university of hawai'! library - ScholarSpace


By Terry Klafehn

Dissertation Committee: Ann M. Peters, Chair Niklaus R. Schweizer John H. Haig William D. O'Grady Benjamin K. Bergen

© Copyright 2003

By Terry Klafehn



There are a number of people I wish to thank for making this dissertation possible. I could not have completed my research and writing without their assistance, professional and technical expertise, and warm encouragement. First of all there is my wife Sharyn, who patiently urged me to keep going when I was discouraged and overwhelmed by the task. I wish to thank my friend Mr. Kazuki Daimon, president of the International Sports Education and Cultural Foundation (ISEC Japan), for making arrangements in Japan so that I was able to conduct research at the Ushigomi Nakano Kindergarten in Shinjuku. My friend Jeffrey Hayden from the EALL department spent many long hours preparing the computerized test that I used with the Japanese children. He continues to patiently help me with computer problems. Wendy Onishi from the office of the Department of Linguistics processed any number of important documents (including a number of extensions) for me in a timely manner. Her "Well, are we making progress?" was more encouraging than she realizes. Evelyn Nakanishi, student services specialist at EALL was always willing to listen to my complaints, offer some encouragement, and conclude our conversation with the friendly admonition "Get back to work!" Stan Starosta made a very important contribution this dissertation. We seemed to always bump into one another on Sunday mornings near the Natatorium during the early spring of 2002. He was on his way from brunch at the New Otani and I had just finished swimming. One morning he said, "I have a book you might be interested in." The Bybee book that he recommended later became very important to my dissertation. I wish I could have shown him this completed version.


Finally, I am deeply grateful to my dissertation committee and wish to thank them all for their guidance and helpful comments. A chance meeting on an airplane to the mainland some years ago led to my acquaintance with Nicklaus Schweizer. I admire his enthusiasm for languages and life and appreciated his encouragement. I was especially happy that Benjamin Bergen agreed to serve on my committee because he enthusiastically represents a newer cognitive approach to linguistics that I hope to be able to adopt in future work. I admire William O'Grady for his dedication to the field of linguistics and for his intellectual rigor. I thank him for his very careful review of my drafts, his insightful questions and helpful comments. I admire John Haig because he is an accomplished linguist and a fluent Japanese reader, writer, and conversationalist. I appreciate the time he took with the details of my drafts. He caught my careless Japanese errors, and at the same time raised questions with serious methodological and linguistic implications. Finally, I cannot say enough to thank my chair Ann Peters. She was so patient and encouraging for so long. She treated each rough draft that I gave her, especially in the beginning, in the same way as some article that she was editing for journal publication. No matter what I gave her, she always made some suggestion that made it better. I really appreciate the effort she made to help me.



This study investigates how speakers of Japanese mentally represent and process verbal inflection. Japanese exhibits an agglutinating inflectional typology, however, morpheme boundaries are not always transparent and there is considerable stem allomorphy. Furthermore, there are no bare stems. The stems of consonant-stem verbs never appear in isolation because they are unpronounceable phonotactic violations. This feature of Japanese presents two problems to a rule hypothesis of verbal processing, whereby regular verb forms are produced by the combination of stem and inflection: 1) How do Japanese speakers compute stems? 2) Are speakers of Japanese able to mentally represent and process forms that they cannot pronounce? An alternative to a rule hypothesis is Bybee's Schema Model, which allows for the

mental representation of fully inflected forms. In this study, Schema Model and rule hypothesis predictions about errors and productivity are compared. First, it is shown that the native analysis of inflection reflects phonological and orthographical constraints consistent with the notion that native speakers do not segment verbs into stems and endings. Second, results of a search of the Miyata database show that: 1) Japanese children under three years of age do not overregularize. 2) Most verb errors are stem errors. 3) There are many more errors with regulars than with irregulars. 4) There is no default error pattern. Third, a written test asks fifty adult native speakers and fifty adult instructed (L2) learners to choose appropriately inflected nonce forms. The learners outperform the native speakers. 76% of the learner group responses are correct, but only 53% of the


native group responses. No evidence is found that learners or natives make use of a default rule. Finally, an oral response, nonce probe test with Japanese children (five and six years of age) finds that the children cannot productively inflect novel verbs. It is concluded that the lack of default error patterns and the inability of native

children and adults to productively inflect novel verbs is best explained by a Schema Model whereby inflectional morphology emerges from use (including verb type and token frequency) and not from the manipulation of abstract verbal stems.


TABLE OF CONTENTS Acknowledgements Abstract

iv vi

Chapter 1: Introduction 1.1. Processing of Japanese verbal inflection 1.2. Typology and processing mechanisms 1.3. Emergent properties 1.4. Chapter outline

1 1 3 4 5

Chapter 2: Japanese verbal inflection 2.1. Outline of Japanese verbal inflection 2.2. Regular vowel-stem verbs 2.3. Regular consonant-stem verbs 2.3.1. Stem-final r verbs Stem identification Stem allomorphy Conjugation of derived forms 2.3.2. Stem-final k, g, and s verbs 2.3.3. Stem-final m, b, and n verbs 2.3.4. Stem-final w verbs 2.4. Irregular verbs and forms 2.4.1. Irregular verbs kuru and suru 2.4.2. Suppletive and irregular derived forms 2.4.3. Irregular nai forms 2.4.4. Irregular form itta 2.4.5. Five irregular honorific verbs 2.4.6. Irregular bound verbal nouns 2.4.7. Irregular free verbal nouns 2.5. Native Japanese analysis of verbal inflection 2.5.1. The six stem system 2.5.2. No-root verbs in native analysis

7 7 8 10 10 10 11 14 17 18 19 20 20 22 23 23 24 26 27 30 32 36

Chapter 3: General morphological models 3.1. Introduction 3.2. Models and theories 3.2.1. Hockett's 1954 models Item and process (IP) Item and arrangement (IA) Word and paradigm (WP) 3.3. Conclusion: Hockett's five evaluation criteria

39 39 39 40 41 42 44 47

Chapter 4: Processing mechanisms 4.1. Introduction 4.2. Rules 4.3. Analogy 4.4. Competing cues 4.5. Connections

49 49 49 52 64 68


Chapter 5: Process models 5.1. Introduction 5.2. Words and rules model (Pinker 1999) 5.3. Connectionist models 5.4. Schema models (Bybee's 2001 network model) 5.5. Japanese inflection and model predictions

