20 .8% (15)

38 .3%• (23) 63 2 0/ /c' (48) 56 .9%' (41)

1.7% (D 0 1 .3% (D 6.6 °/Ό (5) 0 15.3 /'•(Π) 6.9% (5)

15.0% (9) 19 .7% (15) 15 .3% (Π)

83 3 0/ /C• (50) 78 .9%• (60) 72 .2 /c' (52)

0 0 11. 1 0 /0(8)

1.7% Ο) 1.3% (1) 4.2% (3)

Whereas the results of the adult subjects follow the same pattern as in Table 4, i.e., preference of animate plurals with living items and inanimate plurals with non-living items (here the tendency is even stronger than in the previous case), the results of children are strikingly different. Children prefer animate plurals with all items ending in a soft consonant, regardless of their animacy status. However, this tendency is stronger with living items. Before we discuss these data, we turn to the results of the items ending in a hard consonant. 3.2. Results for items with a hard consonant termination Table 6 summarizes the results for the eight items with a hard consonant termination. Four of them, varach, molek, mak, douch, were meant to elicit inanimate plurals (declensional class hrad and plural ending -[i] (spelled -y)). Their expected plural forms are varachy, mol(e)ky,4 maky, and douchy. The other four items, barach, tolek, bak, vouch, were designed to elicit animate plurals (declensional class pan and plural ending -[i] (spelled -/) or -ove). However, if the ending is -[i], the hard consonant requires palatalization into its soft counterpart. Thus, the expected plural forms are: barasi or barachove, told or tolkove,5 bad or bakove, vousi or vouchove. The adult subjects again show the same preferences. The pattern of results for the child subjects is, however, notably opposite to what was observed in Table 5. This time, children preferred inanimate plurals for all items, regardless of their animacy status. Moreover, we can also clearly observe a tendency as suggested by the previous table: the number of animate

318 Denisa Bor dag

plurals increases with age, again irrespective of animacy. This indicates that the unmarked status of "animates" was recognized. Table 6. Results for masculine pseudo-nouns ending with a hard consonant Inanimate plural Animate plural Repetition Non-living Adults (15x4=60) 5; l-6;9( 19x4=76) 3 ;2-4;10 (18x4=72) Living Adults 5;l-6;9 3;2-4;10

Other

53.3% (32) 63.2% (48) 66.7% (48)

45.0% (27) 32.9% (25) 20.8% (15)

0 2.6% (2) 8.3% (6)

1.7% (1) 1.3% (1) 4.5% (3)

11.7% (7) 61.8% (47) 69.4% (50)

86.7% (52) 35.5% (27) 18.1% (13)

0 2.6% (2) 9.7% (7)

1.7% (1) 0 2.8% (2)

3.3. The impact of frequency Having closely scrutinized the data, we can conclude that the plurals formed by children are not randomly produced as it might have appeared from Table 4. However, they are not driven by the same principles as those of the adult subjects either. Adults were clearly sensitive to the factor of animacy: they preferred animate forms for living items and inanimate forms for non-living items. Children, on the other hand, based their preference of animate or inanimate forms on the pseudo-noun's termination, or alternatively its declensional class. But does this provide a complete picture? Considering the results presented in Table 5 and Table 6, our attention should be drawn to plural formation. The animate plural of nouns with a soft termination, which was preferred by children, is formed with the ending -[i]. Crucially, the inanimate plural of nouns with a hard termination, which was preferred by the children in this case, is also formed with -[i]. In both cases the ending is attached to the stem agglutinatively. Obviously, children systematically prefer forming plurals with the ending -[i]. The special status of this ending was discussed above in Section 1. It is the most frequent nominative plural ending, which is not restricted to nouns of a particular gender or declensional class, but covers various noun types. Thus, whereas the factor determining the production of masculine plural nouns by adults is unambiguously animacy, with children it is the frequency of plural endings.

Factors determining the acquisition ofanimacy in Czech

319

The principle 'animacy over frequency' for adults and 'frequency over animacy' for children is illustrated well by the data pertaining to the two items with a neutral termination, namely lam and bam. The pseudo-word lam was combined with a picture of a non-living object and thus the expected inanimate plural ending was -[i]: lamy. The pseudo-word bam was combined with a picture of a living being and the expected animate plural ending was either -[i] (bami} or -ove (bamove). Either ending -[i] or -ove would be correct here, but only the ending -ove was unambiguous in terms of animacy. The ending -[i] could be interpreted either as animate or as inanimate. Table 7 presents the results of these two items. Table 7. Results for the items ending with a neutral consonant Inanimate/animate Animate ending -[i] ending -ove

Repetition Other

Adults Non-living (lam) 5;l-6;9 3;2-4;10

12 18 16

3 1 1

0 0 0

0 0 1

Adults 5;l-6;9 3;2-4;10

4 14 12

11 3 3

0 1 3

0 1 0

Living (bam)

For the item lam there was no choice in plural endings, and both children and adults predominantly selected the ending -[i]. Because of its ambiguous animacy status, the same ending could also have been chosen with the animate item bam, which the children conformed to. The adult subjects, however, changed their preferences and predominantly produced the unambiguously animate ending -ove. Obviously, whereas adults preferred to mark animacy unambiguously, children preferred the frequent, unmarked ending -[i], irrespective of its ambiguity with respect to animacy. The overgeneralization of inanimate plurals with the items ending in a hard consonant reveals further aspects of the children's production. Both the animate and inanimate plurals of such nouns can be formed by attaching the ending -[i] to the stem. If an animate plural is to be produced, an alternation must be performed as well, changing the final hard consonant into its soft counterpart. If the alternation is not performed, the resulting form is identical with the inanimate form. The prevalence of inanimate plu-

320

Denisa Bordag

rals in this group of items may thus be due to the children's failure to perform alternations or, alternatively, to a strategy for avoiding involvement in tasks yet to be fully mastered. The detailed analysis of alternation errors revealed that children indeed often omit them. Two main types of phonological alternations appear within Czech declension: palatalization and epenthesis. Data presented in Table 6 indicates that children might have problems with palatalization, i.e. with transforming the final hard consonant into its soft counterpart.6 Children's performance on a few other items in the experiment suggests that this problem is more general, because they appear to struggle with epenthesis as well. An epenthetic -e occurs in a consonant group before a zero ending. The final group of consonants can consist of a suffix (e.g. -k or -b) and a part of the stem; in this case the epenthesis is obligatory (e.g. kousek χ kousku 'piece', malba χ maleb 'painting'). In cases where the consonant group is morphologically inseparable, the application of epenthesis depends on whether the consonant string is phonetically admissible at the word's end. In loan words, the epenthetic -e may occur if the final group of consonants resembles a Czech suffix (e.g. korek χ korku 'cork', but alba χ alb 'album'). Two situations can occur in declension: 1. The base form contains an epenthetic -e; it must be removed if the form has a non-zero ending, e.g. chlapec 'boy', chlapci dative/locative sg. or nominative pi. 2. The base form has a non-zero ending; an epenthetic -e must be inserted if the ending is zero, e.g. chodba 'corridor', chodeb genitive pi. In our experiments, only the situation described in (1) appeared. Four items, two masculine pseudo-nouns tolek and molek and two feminine pseudo-nouns sukev and sizen, required deletion of the epenthetic -e in nominative plural: -lek —> -Ik-, -kev -—> -kv- and -zen —>· -zh-. Table 8. Subjects' performance on epenthesis tolek Plural

-1k- -lek-

Adults (15) 5;1-6;9(19) 3;2-4;10(18)

15 18 13

0 1 1

molek X

0 0 4

-1k- -lek15 13 11

0 4 3

Sukev X

0 2 4

sizen

-kv- -kev- χ 14 11 5

1 4 4

0 4 9

-zn- -zen14 8 4

1 9 9

X

0 2 5

Factors determining the acquisition ofanimacy in Czech

321

Table 8 shows that the adult subjects performed the epenthesis in almost all the plural forms they produced. Children, on the other hand, not only failed to apply the alternation, but obviously had difficulties with forming the plurals of these items, as can be seen from the rather high number of repetitions and other plurals (marked as "x" in the table) that they produced. Older children, however, performed better than younger children. An additional reason why children omitted the epenthesis could be an attempt to avoid pronouncing consonant clusters, which might be difficult for them. This is supported by the occurrence of plural forms (counted as Other') in which the children simplified the potential consonant clusters into one consonant, like sukove, size, or -si into -s- in plural forms of the feminine pseudo-words blost or klost. On the other hand, children sometimes performed epenthesis where it was not expected and where adult subjects did not, as with the items rapez (rapzi) and lapez (lapzi, lapze). This points to a more general problem children have with alternations, which is not limited to palatalizations, but concerns epenthesis as well. These difficulties might be an additional reason why children produced prevalently inanimate plurals with items ending in a hard consonant. Some of these plurals are very likely failed attempts to produce an animate plural with an alternation. Table 8 also demonstrates that children performed worse with the items sukev and sizen than with tolek and molek. A possible explanation is that children had difficulties with the gender status of these two items. Again, frequency may play a crucial role here. Most Czech masculine nouns end with a consonant and this termination is therefore considered typical for masculine gender. On the other hand, the most frequent feminine termination is -a. Consequently, children might have been confused by the feminine gender of the items sukev and sizen, because their termination is rather atypical for feminine nouns. And indeed, children sometimes used a typical masculine ending -ove to form the plural of these nouns, suggesting that they had problems with gender assignment. An account of the influence of the gender typical, ambiguous and atypical noun terminations on second language gender production which is in line with the observations reported in this study can be found in Bordag et al. (2006) and Bordag & Pechmann (2007). The difficulties children have with alternations should not prevent them from producing animate plurals of masculine nouns with a hard consonant termination, if marking animacy was a priority. They could have used the less frequent animate ending -ove instead. Table 9, however, reveals that

322

Denisa Bordag

overgeneralization of this less frequent ending was not an acceptable solution for children. It also shows that the performance of older children was adult-like regarding the ending -ove and that the deficit lay in the plurals involving alternations. Younger children underproduced both the less frequent ending -ove and the plurals involving alternations. The impact of frequency is clearly so strong that systematic transgressions against the grammatical category of animacy are more acceptable than overproducing a less frequent plural ending. Table 9. Animate plural forms Total

[i]+A

-ove (no A)

Adults 5;l-6;9 3;2-4;10

40.8% (49) 11.2% (17) 6.9% (10)

25.0% (30) 23.0% (35) 12.5% (18)

Finally, we should briefly discuss the data obtained from children with the real Czech words included in the experiment. They were:jezek 'hedgehog' (Ma), vlk 'wolf (Ma), muz 'man' (Ma), hreben 'comb' (Mi), mrak 'cloud' (Mi), nuz 'knife' (Mi), liska 'fox' (F), postel 'bed' (F), kote 'kitten' (N), jehne 'lamb' (N). The words included in the adult and child versions of the experiment overlapped only partially. However, adult subjects produced virtually no incorrect plurals. They supplied only expected plural forms which agreed with the rules of Czech grammar. Older children mostly produced expected plurals as well. An exception were plurals of the masculine nouns vlk andjezek, which require alternation if the plural is formed with the ending -[i]. Here, children sometimes failed to perform the alternation. On the other hand, they sometimes incorrectly palatalized the consonant -k into -c in the plural of the inanimate noun mrak, indicating that they have problems not only with performing the alternation per se, but also with its proper application. This was already demonstrated (above) by the pseudo-word data. The data on existing words support Köpcke's hypothesis that the number of repetition plurals depends on the noun's termination in the singular: the more it resembles a plural morpheme, the more likely it is that subjects will interpret it as such and simply repeat it in the plural condition. Among the 10 experimental real words included in the child tests, there were two neutral nouns, kot-e andjehn-e, which end with a morpheme that appears in Czech in both the nominative singular and plural. The older children pro-

Factors determining the acquisition ofanimacy in Czech

323

duced 13 and the younger children as many as 17 repetition plurals with these nouns. This contrasts sharply with the six repetition plurals produced by the younger children for the other eight existing nouns. Older children produced no repetition plurals with existing nouns except with these two items. Illustrative of the alternation problem was pluralization of the item nuz 'knife', which has the vowel -u- only in the nominative singular and -o- in all other cases, including nominative plural noze. The older children failed to perform this alternation only in three cases and produced no repetition plurals. Younger children failed in six cases and produced three repetition plurals. These data demonstrate the particular difficulty younger children experience with alternations as well as its scope: It ranges from palatalizations and alternations inside the stem (which require qualitative changes), to epenthesis requiring only a quantitative change, namely deletion or insertion of the epenthetic -e.

4. Conclusion This study focused on two main topics: the acquisition of the grammatical category of animacy in Czech and the acquisition of alternations. As far as we know, neither of these topics has been thoroughly discussed within child language research so far. The category of animacy does not seem to be a difficult concept, at least from an adult learner's point of view. It largely reflects reality, as we can generally equate the semantic property "living and moving" with the grammatical feature animate. It is referential like natural gender or number, and we know that the same children who failed to differentiate between living (animate) vs. non-living (inanimate) forms are sensitive to such extra-linguistic (referential) factors, because in most cases they correctly produce plural forms when they see more objects. The acquisition difficulties arise (among other reasons) due to two other characteristics of this category. Firstly, it applies only to a particular group of nouns, namely to those of masculine gender. Moreover, it is morphologically realized only in some cases (e.g. nominative or accusative singular), but not in others. The fact that the distinction animate vs. inanimate does not penetrate the entire nominal system might be a serious hindrance to its acquisition, especially considering that gender acquisition causes problems as well (Henzel 1975).

324

Denisa Bordag

Secondly, the acquisition of animacy as a subgender is probably severely hampered by its linguistic realization, i.e. by its "technical" aspects. There is no one form-one function relationship, which could simplify the detection of the distinction and its acquisition. As seen from Table 5 and Table 6, one and the same ending can signal animacy in one declensional class and inanimacy in another. Moreover, animate forms often (although not always) involve alternations, which are not salient. External flexion, especially in the form of segmental-additive endings, has proven easier than internal flexion, which includes alternations. The crucial factor underlying the acquisition of plural formation of masculine nouns turned out to be frequency. Children preferred the most frequent plural ending -[i] irrespective of animacy. On the other hand, the adult production of nominative plurals was clearly determined by the animacy status of the pseudo-nouns. Though the older children performed similarly to adults in some respects (general preference for unmarked animate plurals, performance on existing nouns), with respect to the animacy distinction the performance of both child groups deviated strongly from that of the adults. This indicates that acquiring the grammatical category of animacy is completed at a rather late age. Regarding the frequency debate, the presented data suggest that children's early plural acquisition depends only very weakly on animacy, i.e. the semantic distinction between living and non-living. Rather, children learn noun endings they hear in the input as a function of their frequency, irrespective of general semantic factors. However, although it does not interact with animacy, the frequency effect of child production interacts with other aspects of plural formation, such as the transparency of form-function mapping and formal aspects of particular plural forms (presence or absence of alternations). As a result, children overproduce the most frequent and easiest plural form (agglutinative addition of the -[i] ending), and in that respect their performance deviates from the distribution patterns of the adult input they hear. Overall, the data reported in this study are in accordance with recent research, which suggests that children do not operate initially with abstract linguistic entities or categories, but instead operate on the basis of concrete, item-based constructions (Tomasello 2001). Children seem to construct more abstract linguistic constructions only gradually - on the basis of linguistic experience in which frequency plays a key role. They also constrain these constructions to their appropriate domains of use only gradually — again on the basis of linguistic experience in which frequency is crucial.

Factors determining the acquisition ofanimacy in Czech

325

Notes 1. The languages in which animacy surfaces in the grammar are sometimes referred to as differential object marking systems (Aissen 2003). 2. Hard consonants are h, ch, k, g, r, d, t, «; any following [i] sound is written y, with the exception of foreign words (e.g., hymna 'anthem'). Soft consonants are z, s, c, f, c, j, d, t', ή; a potentially following [i] sound is written /', with the exception of foreign words (e.g., zidle 'chair'). Note that d, t', ή in front of/' (or e) are always written d, t, n. Neutral consonants are b, f, I, m, p, s, v, z; a potentially following [i] sound can be written either /' ory. 3. German has more complex flexion and is thus better comparable to Czech than English. 4. Epenthesis would be more appropriate here, but both forms were considered correct. 5. Forms toleci, tolekove (without epenthesis) would be also counted as correct animate plurals, but they were not produced by the subjects. However, a few corresponding forms of the non-living item molek (moled, molekove) were produced by children and they were counted as animate plurals. 6. Alternatively, with a selection of a proper stem allomorph, if we assume that alternations are not performed on-line, but that all possible allomorphs of each stem are listed in the mental lexicon and an appropriate one has to be selected depending on the target form, e.g. nominative singular vs. nominative plural.

References Aissen, Judith 2003 Differential Object Marking: iconicity vs. economy. Natural Language and Linguistic Theory 21: 435-483. Bolicova, Helena 1998 Nastin porovnavaci morfologie spisovnych jazyku slovanskych. Karolinum: Praha. Berko, Jean 1958 The child's learning of English morphology. Word 14: 150-177. Bordag, Denisa 2006 Psycholinguistische Erscheinungen in der Flexionsmorphologie des Tschechischen als Fremdsprache. Hildesheim: Olms Verlag. Bordag, Denisa & Thomas Pechmann 2007 Factors influencing L2 gender processing. Bilingualism: Language and Cognition 10(3): 299-314.

