The Acquisition of Lexical and Grammatical Aspect 9783110800715, 9783110166156

Citation preview

The Acquisition of Lexical and Grammatical Aspect

W DE

G

Studies on Language Acquisition

Editor

Peter Jordens

Mouton de Gruyter Berlin · New York

16

Ping Li and Yasuhiro Shirai

The Acquisition of Lexical and Grammatical Aspect

Mouton de Gruyter Berlin · New York 2000

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter 8c Co., Berlin.

The series Studies on Language Acquisition was formerly published by Foris Publications, Holland.

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 The acquisition of lexical and grammatical aspect / Ping Li and Yasuhiro Shirai p. cm. — (Studies on language acquisition ; 16) Includes bibliographical references and indexes. ISBN 3-11-016615-1 (cloth : alk. paper) 1. Language aquisition. 2. Grammar, Comparative and general — Aspect. I. Li, Ping, 1962II. Shirai, Yasuhiro, 1957III. Series. P118.A1424 2000 401'.93-dc21 00-045250

Die Deutsche Bibliothek — Cataloging-in-Publication Data Li, Ping: The acquisition of lexical and grammatical aspect / Ping Li and Yasuhiro Shirai. - Berlin ; New York : Mouton de Gruyter, 2000 (Studies on language acquisition ; 16) ISBN 3-11-016615-1

© Copyright 2000 by Walter de Gruyter GmbH & Co. KG, 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. Printing: Werner Hildebrand, Berlin. Binding: Lüderitz & Bauer GmbH, Berlin. Printed in Germany.

To Rose, Jessie, and Foong Ha

Preface

This book grew out of our continuous engagement in research and career development since our graduate school years. Although the book is intended to represent the most up-to-date understanding of the acquisition of lexical and grammatical aspect, some of the work discussed here goes back to our dissertation research in the late 80's and early 90's. When we first talked about writing this book, we did not anticipate that it could take so long, but now we are pleased to see the outcome of this book in the new millennium. This book represents a joint enterprise: we wrote five of the eight chapters together, although each of us had main responsibilities for specific sections within each chapter. We wrote the following three chapters separately: PL wrote Chapter 5 (Acquisition of Chinese) and Chapter 7 (A connectionist model of the acquisition of aspect), and YS wrote Chapter 6 (Acquisition of Japanese). Even with these three chapters we commented on each other's writing extensively, and revised and rerevised on the basis of several rounds of comments from each other. Without doubt, there may still be inconsistencies both in style and in content, given the inevitable differences between any two authors on what constitutes the best style. However, we are both responsible for any imperfections that may still exist. Bulk of the writing was completed during the summers of 1998 and 1999, when one of us (PL) was a visiting scientist at the Max Planck Institute for Psycholinguistics, Nijmegen. Countless emails and attachments crossed the Atlantic Ocean between Nijmegen and Ithaca. At the dawn of this new era, as we review the greatest achievements of the last century, we must thank the technological advancement of computers (and the internet), without which this book would not have been possible. Although we did not specifically plan to meet each other to discuss our book plan, there were three opportunities when we attended the same scholarly meetings, at which times we were able to discuss the details of the book in person. We are indebted to our colleagues who organized these meetings: John Whitman at Cornell University (workshop on First Language Acquisition of East Asian Languages at the LSA Summer Linguistic Institute, July 1997), Eve Clark at Stanford University (Child Language Research Forum, April 1999), and Mineharu Nakayama at the

Ohio State University (International East Asian Psycholinguistic Workshop, August 1999). The papers that we were urged to present at these forums also served as catalysts to the writing of the book. We owe the idea of this book first to Peter Jordens, who had urged us to produce such a work since our graduate school days. Peter's encouragement has been a constant source of stimulation, without which this book could not have come into existence. We would also like to thank a number of our colleagues who contributed to the development of this book in various ways. Melissa Bowerman, Wolfgang Klein, and Roger Andersen were our graduate mentors, whose insights into language acquisition and whose crosslinguistic studies are reflected in the book in many important respects. Eve Clark kindly read the first four chapters during the summer of 1998, when she was at the Max Planck Institute for her sabbatical leave. She provided many constructive comments that helped us greatly in the revision of the book. Joanna Luks also read the first four chapters and provided extensive comments, and Kevin Gregg read all the chapters of the final version of the book, and provided extensive feedback on both style and content. Other colleagues have also commented on our work presented here, and we are grateful to them: Elizabeth Bates, Bernard Comrie, Mary Erbaugh, Maya Hickmann, Brian MacWhinney, Dan Slobin, Carlota Smith, Dirk Vorberg, and Richard Weist. Risto Miikkulainen gave very helpful comments and suggestions on the use of DISLEX and self-organizing feature maps, and Curt Burgess and Kevin Lund kindly made their HAL semantic vectors available to our modeling. We also thank Kelly Cox and Collin Raymond for helping us with formatting the book to the stylistic specifications of Mouton. Ann Beck and the staff at Mouton provided very helpful editorial assistance on the format of the book. Cecelia Coleman gave us many useful hints on using Microsoft Word. Needless to say, we are solely responsible for any deficiencies that may still remain. We extend our gratitude to the Max Planck Institute for Psycholinguistics for providing summer support as well as a congenial working environment for this project. The writing of the book has also been supported by a Faculty Research Grant from the University of Richmond (#F97419) and a grant from the National Science Foundation (#BCS9975249) to PL, and grants from the East Asia Program (Japan Travel Award) and the Society for the Humanities at Cornell University to YS. Finally, we want to dedicate this book to our loved ones, Rose and Jessie (PL), and Foong Ha (YS). The writing of the book took away

many hours of family time, when our loved ones had to spend lonely weekends and holidays at home. The most difficult times were Christmas Eve and New Year's Eve, when one of us (PL) had to leave home and work on the final touches of the book. Jessie (6;9) said on the last day of 1999 to PL: "Daddy, you are a liar - every day you say you'll come home at 4 o'clock but you always come home at 7. Today if you don't come back at 4,1 will dial 911..." Hopefully, the amount of time and energy we devoted to the book will be proportional to the book's scholastic contribution. January 5, 2000, Richmond, VA and Ithaca, NY

Contents Preface Chapter 1 Introduction Chapter 2 Aspect: Problem of lexicon and morphology

v 1 11

Chapter 3 Theories of language acquisition and the acquisition of aspect 29 Chapter 4 Acquisition of aspect in English

55

Chapter 5 Acquisition of aspect in Chinese

91

Chapter 6 Acquisition of aspect in Japanese

129

Chapter 7 A connectionist model of the acquisition of aspect 149 Chapter 8 Acquisition of aspect: Conclusions and future directions

185

Postscript Notes References Author Index Subject Index

209 211 223 251 255

Chapter 1 Introduction

/./. What is aspect? The expression of time is one of the central conceptual domains of language, and the acquisition of the ability to talk about time is one of the earliest tasks in language acquisition. We speak of situations as being in the past, present, or future, and we talk about events as ongoing or completed. Languages differ in the resources they offer us for expressing temporal meanings, but they can all express these basic concepts about time. Two of the most important grammatical systems for expressing temporal concepts in the world's languages are tense and aspect. In learning to talk about time, the task for child and adult learners is to acquire the systems of tense and aspect. This book is about the acquisition of aspect. It is concerned with the ability of young children and adult second language learners to acquire the meanings and uses of aspect marking. Consider the language learner faced with the following pair of sentences: (1)

a. 5am made a big toy house for Jessie. b. 5am was making a big toy house for Jessie.

A mature understanding of this pair of sentences involves that (la) implies that the toy house was actually completed and ready for use, and that (Ib) simply states the fact that Sam was engaged in making the toy house, without implying that the house was completed. This semantic difference is due to the difference in aspect marking: in (la) the verb is marked with a perfective aspect, while in (Ib) it is marked with an imperfective or progressive aspect. Part of the task for the child or the adult learner in the acquisition of aspect is to figure out this difference.

1.1.1. Tense, aspect, and modality Aspect is one of a trio of categories involving linguistic markings on verbs, often collectively referred to as the tense-aspect-modality (ΤΑΜ)

2 Introduction system. Thus, any in-depth discussion of aspect must take into consideration not only the independent functions of these categories, but also the interrelationships between them.1 Briefly, modality is a linguistic category that characterizes the attitude of the speaker concerning the proposition expressed in an utterance. Examples of such attitudes include the notions of obligation, necessity, ability, possibility, and reality. By contrast, tense is a linguistic category typically used to locate the time of the event being talked about (i.e., event time) with respect to the time at which the speaker utters the sentence (i.e., speech time). Event time can also be specified with reference to some time other than speech time, i.e., reference time. When event time is prior to speech time, we use the past tense; when speech time is prior to event time, we use the future tense; and when the two overlap, we use the present tense. Note that in all of these cases, speech time and reference time overlap. When they do not overlap, we use tense markings such as the pluperfect (event time prior to reference time and reference time prior to speech time) or the future perfect (speech time prior to event time and event time prior to reference time) (see Comrie 1985 for a detailed discussion of the relationships between event time, speech time, and reference time). Finally, aspect is a linguistic category that characterizes how a speaker views the temporal contour of a situation described. In contrast to tense which is concerned with the relationship between situations at different time points, aspect is the means with which speakers discuss a single situation, for example, as beginning, continuation, or completion. Traditionally, aspect is divided into two basic perspectives: perfective versus imperfective (see Comrie 1976). In Chapter 2 we will discuss the various aspectual perspectives in detail. Having identified the essential functions of each of these categories, we must note that the boundaries between them are often not clear-cut. First, the linguistic forms that express each of these notions tend to grammaticize into other categories (Bybee, Perkins, and Pagliuca 1994). For example, it has been observed that across different languages, modal markers denoting notions such as possibility, probability and intention tend to grammaticize into future tense markers, while completive and resultative aspect markers tend to develop into past tense markers. As a result, at various points during grammaticization, it is often not easy to determine whether a form belongs to one category or the other. This is a diachronic reason for boundary obfuscation. Second, a linguistic form can often have more than one of these functions. For example, across

What is aspect 3 languages past tense forms often have irrealis, counterfactual meanings (Lyons 1977; Dahl 1997), which are in the realm of modality, and perfect marking, which is itself intermediate between tense and aspect, often has inferential, evidential modal meanings (Bybee, Perkins, and Pagliuca 1994). This is a synchronic reason for boundary obfuscation. Our approach to the acquisition of aspect assumes that the category boundaries between tense, aspect, and modality are not discrete, and therefore our investigation will not be limited to the acquisition of aspect per se. In particular, we will discuss the acquisition of aspect along with the acquisition of tense, for the following reasons.2 First, the acquisition of tense has been shown to closely parallel the acquisition of aspect. It has been claimed that in various languages children use past tense markers to denote aspectual notions (i.e., perfective aspect, see more in 3.2.1). Second, as outlined above, both tense and aspect are terms that refer to the notion of temporality. In some languages, a given linguistic device can encompass both functions of tense and aspect: the most apparent example might be that of the English past tense, which marks both the past tense and the perfective aspect (Comrie 1976). 7.7.2. Grammatical aspect vs. lexical aspect There are two types of aspect that have played an important role in past research in the acquisition of aspect: grammatical aspect and lexical aspect. Grammatical aspect, also known as viewpoint aspect (Smith 1983, 1997), refers to aspectual distinctions which are marked explicitly by linguistic devices, usually auxiliaries and/or inflectional and derivational morphology. The progressive aspect in English and the perfective/imperfective aspect in languages such as Spanish, Russian, and Greek are examples of grammatical aspect. Lexical aspect, also known as situation(al) aspect, inherent aspect, or Aktionsart,3 refers to the characteristics of what is inherent in the lexical items which describe the situation. For example, know is inherently Stative (i.e., continuous and homogeneous), while jump is inherently punctual (i.e., momentary and instantaneous). Vendler (1957) proposed a four-way classification of the inherent semantics (i.e., lexical aspect) of verbs - achievement (e.g., fall, die), accomplishment (e.g., make a chair), activity (e.g., run, play the guitar), and state (e.g., love, know). While accomplishments have some duration, achievements and accomplishments both share the feature of having a clear endpoint (i.e., they are

4 Introduction telic). Achievement differs from the other categories in that achievement verbs are not durative; state differs from the other categories in that state verbs are not dynamic; see section 2.2 for more details. This four-way classification of the inherent semantics of verbs has become the starting point for any subsequent research on lexical aspect. 1.1.3. Typological and crosslinguistic differences in aspect Although claims have been made for the universality of aspectual categories (Smith 1997), variations exist across languages. First, at the level of lexical aspect, it may be that languages differ in how categories of lexical aspect are realized. For example, Smith (1997) suggests that achievements in Chinese do not include process leading up to the change-of-state in its scope, whereas those in English do. Second, languages differ enormously in how aspect is grammatically encoded. Some languages do not grammatically mark aspect (e.g., Hebrew), whereas others do not encode tense but only aspect (e.g., Chinese).4 In Romance languages, the perfective-imperfective distinction in aspect is restricted to past tense. An added complication to this picture is that aspectual forms with the same label often have different meanings in different languages: for example, the progressive marking in English (be + V-mg) can denote various extended meanings such as futurate and habitual, whereas the Chinese progressive marker zai is mainly restricted to the meaning of action-in-progress. The challenge involved in the acquisition of aspect is that learners need to acquire both lexical aspect, which is relatively similar across languages, and grammatical aspect, which differs widely across languages. How do they learn categories of lexical aspect, and how do they learn the grammatical aspect marking and its interaction with lexical aspect? One of the goals of this book is to present an overview of the various theoretical accounts of how young children and second language learners acquire grammatical and lexical aspect. But before we proceed, let us first examine the potential psycholinguistic insights we can gain from the study of the acquisition of aspect.

Significance of study 5 1.2. Significance of the study of the acquisition of aspect One undercurrent in contemporary linguistics has been to assign an increasingly significant role to the lexicon as the organizing system for the basic structure of language (Bresnan 1982; Levin 1993; Levin and Pinker 1992). The lexicon-based approach advocates that the lexicon contains essential information not only about the semantics of individual lexical items, but also about how the lexical items are organized and combined in a sentence. Parallel to this trend, researchers in language acquisition are asking the questions of how the learning of word meanings can help children acquire grammatical categories (e.g., Pinker 1984, 1987) or how the sentence context can help children acquire lexical semantics (e.g., Gleitman 1990; Landau and Gleitman 1985; Li, Burgess, and Lund 2000). Aspect, as we have defined it, stands at the interface between the lexicon and the grammar. Lexical aspect contains information about the semantic properties of lexical items, and grammatical aspect conveys information, usually expressed by morphological devices, about grammatical categories. Aspect has generated a great deal of interest in the domain of language acquisition. Most previous studies of the acquisition of aspect acknowledge that there is an interesting interaction between inherent verb meanings and tense-aspect markers in child language (see Shirai, Slobin, and Weist 1998 for a recent discussion). An important finding across languages is that children's early use of progressive markers (e.g., English -ing and Chinese ζαΐ) is associated with verbs that name durative, nonresultative events, while past or perfective markers (e.g., English -ed or Chinese -le) are associated with verbs that name punctual, resultative events. Brown (1973) observed that English-speaking children first use past-tense forms with a small set of punctual and resultative verbs, including fell, dropped, slipped, crashed, and broke. Bloom, Lifter, and Hafitz (1980) showed that at around age 2, -ing occurred almost exclusively with verbs such as play, ride, and write (durative, nonresultative), whereas the past tense forms occurred with verbs such as find, fall, and break (punctual, resultative). Similar results have been obtained with older children, and in other languages including Chinese, French, Italian, Japanese, Polish, and Turkish. 5 This type of "undergeneralized" use of tense and aspect markers (i.e., the use of these markers more restrict!vely than in the adult language) may be a direct reflection of the child's awareness of the schematic structure of the lexicon with which tense and

6 Introduction aspect markers co-occur. In this book, we attempt to take the acquisition of aspect as an entry point for understanding the complex relationships between the semantic structure of the lexicon and morphology in language acquisition. How can we clearly define the semantic structure of the lexicon? Linguists have approached this question in various ways. Aspect presents a good example in this regard. Comrie (1976), Mourelatos (1981), Smith (1991, 1997), Vendler (1957), among many others, have attempted to define lexical aspect as individuated categories (e.g., activities, accomplishments, achievements, states), contrastive pairs (e.g., durative vs. punctual, state vs. process), or hierarchical structures (e.g., dynamic vs. non-dynamic, and telic vs. atelic within dynamic). Each of these classifications or categorizations has its own virtues but at the same time presents difficult problems. One of the difficulties is how to deal with many verbs that seem to be borderline cases. For example, the verb understand may be treated as a stative verb, because the understanding of something implies a continuous, homogeneous, and relatively effortless event, as in John understands my problem very well. But it may also be examined as an achievement verb (in Vendler's terms), in the sense that the point of understanding in time may be instantaneous, as in Suddenly John understood the story. Another example is open as in John opened the box: if the box was one that can be opened instantaneously, open indicates an achievement; if the box was carefully wrapped and it took time to open, open indicates an accomplishment (Shirai and Andersen 1995). These examples serve to demonstrate the fuzziness and flexibility of semantic categories. Traditional approaches seem to be insufficient to account for these complex relationships of lexical semantics (Li and MacWhinney 1996; Li and Bowerman 1998). This difficulty has led us to propose that lexical aspect categories may be treated on a par with covert semantic categories or "cryptotypes" (Whorf 1956), and the acquisition of the grammatical categories of aspect can be examined as the emergence of prototypes in connection with that of cryptotypes. In this book, we attempt to outline such a proposal in order to understand the mechanisms that govern the acquisition of lexical and grammatical aspect. The study of the acquisition of aspect is important not only because aspect bears great linguistic significance (e.g., the lexicon-morphology interface) and presents interesting learning problems (e.g., covert semantic categories), but also because it can help us to test a number of influential theories of language acquisition. In the last twenty years, there

Significance of study 1 have been at least two major theoretical proposals that draw heavily on the acquisition of tense and aspect: the Language Bioprogram Hypothesis (Bickerton 1981, 1984), and the Basic Child Grammar hypothesis (Slobin 1985). According to the Language Bioprogram Hypothesis, certain semantic distinctions are biologically pre-programmed and emerge early in language acquisition. For tense and aspect, two such distinctions are state versus process and punctual versus nonpunctual. Bickerton hypothesized that early on in language acquisition, states should be marked differently from processes, and punctual situations differently from nonpunctual situations. According to the Basic Child Grammar hypothesis, children come to the language acquisition task with a prestructured "semantic space" containing a universal, uniform set of prelinguistic semantic notions. Two notions, process and result, define a basic semantic contrast in children's acquisition of tense and aspect and they provide an early mapping point for children to associate grammatical morphemes with content words referring to actions. The Language Bioprogram Hypothesis and the Basic Child Grammar hypothesis both explain the acquisition of tense and aspect by reference to innate or prelinguistically determined factors. However, in recent years, proposals of this nature have been questioned by researchers who emphasize the role of the linguistic input and the extraordinary ability children display in analyzing patterns of the input (e.g., Bowerman 1985, 1989, 1996; Li 1990; Li and Bowerman 1998; Stephany 1981; Shirai and Andersen 1995). In this book, we attempt to present a comprehensive overview of the various theoretical hypotheses about the acquisition of aspect, and evaluate these hypotheses with respect to empirical evidence from crosslinguistic studies of English, Chinese, and Japanese, in both first and second language acquisition. In particular, we present a proposal that emphasizes the remarkable ability of the learner (child, adult, and connectionist network) at extracting patterns from the linguistic input and at forming patterns of association between the lexicon and the morphology. In sum, the study of the acquisition of aspect can provide significant insights into how young children and adult second language learners acquire one of the central conceptual domains of language, the expression of temporal notions through lexical and morphological structures, and can shed light on the psycholinguistic mechanisms underlying the acquisition process. Aspect is one of the earliest devices in child language acquisition (e.g., -ing is the first inflectional morpheme produced

8 Introduction by English-speaking children; Brown 1973), and one of the earliest but probably the most problematic in second language acquisition (Dietrich, Klein, and Noyau 1995). The acquisition of aspect involves the acquisition of the semantic structure of the lexicon and the systematic application of the morphology. The understanding of the acquisition of aspect allows us to test important theoretical models that have been offered to account for language acquisition. 1.3. Goals of the book At the conception of this book, we had two general goals in mind. First, we wanted to provide a comprehensive overview of the state of the art in research on the acquisition of aspect. The last twenty years or so have seen a large amount of research in this area, yet there has been no integrative text on this topic. Second, we wanted to examine the acquisition of aspect from both developmental and crosslinguistic perspectives, with an eye toward the computational mechanisms that govern the learning process. Our academic history reflects these perspectives. Both of us completed our doctoral studies on the acquisition of aspect, one primarily concerned with Chinese and the other with English. We subsequently pursued these studies, largely in a crosslinguistic context, in both first language acquisition and second language acquisition. One of us joined the "connectionist camp" at UCSD and pursued computational studies of the acquisition of semantic structures. These perspectives on the acquisition of aspect have led us to consider specific issues within the framework of three general questions: (a) how children build and organize lexical and morphological systems, (b) how the learning process may vary across languages, and (c) what psycholinguistic and computational mechanisms best capture acquisition patterns such as overgeneralization (e.g., applying -ing to Stative verbs) and undergeneralization (e.g., applying -ed too restrictively). To recapitulate the specific goals of this book as outlined in 1.2., first, we attempt to take the acquisition of aspect as an entry point for understanding the complex relationships between lexical semantics and morphology in language acquisition. Second, we attempt to present a comprehensive review of the various theoretical hypotheses that pertain to the acquisition of aspect, and evaluate these hypotheses with respect to empirical evidence from crosslinguistic studies of English, Chinese, and Japanese, in both first and second language acquisition. Finally, we

Goals of book 9 attempt to examine the acquisition of aspect as a crucial case in understanding the learner's remarkable ability to extract input patterns and form linguistic associations. The book is organized as follows. Chapter 1 is an introduction to aspect and the significance of the study of the acquisition of aspect. Chapter 2 takes a detailed look at linguistic research on grammatical and lexical aspect. Chapter 3 gives an overview of current debates in developmental psycholinguistics and in second language acquisition, and examines several specific hypotheses concerning the acquisition of aspect. Chapters 4 through 6 are in-depth examinations of the acquisition of aspect in English, Chinese, and Japanese, respectively. Chapter 7 presents a connectionist model of the acquisition of aspect. Finally, Chapter 8 provides a summary and evaluation of our studies, an overview of our theoretical stance, and an examination of the implications that the study of the acquisition of aspect has for theories of language acquisition.

Chapter 2 Aspect: Problem of lexicon and morphology

In this chapter we outline several of the influential linguistic analyses of aspect in the literature, and examine how these analyses treat grammatical aspect, lexical aspect, and the interaction between the two. We will begin by looking at Comrie's analysis of perfective and imperfective aspect, and then proceed to Vendler's categorization of verbs and times, Smith's two-component theory of aspect, and Klein's view on aspect in terms of the relationship between topic time and situation time. We end with a discussion of how these various linguistic analyses bear on children's acquisition of aspect. 2.1. Grammatical aspect 2.1.1. Perfective and imperfective The grammatical encoding of aspectual notions, which we call grammatical aspect, is realized in different languages in different ways, for example, through the use of inflectional morphology, derivational morphology, auxiliary, or periphrastic constructions. This variation does not mean, however, that grammatical aspect is wholly idiosyncratic and language-specific in the way it is encoded. Typological studies of how languages of the world encode aspectual notions have uncovered recurring patterns of aspectual marking (Comrie 1976; Bybee 1985; Dahl 1985; Bybee and Dahl 1989; Bybee, Perkins, and Pagliuca 1994). In this section, we characterize these grammatical aspectual patterns and show how they are related to each other. The most basic aspectual opposition that is often encoded grammatically is that of perfective and imperfective. As noted in Chapter 1, perfective aspect presents a situation as an unanalyzed whole (external view), whereas imperfective aspect presents a situation from within (internal view). In the following example from English, (1) (2)

John built a house. John was building a house.

12 Aspect sentence (1) is aspectually perfective, whereas sentence (2) is aspectually imperfect!ve. In sentence (1), the perfective aspect of the verb reports the situation in its entirety, with the speaker presenting the situation including its initial and terminal points. Thus, sentence (1) entails completion: John finished building a house. In contrast, the imperfective aspect in sentence (2) communicates the internal structure of the event, without regard to its beginning and end points. Thus the meaning of the sentence is non-committal as to whether or not John has finished building the house. Comrie (1976: 25) proposed the following hierarchical classification of aspectual categories.

I

I

Perfective

Imperfective

I

Habitual (Non-progressive)

I

(Continuous) Progressive

Figure 2.1. Classification of aspectual oppositions (adapted from Comrie 1976: 25)

Comrie divided imperfective aspect into habitual and continuous, and further subdivided continuous into progressive and non-progressive. Let us characterize these categories from the bottom up, following Comrie (1976). Progressive aspect is a category of imperfective that has the properties of dynamicity and change as its defining features: typically it denotes a dynamic, continuously changing action in progress, and is generally incompatible with Stative predicates (*He is knowing the answer). Habitual aspect denotes a situation that spans an extended period of time, typically involving repetition of an action over multiple occasions. English has a habitual aspect marker (used to) which is used in past-time reference only. Imperfective aspect, sometimes referred to as general imperfective, denotes both habitual and continuous (i.e., progressive and nonprogressive) qualities. For example, Romance languages grammatically encode the perfective-imperfective distinction in the past, and the imperfective past can describe a past action in progress (i.e., progressive), a

Grammatical aspect 13 past state (i.e., non-progressive),6 and a past habit. For instance, the Spanish imperfective past form can express the equivalents of he was dancing, he loved Mary, and he used to dance? It is interesting to note here that Comrie failed to explicitly characterize the categories of continuous and non-progressive. He defined continuousness in the negative, that is, as imperfectivity that is not habitual, and he did not define "non-progressive" (Bybee, Perkins, and Pagliuca 1994: 138). This may be related to the fact that the aspectual markers that specifically denote these categories are extremely rare.8 For this reason, the two categories appear in parenthesis in Figure 2.1.

2.1.2. Grammaticization of aspect markers As noted in Chapter 1, there is a tendency for the meanings and functions of tense-aspect markers to change over time. For example, it is claimed that progressive aspect markers can develop into more general imperfective aspect markers by generalizing their applicability to habitual and non-progressive situations. In this subsection, we discuss grammaticization of tense-aspect markers to explain how different tenseaspect markers are related to each other, both semantically and historically. Building on research from historical linguistics and crosslinguistic survey of typologically diverse languages, Bybee, Perkins, and Pagliuca (1994) proposed hypotheses concerning the grammaticization of tenseaspect-modality markers. Particularly relevant to our research are two hypotheses concerning the universal paths of the development of tenseaspect markers. According to these authors, first, there is a universal tendency for resultative and completive markers to grammaticize into perfect markers, which in turn become perfective or simple past markers. Second, there is a tendency for a progressive marker to develop into a general imperfective marker by expanding its reference to habitual and stative situations; in terms of Comrie's hierarchy (see Figure 2.1), imperfective aspect markers develop from the bottom up.9 Here, we would like to discuss the perfective path in more detail, primarily to explicate the nature of the category perfect. Bybee, Perkins, and Pagliuca (1994) claimed that completive and resultative aspect markers grammaticize into perfect markers. Completive markers denote an action performed completely and thoroughly, and resultative denotes a state that has been brought about by a prior action. These markers then

14 Aspect extend their meaning and become perfect markers, which denote "past action with current relevance". English perfect (have + past participle) is a typical example of this category (Dahl 1985), which has such uses as "perfect of result", "perfect of experience", etc. Perfect further grammaticizes and loses its "current relevance" restriction, and becomes a perfective aspect marker and/or a simple past tense marker. It is often the case that a past tense marker also has an aspectual value. For example, many European languages (e.g., French, Spanish, German, Dutch) have the perfect form grammaticizing into a past tense form which also has a perfective value, although the degree of grammaticization differs. Thus, in these languages, they have two past tense markers older past forms (often referred to as preterite), and more recent past forms (auxiliary plus participle). As noted in Chapter 1, what we see here is a continuous development of tense-aspect markers, which often makes it very difficult to explicitly determine whether a grammatical form is a tense marker or an aspect marker. Interestingly, this historical development charted for tense-aspect markers has been found to have a parallel in language acquisition. In Chapters 3 through 6, we will discuss in more detail the claim that children use past tense marking first as a resultative aspect marker and then as a pure past tense marker, and that progressive markers are initially restricted to the action-in-progress meaning and only later develop more varied meanings (e.g., Antinucci and Miller 1976; Bloom, Lifter, and Hafitz 1980; Shirai and Andersen 1995). This claim about the acquisition of tense-aspect morphology has also been discussed within the larger issue of how and why there are parallel processes in language acquisition and historical change (Slobin 1977, 1997; Ziegeler 1997).