72 72 73 77 82 88

Chapter 6: Acquisition of inflection by Japanese children 6.1. Introduction 6.2. Contrastive English and Japanese acquisition 6.2.1. English acquisition order 6.2.2. Japanese acquisition order 6.3. Database study 6.3.1. Japanese equivalent of *goed 6.3.2. Japanese equivalents of English *doed and *comed 6.3.3. Overregularization of irregular *nai 6.3.4. Overregularization and overgeneralization 6.3.5. Stem segmentation 6.3.6. Regular verb and irregular verb error frequency 6.3.7. Error patterns 6.3.8. Japanese default error pattern

90 90 90 91 92 94 94 97 100 102 105 108 109 112

Chapter 7: Experimental studies: Nonce probe tests 7.1. Introduction 7.2. Previous Japanese nonce probe experiments 7.2.1. Oral nonce probe experiments Sentence completion (de Chene 1982) Paradigm matching (Batchelder 1999) 7.2.2. Written nonce probe experiments Paradigm matching (Yokomizo 1990) Forced choice nonce test (Vance 1991) 7.3. New replication of the Vance test 7.3.1. Purpose of the replication 7.3.2. Questionnaire form with sample responses 7.3.3. Schema and strong rule hypothesis predictions 7.3.4. Method Japanese subjects Instructed learner subjects 7.3.5. Results 7.3.6. Discussion Default rule versus type frequency Type frequency and productivity Low type frequency and low productivity High type frequency and high productivity.............. Low type frequency and low productivity (k stems) Differential performance: natives and learners 7.4. Oral nonce probe test with Japanese children 7.4.1. Purpose of the experiment and research questions 7.4.2. Subjects 7.4.3. Equipment 7.4.4. Procedure 7.4.5. Test format

117 117 118 118 118 119 122 122 125 127 127 129 130 131 131 132 132 135 135 137 138 138 140 141 142 142 145 146 146 146


7.4.6. Results 7.4.7. Analysis 7.4.8. Conclusion

149 149 150

Chapter 8: General discussion 8.1. Summary of results 8.2. Implications of results 8.2.1. First language acquisition of inflection 8.2.2. Experimental studies 8.3. Future directions 8.3.1. Developmental patterns 8.3.2. Experimental studies 8.3.3 Latency of response experiments

152 152 155 155 158 159 159 160 161

Appendix I: Japanese dictionary verb type-frequency Appendix II: Pokemon test frames (English translation)

164 174






This dissertation investigates how speakers of Japanese mentally represent and process verbal inflection. It is widely believed that speakers of languages with inflecting verbs commit only the stems of verbs to memory and then combine stems with appropriate inflectional endings as needed. An English speaker might very economically store only the stem walk in memory and then make use of combinatory rules to produce walk-ed or walk-ing. This type of processing would eliminate the redundant storage of walk, walked, and walking. It seems almost obvious that this would be the most economical way for speakers all languages to represent and process verbs. After all, for agglutinating languages like Turkish, it is possible for a number of different inflectional endings to accumulate after a stem, and it is estimated that a single verb might have as many as a million possible different variations (Hankamer 1992).

Japanese, like Turkish, exhibits an agglutinating inflectional typology. Unlike the three or perhaps four forms of an English verb (walk, walked, walking, or prove, proving, proved, proven), every Japanese verb has close to two dozen possible

inflected forms. However, there is another typological feature of Japanese that makes it quite unlike Turkish. Turkish has very clear morpheme boundaries. Japanese verbs, on the other hand, demonstrate considerable stem allomorphy, and morpheme boundaries are not always transparent. Furthermore, unlike English, there are no bare stems in Japanese. The stem of aruk-u 'walk' is aruk-, but this form never appears in isolation because it constitutes an unpronounceable phonotactic violation. This typological feature of Japanese presents two problems to the rule hypothesis of verbal


processing: 1) How do Japanese speakers compute the stems of verbs? 2) Are speakers of Japanese (or any language) able to mentally represent and process forms that they cannot pronounce? Some theories of language have proposed that the principles that apply to syntactic and morphological structures take place at very abstract levels of representation. Within such theoretical frameworks it makes no sense to ask whether a form is pronounceable or not and the question is generally considered to be uninteresting or trivial. However, the present dissertation is concerned with the actual mental processing of language and considers models of processing that make specific predictions that can be tested. The testing of such predictions is not an easy task. Mental processes are not subject to direct observation and even the most recent experimental methods (for example, neuroimaging with functional magnetic resonance imaging, fMRI) are subject to severe economic, technical, ethical, and theoretical constraints. At the present time neuroimaging remains costly and interpretation of results often remains unclear. Therefore, this dissertation considers two related types of indirect data that represent the results of processing: errors and productivity. First, the speech errors of Japanese children will be examined in an attempt to find patterns of error. Second, the ability of Japanese children, Japanese adults, and instructed learners of Japanese to recognize and produce inflected forms of novel verbs will be contrasted and compared. There are two recurrent themes in this dissertation, the relationship between typology and mechanisms of processing, and the representation and processing of verbal inflection as an emergent property. A brief summary of each theme follows.



In 1957 the linguist Martin Joos wrote that languages could "differ from each other without limit and in unpredictable ways". In that same year Chomsky's Syntactic Structures was published and established the notion that there were in fact profound similarities among all languages that were more significant than any superficial typological variation. By 1963 Greenberg was able to show that a small number of descriptive categories could accommodate a wide range of syntactic and morphological variation. Furthermore, it appeared that those categories could be used to predict other grammatical properties. For example, SVO languages (English) tended to also be prepositional languages, while SOY languages (Japanese) tended to be postpositional. These implicational universals seemed to be even more evidence that languages were more alike than different, and for more than forty years it has been fashionable in linguistics to point out how much progress has been since Joos's time. The following quote is from Sells (1985): Languages differ (within certain specifiable limits) in constituent and word ordering, in which unbounded dependencies will be permitted, in which constituents can be omitted, in which words have syntactic idiosyncracies, and in few other ways. They do not, as an earlier generation of linguists maintained, differ without limit.