326

Denisa Bordag

Bordag, Denisa, Andreas Opitz & Thomas Pechmann 2006 Differences in LI and L2 Gender Processing: The Role of Noun Termination. Journal of Experimental Psychology: Learning, Memory, and Cognition 32: 1090-1101. Corbett, Greville 1991 Gender. Cambridge: Cambridge University Press. Dahl, Osten & Kari Fraurud 1996 Animacy in Grammar and Discourse. In Reference and Referent Accessibility, Thorstein Fretheim & Jeanette K. Gundel (eds.), 47-64. Amsterdam/Philadelphia: Benjamins. Henzel, V. M. 1975 Acquisition of grammatical gender in Czech. Reports on Child Language Development 10: 188-200. Kopeke, K. Michael 1987 Die Beherrschung der deutschen Pluralmorphologie durch muttersprachliche Sprecher und L2-Lerner mit englischer Muttersprache. Linguistische Berichte 107: 23^3. 1988 Schemas in German plural formation. Lingua 74: 303-335. 1998 The acquisition of plural marking in English and German revisited: Schemata versus rules. Child Language 25: 293-319. Mills, Anne E. 1986 The Acquisition of Gender: A Study of English and German. Berlin: Springer. Ramge, Hans 1975 Spracherwerb: Grundzüge der Sprachentwicklung des Kindes. Tübingen: Niemeyer. Mugdan, Joachim 1977 Flexionsmorphologie und Psycholinguistik. Tübingen: Narr. Park, Tschang-Zin 1978 Plurals in child speech. Journal of Child Language 5: 237-250. Schaner-Wolles, Chris 1988 Plural vs. Komparativerwerb im Deutschen: Von der Diskrepanz zwischen konzeptueller und morphologischer Entwicklung. In Experimentelle Studien zur Flexionsmorphologie, Hartmut Günther (ed.), 155-186. Hamburg: Buske. Stankiewicz, Edward 1968 The grammatical genders of the Slavic languages. International Journal of Slavic Linguistics and Poetics 11: 27-41. Smilauer, Vladimir 1972 Nauka ceskemjazyku. Praha: SPN. Theakston, Anna L., Elena V. M. Lieven, Julia M. Pine & Caroline F. Rowland 2004 Semantic generality, input frequency and the acquisition of syntax. Journal of Child Language 31:61 -99.

Factors determining the acquisition ofanimacy in Czech

327

Tomasello, Michael 2001 A Usage-Based Account of Child Language Acquisition. Paper presented at Amlap Conference 2001. Walter, S. 1975 Zur Entwicklung morphologischer Strukturen bei Kindern. Unpublished MS, Psychologisches Institut, Universität Heidelberg.

(Non-)Frequentist perspectives within UG

A note on acquisition in frequency-based accounts of Binding Phenomena Jason Mattausch andlnsa Gülzow

This article addresses the so-called 'pronoun interpretation problem' or 'delay of Principle B effect' - an observation in the study of language acquisition that challenges classical Binding Theory. We show that a recent, frequentist theory of binding which is empirically superior to the classical Binding Theory can, with a minor adjustment, address the pronoun interpretation problem and thus explain why children acquire interpretational restrictions on pronouns later than they acquire such restrictions on reflexives and why the acquisition of interpretational restrictions lags behind restrictions on generation. 1. Introduction The purpose of this article is to reconcile a recent, novel approach to binding phenomena with data from acquisition experiments, namely the so-called 'pronoun interpretation problem' or 'delay of Principle B effect', which has been noted in several acquisition studies, e.g. Wexler & Chien (1990) and Grimshaw & Rosen (1990). The structure of the paper is as follows. In the following section, classical binding theory is introduced and we discuss three problems related to the theory. We then outline a recent alternative to classical binding theory, which solves two of these three problems. Section 4 discusses a solution to the third problem. 2. Three challenges to classical binding theory The following examples illustrate a very common pattern in natural language binding phenomena. (1)

a. * Hei pleases him( b. He·,· pleases

332

Jason Mattausch andlnsa Giilzow c. Het pleases himselft d. *Het pleases himselfj

Below are two principles of Chomsky's Binding Theory (BT), the most well-known approach to explaining the pattern manifested in (1), (Chomsky 1981). (2)

a. Binding Principle A: a reflexive must be locally bound, b. Binding Principle B: a pronoun must be locally free.

Principles A and B account for the ungrammaticality of (1 a), where a pronoun is locally bound and for the ungrammaticality of (Id), where a reflexive is locally free.' The BT analysis of the pattern exemplified in (1) both undergenerates and overgenerates. Firstly, as pointed out in Levinson (2000), there are a considerable number of languages which appear to lack morphological means of encoding reflexivity altogether and use pronouns reflexively, thus disobeying Principle B systematically and obeying Principle A only vacuously. One example is English itself, though not its modern form. Specifically, evidence from Old English (cf. Visser 1963: 420^39; Mitchell 1985: 115189; Keenan 2000, 2001) shows that the opposition between the OE pronoun hine and the emphatic hine selfne is not comparable to the opposition between the modern counterparts him and himself, since hine could appear locally bound and hine selfne, though often used as a reflexive, did not necessarily take a local antecedent. (3)

Old English (Siemund 2000) Hinei he, beweraö mid wwpnum. Him he defended with weapons 'He defended himself with weapons.'

(4)

Old English (Mitchell 1985:115) MoyseSi, se Moses he hine selfnej him self 'Moses was

de wees Godej sua weord dcet het oft wiö who was to-God so dear that he often with sprcec. spoke so dear to God that he often spoke with him.'

A note on acquisition in frequency-based accounts of Binding Phenomena

333

A second challenge to BT can be illustrated by imagining a hypothetical language which manifested the pattern shown below in (5). (5)

'Anti-English' a. Heι pleases /z/w, b. *He, pleases hinij c. * He i pleases himself} d. He i pleases himself,

We call this hypothetical language 'Anti-English' because it exhibits exactly the opposite pattern as modern English in that the self -marked form is only grammatical with a non-local antecedent whereas the bare pronoun demands a local one. One working in the BT framework would have no problem accounting for such a pattern - he would surely just say that, in Anti-English, him is a reflexive whereas Anti-English himself 'is a pronoun. The question then becomes why there are no languages like Anti-English, i.e., why do languages mark pronominal objects of reflexive predicates instead of marking those of non-reflexive ones? Standard BT offers no answer to this question.2 Finally, classical BT faces difficulty accounting for data observed in the study of language acquisition. Several studies have observed a 'pronoun interpretation problem' or 'delay of Principle B effect', whereby children (a) appear to interpret and produce reflexives in accordance with the binding principles, (b) appear to produce pronouns in accordance with the binding principles and (c) do not appear to interpret pronouns in accordance with the binding principles, but rather interpret pronouns reflexively about 50% of the time in experiments (e.g. Shipley & Shipley 1969; Charney 1980; Chiat 1981; Loveland 1984; Chien & Wexler 1990; Grimshaw & Rosen 1990; Girouard, Richard & Decarie 1997). The issue is discussed in more detail below, but it shall suffice here to note that if, as classical BT holds, Principles A and B are innate, universal tenets of grammar then the 'delay of Principle B' effect is not predicted to occur. We shall proceed by outlining an alternative to classical BT due to Mattausch (2004, 2006), which addresses the problems of undergeneration and overgeneration, then turn to a novel solution to the acquisition problem based on that approach.

334 Jason Mattausch andlnsa Giilzow 3.

An alternative to BT

This section outlines the approach to binding phenomena proposed by Mattausch (2004, 2006) and shows how the approach addresses the first two challenges to classical BT discussed above. The account is based on Bidirectional-Evolutionary Optimality Theory, which is introduced below.

3.1. Optimality Theory Optimality Theory (OT) is a theory of grammar that gives up the idea of absolute principles of grammar in favor of conflicting, violable constraints, which can be ranked in various possible ways to reflect their strength in a particular language. In OT, a certain input gets associated with a multitude of possible outputs or candidates. Each candidate is then evaluated with respect to a series of ranked constraints, of which there are two basic types faithfulness constraints, which penalize divergence of the output candidate from the original input and markedness constraints, which militate against certain features or properties of the output. The various possible outputs are compared to one another on the basis of which constraints they violate, the relative violability (i.e., ranking) of the constraints, and the number of violations committed in order to determine the Optimal' or 'maximally harmonic' candidate relative to the original input.

3.2. Bidirectional Optimality Theory In generative grammars whose essence is to produce morphological or syntactic expressions for some underlying meaning, the definition of optimality is as follows. (6)

Optimality (production) A form/is an optimal expression, given a meaning m, iff there is no /' such that /' is more harmonic than / (write: /' > f), given m as an input.

In comprehension grammars whose essence is to interpret morphological or syntactic expressions, the definition of optimality is as below.

A note on acquisition in frequency-based accounts of Binding Phenomena

(7)

335

Optimality (comprehension) A meaning m is an optimal interpretation, given a form/ iff there is no m' such that m' > m, given/as an input.

Bidirectional OT (championed by Blutner 2000; Wilson 2001; Zeevat 2001; Jäger 2003a) is a variation of OT meant to incorporate both production and comprehension aspects of language into one grammar and capture the interdependency of the two. The issue of interdependency is crucial, since it is commonsensical to capture the idea that, in a communication situation, an expression should, first and foremost, allow the hearer to recover the intended meaning of the expression. Such an idea is captured by formulating a definition of bidirectional optimality as below. (8)

Bidirectional optimality (Jäger, 2003a: 19) a. A form-meaning pair is hearer optimal iff there is no pair > . b. A form-meaning pair is optimal iff either (i) is hearer optimal and there is no distinct pair such that > and such that are increased by the constraints that favor are increased by the constraints that favor > > >

Language type *self,co, *pro,co, *pro,co, *self,co,

*pro,dis *pro,dis *self,dis *self,dis

Modern English Old English (no reflexives) Anti-English Anti-Old English (no simplex pronouns)

Moreover, we can employ stochastic OT to illustrate how languages like Middle English, where pronouns and reflexives were both attested but did not appear in complementary distribution, can be represented by a stochastic ranking of the constraints under consideration. Consider an extreme example where a language made no distinction at all between pronouns and reflexives with respect to where they could appear, and no distinction in their interpretation. We will take it for granted that sentences in which the subject and object refer to distinct entities constitute the vast majority we'll say 98% - of sentences used by speakers of all grammars. (Note that this assumption is similar to the Disjoint Reference Presumption (DRP) of Farmer and Harnish (1987) but rather than seeing it as a interpretational presumption made by language users, we take it as a simple fact of life

A note on acquisition in frequency-based accounts of Binding Phenomena

341

about language use.) A speaker who spoke the hypothetical language we are considering would produce corpus frequencies like those below. Table 2. Hypothetical frequencies of pronoun/reflexive distribution

disjoint conjoint

pro

pro+self

% marked

49% 1%

49% 1%

50% 50 %

We can use the frequencies in Table 2 to simulate a grammar learned based on those frequencies. Feeding BiGLA with twenty thousand form-meaning pairs drawn at random based on the frequencies in Table 2 resulted in the learning curves in Figure 2.6

4H +self,co = + 2.86

20-2'*pro,dis=-2.95

0

5000

10000

15000

20000

Figure 2. Learning curves (20K inputs) per Table 2

The resulting grammar shown in Figure 2 is roughly what one should expect under the circumstances. The large gap between the highly ranked constraints *pro,co and *self,co on the one hand and the low ranked *pro,dis and *self,dis on the other represent the preference for interpreting the arguments of predicates as disjoint. Note that this is basically a stochastic version of the Disjoint Reference Presumption, but rather than stipulating it as a pragmatic presumption ä la Farmer & Harnish (1987), a pragmatic implicature toward stereotypicality ä la Levinson (1991, 2000), or a

342

Jason Maitausch andlnsa Gülzow

'derivative of world-knowledge' ä la Huang (1994, 2000), a statistically sensitive bidirectional learning algorithm like the BiGLA can provide a functional explanation for how and why DRP-like effects came to be. The preference for disjoint interpretations is derived directly from a statistical asymmetry in the training corpus and the application of hearer-mode learning to constraints which 'record' that asymmetry. On the other hand, the constraints *pro,co and *self,co have been learned as having almost exactly the same ranking value, and the same is true for *pro,dis and *self,dis. This reflects the fact that, in the training corpus, reflexives and pronouns were in perfectly parallel distribution. Various degrees of variation can be also be captured, though we leave the reader to prove to him- or herself that the four constraints at our disposal in a StOT framework give us all we need to easily handle the problem of undergeneration that faced standard BT.7

3.4. Addressing overgeneration: Evolutionary Optimality Theory With respect to the overgeneration problem in standard BT, one faces the task of explaining why there are no languages like 'Anti-English', where morphologically complex expressions play the role of pronouns and morphologically simplex expressions are reflexive. Fortunately, aside from the advantages already mentioned, stochastic, bidirectional OT and bidirectional learning offer an interesting opportunity to describe language change. Moreover, explanations about the direction of language change can be found when one considers what types of constraints grammars consist of and how these constraints interact. The Iterated Learning Model (ILM) of language evolution due to Kirby & Hurford (1997) takes each generation of learners to be one turn in a cycle of language evolution and, by applying a learning algorithm to the output of one cycle, one may produce a second cycle, and then a third, a fourth, and so on. In the context of bidirectional gradual learning of a StOT grammar, the first-generation learner would be exposed to a set of corpus frequencies, he would adjust his grammar accordingly until it converged into an appropriate set of ranking values. He would produce his own speech in accordance with the grammar he had acquired and the frequencies of his own speech would serve as the corpus frequencies for the second-generation learner. Thus, per the ILM, a learner who acquired the grammar in Figure 2 would himself become a 'teacher' to the next generation of learners

A note on acquisition in frequency-based accounts of Binding Phenomena

343

and the frequencies that he produced would serve as a training corpus for others. The actual simulated output frequencies for a speaker whose grammar was the one in Figure 2 are given in Table 3. Table 3. Output frequencies per Figure 2

disjoint conjoint

pro

pro + self

% marked

50.95 % 0.88 %

47.05 % 1.12%

48% 56%

There has obviously been some cross-generational fluctuation between the (hypothetical) grammar that generated the training corpus and the learned grammar. This is not at all uncommon; in fact, perfect statistical replication of a non-categorical marking pattern from one generation to the next is very rare. Based on the frequencies above, we see that the first generation learning has taken a step toward the Modern English pattern, since self-marked outputs have decreased for disjoint inputs and increased for conjoint ones. However simulations of language evolution from a neutral starting point of grammars comprised of the bias constraints above were unpredictable. There are three possible scenarios: (a) evolution into English, (b) evolution into Anti-English and (c) neither (a) nor (b), i.e., persistent variation. Conducting multiple simulations showed that all of these results were achievable and thus, so far, nothing explains why Anti-English-type grammars are unattested in natural language. However, adding a markedness constraint to represent some universal force of structural economy causes this picture to change significantly.8 Let us assume that an additional constraint represents a universal force of articulatory economy. (11) *Stmct: Avoid morphological structure. The inclusion of a constraint like *Struct will be very significant. The general reason: generative optimization in a grammar with both bias constraints and markedness constraints will be determined not only by the ranking values of bias constraints, but also by how the markedness constraints are ranked among them. With respect to the case at hand, (ignoring blocking effects for the moment) the probability that a se/^marked output is the optimal output for, say, a conjoint input is now no longer equal to the probability that *pro,co outranks *self,co, but rather to the probability that *pro,co outranks both *self,co and *Struct. Moreover, because of the

344

Jason Mattausch andlnsa Giilzow

mechanics of the (Bi)GLA, there will be a strict relationship between the ranking value of *Struct and the various bias constraints.9 Thus, a grammar like the one under consideration this needs to converge in a way such that the markedness constraint and the bias constraints 'share the labor' in the prevention of self-marked forms. (Jäger & Rosenbach 2003 call this effect 'ganging-up cumulativity' - each constraint is relevant to the evaluation regardless of its ranking value.10) To see the difference between learning a grammar with bias constraints only, as above, and a grammar with bias constraints and a markedness constraint, we can again feed BiGLA with twenty thousand form-meaning pairs drawn at random based on the frequencies in Table 2. The result was the learning curves in Figure 3.

3H

*self,co =

•-*pro,dis=—].44

•*Struct=-2.\4

-3 *self,dis =-4.28

-5 0

5000

10000

15000

20000

Figure 3. Learning curves (20K inputs) per Table 2

While it may be difficult to see with the naked eye, the learned grammar in Figure 3 is very different from the one in Figure 2. Briefly stated: because *Struct is strictly a generative constraint (i.e., it is neither promoted nor demoted in the hearer-mode), hearer-mode and speaker-mode will be learning a different number of constraints. Hearer-mode learning will be struggling to keep ranking values exactly as they were in Figure 2 whereas speaker-mode learning will be struggling to find a proper balance between the bias constraints and the markedness constraint. But a proper balance cannot be found and the compromise that is reached will favor generative accuracy for the more common type of learning data, i.e., form-meaning

A note on acquisition in frequency-based accounts of Binding Phenomena

345

pairs where the subject and object are disjoint.11 The resulting output frequencies are in Table 4. Table 4. Output frequencies per Figure 3

disjoint conjoint

unmarked

marked

% marked

52.41 % .64%

45.59 % 1.36%

46.5 % 68 %

One can see that - in the spirit of Shannon's (1948) Optimal coding' and what Horn (1984) called a 'division of pragmatic labor' - marked forms have gravitated toward rare meanings. The marked-forms-for-rare-meanings pattern taking shape here can be seen as a direct consequence of four things: (a) bias constraints (b) markedness constraints (c) the mechanics of the GLA and (d) the bidirectional application of those mechanics. The new asymmetry that has shown up in the first-generation learner's corpus frequencies will have important consequences for future generations. Per the ILM, the student who produces a greater percentage of self-marked outputs for conjoint inputs than he does for disjoint inputs will eventually become a teacher to the next generation and thus a second-generation learner will be exposed to a training corpus in which the tendency to selfmark locally conjoint pronouns is greater than the tendency to mark locally disjoint ones. Without going into detail, the inevitable result of evolutionary simulations using a grammar with bias constraints in (10), plus *Struct, beginning with the corpus frequencies in Table 2 is illustrated in Figure 4.12 64•*self,dis=

2•*Struci-

0-2-

"self.co =-2.63

-Λ*pro,dis=-5.\4

-6 0

5

10

Figure 4. Evolution (20 generations)

15

20

346

Jason Mattausch and Insa Gülzow

The evolved grammar strictly follows the Principle A and B patterns of Modern Standard English, see Table 5. Table 5. Output frequencies (100th generation)

disjoint conjoint

pro

pro+self

% marked

98% 0%

0% 2%

0% 100%

This result was, as noted, the only result achievable using the constraints and frequencies under consideration and thus the overgeneration problem of standard BT can be solved by considering a frequentist, evolutionary account of binding phenomena that hinges on the interaction of bias constraints and markedness constraints, which guarantees a marked-form-forrare-meaning strategy.