2.2. Lexical Aspect: Vendler, Smith, and Comrie Grammatical aspect provides a certain amount of information for the interpretation of the aspectual meaning of verbal predicates, but most linguists accept that we need to consider another type of aspect - lexical aspect. Lexical aspect (also known as inherent aspect, situation aspect, or Aktionsart) refers to the semantic characteristics inherent in the lexical content of words, usually verbs or verb phrases, that are defined in terms of the temporal properties of given situations that the verbs describe. Because lexical aspect deals with lexical semantics, it is generally re-

Lexical aspect 15 garded as a lexical category as opposed to the grammatical category of "grammatical aspect". As a starting point for our illustration, Vendler's (1957) four-way classification represents an early attempt to categorize lexical aspect.10 He classified verbs or verb phrases into four categories with respect to the temporal properties that they encode: activities, accomplishments, achievements, and states. According to Vendler, activity verbs encode situations as consisting of successive phases over time with no inherent endpoint, for example, walk, run, and swim. Accomplishment verbs also characterize situations as having successive phases, but they differ from activity verbs in that they encode a natural endpoint and often a change of state, for example, paint a picture, build a house, and run a mile. Vendler illustrated the contrast between activities and accomplishments by the difference between run and run a mile. The sentence John was running entails that John ran, whereas the sentence John was running a mile does not entail that John ran a mile, that is, John might have stopped halfway. Like accomplishments, achievement verbs also encode a natural endpoint, but they differ from accomplishments and activities in that they encode events as punctual and instantaneous, that is, as having no duration, such as in fall, win the race, and reach the summit. Strictly speaking, every event occupies time. But speakers can construe given verbs as denoting situations having no time duration. Finally, state verbs, in contrast to the other three categories, encode situations as homogeneous, with no successive phases or endpoints, involving no dynamicity, such as know and love. Thus, state verbs cannot usually be combined with progressive aspect that marks change and development from one phase to the next (e.g., *John is knowing the story is odd).11 These categories can be schematically represented as follows (Andersen 1990). State Activity Accomplishment Achievement

X X

love, contain, know run, walk, swim paint a picture, build a house fall, drop, win the race

In this schematization, a solid line is used to represent states, because states have no apparent beginning point or endpoint and endure indefinitely unless some external force changes them. The wavy lines for activities and accomplishments indicate the dynamic duration of an action, while χ for accomplishments and achievements represents a

16 Aspect punctual point of change of state, signaling telicity (i.e., natural endpoint). Vendler's analysis, now probably the most widely accepted and the best known, has become an important starting point for subsequent research on lexical aspect. For example, Smith (1997) recently modified this system and applied it to an analysis of English, French, Chinese, Russian, and Navajo. Smith's modification involved mainly the addition of "semelfactive" verbs, such as cough, tap, and knock in English and their equivalents in other languages. She argued that semelfactives resemble achievements in being punctual, but differ from achievements in that semelfactives encode no endpoint. In particular, semelfactives and achievements behave differently with progressive aspect marking: when semelfactives are marked with progressive, they are interpreted as specifying a repeated event (e.g., coughing or knocking several times); when achievements are marked with progressive, they are interpreted as indicating preliminary, detachable stages of the event rather than a repeated event (e.g., John was reaching the summit means that John was at a stage just prior to being at the summit, not that he arrived at the summit several times). In addition to Smith, other researchers have extended Vendler's analysis in other ways. For example, instead of using a categorical fourway classification, some linguists have classified verbs as pairs of contrasts, such as stative versus dynamic verbs, telic versus atelic verbs, and punctual versus durative verbs (Comrie 1976). These pairs of contrast have the advantage of making the semantic features of lexical aspect explicit and transparent. These contrasts can be reanalyzed as features, for example, [±dynamic], [±telic], and [±punctual]. Table 2.1 presents such an analysis, adapted from Smith (1991), to characterize Vendler's four categories plus semelfactives. Table 2.1. Semantic features for the five categories of lexical aspect (adapted from Smith 1991: 30) states dynamic punctual telic

.

activities accomplishments semelfactives + + + . . + + -

achievements + + +

We should point out a confusion that has been in the literature on lexical aspect since Vendler (1957). Vendler did not use the terms lexical aspect, Aktionsart, situation aspect, or inherent aspect. He was speaking

Lexical aspect 17 of "time schemata". However, it is not clear whether time schemata are part of the lexical semantic contents of verbs, or part of the temporal properties of situations to which the verbs refer. Admittedly, it is not always easy to separate lexical semantics from the objects and situations to which the lexical items refer. But we would like to make this distinction here, for reasons that will become clear later on. Our intention in this book is to treat lexical aspect as lexical categories according to the kinds of temporal properties that the lexical items inherently encode, not as situational categories according to the temporal properties that the situations typically display. We contend that situational properties are not necessarily reflected directly in lexical semantics, and that the confusion between the two often leads to negative consequences in the study of language acquisition (see further discussion in Chapter 3). The abstract concepts of, for example, states and activities can be viewed as universal semantic notions, independent of particular languages. In other words, every language may have lexical means to encode states and activities. However, which verbs encode states and which encode activities is language-specific, and can be determined only within the context of a given language. Accordingly, children have to learn to identify which verbs encode which temporal properties of a situation in the language being acquired. An example from Comrie (1976) clearly demonstrates the need to distinguish between inherent verb meanings and temporal properties of situations. According to Comrie, English and Portuguese treat perception verbs (e.g., see, hear} differently. English treats them as stative, and these verbs consequently do not accept progressive marking, while Portuguese treats them as nonstative, so they can naturally accept progressive marking. Take Japanese for another example. Japanese does not have a stative verb corresponding to the English know. In order to express a notion like / know him, Japanese speakers use the verb siru which means 'come to know' and attach to it the durative aspect marker -teiru, to denote the resultative state after coming to know him. Thus, in Japanese, siru is an achievement verb. These examples show that verbs in different languages may differ with respect to their lexical characteristics even though they refer to the same situations. Such crosslinguistic differences show that there is no one-to-one mapping between given types of situations and given types of lexical items. Having emphasized the importance of this distinction, we must acknowledge that correspondences between lexical contents and situational

18 Aspect properties are not totally arbitrary. There are clear cases where languages do not differ. For example, there is probably no language that treats the verb referring to killing someone as stative, and the verb "jump" in any language has to somehow encode the punctuality of the action as part of its meaning. Thus, while it is important for theoretical and empirical reasons to distinguish classifications of situations from classifications of lexical aspect categories, in practice, they often coincide.

2.3. Lexical aspect and grammatical aspect As discussed above, it is clear that interpretation of the aspectual properties of a sentence cannot be determined without considering the contribution of the lexical aspect of the verb. It is also clear that the aspectual meaning of a sentence can, by and large, be predicted on the basis of which grammatical aspect is combined with which lexical aspect category. One explicit proposal in this regard is that of Smith (1991, 1997). In this section, we present our view of how aspectual meaning is determined as a function of the combinatorial properties of lexical aspect and grammatical aspect on the basis of Smith's theory of aspect and the grammaticization theory of Bybee, Perkins, and Pagliuca (1994). Smith (1997) proposed a comprehensive theory of aspect, which she labeled as a "two-component theory". The two components are what she calls "situation type" and "viewpoint aspect", which roughly correspond to what we have described so far as "lexical aspect" and "grammatical aspect", respectively. Smith's theory attempts to account for diverse aspectual phenomena related to the interaction between situation type (achievement, accomplishment, activity, state, and semelfactive) and viewpoint aspect (perfective, imperfective, and neutral)12. We have been using the term "lexical aspect of verbs", but strictly speaking, the lexical aspect value is determined by both the verb and its arguments, which Smith (1997) calls the verb constellation. Examples include [John love Mary] (state), [John run] (activity), [John run a mile] (accomplishment), [John reach the summit] (achievement), and [John jump] (semelfactive). Note that what is inside [ ] is not a linguistic form but the proposition underlying it, without any verb morphology to signal the viewpoint aspect. In such instances, we are talking about the semantic structure of verb-plus-arguments without any value imposed by grammatical aspect.

Lexical and grammatical aspect 19 Another component of the two-component theory is viewpoint aspect. Smith used the term "viewpoint" since it is essentially the speaker's choice as to which aspectual perspective should be used. To describe the fact that John ran a mile yesterday, the speaker can choose to present it from either a perfective viewpoint (John ran a mile) or an imperfective viewpoint (John was running a mile). If the story is about what happened while John was jogging yesterday, imperfective viewpoint would be used. If the story is about what John did after he ran a mile, then perfective viewpoint would be used.13 The speaker chooses to use a certain combination of lexical items (with lexical aspect value) and morphology (with grammatical aspect value), which results in a particular intended meaning, such as action-in-progress, completion, stative, etc. 2.3.1. Interaction between lexical aspect and grammatical aspect How do particular combinations of lexical aspect and grammatical aspect result in particular interpretations of a sentence? What are the principles behind it? In this subsection, we present an account of how the interaction between the two levels of aspect works. Our account differs from Smith's in minor details, but the principles are essentially the same. Imperfective and progressive aspect

As discussed in 2.1, both general imperfective and progressive markers have the same viewpoint aspect value, except that the progressive viewpoint is [+dynamic]. Therefore, they show very similar distributions in their interaction with lexical aspect. Since imperfective viewpoint presents a situation from within, disregarding its beginning or endpoint, it requires duration upon which the internal view can be imposed. If there is no duration, it is impossible to treat the beginning point or endpoint separately. This requirement predicts that some achievements are incompatible with imperfective aspect, as seen in the ungrammaticality of *He is noticing a friend. To see the properties of lexical aspect clearly, we repeat the following diagram from section 2.2 here, with the addition of semelfactives. The punctual point of action for semelfactives is represented by ·, since it is not telic, but atelic.

20 Aspect State Activity Accomplishment Achievement Semelfactive

~—„~~

X X ·

love, contain, know run, walk, swim paint a picture, build a house fall, drop, win the race jump, knock, cough

Although achievements do not normally allow for an internal view of imperfective aspect, it is sometimes possible to impose an internal view on an achievement verb by focusing on the preliminary stages of an event (i.e., the process leading up to the endpoint). To elaborate, John was reaching the summit means that John was at a stage just prior to being at the summit, not that he was in the middle of reaching the summit. As another example, John is winning the race is appropriate when John's horse is ahead of all other horses in the race, even though he has not actually reached the finishing line. Semelfactive is another punctual, instantaneous category that implies no duration. In this case, the speaker can only impose an internal view on the situation when the duration is created by repetition. Thus, John is jumping normally refers to John's repetitive acts of jumping, that is, multiple jumps in a single activity, which Smith (1997) called Multipleevent Activity.14 In contrast to achievements and semelfactives, categories that have duration, i.e., activities, accomplishments, and states, can readily accept an internal view. Both general imperfective and progressive markers yield the action-in-progress meaning when attached to activity and accomplishment verbs, focusing on the dynamic duration of the situation denoted by the verbs. With respect to stative verbs, the picture is a little different. Stative verbs do not usually combine with progressive aspect: a large set of English stative verbs and their translation equivalents in other languages do not normally accept progressive marking, including psychological and cognitive verbs such as want, need, like, love, believe, and know, perception verbs such as see, hear, and feel, and relational verbs or verbs of existence such as resemble, possess, have, and be. The incompatibility seems to be that since progressive aspect presents a situation as ongoing, it requires that the situation have successive phases, i.e., that it be dynamic, whereas the cognitive, perception, and relation or existence verbs indicate only undifferentiated and homogeneous situations. "Progress" presupposes the dynamic development of a situation. Thus, progressive

Lexical and grammatical aspect 21

aspect combines naturally with activity and accomplishment verbs, but not with stative verbs. However, under certain circumstances some Stative verbs do appear in the progressive, which we will call "stative progressive". For example, Smith (1983) showed that progressive aspect can be used with stative verbs to present a state as an event, such as in John is believing in ghosts these days. Another example is John is being stupid, in which the stative progressive refers to a particular behavior of John' s at the time of speech, in contrast to John is stupid which refers to a general characteristic attributed to John. John is being stupid does not make any claim about John's intelligence before or after the moment of speech, but rather that John is behaving stupidly. General imperfective aspect, in contrast to progressive, is naturally compatible with stative verbs and does not yield special meanings.15 Note that general imperfective has the features of [+intemal view] and [-dynamic], and therefore, it can easily find duration in a state, and it does not trigger the action-in-progress meaning because stative verbs also have the feature [-dynamic]. The obtained meaning of general imperfective with stative verbs is thus 'state continuing', as in La mer etait calme 'the sea was calm' in French (Smith 1997: 197) or Wuli-de deng kai-zhe 'the light in the room is on' in Chinese. Perfective aspect Perfective aspect combines naturally with achievement verbs, because by definition, perfective aspect presents a situation as a single whole, and achievement verbs provide ideal instantiation of such a viewpoint in that they depict punctual situations as single points without internal structure. Because achievements involve an endpoint, their combination with perfective aspect denotes the completion of a situation, although in this case the beginning point of the situation coincides with its point of completion. Perfective aspect also naturally combines with accomplishments and is normally interpreted as indicating the completion of a situation, since by definition accomplishments incorporate an endpoint, and perfective aspect views the situation as a single whole externally, with both initial and terminal points. For example, John built a house not only presents the building process as a single whole, but also indicates that the house was indeed finished. The notion of completion is not applicable to uses of the perfective with activity verbs, since activities encode situations with no inherent

22 Aspect endpoints (Comrie 1976; Lyons 1977). For example, John ran cannot indicate that John "completed" his running. To ask Has John finished running? is not appropriate unless the speaker has in mind in advance some delimiter to the distance or time John ran. John ran, then, indicates simply that John engaged in the activity of running for a while and that at some arbitrary (unspecified) point this activity terminated. In the same vein, perfective aspect with semelfactive verbs conveys a termination and not a completion of the situation (e.g., John coughed). Earlier we pointed out that the progressive aspect is not compatible with stative verbs because of the verbs' lack of dynamic meaning. Perfective aspect is not compatible with stative verbs, either. This is because stative verbs do not include either a beginning point or an endpoint in their temporal structure. Therefore, perfective aspect, which includes both the beginning and end points of a situation in its focus, cannot normally combine with stative verbs. In rare cases when the combination is possible, it indicates an entry into a state as in Then suddenly I knew it! by focusing on the punctual point of entry into that state. This in effect translates a stative verb into an achievement verb. The interaction discussed in this subsection is summarized in Table 2.2, which covers the most common grammatical aspect types, although it is not meant to be exhaustive or without exception. Excluded from our analyses here are the more marked cases of habitual reference (e.g., He is walking to school these days) and futurate (e.g., We are eating out tonight), which can be obtained with any of the dynamic verb classes. Table 2.2. Interaction of lexical aspect with grammatical aspect State state

Activity Accomplishment Achievement prelim, stage prog Imperfective prog ?? prog prelim, stage Progressive prog 99 stative prog completion ?? termination completion Perfective inchoative

Semelfactive iterative prog iterative prog termination

prog = progressive prelim, stage = preliminary stages of event ?? = questionable combination

Our discussion in this subsection, as summarized in Table 2.2, shows clearly that there are combinatorial constraints or compatibility between certain grammatical aspect and certain lexical aspect. Comrie (1976)

Lexical and grammatical aspect 23 discussed this kind of lexical-grammatical relationships with the "naturalness of combination" principle. According to Comrie, perfective aspect combines naturally with punctual verbs, for example, because perfective aspect presents a situation as a whole without regard to its internal structure and punctual verbs encode a situation as a single point lacking internal structure. Conversely, imperfective aspect is not compatible with punctual verbs (except semelfactives, which can be construed iteratively), because imperfective aspect presents a situation as having an internal structure while a punctual verb encodes it as a point lacking internal structure. Activity verbs lend themselves naturally to the internal perspective of imperfective aspect because they encode the successive phases of an event over time. Such interactions between lexical and grammatical aspect presumably originate in certain natural relationships between events in reality. For example, when we know that a situation comes to its end with a clear result, this situation has probably already become a past event, and we are therefore likely to comment on its completion by means of perfective aspect (Brown 1973). Many situations with an end result last for such a brief time that they are almost certain to have ended before one can comment on them, such as situations denoted by verbs like drop, fall, and crash. These situations almost exclude the internal perspective of imperfective aspect. In sum, there are likely to be strong associations in the real world between resultativity or telicity and the use of past and perfective verb forms, and between atelic activities and the use of present and progressive verb forms. 2.3.2. Variation across languages in grammaticization of aspect As noted earlier, there is variation across languages in how aspectual notions are grammaticized. There are two types of variation to consider: in the pattern of grammaticization and in the degree of grammaticization. First, different languages develop different types of grammatical aspect, and some languages do not have grammatical aspect (e.g., Hebrew, Finnish). In terms of combinatorial properties of aspect and tense, Romance languages have a perfective-imperfective distinction only in the past tense. Many Slavic languages, although grammatically marking both tense and aspect, do not have perfective aspect in the present tense (e.g., Polish, Russian). In terms of combinatorial properties of grammatical aspect and lexical aspect, the Japanese imperfective marker -te i- cannot

24 Aspect focus on the preliminary stage of an event as in English; instead it focuses on the duration after the terminal point of a situation, and has the perfect and resultative meaning. The Chinese perfective -le does not necessarily yield the meaning of completion with telic verbs (see Chapter 5 for details). These variations are not random. Typological research (Dahl 1985; Bybee, Perkins, and Pagliuca 1994) has shown that some of these tendencies reflect common patterns of grammaticization. For example, it has been suggested that the reason many languages do not have perfective aspect for present is that the most natural way of talking about situations existing at speech time is imperfective (state, progressive, or habitual) and not perfective, because perfective indicates something has been completed or terminated. This pattern indicates the importance of functional motivation in shaping systems of grammatical aspect in natural language. Second, grammatical aspect markers in different languages also vary in their degree of grammaticization, which is in turn reflected in their different combinatorial patterns with lexical aspect. As noted earlier, it has been suggested that progressive markers grammaticize into general imperfective markers (Bybee, Perkins, and Pagliuca 1994: 140-149). In English, the progressive aspect is highly grammaticized. In addition to its unmarked progressive meaning (i.e., action in progress), it can also be used for stative progressive, preliminary stages, habitual, and futurate. Progressive markers in other languages may be less grammaticized, and practically limited to the typical action-in-progress meaning, for example, in Chinese, Malay, and Thai. Thus, in these languages, progressive markers cannot be used for other occasions (e.g., with states or achievements). Similarly, perfective aspect is also sensitive to the degree of grammaticization in its interaction with lexical aspect. As discussed above, perfective aspect is either incompatible with stative verbs, or has an inchoative meaning with some stative verbs. As perfective aspect grammaticizes, however, it can be used as past tense marker, that is, to describe a state that existed in the past (e.g., The book was there yesterday; see Bybee, Perkins, and Pagliuca 1994: 92). For example, the English past tense is aspectually perfective (Smith 1997), signaling completion or termination with dynamic verbs or inception with some stative verbs (e.g., Then I knew //!). But as a highly grammaticized form in the function of simple past tense, it can also be freely applied to stative verbs to locate states in the past time. The Japanese past tense marker -ta has a similar

Lexical and grammatical aspect 25 property, and is considered to be in a grammaticization transition from perfect to perfective and further to simple past (Horie 1997). Finally, a theoretically important variation concerns that of lexical aspect. As we have discussed, some languages (e.g., English) accept imperfective/progressive with achievement verbs to focus on preliminary stages of an event (e.g., He is reaching the summit) whereas others (e.g., Chinese and Japanese) do not. On the basis of this observation, Smith (1997) suggested that achievements in Chinese do not include preliminary stages of an event, whereas those in English do. An alternative interpretation of this phenomenon is that the Chinese progressive markers zai or the Japanese marker -te i- simply cannot focus on preliminary stages of an event. In other words, it is not the lexical aspect of the verb that does not include preliminary stages, but the grammatical aspect marker that cannot denote such stages (Shirai 1998a). This interpretation attributes the difference between English and Chinese to variation in grammatical aspect, not in lexical aspect. Of course, this issue - whether the differential behavior of progressives in different languages is due to grammatical aspect or lexical aspect - needs further empirical investigation. 2.4. A time-relational analysis of aspect Most linguistic analyses, including ours, have adopted the definition of grammatical aspect given in Comrie (1976). That is, grammatical aspect involves different ways of "viewing" the temporal contour of a situation: perfective aspect presents an external view of the situation as a single whole in its entirety without reference to its internal structure, and imperfective aspect an internal view of the inner constituency of the situation without regard for the situation's initial or final boundaries. Smith (1983) treats this definition of aspect as "viewpoint aspect", in contrast to "situation aspect" (see 2.3). Much the same definition of aspect has been used in research by other linguists (e.g., Dahl 1985; Bybee 1985; Bybee, Perkins, and Pagliuca 1994). This approach to aspect has been challenged recently by Klein (1994, 1995, in press). According to Klein, the definition of aspect as different ways of viewing a situation is entirely metaphorical, and thus vague: what does it mean exactly to "view a situation" in its entirety, or as a single whole, or with or without reference to its internal constituency? For example, the difference between John stood on his toes and John was

26 Aspect standing on his toes is characterized as that between the perfective and the imperfective, yet it is difficult to grasp the metaphorical differences of entirety, wholeness, or reference to internal structure in these instances. In view of this problem, Klein proposed that aspect should be examined on a par with tense, strictly in terms of temporal relations, such as "prior to" (>), "contained in"(c), or "posterior to" ( 4

In Table 4.1, the cell numbered T is the prototype, and the acquisition of the morphological marking spreads from this prototype to the

80 Acquisition of English peripherals ('2' through '4'). To test whether the predicted developmental sequences are consistently observed, in what follows we will review case studies, cross-sectional studies, and longitudinal studies in the acquisition of English as L2. These studies have specifically reported quantitative data regarding progressive and past marking in connection with lexical aspect.45 We will first review small-scale case studies, then more extensive cross-sectional studies, and finally longitudinal studies. Wenzell (1989) investigated the stative-dynamic distinction and its relationship with past marking in conversational data from two of the three Russian learners of English that she studied. One informant's 36 past forms were limited to dynamic verbs (which Wenzell called active); there was no occurrence of stative verbs with past tense. Another informant used 96 past forms, among which 67 were dynamic verbs, and only 4 were stative verbs. (For the rest, stative-dynamic values were not reported.) This finding supports the prototype hypothesis, which predicts that past tense morphology is less likely to be used with stative verbs than with dynamic verbs. Robison's case study (1990), as noted earlier, looked at the correlation between morphology and two semantic oppositions: stative versus dynamic, and punctual versus durative. He found that punctual verbs were more likely to be given past marking than durative verbs (21 vs. 7, by token counts). He also found that dynamic verbs received more past markings than stative verbs. In fact, no stative verbs received past marking, whereas 28 dynamic verbs did. These results are highly consistent with the above predictions. Huang (1993) investigated the interlanguage of five Chinese learners of English living in the US. In conversational interviews, all learners used past marking most frequently with achievements (61-89%, average 80%, token counts). Four out of the five learners used progressive -ing most frequently with activity verbs (58% to 100%). One learner, however, used it more evenly, 1 token for state, 2 for activities, and 2 for achievements, without showing any association with particular classes of verbs. This learner used fewer progressive forms than other learners (5 tokens, as compared to 10 - 20 tokens used by other learners). Except for this learner, whose use of progressive is too limited for any conclusion, Huang's study is consistent with our predictions. Robison's (1995) cross-sectional study (N=30) is important since it includes learners from four different proficiency groups, and therefore we can examine more clearly the spread of morphological marking from

Second language acquisition 81 the prototype to the non-prototypes. He used conversational data for his analyses, although he classified the learners into different proficiency groups on the basis of the correct use of verb morphology in obligatory contexts in their writing samples. First, the percentages of progressive marking used with activities from Level I (lowest) to Level IV (highest) were: 57% -> 70% -» 79% -> 80%. Second, the percentages of past marking used with achievements46 were: 43% -> 60% -> 54% -> 55%.47 These results show that (a) at any given time, the prototype progressive/past has the highest percentage of use compared with nonprototypes, which is consistent with our hypothesis, and (b) less advanced L2 learners are not necessarily more restricted to the prototypes than more advanced learners, which goes against the prototype hypothesis. Bardovi-Harlig (1992) reported a cross-sectional study of ESL students (N=135) at Indiana University using a cloze-type passage in which learners were asked to change the base forms of verbs into appropriate morphological forms in context. She provided detailed results concerning the items where simple past tense was obligatory, which makes a reanalysis possible. Our hypothesis predicts that learners will find it hardest to apply a simple past marker to states, but easiest to apply it to achievements. Table 4.2 shows the accuracy rate for three lexical aspect classes, calculated on the basis of her Tables 3 to 8. (Her test items did not include any accomplishment verbs.) Table 4.2. Percent accuracy for verb types in obligatory simple past contexts in Bardovi-Harlig (1992)

Level- 1 Level-2 Level-3 Level-4 Level-5 Level-6

State dive} 31.6% 50.0% 25.0% 48.3% 75.8% 66.7%

Activity (work, take care of. stav) 35.1% 48.3% 54.2% 60.9% 78.8% 70.4%

Achievement (tell, die) 63.2% 82.5% 78.2% 88.0% 87.9% 86.1%

As is clear from the table, our prediction is borne out. At all levels, it is more difficult to inflect state verbs for past tense than activity verbs (with the exception of Level-2, where states had a slightly higher past marking rate than activities, by 1.7%), and it is the easiest for learners to apply past inflection to achievement verbs. Also important is that the

82 Acquisition of English differences in past marking rate between states and activities are smaller compared to those between achievements and activities. Activityachievement differences are more than 20% for four levels, whereas state-activity differences are less than 5% for four levels, including Level2, for which states had a higher past marking ratethan activities. This result can also be nicely explained by the prototype hypothesis. Achievements, the prototype past, are [+punctual, +telic, +dynamic], and share only one feature with activities, which are [-punctual, -telic, +dynamic], whereas activities share two features with states, which are [-punctual, -telic, -dynamic]. It would be interesting to replicate the study with accomplishment verbs, which the prototype hypothesis predicts would be placed in between activities and achievements in terms of past marking difficulty. Bardovi-Harlig and Reynolds (1995), in a similar cross-sectional study (N=182), also found that achievements and accomplishments, which they called "events", showed a consistently higher past marking rate in obligatory past contexts than activities and states. However, in this new study, learners at Levels 4 and 5 clearly showed a higher past marking rate for states than for activities. This is in a sense congruent with Bardovi-Harlig (1992) in that the difference between activities and states is not so great, but the result goes against the prediction of the prototype hypothesis. Bardovi-Harlig and Reynolds's (1995) study is also important in that it examined the effect of adverbs of frequency on tense-aspect marking by L2 learners. The prototype hypothesis predicts, as proposed by Shirai (1991), and Andersen and Shirai (1994), that habitual past events are less prototypical than unitary past events. Thus the habitual interpretation triggered by adverbs of frequency, such as often, usually should make it more difficult for learners to give past marking on state and activity verbs. This was indeed supported: the learners mostly showed a higher ratio of past marking without adverbs of frequency (Bardovi-Harlig and Reynolds's Tables 3 and 4). However, Level-1 (lowest) and Level-6 (highest) learners were not influenced by the presence or absence of habituality. This is probably because Level-1 learners were not yet sensitive to such adverbs, whereas Level-6 learners already had a good control of past marking, and were able to mark past tense even if it was habitual past - a less prototypical past. Bardovi-Harlig and Bergström's (1996) cross-sectional study (N=23) on the same student population (ESL students at Indiana University)

Second language acquisition 83 reported results similar to Robison's (1995) cross-sectional study. First, the percentages of progressive marking used with activities from Level I to Level IV were: 63%->84%-»100%->52%. Second, the percentages of past marking used with achievements were: 62%->59%->67%->61%.48 Thus, the most frequent verb classes are activities for progressive and achievements for past, a result consistent with the prototype hypothesis. But the association is not necessarily stronger at the lower level of proficiency, and it is strongest at Level III, which goes against the prototype hypothesis. Also noteworthy is the fact that for each level, the association of past marking and lexical aspect is in the order of achievement -^ accomplishment -> activity, which is consistent with the prediction,49 for progressive, the order is activity -> accomplishment -> achievement -> state, for each level.50 Finally, there are two longitudinal studies that are relevant. Rohde (1996) analyzed data from two German children acquiring English in the US, and found that their use of past tense was strongly associated with achievement verbs. This association was particularly conspicuous for the regular past (-ed): for the 6-year-old child, only achievement verbs during the first four months of recording received -ed marking, and in the fifth month, non-achievement verbs (in this case, states) appeared. For the 9-year-old, the trend was not that clear, but only achievements received -ed marking in the first month, and during the next three months, achievements still received by far the most frequent -ed marking; in the fifth month other verbs appeared. Overall, the development of regular past is very consistent with the prototype hypothesis. However, irregular past showed some deviation from the prediction. In the first month, both achievement verbs and non-achievement verbs (states) appeared for both children, although overall achievements were the most frequent category. Rohde also found that the earliest progressive marking was strongly associated with achievements as well as activities, and for the 9-year-old, achievements showed the highest association with progressive inflections during the first two months. He suggested that children were using progressive with achievements to express future tense (e.g., I'm stealing. Why is John not coming?). This goes against the prototype hypothesis, which predicts that progressive morphology is initially associated with activity verbs to denote action in progress. Another longitudinal study by Lee (1997) examined the acquisition of progressive and past morphology with lexical aspect by two Korean learners of English in Hawaii. The following table, calculated from her

84 Acquisition of English Tables 18 to 20, shows the pattern of development for the two learners (a 10-year-old and a 14-year-old). The data were based on token counts. Table 4.3. Percentages of lexical aspect by morphological marking in Lee (1997)

Tl*

T2

T3 T4 T5 T6

14-year-old 100 71 78 75 83 77 10-year-old NA** NA 60 50 62 67 progressive marking with activity verbs> 14-year-old 100 50 80 80 57 50 10-year-old NA 75 71 83 89 0

T7 53 63

T8 48 50

T9 69 62

T10 54 63

57 100 100 78 71 100 83 93

* Tl=Time 1, T2=Time 2, etc. ** N A indicates there were no data in the study on the token of the marking in question.