More recently, Pinker (1994:233) has emphasized that the 4,000 to 6,000 different languages in the world are much better characterized in terms of structural similarities rather than idiosyncratic differences. Pinker notes that Greenberg found no fewer than forty-five language universals. On the face of evidence like this, it would seem that the few relatively minor typological idiosyncrasies of Japanese (no bare stems, basically no syllable final consonants, and many inflected forms) should not present a problem to a carefully


constructed universal model of processing. A good model is designed to accommodate structural variation. However, other investigators have pointed out that most so-called universals are tendencies rather than absolute universals. In phonology, for example, (Maddieson 1984) it may be that the strongest universal statement one can make is to say that all languages have low vowels and plain stop consonants. Even if languages do not vary without limit, attempts to explain syntax, phonology, and morphology in terms of structure have been confounded by the realization that the typological variation among languages that does exist seems to require an ever increasing number of descriptive categories.


Rather than assuming that structures determine processes, the alternate position considered in this dissertation is that grammatical, morphological, and phonological structures emerge from more basic mechanisms. The notion of emergence comes from Bybee's (2001: 3) use of the term to describe her own schema model. The basic idea behind emergence as it will be applicable here is that certain simple properties of a substantive nature, when applied repeatedly, create structure.

In the schema model morphological structure is considered to be emergent because semantically and phonologically similar words and phrases are mentally represented as unsegmented forms. Structure arises from generalizations across forms represented in memory. Within this theoretical framework, Bybee (2001:6-7) lists six basic principles, two of which are especially relevant to the processing of morphology. 4. Generalizations over forms are not separate from the stored representation of forms but emerge directly from them. In Langacker's terms, there is no 'rule/list separation' (see Chapter 2). Generalizations over forms are expressed as relations among forms based on phonetic and lor


semantic similarities. New forms can be produced by reference to existing forms, but most multimorphemic words are stored whole in the lexicon. 5. Lexical organization provides generalizations and segmentation at various degrees of abstraction and generality. Units such as morpheme, segment, or syllable are emergent in the sense that they arise from the relations of identity and similarity that organize representations. Since storage in this model is highly redundant, schemas may describe the same pattern at different degrees of generality (Langacker 2000).


In Chapter 2, the linguistic description of Japanese inflectional morphology is compared and contrasted with the traditional native description. Attention is paid to the fact that the traditional native analysis appears to be limited by orthographic constraints. With the exception of the mora obstruent and the moraic nasal, there is no way to represent a syllable final consonant using the native Japanese script (kana). However, unlike the economical linguistic analysis, the structure of the complicated multi-stem analysis that emerges is consistent with the phonotactics of Japanese. The implications for process models of inflection are considered. In Chapter 3, a review of Hockett's classic 1954 paper on three descriptive models of morphology concludes with criteria for the evaluation of models. The significance of "economy" in the evaluation of process models is questioned. In Chapters 4 and 5 the distinction is made between descriptive and process models. The details of alternative processing mechanisms and models are compared. While process models are designed to account for what speakers are able to do correctly, they also offer explanations for morphological error (for example,


overregularization as a result of rule application in English). Rule Model and Schema Model predictions about productivity and errors are contrasted. In Chapters 6 and 7 Rule Model and Schema Model predictions are tested by examining errors made by Japanese children during acquisition and by the experimental testing of both children and adults. Verb productivity of young Japanese children is assessed with the first Japanese wug test. Appendix I summarizes the results of original research which compiles Japanese verb type-frequency on the basis of entries in the Daijirin Japanese dictionary. A brief summary and conclusion is offered in Chapter 8.



2.1. OUTLINE OF JAPANESE VERBAL INFLECTION. Japanese verbs are remarkable for their predictability. They are traditionally classified into two regular groups. Both descriptive (Bloch 1946, Martin 1952) and generative accounts (McCawley 1968) recognize vowel-stem and consonant-stem verbs. There are also a few exceptional forms, but only two irregular verbs. The irregular verb su-ru 'do' and the irregular verb ku-ru 'come' both show stem vowel alternation not found in the regular verbs. The verb su-ru has inflected forms si-ta 'did', su-reba 'if do' and si-nai 'don't do'. The verb ku-ru has inflected forms ki-ta 'came', ku-reba 'if come', and ko-nai 'don't come'. The two groups of regular Japanese verbs do not show this stem-vowel alternation and the distinction between vowel-stem verbs and consonant-stem verbs is used to describe suffix allomorphy economically. One set of suffixes occurs with vowel-stem verbs and another set with consonant-stem verbs. For example, the form of the volitional suffix that occurs with vowel-stem verbs is -yoo: tabe-yoo 'let's eat', but the form of the volitional suffix that occurs with consonant-stem verbs is


kaer-oo 'let's go home'. A partial list showing the

contrast between some common vowel-stem and consonant-stem suffix alternations is given in Table 1. Vowel-stem suffixes Consonant-stem suffixes Non-past Conditional Volitional Desiderative

do, does, will do if do let's want to

-ru -reba -yoo -tai

-u -eba


don't, won't





TABLE 1. Vowel-stem and consonant-stem suffix allomorphs.


In addition to suffix allomorphy, there is also allomorphy at the stem boundary of consonant-stem verbs. For example, consonant-stem verbs with stem-final k predictably lose the stem-final consonant in some inflected forms. The past form of kak-u 'write' is ka-ita 'wrote' and not *kak-ta. There are a number of other

phonologically predictable alternations. The past form of yob-u 'call for' is yon-da 'called for' and not *yob-da or *yob-ta. The past form of kaer-u 'go home' is kaet-ta 'went home' and not *kaer-ta or *kaer-da. While the traditional linguistic descriptions of Bloch, Martin, and McCawley make use of a basically invariant-stem analysis, the native Japanese analysis, on the other hand, recognizes a peculiar system of complex stems with widely varying stem shapes. The following sections in this chapter describe and summarize these two contrasting treatments. The purpose of these sections is to consider how these two different descriptions of Japanese verbal inflection, linguistic and native, might also serve as models of inflectional processing by speakers of the language and to point out some of the difficulties associated with stem and suffix computation. A more detailed account of both the traditional and the native analysis can also be found in Vance (1987). Detailed English descriptions of the native analysis can be found in Ikeda (1975), and McClain (1981). Sandness (1999) provides additional details about the historical development of the past and perfective forms.