4. The pronoun interpretation problem As already mentioned above, it has been demonstrated in studies conducted within the standard BT framework that children acquiring English as their first language disobey Binding Principle B for a relatively long time in their interpretation of pronouns. Chien & Wexler (1990) have shown that English children younger than four years of age have a great tendency to interpret sentences like (12a) as if they meant (12b). (12) a. Mama Bear is touching her. b. Mama Bear is touching herself. In an experiment testing English children's knowledge of the Binding Principles, children were shown pictures with the characters Goldilocks and Mama Bear. If shown a picture in which Mama Bear is touching herself, children younger than four years of age tend to answer the question Is Mama Bear touching her? with yes. Performance becomes better with increasing age, although children between five and six years of age still perform at chance level and children in the age group between six and seven years of age reacted in a target like manner only in 76% of the cases (Chien & Wexler 1990: 269, 273).13 Similar results are reported by Grodzinsky & Reinhart (1993) who tested ungrammatical coreference in sentences like

A note on acquisition in frequency-based accounts of Binding Phenomena

347

Grover touches him with pictures in which Grover is touching himself. No evidence exists suggesting that children exhibit the same disobedience of Principle B in their production of anaphoric expressions. Obviously, the results of such comprehension studies present a challenge to classical BT. Some previous attempts to resolve the problem are outlined below. Chien & Wexler (1990) for instance claim that until a comparatively late age, English children overgeneralize or misinterpret the rare occurrences of a coreferential interpretation of a personal pronoun and a preceding noun phrase. They follow Reinhart (1983, 1986) in arguing that children know Principle B but lack a Pragmatic Principle P. In (13a) both he and him are taken to be John, thus he and him are coreferential.14 The indexing in these sentences must be as represented in (13b) and (13c), as (13d) would suggest that him is referentially dependent on he which would violate Principle B. (13) a. b. c. d.

That must be John. Thatj must be Johrij. At least hei looks like himj. *At least he, looks like him,.

Before children have a Pragmatic Principle P they are unable to realize that coreference of he and him in (13c) is only possible in very specific contexts. Therefore, they overgeneralize this rather rare occurrence and also allow non-target coreference in other contexts. Grodzinsky & Reinhart (1993) offer a solution that is based on the subtle interpretational differences of a sentence such as (14a), represented in (14b) and(14c). 15 (14) a. Alfred thinks he is a great cook. b. Alfred (λχ (χ thinks χ is a great cook)) c. Alfredj (λχ (χ thinks hef is a great cook)) In (14b) the pronoun is a bound variable while in (14c) the pronoun represents an instance of coreference. Children confronted with sentences like Grover touches him have to assess whether the pronoun represents a bound variable or is coreferential in order to find out if the two noun phrases have an identical referent or if they have distinct referents. One further explanation of the pronoun interpretation problem of English children is given by Grimshaw & Rosen (1990) who claim that if used non-deictically, third person pronouns are difficult to interpret. Assuming

348

Jason Mattausch andInsa Giilzow

that children do know that pronouns normally have discourse antecedents and that normally pronouns cannot be locally-bound, experimental test sentences of the sort Annika is talking to her will cause a conflict for the children. A non-target interpretation of her having the same referent as Annika may arise when children respect the pragmatics of pronouns but violate their syntactic requirements. Finally, Grodzinsky & Reinhart (1993) claim that for young children this task may prove to be too complex and thus they end up guessing. Note that all of the solutions mentioned above hinge crucially either on reference to pragmatic information distinct from the syntax or on speculation about children's abilities to distinguish bound variable readings from coreference. None of these proposed solutions offers a straightforward syntactic explanation of why a supposedly innate syntactic principle should, in the acquisition phase, be systematically violated in the interpretation of anaphoric expressions but not in the production of these expressions. Below we suggest how the alternative to classical BT in section 3 can solve the pronoun interpretation problem without reference to pragmatics or processing difficulties. As noted, the pronoun interpretation problem as described above presents an equally serious challenge to the alternative approach to binding phenomena advocated in section 3. To see why, consider a corpus like Table 5, i.e., one in which the Principle A and B pattern is strictly obeyed. We can use this corpus as a training corpus to simulate a grammar learned by a child learning modern English. Feeding BiGLA with twenty thousand pairs drawn at random based on the frequencies in Table 5 yielded the learning curves in Figure 5. One can make the following observation: The constraints *pro,co and *pro,dis — the constraints which regulate the interpretation of pronouns have distanced themselves from each other more quickly and to a greater degree than the constraints which regulate the interpretation of reflexives, *self,co and *self,dis. On the one hand, this is exactly what we should expect - because the vast majority of learning data were pronouns, the learner has learned the correct way to interpret of these expressions faster and more veraciously than he has learned to interpret the much rarer reflexive expressions. On the other hand, it contradicts what might be a commonsense intuition - that more common expressions might tend to be less restrictively interpreted - and, in fact, also contradicts the experimental data that constitute the pronoun interpretation problem. In this way, a frequency-based approach to binding phenomena is very seriously threatened by the pronoun

A note on acquisition in frequency-based accounts of Binding Phenomena

349

interpretation problem, since it not only fails to predict that phenomenon but actually predicts exactly the opposite.

3H

s

/

1-1 -

s

i / _..·· /.··' _^f

•\

-3-

\ χ \ '-«... ^ ' v \

\

·

*se(£co = -1.74

v

-*pro,dis=-4.77

0

5000

10000

15000

20000

Figure 5. Learning curves (20K inputs) per Table 5

4.1. Addressing the pronoun interpretation problem Below we present a solution to the pronoun interpretation problem.16 The solution will hinge on an alternative definition of bidirectional optimality. In particular, we propose the following, revised definition. (17) Revised bidirectional optimality a. A meaning m is recoverable from a form / iff there is no formmeaning pair > . b. A form-meaning pair is speaker optimal iff either (i) m is recoverable from /and there is no pair > , or (ii) no form χ is such that m is recoverable from χ and there is no pair > . c. A form-meaning pair is hearer optimal iff there is no pair such that Can Bill play the sax? b. The sky is blue. => Is the sky blue?

Over thirty years ago, Chomsky (1971, 1975) observed that a simple 'structure-independent' hypothesis yields the correct results for much of the input that children receive. It can be assumed that much of the input to children is simple sentences like the declarative ones in (1). The following structure-independent hypothesis, for example, will generate the yes/no-questions in (1): (2)

Structure-Independent Rule: To form a yes/no-question, move the first verbal element {is, can, has, ...} of the declarative statement to the front.

The inadequacy of this structure-independent rule is revealed when it is applied to complex examples with a modifying relative clause (who is riding the skateboard), as in (3a). The first is in the declarative sentence (3a) appears in the modifying clause; if this is is moved to the front, the result is an unacceptable yes/no-question in (3b). (3)

a. Declarative: The boy who is riding the skateboard can do flips b. Yes/no-question: Is the boy who _ riding the skateboard can do flips?

To produce the correct yes/no-question corresponding to (3a), the can following the entire subject phrase the boy who is riding the skateboard is moved to the front, yielding (3c). c. Yes/no-question: Can the boy who is riding a skateboard _ do flips?

362

Rosalind Thornton, Stephen Grain and Graciela Tesan

A rough formulation of the structure-dependent rule that gives the correct results is something like the following: (4)

Structure-Dependent Rule: Move the auxiliary verb in the main clause to a sentence-initial structural position.

Chomsky (1971) maintained that children would never adopt structureindependent hypotheses, even if the data available to children were consistent with both structure-independent and structure-dependent rules. In other words, children would not be expected to make certain kinds of mistakes in forming yes/no-questions at any stage in language development. Thus, for example, they are not expected to produce questions like (3b): Is the boy who riding the skateboard can do flips! In an elicited production study, Grain & Nakayama (1987) evoked yes/no-questions from thirty 3-5 year old children, to see if they ever made such mistakes. Although children made errors of certain kinds, they never produced questions that were consistent with structure-independent rules. Such findings suggest that children converge on a system of linguistic principles that is equivalent to that of adults at least by their third birthday, and possibly much earlier. Lasnik & Grain (1985) argue, moreover, that if convergence of the adult system depends on there being evidence, then evidence must be available in abundance. Suppose to the contrary, that evidence falsifying the structure-independent hypothesis for forming yes/noquestions is not available to children before they reach their third birthday. In other words, suppose children are not exposed to questions like (3c). One should then expect to observe that many children make structureindependent errors in situations that call for complex yes/no-questions. But Grain & Nakayama (1987) did not find any evidence that children were adopting structure-independent rules when they elicited yes/no-questions with complex subjects. So we are led to the following conclusion: either children never form structure-independent hypotheses, or there is abundant evidence available to and used by very young children. According to Cowie (1999), "... something like the requisite guarantee can be provided when one reflects on the sheer size of the data sample to which a learner has access." (p. 219). If the requisite evidence includes sentences like (3c), however, then the evidence is not readily available to children. A search of the input to English-speaking children turned up only one example of a structure like (3c) out of about 3 million utterances (reported by MacWhinney 2004 using the CHILDES database; see MacWhinney & Snow 2000).

Principles, parameters and probabilities

363

Advocates of the experience-based approach have therefore proposed other sources of evidence for children. For example, it has been suggested that w/z-questions like (5) can provide children with the relevant data. The w/z-question in (5) is assumed to be derived from the declarative sentence represented in (6). (5) Where 's the other dolly that was in there? (6) [ the other dolly that WOSAUX /w there ] [ /SAUX where ] Notice that the representation that is assigned to the declarative sentence in (6) is partitioned into a predicate [isAUx where] and a relative clause [the other dolly that wasAUX in there]. To form the corresponding w/z-question, (5), the wh-question word where is moved to initial position, and the verb from the predicate is raised in the structure.1 Despite the absence of sentences like (3c=7) in the input to children, if questions like (5) are readily available in the input, then these questions would be subject to a similar analysis (compare examples (6) and (8)). In forming both kinds of questions, the isAux in the subject phrase appears first in the declarative sentence, but it remains in place, whereas the isAUX or canAUX in the predicate is moved to the front in order to form the corresponding question. (7)

Can the boy who is riding the skateboard do flips?

(8)

[ the boy who /SAUX riding the skateboard ] [ cawAUX do flips ]

As MacWhinney (2004) acknowledges, this experience-based account "requires children to pay attention to relational patterns, rather than serial order as calculated from the beginning of the sentence" (p. 891). If even advocates of the experience-based approach admit that children analyze linguistic expressions into relational patterns, we should ask what distinguishes the experience-based account from the nativist account. The difference is that on the experience-based account, a child could, in principle, have learned to move the first is to the front, as in the structure-independent rule described earlier, even if no child happens to take up this option. On the nativist account, children are incapable of any such structure-independent analyses. Regardless of the input, children should be compelled to impose a symbolic analysis onto sequences of linguistic expressions. According to the experience-based account, on the other hand, children have no such predisposition; the system that children acquire depends on the statistical regularities of the input.

364

Rosalind Thornton, Stephen Grain andGraciela Tesan

It remains to determine whether or not questions like (5) [Where's the other dolly that was in there?] are available in sufficient quantity in the input to children, to ensure that every child converges on a grammar that conforms to structure-dependent operations. In the present case, Legate and Yang (2002) conclude that the input does not suffice. They report the results of a search through transcriptions of the input to two young children, Nina and Adam (in the CHILDES database). The input to Nina consists of 14 critical w/z-questions out of 20,651 questions overall. For Adam, there were just 4 critical examples out of 8,889 questions in the input. The paucity of critical input for these children bears out Chomsky's (1975) conjecture that a child could "go through a considerable part of his life without ever facing relevant experience." Moreover, such low frequencies of relevant input make it unlikely that every child encounters the requisite evidence by the age at which they are found to adhere to structure dependence. This is a problem for the experience-based account because, without relevant input, some (perhaps many) children would be expected to commit structure-independent errors, by asking questions like (3b). But this is contrary to fact, as the findings of Grain & Nakayama (1987) demonstrate. As a final remark, the example of structure-dependence exposes a faulty assumption of the experience-based approach, namely that 'core' phenomena are more frequently attested in the primary linguistic data. This point is underscored in the next section. 3. A constraint on contraction The second example of a putative universal principle is a constraint that governs where contraction processes may and may not occur. As in the previous example, we are concerned with investigating whether or not children's productions closely follow the available input. In English, contraction constrains a number of verb plus infinitive sequences, but is best known for inhibiting contraction of the verbal elements want and to to wanna in certain kinds of sentences. Inspection of data makes it clear that the wzrwa-contraction is permitted most of the time. Examples (9)-(12) illustrate a range of linguistic environments that permit the contraction of want and to. It can be seen that the wanna-contraction is permitted in (most) w/z-questions, in yes/no-questions and in declarative utterances. An example of a w/z-question in which contraction is not permitted is given in (13a). It is pertinent to ask what sets this example apart from the others, especially other w/z-questions, such as (9) and (11)?

Principles, parameters and probabilities

(9)

365

a. Who does Arnold wanna make breakfast for? b. Who does Arnold want to make breakfast for?

(10) a. Does Arnold wanna make breakfast for Maria? b. Does Arnold want to make breakfast for Maria? (11) a. Why does Arnold wanna make breakfast? b. Why does Arnold want to make breakfast? (12) a. / don't wanna make breakfast for Arnold or Maria. b. / don V want to make breakfast for Arnold or Maria. (13) a. *Who does Arnold wanna make breakfast? b. Who does Arnold want to make breakfast? According to a generative-transformational grammar, w/z-questions are formed by movement of a w/z-phrase from one position at an underlying level of representation to another position at another level of representation, where the w/z-phrase is pronounced. A further assumption of the account is that a record, which we abbreviate as / (for 'trace'), is left behind at the site of the origin of the w/z-movement. In (14) the w/z-phrase originates in the subject position of the embedded infinitival clause t to kiss Bill. When the w/z-phrase starts out in subject position, between want and to, as in (14), the trace left behind by the w/z-movement operation blocks the contraction of want and to. This explains why (14b) is ruled out. The same account explains the unacceptability of (13a). (14) a. Who do you want t to kiss Bill? b. *Who do you wanna kiss Bill?

Subject Extraction

By contrast, in (15), the formation of the w/z-question requires the movement of the w/z-phrase from the object position of the embedded infinitival clause. In that case, the trace does not originate between want and to, so wanna-contraction is permitted. (15) a. Who do you want to kiss t ? b. Who do you want to kiss t ?

Object Extraction

These facts allow us to make the following generalization: Contraction of the verbal elements want and to is blocked if the trace of w/z-movement originates between them. In declaratives and in yes/no-questions, there is

366

Rosalind Thornton, Stephen Grain and Graciela Tesan

no trace of w/z-movement, and so the constraint on contraction is irrelevant. Contraction is therefore not inhibited in these structures. If the grammars of English-speaking children lacked the constraint on contraction of -want and to (across the trace of a moved w/z-word), then child English would include more sentences than adult English does. In other words, without the constraint, children would Overgenerate', and would produce sentences like (13a) and (14b) with illicit contraction of want and to. Children who lack the constraint on contraction across a trace should permit contraction to a similar extent in both subject and object extraction questions. To test children's adherence to the constraint, an experiment was designed to elicit relevant questions from children (Grain & Thornton 1998). A total of 26 children between 2; 10 and 5;5 were interviewed (mean age 4;3), 14 of whom produced 2 or more subject extraction questions like (14), and 2 or more object extraction questions like (15). These children permitted a comparison of the proportion of contraction in subject and object extraction questions. Close analysis of the data from these 14 children showed that when the w/z-phrase was moved from object position, the children contracted 88% of the time (60/68 questions). A different pattern was found in subject extraction questions, however. Here, children failed to contract 92% of the time (6/74 questions had contraction). Thus, children were found to act in accordance with the constraint.2 This concludes our discussion of principles, and brings us to parameters. According to the perspective of linguists working within the generativetransformational tradition, children should be expected to sometimes follow developmental paths to the adult grammar that would be very surprising from a data-driven perspective. Of course, children quickly internalize a grammar equivalent to that of adults around them. But a child who has not yet achieved, say, a dialect of American English can still be speaking a natural language - albeit one that is (metaphorically) a foreign language, at least somewhat, from an adult perspective. And interestingly, the children of English speakers often do exhibit constructions that are not available in English - but ones that are available in other languages spoken by actual adults. This is unsurprising if children project beyond their experience, rather than being inductively driven by it. From a nativist perspective, children are free to try out various linguistic options (compatible with Universal Grammar) before 'setting parameters' in a way that specifies some particular natural grammar, like that of Japanese or American English. An extension of this line of reasoning is the Continuity Hypothesis (Pinker 1984; Crain 1991; Grain & Pietroski 2001).