The table indicates that (a) there are predicted associations between telic verbs and the past tense marking, and between activity verbs and the progressive marking, and (b) the spread from prototype to nonprototype is not transparent. One might argue for such a spread for the past tense marking for the 14-year-old learner, but not for the 10-yearold, and not for the progressive for either child. To summarize, the following generalizations can be derived from the above studies of the acquisition of tense-aspect morphology and lexical aspect in English as a second language: (1)

(2)

(3)

Learners tend to use the prototypical combinations (past with punctual/telic verbs, and progressive with activity verbs) most frequently. The development of prototype to non-prototype is not observed in the cross-sectional and longitudinal studies that used oral or written production data. In fact, for many cases the association is stronger for the intermediate than for the elementary level learners. The results from the paper-and-pencil tests on the use of past tense forms in obligatory contexts are generally consistent with the prediction of the prototype hypothesis.

Apparently, these generalizations are not completely consistent with the patterns from LI acquisition, especially with regard to the spreading

Second language acquisition 85 of prototypes to non-prototypes. In the next section, we will discuss general issues in L2 acquisition, some of which are relevant to such discrepancies between LI acquisition and L2 acquisition.

4.4.3. General issues in L2 acquisition of tense-aspect morphology in English Effects of LI transfer None of the studies reviewed above specifically addressed the issue of L l transfer, and thus how learners' first language influences the L2 acquisition process is still a wide open question. However, a recent study by Quick (1997) was specifically designed to address the question. In a cross-sectional study Quick (1997) examined ESL learners in the US from three different LI groups (Chinese, Japanese, and Spanish; N=103) to test how their LI influenced their acquisition of progressive marking in English. She hypothesized that Spanish learners, as compared with Chinese and Japanese learners, would more easily acquire progressive marking with achievement verbs (e.g., I'm leaving) since the combination is allowed in Spanish, but not allowed in Chinese and has a different meaning in Japanese (i.e., resultative meaning; see 2.3.2 and Chapters 5 and 6 for details). The results from two tasks (verb form change and grammaticality judgment) supported the hypothesis: the Spanish learners consistently (i.e., at all three levels of proficiency) performed better than the Chinese or Japanese learners with respect to progressive marking with achievement verbs. In a third task, Quick asked learners to describe pictures, for which the obligatory context is more difficult to determine and harder to interpret than in the other two tasks. Here too, the association of the progressive with achievements was consistently higher for the Spanish learners than for the Chinese and Japanese learners, except at the advanced level where the Japanese learners showed a higher association than the Spanish learners. Quick's data suggest that non-prototypical uses of the progressive (i.e., with achievements) are facilitated by the similarity between LI and L2 aspectual systems. Needless to say, further research is needed to determine the detailed effects of LI transfer in many other cases of tense-aspect acquisition.

86 Acquisition of English

Discourse grounding and past/progressive morphology Research in L2 acquisition of English has reported correlations between discourse grounding and learners' use of tense-aspect morphology. Briefly, foreground clauses narrate the main story-line, which moves the story forward, whereas background clauses describe background information relevant to the main story-line. In her cross-sectional study (N=37) Bardovi-Harlig (1995b) showed that ESL learners tend to produce past marking more frequently in foreground clauses than in background clauses in oral or written narratives. Thus, in addition to the semantic features of [+dynamic, +telic, +punctual], the discourse notion of [+foreground] appears to contribute to L2 learners' past tense marking. This type of result is in a sense expected, given that it has been suggested that foreground clauses are universally associated with telicity, punctuality, and completedness, while background clauses are universally associated with durativity and lack of telicity (Hopper 1979; Reinhart 1984). A natural question arises here: Which factor is more important for L2 learners - lexical aspect or discourse grounding? Bardovi-Harlig (1998) investigated this question using the same data from BardoviHarlig (1995b). She showed that lexical aspect and discourse grounding contribute equally. Her conclusion is quite compatible with our prototype/connectionist account: in L2 speech production, when the semantic and discourse cues converge and the activation reaches the threshold level, then a past tense form is produced (see also Bates and MacWhinney 1982, 1987 for a cue-based account of language acquisition).

Child-adult differences in tense-aspect acquisition Here, we attempt to make sense of the difference observed between LI and L2 acquisition of tense-aspect morphology in English discussed above. A consistent finding from LI acquisition is that children acquiring English initially start with the prototype and then gradually expand the application of the inflections to non-prototypes (Shirai 1991; Shirai and Andersen 1995). This, however, was not found in L2 acquisition. Although the L2 data reviewed in 4.4.2 were consistent with the Aspect Hypothesis almost without exception in that the prototypical combinations are more frequent than non-prototypical combinations at any particular level of development, the cross-sectional studies and longitudinal studies that analyzed production data did not support the hypothe-

Second language acquisition 87 sized spread from prototype to non-prototypes (Robison 1995; BardoviHarlig and Bergström 1996; Rohde 1996 especially for progressive; Lee 1997). These studies indicate that the associations often strengthen as learners' proficiency level increases, in sharp contrast to the one-way relaxation of the initial restriction to the prototype observed in spontaneous production data of LI English. However, one type of L2 study that did show data consistent with the prototype hypothesis is the crosssectional study in which the learners were tested for their ability to supply past tense for obligatory contexts. There we almost always see the predicted accuracy order of achievement/accomplishment -> activity -> state for each proficiency group, and the accuracy rate increases as proficiency level increases51 (Bardovi-Harlig 1992; Bardovi-Harlig and Reynolds 1995). We suggest that the difference between LI and L2 acquisition comes from the higher capacity of L2 learners who use rote-learned forms in early acquisition. In both LI and L2 acquisition, learners use both holistic rote learning and analytic learning (Bates, Bretherton, and Snyder 1988; Bloom, Lifter, and Hafitz 1980; Karmiloff-Smith 1986; Krashen and Scarcella 1978; MacWhinney 1978; Peters 1983). Some researchers suggest that L2 learners with a higher memory capacity and possibly a lower analytic ability tend to rely more on the data-driven, rote-learning strategy in acquisition (Hakuta 1974; Huang and Hatch 1978; Klein, Dietrich, and Noyau 1995). For example, Huang and Hatch (1978) reported that their 5-year-old Chinese learner acquiring English in the US produced long sentences such as "It's time to eat and drink" as a memorized phrase when his productive language was still lacking correct use of copula (e.g., "This... kite."). In conversation and writing tasks for L2 learners, where learners occasionally need to produce forms beyond their capacity, they sometimes access the form that is easily available to them (probably through rote memory). Since highfrequency forms are easily activated in learners' mental lexicon, they tend to produce such forms more frequently. Thus we suggest that the weaker association between past and achievements (or between progressive and activities, at the early stage of acquisition) observed with L2 learners at lower proficiency levels is due to the use of these highfrequency forms that learners access and produce without really knowing or controlling the semantics of the morphological forms associated with them. In other words, we claim that early on these are haphazardly produced forms before the actual form-meaning relationship is acquired.

88 Acquisition of English Robison (1995) reported an example that supports this claim. He stated that an unexpectedly high correlation between achievement and progressive marking is due to the form going - a high frequency form in the input. He states that "[t]he elevated application of -ing to punctual events thus appears to be artificially induced. With increasing proficiency level, this spurious inflation reverses to a deficit" (Robison 1995: 357). Thus, L2 learners can easily produce inflected forms due to their higher memory capacity even when they are not sure about the semantics of the inflections.52 Of course this is not the whole story. Second language acquisition involves many external factors that can come into play. Rhode's (1996) finding that early progressive was associated with achievements is a case in point. His data showed that older children as L2 learners use the progressive marker with achievement verbs as a convenient modal marker for expressing intention in interactional contexts.53 These L2 learners may eventually recover from this type of overgeneralization, but such an incorrect form-meaning mapping at some point in development is not surprising. In contrast, in LI acquisition, this type of overgeneralization is rare. LI children may be highly constrained by the prototype and unable to use -ing with achievements, a non-prototypical use. Older children, especially in L2 contexts, can use their metalinguistic knowledge, and once they notice that use of -ing in the input includes the expression of future intention, they may start to use it productively. Of course, at this point our analysis is highly speculative; further research is needed, especially on conditions under which L2 learners use progressive markers with achievements to express modal meaning (see Shirai and Kurono 1998 for further discussion). Another difference between LI and L2 acquisition is the absence of the role of discourse grounding in LI acquisition of tense-aspect morphology, particularly during early stages of LI acquisition. Obviously, this is due to the absence of narrative structure in the speech of the young children whose morphological development has been studied. Whereas L2 learners already have their narrative skills in their LI and can mark foreground and background by some means, very young children generally do not have such a skill. Young children whose past and progressive marking is still restricted to prototypes will find it difficult to tell a narrative. However, Herman and Slobin (1994: 4-5) showed that the backgrounding function of the progressive was already emerging for a 3-year-old child telling a frog-story narrative. If their

Second language acquisition 89 result is not an exception, then it may be possible for us to assess the relationship between discourse grounding and past-progressive marking in future studies. Finally, a major difference between LI and L2 acquisition is that LI learners eventually create native semantic representations of tense-aspect markers, whereas adult L2 learners may not be able to do so. L2 learners' ultimate attainment of the language may not reach native competency, even though the development from prototype to non-prototype may be similar for both LI and L2. For example, Coppieter's (1987) study showed that native-like L2 speakers of French still do not have native-like control of aspectual marking. The above review also shows that L2 learners either overuse aspect markers (e.g., incorrect stative progressives), or underuse them (e.g., display higher association of the progressive with activities than do native speakers), whereas native speakers have more flexible control of morphology and can use it in nonstandard, non-prototypical ways - for example, correct control of the stative progressive (Andersen and Shirai 1994). The difference between LI and L2 learners in their ultimate level of acquisition of tense-aspect morphology should be considered along with the Critical Period Hypothesis. We will return to this point in Chapter 8. 4.5. Summary In this chapter, we have presented empirical data on LI and L2 acquisition of progressive aspect and past-perfective morphology in English, in connection with the acquisition of the lexical aspect of verbs to which the morphology applies. For both child and adult learners we demonstrated the existence of an interaction between morphology and the lexicon in the patterns of association between uses of grammatical morphemes and lexical aspect categories. In particular, we have shown that the progressive marker -ing has a tendency to occur initially with only atelic, activity verbs, while the past-perfective forms (including both regular -ed and irregular forms) occur with only telic, resultative verbs. We advocated a prototype account of why there are such initial associations, and how prototypical uses of morphology gradually spread to non-prototypes. We also examined alternative explanations of the bootstrapping and nativist perspectives on the acquisition of grammatical and lexical aspect. Finally, we discussed empirical and theoretical issues in the acquisition of

90 Acquisition of English English as a second language, and how LI acquisition and L2 acquisition are similar and different. In this chapter we have proposed an account of how aspect, in connection with tense, is acquired by LI and L2 learners of English. We do not propose by this, however, that exactly the same processes should be observed in the acquisition of other languages. We expect to see similarities, but also major differences between languages, which probably result from typological, structural differences between the languages to be learned. In the next two chapters, we will discuss how the acquisition of Chinese and Japanese resembles and differs from that of English, and will show how our analysis in English can be applied to explain the learning of these languages.

Chapter 5 Acquisition of aspect in Chinese The last several decades have witnessed a rapidly growing interest in the study of the Chinese language, not only the structural properties, but also the processing and the acquisition of this language.54 Chinese has become a major challenge to many theories of linguistics and psycholinguistics, because the language offers unique properties of grammatical structures and new windows on language processing and acquisition (see Wang and Tzeng 1983; Hoosain 1991; Li, Bates, and MacWhinney 1993; Li 1996, 1998; Li and Yip 1998, for overviews and discussions). These properties and windows, quite often, do not fit with standard, classic, linguistic or psycholinguistic explanations. They tend to point to linguistic variation or language specificity rather than linguistic universality (see Li 1996, 1998 for an argument). Chinese differs significantly from most Indo-European languages with respect to phonological, lexical, and syntactic structures. For example, the language does not have morphological markers that indicate tense, number, gender, or case; it involves a high degree of ellipsis, allowing both null subjects and null objects; and it has a relatively free word order - in sum, a typical Chinese sentence in natural speech could sound telegraphic in a richly inflected language when literally translated. Despite these properties that point to variation and specificity and the processing consequences therein, Chinese does offer a rich set of grammaticized aspect markers, in the form of free-standing morphemes, or what are often called "particles". These markers carry the functions of inflectional morphemes, but differ from inflections in that they are not phonologically bound to the verb stems as they are in inflectional IndoEuropean languages (hence "free-standing"). However, they are not entirely free morphemes either, as they are often pronounced in a neutral tone, one signal of binding.55 As in other languages that lack inflectional morphology (e.g., Thai and Vietnamese), verbs in Chinese are not marked for event time difference, because there is no grammatical category of tense in this language. Thus, aspect in Chinese is not confounded with tense, as is in some languages (e.g., -ed in English encodes both perfectivity and pastness, and imparfait in French both imperfectivity and pastness).56

92 Acquisition of Chinese Three basic points about aspect in Chinese should be noted here, since they affect our discussion in the rest of this chapter. First, the grammatical and semantic functions of Chinese aspect markers have been notoriously difficult to characterize. Numerous publications have addressed the linguistic functions of the aspect markers. In spite of these efforts, however, there is still no general consensus on the use of these markers. In the absence of such general agreement, we devote the first part of this chapter to linguistic analysis of aspect in Chinese, which will serve as a basis for our study of acquisition.57 Second, most previous studies of aspect in Chinese have concentrated on grammatical aspect, but not lexical aspect. Because we are concerned in this book with the role of both lexical and grammatical aspect, we will present our analysis of lexical aspect on the basis of a few existing accounts (e.g., Tai 1984; Li 1990; Smith 1997). Third, the acquisition of aspect has received relatively little attention in the empirical literature, in contrast to other grammatical or lexical domains that have attracted more attention (see Erbaugh 1992; Lee 1996, for an overview of these other domains). This is a quite odd situation, given that aspect stands out as one of the few distinctive grammatical categories in Chinese and that it has received much attention in the linguistic literature. To date, there are only a few studies of first language acquisition of aspect in Chinese, and there are very few published studies of second language acquisition of aspect in Chinese.58 Our empirical coverage of the acquisition facts in Chinese will thus be limited to first language acquisition only. 5.1. Grammatical and lexical aspect in Chinese 5.1.1. Grammatical aspect Grammatical aspect in Chinese has been extensively discussed in the linguistic literature (cf. Chao 1968; Li and Thompson 1981; Smith 1997). Several aspect markers have been studied in detail, including the progressive marker zai, the durative marker -zhe, the perfective marker -le, and the experiential marker -guo. Roughly speaking, the first two are imperfective aspect markers, and the last two are perfective aspect markers. We discuss each of these two types of markers in turn.

Grammatical and lexical aspect 93 The imperfective aspect markers As discussed in 2.1.1, imperfective aspect is not a single category in language. Comrie (1976) divided imperfective into habitual versus continuous, and further subdivided continuous into progressive versus nonprogressive. Since Chinese does not have grammatical markers for habitual aspect, there is only the contrast between progressive and nonprogressive under the imperfective category. Progressive aspect in Chinese is marked by the morpheme zai, nonprogressive by the morpheme -zhe. Here we use "durative" instead of "nonprogressive" for -zhe, given the nature of the contrast between the two markers.59 Zai has had a long historical development, appearing first as a verb, then as a locative preposition, and only recently as an imperfective aspect marker (see Li 1988, 1993a, for a discussion).60 As a preposition, zai can occur both preverbally and postverbally, but as an aspect marker it can occur only preverbally (Zhu 1981; Li 1990, 1993a). Its main function as an aspect marker is to indicate that an action or event is in progress, hence a progressive marker. On the other hand, -zhe indicates that a situation is viewed as enduring or continuing, often as a background event in a discourse (cf. the function of be + (V)-ing in a sentence such as "While Mary was reading a book she heard a noise in the kitchen"). As a durative marker, -zhe is used most naturally with verbs that specify a state. Zai, in contrast, cannot be used with stative verbs.61 Chinese also has a third imperfective marker, -ne, which is not traditionally treated as an aspect marker (note that this marker is identical in form to the sentence-final particle -ne}. Its aspectual function has been recognized only recently (Chan 1980; Liu 1985; Ma 1987). The neglect of -ne as an aspect marker might be partly due to the fact that -ne frequently co-occurs with zai and -zhe, and partly due to pragmatic constraints on its use in spoken language. Liu (1985) suggested that -ne is used to mark progressive aspect only in answers to questions in colloquial dialogues. However, Ma (1987) argued that -ne is actually the main device for imperfective aspect in the spoken Beijing dialect. In our opinion, -ne encompasses both progressive (zai) and durative (-zhe) meanings, and is a general imperfective marker. In this book, we treat zai, -zhe, and -ne all as imperfective aspect markers, although we recognize the constraints on their use (e.g., zai cannot be used with stative verbs and -ne occurs only colloquially). Sentences (1) to (3) are examples of the three aspect markers in use, respectively:

94 Acquisition of Chinese (1)

Zhangsan zai chi-fan. Zhangsan ZAI eat-rice Zhangsan is eating his meal.

(2)

Zhangsan ai-zhe Lisi. Zhangsan love-ZHE Lisi Zhangsan loves Lisi.

(3)

Zhangsan (zai) kan baozi -ne. Zhangsan (ZAI) read newspaper -NE Zhangsan is reading a newspaper.

The perfective aspect markers In contrast to zai, -the, and -ne, the aspect marker -le is perfective: it presents a situation in its entirety, as an event bounded at the beginning and the end, with particular focus on the endpoint of the situation.62 Even though -le is the most intensively studied aspect marker in Chinese, there is still no general consensus on its grammatical and semantic functions. Traditionally, -le has often been characterized as marking completion (e.g., Chao 1968). But some researchers argue that -le does not by itself indicate a completed event or action, and the meaning of completion often comes from its perfectivity meaning (Li and Thompson 1981), or from the meaning of the verb with which -le occurs (Li 1990; Shih 1990). For example, when the verb encodes a situation with a clear temporal boundary, -le indicates that the situation comes to its natural endpoint, that is, it is completed, as illustrated in (4). But when the verb encodes a situation with no natural boundary, -le signals the termination rather than completion of a situation, as in (5). (4)

Qiche zhuang-dao -le fangzi. car hit-break -LE house The car knocked down the house.

(5)

Xiaoyazi you -le yong. duckling swim -LE stroke The duckling swam.

The example in (4) contains a so-called "resultative verb construction" (RVC; more discussion below) that encodes a telic, resultative

Grammatical and lexical aspect 95 endpoint (i.e., the break-down of the house), and the perfective -le indicates that the end result has been achieved and the event is completed. In contrast, (5) contains an atelic activity verb you-yong 'swim' that encodes no natural endpoint, and -le indicates that the event took place and terminated at some indefinite point. Note that the completion or the termination meaning of -le is the natural consequence of the perfective viewpoint: by focusing on the end rather than the internal structure of the situation, the use of aspect markers realizes particular components of the lexical aspect of verbs (e.g., endpoint).63 Finally, traditional analyses (e.g., Chao 1968; Smith 1991) have also designated -le as indicating the inception or inchoativity of a stative situation, for example, as in Zhangsan mingbai -le (Zhangsan started to know). Variations of this kind in the interpretation of -le provide a good example of how grammatical aspect interacts with lexical aspect to determine the final aspectual interpretation of a sentence; we will return to this point in more detail in 5.1.3. Another perfective marker, -guo, has generally been characterized as an experiential marker: it indicates that an event has been experienced at some indefinite time, usually in the past,64 and that the resulting state no longer obtains at the time of speech (Chao 1968; Li and Thompson 1981). As a perfective marker, it is concerned with the external rather than the internal structure of the situation. According to some researchers, -guo is more of a "perfect" than a "perfective" marker, because of its involvement with two distinct times, reference time and speech time, and its indefmiteness characteristic that differentiates -guo from -le. Examples (6) and (7) serve to illustrate the differences between -le and -guo. (6)

Lisi da-po -le yi-ge beizi. 65 Lisi hit-break -LE one-CL cup Lisi broke a cup.

(7)

Lisi da-po -guo yi-ge beizi. Lisi hit-break -GUO one-CL cup Lisi once broke a cup.

Sentence (6) may refer to a situation in which the broken pieces of the cup are still laying on the ground; -le indicates a completed action of breaking and the result therein. Sentence (7), however, is appropriate

96 Acquisition of Chinese only when it refers to an experience that Lisi had - that he or she has once broken a cup (at some indefinite time in the past), and that the resulting state of breaking no longer holds at the time of speech. This last characteristic of -guo — the resulting state no long obtains - is what Chao (1968) and Smith (1997) called the "discontinuity" meaning of -guo. 5.1.2. Lexical aspect As we pointed out, traditional linguistic analysis of aspect in Chinese has focused on grammatical aspect, and lexical aspect has received relatively little attention. Traditional analyses of verb categories do exist, but they have been mostly concerned with the grammatical functions of verbs rather than their semantic classification. There are two earlier studies that bear some relevance to our questions here: Teng (1974), though not concerned with aspect, attempted to divide Chinese verbs into actions, states, and processes within the semantic framework of Chafe (1970). However, Teng's classification of verbs is not based on inherent temporal features such as telicity or durativity, and cannot be regarded as a categorization of lexical aspect in Chinese. Tai (1984), in a more systematic analysis, attempted to apply Vendler's categorization scheme to the study of lexical aspect of verbs in Chinese. According to Tai, Chinese has roughly the same types of verbs as English. But one striking difference between the two, Tai noted, is that Chinese often uses resultative verb constructions (RVC) to describe events that are specified in English with accomplishment and achievement verbs. Unlike English accomplishment verbs, Chinese RVCs cannot be marked with progressive aspect. For example, in sentence (8), xue 'study' is an activity verb, so it is compatible with the progressive marker zai; but xue-hui 'study-know' in sentence (9) is an RVC, and so it cannot be combined with zai. (8)

Yuehan zai xue Zhongwen. John ZAI study Chinese John is studying Chinese.

(9)

*Yuehan zai xue-hui Zhongwen. John ZAI study-know Chinese John is learning Chinese.

Grammatical and lexical aspect 97 The incompatibility of RVCs with ιαί probably reflects the construction's increasing emphasis on the result rather than the action component of the verb in its historical development (Li 1987). This incompatibility prompted Tai to argue that Chinese does not have accomplishment verbs. Tai further argued that Vendler's achievements are also realized in Chinese in the form of RVCs, for example, zhao-dao (seek-reach) 'find', shou-dao (collect-reach) 'receive', and kan-jian (look-perceive) 'see'. Recently, however, Tai's position that Chinese does not have accomplishments has been challenged (e.g., Li 1990; Smith 1997; Yang 1995). Li (1990) proposed that Chinese has accomplishment verbs similar to those of English, for example, hua yi-fu hua 'draw a picture', gai yi-suo fangzi 'build a house', and pao yi-bai mi 'run 100 meters', in that they can take progressive marking zai to denote the action-in-progress meaning and at the same time indicate the temporal boundary of the event. In this chapter, we retain the category of accomplishments for Chinese, but we agree with Tai that the RVCs lack a progressive meaning, so we treat them as a subclass of achievement rather than accomplishment verbs (cf. Smith 1997, who argued that all the Chinese RVCs belong to accomplishments). Achievement verbs also include, in addition to RVCs, other monosyllabic telic or resultative verbs, such as dao 'arrive', sa 'spill', and diao 'drop'. Thus, achievement verbs in Chinese incorporate both punctuality (lack of progressive meaning) and resultativity (marking endpoint or end result), as is the case in other languages. Verbs that mark only punctuality but not resultativity can be found in Chinese with what Smith called "semelfactives" (see 2.2 and 2.3). According to Smith, semelfactive verbs are punctual verbs that indicate repeated events when combined with progressive aspect; this is unlike achievement verbs that indicate preliminary stages of the event when combined with progressive marking (see 2.3). Examples like cough, tap, and knock qualify as good semelfactive candidates in English. Chinese also has semelfactive verbs, such as tiao 'jump', qiao 'knock', ti 'kick', and, consistent with Smith's analysis, they indicate repeated events when they are combined with zai. For example, the sentence Zhangsan zai qiao men 'Zhangsan is knocking on the door' indicates a continually repeated action of knocking the door, not a one-time knocking action. In contrast to semelfactive verbs, achievement verbs in Chinese cannot be combined with zai at all. Still another lexical aspect class in Chinese is discussed by Li (1990): mixed telic-stative verbs, including chuan 'put on/wear', na 'take/hold',

98 Acquisition of Chinese ti 'pick up/carry', bei 'put on shoulder/carry', gua 'hang', ji 'tie', and kun 'bind' (for a complete list of related verbs, see Teng 1979 and Jaxontov 1988). These verbs encode either the process of a telic action or the state resulting from that action, depending on which aspect markers they take. In English, separate lexical items are usually needed to express such a pair of meaning. For example, with the progressive marker zai the verb chuan corresponds to English "put on", whereas with the durative marker -zhe or the perfective marker -le, chuan corresponds to "wear", as shown in sentences (10) to (12). Mixed telic-stative verbs demonstrate clearly the interaction between lexical and grammatical aspect. (10)

Ta zai chuan yi-jian xin yifu. he ZAI put-on one-CL new garment He is putting on a new garment.