Japanese vowel-stem verbs have either stem-final i or stem-final e and the shape of the stem is invariant throughout the paradigm. There are no stem-final a, u, or 0 verbs. Table 2 provides a representative partial paradigm of the two verbs mi-ru 'see' and tabe-ru 'eat'. The segmentation shown follows Bloch's 1946 analysis where some

forms (non-past, past, conditional, volitional) are considered basic inflected forms


while others: the infinitive, negative, desiderative, potential, passive, and causative are considered derived forms. Stem-final e Base tabeNon-past tabe-ru eat Past tabe-ta ate Conditional tabe-reba if eat Volitional tabe-yoo let's eat' Infinitive [email protected] eating Negative tabe+na-i don't eat Desiderative [email protected]+ta-i want to eat tabe+rare-ru can eat Potential Passive tabe+rare-ru is eaten tabe+sase-ru make eat, let eat Causative TABLE 2. Regular vowel-stem verbs.

Stem-final i mimi-ru mi-ta mi-reba mi-yoo [email protected]

watch watched if watch let's watch watching don't watch mi+na-i [email protected]+ta-i want to watch mi+rare-ru can watch mi+rare-ru is watched mi+sase-ru make eat, let watch

In Bloch's analysis, the infinitive form of vowel-stem verbs is considered to be a derived form, even though the shape of the infinitive form appears to be identical to the shape of the base. The infinitive form of vowel-stem verbs is analyzed as consisting of the base plus the zero morph (0). This analysis provides pattern symmetry between vowel-stem verbs and consonant-stem verbs. The infinitive form of consonant-stem verbs consists of the stem + i: yom + i 'reading', kaer + i 'returning', oyog + i 'swimming'. In other words, the infinitive of a Japanese verb can be

described as the base plus an additional morpheme. In the case of consonant-stem verbs, the additional morpheme is i. In the case of vowel-stem verbs, the additional morpheme has no phonological realization. Vance (1987: 176) points out that Japanese fits very well the description of an agglutinating language (Lehmann 1983) where morphemes are successively strung together, one after another. And, even in the case of longer Japanese derived forms such as tabe-rare-reba, stem-potential-conditional, 'if you are able to eat' or mi-rarena-katta, stem-potential-negative-past, 'was not able to watch', the morphology of


vowel-stem verbs is relatively transparent. The morphology of regular consonant-stem verbs, however, is more complex and opaque.

2.3. REGULAR CONSONANT-STEM VERBS. A regular Japanese consonant-stem verb has a stem form ending in w or one of the following consonants: k, g, s, t, b, m, n, or r. There is much allomorphy within the paradigms of these verbs. Stem and suffix shapes vary with the stem-final consonant and it is not possible to make general descriptive statements that will be valid for all consonant-stem verbs. In the following sections, the significant details of allomorphy in each paradigm are described.

2.3.1. STEM-FINAL R VERBS. STEM IDENTIFICATION. A partial paradigm of the regular consonant-stem verb kaer-u 'return, go back, come back', with stem-final r, is contrasted with a partial paradigm of the vowel-stem verb tabe-ru from Table 3. Stem-final r

Stem-final e

kaertabe Stem tabe-ru return Non-past kaer-u tabe-ta kaet-ta returned Past if return tabe-reba Conditional kaer-eba tabe-yoo kaer-oo let's return Volitional [email protected] return kaer+i Infinitive not return tabe+na-i kaer+ana-i Negative [email protected]+ta-i want to return Desiderative kaer+i+ta-i tabe+rare-ru can return Potential kaer+eru tabe+rare-ru kaer+are-ru be returned Passive tabe+sase-ru kaer+ase-ru make, let return Causative TABLE 3. Stem-final r verbs and stem-final e verbs.


eat ate if eat let's eat eating don't eat want to eat can eat is eaten make, let eat

These contrastive paradigms highlight some potential complications posed for a model of mental processing that assumes that speakers isolate stems and successively add on suffixes as necessary. First, the bare stems of consonant-stem verbs never appear in isolation. The infinitive form tabe of the vowel-stem verb tabe-ru, exposes the bare stem tabe, but the bare stem of the consonant-stem verb kaer-u is hidden in the derived infinitive form kaer-i. Note that in general, syllable final consonants violate Japanese phonotactic constraints and the consonant-final bare stem kaer is not a pronounceable form for native speakers. Generally, a Japanese syllable has the form (C) (y)V(V/N/Q), where N represents the moral nasal and Q represents the mora obstruent. In other words, Japanese syllables end in a vowel, the mora obstruent, or the mora nasal. One motivation for providing the details of how Japanese verbal morphology has been described is to consider the appropriateness of these descriptions as models of representation and processing. For example, if Japanese speakers actually made use of combinatory rules that combine stems and endings, then they would be obligated to compute the stems of consonant-stem verbs from the fully inflected forms that do occur. However, even in the case of vowel-stem verbs, the importance of the infinitive form as the source of bare stems is questionable because the infinitive is a form used mainly in formal writing and has relatively low type frequency in speech. For all practical purposes, neither the bare stems of consonant-stem nor the bare stems of vowel-stem verbs occur in colloquial speech. STEM ALLOMORPHY.

A second potential complication in the representation and processing of consonant-stem verbs is allomorphy at the stem boundary of the past form. (The same


type of allomorphy is found in related -te, -tara, and -tari forms.) The past form of

kaer-u 'return' is not the expected stem plus ending form *kaer-ta, but the form is kaet-ta. In the orthography used in this paper, kaet-ta represents kaeQ-ta. The appearance of the mora obstruent allomorph Q in the past form is completely predictable, but obscures the shape of the stem (kaer-). On a generative account, the stem-final consonant in the unacceptable C#C sequence at the morpheme boundary is regularly realized as the mora obstruent. Therefore, the shape of the past form is not always a reliable predictor of paradigm identity. The negative form of a Japanese verb is a better paradigm diagnostic. When the negative stem of a verb ends in r or t or w the allomorph Q appears in the stem final position of the past form. Table 4 presents a brief comparison of these three paradigms where the mora obstruent appears in the stem final position of the past form. Non-past



return kaer-anai not return kaef-fa returned a-u meet aw-anai not meet at-fa met maf-u wait maf-anai not wait mat-fa waited TABLE 4. Mora obstruent allomorphy in stem-final r, w, and f verbs. kaer-u