Principles, parameters and probabilities

367

According to the Continuity Hypothesis, child language can differ from the local adult language only in ways that adult languages can differ from each other. The idea is that at any given time, children are speaking a possible (though perhaps underspecified) human language - just not the particular language spoken around them. If this is correct, we should not be surprised if children of monolingual Americans exhibit some constructions characteristic of Germanic, Romance or East Asian languages, even in the absence of any evidence for these properties in the primary linguistic data. Indeed, such mismatches between child and adult language may be the strongest argument for Universal Grammar. We report two examples, both involving w/z-questions. The first example reveals a trace of Romance in child English, and the second reveals a trace of Germanic in child English. In each case, the relevant facts come into view only when they are framed within a detailed theory of some non-English phenomena, alongside some otherwise puzzling observations about child English.

4. Children's why-questions In most English w/z-questions the question word must be immediately followed by an inflected auxiliary verb (i.e., a tensed form of be, do, can, have, etc.). Hence, the examples in (16a)-(16d) are acceptable, whereas the examples in (16e)-(16h) are not acceptable.3 (16) a. b. c. d. e. f. g. h.

Why are you here? What do you want to do? Where is he going? Who don 'tyou want to win the game? * Why you are here? *What you want to do? *Where he is going? *Who you don't want to win the game?

It has frequently been noted that for some children learning English, w/zquestions with why lack subject-auxiliary inversion to a greater extent than other w/z-questions. The absence of inversion for w/zy-questions persists in these children's speech well after inversion is consistently present in other w/z-questions (e.g., Labov & Labov 1978; Rowland & Pine 2000). Naturally, this observation is out of step with adult grammars, in which all w/z-ques-

368

Rosalind Thornton, Stephen Grain and Graciela Tesan

tions feature subject-aux inversion.4 Adopting the Continuity Hypothesis, de Villiers (1990) and Thornton (2004) have both noted that why behaves differently from other w/z-phrases in French and Italian in not enforcing inversion. Thornton looks to the details of the Italian pattern to explain children's linguistic behavior. In Italian, the w/z-word corresponding to English why is perche. Italian perche differs from other Italian w/z-words in simple questions (for analysis, see Rizzi 2001). As the example in (17) illustrates, the adverb gia as well as an entire subject phrase (/ tuoi amid) can intervene between perche and the inflected verb (hanno). No linguistic material can intervene with other w/z-words in Italian.5 In questions with other w/z-phrases, the inflected verb must be raised in the structure to a position higher than the subject. (17) Perche (I tuoi amid) gia hanno flnito il lavoro? Why (the-PL your friends) already have-3PL finished the-SG work 'Why (your friends) already have finished the work?' However, in complex w/z-questions with perche, the intervention of short adverbs is sometimes prohibited, as shown in (18). This prohibition is enforced if the w/z-phrase is being moved a 'long distance' from the embedded clause, that is, if the question is asking about the reason for someone's resignation. This movement triggers verb raising. (18) Perche ha detto ehe si dimetterä? Why have-3SG said that self resign-3SG.FUT 'Why did he say that he would resign?' Thus, complex w/zy-questions like (18) pattern the same way in both English and in Italian, whereas the simple w/z_y-questions differ, with inversion obligatory in English, but not in Italian. In both languages, complex questions require the inflected verb to raise to a position immediately following the w/z-phrase. If the Continuity Hypothesis is correct, therefore, Englishspeaking children may differ from English-speaking adults in the way they form simple wAy-questions, but they should parallel English-speaking and Italian-speaking adults in their long-distance w/z^-questions. From an experience-based perspective, this pattern is not anticipated. Since simple questions are more frequent in the input, these should become adult-like in advance of more complex questions. Until recently, only data corresponding to children's simple w/z-questions were available, so it was difficult to adjudicate between the experience-

Principles, parameters and probabilities

369

based account and an account invoking the Continuity Hypothesis. In an experimental and longitudinal diary study, however, Thornton (2004) recorded both simple and complex questions uttered by one child, AL, between the ages of 2 to 5 years. By age 3, AL was producing adult-like whquestions with w/z-phrases other than why, correctly carrying out subjectaux inversion 87% of the time between 2 and 3 years of age. In the same time period, why resisted inversion, with the inversion rate at only 40% for positive questions, and even lower for negative questions.6 The low rate of inversion for simple w/zy-questions persisted until AL was 5 and a half years old. (19) Why that boy is looking at us? Why the monster goed away and never corned back? Why the lights are on in my school?

(2;4) (3;3) (4; 10)

What about AL's complex Wz-questions? From the time AL was 3 years old until she reached 5;6, AL produced 83 complex w/z-questions. Sixtytwo w/z-questions contained words other than why, and all were adult-like, with inversion. Moreover, all but 4 of the 21 remaining w/z-questions, with why, were adult-like, with inversion. Some examples are in (20). (20) Why do you think you think Santa 's not coming this year? (3; 10) Why do you think mummy would not wanna watch the show? (4;6) What do you think is under daddy 's chair? (3 ;5) How do you think he can save his wife and her at the same time? (4;9) In short, the production data suggest that AL analyzed w/zy-questions like the corresponding questions are analyzed in Romance languages, such as perche in Italian. AL failed to invert in simple wAty-questions, but AL inverted in complex wAy-questions, where inversion is required only when the wAy-question is formed by long-distance wA-movement. In producing simple wAj-questions without inversion, moreover, AL was ignoring abundant evidence in the input, indicating a mismatch between her grammar and that of adult speakers in the same linguistic community. However, AL adhered to the grammatical principles that govern all natural languages, producing adult-like complex wAy-questions, but non-adult simple whyquestions. See Thornton (2004) for several further parallels between AL's wAy-questions and those of adult speakers of Italian; see Rizzi (2001) for an analysis of questions in Italian.

370 Rosalind Thornton, Stephen Grain and Graciela Tesan 5. Extra words in children's questions Another example of children's non-adult (but UG-compatible) productions is the medial-w/z phenomenon, which reveals a trace of Germanic in child English. Using an elicited production task, Thornton (1990) found that about one-third of the 3-4 year-old English-speaking children she studied consistently inserted an 'extra' w/z-word in their long-distance questions, as illustrated in (21) and (22) (also see Grain & Thornton (1998) and Thornton (1996) for a description of the experimental technique used to elicit longdistance w/z-questions from children). (21) What do you think what pigs eat? (22) Who did he say who is in the box? This 'error' by English-speaking children is presumably not a response to the children's environment, since medial-w/z constructions are not part of the primary linguistic data for children in English-speaking environments. However, structures like (21) and (22) are attested in dialects of German, as the example in (23) illustrates (from McDaniel 1986). (23) Wert glaubst du wer, nach Hause geht? who-NOM think-2SG you who-NOM towards house go-3SG 'Who do you think who goes home?' Further investigation shows that the similarity of child English to a foreign language runs deep. For both adult Germans and American children, lexical (full) w/z-phrases cannot be repeated in the medial position. For example, German-speaking adults judge (24) to be unacceptable, and English-speaking children never produced strings like (25), as indicated by the '#.' Instead, children reduced the w/z-phrase or omitted it altogether, as in (26). (24) * Wessen Buch i glaubst du wessen Buchj Hans liest? who-GEN book think-2SG you who-GEN book Hans read-3SG 'Whose book do you think whose book Hans is reading?' (25) #Which Smurfdo you think which Smurfis wearing roller skates? (26)

Which Smurfdo you think (who) is wearing roller skates?

Principles, parameters and probabilities

371

Finally, children never used a medial-w/? when extracting from infinitival clauses, so they never asked questions like (27). This is not permissible in languages that allow the medial-w/z either. (27) #Who do you want who to win? This complex pattern of linguistic behavior suggests, again, that many children of English-speakers go through a stage at which they speak a language that is like (adult) English in many respects, but one that is also like German in allowing for the medial-w/?. There is nothing wrong with such a language - it just happens that adults in the local community do not speak it. If the ways in which child and adult language can differ is limited to ways in which adult languages can differ from each other, then this would be compelling evidence in favor of the theory of Universal Grammar. On the account we envision, children's linguistic experience drives children through an innately specified space of grammars, until they hit upon one that is sufficiently like those of adult speakers around them, with the result that further data no longer prompts further language change. On this view, children's 'errors' are not simply failures to match adult input. Indeed, speaking of 'errors' here may have outlived its usefulness, except as a way of noting that a child's course of language acquisition - achieving the stable adult state - has not yet ended. In a more interesting sense, children are not merely speaking adult English badly; like monolingual speakers of Japanese, they are speaking a foreign language. An explanation is owed by advocates of the experience-based approach to language development as to why children project beyond the data in certain ways but not others.

6. Resetting a parameter for negation Another parameter concerns negation. This parameter is a choice between two structural positions in which a lexical item that expresses negation can reside. It can either be the 'head' of a phrase, or the 'specifier'. The choice between head or specifier interacts with the verb movement properties of the language. Roughly, negative expressions that are heads can move along with the verb to a higher structural position in verb-raising languages. But a verb cannot raise over a negative head in such languages. This contrasts with the status of negative expressions that are specifiers in a language; such expressions do not block the movement of a verb, which is free to

372

Rosalind Thornton, Stephen Crain and Graciela Tesan

raise or lower over expressions in specifier position. We will illustrate these parametric options with cross-linguistic examples. Spanish is a verb-raising language, and the negative item no is a head. Because the form of negation is a weak form, a clitic, it raises along with the verb to the INFL position, as shown in (28). In the example, the origins of the elements that have been moved are indicated by strikethrough. (28) [IP Juan no habla [NEG n& [Vp habla italiano]]] Juan NEG speak-3SG Italian 'Juan doesn't speak Italian.' In French, as in Spanish, the weak clitic form of negation (ne) is a head (although this form is often omitted in colloquial language). It also raises with the verb to INFL. But there is a second form of negation in French, namely the lexical item pas. The expression pas is obligatory in negative sentences and is positioned in specifier position. The example in (29) illustrates a sample derivation in which the main verb raises to pick up the negative element ne in the head position, and passes over pas as it raises to INFL. (29) [IP Jean ne-parle [NEG POS *e [Vp parlc grec ]]]

Jean NEG -speak not 'Jean doesn't speak Greek.'

Greek

Another language that positions negative expressions in the specifier position is Swedish. However, in Swedish, the verbal affix lowers over the negative item inte to merge with the main verb. This is most transparent in embedded clauses. The example in (30) (taken from Tesan 2005) illustrates the word order of Swedish in embedded sentences. (30) ...att Lena inte kopte en ny bok igär ...that Lena not bought a new book yesterday ' .. .that Lena didn't buy a new book yesterday. (adapted from Vikner 1995: 45) Turning to English, the negative expressions not and n 't are categorized as heads (cf. Chomsky & Lasnik 1993). It is useful to compare these expressions with the negative adverb never, which resides in the specifier position of the negation projection and, therefore, functions much like pas in French

Principles, parameters and probabilities

373

and inte in Swedish. Since never is a specifier, the verbal affix can lower across the adverb never and attach to the verb. This is illustrated in (31) where the 3rd person agreement affix 5 lowers onto the main verb to yield He never speaks French. (31) [IP He s [MEG never [Vp speak-s French ]]] The verbal affix cannot lower across the head not, however. This prevents the derivation of utterances like He not speaks French. A rescue operation, 'do-support', is needed to salvage the derivation. Insertion of do provides a host for the stranded affix that is prevented from lowering to the verb. This yields acceptable utterances such as He does not speak French. (32) [IP He doe -s [NEG not [ V p speak French ]]]

The most common form of negation in English is n 't. This contracted form of negation is a clitic that joins with a host auxiliary verb or modal (e.g., doesn 't, can't, haven't, isn 't etc.), as in (33) (cf. Zwicky 1983). (33) [IP He doe -s -n 't [NEC *& [VP speak French ]]] With this survey of parametric options in mind, let us see how the negation parameter is set in the grammars of English-speaking children. The fact that n't is attached to modals and auxiliaries (can't, shouldn't, haven't, isn't etc.) informs children that this form of negation is a head. However, the knowledge that n't is a head doesn't help children implement this value of the negation parameter in sentences with main verbs. Children must be exposed to the specific lexical item doesn't to see that n 't must remain higher than the verb phrase, so the negation parameter may not be set by children until they discover the lexical item doesn't.1 Until this time they may treat not as a specifier and produce he not speak french. Having established that doesn't constitutes unambiguous data for learners to set the negation parameter, we can use the frequency of occurrence of doesn't in the input to establish whether the parameter is expected to be acquired early or late. To obtain an estimate of the frequency of doesn't in the input, we conducted a search of the adult input to Adam and Eve in the CHILDES database (MacWhinney & Snow 2000). Of the 30,882 adult sentence utterances that were checked, only 296 (0.95%) contained doesn't. According to the statistical learning model proposed by Yang, parameters

374

Rosalind Thornton, Stephen Grain and Graciela Tesan

whose unambiguous input appears with a frequency of occurrence of 1.2 % or less are expected to be set late in the course of acquisition. The expectation is, therefore, that the negation parameter will be set late. To test this expectation, an elicited production experiment was conducted with four English-speaking children who attended our language acquisition laboratory starting at about age 2, and continued to visit the lab every two weeks for roughly a year, at which point the verbal morphology of all of the children was close to adult-like. The elicited production experiment encouraged children to use constructions that they would otherwise have avoided. This is particularly true of negative utterances, which are sparse in the CHILDES database. For example, Harris & Wexler (1996) searched the transcripts of 10 children who ranged in age ranged from 1;6 to 4;1, and found 52 sentences with 3rd person subjects in structures that contained no or not and a main verb (cf. H&W, Table 5, p. 16). Our study evoked 204 comparable structures from the four 2-year-old child subjects over a considerably shorter period of time (see Tesan 2005 for more details). Procedures to evoke negative sentences included a range of games to see, for example, where various objects would fit. For example, a puppet might try to complete a puzzle, but would end up putting pieces in the wrong place. The child was encouraged to correct the puppet ("It not goes there!"). In an alternative task setting, the child subject was assigned the task of performing 'experiments' with a group of various objects, to see if they float, or squeak, of would stick to a magnetic board, and so forth ("This one squeaks. This one not squeaks"}. The inclusion of these elicitation procedures resulted in a robust set of data for each child. One child, SL, initially selected the adult value of the parameter, with negation residing in head position. By contrast, the other 3 children began with the other value, with negation in Specifier position. This means that these 3 children treated not like never, and produced utterances like He not fits, with an inflected main verb. Since SL began with the adult value, no parameter resetting was needed, thus this child's pattern of behavior was essentially flat. The pattern of behavior across time was quite different for the other 3 children who initially adopted the non-target value of parameter. These children exhibited an abrupt change in values, with different children initiating and completing their own precipitous change at different ages. There was no indication that the statistical distribution of structures or lexical items in the input was responsible for the trajectories of any of the children. As illustration, Figure 1 provides the data for CM, one of the 3 children who initially mis-set the negation parameter, and treated not as a Specifier.

Principles, parameters and probabilities

375

CM initially used only this option in the session at 1;10;27, when she produced 3 instances of sentences like He not fits. Grammatical change was initiated almost immediately, and this treatment of negation rapidly disappeared, and was completely gone within 3 months. CM took heed of the input early, acquiring does n 't by age 2;1. Another interesting observation is that, at 1;11;11, CM produced 50% sentences like He not fits, in which negation was treated as a Specifier, and 50% adult-like sentences, with Josupport, in which n 't was behaving as a head. In fact, in the first half of the session, CM produced 6 utterances like He not fits and, in the second half of the session, CM produced 6 adult-like utterances. In other words, there was an abrupt change from one parameter value to the other - within a single testing session in the laboratory.

Age

x ..\SS* s;,·

s;.·

v

^

Figure 1. The trajectory for the two negation values in CM's data

Beginning with Saffran, Aslin & Newport (1996), the last decade has witnessed a series of research studies showing that children are endowed with learning mechanisms that are sufficiently powerful to assist them in word segmentation, and even in the detection of phrasal units (Saffran 2001, 2002). Yang (2002, 2004) has proposed that such learning mechanisms can be paired with Universal Grammar to assist language learners in setting parameters (see Roeper, this volume). Granting that learners employ a statistical learning mechanism for certain tasks, we investigated children's acquisition of a negation parameter, to see whether the learning path in child language development assumed the gradual curve associated with statistical learning over time or, instead, if the path of language development resembled the sharp edges associated with setting and resetting parameters. The

376

Rosalind Thornton, Stephen Grain and Graciela Tesan

empirical findings from a longitudinal study of four children's development of negation do not support the proposal that statistical learning is driving children's parameter setting. Our empirical findings show, instead, that (a) children initiate grammatical change at some point in time, (b) change takes hold quickly, and (c) change is brought to closure quickly. We do not fully understand the mechanisms that set grammatical change in motion, but they are apparently responsive to the child's internal grammatical development, and do not always reflect children's linguistic experience.

Notes 1. Roughly, a ^//-question is one that begins with a Wz-word: why, what, where, who, etc. But questions that begin with how or how come also qualify. 2. In contrast to the elicitation study, a search of CHILDES (MacWhinney & Snow 2000) reveals few contractions in questions, either by adults or by children. For example, there are just 7 occurrences of contracted wanna in the yes/no-questions posed to Adam by adults; nevertheless, Adam produced 50 yes/no-questions with wanna (21 % of the 237 questions that contained either contracted or uncontracted forms of want + to). Unfortunately, the entire record of Adam's spontaneous productions lacked any subject extraction wh-questions, so there is no evidence one way or the other regarding Adam's adherence to the relevant constraint. 3. An exception is how come (cf. note 5). 4. For example, the Brown corpus in the CHILDES database contains 475 whyquestions by children, and 699 by adults. Ninety-two percent of adult why-questions show inversion. Despite this, only 13% (62) of children's vWzy-questions show inversion. The remaining 87% of children's w/fy-questions either lacked a modal/auxiliary verb (20%), or contained a modal/auxiliary verb that was not inverted with the subject (67%). (NB: wAy-questions with me as the subject were excluded from the analysis, as were ones in which why was immediately following by not.) 5. An exception is come mai (how come in English; cf. footnote 3). 6. There are independent reasons for the lower proportion of inversion in negative questions (see Guasti, Thornton & Wexler 1995). 7. Of course, the discovery that n 't is a head still does not guarantee that the negative morpheme not is also a head; it could be a specifier. Therefore, children could use doesn 't in the same way as adults do but, at the same time, they could analyze not as a specifier. Our empirical findings suggest that once children acquire doesn't, they cease to use not, at least for a time. For now, we will simply assume that doesn't is the critical data that children need, and leave the continuing status of not in children's grammars as an open question.