(11)

Ta chuan-zhe yi-jian xin yifu. he wear -ZHE one-CL new garment He is wearing a new garment.

(12)

Ta chuan -le yi-jian xin yifu. he wear -LE one-CL new garment He is wearing (as a result of having put on) a new garment.

Integrating the analyses of Vendler, Tai, Smith, and Li, we arrive at six different categories of lexical aspect in Chinese: activities, states, accomplishments, achievements (including RVCs), semelfactives, and mixed telic-stative verbs. Note that although the first four categories are the same as in Vendler's categorization, it does not mean that they have the same syntactic and semantic consequences as they do in other languages (see Smith 1997). For example, achievement verbs accept imperfective -ing marking in English to denote the preliminary stages, but they do not accept imperfective marking with either zai or -zhe in Chinese. Similarly, stative verbs do occasionally accept -ing in English, but they do not accept zai at all in Chinese. Thus, examples (13a-b) are ungrammatical in Chinese, while their translations in English are perfectly okay.

Grammatical and lexical aspect 99

(13) a.

b.

*Zhangsan dao -zhe jia (or zai dao jia). Zhangsan arrive -ZHE home (or ZAI arrive home) Zhangsan is arriving home. *Zhangsan zhe-ji tian zai xin gui. Zhangsan this-CL day ZAI believe ghost Zhangsan is believing in ghosts these days.

These language-specific characteristics of lexical aspect of verbs are clearly reflected in the relationship between lexical and grammatical aspect, which brings us to the next section.66

5.1.3. Interactions between lexical and grammatical aspect Our previous discussion has indicated that lexical aspect and grammatical aspect interact with each other in Chinese. On the one hand, the interpretation of a grammatical aspect marker often depends on the lexical aspect of the verb it marks. For example, the perfective marker -le may indicate completion, termination, or inception of a situation, depending on whether the lexical aspect of the verb encodes activities (termination), accomplishments or achievements (completion), or states (inception). On the other hand, the interpretation of the lexical aspect of a verb often depends on the co-occurring grammatical aspect marker. For example, the two meanings of the mixed telic-stative verbs are brought out by the use of different aspect markers, as discussed above. These interactions, when viewed from a slightly different perspective, suggest ways in which lexical aspect and grammatical aspect are correlated or associated. Thus, the telic meaning of a mixed telic-stative verb is compatible with the progressive marker zai, whereas its stative meaning is most compatible with and brought out by the durative marker -zhe. There are several natural associations between lexical and grammatical aspect in Chinese: in particular, between the perfective marker -le and accomplishment/achievement verbs, between the progressive marker zai and activity verbs, and between the durative marker -zhe and stative verbs. In fact, as noted above, zai is ungrammatical with stative verbs and with RVCs (a subtype of achievement verbs). These combinatorial constraints are naturally explained by Comrie's "naturalness of combination" principle (see 2.3.1). Although there is yet no empirical re-

100 Acquisition of Chinese search using quantitative methods to determine the exact frequency with which given grammatical aspects combine with given lexical aspects, the general tendency for the "natural combinations" is clear in Chinese.67 Combinations of grammatical and lexical aspects that are relatively more natural, in the sense just described, occur more often than those that are less natural. For example, in Chinese, the perfective -le may occur more often with accomplishment/achievement verbs than with activity verbs because the former, which encode situations with a clear end result or endpoint, provide ideal instantiations of the perfective meaning. A given combination may not be altogether prohibited, such as perfective marking on stative verbs, but it may occur less frequently. In this study, we are interested in whether Chinese-speaking children are sensitive to these natural (and less natural) combinations, and why, if they are. 5.2. Lexical and grammatical aspect in early child Mandarin: A longitudinal study As we pointed out earlier, there are, oddly, few studies that have directly examined the acquisition of aspect in Chinese. One exception was Erbaugh (1978, 1982; see also Erbaugh 1992). In a longitudinal study of four Chinese-speaking children from ages 2 to 3, Erbaugh found that the perfective marker -le emerged first, and the number of early -le tokens was immense: in 64 hours of her data, there were 2,300 instances of -le, an average of one -le every 1.7 minutes. In contrast, there were relatively fewer uses of zai (108 tokens), -zhe (50 tokens), and -guo (34 tokens) in the same data. Erbaugh suggested that a similar distributional bias is also present in the adult speech. According to Erbaugh (1992), young Chinese-speaking children acquire the use of aspect in four cumulatively overlapping stages. The first stage, a general boundedness stage, occurs before age 2;4, during which time children focus on the completion and current relevance of situations, as indicated by their use of the perfective marker -le. The second stage occurs between 2;4 and 2;9, and is characterized by the child's double-marking of completion and result. The third stage is a stage of expressing sequenced temporal relations, occurring between 2; 10 and 3;4. At this stage children are able to coordinate two or more events within a sentence, and the progressive aspect becomes important for the child in describing the sequence of events. The final stage,

Longitudinal study 101 starting from about 3;4, marks the child's ability to describe backgrounded events and use narratives. Chinese-speaking children's early use of the perfective -le at the first stage is consistent with our previous discussion of the acquisition of English and other languages. Erbaugh reported that children's earliest uses of -le are to indicate visible actions, often resultative or a change of state, as in da-po-le 'hit-break-LE' (I have broken [it]). About 96% of children's uses of -le in Erbaugh's data referred to the "immediate past", many of which are situations like breaking a cup or finishing a block tower. Most interestingly, some novel uses of -le also conformed to this pattern, even though the particular word to which -le was applied is not a change of state verb, as in (14) (produced by a child, at age 2; Erbaugh 1992: 426). (14)

Wo jiqiren -le. I robot -LE I have become a robot.

In this example, the child used jiqiren 'robot' as a verb, treating the noun as marking an achievement of his action. There were several examples of this type in Erbaugh's data.68 Although Chinese children's early use of -le was selective on the lexical aspect of verbs, Erbaugh reported that children did not restrict -le to only punctual verbs;69 they also used -le with nonpunctual verbs equivalent to roll, fly, talk, cry, draw, and play. However, her data suggested that -le occurred more often to describe punctual events. In addition, many activity verbs with which -le occurred were part of the RVC construction, such as he-wan-le 'drink-finish-LE', hua-hao-le 'draw-finishLE', and na-dao-le 'grasp-get-LE', hence both punctual and resultative (i.e., achievement) according to our above analysis (see 5.1.2). In contrast to the early emergence of the perfective -le, the use of the progressive marker zai was rare before age 3 in children's productive speech. This also contrasts with the early emergence of the progressive marker -ing in English (see 4.1.1). Erbaugh noted that only nine instances of zai occurred before age 2;4. Her data showed that children in most cases respected the linguistic constraint on the use of zai from the beginning: no child used zai with stative verbs such as in *zai tang (ZAI hot) or *zai guai (ZAI well-behaved), even though such errors are common among foreign learners of Chinese. One overgeneralization of

102 Acquisition of Chinese zai did occur; a child at age 2;4 used zai with a reduplicated verb form of an achievement verb. The child held as she also said slowly *zai diaodiao 'is dropping' to act out a slow fall; the child apparently treated diao, a punctual and resultative verb, as a durative verb. As compared with -le and zai, the aspect marker -guo occurred much later in Erbaugh's data, after about 2;6. Most of children's uses of this marker described extended, recurring activities, such as sleeping, eating, and peeing (e.g., I have already had a sleep, or I have already eaten), rather than unique experiences. Guo was also rare in general, and there was only a total of 34 instances in all of her data. Erbaugh (1992) suggested, along with Li (1990), that the distinction between process and result, as postulated in Slobin's (1985) Basic Child Grammar, develops gradually for Chinese-speaking children, and does not exist prelinguistically. One piece of her evidence was that children at first would not respect the constraint on the use of the Chinese object marker ba with resultative verbs; for example, they would use ba with non-resultative process verbs, such as ca 'wipe', xi 'wash', and shu 'comb'. Erbaugh explained that this incorrect combination might come from the strong transitivity involved in the manipulation of toys in actions such as washing and combing, which overrode the linguistic constraint on the requirement of the ba construction to be resultative. This result, however, differs somewhat from Li's (1990, 1993b) experimental data in which children's use of ba always respected the constraint (see 5.3.2 below). It is possible to reconcile this difference, because Erbaugh's data come from younger (2- to 3-year-old) children, while Li's data come from older children (3- to 6-year-old) in an experimental setting.

5.3. Lexical and grammatical aspect in child Mandarin: An experimental study Previous crosslinguistic studies have provided important information about how children acquire lexical and grammatical aspect in various languages. However, most of these studies have been based on naturalistic observations, in which only children's productive, but not receptive, ability is examined. But as Clark and Hecht (1983) pointed out, evidence from different modalities may give rise to qualitatively different pictures of what children know about a particular domain of language; thus, it is desirable to study both production and comprehension in child Ian-

Experimental study 103 guage. In the following, we present an experimental study of Chinesespeaking children's acquisition of aspect, investigating both children's production and their comprehension of lexical and grammatical aspect. Our discussion of the language-specific properties of Chinese in 5.1 shows that Chinese provides an interesting test case in light of current theorizing about the acquisition of aspectual systems. For example, Chinese provides a set of semantic distinctions in the lexical aspect of verbs that we can use to further test the Language Bioprogram Hypothesis and the Basic Child Grammar hypothesis. For instance, "process" in the process-state distinction (Language Bioprogram) and the processresult distinction (Basic Child Grammar) is encoded by activity verbs, while "state" and "result" are encoded by stative and accomplishment/achievement verbs (especially RVCs), respectively. The term process is not used in exactly the same way in the Language Bioprogram Hypothesis as in the Basic Child Grammar hypothesis, with neither hypothesis explicitly adopting Vendler's classification. The process side of the state-process contrast remains somewhat vague, but it seems to have to do with the presence of action or dynamicity in the meaning of a predicate. On the other hand, the process side of the process-result contrast is defined as a temporal perspective (nonpunctual, noncompletive, and ongoing) that contrasts with the result perspective (punctual, completive) - both perspectives are superordinate to language-specific grammatical or semantic categories. While recognizing that the notion of process is not identical in the two hypotheses, we would argue that "activity" verbs are good representatives of what is meant by the term in both hypotheses. This means that by comparing activity verbs with stative verbs in children's acquisition of aspect markers we can test the Bioprogram's state-process distinction, and by comparing activity verbs with accomplishment and achievement verbs we can test the Basic Child Grammar's process-result distinction. If the state-process or the processresult distinction is bioprogrammed or otherwise prelinguistically salient, and serves as semantic template that strongly attracts grammatical mapping, children acquiring Chinese should show early sensitivity to the differences between activity and stative verbs, or between activity and accomplishment/achievement verbs; that is, they should mark verbs from these different classes differently. Evidence for such sensitivity would be compatible with hypotheses about prelinguistically salient aspectual distinctions, but not conclusive, since differential marking patterns could also be learned on the basis of

104 Acquisition of Chinese the distribution of aspect markers in the input. So it is also important to examine how children handle distinctions that are marked in Chinese differently from the way they are marked in other languages. For example, unlike English, Chinese does not accept imperfective markings with achievement verbs (e.g., da-po "hit-break") or with posture stative verbs (e.g., zuo "sit" and zhan "stand"). If children's obedience to aspectual distinctions is actually caused by their sensitivity to patterns in the input rather than to the supposedly prelinguistic factors, they should be just as sensitive to language-specific constraints as to the distinctions of the Language Bioprogram or Basic Child Grammar. In what follows, we report three experiments that examined children's skills in comprehension, production, and imitation of lexical and grammatical aspect in Chinese.70 5.3.7. Experiment 1: Comprehension In this experiment, we systematically tested young Chinese-speaking children's comprehension of three aspect markers: the perfective marker -le, and the imperfective markers zai and -zhe. We did not test children's acquisition of the perfective marker -guo in this study, because the indefinite, experiential meaning of this marker (see 5.1.1) makes it difficult to model in an experimental setting. Method Participants. A total of 135 children from several kindergartens of Beijing University and Qinghua University in Beijing, China took part in the experiment. In a pilot test with 20 children between ages 2 and 6, it was found that children younger than 4 years of age were unable to perform the task adequately. Children in this study were thus divided into three age groups of 45 each: 4-year-olds (mean age = 4;2, range = 3;11 to 4;4), 5-year-olds (mean age = 5;1, range = 4;10 to 5;4), and 6year-olds (mean age = 6;0, range = 5; 10 to 6;4). Half the children were boys and half girls. Although no attempt was made to test the children's general linguistic and cognitive abilities, all child participants appeared normal. Materials. Our materials and procedure were adapted from a sentence-picture matching task used by Weist (1983) and Weist, Wysocka, and Lyytinen (1991) who tested the comprehension of aspectual distinc-

Experimental study 105 tions in child Polish and English. In our study, aspectual distinctions were represented by pairs of contrasting picture stories, with each story made up of two pictures. Both stories of a pair could be described by the same verb; the difference between them corresponded to the contrast between two aspect markers, one perfective (-le) and the other imperfective (zai for progressiveness or -zhe for durativity). The paired stories were bordered with different colors, one red and one green, and some details of the pictures were also correspondingly painted red or green. One was called the "red story" and the other as the "green story", which allowed the two pictures of each story to be referred to as a unit (see Figure 5.1 for an example). These pairs of two-picture stories have two methodological advantages over pairs of single pictures that were commonly used in previous studies, including studies by Weist and his colleagues. First, aspect often has to do with the contour of a situation over time, and the stages of a situation can be suggested more clearly in successive frames than in a single picture. Second, pilot testing revealed a methodological problem with the use of single pictures. When one picture shows an imperfective situation (i.e., an ongoing or enduring situation, such as people skating) and the other shows a perfective situation (i.e., the aftermath of a justcompleted event, such as people leaving the ice), children who ignore the aspect markers or who do not understand them will tend to choose the imperfective picture because it best represents the meaning of the verb itself- e.g., skate. To solve this problem, the pictures in our study were combined into stories in such a way that one picture - which optimally represented the situation named by the verb - was shared by both stories. In the imperfective story this picture was combined with another picture showing an earlier phase of the same situation, and in the perfective story it was combined with another picture showing the event just ending. (Hints were built into both stories to indicate the passage of time; e.g., in the story shown in Figure 5.1, a bystander cat has just entered a room in the first picture and has moved to the center of the room in the second picture.) This arrangement encouraged children to compare stories rather than individual pictures and eliminated their bias for choosing imperfective pictures. In the overall design of the experiment there were three sentences associated with each story pair (the verb plus -le, zai, or -zhe). (see Li and Bowerman 1998 for the complete set of sentences used in the experiment.) Each child heard only two sentences with each pair: the perfective

106 Acquisition of Chinese sentence with -le and, as its imperfective counterpart, a sentence with either zai or -zhe. Across all children, each story was presented with -le I zai versus -le /-zhe an equal number of times (this is done by pseudorandomly arranging all sentence pairs so that the total number of -le I zai and -le /-zhe pairs are balanced). Which of the two sentences - perfective or imperfective - was designated as the target varied systematically, so that for each story the target sentences contained -le, zai, and -zhe an equal number of times and for each child there was an equal number of -le, zai, and -zhe target sentences (with the exception that only -le and -zhe appeared in sentences with stative verbs, since zai cannot be combined with stative verbs in Chinese; see 5.1.2). For each lexical-aspectual verb class (except achievement verbs) there were three stories, each with a different verb. We did not test achievement verbs because they cannot be combined with the progressive marker zai or the durative marker -zhe. We divided accomplishment verbs into two subcategories: resultative verbs that encode the end result of an event (e.g., gai yi-suo fangzi 'build a house' [note that these are not RVCs which we have classified as achievements]), and locative verbs that encode the endpoint of a change of location (e.g., pao yibai mi 'run 100 meters'). The purpose of this distinction was to allow us to assess whether children would treat these subcategories alike; Slobin's temporal perspective "result" encompasses both change of state and change of location, but most of his examples involve resultative verbs, which might well be considered more prototypical than change-of-location verbs for this temporal perspective.71 Each child was presented with a total of 18 pairs of stories, three for each of the following six verb types (see 5.1.2 for how they are defined in Chinese): activity, semelfactive, stative, mixed telic-stative, accomplishment: resultative, and accomplishment: locative. Figure 5.1 presents a sample story pair, which illustrates the stative verb kai "open". Procedure. All children were tested individually. For each story pair, the two stories were laid out side by side on a table in front of the child. The experimenter briefly explained the stories without using the target verb, pointing out that one was the "red story" and the other was the "green story". Then the test procedure was administered. For example, the child was told: "In these two differently colored stories, one story tells that Wu-li-de chuanghu kai-LE [The window in the room opened], and the other tells that Wu-li-de chuanghu kai-ZHE [The window in the room is open]. Now, tell me which story shows Wu-li-de chuanghu kai-

Experimental study 107 LE (or Wu-li-de chuanghu kai-7ME)T (These instructions are translated from the Chinese original.) The child could point to the pictures as in Weist (1983) and Weist, Wysocka, and Lyytinen (1991), but he or she was required to say whether the sentence described the "red story" or the "green story". This procedure was designed to forestall responses based purely on the child's assessment of individual pictures (on the basis of which they would point to only individual pictures).

(la)

(Ib)

Figure 5.1. Sample pair of picture stories used in experiment 1 to test the comprehension of the contrast between perfective and imperfective aspect. Story (la) is described by the perfective sentence Wu-li-de chuanghu kai-LE (The window in the room opened), whereas story (Ib) is described by the imperfective sentence Wu-li-de chuanghu kai-ZHE (The window in the room is open).

Before the testing began, the child was asked to label the color of two identical toys, one green and one red, and those who could not do this were replaced by children who could. Children were also given two warm-up sets of stories and sentences, and the practice was repeated until they clearly understood the procedure. Each experimental session lasted about 30 minutes. Data analysis. A child's response was counted as "correct" if it picked out the picture story described by the test sentence and "incorrect" if it picked out the other story. When children gave no verbal

108 Acquisition of Chinese response (e.g., they only pointed), or changed their answer, their responses were counted as missing. Children who had more than three missing responses were replaced. Our analysis will focus on the contrasts between (1) the progressive marker zai and the perfective marker -le, and (2) the durative marker -zhe and the perfective marker -le. We excluded from the analysis target sentences that combined -zhe with verbs other than Stative and mixed telic-stative, because these combinations are generally used to signal backgrounding information, and so are incomplete on their own. We did not test the combination of zai with stative verbs, as the combination is ungrammatical. The contrasts we are interested in, then, are between zai and -le for all verb types except statives, and between -zhe and -le for statives. Only one category of verbs, the mixed telic-stative verbs, was analyzed with all three aspect markers, zai, -zhe and -le. Responses to the three stories for each verb type were pooled and a loglinear analysis was performed on the frequencies of correct responses for each verb type. Loglinear analysis has become an increasingly important tool for developmental researchers because of its usefulness in analyzing complex frequency data (Gilbert 1981; Green 1988; Kennedy 1992; Knoke and Burke 1980).72 Loglinear analysis fits data to various candidate models that incorporate the effects of one variable (denoted below with {X}), two variables ({X}{Y}), or three variables ({X}{Y}{Z}). For example, the model that incorporates the effect of age is designated {A}, whereas the model that incorporates the individual effects of age and aspect marker is designated {A}{M}, with each variable name in separate brackets. A model that incorporates the interaction effects as well as the individual effects is labeled with all variable names enclosed in one bracket (e.g., {AM}, for the interaction between age and aspect marker). Because the models are hierarchically organized, models with higher-order effects (e.g., interactions between age, aspect marker, and verb type) presuppose the inclusion of the corresponding lower-order effects (e.g., individual effects of age, aspect marker, and verb type, plus three possible two-way interactions).

Results and discussion Table 5.1 shows the percentage of correct responses broken down by age, obtained by dividing the number of correct responses by the total number of responses for each verb type (Accomp/Res = accomplishment resultative verbs; Accomp/Loc = accomplishment locative verbs; Mixed

Experimental study 109 Tel-Sta = Mixed telic-stative verbs).73 It can be seen that as a general tendency, children's comprehension steadily increased across age. In fact, there is no exception to this trend; that is, there is no U-shaped developmental curve here. Table 5.1. Percent correct in children's comprehension of aspect markers with each verb type across age groups

VERBS

4-Year-Olds -le -zhe zai * 77 38 * 70 51 * 53 68 * 62 56 56 78 89

Aspect Markers 5-Year-Olds zai -Je -zhe * 78 49 * 77 52 * 61 80 * 75 65 66 80 91

Activity Semelfactive Accomp/Res Accomp/Loc Mixed Tel-Sta ** ** 44 Stative 78 51 87 * incomplete sentences, tested but not analyzed in this study ** ungrammatical combinations, not tested

6-Year-Olds zai -le -zhe * 96 51 * 84 67 * 70 82 * 76 78 68 96 89

**

58

87

Table 5.2 presents the results of a loglinear analysis performed on the raw frequency data from which the percentages of Table 5.1 were calculated. This table shows all the possible models and their fit to the data. In a loglinear analysis we select the best-fitting model that is at the same time the most parsimonious in terms of how many effects are involved and explained. Each model's fit to the data is indicated by a p value in Table 5.2. A p value above .05 indicates an adequate fit to the data, while a p value below .05 indicates an insufficient fit.74 All models except (3), (9), and (14) are above this significance level. To determine which of the acceptable models accounts for the data best, we evaluated pairwise all combinations of models that differed in only one effect. Through a forward-selection method (i.e., starting with the simpler models and moving on to the more complex ones),75 we identified (13) as the best-fitting and the most parsimonious model. This model shows that the interaction between verb types (lexical aspect) and aspect markers (grammatical aspect) and the effect of age account for the most important relationships in the data. The interaction effect of model (13) from the loglinear analysis shows that the children understood given aspect markers better with some verb types than with others. In particular, children of all ages comprehended progressive zai better with activity and semelfactive verbs than with

110 Acquisition of Chinese resultative and locative verbs, and they comprehended perfective -le better with resultative and locative verbs than with activity or semelfactive verbs (see Table 5.1). Notice that activity and semelfactive verbs are both atelic (no inherent endpoint), while resultative and locative verbs are both telic (with inherent endpoint). Within the category of atelic verbs there was no difference between activity verbs (nonpunctual) and semelfactive verbs (punctual), and within the category of telic verbs there was no difference between resultative verbs (end result) and locative verbs (endpoint). The property of lexical aspect that seems to be critical for these young learners is, then, telicity rather than punctuality. For the Stative verbs, children understood -zhe much better than -le, whereas for the mixed telic-stative verbs, they understood both -zhe and -le very well (both above 80%). Table 5.2. Loglinear models fitted to the comprehension data in Experiment 1 Model Effect Names* df 36 {M} (1) 36 {A} (2) 33 {V} (3) 34 {M} {A} (4) 31 {V} {A} (5) {V} {M} 31 (6) 29 {V} {M} {A} (7) 30 {MA} (8) 21 {VA} (9) 21 (10) {VM} 25 {MA} {V} (ID 19 (12) {VA} {M} {VM}{A} 19 (13) (14) 15 {VA} {MA} 15 {VM}{MA} (15) 9 (16) {VM} {VA} 5 {VM} {MA} {VA} (17) 0 {VMA} (18) *A = Age, M = Aspect Marker, V = Verb Type

L2

P

43.68 48.81 51.49 34.33 42.14 39.19 29.84 32.93 41.09 12.33 28.44 28.79 2.97 28.01 1.57 1.92 1.12 0.00

.18 .08 .02 .45 .09 .15 .42 .33 .01 .93 .29 .07 1.00 .02 1.00 .99 .95 1.00

The significant main effect of age in model (13) indicates that there was a developmental effect: the comprehension of aspect markers increased steadily with age, not surprisingly, and there was no U-shaped pattern. The differences between different verb types were more pronounced for younger than for older children, but the pattern of interaction between aspect markers and verb types was similar across age

Experimental study 111 groups, with zai associated with atelic verbs and -le with telic verbs. In contrast to zai and -le, the comprehension of -zhe showed no clear development across the age range studied. Even the youngest children, the 4-year-olds, responded correctly to -zhe 78% of the time with stative verbs, and 89% of the time with mixed telic-stative verbs. Experiment 1 shows that grammatical aspect and lexical aspect interact to determine the pattern of children's correct responses, suggesting that the lexical aspect of verbs plays a significant role in children's comprehension of grammatical aspect in Chinese. The associations between zai and atelic verbs, and between -le and telic verbs, are highly consistent with the patterns found in child English (see Chapter 4) and other languages (e.g., Aksu 1978; Antinucci and Miller 1976; Bloom, Lifer, and Hafitz 1980; Stephany 1981; Weist et al. 1984). But our study shows that the interaction between grammatical and lexical aspect exists not only in production as observed in English and other languages, but also in comprehension. This is important, because arguments have been made (e.g., Weist 1989a) that correlational patterns in production only mirror how language is used, given that similar skewing is observed in adult speech. Comprehension studies such as this one clearly show that children's knowledge or control of aspect differs from that of the competent adult speaker. The finding that children understand the progressive marker zai better with atelic verbs (activities and semelfactives) and the perfective marker -le better with telic verbs (accomplishments) is consistent with the predictions of the Basic Child Grammar hypothesis. Recall that Slobin (1985) argued that children are prelinguistically attuned to the distinction between process and result (see 3.2.2). "Process" includes both activity verbs and semelfactive verbs, and "result" encompasses both changes of state (our resultative verbs) and changes of location (our locative verbs). However, we found no significant difference in the comprehension of -le with verbs of the two kinds of changes, so for children changes of state are apparently no more central to the result category than changes of location. In contrast, these data are not consistent with the Language Bioprogram Hypothesis. In particular, children's comprehension of zai and -le did not differ as a function of the punctuality of the verb (zai was not understood better with activity verbs [nonpunctual] than with semelfactive verbs [punctual] nor, conversely, was -le understood better with semelfactive verbs [punctual] than with activity verbs [nonpunctual]). 76

112 Acquisition of Chinese Similarly, there was no evidence for the bioprogram "state-process" distinction, since children understood the perfective marker -le poorly with both activity and stative verbs, and they understood imperfective markers equally well with both activity verbs (which take zai) and stative verbs (which take -zhe). 5.3.2. Experiment 2: Production Experiment 1 was a comprehension study that assessed children's understanding of grammatical aspect when aspect markers were combined with verbs that differ in lexical aspect. Experiment 2 was a production study that investigated how children use aspect markers with verbs of different kinds in an experimental (elicited production) setting. Method

Participants. Children for this experiment came from the same kindergartens as in Experiment 1, and some children participated in both experiments (with at least a 48-hour interval between the two experiments). There was a total of 99 children across four age levels: 3-yearolds (mean age = 3;2, range = 2;9 to 3;6), 4-year-olds (mean age = 4;1, range = 3;8 to 4;4), 5-year-olds (mean age = 5;1, range = 4;11 to 5;4), and 6-year-olds (mean age = 6;1, range = 5;11 to 6;4). Each age group had 25 children, except for the 3-year-old group which had 24. Half the children were boys and half girls. Materials. Children were asked to describe 18 situations enacted with toys. There were three situations for each lexical aspect categories - the same categories as in Experiment 1 except for the addition of achievements and the omission of mixed telic-statives, since it was difficult to elicit the use of a mixed telic-stative verb in both its senses (e.g., both "wearing" and "put on" for chuari). Sample enactment situations include: for activities, a doll canoeing; for semelfactives, a doll knocking once on the back of a turtle (single punctual event) and a rabbit jumping around (iterated event); for achievements, a car knocking down a bridge; for accomplishments (again split into two subcategories), a doll catching a fish with a hook (resultative) and a penguin climbing to the top of a staircase (locative); and for (postural) states, a monkey standing on a table.