It is important to note that the appearance of the mora obstruent allomorph in the inflected past forms of these verbs poses another complication in the computation of consonant-stem verbs. Since r, w, and t all alternate with Q in the inflected past form, it is possible to have a past form with a possible three-way paradigm ambiguity. A given form might belong to one of three different paradigms: stem-final r, stem-final

w, or stem-final t. The form katta, when written in the native Japanese syllabary or in the romanized script used here could represent the past form of cut, the past form of bought, or the


past form of won. Contrastive partial paradigms of the three verbs kar-u 'cut, crop, shear, harvest', kat-u 'win', and ka-u 'buy' are given in Table 5. Non-past



cut kar-anai not cut kat-ta cut buy kaw-anai not buy kat-ta bought kdt-u win kat-o.nai not win kat-ta won TABLE 5. Overlapping consonant-stem paradigms (accented and unaccented forms). kar-u ka-u

However, there is a spoken distinction between accented and unaccented Japanese verbs. Accented non-past verb forms have accented past forms, while unaccented nonpast verb forms have unaccented past verb forms. The unaccented past form kat-ta 'cut or bought' is minimally distinct from the accented past form kiit-ta 'won'. Stem-final t verbs are all accented. The past form of mot-u 'have' is mot-ta 'had', the past form of but-u 'strike' is but-ta 'struck', and the past form of tat-u 'stand' is

tat-ta 'stood'. However, there are some variable forms. The NHK accent dictionary (1992) lists both accented torimot-u and unaccented torimot-u alternatives of the verb

torimot-u 'mediate, entertain, serve'. In the specific case of the three verbs ka-u, kar-u, and kat-u, the accented past form kat-ta can be identified with the paradigm of kat-u 'win' because verbs with accented past forms also have accented non-past forms. However, whether a verb is accented or unaccented is lexically determined and is not phonologically predictable. Examples of contrasting accented and unaccented verbs from Vance (1987:86) are provided in Table 6. Accent type




Unaccented nar-u ring nar-anai not ring nat-ta Accented no.r-u become nar-o.nai not become no.t-ta Unaccented okona-u perform okonaw-anai not perform okonat-ta Accented tetudo.-u help tetudaw-o.nai not help tetudo.t-ta TABLE 6. Accented and unaccented stem-final r verbs and stem-final w verbs.

rang became performed helped

Some stem-final r verbs are accented and some are not. Some stem-final w verbs are accented and some are not. The significance of this discussion to the problem of 13

stem computation should now be clear. If Japanese speakers must compute the stems of verbs, knowledge of a single inflected form is not always sufficient to establish a paradigm identity and extract a stem. CONJUGATION OF DERIVED FORMS. A third complication in the extraction of consonant-stem verb stems is the fact that there are productive processes that derive vowel-stem verbs, verbs that occur with word final vowel-stem suffixes, from consonant-root verbs. As a result of this process, basic consonant-root verbs can also occur in derived forms with word final vowelstem suffixes. The potential, passive, and causative forms of all Japanese verbs are derived forms that are inflected with word-final vowel-stem suffixes. Table 7 provides contrastive partial paradigms. Fonn

Vowel stem + vowel suffix

Potential Passive Causative TABLE 7.

tabe+ rareru can eat kaer+eru can return tabe+rareru is eaten kaer+areru be returned (adversative interpretation) tabe+saseru make, let eat kaer+aseru make, let return Derivational suffixes for vowel-stem and consonant-stem verbs.

Consonant stem + consonant suffix

In general, Japanese vowel-stem verbs consistently occur with suffixes that are consonant-initial and consonant-stem verbs consistently occur with suffixes that are vowel-initial. For example, the potential suffix is -rareru for vowel-stem verbs, but -

eru for consonant-stem verbs. As described in a previous section, in cases where consonant-stems are forced to accommodate consonant-initial suffixes, adjustments are made to preserve the preferred syllable structure. The stem kaer- and the simple past suffix -ta have the allomorphic realization kaet-ta 'returned' and not *kaer-ta. The derived potential, passive, and causative forms are of interest because these derived forms conjugate like vowel-stem verbs. The derived potential form of the consonant-stem verb kaer-u is kaer-eru, and the negative of this potential form is 14

kaere-nai 'not able to return'. The derived potential stem kaere- occurs with the

vowel-stem suffix -nai and not the consonant-stem suffix -anai. The derived causative stem also occurs with the vowel-stem suffix -nai and not the consonant-stem suffix anai. The derived negative-causative form is kaerase-nai. If verbs with consonant

roots consistently occurred only with consonant-stem suffixes in all inflected forms, the past-potential form of kaer-u would be *kaeret-ta, 'was able to return', but this form does not occur. To illustrate this point, a contrastive partial paradigm is given in Table 8, where derived stems are conjugated as both consonant-stems and vowelstems. The derived consonant-stem forms do not occur and are marked with the symbol *. The transparent and categorical distinction between a consonant-stem verb and a vowel-stem verb, based on the word-final suffix, is lost in these derived forms. Furthermore, it is not just these derived forms which are problematic. There are some basic verb shapes which are always ambiguously vowel-stem or consonant-stem. All verbs with the word-final non-past shape -eru or -iru are potentially either consonantstem or vowel-stem verbs. These ambiguous shapes are considered next. Form

Consonant-stem + consonant suffix

Vowel-stem + vowel suffix

kaere-na-i not able to return Negative-potential *kaerer-anai *kaerer-ta > *kaeret-ta kaere-ta was able to return Past-potential Negative-causative *kaeraser-anai kaerase-na-i not cause to return TABLE 8. Derived stems of consonant-root verbs conjugated as both consonant and vowel stems.