Principles, parameters and probabilities

377

References Chomsky, Noam 1965 Aspects of the theory of syntax. Cambridge, MA: MIT Press. 1971 Problems of knowledge and freedom. New York: Pantheon Books. 1975 Reflections on Language. New York: Pantheon Books. 1980 Rules and Representations. New York: Columbia University Press. 1981 Lectures on government and binding. Dordrecht: Foris. 1986 Knowledge of language: Its nature, origin and use. New York: Praeger. 1991 Some notes on economy of derivation and representation. In Principles and parameters in comparative grammar, Robert Freidin (ed.), 417^54. Cambridge, MA: MIT Press. 1995 The minimalist program. Cambridge, MA: MIT Press. 2002 On nature and language. Cambridge: Cambridge University Press. Chomsky, Noam & Howard Lasnik 1993 The theory of principles and parameters. In Syntax: An international handbook of contemporary research, Joachim Jacobs, Arnim von Stechow, Wolfgang Sternefeld & Theo Vennemann (eds.). Berlin/ New York: Mouton de Gruyter. Cowie, Fiona 1999 What's within? Nativism reconsidered. New York: Oxford University Press. Grain, Stephen 1991 Language acquisition in the absence of experience. Behavioral and Brain Sciences 14: 597-650. Grain, Stephen & Mineharu Nakayama 1987 Structure dependence in grammar formation. Language 63: 522-543. Grain, Stephen & Paul Pietroski 2001 Nature, nurture and universal grammar. Linguistics and Philosophy 24: 139-186. Grain, Stephen & Paul Pietroski 2002 Why language acquisition is a snap. The Linguistic Review 19: 163183. Grain, Stephen & Rosalind Thornton 1998 Investigations in universal grammar: A guide to experiments of the acquisition of syntax and semantics. Cambridge, MA: MIT Press. De Villiers, Jill 1991 Why questions? In UMOP Special Edition: Papers in the Acquisition of Wh, Thomas Maxfield & Bernadette Plunkett (eds.), 155-175. Amherst, MA: GLSA. Goldberg, Adele 2003 Constructions: A new theoretical approach. Trends in Cognitive Science 7:219-224.

378

Rosalind Thornton, Stephen Grain and Grade la Tesan

Guasti, Maria Teresa, Rosalind Thornton & Kenneth Wexler 1995 Negation in children's questions: The case of English. In Proceedings of the Boston University Conference on Language Development 19, Dawn MacLaughlin & Susan McEwen (eds.), 228-240. Somerville, MA: Cascadilla Press. Harris, Tony & Kenneth Wexler 1996 The optional infinitive stage in child English: Evidence from negation. In Generative perspectives on language acquisition, Harald Clahsen (ed.), 1^42. Amsterdam/Philadelphia: Benjamins. Hyams, Nina 1986 Language acquisition and the theory of parameters. Dordrecht: Reidel. 1987 The theory of parameters and syntactic development. Parameter setting, Thomas Roeper & Edwin Williams (eds.), 1-22. Dordrecht: Reidel. Labov, William & Teresa Labov 1978 Learning the syntax of questions. In Recent Advances in the Psychology of Language HI:4b, Robin Campbell & Philip Smith (eds.), 1-44. New York: Plenum Press. Lasnik, Howard & Stephen Grain 1985 On the acquisition of pronominal reference. Lingua 65: 135-154. Legate, Julie & Charles Yang 2002 Empirical re-assessment of stimulus poverty arguments. The Linguistic Review 19: 151-162. MacWhinney, Brian 2000 The CHILDES project: Tools for analyzing talk. 3rd ed. Mahwah, NJ: Erlbaum. 2004 A multiple process solution to the logical problem of language acquisition. Journal of Child Language 31: 883-914. McDaniel, Dana 1986 Conditions on wh-chains. Ph.D. diss., City University of New York. Pinker, Steven 1984 Language learnability and language development. Cambridge, MA: Harvard University Press. Rizzi, Luigi 2001 On the position of 'lnt(errogative)' in the left periphery of the clause. In Current Studies in Italian Syntax, Guglielmo Cinque & Giampaolo Salvi (eds.), 287-296. Oxford: Elsevier. Rowland, Caroline & Julian Pine 2000 Subject-auxiliary inversion errors and wA-question acquisition: 'what do children know?' Journal of Child Language 27: 157-181.

Principles, parameters and probabilities

379

Saffran,Jenny 2001 The use of predictive dependencies in language learning. Journal of Memory and Language 44: 493-515. 2002 Constraints on statistical language learning. Journal of Memory and Language 47: 172-196. Saffran, Jenny, Richard Aslin & Elissa Newport 1996 Statistical learning by 8-month old infants. Science 274: 1926-1928. Tesan, Graciela 2005 What do children have in their heads? Functional heads and parameter setting in child language. Ph.D. diss., University of Maryland. Tesan, Graciela & Rosalind Thornton 2004 Overspecification in child language. In The Proceedings of GALA (Generative Approaches to Language Acquisition Conference, 2003), Jacqueline van Kämpen & Sergio Baauw (Eds.). LOT Occasional Series 3: The Netherlands. Thornton, Rosalind 1990 Adventures in Long-Distance Moving: The Acquisition of Complex Wh-questions. Ph.D. diss., University of Connecticut. 1996 Elicited production. In Methods for Assessing Children's Syntax. Dana McDaniel, Cecile McKee & Helen Smith Cairns (eds.). Cambridge, MA: MIT Press. 2004 Why continuity? In Proceedings of Boston University Conference on Child Language Development (BUCLD) 28, Alejna Bruges, Linnea Micciulla & Christine E. Smith (eds.), 620-632. Somerville, MA: Cascadilla Press. Tomasello, Michael 2000 Do children have adult syntactic competence? Cognition 74: 209253. 2003 Constructing a language: A usage-based theory of language acquisition. Cambridge, MA: Harvard University Press. Vikner, Sten 1995 Verb movement and expletive subjects in the Germanic languages. New York: Oxford University Press. Yang, Charles 2002 Knowledge and learning in natural language. New York: Oxford University Press. 2004 Universal grammar, statistics, or both? Trends in Cognitive Science 8: 451^56. Zwicky, Arnold & Geoffrey Pullum 1983 Cliticization vs. inflection: English n't. Language 59: 502-513.

Comments

The role of frequency in language acquisition Katherine Demuth

The role of frequency in language acquisition has largely been ignored, in part by early disclaimers by both psychologists and linguists. For example, Brown (1973), in his study of morphological development by Adam, Eve and Sarah, declared that the order of grammatical morpheme acquisition was not determined by frequency. Likewise, Chomsky (1965), in defining the goal of modern formal linguistics to focus on 'competence' rather than 'performance', suggested that the collection and analysis of actual speech samples be left to the field of 'sociology'. As a consequence, the possible role of frequency effects on the formation of linguistic representations has been largely missing from both formal linguistic and more psychological approaches to language learning. This has been consistent with a more categorical approach to the nature of grammatical representations, where sentences are either grammatical nor not. The field began to change in the late 1980s. With the increasing availability of computers, it was possible to collect and analyze large corpora of speech. This was especially relevant for the field of acquisition, since longitudinal spontaneous speech corpora provide much of the primary data on individual children's course of language development. It is during this time that contributions to the CHILDES database (MacWhinney 2000) also increased. A rough count (excluding recent Talkbank contributions) indicates that the database included 4 corpora in 1984, 104 in 1995, and 210 in 2006 (MacWhinney p.c.). Although each corpus varies in size, it is clear that many more researchers have access to various language acquisition corpora today. The increased availability of computerized corpora also coincided with developments in linguistic theory that began to change the notion of categorical 'rules' that were never violated. Optimality Theory (Prince & Smolensky 1993, 2004) provided the opportunity for thinking about surface output forms in terms of a constraint satisfaction problem. Under an optimal ity-theoretic view grammatical judgments assigned a '?' could be interpreted as 'acceptable' under certain (discourse, morphological) conditions. This notion of interacting constraints was particularly appealing to researchers of language acquisition, where change was documented over time. Thus,

384

Kathenne Demuth

rather than an abrupt parameter-setting the gradual learning curve could be understood in terms of a continual process of constraint reranking, gradually approximating the adult form. This constraint-reranking approach to acquisition turned out to be particularly fruitful for dealing with problems of higher-level phonological acquisition, where issues of syllable and word structure could be dealt with in terms of reranking constraints, with Markedness constraints initially ranked higher than Faithfulness constraints (e.g., Demuth 1995; Gnanadesikan 1995, 2004; Pater 1997). Some scholars have now taken this further, suggesting that the constraints should be 'weighted', with frequency information playing a role in determining weights and when/how they are reranked (e.g. Boersma & Levelt 1999). Appealing to frequency effects in syllable structure was particularly useful in understanding the course of acquisition of syllable structure in Dutch (Level, Schiller & Levelt 2000). It has also been proposed to help account for language-specific differences in the timing of when coda consonants are acquired (earlier in English and German than in Spanish) and in when initial weak syllables appear (earlier in Spanish than in English or German) (e.g., Roark & Demuth 2000). Thus, both languageinternal and cross-language developmental paths have been shown to be sensitive to the frequency of language-specific structures - at least in the domain of prosodic phonology. It has also been shown to occur not only in production, but also in perception, at the level of segments (e.g., Anderson, Morgan & White 2003). Thus, at least for phonology, there appears to be some effect of frequency, where the more frequent structures are earlier acquired. This raises the possibility that more frequent structures may be easier to be learned because they are typically also 'unmarked'. Stites, Demuth & Kirk (2004), however, show that markeness and frequency do not always coincide. For example, more sonorant consonants (e.g., /I, m., n/) are less marked than less sonorant consonants (e.g., /t, d, k, g, b, p/) in the coda of a syllable, but less sonorant /t, d/ are much higher in frequency as codas in English than are the sonorant consonants. This provides a context for teasing apart the markedness and frequency issues. The authors show that most children acquire the most frequent codas first, though a few acquire the more sonorant, unmarked codas first. This suggests that some individual variation may be accounted for by adhering to either frequency or markedness patterns. Zamuner, Gerken & Hammond (2004) take this further, showing that frequency effects within the syllable rhyme are better at predicting order of acquisition. This again raises the question of what units the learner (and linguist) are computing over.

The role of frequency in language acquisition

385

But why might children acquire more frequent structures first? One possibility is that the more frequently a certain linguistic unit occurs, the harder it is for the learner to ignore it. That is, from an information-theoretic perspective, if it occurs often, it must be attended to, it is not 'noise' in the system. This is easy to capture in optimality theoretic terms. Imagine a young learner who produces no codas, such that dog is realized as /da/. If 60% of the syllables the child hears contain a coda consonant, this means that the majority of the syllables the child utters will be ill-formed, resulting in massive Faithfulness violations. Since this is a lot of ill-formedness, the English-speaking child will probably be faster to start producing coda consonants than a child learning Spanish, where only 25% of syllables have codas. Thus, we predict that, for most English learners, the acquisition of complex syllable structure will proceed more rapidly than it does in Spanish, and this appears to be the case (cf. Demuth, Culbertson & Alter 2006; Lleo 2003). Thus, perhaps the high frequency of a particular form merely forces the child to attend to a particular grammatical structure earlier than they otherwise might. This is known in the adult psycholinguistics literature as a 'priming' effect (e.g., Bock 1986; Bock & Loebell 1990), which has been shown to effect adults' use of syntactic constructions. That is, if adults are previously exposed to a certain syntactic construction, they are more likely to use it in subsequent speech. This raises the possibility that frequency effects occur not only in children's developing phonologies, but also in their syntax. A number of results suggest that this is the case. One of the first studies to suggest that frequency played a role in children's syntax came from Demuth's (1989) study of passive acquisition in the Bantu language Sesotho. This study showed that passives are much more frequent in Sesotho child-directed speech than they are in English. Demuth argued that this much higher frequency of Sesotho passives could help explain why passives are acquired in this language by the age of 2;8. Furthermore, this also predicted that even English-speaking children could show competence with such grammatical structures under appropriate priming conditions. Brooks & Tomasello (1999) have recently shown that this is the case, eliciting passives from 3-year-olds with novel verbs. These findings are not surprising for those who work on adult psycholinguistics, but they have been less routinely explored in the domain of language acquisition. The frequency of different grammatical structures is part of our linguistic competence, and influences not only how we process and produce language, but also which aspects of language will be learned first, and which aspects are most likely to change (e.g., Bybee & Hopper 2001).

386

Katherine Demuth

We should therefore be able to make frequency-based predictions about when certain grammatical structures might appear in different languages all else being equal. But all else is not always equal. For example, French also has codas in only 25% of syllables, but coda consonants appear to be acquired earlier than in Spanish. However, coda consonants are more likely to be preserved in stressed rather than unstressed syllables and in monosyllables rather than disyllables - at least by English speaking children (Demuth, Culbertson & Alter 2006; Kirk & Demuth 2006). Thus, frequency alone cannot explain this cross-linguistic difference. We suggest that, in this case, the longer average word length of Spanish, plus the stress differences between the two languages, account for the differences here. This again raises the problem how to count. That is, are learners (which we DO assume are keeping track of all kinds of statistics) computing statistics over tokens (as often assumed in the acquisition literature) or types? And what are the linguistic units over which they are keeping these statistics? Recent modeling work indicates that in the domain of morphological segmentation, type information is more useful than token information (Goldwater 2006). Obviously, further modeling and empirical research is needed to determine what learners are counting, how they are counting, and how this triggers perceptual awareness and comprehension/product! on competence in different linguistic domains. All these findings suggest the need for an intergraded model of language acquisition, where competing constraints (syntactic, semantic, discourse, phonological, prosodic, processing, frequency effects, etc.) all play a role in determining how and when which aspects of language are acquired. The frequency with which a particular grammatical construction occurs in a child's experience is therefore also part of that child's knowledge of language. Only when all these other constraints are controlled for can we really look at 'syntactic' competence in isolation. Syntactic experiments are much better designed today than they were in the past, controlling for more of the possible discourse/pragmatic, semantic, and processing confounds. Researchers are also more aware of processing constraints, controlling for sentence length and complexity. Some also take lexical frequency into account, ensuring that children know the vocabulary items to be used. Less controlled are prosodic effects, which play a significant role in when and how certain grammatical function items (such as determiners) are acquired (e.g., Gerken 1996; Demuth 2007; Tremblay & Demuth 2007). Frequency is thus only one of the factors that can influence when and how learners demonstrate knowledge of grammar.

The role of frequency in language acquisition

387

References Anderson, Jennifer, James L. Morgan & Katherine S. White 2003 A statistical basis for speech sound discrimination. Language and Speech 46: 155-182. Bock, Kathryn 1986 Syntactic persistence in language production. Cognitive Psychology 18:355-387. Bock, Kathryn & Helga Loebell 1990 Framing Sentences. Cognition 35: 1-39. Boersma, Paul & Clara Levelt 1999 Gradual Constraint-Ranking Learning algorithm predicts acquisition order. In Proceedings of the 3tfh Child Language Research Forum, Eve V. Clark (ed.), 229-237. Stanford: The University of Chicago Press (CSLI: Center for the Study of Language and Information). Brooks, Patricia & Michael Tomasello 1999 Young children leam to produce passives with nonce verbs. Developmental Psychology 35: 29—44. Brown, Roger 1973 A first language: The early stages. Cambridge, MA: Harvard University Press. Bybee, Joan & Paul Hopper (eds.) 2001 Frequency and the emergence of linguistic structure. Amsterdam: Benjamins. Chomksy, Noam 1965 Aspects of the theory of syntax. Cambridge, MA: MIT Press. Demuth, Katherine 1989 Maturation and the acquisition of Sesotho passive. Language 65: 56-80. 1995 Markedness and the development of prosodic structure. In Proceedings of the North East Linguistic Society 25, Jill Beckman (ed.), 1325. Amherst, MA: GLSA, University of Massachusetts. 2007 Acquisition at the prosody-morphology interface. In Proceedings of the 2nd Conference on Generative Approaches to Language Acquisition North America (GALANA), Alyona Belikova, Luisa Meroni & Mari Umeda (eds.), 84-91. Somerville, MA: Cascadilla Proceedings Project. Demuth, Katherine, Jennifer Culbertson & Jennifer Alter 2006 Word-minimality, epenthesis, and coda licensing in the acquisition of English. Language and Speech 49: 137-174. Gerken, LouAnn 1996 Prosodic structure in young children's language production. Language 72: 683-712.