Experimental study 113 Procedure. Each child participant was brought individually to a room where the toys were laid out on the floor, and told by the experimenter that they were going to play games with the toys. There were two experimenters. One acted out the situations one at a time and, after each enactment, asked the child to describe what had happened for the benefit of the other experimenter, who was blindfolded (see Hickmann and Liang 1990, for a similar procedure). The child was told to look carefully at the situations that were about to be shown and to make sure that the blindfolded experimenter could understand from his or her descriptions. The instruction given to the child was "Gaosu shushu/ayi, X zenme la?" (literally: tell uncle/auntie [i.e., the blindfolded experimenter] X [i.e., the manipulated toy] how question-marker; English translation: tell the uncle/auntie how X is). This instruction contains no aspect marker, and was chosen because, as Bronckart and Sinclair (1973) and Weist et al. (1984) have noted, children's use of tense-aspect markers in experimental situations can be influenced by biased elicitations (e.g., tenseaspect markers used in the instructions). Note that because there is no tense in Chinese, the imperfective aspect markers zai and -ne can be used to mark events in either the present or the past. In our experiment, then, children could use zai and -ne to refer to enactment situations that were either ongoing at the time they described them (e.g., a doll was canoeing or a rabbit was jumping) or completed (e.g., a doll had been canoeing or a rabbit had been jumping earlier). Children were given a few practice trials to make sure that they understood the procedure. The whole testing session lasted about 20 minutes, and was audio-taped for later transcription and analyses. Data coding and analysis. Children's descriptions of the enacted situations were all transcribed, coded, and entered into the computer by the first author according to the format of the CHILDES database (MacWhinney 1995). The transcriptions were double-checked by James Liang of the Sinological Institute, Leiden University. The CLAN program designed for CHILDES data was used for our computational analyses (including lexical search and frequency counts). Although the enactment situations were designed to elicit verbs belonging to particular categories of lexical aspect, it was impossible to ensure that the children would use the target kinds of verbs, since they were free to focus on any part of the situation. Thus, it would not be meaningful to count how many times a given aspect marker was used in a given category of situation (as has been done in some studies, e.g.,

114 Acquisition of Chinese Bronckart and Sinclair 1973). Instead, we classified the actual verbs used in children's descriptions irrespective of the type of situation they were used to describe. From these descriptions we derived five classes of lexical aspect: (1) activity verbs that encode an action with no endpoint or end result (e.g., hua-chuan 'row-boat'; youyong 'swim'); (2) semelfactive verbs that encode a punctual but not resultative situation (e.g., tiao "jump"; zhayan 'blink'); (3) achievement verbs that encode the end result of a punctual situation (e.g., zhuang-dao 'hit-break'; diao 'drop'); these were mostly RVCs; (4) accomplishment verbs that encode a durative process with a locative endpoint (e.g., pao xiao fangzi-li 'run into the little room'; shang louti 'go upstairs'); and (5) stative verbs that encode the posture of the actor in a situation (e.g., zuo zai yizi-shang 'sit in the chair'; zhan zai zhuozi-shang 'stand on the table'). Eighty-five percent of all the sentences in the children's descriptions contained an aspect marker, either zai, -ne, -the, or -le. Of the remaining 15%, half had stative verbs that did not require an aspect marker because a prepositional phrase occurred postverbally.77 We will report only on those sentences in which the verb was accompanied by either the imperfective marker zai or -ne or the perfective marker -le. The marker -zhe was excluded from the analysis because it occurred rarely in the children's speech.78 A total of 1007 sentences were included in our analyses: 213 sentences from the 3-year-olds, 254 sentences from the 4-year-olds, 254 sentences from the 5-year-olds, and 286 sentences from the 6-yearolds. Quantitative analyses Table 5.3 presents the percentages represented by zai, -ne, and -le of the total number of aspect markers used with each verb type, broken down by age group. As in Experiment l, a loglinear analysis was performed on the frequency data from which these percentages were calculated. Table 5.4 presents all the possible models and their fit to the data. In Experiment 1, many simple models fit the data well, but in this experiment, models simpler than (14) did not fit the data adequately; only models (14), (17), and (18) were adequate. This indicates that the relationships among the variables are much more complex than they were in Experiment 1, and that more complex models are required to account for the relationships.

Experimental study 115 Table 5.3. Percentage represented by -ne, zai, and -le of the total number of aspect markers used with each verb type across age groups

VERBS Activity Semelfactive Achievement Accomp/Loc Stative

-ne 48 55 0 9 47

ASPECT MARKERS 3-Year-Olds 4-Year-Olds T* -le -le -ne mi zai 18 24 20 32 (66) 58 24 24 20 25 (20) 52 0 100 (71) 0 1 99 (22) 0 95 0 91 5 9 44 (34) 5 40 55

5-Year-Olds -ne -le T -ne mi Activity 42 33 25 (72) 61 40 Semelfactive 37 23 (40) 56 Achievement 0 1 99 (72) 0 0 Accomp/Loc 5 95 (21) 0 49 Stative 22 29 (49) 75 *Total token frequencies of aspect markers used with a given

6-Year-Olds -le 31 8 36 8 1 99 8 92 10 15 verb type.

zai

T (78) (29) (86) (19) (42) T (85) (36) (104) (22) (39)

Although there was no obvious best model in terms of both fit and parsimony, we again used the forward-selection method to discover - by systematic pairwise comparisons of all models that differed in only one effect - the significant effects that account for most of the structure in the data. Through this method we found three significant main effects, verb type, aspect marker, and age, and two significant two-way interaction effects, verb type by aspect marker, and aspect marker by age. The interaction between verb type and aspect marker was by far the most important effect: models that included this interaction provided a significantly better fit to the data than models that did not. As in Experiment 1, the strong effect of verb type by aspect marker emphasizes the importance of the interaction between lexical aspect and grammatical aspect in children's acquisition of Chinese. In Experiment 1, we observed an association in children's comprehension between the progressive marker zai and atelic verbs (activities and semelfactives), and between the perfective marker -le and telic verbs (accomplishments, including resultative and locative). In this experiment we found that children used the imperfective markers zai and -ne almost exclusively with atelic verbs (activities and semelfactives), and they used the perfective marker -le predominantly with telic verbs (accomplishments and achievements). This pattern held for all age groups, and became more pronounced with older chil-

116 Acquisition of Chinese dren (except that the 5-year-olds did not produce a higher proportion of -ne with activities and semelfactives than the 3- or 4-year-olds, but they produced a higher proportion of zai with these verbs than did the 3- or 4-year-olds). Table 5.4. Loglinear models fitted to the production data in Experiment 2 Model

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Effect Names*

{M} {A} {V}

df 57 56 55 54 53 52 50 48 45 40 44 42 38 36 32 30 24 0

{M}{A} {V} {M} {V} {A} {V} {M} {A} {MA} {VM} (10) {VA} {MA} {V} (Π) {VM} {A} (12) {VA} {M} (13) (14) {VM} {MA} {VA} {MA} (15) (16) {VM} {VA} {VM} {MA} {VA} (17) {VMA} (18) * A = Age, M = Aspect Marker, V = Verb Type

L2 1010.00 1256.92 1024.46 999.20 766.75 1013.66 755.95 975.78 76.78 1002.55 732.52 65.98 744.83 42.56 721.41 54.87 15.21 0.00

P .00 .00 .00 .00 .00 .00 .00 .00 .00 .00 .00 .01 .00 .21 .00 .00 .91 1.00

To establish more clearly the difference between pairs of verb types of our concern, we conducted several separate chi-square analyses. For each age group, we contrasted activity verbs with semelfactive verbs and activity verbs with stative verbs with regard to the frequency of use of aspect markers, in 2 χ 3 (verb type χ aspect marker) contingency tables. The only significant differences were found between activity and stative verbs in the 4-year-olds (χ2 (2) = 5.97; ρ =.05) and 6-year-olds (χ2 (2) = 6.58, ρ < .05). There is, then, no difference between activity and semelfactive verbs in this experiment with respect to children's use of aspect markers, nor is there a stable difference between activity and stative verbs in general. These results converge with those from Experiment 1.

Experimental study 117 Qualitative analyses In addition to the quantitative analyses above, we also analyzed children's utterances qualitatively. The qualitative analyses showed the same basic patterns as the quantitative analyses, but also revealed new information about children's productive use of aspect markers. First, as shown in Table 5.3, children mostly used -ne as the imperfective (durative) marker with the stative posture verbs. However, children's use of zai with these verbs increased from 5% at age 4 to 22% at age 5. We can see that between ages 4 and 5, the use of zai increased rapidly with activity and semelfactive verbs, and this increase, as compared with the predominant use of -ne as the imperfective marker, may indicate that zai is gaining strength over -ne as marker of imperfective aspect, and the increased use of zai on stative posture verbs may be the result of its overgeneralization across verb types. Examples of such overgeneralization errors are shown in (15) and (16). -ne, (LID; 5;4)79 -NE

(15)

*Xiao houzi zai zhan zai zhuozi-shang little monkey ZAI stand at table-on The little monkey is standing on the table.

(16)

*Yangwawa zai gui zai yizi-shang. (XUQ; 5;4) doll ZAI kneel at chair-on The doll is kneeling on a chair.

In these cases, Chinese-speaking children apply the progressive aspect marker to stative verbs, regardless of the language-specific constraints on the use of zai with posture verbs. Recall that the state-process distinction, as postulated in the Language Bioprogram Hypothesis, should be salient to children in their acquisition of lexical and grammatical aspect. The fact that Chinese-speaking children overgeneralize zai to stative verbs is incompatible with a language bioprogram explanation of the nonovergeneralization of -ing to stative verbs in English (see discussion in 4.1.1). A second finding from our qualitative analyses is that the children had a good command of the associations between grammatical aspect and the locative phrases. As we mentioned earlier (5.1.1), zai can occur as a locative preposition, both preverbally and postverbally. When it occurs preverbally, it indicates the location where an event takes place; when it occurs postverbally, it indicates the locative goal of the event (see

118 Acquisition of Chinese Li 1993b; Zhu 1981), which in effect provides an endpoint to the verb phrase. Interestingly, young children in this study typically associated their use of the imperfective marker -ne with the preverbal zai phrase, and the perfective marker -le with the postverbal zai phrase, as examples (17) and (18) illustrate. (17)

Yazizai shui-li you -ne. (CHH; 4;0) duck ZAI water-in swim -NE The duck is swimming in the water.

(18)

Wugui pa zai chuang dixia qu -le. (LIY; 5;0) turtle crawl ZAI bed bottom go -LE The turtle crawled to the bottom of the bed.

Children clearly distinguished the different functions of the locative phrase at different positions of the sentence, and capitalized on this difference by associating different aspect markers with the phrase at preverbal versus postverbal positions. This association is another example of children's sensitivity to the association between grammatical aspect and lexical aspect in Chinese. Finally, in the adult language, the use of the object marker ba is associated with the perfective marker -le and with highly transitive, resultative events (usually in the form of RVCs; see Chao 1968; Li and Thompson 1981; Tenny 1994). Children in this study conformed to this association, in that 90% of the 296 sentences that contained the ba construction had an RVC as the main verb, such as in (19). (19)

Qiche ba qiao gei zhuang-dao -le. (RAO; 3;3) car BA bridge give bump-collapse -LE The car knocked down the bridge.

The remaining 10% of the 296 sentences that contained ba had primarily monosyllabic achievement verbs, such as sa 'spill', diao 'drop', and reng 'throw', which are also highly transitive and resultative. Furthermore, almost all of these sentences were marked with the perfective -le; only 3 sentences involved the imperfective markers zai and -ne. The almost perfect association of the ba construction, resultative verbs, and the perfective aspect in children's productive speech suggests that

Experimental study 119 children respect the lexical and grammatical co-occurrence constraints from early on.80 To summarize, results from Experiment 2 provide converging evidence for the patterns observed in Experiment 1. Children's productive speech is characterized by a strong association between lexical and grammatical aspect at an early stage. In particular, the results provide further support for the Basic Child Grammar hypothesis that the distinction between process and result (i.e., the distinction between atelic and telic) is important in children's early acquisition of tense-aspect markers. From at least age 3 on our children almost always combined achievement verbs with -le, and not zai or -ne, which indicates that they must have integrated the meaning of result into their knowledge of these verbs. In contrast, the findings from Experiment 2 are not consistent with predictions of the Language Bioprogram Hypothesis: (a) in no age group did the children use aspect markers differently with activity (nonpunctual) and semelfactive (punctual) verbs; (b) in all age groups the children used aspect markers differently with achievement verbs (punctual and resultative) and semelfactive verbs (punctual but non-resultative) - in other words, punctual verbs did not hang together for our children as a category; and (c) the pattern with respect to activity versus stative verbs went, if anything, counter to the predictions of the Language Bioprogram Hypothesis, because the youngest children (3-year-olds) did not mark verbs belonging to these two categories differently while the oldest (6year-olds) did. 5.3.3. Experiment 3: Imitation The previous two experiments were designed to examine children's comprehension and production of grammatical and lexical aspect in Chinese. In Experiment 3, we used an elicited imitation task (Slobin and Welsh 1973) to test more specifically children's sensitivity to the semantic distinctions of the Basic Child Grammar and the Language Bioprogram Hypothesis. We adopted Slobin and Welsh's assumption that in imitating a sentence, children try to retrieve the meaning of the sentence and filter it through their own productive system. This means that elements in the model that are ungrammatical or unusual within the child's own system will tend to be modified (see Kuczaj and Maratsos 1975; Budwig 1991 for systematic applications of the elicited imitation technique). Applying this reasoning to the problem at hand, we can predict

120 Acquisition of Chinese that if Chinese-speaking children understand the use of aspect markers, they should have more difficulty imitating ungrammatical than grammatical combinations of aspect markers with verbs. In particular, when faced with an ungrammatical combination of verb and aspect marker they would tend to omit the marker or change either the marker or the verb so that the combination becomes grammatical. In this experiment, we used elicited imitation to see whether children are sensitive to two combinations of aspect marker and verb that are ungrammatical in Chinese: 1) progressive zai with achievement verbs (testing the process-result distinction), and 2) progressive zai with stative verbs (testing the state-process distinction).

Method Participants. A total of 72 children participated in this experiment immediately after completing the production experiment: 22 3-year-olds (mean age = 3;2), 25 4-year-olds (mean age = 4;1), and 25 5-year-olds (mean age = 5;1). Materials. Eight model sentences were constructed. All the sentences were about the same length (9 to 10 syllables, 6 to 8 words). The sentences were intended to exceed the children's short-term memory capacity so that children could not simply repeat the sentences by rote. Because no research has assessed 3- to 5-year-old Chinese children's short-term memory span for syllables and words, we settled on 9 to 10 syllables on the basis of a pilot test. A pre-test showed that our sentences were too easy for 6-year-olds, so we did not test children of this age further. In the production task of Experiment 2, children almost never used the progressive marker zai with achievement verbs (RVCs) - rather, they used it almost exclusively with activity and semelfactive verbs, which are non-resultative atelic. We interpreted this as showing that children are sensitive to the meaning of result inherent in achievement verbs. On the other hand, children did produce some ungrammatical combinations of zai with stative posture verbs, although it was in general rare. In the present experiment we created an artificial environment in which both achievement verbs (RVCs)81 and stative verbs were combined ungrammatically with zai in the model sentences for imitation. Achievement verbs and stative verbs were also combined grammatically with the perfective -le to provide a baseline for comparison. Two instances of each verb type were combined with zai and another two with -le for a

Experimental study 121 total of eight test sentences, all of which were administered to each child. The four RVCs were qie-kai 'cut-open', da-hao 'build-ready', zuo-hao 'make-ready', and shuai-po 'throw-break'. The four stative verbs were xihuan 'like', mingbai 'understand', you 'have', and zhidao 'know'. Two examples are shown in (20) and (21) (see Li and Bowerman 1998 for the complete set of sentences). (RVC + -le; grammatical): (20) Xiaopengyou da-hao -le neixie jimu. child build-ready -LE those blocks The child built (i.e., stacked up) the blocks. (zai + RVC; ungrammatical): (21) *Mama zai zuo-hao yi-guo mifan. mother ZAI make-ready one-pot rice Mother is making a pot of rice ready. Procedure. After completing Experiment 2 each child was asked to continue with another game called "teaching the puppet to speak". The experimenter held out two puppets and asked the child which one he liked better. The child then took the preferred puppet and the experimenter kept the other. Next, the experimenter explained that the puppets could not speak and that she and the child would play a game to teach them how to speak. The child should follow the experimenter and teach his own puppet in exactly the same way as the experimenter would do. Then the experimenter read the model sentences one by one for her puppet, and the child imitated each one for his puppet. The order of presentation of the model sentences was pseudo-randomized for each child, with the condition that no more than two ungrammatical sentences occurred consecutively. Before testing began, the child practiced with three warm-up sentences. The whole session was audio-taped for later transcription and analysis. Data coding and analysis. The data were transcribed and coded in the same way as in Experiment 2. An imitation was counted as successful if the child retained the main verb and the aspect marker of the model sentence, irrespective of other changes; otherwise it was counted as erroneous. The errors were either omissions or substitutions of the aspect marker or the verb, and they occurred most typically when the imitated combination was ungrammatical.

122 Acquisition of Chinese Results and discussion Table 5.5 presents the successful imitations as percentages of the total number of imitations in each age group for each combination of aspect marker with verb type (there were 22 children in the 3-year-old group and so a total of 44 imitations for each combination; there were 25 children in both the 4- and 5-year-old groups and so 50 imitations for each combination). Table 5.5 shows that children's successful imitations increased steadily with age. Because the data shown in Table 5.5 have a much simpler structure than the data of the previous experiments, we used chi-square analyses rather than loglinear analyses. A 2 χ 2 chi-square analysis was conducted for each age group, treating aspect marker (zai vs. -le) and verb type (achievements vs. statives) as the classifying variables. For the 3- and 4year-olds, success in imitating sentences with zai versus -le differed significantly as a function of whether the verb encoded a result (achievement verbs) or a state (stative verbs) (age 3: χ^ (1) = 3.95, ρ < .05; age 4: χ2 (1) = 3.71, ρ - .05). For the 5-year-olds, however, the difference did not reach significance (χ^ (1) = 1.90, n.s.). This may be due to the increased short-term memory span of the 5-year-olds: of the ungrammatical combinations with zai, they could successfully imitate over 60% of those with achievement verbs and over 80% of those with stative verbs. Table 5.5. Percent successful imitations of the model sentences (verb type by aspect marker by age group) Aspect Markers in Model Sentences 4-Year-Olds 5-Year-Olds 3-Year-Olds -le -le -le zai* zai* zai* 64 40 72 82 Achievement 23 57 88 72 78 70 Stative 46 43 * ungrammatical combinations with achievement and stative verbs VERBS

The ungrammatical combination of zai with achievement verbs (RVCs) presented a particular imitation difficulty, especially for the younger children. Note that RVCs consist of two components, the first indicating the action and the second the result (e.g., qie-kai 'cut-open'). In adult language, the meaning of result is dominant in these constructions: it eclipses the meaning of action to such an extent that the con-

Experimental study 123 struction cannot be marked with zai (Li 1987; Klein, Li, and Hendriks, in press; Tai 1984). Children's difficulty in imitating the ungrammatical combination of zai with these verbs shows that they are sensitive to the salience of the meaning of result and its clash with the progressive meaning of zai. This finding is consistent with the findings in Experiment 1 that children understood -le better than zai with resultative accomplishment verbs, and in Experiment 2 that they used -le and almost no zai with achievement and locative accomplishment verbs. Taken together, these findings indicate that Chinese children are indeed sensitive to the distinction between process and result, as predicted by the Basic Child Grammar hypothesis. If the process-state distinction were just as salient to learners as the process-result distinction, children should be just as resistant to imitating combinations of zai with Stative verbs as combinations of zai with achievement verbs; recall that both are ungrammatical. But these combinations did not present a particular difficulty to children - even the 3year-olds imitated the ungrammatical stative combinations just as well as the grammatical ones, and the 4- and 5-year-olds imitated the ungrammatical ones even better. Children are not, then, sensitive to grammaticality per se, but to particular combinations of grammatical marker and lexical aspect of verbs. This result constitutes a further challenge to Bickerton's proposal that the process-state distinction is bioprogrammed and so should be salient to children, and to his explanation for why English-speaking children do not overgeneralize progressive -ing to stative verbs. 5.3.4. General discussion of the experimental research The experimental research reported here has examined children's comprehension and use of grammatical and lexical aspect in the acquisition of Mandarin Chinese. In Experiment 1, children showed that they understood progressive zai better with activity and semelfactive verbs than with accomplishment verbs, and perfective -le better with accomplishment verbs than with activity and semelfactive verbs. There was no difference in their comprehension of either zai or -le with activity versus semelfactive or stative verbs. In Experiment 2, children produced imperfective aspect markers (zai and -ne) mostly with activity and semelfactive verbs and rarely with accomplishment or achievement verbs, and they produced the perfective marker -le more frequently with accomplish-

124 Acquisition of Chinese ment and achievement verbs than with activity and semelfactive verbs. There was no significant difference in their use of either perfective or imperfective markers with activity as opposed to with semelfactive or stative verbs. In Experiment 3, children imitated the combination of achievement verbs with the perfective marker -le (grammatical) better with than with the progressive marker zai (ungrammatical), but their imitations of stative verbs with -le (grammatical) and zai (ungrammatical) did not differ. In sum, there was a consistent association of imperfective markers with atelic verbs (activity) and an association of the perfective marker with telic verbs (accomplishment and achievement). Stative and semelfactive verbs patterned in general like activity verbs. These results have specific implications for the Language Bioprogram Hypothesis. According to this hypothesis (Bickerton 1981, 1984), children are innately sensitive to the distinctions between process and state and between punctual and nonpunctual meanings. Our results are not consistent with this claim. In general, children do not distinguish between activity and stative verbs in their use of aspect markers. They do treat achievement verbs differently from activity and stative verbs, which might at first glance be taken as evidence for the punctual-nonpunctual distinction (Bickerton indeed used such evidence to support the distinction). But achievement verbs are not only punctual but also resultative, for example, as reflected in fall, break, spill in English and in all the achievement verbs in Chinese. Interestingly, Bickerton viewed them as "punctual" (see discussion in 8.4.2), while Slobin interpreted them as "resultative". Thus, to disentangle punctuality from resultativity in Chinese and to identify which meaning component is responsible for the difference between activity verbs and achievement verbs in our data, we must look at verbs that are punctual but not resultative, such as semelfactive verbs (see 5.1.2). In our studies children do not distinguish semelfactive verbs from activity verbs, but do distinguish them from achievement verbs. This is strong evidence that the difference between achievement and activity verbs in our experiments comes not from the punctual but from the resultative meaning. Our results also have implications for the Basic Child Grammar hypothesis. According to this hypothesis, children orient from the beginning to a major temporal contrast between process and result (Slobin 1985). This contrast is claimed to belong to an initial set of basic semantic notions that are independent of experience with any particular language, and that serve as initial magnets for the grammatical mor-

Experimental study 125 phemes of the input language. Our findings are consistent with the claim that children are highly sensitive to this contrast: in all three experiments children treated activity verbs differently from verbs that encode telicity or resultativity. But must we concur with Slobin (1985) that the cleavage between process and result is critically salient for the child before language learning even begins? At the end of Chapter 3, we noted that in a recent rethinking Slobin (1997, in press) has revised his earlier proposal that there is a universal set of semantic notions that are "privileged" for mapping to grammatical morphemes. He now suggests that the grammatical morphemes of a language and the meanings they express are best seen as the outcome of continuous diachronic processes of grammaticization, and the learning of these morphemes is similar to the learning of content words: both belong to the general domain of concept formation. Although our data confirm the saliency of the process-result distinction, we suggest that children could develop such sensitivity over time - for example, the developmental patterns in our data show that the associations between lexical and grammatical aspect in children's production become stronger, rather than weaker over time; this constitutes evidence that the relevant distinctions for children are not prelinguistic (see also Erbaugh 1992 for a similar argument). We think that it is possible to account for children's early sensitivity by appealing not to prelinguistic constructs, but to the learners' analysis of the distributional properties in the speech they hear, and to their ability to extract patterns of association between lexical and grammatical aspect. Such analytical abilities could be conceptualized as, or operationalized through, connectionist principles, to which we will turn in Chapter 7 in detail. 5.4. Crosslinguistic characteristics in the acquisition of aspect in Chinese In Chapter 4 we examined the acquisition of aspect in English, on the basis of which we laid out a general model of the acquisition of aspect, in particular, the prototype account. As noted earlier, the prototype account assumes that children's associations between tense-aspect markers and categories of lexical aspect are probabilistic rather than absolute (3.2.3; 4.1.3); for example, in our production experiment (5.3.2), the perfective aspect marker -le was used most often with telic verbs, but it was also occasionally used for activity and semelfactive verbs. Probabilistic

126 Acquisition of Chinese associations are problematic for theories that attribute children's biases to the powerful force of prelinguistic or innate semantic distinctions, but are compatible with our hypothesis that children's biases reflect patterns of association in the linguistic input. The data we presented in this chapter are generally consistent with patterns observed in English, in that we found clear associations of telic verbs (i.e., accomplishments and achievements) with perfective aspect (-le), and atelic verbs (i.e., activities and semelfactives) with imperfective aspect (zai, -zhe, and -ne). But the Chinese data differ from the English data in several important ways. First, our experimental study indicates that these associations exist not only in production, but also in comprehension. This provides evidence that the associations reflect the child's overall linguistic capacity, and are not due to contextual constraints. Second, our production data reveal a relatively more absolute picture of the associations than do the English data. In particular, Chinese children rarely, if ever, use imperfective aspect markers with telic verbs; telic verbs occur almost exclusively with perfective aspect across all age groups of our study. Finally, our results do not seem to be consistent with the idea that the associations are initially based on prototypical uses, and are later expanded to non-prototypical cases. For example, in our production data, the associations remain strong across all age groups: if anything, the association between atelic verbs and imperfective markers becomes stronger, rather than weaker, while the association between atelic verbs and perfective aspect becomes weaker, rather than stronger over time. How do we explain these discrepancies between the acquisition of aspect in Chinese and that in English? Our explanation invokes crosslinguistic differences between the two languages and the language-specific properties of Chinese. Language-specific properties have important implications for our understanding of the crosslinguistic variations that determine patterns of language acquisition. In 5.1.2, we discussed the lexical aspect of Chinese verbs. One particular characteristic of Chinese verbs is the use of RVCs, such as ti-dao 'kick-down', qie-kai 'cut-open', or reng-diao 'throw away', which are achievement verbs that encode a clear end result. These verbs do not accept the imperfective marking with zai, -zhe, or -ne at all, unlike resultative verbs in English that can be easily marked with the progressive -ing. In other words, achievement verbs in Chinese can only be combined with the perfective marker -le. The more abso!" associations that we see in the acquisition of Chinese aspect might be due to this constraint: they show that Chinese-speaking

Crosslinguistic characteristics 127 children respect the incompatibility between imperfective markers and achievement verbs from early on and, given that this constraint is preserved in the adult language, there is hardly any reason for children to relax such a constraint. The imitation experiment also shows that from age 3 on, children are sensitive to the incompatibility between zai and achievement verbs. Without doubt, language-specific properties also interact with language-general principles. In Chinese, the imperfective marker zai cannot be combined with stative posture verbs, such as zhan 'stand', zuo 'sit', or tang 'lie', whereas posture verbs in English accept the -ing marking. Our data show that despite the language-specific constraint on the use of zai, imperfective markers predominate over the perfective marker with posture verbs, in both comprehension and production (and in elicited imitation children have no difficulty imitating zai with stative verbs). The fact that the association between perfective aspect and stative verbs does not gain strength over time, as the prototype account would predict, is probably due to a general constraint: it is natural to describe the state of affairs of events in their current status, as we view them imperfectively. This is in direct contrast with the constraint between resultativity and imperfectivity - it is difficult to comment on the ongoingness of a resultative action, especially when the action occurs instantaneously, such as spilling or falling. Thus, the persistence of the weak association between perfective aspect and stative verbs in our Chinese data, despite the restriction on the use of zai with stative verbs, reflects a general rather than language-specific constraint between stativity and perfectivity. Some of these general compatibility constraints or tendencies point to Comrie's (1976) "naturalness of combination" principle: some aspect morphemes combine naturally with some verb types but not others (see discussion in 5.1.3; also Bybee's 1985 discussion of Local Markedness, and Andersen's 1993 Distributional Bias Hypothesis). For example, perfective aspect does not go naturally with stative verbs because it presents a situation with a boundary whereas stative verbs encode indefinite situations; in contrast, imperfective aspect combines naturally with stative verbs.82 Many situations with an end result last for such a short period of time that any comment on them will have occurred after its ending (e.g., situations denoted by verbs like drop, fall, or crash). Such combinatorial constraints are clearly reflected in our experiments. For example, because children describe situations as the situations take place in the experiment, the posture states are more likely to be described in

128 Acquisition of Chinese their current status by imperfective aspect markers. In contrast, the telic and resultative situations last for only a brief time, and children can comment on the situation only after it has ended. This might explain why accomplishment verbs also occur almost exclusively with the perfective marker -le in children's productive speech, even though these verbs can occur in the adult language with imperfective markers. Finally, although our production data are not too informative about children's development over time, the comprehension data from Experiment 1 indicate clearly that children's comprehension of the function of aspect markers increases over time: by age 6, children's comprehension of both zai and -le has reached 70% or more with the resultative and locative accomplishment verbs, and although their comprehension of -le is still depressed with activity and stative verbs, comprehension of this marker has also been on the rise. Thus, these data suggest that even though children may well adhere to the constraints of lexical and grammatical aspect in productive speech, they could start to relax these constraints in comprehension, with an increasingly solid command of the grammatical meaning of aspect markers and the separation of it from inherent lexical aspect of verbs.