The accented verb kaer-u 'return' conjugates as a consonant-stem verb, but the unaccented verb kae-ru 'change, replace' conjugates as a vowel-stem verb. The unaccented verb kir-u 'wear' conjugates as a vowel-stem verb, but the accented verb kfr-u 'cut' conjugates as a consonant-stem verb. Coincidentally, both of the accented

forms kaer-u and kfr-u happen to conjugate as consonant-stem verbs, but recall from an earlier section that accent is lexical. For example, the accented verb mf-ru 'see, watch' conjugates as a vowel-stem verb. 15

A partial paradigm of -iru verbs showing overlapping forms is given in Table 9. A similar pattern of overlapping forms is found among these -eru verbs. The distinction between the consonant-stem paradigm and the vowel-stem paradigm is thus lost in the non-past, provisional, passive, and short formal potential forms. There is no short potential form for consonant verbs, but the short potential vowel-stem verb form overlaps with the potential form of the consonant-stem verb. Consonant-stem


kfr-u ki-ru cut Non-past kit-ta ki-ta Past cut kir-oo let's cut ki-yoo Volitional kfr-eba Provisional if cut ki-reba kir-i cut and Continuative ki kit-te ki-te Gerund cut ki-nai kir-anai not cut Negative kir-areru ki-rareru Passive be cut kir-eru can cut ki-rareru Potential ki-re-ru Short potential 0 ki-saseru kir-aseru make cut Causative kir-itai ki-tai want to cut Desiderative kir-imasu ki-masu Formal cut Short formal potential kir-emasu can cut ki-remasu TABLE 9. Overlapping forms in the paradigms of -iru verbs.

wear wore let's wear if wear wear and wear not wear be worn can wear can wear make wear want to wear wear can wear

Accent is lexical and there are both accented and unaccented consonant-stem verbs as well as accented and unaccented vowel-stem verbs. While certain inflected forms, such as the simple negative kiranai are useful as a paradigm diagnostic, the single inflected forms kiru, kireba, kirareru, and kiremasu, by themselves, give no suggestion about their paradigm classification. This is exactly the problem that makes the stem identification of r stem verbs particularly problematic.



There is phonologically predictable allomorphy in the paradigms of stem-final k, g, and s verbs that complicates the identification of both verb stem and the past morpheme. The loss of stem-final k in the past, gerund, and related forms is accompanied by i epenthesis. For example, the past form of kik-u 'listen, ask' is ki-ita 'listened to, heard' and not *kik-ta or *kit-ta. The paradigms of stem-final g verbs show this same type of patterning. The past form of oyog-u 'swim' is oyo-ida 'swam' and not *oyog-ta or *oyot-ta or *oyog-da. A somewhat similar situation occurs in the paradigm of stem-final s verbs where the stem final s is not lost, but automatically palatalized preceding the epenthetic i in the gerund, past and related forms. The verb hanas-u 'speak' has the past form hansita 'spoke' and not *hanas-ta or *hanat-ta. A partial paradigm demonstrating gerund

and past form allomorphy is given in Table 10, where stem-final k, g, and s verbs are contrasted with stem-final r, t, and w verbs. Notice that the non-appearance of k in the gerund and past forms of kik-u and the non-appearance of g in the gerund and past forms of oyog-u contrast with the paradigms of other consonant-final verbs. Stem-final





Non-past Negative

kik-u kik-anai ki-ite ki-ita

oyog-u oyog-anai oyo-ide oyo-ida

hanas-u hanas-anai hanas-ite hanas-ta

kaer-u kaer-anai kaet-te kat-ta

Gerund Past


mat-u mat-anai mat-te mat-ta

ka-u kaw-anai kat-te kat-ta

TABLE 10. Allomorphy in the paradigms of stem-final k, g, and s verbs.

The gerund and past forms of stem-final r, t, and w verbs have word final sequences that look quite different. To any trained linguist this allomorphic variation is transparent and these forms are easily and economically segmented and arranged into neat paradigms. However, it remains to be seen whether speakers and learners of Japanese perceive and analyze these alternating forms in the same way. The results


from the longitudinal study of acquisition in Chapter 6 and the experimental nonceprobe tests in Chapter 7 suggest that they do not.

2.3.3. STEM-FINAL M, E, AND N VERBS. Modern spoken Japanese has only one verb with the stem-final consonant n. However, in terms of phonological similarity and morphological patterning, the verb

sin-u 'die' can be grouped with stem-final verbs like yorokob-u 'be happy' and yom-u 'read'. These verbs are of interest because they present yet another complication in the task of verb stem identification. The paradigms of these verbs demonstrate what has been called reciprocal conditioning (Wells 1949:109). The past form of sin-u is sin-da 'died' and not *sin-ta. The past form of yorokob-u is yorokon-da 'was happy' and not

*yorokob-ta or *yorokob-da. The past form of yom-u is yon-da 'read' and not *yom-ta or *yom-da. There is a potential three-way ambiguity in stem identification since the stem-final n, b, m distinction is neutralized in the past form. The verb yom-u has the past form yonda. The verb yob-u 'call for' also has the past form yon-da 'called for'. A contrastive partial paradigm of these three verbs is provided in Table 11. Stem-final n

Stem-final b

Stem-final m

Non-past sin-u yorokob-u yom-u Negative sin-anai yorokob-anai yom-anai Gerund sin-de yoroko-de yon-de Past sin-da yorokon-da yon-da TABLE 11. Allomorphy in the paradigms of stem-final n, b, and m verbs.

The verb yom-u has the past form yon-da. The verb yob-u 'call for' also has the past form yon-da 'called for'. The two past forms differ only in that yon-da 'called for' is unaccented, while yon-da 'read' carries a distinctive accent. The only occurring stem-final n verb in modern Japanese is unaccented: sin-u 'die'. However, there are no phonological constraints against a verb with the shape *yon-u. If there were an


actually occurring verb *yon-u, there would be a three-way ambiguity with the form yon-da. Recall from the earlier discussion that accent is lexical in Japanese, and since

there are both accented and unaccented stem final b and m verbs accent would not be helpful in the identification of the stem of an unknown verb. The potential three-way ambiguity of the past form yonda is illustrated in Table 12. Stem final m

Stem final n (non-occurring verb)

Stem final b

Non-past yom-u read yob-u call for Negative yom-dnai not read yob-anai not call for Past yon-da read yon-da called for TABLE 12. Potential three-way ambiguity of past form yon-da.

*yon-u *yon-anai *yon-da

2.3.4. STEM FINAL W VERBS. Due to the diachronic change w > 0, a modem Japanese verb stem that ends in Vw before the vowel a now ends in V elsewhere. This change has led to yet another allomorphic variation in the paradigm of verbs with the word-final non-past form Vu. The negative form of ka-u 'buy' is kaw-anai 'not buy' and not *ka-anai. The w also appears in the causative and passive forms of these verbs. A representative partial paradigm of ka-u is given in Table 13, along with corresponding forms for stem-final r verbs and stem-final e verbs. Since verbs like ka-u inflect with the same suffixes used by consonant-stem verbs, these verbs are classified as consonant-stem verbs, even though no stem-final consonant appears in most inflected forms of the paradigm. Stem-final

Consonant w

Vowel e

Consonant r

Non-past ka-u kaer-u tabe-ru Past kat-ta kaet-ta tabe-ta Polite non-past ka-imasu kaer-imasu tabe-masu Negative kaw-anai kaer-anai tabe-nai Causative kaw-aseru kaer-aseru tabe-saseru Passive kaw-areru kaer-areru tabe-rareru TABLE 13. Stem alternation in the paradigm of stem-final w verbs.