388 Katherine Demuth Gnanadesikan, Amalia 2004 Markedness and faithfulness constraints in child phonology. In Constraints in phonological acquisition, Rene Kager, Joe Pater & Wim Zonneveld (eds), 73-109. Cambridge: Cambridge University Press. Goldwater, Sharon 2006 Nonparametric Bayesian Models of Lexical Acquisition. Unpublished Ph.D. diss., Brown University. Kirk, Cecilia & Katherine Demuth 2006 Accounting for variability in 2-year-olds' production of coda consonants. Language Learning and Development 2: 97-118. Levelt, Clara, C., Niels O. Schiller & Willem Levelt 2000 The acquisition of syllable types. Language Acquisition 8: 237-264. Lleo, Conxita 2003 Prosodic licensing of codas in the acquisition of Spanish. Probus 15: 257-281. MacWhinney, Brian 2000 The CHILDES Project: Tools for analyzing talk. 3rd Edition. Vol 2: The Database. Mahwah, NJ: Erlbaum. Pater, Joe 1997 Minimal violation and phonological development. Language Acquisition 6: 201-253. Prince, Alan & Paul Smolensky 1993/ Optimality Theory: Constraint Interaction in Generative Grammar. /2004 Cambridge, MA: B lackwell. Roark, Brian & Katherine Demuth 2000 Prosodic constraints and the learner's environment: A corpus study. In Proceedings of the 24th Annual Boston University Conference on Language Development, S. Catherine Howell, Sarah A. Fish & Thea Keith-Lucas (eds.), 597-608. Somerville, MA: Cascadilla Press Stites, Jessica, Katherine Demuth & Cecilia Kirk 2004 Markedness versus frequency effects in coda acquisition. In Proceedings of the 28th Annual Boston University Conference on Language Development, Alejna Brugos, Linnea Micciulla & Christine E. Smith (eds.), 565-576. Somerville, MA: Cascadilla Press. Tremblay, Annie & Katherine Demuth 2007 Prosodic licensing of determiners in children's early French. In Proceedings of the 2nd Conference on Generative Approaches to Language Acquisition North America (GALANA), Alyona Belikova, Luisa Meroni & Mari Umeda (eds.), 426-436. Somerville, MA: Cascadilla Proceedings Project. Zamuner, Tania, LouAnn Gerken & Michael Hammond 2004 Phonotactic probabilities in young children's speech production. Journal of Child Language 31:515-536.

Counting grammars Charles Yang

1. Pedestrian intuition Language learning is a matter of imitating parents, says the pedestrian on the street. Among them is apparently "Adam" (Brown 1973), the boy that has taught us much about language acquisition: when a documentary filmmaker quizzed him about how children learn languages, Adam said, "By imitating their mothers, I guess." Frequency effects is simply the professional's preferred term for this pedestrian intuition: the greater the exposure, the better the imitation. At some level, to say that there are frequency effects in language learning is perhaps close to stating a truism: I don't know how else a two-year-old would utter "mama" but not "data" - other than the plain fact that he hears a lot more "mama" than "data". The problem is, though, learning by imitation - and by implication, frequency effects - cannot be the whole story: after all, Adam taught us language acquisition through his "Where go?", "I like hit ball", "He should have holded his horses", and other innovations, as every consumer of the CHILDES database quickly finds out. And we have been busy purging learning via imitation from the thoughts of linguistics undergraduates ever since. Frequency effects, as I shall argue in this chapter, are part of the collateral damage when theories replaced intuitions in language acquisition research. As the very existence of the present volume illustrates, it is of considerable interest to study the role of frequency effects in language and language learning. The issue at hand is in fact broader in scope: the interpretation and integration of quantitative factors into linguistic theories has becomes an important - and controversial - research topic in recent years. As I suggested elsewhere (Yang 2002, 2004, and more directly Yang in press), much of the contention can be attributed to a lack of historical perspective on the development of modern linguistics, compounded with specific misunderstandings of the technical tools in the quantitative study of language. I submit that these confusions spill over into the study of Ian-

390

Charles Yang

guage acquisition as well - on the part of both advocates and critics of generative grammar - engendering what appears to be a unnecessary dispute over frequency effects. My discussion of frequency effects in acquisition is built around the unifying theme of computation: without a concrete model of language learning mechanisms, no proper assessment of frequency effects is possible. Moreover, learning mechanisms can, and perhaps should, be distinguished from the representations of linguistic structures: they play distinct roles in the explanation of child language and the present controversy over frequency effects stems from the failure to draw such a distinction. I conclude with a review of frequency effects under generative models of grammar, provided that an adequate model of learning is in place.

2. Frequency and learning Although the present volume focuses mostly on the acquisition of syntax, some of the most robust findings of frequency effects come from the development of speech, phonology and the lexicon (see deBoyssyon-Bardies 1999 and Yang 2006 for reviews). Babbling generally starts with the canonical form of CV alternations along with what appears to be a fairly universal set of phonetic inventories (MacNeiledge & Davis 1993; Locke 1993), but babies soon start using the statistically dominant vowels and syllabic structures specific to the native language (de Boysson-Bardies etal. 1989; de Boysson-Bardies 1993). Infants are equally adept at extracting the prominent lexical stress patterns, which can be used as cues for word segmentation (Jusczyk et al. 1993). In lexical development, at least in some languages, the most frequent lexical category (e.g., nouns, verbs) tends to form the majority of the early vocabulary (Tardif et al. 1997). Humans - and indeed, many other species - are fantastic machines in tune with the statistical patterns in the environment (Gallistal 1990), of which frequency effects in language learning are just one of the many manifestations. But frequency effects in language learning are selective: obviously not all frequent patterns are kept track of, and one naturally wonders what delimits the learner's choices (see Yang [2004] for a case study of statistical learning in word segmentation). In addition, frequency effects are often only part of the story. For instance, and keeping to the above discussion of lexical development, Naigles & Hoff-Ginsburg (1998) find that the input frequency is a strong predictor of the acquisition of verbs, but the diversity

Counting grammars

391

of syntactic frames in which verbs appear is also extremely important. Such complications are likely to multiply quickly once we move to more complex linguistic levels such as morphology and syntax, where the combinatorial power of language and interaction of multiple linguistic components - and the subtleties and limitations of frequency effects - are under a better light. Frequency effects are statements about Ε-languages (of children and adults), which are only partial extensions of the grammar, or the I-language (Chomsky 1986). The grammar is far more than the totality of utterances in language use, however large the sample may be. After all, "colorless green ideas sleep furiously" is to draw the distinction between the knowledge of the grammar and the attestation of the utterance. The argument from the Poverty of Stimulus (Chomsky 1975) pushes this demonstration further, highlighting the presence of linguistic knowledge despite the absence of the relevant linguistic experience. The richness and complexity of I-language knowledge is fundamentally beyond the scope of Ε-language, and by logic, any conclusion that can be drawn from frequency effects. Not everything that can be said will be said, and there are plenty of things that will never be said - Universal Grammar forbids - that are nevertheless part of the linguistic knowledge. When frequency effects are found, the question still remains whether they are due to general linguistic/cognitive factors or to the statistical properties of the particular language under learning. For instance, Rus (in preparation) explores the crosslinguistic acquisition of morphosyntactic structures. A surprising result emerges from this work that, contrary to the earlier claims of Brown (1973), the development of the English morphemes does show frequency effects. But the implications of this finding will not be understood unless one has a specific model of how children acquire morphologies, one which includes the problem of morpheme segmentation, the identification of the underlying form, etc. Only then can we know whether the frequency effects in English are genuine, and only then can one generalize to the study of other languages as to understand whether the order of morpheme acquisition is invariant and universal (Pye 1979; James & Khan 1983). Yet to the best of my knowledge, no such model presently exists: frequency effects, then, only add to the puzzle of language acquisition. Another complication arises when frequency effects are found, as is often the case, in co-existence with non-frequency effects. Again we need a specific model of learning that can tell us what shows up when. To give a specific example, consider one of the best known frequency effects in language acquisition: that high frequency irregular verbs in English are learned

392

Charles Yang

better than low frequency irregular verbs when measured by the rate of overregularization (Marcus et al. 1992; see Yang 2002 for a more detailed look). The problem, however, is that the acquisition of the regular verbs does not show frequency effects (Prasada et al. 1990), and indeed, the lack of frequency effects is generally taken as a diagnostic for morphological regularity (Clahsen 1999).1 So yet again, a model of morphological learning is required to explain frequency effects and lack thereof, one which includes a component that recognizes the productivity of morphological processes that "add -d" is productive while other forms are not - and then proceeds to store the irregulars accordingly in a frequency-sensitive manner, but prevents that of regular forms. Again, no such model is on the offer.2 Frequency effects, then, raise plenty of questions - perhaps more than they answer. They are statements of correlations rather than causations: it is a cliche to say that the former does not imply the latter. While the discovery of frequency effects is potentially important and illuminating, it is only the beginning. The remaining part is to develop a precise model of learning that facilitates the mapping between adult and child languages - which involves more than numerical tabulations of the data. 3. Frequency and Universal Grammar The perception that frequency effects need to be reconciled with generative grammar requires some commentary. The earlier proponents of generative grammar did not disregard the quantitative use of data in linguistic theorizing. In Logical Structure of Linguistic Theory (1955/1975), widely viewed as the founding document of generative grammar, Chomsky outlines a broad framework in which the grammar is measured against the linguistic corpus in an information-theoretic sense. Specific proposals are made as to how the input data is mapped into the grammar through the means of distributional analysis.3 And as Chomsky remarks in Aspects (1965: 61 f.), the goal of explanatory adequacy must be complemented with the feasibility requirement: a theory of grammar that contains "hypotheses compatible with fixed data to be 'scattered' in value, so that choice among them can be made relatively easily." The design of Universal Grammar is to keep the quantitative aspects of learning in mind. The uneasy feeling about frequency effects, I believe, has less to do with the nature of Universal Grammar, and more to do with the nature of learning (Yang 2002, 2006). To put it bluntly, several prominent approaches

Counting grammars

393

in the generative tradition (see Clahsen 1996 for a summary) have failed to adequately address the balance between the representation of linguistic knowledge and the mechanisms of learning. And they have left no wiggle room for quantitative considerations such as frequency effects. For a good while now, language acquisition research has been dominated by two competing approaches that appeal to, respectively, competence and performance factors to account for the divergence of children's language from adults'. But they are united in denying experience-based learning as an explanatory force. The performance-based approach turns to processing, memory, pragmatics, and other perceptual/cognitive constraints (Valian 1991; Weissenborn 1992; Bloom 1993), which presumably develops independent of linguistic experience. The competence approach, which has uncovered many informative patterns in children's language, has in recent years been relying on the postulation of distinct grammatical systems between children and adults (Wexler 1994, 1998; Rizzi 1994). Both approaches, then, relegate the burden of explanation to the ill-understood problem of biological maturation: the choice is either the cognitive system at large or the linguistic system in particular. And both approaches claim that the task of learning - that is, the mastery of the specific features of the native language - is complete: child's competence is identical to the adult's (Pinker 1984; Bloom 1993), or the setting of parameters is done correctly very early, even prior to the multiword stage (Poeppel & Wexler 1993; Wexler 1998). While many elements of grammar are undoubtedly on target from early on (Brown 1973; Pierce 1993), it is a leap of faith to make a blanket statement about the integrity of the entire grammatical system - even if we are only concerned with the core parameter system (see Section 4 for additional discussion). Moreover, these assertions are made in the absence of a working model of learning mechanisms - one that explains how the child's grammar gets to be like the adult's, or how the parameters are set correctly. How do we brush up this picture to accommodate frequency effects? One approach is to modify the assumptions about linguistic structures, or to abandon Universal Grammar altogether, replacing it with inductive learning mechanisms in conjunction with other cognitive and cultural constraints (Tomasello 2003; MacWhinney 2004). It is also believed that by directly introducing a probabilistic component to learning, the task of language acquisition is simplified (Abney 1996; Bod et al. 2003); presumably, the role of Universal Grammar would be weakened. Such optimism is misplaced, resulting in part from misunderstandings of formal learning theory. The

394

Charles Yang

technical discussion of learnability can be found elsewhere (Yang in press, and in particular, Niyogi 2006). The main result that has emerged from formal learning theory is that no successful/feasible learning can take place unless the hypothesis space is constrained in some finite fashion, which comports well with the linguist's approach to Universal Grammar. This result holds independently of the specific learning algorithm: that is, no amount of semantic information, cultural learning, or conservative generalization can substitute for the constrained hypothesis space. Empirically, the motivation for the renewed interest for inductive learning has not gone unchallenged; see Tomasello (2000), Fisher (2000), Lidz et al. (2003) for recent discussion. And let us not forget that the assessment of children's linguistic knowledge may require more than frequency effects and other textual analysis - children may know more than they say - as the classic study by Shipley et al. (1969) already made clear.4 The alternative, keeping to the division between knowledge and learning, is to modify the traditional view of acquisition mechanisms. Again, a look at history proves instructive, for the connection between the quantitative and formal aspects of grammar has been extensively explored before. The variationist study pioneered by Labov (1969), after all, is to attach both linguistic and social conditions to formal grammars - in that study, the SPE system - with probabilistic variables that describe the linguistic knowledge of speakers embedded in linguistic communities. The goal of variational analysis, as David Sankoff (1988) put it, concerns the "distributions of discrete linguistic choices." No formal linguistic model that I know of prohibits a probabilistic enhancement, and thus no formal linguistic model inherently precludes quantitative factors such as frequency effects. In this next section, we turn our attention to a particular linguistic model, that of syntactic parameters. 4. Frequency and parameters In the fallout of frequency effects, the most unfortunate victims are the syntactic parameters. A highly successful tool for the description of the linguistic diversities around the world, the parameters presently play no role in the explanation of language acquisition - a task set out at the conception of the Principles and Parameters framework (Chomsky 1981). Parameters certainly have seen better days, when the idea of triggering (Gibson & Wexler 1994) was viewed as a viable model of learning. Under triggering, the learner is identified with a specific parameter setting. De-

Counting grammars

395

pending on the how the current grammar fares with the input data, the learner may modify some parameter values and obtain a new grammar. Hyams's ground-breaking work was an attempt to link null subjects in child English ("Tickles me") to the pro-drop grammar such as Italian (1986) or topic-drop grammar such as Chinese (1992), for which missing pronoun subjects are grammatical. However, quantitative comparisons with Italian and Chinese children, whose frequencies of pronoun use are close to Italian and Chinese adults (Valiant 1991; Wang et al. 1992), reveal irreconcilable differences. Furthermore, if the child changes her grammar by changing parameter values, one expects dramatic - at least noticeable changes in the learner's language use. Unfortunately, these predictions have not borne out: for instance, subject drop disappears only gradually (Bloom 1993). The idea of triggering is further undermined under computational considerations: convergence to the target grammar cannot be guaranteed, as shown by both formal (Berwick & Niyogi 1996) and numerical analysis (Sakas & Fodor 2001). Rather than relegating parameters - knowledge - to the background, we advocate a reworking of parameter learning. In an approach dubbed variational learning (see Kroch 1989; Clark 1992; Roeper 2000 for related work), we view the child's language as a population of hypothesis whose composition changes during the course of learning. This population is an (innate) space of syntactic parameters provided by UG. Each grammar, or more specifically each parameter, is associated with a probability, and it is this probability distribution that changes adaptively to the linguistic data in the environment. Schematically, variational learning works as follows: (1)

For an input sentence s, the child a. with probability Pi selects a grammar G, b. analyzes s with G, c. - if successful, reward G, by increasing p, - otherwise punish Gj by decreasing p,

It is notable that this class of learning models is literally a throwback to the study of animal learning, which used to dominate the American scene of psychology (Bush & Mosteller 1951). The model outlined in (1) is equally applicable to an animal evaluating multiple choices of natural resources or behavioral strategies, where the probabilities/weights signify the relative success of these choices. Species ranging from rodents to fish to birds have demonstrated such capacity of learning (Herrstein & Loveland 1975; Gallistel 1990). The application here necessitates a sharp division between the

396

Charles Yang

mechanisms of learning, which is clearly a general process with ancient evolutionary underpinnings, and the objects of learning: the syntactic parameters that are in the unique possession of the human language learner. The basic learnability result is straightforward (see Yang 2002 for details). Clearly, the target grammar, by being consistent with the input data, will never be punished. All other grammars in the UG space, however, are at least inconsistent with some portion of the input data. Thus, non-target grammars will necessarily be driven to extinction by the target grammar, ensuring convergence. As a concrete example, consider the parameter of verb raising to tense. In the variational model, the learner initially has probabilistic access to both the + and the — value of the parameter. In a French-speaking environment, however, the - value will be punished. This is not to say that it will be punished all the time. For instance, a sentence such as "Jean voit Marie" obviously is consistent with both values of the parameter: in other words, the child learner will succeed regardless whether she has selected the + or - value to analyze this sentence. When positional markers are present, however, as in the case of (2)

Jean voit souvent Marie. Jean sees often Marie 'Jean often sees Marie.'

where the adverb follows the tensed verb, only the + value will succeed in analyzing the sentence. Hence, if the learner has probabilistically selected the + value for syntactic analysis, it will result in an increment of the probability associated with +. On the other hand, if the - value is selected, its failure in analyzing the sentence decreases its associated probability, which increases the probability of the + value as well. The probabilistic nature of the variational model has a number of features that distinguish it from the traditional conception of parameter setting. First, unlike triggering, even unambiguous evidence such as (2) does not settle learning decisively, but only nudges the learner toward the target value. The rise of the target grammar is gradual, as its probability gradually approaches 1; this appears to be characteristic of language development in general. Second, the demise of non-target grammars is also gradual: These will be accessed with decreasing probabilities as they are gradually driven out by the target grammar. This leads to a principled interpretation of "errors" in child language, as potential adult grammars sanctioned by UG thus bring the variational model in line with the guiding principle of the Continuity Hypothesis (see Yang [2006] for extensive discussion and Grain &

Counting grammars

397

Thornton [1998] for similar views and results). More pertinent to the present discussion, the variational model makes it possible to integrate the theory of parameters into a quantitative model of language learning, with frequency effects as a natural by-product. This is achieved by tying grammars with fitness values, similar to models of natural selection in biological evolution. A useful measure is the probability that a grammar is penalized in a specific linguistic environment, i.e., the percentage of sentences in the input that the grammar is inconsistent with. Adapting the formulation of Bush & Mosteller (1951), we have: (3) The penalty probability of grammar G, in a linguistic environment E is

C|=Pr(Gj-/»s| seE)5 Ceteris paribus, the speed with which a grammar (or a parameter value) rises to dominance is correlated with its competitor's penalty probability. More formally, consider two grammars, target G, and the competitor G 2 , with c, = 0 and C2 > 0. At any time, Pi+p 2 = 1 · With the presentation of each input sentence, the expected increase of ρ,, Ε[Δρ,], can be computed as follows (see Yang (2002) for details of the learning model): (4) Ε[Δρ,] - ρ,γΟ-ρ,Η with Pr. pi, Gi is chosen and G, —> s c + w m Γ P20~ 2)(~Y)Pi i Ρ · P20~ C 2)» ^2 is chosen and G 2 —> s p 2 c 2 y(1-p 1 ) with Pr. p2c2, G 2 is chosen but G 2 -/> s = C 2 y(1-p,) That is, the increase of the probability of the target grammar (p,) is correlated with the penalty probability of the competitor (C2), which can be directly estimated from corpus studies of child directed speech. (Note, however, that these fitness measures are not statistics that the child learner needs to explicitly keep track or make use of.) This, then, allows one to make quantitative comparisons for the development of different aspects of syntactic development. In Table 1 , we summarize the results from Yang (2004). The crucial data for the correct setting of these parameters - such as the discussion of verb raising in (2) - is largely self-evident.