Chapter 6 Acquisition of aspect in Japanese In this chapter, we will review research on the acquisition of aspect (along with tense) in Japanese as a first and second language. There has recently been a growing interest in this area, both in LI and L2 research, to the extent that research on Japanese acquisition can contribute to the theory of tense-aspect acquisition, and of morphological development in general. We will review research in this area and discuss it in relation to the general theoretical framework that we have presented so far. But first, we need to briefly discuss how the Japanese tense-aspect system is organized. 6.1. Tense-aspect system in Japanese The Japanese tense-aspect system is similar to that of English. All finite indicative predicates are marked for tense (past -ta vs. non-past -ru) as in English.83 In both languages, the past tense marker can be attached to any verb without any systematic restriction. A major difference between the two languages is that the Japanese past tense marker is often considered to have the sense of 'perfect' or 'perfective' because it is still in the process of grammaticizing from a perfect marker into a perfective aspect marker and then into a simple past tense marker (Horie 1997). Due to this transitional nature of the Japanese past tense marker, there have been disagreements on whether the past tense marker -ta is an aspect marker or a tense marker (e.g., Takahashi 1976; Ando 1986: Ch. 8). The nonpast form -(r)u normally refers to present state with stative verbs, and to future action or habitual action with dynamic verbs. With regard to aspect, Japanese has an obligatory durative imperfective aspect marker -te i- which must be used in referring to action in progress at the reference time.84 However, the semantic scope of this marker is different from that of the English progressive aspect. Although it shows an interaction with lexical aspect similar to that of the English progressive, it exhibits an important difference concerning achievements. Some examples of the use of -te i- are given below in sentences with the four lexical aspect types:

130 Acquisition of Japanese Activity: Action in progress (1) Ken-ga utat-te i-ru. Ken-Norn sing-Asp-Nonpast 'Ken is singing.'

85

Accomplishment: Action in progress (2) Ken-wa isu-o tukut-te i-ru Ken-Top chair-Ace make-Asp-Nonpast 'Ken is making a chair.' Achievement: (a) Resultative state (3) Booru-ga oti-te i-ru. ball-Norn fall-Asp-Nonpast The ball has fallen (and it is there).'

(4)

(b) Iterative action-in-progress Ken-wa doa-o tatai-te i-ru. Ken-Top door-Ace bang-Asp-Nonpast 'Ken is banging on the door.'

State: Vividness; temporariness (5) Huzisan-ga mie-te i-ru Mt. Fuji-Nom be visible-Asp-Nonpast 'We can see Mt. Fuji (at this moment).' cf. (6)

*Okane-ga it-te i-ru Money-Nom be necessary-Asp-Nonpast (intended meaning 'Money is being needed.')

An important difference between English and Japanese is that the Japanese imperfective -te i- combined with achievements can refer to a resultative state, but not to 'a preliminary stage of an event' as does the English progressive combined with achievements. In imposing an internal view, English can focus on the process leading up to the punctual point of achievement (e.g., The horse is winning the race), whereas Japanese cannot; instead, Japanese focuses on the duration of the resultant state that obtains as a result of the punctual action.86 Thus, the morphological equivalent of Ken is dying in Japanese (Ken-wa sin-de

Japanese tense-aspect system 131 i-ruf1 means 'Ken is dead.' This contrast can be schematically represented as follows: English

Japanese time axis

(die) Ken is dying.

Ken-wa sin-de i-ru. 'Ken is dead.'

Figure 6.1. Imperfective aspect combined with achievements in English versus Japanese

Some achievement verbs that are anomalous with progressive marking in English (e.g., notice, find) are compatible with -te i- in Japanese because their combinations refer to resultative states. For example, the direct translation of *Ken is noticing the picture (Ken-wa sono e-ni kizuite i-ru) is OK in Japanese, and means that he has already noticed (the existence of) the picture. 6.2. Acquisition of tense-aspect morphology in Japanese Since the early 1980's the acquisition of Japanese tense-aspect morphology has been investigated in relation to important theoretical issues discussed in earlier chapters. This is primarily because of a unique feature of the Japanese imperfective marker -te i-. As noted in Chapter 2, the perfective-imperfective distinction is the most basic aspectual distinction crosslinguistically. However, Japanese -te i- combines these two distinct notions. That is, when it is attached to durative verbs (accomplishment, activity, and state88), it denotes progressive meaning, which is a type of imperfective, but when it is attached to achievement verbs, it denotes resultative meaning, which is closely associated with perfective aspect (Shirai 1998b).89 This makes it possible for language researchers to address issues such as the role of input vs. universals, and Bickerton's Language Bioprogram Hypothesis (see 6.2.2.).

132 Acquisition of Japanese In what follows, we review LI and L2 acquisition literature with respect to the close relationships between lexical aspect and the development of tense-aspect morphology that have been observed in various languages discussed so far. As we shall see, Japanese data show some universal tendencies as have been observed in other languages, but at the same time also show language-specific properties due to the typological characteristics of Japanese. We will first review how the Japanese data fit the prototype model proposed in Chapter 4 at a descriptive level, and then discuss the theoretical implications of the data. At the descriptive level, the prototype hypothesis predicts that past tense marking is initially associated with achievement and accomplishment verbs. This prediction can easily be tested in Japanese because Japanese has a past tense marker, -ta. On the other hand, the hypothesis predicts that progressive marking is initially associated with activity verbs, and only later attached to other classes, and that it is dissociated from state verbs. This prediction cannot be easily tested in Japanese because the Japanese imperfective marker -te i- can also denote resultative state. Therefore, there are three possibilities for the development of -te i-: (1) it is associated with activity verbs to denote action in progress, (2) it is associated with achievement verbs to denote resultative state, and (3) no particular association is observed. Here, we review empirical studies to determine how LI and L2 learners acquire Japanese past tense and imperfective aspect markers. 6.2.1. First language Rispoli and Bloom (1985) were probably the first to discuss the issue of lexical aspect in relation to tense-aspect acquisition in Japanese. Citing data from Rispoli (1981), they argued that Japanese children have no problem acquiring -te i- regardless of whether it is used for progressive or resultative. Rispoli and Bloom (1985: 473) stated: "In Japanese, the feature continuative... can be coupled with the feature punctual... This arrangement causes Japanese children absolutely no problem. From as early as 22 months the same inflection is used to signify both continuing incomplete aspects and continuing complete aspects of events." This is an important observation, which Shirai (1993, 1998b) further attempted to test empirically (to be discussed below). Several other studies investigated the acquisition of Japanese tenseaspect morphology in relation to lexical aspect. Cziko and Koda (1987)

Acquisition of tense-aspect 133 analyzed the data from a longitudinal corpus of a Japanese boy (Sumihare) from age 1;0 to 4; 11, and claimed to have found that the child never overextended the -te i- form to stative verbs, which is in line with the prototype hypothesis. However, they also found that punctuality (both at the level of lexical aspect and the level of situational properties) did not necessarily correlate with the past inflection -ta (see related discussion in 3.2.3.). This clearly contradicts the claim that the early past inflections are used with punctual/telic verbs. Another study, Rispoli (1981), provided detailed information about the longitudinal development of a Japanese child, and our reanalysis of his data seems to support the association of past marking with punctuality/telicity, contrary to Cziko and Koda's finding. In Rispoli's earliest samples (1;6 and 1;8), verbs used with the past marker -ta are almost exclusively accomplishment and achievement verbs, with the one exception of hiita 'played (an instrument)', which should be classified as an activity. In later samples (1;10 to 2;0), the past marking extends to more activity and state verbs. This supports the prototype hypothesis regarding the past tense form. On the other hand, Rispoli's (1981) data do not suggest any special effects of lexical aspect with respect to the development of -te /'-, as noted by Rispoli and Bloom (1985). The data show that resultative meaning and progressive meaning emerge at the same time, at age 2;0. It is interesting to note, however, that there is one use of -te(i)ta (i.e., a combination of the durative marker with the past form -ta, signifying past durative) at age 1;11, even before the emergence of -te i-ru (i.e., present durative) and it was used with an activity verb (aruku 'walk') to denote action in progress. There was a one-month period when there was no recording between the two samples (i.e., between 1;11 and 2;0). Therefore, it is possible that the child Rispoli studied might have restricted the use of -te i- to activity verbs during that brief period. Horiguchi (1981) also reported on the longitudinal development of verb morphology of a boy, who used -te i- to denote progressive earlier (at 1;10) than resultative (at 2;0). The boy also used the form to denote progressive more frequently at the earlier period (1;10 to 2;5, 68.9% and 2;6 to 2; 11, 72.9%) than at the later period when this preponderance was lost (3;0 to 3;5, 52.2%).90 Horiguchi also reported an incorrect use of -te i- with the state verb aru 'exist' (i.e., *atteiru). She further reported on a girl studied by Okubo (1967), whose -te i- form appeared at 1;8, during which month it referred to both progressive and resultative situations.

134 Acquisition of Japanese Horiguchi also analyzed all 60 tokens of -te i- reported in Okubo (1;8 to 3;5) and found that 61.7% were used for progressive meaning. Shirai (1993, 1998b) specifically attempted to test the relationship between verb morphology and lexical aspect. Shirai (1993) analyzed the same published longitudinal data (Noji 1976, 1977) from the Japanese boy Sumihare that Cziko and Koda analyzed. His finding was that the boy exhibited a period in which the use of -te i- was more frequent with activity verbs to denote action in progress than with achievement verbs to denote resultative state. No state verbs were used with -te i-. Shirai (1993) also reported a strong association between the past form -ία and achievement verbs. This finding contradicts Cziko and Koda (1987), who found no correlation between past tense marking and punctuality. Since both studies used the same published corpus, the discrepancy must be due to differences in methodology. Shirai (1993) analyzed all the uses of -ta from its emergence at 0;11 to 1;8 (139 tokens) whereas Cziko and Koda (1987) analyzed randomly sampled pages of the corpus from 1;0 to 4; 11, yielding 114 tokens of -ta, among which 50% were used with punctual verbs. They also analyzed early uses of -ta (1;0 to 2;0, 55 tokens) from the same sampled pages, but still only 51% were with punctual verbs. Shirai's analysis showed that after a period of sporadic uses of -ta (only 10 tokens between 0;11 and 1;5), the correlation became high (73% at 1;6), and gradually declined (66% at 1;7 and 59% at 1;8). Since Cziko and Koda's 'early use' analysis included 1;9, 1;10, 1;11, and 2;0, during which time correlation is expected to become weak, it appears their analysis was not able to identify the earliest correlation. This is probably why they did not find a correlation between punctuality and the past tense form -ta. Shirai (1998b) analyzed the longitudinal data of three children: Aki (Miyata 1995), Yocchan (Clancy 1985), and Taachan (Kokuritsu kokugo kenkyuujo 1982a, 1982b). Regarding the use of past tense -ία, a strong correlation with achievement verbs was found for all three children. Except for Aki's early uses (1;8, 2;1 and 2;2; between 33% and 40%), all other samples showed a high correlation of past -ία with achievement verbs (between 57% and 76%). However, the earliest uses of -ία by the three children were not necessarily limited to achievements, as in English acquisition, but included many atelic verbs, and state verbs in particular, and this needs to be explained. We will address this issue in 6.2.2.

Acquisition of tense-aspect 135 On the other hand, individual differences were found for the emergence of -te i-. Yocchan's data did not include the emergence of -te i-, and therefore our discussion focuses on the other two children. First, Aki's early use of -te i- was predominantly with activity verbs to denote action in progress. This is in line with the finding in Shirai (1993). However, Taachan used -te i- with achievement verbs with resultative state meaning very frequently (84%) when his use of -te i- began to emerge at age 2;2. None of the children used -te i- with stative verbs. On the basis of the studies reviewed above, we can propose the following generalizations: (1) (2)

(3)

The past tense form -ta is strongly associated with achievement verbs. The durative/imperfective marker -te i- is not used with state verbs although it is not totally impossible in adult speech. Although -te i- tends to be used with activity verbs at the emergent stage, individual differences exist.

An elaboration on (3) may be necessary here. Among the studies reviewed above, child H (Horiguchi 1981), Sumihare (Shirai 1993), and Aki (Shirai 1998b) showed early preference for -te i- with activity verbs to denote action in progress, child Υ (Okubo 1967) and child T (Rispoli 1981) showed no preference, and Taachan (Shirai 1998b) preferred resultative meaning. Thus it appears that there are individual variations.

6.2.2. Theoretical issues and crosslinguistic comparison in LI acquisition In what follows, we discuss the above descriptive generalizations in relation to the issues raised in Chapters 3 and 4. Three of the four generalizations of the prototype hypothesis that are relevant to Japanese will be discussed: association of past marking with telicity, association of progressive marking with activity, and lack of incorrect use of stative progressives. In so doing we compare the Japanese data with the English data discussed in Chapter 4.91

136 Acquisition of Japanese Past tense -ta One major difference between the Japanese data and the English data discussed in Chapter 4 is that whereas the children acquiring English followed the prediction of the prototype hypothesis exactly, the Japanese children showed a very early use of the past tense form with stative verbs, which should occur very late according to the predictions of the hypothesis. How can this difference be accounted for? There are two possible interpretations, both of which are based on the typological facts of Japanese. First, the early -ta forms appear to be the product of lexical learning. As Clancy (1985) suggested, these early forms are probably unanalyzed forms learned by rote (MacWhinney 1978). This is especially likely in the case of Taachan, who used atta (state verb aru 'exist' + past tense -to) very early at 1 ;0, and did not produce the non-past form -ru until 1;11. This type of early use of verbal inflection in Japanese has been attributed to the fact that "there is no 'base' form of Japanese verbs, so that every verb which the child hears is necessarily inflected for some tense-aspect" (Clancy 1985: 427). We suggest that this typological fact also enhances the tendency of early reliance on lexical rote learning of inflected forms, in contrast to languages that allow the 'base forms' to appear frequently, which facilitates the use of these forms early on before children start to use verbal morphology. The English children discussed in Chapter 4 (Shirai 1991) may thus have closely followed the prototype hypothesis because of their relatively low reliance on rote learning of inflected forms. Japanese children could not go through this bare-verb stage even if they wanted to (see also Weist 1989b: 66). Another relevant typological fact about Japanese is the status of the past form -ta. Although it is generally called the past tense form, it is not a full-fledged past tense marker. Historically, it comes from the perfect marker -tan, which has been phonologically reduced to -ta over time. It is still on its way to becoming fully grammaticized as the deictic past tense marker, as noted above (Takahashi 1976; Kinsui 1997; Horie 1997). Therefore -ta can still be used to refer to a situation at speech time, a characteristic of a perfect marker. The stative verbs with -ta attached (i.e., atta and ita), which appeared frequently in early speech, may be examples of such use. These forms are used when the speaker has found something that he or she has been looking for, through direct perception, and in many cases they can be uttered monologically; that is,

Acquisition offense-aspect

137

the utterances do not require the presence of an addressee. Indeed, most of the stative verbs these children used with -ta were of this type. Other than the fact that stative verbs were often used with -fa, the above review reveals that there was a predominance of achievement verbs with the past tense marker, as predicted by the prototype hypothesis. Also notable is the fact that most of the activity verbs used with -ta early on were those that share similarity with achievement verbs, i.e., verbs that are on the borderline between activity and achievement. For example, neru 'sleep' is coded as activity since it primarily refers to the activity of sleeping, but it can also refer to the punctual entry into the activity of sleeping (i.e., going to sleep), depending on the context. Many of the activity verbs used with -ta by children early on were of this kind, and therefore the characterization of the prototypical past tense marking (i.e., to denote punctual action that results in change of state, e.g., Slobin 1985) may still hold true. In this connection, the past tense forms of stative verbs that are often used with -ta, atta and ita can also be considered to indicate change of state, in the sense that there is a change of state in the speaker's mind from not being able to find something to perceiving it. Durative aspect -te iIn English acquisition, progressive marking is first used mostly on activity verbs, and is then extended to accomplishment and achievement verbs. Does the acquisition of Japanese exhibit the same trend? This question cannot be answered directly, since the Japanese durative aspect marker -te i- can also denote resultative states when attached to achievement verbs; e.g., -te i-ru- can mean either "X is V-m#" or "X has V-ed". In the case of English acquisition, there is a brief period when progressive marking is virtually restricted to activity verbs and to iterative use of punctual verbs, to denote action in progress. Do Japanese children use the durative aspect marker -te i- only for progressive meaning at the emergent stage? The data from Japanese are not conclusive. As discussed in 6.2.1, some children showed a preference for activity verbs (and accomplishment verbs) in early use of -te i-, to denote action in progress, but others did not show any preference for either progressive or resultative meaning. Taachan even showed a preference for achievement verbs right from the beginning, and even used the past progressive/resultative (-te i-ta) in

138 Acquisition of Japanese the same month when -te i- was first used. That is, his first use of the -te i- morphology was already productive. How do we make sense of these observations? First of all, based on the data from Taachan, who showed a preference for achievement verbs with -te i-, we can reject the strong version of the universalist explanation, which would predict that all children prefer the use of grammatical functors to mark notions that have special status. In the case of Bickerton's Language Bioprogram Hypothesis, these are the notions "punctual" and "nonpunctual", which would predict, as noted above, that children associate -te i- with the lack of telicity and therefore use it for action in progress with activity verbs, since telicity is already marked by the past tense marker -ta, which develops earllier than -te i- (Shirai 1993, 1998b). Taachan is a genuine counterexample to a strong version of this hypothesis that does not allow for individual variation. But does this force us to reject the universalist explanation altogether? We should also keep in mind that at least three children (Sumihare, child H studied by Horiguchi 1981, and Aki) showed a preference for activity verbs during the emergence of -te i-. The use of -te i- by Taachan is reminiscent of the anecdotal cases of children who do not start speaking for a long time, but once they start, they utter full sentences, i.e., these children do not go through the one-word and two-word stages. These children may be conservative, not speaking out until they feel comfortable. Taachan's case may be similar, although his conservatism is domain-specific. He did not produce a -te i- form at all until 2;2, and once he started he was very productive, using it even in the past form. If this is the case, this child should then be considered as an exception rather than the rule and the case is not necessarily problematic for the hypothesis that children's -te i- morphology emerges with activity verbs.92 Needless to say, we still need to analyze data from more Japanese-speaking children. A possible reason for the early preference of -te i- with activity verbs is the effect of input distribution in parental speech. As discussed in Chapter 4, Shirai (1994) found that a child acquiring English showed frequent uses of stative progressive, including incorrect uses such as *Georgie is seeing light, in contrast to two other children who rarely used stative progressives. The mother of this child used stative progressives quite frequently whereas the other two mothers did not use them at all. The use of Aki's mother's -te i- showed a relatively high frequency of progressive meaning at early stages (Shirai 1998b), and this could be

Acquisition of tense-aspect 139 a factor for Aki's early use of -te i- with progressive meaning. But Aki's mother's use was already skewed toward achievement at 1;10, and Aki started to use -te i- only at 2;1. Is it possible that there is such a delayed effect of skewed distribution in the input (Moerk 1980)? We need more data on this point, as well. Is it the case that progressive inflections are rarely overextended to Stative verbs? None of the four children (Sumihare in Shirai 1993; Aki, Yocchan, Taachan in Shirai 1998b) whose verb forms were analyzed for lexical aspect produced a stative progressive, i.e., there was no token of -te i- used with stative verbs. Stative progressive is not totally ungrammatical in Japanese, as noted above. Shirai (1995) reported that 5% of the -te i- forms used in a sample of adult-adult discourse were on stative verbs. Therefore, the complete absence of stative progressive by Aki's mother (Shirai 1998b) is noteworthy. As discussed in Chapter 4, in English, two mothers whose data Shirai (1994) analyzed never used stative progressive, and this parallels the result from Aki's mother. Since stative progressives are a marked, non-standard use of the progressive (Smith 1983), they may tend not to be used in child-directed speech, either because of unconscious avoidance by mothers or because of discourse context - i.e., there is a preponderance of "here and now" contexts in caretakers' interactions with children, which may not require progressive uses of statives. Thus it is not surprising that there were no tokens of stative progressive in the speech of any of the four Japanese children. But it should be pointed out again that individual variations are still possible, as shown in the English data in Shirai (1994), and in a Japanese child who used one token of incorrect stative progressive (Horiguchi 1981). In the next sections, we will review studies of L2 acquisition of the Japanese tense-aspect system, and then consider the prototype hypothesis and theoretical issues related to both LI and L2 acquisition. 6.2.3. Second language There has been a growing interest among L2 researchers on the acquisition of tense-aspect morphology in Japanese. Most of the studies involve classroom learners, and there are two studies on untutored learners.

140 Acquisition of Japanese Instructed learners Kurono (1994, also 1995) was the first to systematically investigate the effect of verb semantics in L2 acquisition of tense-aspect in Japanese. Based on a grammaticality judgment task administered to 14 classroom learners (of various LI backgrounds) at 3, 6 and 9 months after their arrival in Japan, Kurono found that the learners had more difficulty in correctly accepting or rejecting sentences with resultative meaning (i.e., achievement verb + -te i-) than those with progressive meaning (i.e., activity verb + -te i-). Shirai (1995) analyzed conversational interviews with three Chinese learners of Japanese studying at a Japanese university, and found that they showed stronger associations between past tense -ta and achievement verbs, and between durative -te i- and activity verbs than comparable data taken from native speakers' speech. (See also Shirai and Kurono 1998, which reported and discussed major findings from Kurono 1994, 1995 and Shirai 1995.) Sheu (1997b) investigated two groups of Chinese learners of Japanese in a classroom setting, one in Japan (N=30), and the other in Taiwan (N=30), using a picture description task, and a paper-and-pencil test requiring learners to provide appropriate verb forms in context. The picture description task showed that for both groups, the progressive use of the durative aspect marker -te i- was easier than its resultative use. This is consistent with Shirai's (1995) and Kurono's (1994, 1995) results. What is interesting is her analysis of various verbs that learners produced. She singled out the verbs for which the correct answer is the resultative use of -te i-, and divided them into easy items (accuracy higher than 80%) and difficult items (accuracy lower than 60%). It turned out that all the easy items can be translated into Chinese by verbs combined with the durative marker -zhe (e.g., qiche ting-zhe 'the car has stopped [and it is there]'; see discussion of -zhe in 5.1.1). In contrast, the difficult items cannot be translated into Chinese in this way, but has to be translated by a verb with the perfective marker -le (see 5.1.1). Sheu suggested that since the learners associate -le so closely with the past tense marker -ta, they had difficulty using -te i- when the resultative use of -te i- involves Chinese equivalents that are associated with the use of perfective marker -le. Indeed, most of the errors that the learners produced in relation to these types of verbs involved the overuse of the past tense marker -ta, that is, they used -ta when -te i- is obligatory. This is very strong evidence for the influence of the LI.

Acquisition of tense-aspect 141 Sheu's paper-and-pencil test yielded a similar result. She compared eight different meanings that -te i- can take (e.g., experiential perfect, progressive, resultative, habitual/iterative, etc.) in terms of learners' accuracy scores. Regarding the progressive vs. resultative meanings, it was found that for both groups, learners had more difficulty in using resultative -te i- than progressive -te i-. Another interesting finding was that although there was no difference in proficiency as measured by a cloze test, the learners in Japan consistently outperformed the learners in Taiwan, and the difference was more conspicuous in the picture description task than in the paper-and-pencil test. This may suggest that the added element of a naturalistic environment, where both ample exposure to input and frequent output opportunities are available, may be more effective than classroom learning only for the successful acquisition of core grammatical morphology such as aspect. Nishikawa (1998) also reported data which supports the claim that the progressive meaning -te i- is easier than its resultative meaning for L2 learners of Japanese. In both written and oral elicitation tasks administered to 18 learners93, she found that learners showed a higher accuracy rate in producing the -te i- form in obligatory contexts for progressive meaning than for resultative meaning. Koyama (1998) reported data that clearly argue for the role of LI transfer in the acquisition of -te i-. In his cross-sectional study of classroom learners of Japanese using a grammaticality judgment test, he found that both Korean and Chinese learners had more difficulty acquiring the resultative state meaning than the progressive meaning, which is consistent with the data reported in previous studies. In addition, he found that Chinese learners had more difficulty with -te i- attached to achievement verbs than Korean learners. This appears to be due to LI transfer. As noted above, in Chinese, the imperfective marker -the cannot be attached to achievement to denote resultative state; rather, for achievement verbs the perfective aspect marker -le must be used, which presumably makes it difficult for Chinese learners to accept some resultative uses of -te i-. Korean appears to correspond more closely to Japanese in that the semantic space covered by the Korean resultative marker -o iss- is quite similar to the semantic space covered by the Japanese durative marker -te i- used with achievement verbs (Shirai 1998a).

142 Acquisition of Japanese Koyama (1998) also reported an interesting result concerning the judgment on -te i- for different time references. The learners' accuracy order for three types of progressive use was (1) present progressive (-te iru; 85%), (2) past progressive (-te i-ta; 61%), and (3) future progressive (-te i-ru; 37%).94 It appears that the prototype of -te i- is the present reference. Shibata (1999) analyzed a frog-story narrative (Herman and Slobin 1994) elicited from four classroom learners of Japanese at a university in the US. She found that there was a strong relationship between past tense form -ta and the achievement verbs (88%, type count). The durative aspect marker -te i- was also more often used with achievements (48%) than with activities (41%) based on type count. She also noted that some of her learners exhibited inappropriate uses of -te i- with achievement verbs: in Japanese, achievement used with -te i- denotes resultative state, but some of the learners used it to refer to ongoing action: (7)

bin kara de-te i-masu jar from come:out-Asp-Pol (intended meaning) '...is coming out of the jar'

Shibata (1999) discussed four examples of this type, which she attributed to LI transfer; that is, English allows the progressive marking combined with achievement to refer to the process leading up to the endpoint, and the learners were using the durative marker -te i- by analogy to the English progressive. The above example (7) in Japanese must mean that the frog has come out of the jar, not that it is coming out as its morphological equivalent in English would. Naturalistic learners Although most of the studies on L2 Japanese tense-aspect have looked only at classroom learners, Kurono (1998) and Shibata (1998) have examined naturalistic learners. Kurono (1998) analyzed the speech of eight untutored factory workers whose LI was Brazilian Portuguese. The data, which consisted of recordings at two different times, separated by 10 months,95 came from various tasks such as interviews, picture descriptions, and role plays. Kurono reported that at Time 1, 13 out of 16 tokens of the past tense form -ta (81%) were attached to achievement verbs, whereas the ratio was 75% (33 out of 44) at Time 2. Regarding the durative aspect marker -te i-, 6 out of 9 tokens were used with activities at

Acquisition offense-aspect

143

Time 1. However, at Time 2, only 3 out of 10 tokens (33%) were used with activities, and achievements predominated (70%). Kurono noted, however, that 5 of the 7 uses of achievement with -te i- at Time 2 were with suwaru 'sit' and were used by the same learner, and therefore the result concerning -te i- is inconclusive. Shibata (1998) analyzed a one-hour interview with an untutored factory worker (LI Portuguese) acquiring Japanese in a naturalistic setting. She reported that past tense -ta did not correlate with achievements (only 31% by type count, and 15% by token count), and that -torn, the dialectal equivalent of -te i-ru,96 did not correlate with activities. However, it appears that her classification of verbs was slightly different from that in other studies, although she used Shirai's (1993) test for classifying verbs. Since she reported all the verb types used by the learner, it was possible for us to reanalyze her data. Our reanalysis yielded the following results (see Table 6.1). Table 6.1. Frequencies of past (-ta) and imperfective (-toru) forms used by lexical aspect by Shibata's (1998) learner Lexical aspect

Tense-Aspect Markers -toru type token type token

State

3

8

0

0

Activity

2

28

2

8

Accomplishment

1

3

0

0

Achievement

6

25

0

0

Verb types kakeru'can paint', iru 'exist', wakaru 'understand' hanasu ' speak ' , miru 'watch', neru 'sleep', okoru ' get angry' kaku 'write' kuru 'come', ireru 'put in', iku 'go^yameru 'quit', tukareru 'get tired', bikkurisuru 'get surprised'

The two verbs used with the imperfective -toru were neru 'sleep' and okoru 'get angry' both of which are admittedly borderline cases, but if we apply the tests they are achievements. Shibata also classified kuru 'come' as activity, but it should be achievement, according to Step 2 of Shirai's tests. Thus, even though Shibata (1998) claimed that her results are not consistent with the Primacy of Aspect hypothesis, or the prototype hypothesis as we call it in this book,97 this is not necessarily the case.