Note that unlike unpronounceable forms which violate the phonotactic constraints of Japanese, such as *kaer-ta for kaet-ta 'returned' and *oyog-ta for oyoi-da 'swam', forms without w in the paradigm of ka-u would be completely acceptable. While there is a preferred (C)V syllable structure in Japanese, forms like *kaanai, *kaaseru, *kaareru are completely pronounceable. While forms with w in the paradigm of Vu verbs contrast with forms in the paradigms of other consonant-stern verbs, they are clearly regular forms in the sense that they can be predicted by rule.


The processing of English verbal inflection has been characterized as the contrast between a productive rule for regular forms and the retrieval from memory of irregular forms (Marcus et al. 1992). This model of representation and processing accommodates the rather atypical inflectional typology of English where there is a single productive past rule and about 180 irregular verbs. The characterization of Japanese verbal inflection is quite different. As demonstrated in the first part of this chapter, Japanese contrasts a number of productive regular rules against two irregular verbs and a small number of irregular forms.


Although regular Japanese verbs do not show stern-vowel alternation, the two irregular verbs suru 'do' and kuru 'corne' show unpredictable stern vowel alternation across the paradigm. The stern vowel in the paradigm of su-ru alternates between u, and i: su-ru 'do', si-ta 'did'. The stern vowel in the paradigm of ku-ru alternates between u, i, and 0: ku-ru 'corne', ki-ta 'carne', ko-nai 'not corne'. A contrastive partial paradigm of a regular consonant-stern verb, a regular vowel-stern verb, and


these two irregular verbs is provided in Table 14. Notice that the pattern of stemvowel alternation in the paradigm of suru is different than the pattern of alternation in the paradigm of kuru and these two patterns cannot be conflated into a single irregular pattern. Form

Regular consonant-stem

Regular vowel-stem

Irregular stem

Irregular stem

hanas--u tabe-ru ku-ru su-ru Non-past Provisional hanas-eba tabe-reba ku-reba su-reba Past hanas-ita tabe-ta ki-ta si-ta Gerund hanas-ite tabe-te ki-te si-te Negative hanas-ana-i tabe-na-i ko-na-i si-na-i Volitional hanas-oo tabe-yoo ko-yoo si-yoo TABLE 14. Regular verbs and irregular verbs kuru and suru showing stem vowel alternation.

It is therefore not possible to say that there is a single pattern of Japanese

irregularity that contrasts with a single pattern of Japanese regularity. The irregular verb kuru makes use of the stem vowel 0 in the high frequency negative and volitional forms, while suru makes a more economical use of i in these forms. It is also of interest to note that the typology of Japanese irregularity is quite

different than the typology of English irregularity. English past tense irregularity either makes use of suppletion (go/went), zero-change (hit/hit), or a small number of semi-productive stem-vowel changes such as swim/swam, hold/held, run/ran, and

string/strung. Japanese on the other hand, employs the same morphological mechanism with irregulars as with regulars. There is a semantically empty stem-vowel alternation, but in addition, the two Japanese irregular verbs are inflected with the same suffixes as regular vowel-stem verbs. We have already noted that there are productive processes which derive vowel-stem forms (the passive, the causative, and the potential) from consonant-stem forms. Since suru 'do', also makes use of vowelstem suffixes, and is often informally recognized as the Japanese verb with the highest token frequency, one possibility to be considered is that the vowel-stem suffixes might


end up as the default paradigm for Japanese. In first language acquisition, for example, a child might make exclusive use of the vowel-stem suffix -yoo, rather than the consonant-stem suffix


to express the volitional form when the stem of a verb is

unknown or difficult to extract. This possibility is investigated in more detail in Chapters 6 and 7.


In addition to unpredictable stem-vowel changes in basic inflected forms, there are also irregular derived forms and one suppletive form in the paradigm of suru. In basic forms the stem vowel of su-ru is either u or i. Derived causative and potential forms show yet another alternation. The causative form of su-ru is s-ase-ru 'let or make do' and not *si-sase-ru or *su-sase-ru. The passive is s-are-ru 'is done' and not *su-areru or *si-are-ru. The causative passive makes use of both s-ase-rare-ru and the short

form s-as-are-ru 'is caused to do'. The potential form of su-ru is not the expected form *s-are-ru, but the suppletive form deki-ru 'able to do'. The possible potential forms *si-sase-ru and *su-sase-ru do not occur. A partial contrastive paradigm is provided in Table 15. Regular consonant-stem Non-past Provisional Past Negative Causative Potential Passive TABLE 15.

Regular vowel-stem

Irregular stem and suppletive form

nom-u tabe-ru su-ru do nom-eba tabe-reba su-reba if do non-de tabe-te si-ta did nom-anai tabe-nai si-nai not do nom-aseru tabe-sase-ru sase-ru make do, let do nom-eru tabe-(ra)re-ru deki-ru able to do nom-areru tabe-rare-ru sare-ru be done Irregular derived forms and suppletive forms in paradigm of the irregular verb suru.