398

Charles Yang

Table 1. Frequency effects of parameter setting. (References: a. Brown 1973; b. Pierce 1992; c. Valian 1991, Wang et al. 1992; d. Lightfoot 1999, Yang 2002; e. Clahsen 1986, Yang 2002; f. Thornton & Grain 1994, McDaniel et al. 1995.) Parameter Wh fronting Verb raising Obligatory subject Verb second Scope marking

Target Language

Evidence

Input

Time of acquisition

English French English German/Dutch English

Wh questions verb adverb (2) expletive subjects d

30% 7% 1.2% 1.2% 0.25 %

Very early3 l;8b

ovs

Long-distance Wh questions

3;0C 3;0-3;2e 4;0+f

The quantitative study of parameter setting explains just when frequency effects are expected and when they disappear. Again, null subjects provide a perfect illustration. Despite an overwhelming amount of input data containing overt subjects, only the relatively rare expletive subjects (1.2% in Table 1) serve to rule out the topic drop option. This particular property of UG is independently established through the study of crosslinguistic variations (Huang 1984; Rizzi 1986); the acquisition data, when interpreted under a suitable framework of learning, provides the additional confirmation. In what follows, we present an account of the well-known Root Infinitives (RI) phenomenon developed recently in Legate & Yang (2007), which also illustrates the methods and benefits of incorporating frequency effects in the generative approach to acquisition. Despite the considerable effort in recent years (Rizzi 1994; Wexler 1994, 1998; Hoesktra & Hyams 1998), there are several aspects of RI that have not received adequate theoretical considerations. First, RI is optional: children use a mixture of RI, decreasing in frequency, as well as finite verbs. This suggests that RI cannot be due to a categorical discontinuity in the grammar. Second, crosslinguistic distribution of RI is gradient (Guasti 2002): in the acquisition of languages such as Italian, Catalan, and Spanish, there is a very short RI stage if at all, whereas English, Dutch, and German children have a much longer period of RI use, sometimes extending over three years. This suggests that RI cannot be due to performance or maturational constraints, which would presumably affect all populations equally. Finally, the distribution of RI is grounded in the verbal morphology of the target language: the propensity and duration of RI across languages is strongly correlated with the morpho-

Counting grammars

399

logical "complexity" of the verbal system (Phillips 1995). This suggests that morphological learning must play a crucial - and quantitative - role in the explanation of RI. We approach the RI problem in the general framework of variational learning with two specific suggestions. First, the optionality and gradual decline of RI reflect the presence of a grammar - call it a [-T] grammar such as Chinese, which does not manifest tense marking. A [-T] grammar typically uses aspect for temporal anchoring (Enc 1987), which may explain the aspectual interpretation of RI in child language (Hoestra & Hyams 1998; Becker 2000). Second, the elimination of the [-T] grammar is facilitated by the learning of the morphosyntactic system of the [+T] target language. This is where quantitative differences in morphological complexity lead to the crosslinguistic variations of RI in acquisition. When the tense marker is null - such as "I run" in the English present tense - the [-T] grammar will be rewarded. We conjecture that the [+T] option may be punished, for it requires the postulation of a phonologically null morpheme for present tense, which, for independent reasons, must be dispreferred. What kind of morphological patterns would punish [-T] grammar, thus driving the learner toward the target? Here both primary and secondary exponence of tense features must be taken into account (e.g., Harley & Noyer 1999). Obviously, any overt marking of tense (e.g., English past tense morpheme "ed") pushes the learner toward [+T]. However, [+T] gains prominence in morphological environments that do not mark but nevertheless implicate the tense feature. For example, the English third person singular maker ("s") does not contain an overt morpheme for tense, but surfaces in the present tense. Both kinds of morphological data would contribute to the elimination of the [-T] grammar. We present corpus studies of child-directed speech of Spanish, French, and English (MacWhinney 1995). We measure the quantity of morphological evidence for [+T] as the difference between the percentage of input sentences that unambiguously implicate [+T] (i.e., the primary and secondary exponence discussed above), and the percentage of input sentences that favor [-T] (e.g., "I run"). The results are summarized in Table 2.

400

Charles Yang

Table 2. Frequency effects in Root Infinitives. (References: a. Gristead 1994; b. Pierce 1992; Rasetti 2003; c. Phillips 1995; Hoekstra & Hyams 1998.) Language Spanish French English

([+T] - [-T]) / %

RI duration

54.4% 39.8% 8.8%

~2;0a ~2;8b >3;5C

It is evident that the quantitative evidence for the [+T] grammar correlates strongly with the duration of the RI stage in these languages. Taken together, we have findings which suggest that parameters are more than elegant tools for syntactic description; their psychological reality can be observed by the developmental correlates in children's grammar, that the learner is sensitive to specific types of input evidence relevant for the setting of specific parameters. On the basis of corpus statistics, this line of thinking can be used to quantify the argument from the Poverty of Stimulus (Legate & Yang 2002). Under variational learning, the proponents of generative grammar can no longer be accused of ignoring the input - and frequency effects no longer require "reconciliation" with Universal Grammar. 5. Frequency redux The conventional view of frequency effects needs to be broadened. Language learning can be observed as a mapping from one set of linguistic data (the input) and another set (the output), yet one needn't subscribe to the view that this mapping is transparent - the vestige of learning by imitation. If the learning mechanisms themselves are sensitive to the quantity of linguistic data, the linguistic representations provided by Universal Grammar can be viewed as the sieves that allow for frequency effects to emerge but also provide an explanation of when frequency effects break down. Pursuing this line of thinking, frequency effects may be used as an additional constraint in the design of UG, as Chomsky suggested in Aspects; more directly, they may provide a novel tool for the evaluation of linguistic theories. The conclusion of our discussion is that under the regulation of UG, the mechanisms of language learning may be quite simple and not unique to language - recall that even rodents are variational learners - with ample room to accommodate frequency. Learning is a probabilistic process in

Counting grammars

401

which the learner matches the hypotheses on the outside against the hypotheses from the inside. We can even maintain the intuition that language acquisition is a matter of children imitating adult language, as long as the language comes not with an E- but an I-.

Acknowledgement Much of the language acquisition work reported here is done in collaboration with Julie Legate, whose contribution is gratefully acknowledged.

Notes 1. We do not consider the storage of regular forms, which presumably would show frequency effects, prior to the emergence of the "add -d" rule - the initial part of the so-called U-shape learning curve (Marcus et al. 1992). 2. See Yang (2005) for additional discussion. 3. Ironically, some of these very proposals were apparently rediscovered decades later, on the charge that quantitative considerations had been systematically excluded from generative enterprise; see Bod et al. (2003), Yang (in press). 4. We do not wish to imply that inductive learning plays no role in language acquisition. While we believe the core parameter system is provided by UG, the periphery system, specifically the partial generalizations over specific lexical items, are undoubtedly learned inductively under suitable linguistic and learning constraints; see Yang (2005) for preliminary discussion of these issues. 5. We write s e E to indicate that s is an utterance in the environment E, and G —> s to mean that G can successfully analyze s. Formally, the success of G —>· s can be defined in any suitable way, possibly even including extra-grammatical factors; a narrow definition that we have been using is simply parsability.

References Abney, Steven 1996 Statistical methods and linguistics. In The balancing act, P. Resnick & J. Klavins (eds.), 1-26. Cambridge, MA: MIT Press. Becker, Misha 2000 The development of the copula in child english: The lightness of Be. Doctoral diss., UCLA.

402

Charles Yang

Berwick, Robert C. & Partha Niyogi 1996 Learning from triggers. Linguistic Inquiry 27: 605-622. Bloom, Paul 1993 Grammatical continuity in language development: the case of subjectless sentences. Linguistic Inquiry 24: 721-34. Bod, Rens, Jen Hay & Stefanie Jannedy (eds.) 2003 Probabilistic linguistics. Cambridge, MA: MIT Press, de Boysson-Bardies, Benedicte 1993 Ontogeny of language-specific phonetic and lexical production. In Changes in speech and face processing in the first year of life, B. Boysson-Bardies, S. de Schonen, P. Jusczyk, P. MacNeilage & J. Morton (eds.), 353-363. Dordrecht: Kluwer. 1999 How language comes to children. Cambridge, MA: MIT Press, de Boysson-Bardies, Benedicte, Pierre Halle, Laurant Sagart & Catherine Durand 1989 A cross-linguistic investigation of vowel formants in babbling. Journal of Child Language 16: 1-17. Brown, Roger 1973 A first language. Cambridge, MA: Harvard University Press. Bush, Robert & Frederick Mosteller 1951 A mathematical model for simple learning. Psychological Review 68: 313-323. Chomsky, Noam 1955/75 The logical structure of linguistic theory. MS (1955), Harvard/MIT. New York: Plenum (1975). 1965 Aspects of the theory of syntax. Cambridge, MA: MIT Press. 1981 Lectures on Government and Binding. Dordrecht: Foris. 1986 Knowledge of language. New York: Praeger. Clahsen, Harald 1986 Verbal inflections in German child language: Acquisition of agreement markings and the functions they encode. Linguistics 24: 79-121. 1996 Introduction. In Generative perspectives on language acquisition, H. Clahsen (ed.). Amsterdam: Benjamins. 1999 Lexical entries and rules of language. Brain and Behavioral Sciences 22:991-1013. Clark, Robin 1992 The selection of syntactic knowledge. Language Acquisition 2: 83149. Grain, Stephen & Rosalind Thornton 1998 Investigations in Universal Grammar. Cambridge, MA: MIT Press. Mürvet 1987 Anchoring conditions for tense. Linguistic Inquiry, 633-657.

Counting grammars

403

Fisher, Cynthia 2002 The role of abstract syntactic knowledge in language acquisition: A reply to Tomasello (2000). Cognition 82: 259-278. Gallistel, C. Randy 1990 The organization of learning. Cambridge, MA: MIT Press. Gibson, Edward & Kenneth Wexler 1994 Triggers. Linguistic Inquiry 25: 355—407. Grinstead, John 1994 Consequences of the maturation of number morphology in Spanish and Catalan. MA thesis, UCLA. Guasti, Maria Teresa 2002 Language development: The growth of grammar. Cambridge, MA: MIT Press. Harley, Heidi & Rolf Noyer 1999 Distribution morphology. Glot International 4: 3-9. Herrnstein, Richard & Donald Loveland 1975 Maximizing and matching on concurrent ratio schedules. Journal of the Experimental Analysis of Behavior 24: 107-116. Hoekstra, Teun & Nina Hyams 1998 Aspects of root infinitives. Lingua 106: 81-112. James, Sharon& Linda Khan 1982 Grammatical morpheme acquisition: An approximately invariant order? Journal of Psycholinguistic Research 11: 381-388. Jusczyk, Peter, Ann Cutler & Nancy Redanz 1993 Preference for the predominant stress patterns of English words. Child Development, 64, 675-697. Kroch, Anthony 1989 Reflexes of grammar in patterns of language change. Language Variation and Change 1: 199-244. Labov, William 1969 Contraction, deletion, and inherent variability of the English copula. Language 45: 715-762. Legate, Julie Anne & Charles Yang 2002 Empirical reassessment of poverty stimulus arguments. LinguisticReview 19: 151-62. 2007 Morphosyntactic learning and the development of tense. LanguageAcquisition, 14, 1-30. Lidz, Jeffrey, Henry Gleitman & Lila Gleitman 2003 Understanding how input matters: verb learning and the footprints of Universal Grammar. Cognition 87: 151-178. Lightfoot, David 1999 The development of language: Acquisition, change, and evolution. Oxford: Blackwell.

404

Charles Yang

Locke, John 1993 The child's path to spoken language. Cambridge, MA: Harvard University Press. MacNeilage, Peter & Barbara Davis 1993 Motor explanations of babbling and early speech patterns. In Changes in speech and face processing in the first year of life, B. de Boysson-Bardies, S. de Schonen, P. Jusczyk, P. MacNeilage & J. Morton (eds.), 341-352. Kluwer: Dordrecht. MacWhinney, Brian 1995 The CHILDES Project: Tools for analyzing talk. Hillsdale: Erlbaum. 2004 A multiple process solution to the logical problem of language acquisition. Journal of Child Language 31: 883-914. Marcus, Gary, Steven Pinker, Michael Ullman, Michelle Hollander, John Rosen & FeiXu 1992 Overregularization in language acquisition. Monographs of the Society for Research in Child Development 57. McDaniel, D., B. Chiu & T. Maxfield 1995 Parameters for Wh-movement types: Evidence from child language. Natural Language and Linguistic Theory 13: 709-753. Naigles, Letitia & Erica Hoff-Ginsburg 1998 Why are some verbs learned before other verbs? Effects of input frequency and structure on children's early verb use. Journal of Child Language 25:95-120. Niyogi, Partha 2006 The computational nature of language learning and evolution. Cambridge, MA: MIT Press. Phillips, Colin 1995 Syntax At Age 2: Cross-Linguistic Differences. MIT Working Papers In Linguistics 26: 325-382.. Pierce, Amy 1992 Language Acquisition and Syntactic Theory: A Comparative Analysis of French and English Child Grammars. Boston, MA: Kluwer. Pinker, Steven 1984 Language learnability and language development. Cambridge, MA: Harvard University Press. Poeppel, David & Kenneth Wexler 1993 The Full Competence Hypothesis. Language 69: 1-33. Prasada, Sandeep, Steven Pinker & William Snyder 1990 Some evidence that irregular forms are retrieved from memory but regular forms are rule generated. Poster presented at the Psychonomic Society Meeting, Nov. 1990. New Orleans.

Counting grammars

405

Pye, Clifton 1979 The Acquisition of Quiche (Mayan). Current Anthropology 20: 45960. Rasetti, Lucienne 2003 Optional categories in early French syntax: A developmental study of root infinitives and null arguments. Doctoral dissertation, Universite de Geneve, Switzerland. Rizzi, Luigi 1986 Null object in Italian and the theory of Pro. Linguistic Inquiry 17: 501-557. 1994 Some notes on linguistic theory and language development. Language Acquisition 3: 371-393. Roeper, Thomas 2000 Universal bilingualism. Bilingualism: Language and Cognition 2: 169-185. Rus, Dominik in prep. The acquisition of tense and agreement in child grammar. Doctoral diss., Washington, DC: Georgetown University. Sakas, William & Janet Dean Fodor 2001 The Structural Triggers Learner. In Language acquisition and learnability, S. Bertolo (ed.) Cambridge, UK: Cambridge University Press. Sankoif, David 1988 Variable rules. In Sociolinguistics. An international handbook of the science of language and society, U. Ammon, N. Dittmar & K. Mattheier (eds.), 984-997. Berlin: Walter de Gruyter. Shipley, Elisabeth, Carlota Smith & Lila Gleitman 1969 A study in the acquisition of language: Free response to commands. Language, 45, 322-342. Tardif, Twila, Marilyn Shatz & Letitia Naigles 1997 Caregiver speech and children's use of nouns versus verbs: A comparison of English, Italian, and Mandarin. Journal of Child language 24:535-565. Thornton, Rosalind & Stephen Grain 1994 Successful cyclic movement. In Language Acquisition Studies in Generative Grammar, T. Hoekstra & B. Schwartz (eds.), 215-53. Amsterdam/Philadelphia: Benjamins. Tomasello, Michael 2000 Do young children have adult syntactic competence? Cognition 74: 209-253. 2003 Constructing a language. Cambridge, MA: Harvard University Press. Valian, Virginia 1991 Syntactic subjects in the early speech of American and Italian children. Cognition 40: 21-82.

406

Charles Yang

Wang, Qi, Diane Lillo-Martin, Catherine Best & Andrew Levitt 1992 Null subject vs. null object: Some evidence from the acquisition of Chinese and English. Language Acquisition 2: 221-54. Weissenborn, Jürgen 1992 The role of null subjects in early grammar. In Theoretical issues in language acquisition, J. Weissenborn, H. Goodluck & T. Roeper (eds.). Hillsdale, NJ: Erlbaum. Wexler, Kenneth 1994 Optional infinitives, head movement, and the economy of derivation in child language. In Verb movement, D. Lightfoot & N. Hornstein (eds.), 305-350. Cambridge: Cambridge University Press. 1998 Very early parameter setting and the Unique Checking Constraint: A new explanation of the Optional Infinitive stage. Lingua 106: 23-79. Yang, Charles 2002 Knowledge and learning in natural language. New York: Oxford University Press. 2004 Universal grammar, statistics, or both. Trends in Cognitive Sciences 8:451-456. 2005 On productivity. Yearbook of Language Variation 5: 333-370. 2006 The infinite gift. New York: Scribner. In press The great number crunch: Reflections on Probabilistic Linguistics. Journal of Linguistics.