144 Acquisition of Japanese That is, past forms are used with achievements very frequently, and -te i* ι · · · OR with activities. One interesting observation that Shibata (1998) made was the association of the nonpast -ru form with state verbs in her data. This is similar to observations in the LI (Bloom, Lifter, and Hafitz 1980; Shirai 1991) and L2 (Robison 1995) acquisition of English, where there is an association between the third person singular present -s inflection and state verbs. The studies of L2 Japanese tense-aspect morphology reviewed here show that learners of Japanese as a second language generally follow the universal tendency predicted by the prototype hypothesis. This is interesting, because in native Japanese speech the association is not simple, in that achievement is associated with both past and durative imperfective morphology. Shirai (1995) showed that L2 learners follow the prototype hypothesis, predominantly using the past tense form -ta for achievements and the imperfective form -te i- for activities. The study also showed that native speakers associate -ta with achievements, but not as strongly as L2 learners because they have more flexibility in using verb morphology. L2 learners also show much stronger association of -te i- with activities than native speakers do; that is, learners follow more closely the prediction of the prototype hypothesis. Kurono (1994, 1995) showed that the progressive meaning of -te i-, which the prototype hypothesis predicts will be acquired earlier, is accepted more easily by learners than the resultative meaning in their obligatory contexts. Subsequent studies by Sheu (1997a, 1997b), Nishikawa (1998), and Koyama (1998) all confirmed that the resultative meaning is more difficult for learners than the progressive meaning, supporting the prototype hypothesis. What has also become clear in recent research on Japanese as L2 is that there is a clear effect of LI influence. Sheu (1997a, 1997b), Koyama (1998), and Shibata (1999) all suggest that there is an effect of LI transfer as seen in the acquisition of Japanese by Chinese, Korean, and English speakers. In particular, Sheu (1997b) showed a strong influence of LI in Chinese learners' production of -te i- when referring to resultative state: the learners had much more difficulty in producing the correct form when the LI equivalent is not expressed by the imperfective marker -zhe.

Prototype hypothesis 145 6.3. Prototype hypothesis and the acquisition of Japanese as LI and L2 How do we interpret and integrate these various findings? We have so far argued that first and second language acquisition of tense-aspect morphology in various languages starts with the prototype for each aspectual morpheme; that is, with achievement verbs for past morphology, and with activity verbs for progressive morphology. Then the use of each morpheme spreads to other classes of verbs as the learner's language approximates the native adult norm. What we have found in Japanese L 2 acquisition can be interpreted along this line. The learners' use of the morphemes is seen to be restricted to the prototypes when compared to that of native speakers, who are capable of using the morphology in nonprototypical ways. If we accept this claim as the description of learner behavior, we still need to address the question of where the prototypes come from. One argument that we proposed was that learners acquire the prototypes from the input (see 4.2.2). For example, in English, the majority (approximately 60%) of progressive and past inflections are attached to activity verbs and achievement verbs, respectively, in speech addressed to children (Shirai and Andersen 1995), and thus children can create the initial prototypes from the input. However, this scenario may not work for Japanese L2 with respect to the acquisition of the durative aspect marker -te i- since the results of Shirai (1995) reveal that in native speech -te i- is more frequently attached to achievement verbs than to activity verbs. This casts doubt on the universality of the distributional bias in the input as the only causal explanation for prototype formation. On what basis, then, do learners construct their prototype (i.e., -te iwith activity verbs)? One interpretation is that it comes from the influence from LI. Here, most probably, the Chinese learners have found a close affinity between the Chinese progressive marker zai and the Japanese -te i-. Zai can in most cases be translated into Japanese -te i- when denoting progressive meaning, but this is not the case for -te i- when denoting resultative meaning. In other words, there is a straightforward mapping between Chinese zai and Japanese -te i- for progressive meaning, but not for resultative meaning. This is one possible explanation of the early uses by Chinese learners of Japanese of the progressive meaning of -te i-. Although it is apparent that the learner's LI does influence the acquisition of -te i-, it may not be the only reason why Chinese learners find it

146 Acquisition of Japanese easier to acquire the progressive meaning as the prototype of the durative aspect marker -te i~. Another possibility is that the prototype comes from the effect of instruction. It is often the case that when Japanese -te i- is introduced in the classroom, the sequence is from progressive meaning to resultative meaning, and in fact, for the learners in Shirai (1995), that was the case. If the progressive meaning is introduced first, it is not surprising that learners treat the progressive meaning as the prototype of -te i-. However, this interpretation may not be tenable, considering that the learners in Kurono (1995) had not yet studied -te i- at the time of Test 1 (3 months after arrival in Japan), when they already found resultative more difficult than progressive, and only eight of them continued on with the formal instruction which could influence the results of Test 2 (6 months) and Test 3 (9 months). Therefore, the effect of instruction seems to be minimal in such cases. In any event, the effect of instruction needs to be tested empirically. For example, one possibility for testing the effect of instructional sequencing is an experiment where different groups are taught two meanings of -te i- in different order, and where the two groups' knowledge of the two different meanings of -te i- is compared. The results of Shirai (1995) regarding the native speaker's use of -te i- also present a puzzle for the LI acquisition of Japanese, and in particular the distributional account. If children start to use -te i- with its progressive meaning by attaching it to activity verbs (as observed for some children), the higher percentage of -te i- used with achievement verbs in adult native speech is problematic, since children cannot induce from the input that the prototype of -te i- is action in progress. It appears that there are three possible solutions to this puzzle. The first possibility is that there is an a priori conceptual/linguistic predisposition on the part of the learner to map morphology onto "action in progress" (or "process") rather than onto "states", other things being equal (Shirai 1997b). That is, an ongoing action is more salient than a static situation, and other things being equal, grammatical functors tend to be attached to the former rather than latter. The second possibility is that the distributions of the tense-aspect forms in child-directed speech are different from those in adult-directed speech. In fact, Shirai (1993) looked at a portion of adult speech addressed to a child acquiring Japanese, and 78% of the tokens of -te iwere attached to activity verbs. Shirai (1998b) reported a similar skewing toward activities in the use of -te i- by Aki's mother at early stages.

Prototype hypothesis 147 However, the conclusion is still tentative since the relevant portion of the diary data Shirai (1993) analyzed only recorded the adult utterances that triggered the child's speech, Aki's mother's frequent use of -te i- with activities was over even before Aki started to use -te i- (see 6.2.2). Therefore, we cannot be certain of this possibility until we examine more Japanese LI acquisition data. The third possibility, noted by Shirai (1993), is that the initial restriction of -te i- to activity is due to the alternative form (-to) in Japanese for referring to resultative state. As noted earlier, Japanese past tense form -ta, which used to be a perfect suffix -tan, still retains its old function and is sometimes used to refer to resultative perfect. Since both LI and L2 learners generally prefer to assign one form to one meaning (see 4.2.2), it is not surprising if learners associate -ta with achievement to denote past and resultative perfect and -te i- with activity to denote progressive. In other words, learners do prefer simple mapping, associating -ta with telicity, and -te i- with lack of telicity, even though that is not the case in Japanese usage (i.e., achievement is associated with both -ta and -te i- for native Japanese speakers). Finally, it should be noted that these three possible explanations are not mutually exclusive. In fact, it is possible that in LI acquisition (1) a universal predisposition to attach functors to grammaticizable notions (in this case, action in progress), (2) input distribution, and (3) a one-to-one mapping principle all contribute to the observed phenomenon. In the case of the L2 acquisition of Japanese, the effect of LI transfer, and possibly instructional effects, in addition to (1) (2) and (3) above, may be contributing to the acquisitional patterns observed. The task for future research is to tease out which factors (if any) are possibly less relevant, and which ones are more important, and to understand how these factors interact with each other. Our approach in this book, however, has been to minimize the reliance on universal predispositions and to emphasize input-based distributional learning. Thus, from our perspective, we rely on (2) and (3) for the sake of parsimony to explain LI acquisition of aspect in Japanese. For this input-based scenario to work, more systematic research is needed to investigate the input distribution that children acquiring Japanese are exposed to, and the way such distribution influences children's acquisition. For second language acquisition, too, we can rely on (2) and (3) to explain learners' preference for -te i- with activity verbs, but in this case

148 Acquisition of Japanese the effect of LI and the effect of instruction may subtly interact with (2) and (3), since most of the learners that have been studied so far have a productive progressive morphology in their LI (Chinese, English, Korean, Portuguese, etc.). The real test concerning the effect of LI has to come from learners whose LI does not have a productive progressive marker (e.g., French, German, or Russian). If they show a similar preponderance of activity verbs in the acquisition of -te i-, then we can safely conclude that LI transfer is not important in their initial preference/or -te i- with activity verbs. 6.4. Summary In this chapter, we have discussed how LI and L2 acquisition of Japanese tense-aspect morphology can be characterized by its interaction with lexical aspect. As is the case with English and Chinese reviewed in the previous chapters, Japanese exhibits a strong association between achievement verbs and past/perfective morphology, and a lack of association between state verbs and progressive or imperfective morphology. With respect to the association between activity verbs and progressive/imperfective morphology observed in English and Chinese, the picture is more complex for Japanese, since the Japanese durative imperfective marker denotes both progressive meaning (with activity and accomplishment verbs) and resultative meaning (with achievement verbs). On this point, although LI children generally show an early preference for the progressive use of the imperfective morphology, some children do not. L2 learners, in contrast, consistently exhibit the earlier use of progressive meaning. Explanations of these observations are discussed with respect to the prototype hypothesis, and it is suggested that more focused research is necessary, including the study of inputoutput correspondences in LI acquisition, and experimental studies with L2 learners, in particular with learners whose LI does not have a progressive aspect. Such research will speak to the issue of how lexical aspect and grammatical aspect-tense are associated and interact in LI and L2 learning processes.

Chapter 7 A connectionist model of the acquisition of aspect In Chapter 3 and elsewhere we proposed that our approach to the acquisition of aspect is consistent with a connectionist approach to language acquisition. In this chapter, we argue in detail that the acquisition of aspect can be best described as a connectionist process in which the learner is engaged in computing the probabilistic co-occurrences of semantic properties, lexical forms, and morphological devices. We implement this process with a specific connectionist model, the selforganizing neural network model of language acquisition. 7.1. Connectionist models of language acquisition

7.1.1. Connectionism and language acquisition Connectionist theory, or connectionism, is a computational account of human information processing. A connectionist network is based on the use of massively connected units (or sometimes called "neurons") that process information in parallel. Because of their structural and processing characteristics, these networks are often referred to as Parallel Distributed Processing (PDP) networks or artificial neural networks." Two key assumptions of connectionism have to do with (a) representation knowledge is represented as patterns of activation distributed across highly connected processing units, and (b) learning - new knowledge is formed through the adaptation of the strengths or weights that hold among the connections of multiple processing units. A typical connectionist network consists of several layers of units: the input layer, the hidden layer, and the output layer. The input layer receives information from input representations, the output layer provides output representations produced by the network, and the hidden layer provides the network's internal representations as they are formed during the network's learning of the input-output mappings (comparable to the various processing states of the human brain). In a typical feed-forward network, information processing goes from the input layer to the hidden layer, and then from the hidden layer to the output layer. During learning, the activations as well as the connections of the various units are

150 Λ connectionist model adjusted according to specific algorithms. Different networks use different learning algorithms for computing the input-to-output mapping function. The most widely used learning algorithm is "back propagation". According to this algorithm, each time an input-to-output mapping is learned, the discrepancy (or error) between the actual output (produced by the network) and the desired output (provided by the modeler) is calculated, and is then propagated back to the network so that the relevant connection weights can be adjusted relative to the amount of errors (Rumelhart, Hinton, and Williams 1986). Continuous weight adjustments in this way lead the network to fine-tune its activation of units and strength of connections in response to the regularities in the input-output relationships. At the end of learning, the network derives an optimal set of weight configurations so that it can take on any input and produce the correct output. The connectionist approach to information processing contrasts with traditional symbolic, rule-based approaches in significant ways. Connectionist models rely on notions that are more biologically motivated than notions in symbolic models. For example, the notions of multiple processing units, activation, excitation, inhibition, and connection strengths provide us with more neurally plausible constructs for conceptualizing information processing than do discrete symbols, rules, and abstract categories.100 The human brain consists of a massive network of neurons working together, often in parallel, and the analogy from real neuronal networks to connectionist networks is clear (Spitzer 1999). Connectionist theories assume a high degree of interactivity between various levels of information processing, in contrast to highly modularized, often serial, and "informationally encapsulated" processes assumed by symbolic theories (Fodor 1983). With regard to language, advocates of connectionism argue that linguistic representations (of the lexicon, morphology, and grammar) are "emergent properties" due to the interaction of the learning system with the linguistic environment. Through detecting regularities in the formmeaning mapping process, connectionist networks demonstrate capabilities in inducing syntactic and semantic structures from the learning environment. This view contrasts with the symbolic view that often emphasizes the psychological reality of linguistic rules and the representational innateness for the a priori status of some grammatical and semantic categories. The concept of emergent property is crucial in connectionist theory. A shortcut to the understanding of this concept

Connectionist models of language acquisition 151 may come from the following example. Structured, "rule-like" representations in a connectionist network can emerge in much the same way as a hexagonal structure emerges from the honeycomb: every honeybee packs a given amount of honey to the honeycomb from a given angle, but no honeybee has a grand plan for the formation of the hexagonal structure (Bates 1984). One might at first hypothesize that the honeybees have a symbolic, rule-like representation of "making a hexagonal shape", but such a hypothesis is unnecessary in the connectionist view, given that the dynamics of the honey-packing process (i.e., each honeybee maximizing the packing density of spheres) guarantees the emergence of a hexagonal shape. A classic example that illustrates connectionist versus symbolic views of language acquisition is the acquisition of the English past tense (see also discussions in 3.1.1). The empirical phenomenon is that children overgeneralize the regular past tense -ed to irregular verbs and produce forms like failed, breaked, and corned at some stage (an "overregularization" process), typically after they have used the correct irregular forms for some time (Brown 1973; Bybee and Slobin 1982; Kuczaj 1977). The theoretical debate focuses on what computational and psycholinguistic mechanisms characterize the acquisition of the regular and irregular forms of the English past tense. Symbolic theorists insist that at some stage of acquisition children have internalized a linguistic rule (something like "add -ed to a verb stem to make the past tense form"), and overregularization errors such as blowed, breaked, and corned result from children's application of this rule too broadly (to cover irregular verbs). In particular, Pinker and his colleagues (Pinker 1991, 1999; Pinker and Prince 1988, 1994; Marcus et al. 1992) argue that the acquisition of the English past tense is a dual-mechanism process, one mechanism involving the acquisition of a regular morphological rule, and the other the associative learning of exceptions (i.e., irregular verb forms). In contrast, connectionist researchers (Rumelhart and McClelland 1986; MacWhinney and Leinbach 1991; Plunkett and Marchman 1991, 1993; Seidenberg 1997) argue for a single mechanism for the acquisition of both regular and irregular past tense forms, a mechanism realized as connectionist learning with distributed representations of knowledge and adaptive connection weights. According to this view, overregularization errors thus reflect the child's ability to extract statistical regularities in the input and the ability to use the extracted pattern productively (e.g., flow, glow, and slow all take -ed to make the past, and so should blow). A

152 Λ connectionist model single mechanism would predict that there will be not only regularizations - the child's application of regular forms to irregular verbs (e.g., corned for came), but also irregularizations - the child's application of irregular patterns to regular verbs (e.g., ment for mended). Indeed, both regularizations and irregularizations exist in children's speech. How does a connectionist approach help us understand children's acquisition of aspect? In section 4.2.2, we considered how the acquisition of tense-aspect morphology can be accounted for by a prototype theory. Because aspect stands at the interface between the morphology and the lexicon, children do not acquire the tense-aspect morphology alone. Rather, they acquire it along with the verbs to which the morphology applies. In this perspective, the formation of prototypes of tense-aspect morphemes goes hand-in-hand with the formation of categories of the lexical aspect of verbs. In the next section, we provide an account of how semantic categories of verbs can emerge from learning in connectionist networks.

7.7.2. A connectionist account of the acquisition of lexical semantic categories How do children acquire inherent semantic categories of lexical aspect? Previous theories have assigned a significant role to innate factors with regard to this question, as discussed earlier. In contrast to these theories, our proposal is that lexical aspect categories can emerge from connectionist learning, in much the same way as the representation of "cryptotype" emerges from learning in connectionist networks. According to Whorf (1956), a cryptotype is a covert semantic category that underlies the use of some overt linguistic devices, for example, morphological markers. The precise meaning of a cryptotype is "subtle", "intangible", and hard to pin down, but the linguistic device that it licenses is productive. Cryptotypes are only definable negatively through the constraints that they place on how morphemes can be combined; for example, the use of the prefix un- in English has an underlying cryptotype - that is, the meaning of the set of verbs that can take un- forms a cryptotype. Thus, English allows unbuckle, uncoil, undress, unfasten, and untie, but not *unkick, *unmove, *unpush, or *unswim. Whorf further presented a thought experiment to show that it is the cryptotype that licenses the use of un- in these cases: if we coin a verb flimmick to mean "to tie a tin can to something", then we are willing to accept the

Connectionist models of language acquisition 153 sentence He unflimmicked the dog as expressing the reversal of the "flimmicking" action. However, if flimmick means "to take apart", then we will not accept He unflimmicked the puzzle as describing the act of putting a puzzle back together. Whorf took this as evidence in support of his claim that native speakers of English have an intuitive feel for what verbs can and cannot take un-, even though linguists cannot clearly describe the underlying category with a label. Here we have a case for which language use is conditioned by underlying linguistic principles, but the principles themselves cannot be clearly described by crispy symbolic rules. In which way can we compare the acquisition of lexical aspect to that of cryptotypes? Unlike the prefix un-, which is restricted to a specific set of verbs in the adult language, the tense-aspect suffixes like -ed and -ing can in principle be used with all verbs, and the meaning of the verbs does not in general place a restriction on when the past form -ed or the progressive -ing is allowed in the adult language.101 However, the picture becomes different when we look at child-directed parental speech, in which the use of tense-aspect suffixes is, indeed, constrained by the lexical aspect of verbs: telic verbs typically occur with -ed, while activity verbs typically occur with -ing, and Stative verbs never or rarely occur with -ing (see 4.2). Some of these constraints occur even in adult-toadult speech; for example, stative verbs and some achievement verbs typically resist the use of progressive aspect (Smith 1983; see also discussion in 5.1.3; 5.4 on Comrie's "naturalness of combination" principle). Children might be able to use such distributional constraints to identify the semantic properties shared by the verbs that take -ing versus those that do not or those that take -ed. It is in this sense that we view the acquisition of the lexical aspect categories of verbs as a process similar to that of the acquisition of cryptotypes (see also discussion in 4.3.2 on "morphological bootstrapping"). Bowerman (1982) was the first to point out that cryptotypes might play a significant role in children's acquisition of un-. Her data suggested that children's overgeneralization errors with un- tend to fall into Whorf s cryptotype. What was not clear from her analysis, however, was how the child extracts the cryptotype and uses it as a basis for morphological generalization (when the cryptotype is intangible even to linguists like Whorf). Li (1993a) and Li and MacWhinney (1996) hypothesized that cryptotypes could naturally emerge from connectionist learning. They simulated a connectionist network to learn the reversive cryptotype

154 Λ connectionist model associated with the use of un-. They used a standard three-layer backpropagation network in their model, in which the input consisted of semantic features of verbs, and the output was either un-, its competitor dis-, or no prefixation. Results from the modeling indicate that the network formed internal representations of semantic categories that correspond to Whorf s cryptotypes, on the basis of learning limited semantic features of verbs and morphological classes. The network was able to compute the combinatorial constraints on the co-occurrence of the prefix and the semantic features distributed across verbs, and extracted a meaningful representation of the verbs. Moreover, on the basis of the semantic representations that it developed, the network produced overgeneralization errors similar to those reported by Bowerman (1982) and Clark, Carpenter, and Deutsch (1995), such as *unhold, *unpress, *unfill, ^unsqueeze, and *untighten. The success of the modeling led Li and MacWhinney (1996) to hypothesize that cryptotypes seemed intangible only because of the limitations associated with traditional symbolic methods for analyzing complex semantic structures. The meaning of a cryptotype constitutes a complex semantic network, in which words in a cryptotype can vary in (a) how many semantic features are relevant to the cryptotype, (b) how strongly each feature is activated in the representation of the cryptotype, and (c) how features overlap with each other across members in the cryptotype. These complex structural properties for whole groups of words make a rule-based analysis less effective if not impossible, but lend themselves naturally to distributed representations and statistical learning in connectionist networks. Thus, Li and MacWhinney (1996) proposed that a cryptotype is supported by multiple features connected with different weights in a network, and the emergence of the cryptotype in the network is the product of the network's learning of the appropriate connections that hold among particular semantic features, lexical forms, and morphological markers. The network computes the combinatorial constraints on the co-occurrences of the prefix un- and the set of weighted semantic features distributed across verbs. The end result of this process is that new representations developed at the hidden layer of the network differ from each other in the number of features they share and in the strengths of feature activation for a pattern, as a result of detecting statistical regularities in the input-output mapping process. Meaningful, categorical representations of verbs emerge from these patterns.

Connectionist models of language acquisition 155 Very much the same story can be told about the acquisition of lexical aspect of verb categories. In an attempt to explain tense-aspect acquisition from a connectionist perspective, Li (2000b, in press-a) has proposed that in early child language, features such as [+dynamic], [+punctual], and [+telic] interact collaboratively through summed activation to support the formation of a lexical aspect category that licenses the use of the past or perfective forms, such as the English -ed or the Chinese -le. These features collaborate in the sense that a given verb may be represented with multiple features, and the features themselves often co-occur in the situations to which the verb applies. For example, drop can be viewed as indicating both a punctual and a change-oflocation meaning, and crash as involving both punctual and a changeof-state meaning. A feature may also vary in the strength with which it is represented in different verbs. With varying degrees of connections from semantic features to verb forms, verbs can form clusters or categories that differ overall in lexical aspect. The implication of this feature-based account of lexical aspect of verbs is that for the child, then, the learning of grammatical morphemes like -ing or -ed is not simply the learning of a rule, but the accumulation of the connection strengths between -ing or -ed and a set of weighted features shared by the verbs. At the same time, the child also develops a feature-based organization of verb categories. Within this scenario, the learning process is best described as a correlational, statistical procedure in which the learner implicitly tallies and registers the frequency of co-occurrences (strengthening what goes with what) or the co-occurrence constraints (inhibiting what does not go with what) among morphemes, features, and verbs. Such a process can be captured by a biologically motivated computational principle called Hebbian learning (Hebb 1949); we will explore Hebbian learning in more detail later in a simulated model of tense-aspect acquisition. How does such a statistical, correlational analysis actually work for the learner? There are several ways in which children may be engaged in correlational analyses. First, children may analyze the semantic properties of verbs through a verb's co-occurrences with situational contexts (e.g., some verbs refer to result, others do not). For example, de Lemos (1981) showed how parents intentionally draw infants' attention to event structures, such as the difference between result and process, through carefully modeled input in parental speech (see 8.4.3 for details concerning de Lemos's study). Second, children may extract semantic representations through a verb's co-occurrences with other words (e.g.,

156 Λ connectionist model break typically occurs with glasses, while tear with pieces of paper). Li, Burgess, and Lund (2000) showed how global lexical co-occurrence information in sentences can be used by the learner to derive accurate lexical semantic representations. And finally, children may compute the co-occurrences of particular grammatical morphemes (e.g., -ed and -ing) with sets of semantic features that turn up repeatedly across lexical items (e.g., the recurrent telic and punctual meanings of achievement verbs), as we discussed above. These correlational analyses (forms with contexts, forms with forms, and forms with meanings) also account for the learner's ability to build and later to relax the strong associations between emergent categories of lexical aspect and tense-aspect morphology. As discussed earlier, tenseaspect morphemes are associated with lexical aspect categories in parental speech, and to a lesser extent in adult speech in general. Using the correlational analysis outlined above, children may initially restrict particular tense-aspect forms to particular lexical aspect of verbs because they are sensitive first to only the most frequent or salient associations (prototypes) in the input (see also 4.2.2). As the distributional biases become weaker in the input, and as the child's lexical repertoire grows, the strong associations between lexical aspect of verbs and tense-aspect morphology will gradually relax. Categories of lexical aspect no longer need to be defined by the grammatical markers with which they are used (as are required with cryptotypes), and prototypes of tense-aspect morphology start to spread their uses to different types of verbs. In the end, lexical aspect becomes independent, prototypes spread to nonprototypes, and the associations between the two weaken.102 In the rest of this chapter, we will provide a more technical analysis of this account in a simulated connectionist model. In particular, we explore the self-organizing neural network as a model of lexical and morphological acquisition. We argue that self-organization and Hebbian learning are two important computational principles that can account for the psycholinguistic processes of semantic representation and morphological productivity in the acquisition of aspect.

7.2. A self-organizing connectionist model of language acquisition As discussed above, in the last decade a heated debate surrounding connectionism and language acquisition has been on the acquisition of

Self-organizing connectionist model 157 the English past tense. The core of the debate is whether language acquisition should be characterized as a symbolic, rule-based process or as a connectionist, statistical learning process. Symbolic theorists insist that regularization errors in the acquisition of the past tense result from the child's internalization and application of an overly broad linguistic rule, whereas connectionist researchers argue that such errors reflect the learner's ability to extract statistical regularities from the input. There are two major gaps in this debate that we take as points of departure for our current investigation. First, few studies in this debate have paid attention to the meaning structure of the verbs with which the past tense form is used. Although the change from a verb's present to its past tense is largely governed by phonological properties of the word, as modeled in connectionist networks, empirical evidence indicates, as we have discussed throughout the book, that children's early past tense forms carry an aspectual function and are highly restricted to telic or resultative verbs, due to the interaction between lexical aspect and grammatical aspect. Second, most of this debate has revolved around a specific cluster of connectionist models, the three-layered back-propagation network as a model of language acquisition. Several limitations to the back-propagation algorithm are now known, especially in the context of language acquisition: in particular, back-propagation utilizes error signals from the discrepancy between desired and actual outputs. At each time an input is paired with an output, the network calculates an error between what should be outputted and what it actually produces, and uses this information in further adjustment of connection weights (see 7.1.1). This type of error, feedback, and adjustment procedure seems unrealistic to the language learner (and perhaps to biological systems in general; see Crick 1989). According to the well-known "no negative evidence" hypothesis (Baker 1979; Bowerman 1988; Pinker 1989), children do not receive constant feedback about what is incorrect in their speech, at least not the kind of error feedback on a word-by-word or input-by-input basis as provided in the back-propagation network. Thus, backpropagation networks would seem to be poor candidates as models of language acquisition on grounds of their psychological or biological implausibility. In this study, we explore self-organizing neural networks as a potential class of candidate models for language acquisition. In particular, we use self-organizing feature maps to model the interaction between lexical and grammatical aspect. In contrast to back-propagation networks

158 Λ connectionist model (networks with supervised learning), self-organizing neural networks use unsupervised learning that does not require the presence of a supervisor or an explicit teacher; learning is achieved entirely by the system's selforganization in response to the input. The self-organizing process extracts an efficient and compressed internal representation from a highdimensional input space, and often expresses this new representation in a two-dimensional map structure, as in the feature map networks (Kohonen 1982, 1989, 1995). There are several important properties of selforganizing feature maps that make them particularly well suited to the study of lexical and morphological acquisition: self-organization, representation, and Hebbian learning. Self-organization in these networks typically occurs in a twodimensional topological map, where each processing unit (neuron) is a location on the map that can uniquely represent one or several input patterns.103 At the beginning of learning, the first input pattern activates one of the many units on the map, according to how similar by chance the input word is to the weight vector (a weight vector is the activation pattern of all individual weights going to a given unit). The second input proceeds in the same way to activate another unit on the map, and so on. If the second input is similar to the first input, it will probably activate the same or a nearby unit. This is because when a unit becomes activated in response to the input, the weight vectors of that unit and the units in its neighborhood are adapted such that they become more similar to the perceived input pattern; thus the unit and its neighboring units will respond to the same or similar input even more strongly the next time the network sees the input. In this way, every time an input is presented, an area of units (i.e., a neighborhood) will become activated on the map (the so-called activity "bubbles"), and the maximally active unit is taken to represent the input. Initially, learning occurs in large neighborhoods, so that all units in a neighborhood will become active in response to the input. Gradually, learning becomes more focused as the neighborhood size and the learning rate decrease, so that only the maximally responding units are active. This process continues until all the inputs have found some maximally responding units. In the end, the network develops "specialist" unit or units for each input, and similar inputs will end up activating the same unit or units in nearby regions of the map. As a result of this self-organizing process, the statistical structures implicit in the high-dimensional space of the input are extracted as topological structures and represented on a two-dimensional space.

Self-organizing connectionist model 159 Because the network develops activity bubbles to capture the input space, similar inputs will be represented on the same or nearby regions on the 2-D space, yielding new similarity structures that become clearly visible on the feature map. Ritter and Kohonen (1989) showed how such a map can represent structural relationships among animal categories, after the network has seen and self-organized a handful of surface and functional features of animals. This type of self-organized representation has clear implications for language acquisition: the formation of activity bubbles may capture critical processes of the emergence of semantic categories in children's acquisition of the lexicon. In particular, the network organizes information first in large areas of the map and gradually zeros in onto small areas; this zeroing-in process is a process from diffuse to focused patterns, reflecting the network's continuous adaptation to its representation. This process naturally explains many generalization errors reported in the literature: for example, substitutions of put for give ("put me the bread") or fall for drop ("I failed it"; see Bowerman 1978a) reflect the child's recognition of diffuse lexical similarities but not focused fine distinctions between the words. Miikkulainen (1997) also showed that in a lesioned self-organizing feature map, behaviors of deep dyslexia (confusion of words with related meanings, e.g., dog for sheep), as well as surface dyslexia, can arise from partial damage to the semantic representation (in effect a diffuse representation of meaning).104 Self-organizing feature maps can also be connected via Hebbian learning, such as in Miikkulainen's (1993, 1997) multiple feature-map model: initially all units on one map are connected to all units on the other map; as self-organization takes place, the associations become more focused, such that in the end only the maximally active units on the two (or more) maps are associated. Hebbian learning is a co-occurrence learning mechanism, according to which the associative strength between two neurons is increased if the neurons are both active at the same time (Hebb 1949). The amount of increase is proportional to the level of activation of the two units. Hebbian learning has been considered to be a highly biologically-plausible mechanism, and has been studied extensively in models of neural networks and biological systems (Fausett 1994; Hertz, Krogh, and Palmer 1991). Although Hebbian learning is not an intrinsic property of self-organizing neural networks, its combination with this type of network has strong implications for language acquisition, in particular for the acquisition of lexical and grammatical aspect: it can account for the process of how the child abstracts the

160 Λ connectionist model relationship between word forms, lexical meanings, and suffixes, on the basis of how often they co-occur and how strongly they are co-activated in the representation. We will return to this point in our discussion of the simulation results. Because of these properties - self-organization, representation, and Hebbian learning - self-organizing networks (a) allow us to track lexical development as an emergent property more clearly in the network's selforganization (from diffuse to focused patterns or from incomplete to complete associative links); (b) allow us to model one-to-many or manyto-many associations between forms and meanings in the development of the lexicon and morphology, and (c) provide us with a set of biologically plausible and computationally relevant principles to study language acquisition, without the use of direct negative evidence. These models are biologically more plausible because we could conceive of the human cerebral cortex as essentially a self-organizing map (or multiple maps) that compresses information on a two-dimensional space (Kohonen 1989; Spitzer 1999). They are computationally more relevant because we could argue that child language acquisition in the natural setting (especially organization and reorganization of the lexicon) is largely a self-organizing process that proceeds without explicit teaching (MacWhinney 1997, 1998). Li (1999, in press-b) recently used a self-organizing feature-map network to model the acquisition of cryptotypes and morphological overgeneralization, and obtained interesting results. His results indicate that the network was able to form representations that correspond to Whorf s cryptotype, a covert semantic category that Whorf (1956) believed governs the use of the English prefix un- (see 7.1.2). The network's representations in turn allowed it to generalize to novel instances of verbs as children do. Moreover, the network also recovered significantly from overgeneralization errors, on the basis of its ability to adjust the associative pathways between verb semantics, lexical forms, and prefixes through Hebbian learning. In the following section, we apply this model to the domain of tense-aspect acquisition in English. 7.3. Method As discussed above, self-organizing feature maps learn on the basis of self-organization, produce representations in a map structure, and form associative connections via Hebbian learning. These properties have

Method 161 recently been implemented in DISLEX, a multiple feature-map model of the lexicon (Miikkulainen 1997). In this study, we use the basic architecture of DISLEX to simulate the acquisition of aspect. We think that the self-organizing and Hebbian learning processes as implemented in DISLEX can help us to understand the interaction between lexical and grammatical aspect in language acquisition. 7.3. L Network architecture In DISLEX, different self-organizing maps dedicated to different types of linguistic information (e.g., orthography, phonology, and semantics) are connected through associative links via Hebbian learning. These maps can be further connected to maps dedicated to other sources of information (e.g., sensorimotor memory; see Figure 7.1). During learning, an input pattern activates a unit or a group of units on one of the input maps, and the resulting bubble of activity propagates through the associative links and causes an activity bubble to form in the associated map. If the direction of the associative propagation is from phonology or orthography to semantics, comprehension is modeled; production is modeled if it goes from semantics to phonology or orthography. The activation of co-occurring lexical and semantic representations leads to continuous organization in the corresponding maps, and to the adaptive formation of associative connections between the maps. Figure 7.1 presents a schematic diagram of the architecture of the model. DISLEX has been used successfully to model the mental lexicon in normal and disordered language processing (Miikkulainen 1993, 1997). In this study, we apply it to the study of the acquisition of lexical and grammatical aspect.105 We constructed two self-organizing maps, each of the size of 50 χ 50 units, one for the organization of phonological input (henceforth the phonological map), and the other for the organization of semantic input (the semantic map). We used no orthographic maps since we were modeling acquisition in young children who are preliterate, as lexical organization and morphological acquisition take place well before literacy (Bowerman 1978a, 1982).

162 A connectionist model

:

j

**·*··«·

* 4·*«>··.·».»··««·»·«»·«··»

ÜMMHbai jjr

Figure 7.1. DISLEX: A self-organizing feature-map model of the lexicon (Miikkulainen 1997; reproduced with author's permission)

7.5.2. Input data and representations In order to model the role of linguistic input in children's acquisition of lexical and grammatical aspect, we selected as our input data the parental or caregivers' speech in the CHILDES database (MacWhinney 1995). We extracted all the utterances produced by parents, caregivers, and experimenters from the CHILDES database in about half of the English corpus (from Bates to Korman).106 Although not all of these utterances are child-directed, they form a representative sample of the speech that children are exposed to (e.g., dinner table talk, activities during free play, storytelling, etc.). The total number of lexical items in our corpus is 764,907 words (tokens). A verb from this corpus was chosen as input to the network if it occurred in the parental or caregivers' speech for five or more times in a given age period (see 7.3.3). With this criterion we selected a total of 562 verbs (types) as input to our network. They were inputted to the network in five stages, according to the age groups at which they occurred (see 7.3.3). Previous connectionist models of language acquisition have often relied on the use of artificial input/output representations (e.g., randomly generated patterns of phonological or semantic representations) or representations that are constructed ad hoc by the modeler. For example,

Method 163 in our own previous studies (Li 1993a; Li and MacWhinney 1996) we represented each verb as a pattern of 20 semantic features, selected on the basis of our linguistic analyses. Representations of linguistic information in this way are often subject to the criticism that the network works precisely because of the use of certain linguistic features (see Lachter and Bever 1988). To overcome potential limitations associated with this approach, in this study, we used more realistic input data to simulate the acquisition of aspect. To our knowledge, this is the first simulation of language acquisition that uses real input data to which children are exposed. We represented our inputs in the following ways. Phonological representations. Phonological representations of verbs to our network were based on a syllabic template coding developed by MacWhinney and Leinbach (1991). Instead of a simple phonemic translation, this representation reflects current autosegmental approaches to phonology (Goldsmith 1976), according to which the phonology of a word is made up by combinations of syllables in a metrical grid, and the slots in each grid made up by bundles of features that correspond to phonemes, C's (consonants) and V's (vowels). The MacWhinneyLeinbach model used 12 C-slots and 6 V-slots that allowed for representation of words up to three syllables. For example, the 18-slot template CCC W CCC W CCC W CCC represents a full tri-syllabic structure in which each CCCW is a syllable (the last CCC represents the consonant endings). Each C is represented by a set of 10 feature units, and each V by a set of 8 feature units. The following vector is an example of this representation for a disyllabic verb (1 indicates the presence of a feature, e.g., "dental", while 0 the absence of a feature). The suffixes -ing, -ed, and -s were also represented in the same way. Decide 1010000000000000000000000000000100100000000 000001000001000000000000000000000100010110000000 010100000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000 000000000000000000000000000

In this example, the first 10 digits "1010000000" represent the consonant /d/; the next 20 digits are all zeros because decide has no initial consonant clusters. The next 8 digits (30-38) represent the vowel /i/, and so on. These segments of the representation are filled into a left-

164 A connectionist model justified grid, as in dCC IVsCCal dCC WCCC (see the 18-slot template above). Semantic representations. Semantic representations of verbs to our network were based on the lexical co-occurrence analyses in the Hyperspace Analogue to Language (HAL) model of Burgess and Lund (1997) and Lund and Burgess (1996). According to HAL, the meaning and function of a given word is determined by lexical co-occurrence constraints in a high-dimensional space of language use. HAL focuses on global rather than local lexical co-occurrences: A word is anchored with reference not only to other words immediately preceding or following it, but also to words that are further away from it in a variable cooccurrence window, with each slot (occurrence of a word) in the window acting as a constraint dimension to define the meaning of the target word. In short, global lexical co-occurrence is a measure of a word's total experience in the context of other words (i.e., its contextual history) - what words occur before and after a given word, and how frequently. Thus, in this perspective, a word can be represented by a vector that encodes the multiple lexical constraints in a large-scale language corpus. Note that this type of representation differs from our earlier semantic representations (e.g., Li and MacWhinney 1996) in that here the modeler does not (and cannot) determine which semantic features are relevant to the task at hand; instead, the representation is purely based on lexical cooccurrence information in the speech corpus. For example, the following 100-unit vector represents the semantics of decide; each unit represents the degree of the co-occurrence constraint from another associated word (the degree of constraint is set on a continuous scale from 0 to 1). Decide 0.030 0.295 0.040 0.348 0.040 0.009 0.011 0.103 0.015 0.017 0.017 0.226 0.008 0.029 0.003 0.023 0.020 0.068 0.005 0.006 0.042 0.064 0.008 0.011 0.055 0.010 0.017 0.008 0.008 0.014 0.059 0.016 0.188 0.011 0.152 0.053 0.003 0.074 0.063 0.011 0.020 0.008 0.009 0.034 0.026 0.010 0.019 0.016 0.010 0.041 0.010 0.091 0.026 0.021 0.012 0.010 0.038 0.006 0.038 0.037 0.016 0.005 0.073 0.288 0.010 0.007 0.076 0.010 0.040 0.025 0.556 0.022 0.026 0.008 0.011 0.008 0.023 0.079 0.012 0.034 0.085 0.021 0.025 0.006 0.420 0.041

0.008 0.010 0.007 0.000 0.012 0.022 0.055

0.010 0.068 0.023 0.007 0.040 0.044 0.006

In our network there were no separate semantic representations of the suffixes themselves, but there were semantic representations of verbs plus suffixes. That is, the semantic representations of a verb included both the representation of the verb stem (walk) and that of its inflected forms (walked, walks, walking). During training, the various forms of a verb's

Method 165 representation may be co-activated in the learning process, which brings us to the task and procedure of network training. 7.5.5. Task and procedure Upon training of the network, a phonological representation of a verb was inputted to the network, and simultaneously, the semantic representation of the same verb was also presented to the network. By way of selforganization, the network formed an activity on the phonological map in response to the phonological input, and an activity on the semantic map in response to the semantic input. The phonological representations of the corresponding tense-aspect suffixes were also co-activated with the phonological and the semantic representations of the verb, depending on whether the verb co-occurs with the progressive marker -ing, the regular perfective/past marker -ed, or the third person singular -s in the parental speech in our CHILDES corpus. As the network received input and continued to self-organize, it simultaneously formed associations through Hebbian learning between the two maps for all the active units that responded to the input. The network's task was to create new representations in the corresponding maps for all the input words and to link the semantic properties of a verb to its phonological shape and its morphological pattern. Figure 7.2 presents an example of the two maps along with some co-activations of words on each of the maps. To observe effects of the interaction between lexical and grammatical aspect in the parental input on the network's learning and representation, we designed five stages to train the network, according to the different age groups of our input data (see below). During training, the network received input incrementally (stage by stage), such that input to the network at the current stage included input at all previous stages. At each stage the network was trained for 200 epochs; that is, each phonological and semantic representation of a verb and its suffix was presented to the network 200 times. (1)

Input Age 1;0. Parental/caregivers' data in our CHILDES corpus are available from an age when the child is 6 months old (the Korman corpus). We analyzed all the available input data in the corpora of Korman and Higginson up to Age 1, which had about 1,569 word types

166 A connectionist model (22,661 tokens), among which 85 verbs (type) occurred for five or more times (these were chosen as input to our network). During this stage relatively few verbs in the adult speech occurred with tense-aspect suffixes: out of the 85 verbs 17 (20%) occurred with -ing, 3 (3.5%) with -s and 2 (2.4%) with regular past form -ed. This pattern indicates that children's early predominant uses of bare verb forms might partly reflect the nature of the adult input. At this stage, irregular past tense forms were very rare; only one irregular verb, got, occurred for five or more times, which we included in the training.107

•»•ft* ··* t

bit! τ—I

·ί*·ι ····' τ--~·

·- -ί

-·-··· ------

!—fffaur***

-·'·»··- '

·>~'

i—τ—-'— t-—Seeer--4—"f—

- -· -·

· ].„,

· -- ·"·"

·

Figure 7.2. Phonological and semantic representations in the feature maps after the network was trained on 186 verbs at Stage 2 (see below for training stages). The upper panel is the phonological map (in capital letters), and the lower panel the semantic map. Each unit typically represents one input pattern, a verb in this case, but in some cases a unit may represent multiple inputs when the input words cannot be clearly differentiated by the network (the "more" items on the map). Words longer than four letters are truncated.

Method 167 (2) Input Age 1;6. A total of 2,523 word types (49,408 tokens) from our CHILDES corpus during the period when the child is 13-18 months old were included in our analysis. Many of these verbs overlapped with verbs from Stage 1 (naturally there would be such overlaps), and therefore our input to the network at this stage also included the 85 verbs at Stage 1, yielding a total of 186 verb types. At this stage, the number of tense-aspect suffixes increased, compared to that at Stage 1. Out of the 186 verb types, 34 (18%) occurred with -ing, 9 (5%) with -ed, and 9 (5%) with -s. The progressive marker -ing was the predominant inflection type. At this stage, many irregular past forms occurred in the corpus for five or more times and were included in the training. (3) Input Age 2;0, A total of 4,229 word types (167,623 tokens) during the period when the child is 19-24 months old were included in our analysis. Again, verbs that occurred for five or more times were selected as input to our network, and a total of 324 verb types was used, which included all the new verbs as well as verbs from Stages 1 and 2, and 76 (23%) of these verbs occurred with -ing, 23 (7%) with -ed, and 24 (7%) with -s. (4) Input Age 2;6. A total of 6,452 word types (27,0171 tokens) during the period when the child is 25-30 months old were included in our analysis. A total of 419 verb types which occurred for five or more times was used, including all verbs from Stages 1 through 4, and 82 (20%) of these verbs occurred with -ing, 35 (8%) with -ed, and 31 (7%) with -s. (5) Input Age 3;0. A total of 7,457 word types (25,5044 tokens) during the period when the child is 31-36 months old were included in our analysis. A total of 562 verb types which occurred for five or more times was used, including all verbs from Stages 1 through 5, and 123 (22%) of these verbs occurred with -ing, 70 (12%) with -ed, and 61 (11%) with -s. During training, if a verb occurred at two different stages, the network would receive the verb as input twice, on both the phonological map and

168 Λ connectionist model the semantic map; if it occurred at three stages, the network would receive it three times, and so on. This procedure was also used with the corresponding suffixes of the verb; for example, if -ing occurred with talk at both Stages 1 and 2, then the network would receive, twice, the phonological representation of talk, that of -mg,· and the semantic representation of talk. In other words, the phonological representations of the verb and the suffix and the corresponding semantic representation were co-activated in the network for a given number of times; in essence, this procedure modeled a coarse frequency of the verbs and the tense-aspect markers, even though the frequency variable was not explicitly manipulated in our simulations. Two other parameters were systematically varied in the simulations. The learning rate (a) varied from 0.1 to 0.0 at each stage of training. The neighborhood size (We) started large to cover the entire map and decreased to 1 at the end of each stage (see discussion on selforganization under 7.2). At the beginning of each stage, Nc was slightly increased relative to that at the end epoch of the previous stage, to accommodate the increasing number of input verbs and suffixes at the new stage. All simulations were run on a SUN Spare workstation, using the DISLEX code configured by Miikkulainen (1999).

7.4. Results and discussion In this section we will focus on three levels of analysis of our modeling results: the role of input, the emergence of lexical aspect categories, and the formation and dilution of prototypical associations between lexical and grammatical aspect.

7.4.1. Role of input One important rationale behind the current modeling effort is the understanding of the role of the linguistic input in guiding children's acquisition of lexical and grammatical aspect. In previous chapters, we have emphasized the relationship between patterns observed in children's speech and those in adult speech with respect to the interaction between lexical aspect and grammatical aspect. But a simple correlation between children's and adults' patterns tells us only that the child is sensitive to

Results and discussion 169 the linguistic environment and can analyze and incorporate information from that environment into his or her own speech. It does not tell us how the child actually does this, or what mechanisms allow the child to do this. Thus, we wanted to test whether a connectionist network, endowed with self-organization and Hebbian learning principles, is able to display learning patterns as the child does. Our network receives phonological and semantic representations of input words from actual adult speech, along with morphological information of these words. If the network is able to produce patterns like those we found in children's speech, on the basis of learning from the input, we can then conclude that selforganization and Hebbian learning provide the necessary kinds of mechanisms that drive the formation of patterns in children's acquisition. In this way, our modeling enterprise provides insights into the mechanisms that underlie the learning process. Table 7.1 presents a summary of the major patterns from the network's learning according to the tense-aspect suffixes it produced at the different learning stages. It shows the results of the network's production of three suffixes, -ing, -ed, and -s, with three types of verbs, activity, telic, and stative.108 The table does not include results from the earliest stage, Input Age 1, because there were too few items that had tense-aspect suffixes at that stage (2 -ed and 3 -s forms; see 7.3.3), and it would clearly be misleading to compute the percentages on the basis of these few items. The results for the remaining four different input age groups were based on the analysis of the network's production ability, i.e., how semantic representations activate corresponding phonological representations through associative pathways. The analysis was done by inspecting the units on the phonological map that each verb on the semantic map activated, after the network had been trained for a specified number of epochs at each stage. The results in this table are highly consistent with empirical patterns observed in early child language (see 3.2.1; 4.1): the use of imperfective aspect is closely associated with activity verbs that indicate ongoing processes, while the use of perfective aspect is closely associated with telic verbs that indicate actions with endpoints or end results. In particular, in early child English, the progressive marker -ing is highly restricted to activity verbs, the perfective/past marker -ed restricted to telic verbs, and third person singular -s restricted to stative verbs (Bloom, Lifter, and Hafitz, 1980; Brown 1973; Clark 1996; Shirai 1991). Our network, having taken in input patterns based on realistic parental speech, behaved

170 Λ connectionist model in the same way as children do. For example, at Input Age 1;6, the network produced -ing predominantly with activity verbs (75%), -ed overwhelmingly with telic verbs (82%), and -s exclusively with stative verbs (100%). Such associations remained strong at Age 2, but gradually became weaker though still transparent at later stages (e.g., the association between -s and stative verbs remained strong throughout). Table 7.1.

Percentage of use of tense-aspect suffixes with different verb types across input age groups in the network's production * TENSE-ASPECT SUFFIXES

VERBS Activity Telic Stative Item Totals**

-ίηκ

75 25 0 40 -ing

Activity Telic Stative Item Totals

64 31 0 89

Age 1;6 -ed -s 18 0 82 0 0 100 11 3 Age2;6 -ed -s 26 0 74 0 0 100 19 7

-ing

66 28 0 71 -ίηκ

52 44 4 70

Age 2;0 -ed 16 84 0 19 Age 3;0 -ed 9

77 14 22

-s 0 0 100 9 -s 10 10 80 10

* This table includes only verbs that could be uniquely assigned to one or the other suffixation pattern and does not include instances for which the network produced a given verb with multiple suffixes. See Table 7.3 for the latter. ** These are the total number of verbs that occurred with the given suffix. Note that the percentages within a given column does not always equal 100%, reflecting the fact that some verbs could not be easily classified into one or the other category. This is also true for other tables in this chapter.

Interestingly, when we analyzed the actual input to our network, we found similar patterns. Recall that the input to our network was based on parental speech from the CHILDES database. Table 7.2 presents the percentages of use of suffixes with different verb types in the input data. An analysis of this table indicates that in the input data there were also clear associations between -ing and activity verbs, -ed and telic verbs, and -s and stative verbs, and that these associations are strong throughout the four stages, as also found previously by Shirai (1991) and Olsen et al (1998) (although our analysis here does not include irregular past forms).

Results and discussion 171 Table 7.2. Percentage of use of tense-aspect suffixes with different verb types across input age groups in the input data Tense- Aspect Suffixes Verbs Activity Telic Stative Item Totals

-inf> 69 28 3 29

Verbs Activity Telic Stative Item Totals

-ing

67 25 8 40

Age 1;6 -ed 22 77 0 3 Age 2;6 -ed 23 69 8 13

-s 0 33 67 9 -s 20 20 60 10

-ing

74 24 2 54 -ing

67 31 2 60

Age 2;0 -ed 15 77 8 13 Age 3;0 -ed 23 65 12 26

-5

17 0 83 12 -s 23 8 69 13

The high degree of correlation between the network's production and the input it was exposed to shows that our network was able to learn on the basis of the information of the co-occurrences between lexical aspect (verb types) and grammatical aspect (use of suffixes). This learning ability was due to the network's use of Hebbian learning in computing (a) when the semantic, phonological, and morphological properties of a verb co-occur and (b) how often they do so. Note that the patterns in Table 7.1 and Table 7.2 are consistent and similar, but not identical. This is important because if the learner, child and network alike, simply mimicked what's in the input by recording each individual word and suffix and their co-occurrence, the learner would have no productive control of the relevant linguistic device and would simply produce the patterns verbatim. Our results suggest that the associations between verb types and suffixes are stronger in the network's production than they were in the input to the network. It seems that our network, like the child, behaved more restrictively than what is in the input with respect to the relationship between lexical and grammatical aspect. This pattern is consistent with the prototype account that we proposed in Chapter 4, according to which children's initial uses of tense-aspect suffixes are restricted to prototypical instances (see 4.1.3 and 4.2.1 where we discussed how a non-absolute tendency in the input can lead to a more absolute association in the child's representation). A comparison of Tables 7.1 and 7.2 indicates that the stronger associations were true for all verb types at all stages, especially for stative verbs. For

172 A connectionist model example, the association between stative verbs and -s in the input was not perfect; a few stative verbs occurred with other suffixes, and a few -s's occurred with other verbs. In the production by the network, however, the association between stative verbs and -s was perfect for Stages 2 to 4 (Input Age 1;6 to Age 2;6), and near perfect for Stage 5 (Input Age 3;0). This result shows that the network may have established strong pathways connecting stative verbs and -s through Hebbian learning, on the bases of (a) stative verbs were identified on the map clearly as a transparent cluster of verbs (more on this in 7.4.2), (b) most stative verbs occurred with -s, and (c) the co-occurrence of stative verbs with other suffixes was infrequent. These analyses remind us of two points: First, we need a more detailed understanding of how categories of lexical aspect, for example, activity, telic, and stative verbs, emerge in the network. Second, we need a more detailed understanding of how input interacts with frequency to affect the formation and dilution of prototypical associations in the acquisition process. We deal with the first point in 7.4.2 below, and with the second point in 7.4.3. 7.4.2. Emergence of lexical aspect categories In 7.1.2, we proposed an account of the acquisition of lexical aspect of verbs as a process comparable to that in the acquisition of cryptotypes, i.e., covert semantic categories that govern the use of morphology. We suggested that the emergence of lexical aspect of verbs could be the result of multiple features working together through summed activation. In our simulation, we implemented this idea by feeding our connectionist network with verbs that are represented with multiple semantic features (lexical contextual constraints), and we wanted to see how semantic categories of lexical aspect could emerge from a self-organizing learning process. As discussed earlier, a distinct property of self-organizing feature maps is that the structure in the network's new representations can be clearly visualized as activity bubbles or patterns of activity on a twodimensional map; this property obviates the need of extra steps of mathematical analysis (e.g., cluster analysis or principal component analysis for visualization). In our network, the self-organizing process extracts statistical information from the high-dimensional space of semantic representations of verbs in parental input, so that verbs that are

Results and discussion 173 similar are clustered together on the feature map. Figure 7.3 presents a snapshot of the network's self-organization of the semantic representations of 186 verbs at the end of Stage 2 (Input Age 1;6). pull

rub

tooc

ihou

• cru

fjni

«Irin

Fujie * i

i lee

k»··

»»r«

515

«ore chew

U

wit

Itop

t"* · ·

J«ti

fly

uie

EEE

Lift 1ΓΜΗ

turn

ihut

thro

c«un

•*^t

put

fit

con«

^ff?

»i·»

court

looh

for«

i*Q see

5*3 he«r

Io«t «oi

ffi

jjj

ns

call

•e«r

try

l*«rt* WOfMl

f*«I

tfwu I4W

kr>w tlun

I«H

EH

hurt

lw

cooh

More

fflll

luch

ll«CI

#*t

UorV

>h«h star h«n«

*0

hol.»

ilf«

EU

1'·»

33

-ΤΓΓ

^— 7*79

7ΓΪ"

wife

pick

*Mt