List of contributors

Natalia Gagarina Centre for General Linguistics, Typology and Universals Research (ZAS)

Ute Bohnacker Department of Linguistics and Philology

Schützenstr. 18 10117 Berlin Germany

Uppsala University Box 635 SE-75126 Uppsala Sweden

[email protected]

[email protected]

Insa Giilzow Centre for General Linguistics, Typology and Universals Research (ZAS)

Denisa Bordag Department of Linguistics

Schützenstr. 18 10117 Berlin Germany

University of Leipzig Beethovenstr. 15 04107 Leipzig Germany

[email protected]

[email protected]

Darinka Andelkovic Laboratory of Experimental Psychology Faculty of Philosophy

Nancy Budwig Department of Psychology

University of Belgrade 18-20CikaLjubinaStr. 11000 Belgrade, Serbia

Clark University 950 Main Street Worcester, MA 01610 USA

[email protected]

[email protected]

Kristine Bentzen Center for Advanced Study in Theoretical Linguistics (CASTL)

Stephen Grain Center for Cognitive Science

University of Troms0 Department of Language and Linguistics 9037 Tromse Norway [email protected]

Macquarie University North Ryde, Sydney Australia [email protected]

408

List of contributors

Katherine Demuth Department of Cognitive & Linguistic Sciences

Tanja Kupisch Department of Linguistics

Brown University Box 1978 Providence, RI 02912, USA

McGill University 1085 Dr. Penfield Montreal, Quebec H3A 1A7 Canada

[email protected]

[email protected]

Milja Djurkovic Research Centre for English and Applied Linguistics

Jason Mattausch Centre for General Linguistics, Typology and Universals Research (ZAS)

University of Cambridge English Faculty Building 9 West Road Cambridge CB3 9DP, UK

Schutzenstr. 18 10117 Berlin Germany mattausch@zas. gwz-berl in.de

[email protected] Christina Kauschke Institut für Germanistische Sprachwissenschaft

Tom Roeper Department of Linguistics

Philipps-Universität Wilhelm-Röpke-Str. 6 a 35032 Marburg Germany

University of Massachusetts 226 South College 150 Hicks Way Amherst, MA 01003-9274 USA

[email protected]

[email protected]

Gisela Klann-Delius Institut für Deutsche und Niederländische Philologie

Maja Savio Laboratory of Experimental Psychology Faculty of Philosophy

Freie Universität Berlin Habelschwerdter Allee 45 14195 Berlin Germany

University of Belgrade 18-20CikaLjubinaStr. 11000 Belgrade Serbia

gyzheel@zedat. fu-berl in. de

[email protected]

List of contributors Graciela Tesan Center for Cognitive Science Macquarie University North Ryde, Sydney Australia [email protected]

Marit Westergaard Center for Advanced Study in Theoretical Linguistics (CASTL) University of Troms0 Department of Language and Linguistics 9037 Tromse Norway [email protected]

Rosalind Thornton Linguistics Department

Charles Yang Linguistics Department

Macquarie University North Ryde, Sydney Australia

University of Pennsylvania 608 Williams Hall Philadelphia, PA 19104

[email protected]

[email protected]

Sigal Uziel-Karl Department of Communication Disorders Haifa University P.O. Box 803 Reut 71908 Israel [email protected]

409

Index

accessibility general, 3 Α-chains, 241-242, 257, 264 acquisition, 1-2, 5-15,23,29, 38,41,44, 49-55, 59, 67, 71-72, 76-77, 83-86, 88-90,94-97, 100-101, 103-104, 106-107,117,119,121-123,126128,138-153, 159, 162, 165,169, 171-172, 181-184, 197,205-209, 213, 215, 219, 224-225, 227-229, 237-242, 244, 250-251, 253-254, 256-257, 259-262, 271-273, 276277,280, 285-286, 291, 293-294, 299, 301-308, 310, 315, 323-324, 331, 333, 348, 353,359-360, 371, 374-375, 383-391, 393-394, 398399,401,412 ~ of grammatical elements, 10 item based approach to language ~, 147 lexical-, 181-185, 198 mechanism (of language acquisition, underlying), 3, 7, 15, 23-24, 33, 85, 145, 171,228,375 adjective, 42, 56-57, 69-70, 226 adjustment (of adult language to child's age), 154, 158,162, 170,215,331, 337-339 adverb, 271-272,277-278, 281, 284285, 287-291, 293, 299, 301, 368, 372, 396, 398 agent, 38, 117-120, 124, 129, 139,242, 245, 247-249, 264 algorithm Bidirectional Gradual Learning Algorithm (BiGLA), 4, 338 Gradual Learning Algorithm, 337

Learning Algorithm, 338, 342, 350, 360, 394 article, 9, 10, 31, 39-40, 51-53, 55-72, 75-78, 86-87, 90-92, 95-102, 104, 187,210,227,300,331 ~ omission, 51-53, 57, 67,69-70 expletive ~, 95-96 bare ~ constraint, 4, 340, 343-346 ~ nominal, 53 ~ noun, 11,31,51,55,58-61,67,7077, 83, 86-90,92, 94,96, 103-105 ~ singular, 57-59, 71-72, 74-75, 78 binding anaphoric binding phenomena, 4 Binding Theory, 331-332 case, 39-40,97-99,151,169,246,248249, 263-264 CHILDES, 59, 91, 123-124, 174, 210, 362, 364, 373-374, 376, 383 clitics, 101 enclitic, 11,51, 55-56, 61, 64-65, 68-70,76-78, 102 conceptual transparency, 3, 207 constraint constraints, 4, 14, 107, 148-149, 152, 156, 168, 172, 242, 272, 293, 334, 336-344, 346, 348, 350-354, 383384,386,393,398,401 continuity hypothesis, 15, 366-369, 396 contraction wanna (see wanna contraction), 359, 364-366 core, 24, 42, 238-239, 359-360, 364, 393,401

412

Index

data data-providing factors, 184 datives double object ~, 6 to- ~, 6 declaratives, 13,271,278,281,283,287, 289,292,294, 296-297, 300,365 definite, 10, 31, 51, 53, 55-57, 61, 64, 66-70, 75-78, 86, 95-96, 101-102, 105, 165 ~ pronominal, 51, 55-57, 61, 65, 6870, 75-78 defmiteness, 31, 52, 56, 97-98,107, 303 determination coefficient of ~, 153, 159, 174 double ~, 56-57, 65, 78 determiner -omission, 10,83,86, 106 demonstrative, 56, 69-70, 78, 187 determiners, 10,11,49, 52-53,63,71, 78, 83, 86-87, 90, 92, 95-99, 101, 103-105,386 DP, 53, 59, 89 economy ~ of movement, 293-294, 297, 300 -principle, 13,272,286,291,293, 295,297,300-301 structural ~, 293, 343 elicited production, 362, 370, 374 embedding embedded clauses, 13,271-274,276282, 285-288, 290-293, 296-303, 372 embedded questions, 13, 271-272, 278,280-281,284-285,287-292, 297-298,301,303 exponential growth, 188, 190, 196 factor ~ of language acquisition, 207 cognitive-, 150,391

language-specific-, 148 fine-tuning, 185, 198-199 form-function pairings, 2 frequency, 1-15, 23-31, 33, 35-39,4245, 51-52, 54^55, 64-86, 88-90, 9496,98, 103-106, 117,120-122,126127, 137-140, 145-147, 149-150, 152-155, 157-163, 165-166, 168172, 182-186,196-199,205-210, 224, 227-228, 235, 237, 240-242, 244, 250, 253-260, 262, 271-273, 276, 286-287, 289, 291, 294, 299302, 307-308, 318-319, 321-322, 324, 339, 360, 373, 383-394, 397398, 400-401 -debate, 1-3, 5, 8,10,324 ~ effects, 1-2, 10, 14,23, 55, 77,121, 138-139,183,237,262, 383-386, 390-394,397-398, 400-401 -pattern, 51, 54-55, 65, 67, 76, 122, 237,244, 253, 256-258 input frequencies, 5,6, 11, 52, 54, 67, 70, 77, 115, 208, 272, 286-287, 291,300-302 token ~, 83, 85-87, 90, 94, 96, 103105,196,207,239,359 type-,7,98, 191 full, 6,29, 38, 72, 85,101,119,124-125, 134-137, 151-152, 168, 170, 182, 219,295,370 function grammatical functions, 249, 250251 gender, 52, 56, 97-101, 104, 106, 121, 125, 213-214, 308-310, 312, 315316,318,321,323,325 generative, 14,52,55,85-86,88,95,103, 272, 287, 303, 334-335, 343-344, 350-353, 390, 392-393, 398, 400401 grammar multiple ~s, 9, 23, 36,40, 52, 55

Index Universal ~, 13, 15,42, 52, 84, 147, 180,237,239,272,360-361,366367, 371, 375, 391-393, 400, 412

413

229, 241, 244, 259, 263, 277, 286, 300, 311,315, 353,370-371,384, 398 Germanic, 44, 51, 60, 70-71, 86-87, impersonate, 13,245,247-249,254-255, 90,104-105,367,370 Hebrew, 10, 12, 117-119, 124-126, 263 indefinite, 10, 51, 53, 55-57, 59, 61, 64, 128,132,139, 148,229 66-67, 75-77, 86, 96, 99, 102, 105, Italian, 10-11,31, 53-54, 60, 71-72, 247, 249, 263 76, 78, 83, 86-87, 90-92, 94-101, 103-107, 148, 229, 245, 263, 292, infinitive Russian infinitives, 12 368-369,372, 395,398 infinitives Norwegian, 10, 13, 60, 78, 104, 107, optional infinitives, 205-206,208, 229,271-277, 281, 285-286, 288229,412 289,292-295, 297,299-303 inflectional Polish, 229, 244, 245, 248, 263 -endings, 13,209,307 Russian, 10, 12, 151, 205, 208-214, ~ learning, 7 217,223-225,227,229, 237,240, innate, 2, 8, 13, 52, 54-55, 64, 84, 139, 251, 256-258, 260-261, 264, 310 148,182,206,237-240,333,348,395 Serbian, 10, 12, 145, 148, 150-155, innateness, 5,206, 240 158,165,169-170,174,237,240innateness hypothesis, 5 246, 248,250-251,253, 256-258, input 260-261,263-264 maternal-, 12, 126, 132, 181, 184Swedish, 10-11, 51-53, 55-60, 64, 185,191 66,69, 71-72, 74-77,94, 104, 229, 277, 288, 307, 372-373 language language acquisition conversational-, 154-155, 159, first ~, 5, 272 learning 162-163,173 Danish, 78, 148-149, 229 probabilistic / statistical learning, 7, English, 5-7, 14, 29, 31, 33-36, 3856,88, 172,182,373,375,390 39, 41,45, 53, 59-60, 86,103,122, lexical 148, 150, 185, 199, 229, 241-242, -development, 12, 149-150, 174, 292, 294, 303,311,315, 325, 332181-182, 184-185, 197, 207, 390 333, 340, 343, 346-348, 361, 364, -diversity, 148, 161, 186, 190, 193, 366-368,370,371-373,376, 384196, 199 386,391,395,398^01,412 LF-interpretability, 97 French, 3, 10-11, 53, 60, 71-72, 76, linguistic 78, 83, 85-87, 90-101, 103-107, cross-linguistic, 10, 53, 64, 76, 78, 229, 368, 372-373, 386, 398-tOO 86-87, 94-97, 103, 123, 148, 169, German, 6, 10, 13,29, 31, 33-34, 39, 171, 183,238,261,372,386 41,45, 53, 59-60, 71-72, 76, 78, cross-linguistic variation, 94-96,103, 83, 85-87, 90-94, 96-107, 114, 238 151,181, 184-185-196, 198-199,

414

Index

markedness, 3,4,14,227, 334,340, 343346, 354, 384 ~ constraints, 4, 334, 340, 343, 345346, 384 mental representation, 9, 23-28, 44^45, 52-54,171,260 models connectionist ~, 7, 147 morphological ~ features, 97, 225 -load, 100 movement, 13, 25, 34,43, 45, 241-242, 264,271-283, 285-290, 292-301, 303,365,368,371,412 naming, 31, 64, 74 negation, 15, 271-279, 281-282, 284291,293-302,359,371-375 non-agent, 117-120, 122, 127,129-130, 133,138-140 non-subject-initial clauses, 273, 277,294 noun -bias, 182 nouns, 12-13, 51, 56-63, 67, 70-76, 83,87-88,90,92,94,99-100, 103-104, 106-107, 121-122, 125, 134-137, 149, 151,182, 184,187, 192-196, 198-199,212,307-311, 313-314,318-319,321-324,390 NP, 31,34, 53, 89 null, 119, 124, 127-128, 138-139,209,

395,398-399,412 number, 52, 56, 97-99, 125, 209, 213214,229,251,323 obligatory context, 60-61, 76, 90 opacity, 98, 105 optimality bidirectional ~, 3,4,14,334-335,349, 351-353 Optimality Theory, 334, 342, 383 stochastic ~, 336

overgeneralization, 8, 34, 271-272, 277, 279-281,283, 285, 287,291,297298,300-301,319,322 parameter, 29, 33, 35-36, 52, 55, 84, 88, 103-105, 182, 238, 295, 299, 359, 371, 373-375, 393-398, 401, 412 ~ resetting, 374 ~ setting, 52, 55, 84, 88, 103, 182, 238, 359, 376, 394, 396, 398,412 passive, 5, 12, 213, 237, 240-258, 260, 385 reflexive ~, 243, 245-246 periphery, 238-239, 401 personal-social words, 181, 184, 187188, 191-195, 197-198 plasticity value, 337, 339 prenominal, 51, 56-58, 61, 63-64, 6869, 70, 75-78, 86, 92, 95, 99, 105 prepositions, 11, 58, 99, 145-156, 158159,161-162, 165-170,173,187 principles, 14, 43, 55, 83-84, 147, 150, 182, 238-240, 264, 293, 318, 332334, 339, 346, 359, 360, 362, 366, 369, 394 ~ and parameters, 147,182,238,359360, 394 probability, 4, 9, 89, 336, 337, 343, 351, 353,395-397 production language-, 15,76-77, 145,310 spontaneous-, 15, 55, 64, 76-77,145, 151-152, 168,241,310,376 productivity, 3,33-34,36, 39^0,45,54, 56-57, 63, 184, 205, 214, 216, 219, 224,229,237,251,258,392,412 pronoun, 14, 120, 124-125, 134-138, 226, 331-333, 339, 341, 346-349, 351-354,395 - interpretation problem, 14, 331, 333, 339, 346-349, 352-353 personal-, 14, 187, 315 reflexive-, 14,247,256

Index prosody, 11, 102, 104 prosodic prominence, 3 quantitative approach (to research of language acquisition), 145-146 questions wh- ~, 6, 15, 121, 271, 273-276, 279290, 292-294, 296, 298-299, 301, 363-365, 367-370, 376 Why- ~, 359, 367-369, 376 ranking value, 336-337, 342-344, 354 register, 35, 38, 153 regression analysis, 154, 174 sample (of language), 153, 159, 172 semantic generality, 6 singly determined, ~ definite, 57, 64, 69-70, 75, 78 social-pragmatic factors, 183 speech caregiver ~, 11, 13, 55, 67, 76 child directed ~ (CDS), 11, 99, 121, 145, 153-163, 165, 168-170, 173174, 181,184-188, 193,195-199, 397 child-directed (adult) ~, 51, 88, 99 split-CP model, 272, 291-292, 294, 298-299, 301 spontaneous production, 55, 64, 146, 151-152, 168,241,376 statistical, 6-8, 13, 38,43, 54-56, 88, 93, 100,146-147, 152-154, 168,172, 182, 188, 199,255, 307, 339-340,342-343, 359, 363, 373-375, 390-391, 401 ~ properties (of language input, of language production), 147 quantitiative/distributional properties, 149, 158, 172, 182 structure argument ~, 15, 117, 124, 240, 244247, 249-251, 260-261, 264 biclausal-, 280-281

415

clause ~, 272, 279, 286, 291-293, 295,301 complexity of-, 7, 12, 150, 152-153, 166,168,170-171,205,240-241, 256,259,261-262 innate structures, 2 passive structures, 5 subject, 11, 35-37, 58, 73-74, 95, 117120, 122, 124-125, 127-130,133140, 151, 209, 230, 241-245, 263, 271-274, 276-281, 285-294, 296303, 307, 340, 345, 352, 361-363, 365-369, 374, 376, 395, 398 surface-, 119,247,249 target criterion, 76 tokens, 84, 122, 130, 133, 149,151, 170, 181, 186, 188-198, 207,215-218, 253,386 trend analysis, 188, 190, 196 types, 4, 6, 8, 11-12, 53, 61, 64-67, 69, 72, 75-76, 78, 84, 89, 91, 97, 99-100, 105-106, 117, 120,122,130-132, 134-140, 181, 186, 188-198,214215, 217, 225-226, 245, 256, 271, 282, 290, 292, 294, 296, 299, 301, 318,320,334,342,386,400 type-token ratio, 193-194 unaccusativity, 263 usage-based approach, 8, 84, 139, 145, 147, 182, 238-239, 242, 251, 258 variational Model, 88, 396 verb change-of-state ~, 118, 127, 130 verbs, 6-7, 11-12, 30, 32, 34-36, 38, 45,58,74, 117-119, 121, 122, 124-127,129-131,133-134,137139,148,182, 184,187, 191-196, 198-199,205-206,208-209, 211216, 218-219, 225, 243-247, 249, 252, 256, 263-264, 285-286, 294, 302,373,385,390-391,398

416

Index

verb placement, 271, 276-277, 286, 303 word ~ categories, 181-182, 184-187, 191, 193-195, 197-199 -frequency, 181, 185-186, 196

~ order, 10, 13, 151, 209, 224, 271273, 275-283, 285-287, 289-292, 294-303, 307, 372 parts of speech, 181-182, 188, 197 written language, 154-159, 173