2.4.3. IRREGULAR NAl FORMS In descriptions of Japanese verbal inflection, a distinction is made between irregular verbs and irregular forms. Irregular forms like na-i 'not have, not be' and nakatta 'didn't have, wasn't' in the paradigm of ar-u 'have, be' represent only a few of the possible forms in the paradigm and do not show the unpredictable stem vowel alternation of the irregular verbs. The irregular form nai is an exception to the negative rule for consonant-stem verbs: negative

= stem + -anai. The expected negative form of

ar-u is *ar-anai, which does not occur. The expected past-negative form *ar-anakatta also does not occur and is replaced by the irregular form na-katta 'didn't have, wasn't'. Partial contrastive paradigms of some consonant-final r verbs are given in Table 16. Non-past



wakar-ru understand wakar-anai not understand kaer-u return kaer-anai not return ar-u have, *ar-anai>nai not have TABLE 16. Regular stem-final r verbs and irregular forms

wakar-ana-katta kaer-ana-katta *ar-anakatta>nakatta of ar-u.

didn't understand didn't return didn't have

2.4.4. IRREGULAR FORM lITA. Much research attention has been devoted to study of the overregularization of English irregular forms. It has been noted that English speaking children appear to pass through a stage of linguistic development where the past form of go is sometimes overregularized to *goed or less frequently *wented. A very interesting coincidence occurs in Japanese. The Japanese verb ik-u 'go' also has an irregular past form: itta 'went' and not *i-ita, the form predicted by the productive rule for the past. If the English past rule is past =stem + ed, the Japanese past rule, for verbs with a stem-final Ik/, could be past

= stem minus k + ita.

A partial paradigm of some high frequency

stem-final k verbs is presented in Table 17.


Non-past kak-u kik-u tuk-u ik-u TABLE


write ka-ita wrote ask, listen ki-ita asked, listened be attached tu-ita was attached go *i-ita > itta went 17. Partial paradigms of stem-final k verbs with irregular form itta.

The past form of ik-u predicted by this past rule should therefore be ik minus k > i

+ ita> *i-ita. This form (*i-ita) does not actually occur in the paradigm of iku. The fact that irregular forms tend to be regularized by children in English (for example *goed for went) has led to models of verbal processing that depend on the segmentation of stems and endings in the acquisition process. Also implicit in these rule-based models of processing is the assumption that regularly inflected verbal forms will be less problematic to learners than irregular forms. Whether the same type of overregularization actually occurs in Japanese seems not to have been addressed in the relevant literature. Therefore, this preliminary description of irregular Japanese forms will be followed by a detailed attempt to find evidence of overregularization in a Japanese database in Chapter 6.


There is an interesting and well-known pattern of irregularity that occurs only in the five polite-honorific verbs that are used to refer to the action of out-group superiors. The five polite verbs are irasshar-u 'be, go, come', osshar-u 'say', nasar-u 'do', kudasar-u 'give' and, gozar-u 'be, have'. The particular form gozar-u does not occur in colloquial modem Japanese, even though gozar-u is the citation form in both Japanese and Japanese-English dictionaries. In modem spoken Japanese the irregular form gozaimasu occurs in place of gozar-u. Only these five honorific verbs show an unpredictable loss of stem-final r is some inflected forms. The irregularities in the


paradigms of these five honorific verbs are of interest because the pattern of irregularity is only semi-productive, operating only within the paradigms of these five verbs. Furthermore, the loss of stem final r is not phonologically or morphologically predictable. A partial paradigm contrasting these verbs with the regular forms of the stem-final r verb nor-u 'ride' is given in Table 18. There are also alternating irregular gerund forms nasut-te and irashite. Not shown in the table are the alternating past forms nasat-ta/nasut-ta 'did' and irassyat-ta/irasita 'was, came, went'. Since the honorific-polite forms represent a formal style of speech, they are not typically used by young children in first language acquisition, however there are a few set expressions which are derived from these forms and it is of interest to see if children overregularize these set expressions. For example, the "Set phrase oyasumi nasai 'good night' could be overregularized as oyasumi *nasari. The ritual greeting *ohayoo gozaimasu could be overregularized to ohayoo *gozarimasu. Regular stem-final r verb Non-past nor-u


Regular gerund not-te

Regular polite non-past nor-imasu

Irregular stem-final r polite verbs with loss of stem final r Non-past Regular gerund Irregular polite non-past gozar-u have, be gozat-te gozaimasu nasar-u do nasat-te/nasutte nasaimasu irasshar-u be, go, come irassyat-te/irasite irassyaimasu ossyar-u say ossyat-te ossyaimasu kudasar-u give kudasat-te kudasaimasu TABLE 18. Irregular forms in the paradigms of the five honorific verbs.

Non-occurring forms *gozar-imasu *nasar-imasu *irasyar-imasu *ossyar-imasu *kudasar-imasu

Likewise, the expression arigatoo gozaimasu 'thank you' could be overregularized to arigatoo *gozarimasu. Whether such rule-based overregularizations actually occur in first language acquisition is investigated in some detail in Chapter 6.



In addition to free verbal nouns, there are also bound verbal nouns, which have alternative irregular forms. One well-known example is the verb kanji-ru 'feel'. This verb and others like it of Chinese origin, has some alternate inflected forms, based on non-colloquial literary forms. The alternate non-past form of kanji-ru is kanzuru. The past form of kanji-ru is the expected kanji-ta 'felt', but there are alternate forms for the negative and the passive. The regular negative form kanzi-nai 'not feel' contrasts with the alternate forms kanzezu and kanzizu. The contrasting alternate forms are shown in Table 19. Form

Regular vowel-stem

Irregular alternate form

kanzuru Non-past kanzi-ru Negative kanzi-nai kanzezu/kanzizu Causative kanzi-sase-ru kanzerareru Passive kanzi-rare-ru Honoroific okanzi ni naru Table 19. Regular and alternating irregular forms of kanji-ru.

Martin (1988:877-878) provides a table of 43 different bound verbal nouns of this type with native-speaker acceptability judgements. Based on Martin's table, there is much variation in the acceptability of various inflected forms, even the regular forms. The alternate form hoozuru of the verbal noun hooji-ru 'report' is judged to be acceptable with reservation. Both the regular negative form hooji-nai and the alternative form hoozezu for the intended meaning 'not report' are judged to be of questionable acceptability. The regular causative hoozi-saseru 'make report', the regular passive hoozi-rareru 'be reported', and the regular honorific o-hoozi ni naru are all judged to be unacceptable. It seems strange that both the expected regular negative and passive forms would be unacceptable to native informants polled. Even Martin himself does report one newspaper example of the regular passive use, which his informants would reject as unacceptable.


It appears that many of the questionable and rejected inflected forms of hooji-ru, as reported in Martin (1988), are now acceptable to at least some speakers and that a language change may be in progress. The results of a Google search are reported in Table 20. These bound verbal noun forms as well as the free verbal noun forms in the previous section appear to be highly lexicalized. Their use and acceptability varies from word to word and from person to person. Form

Regular forms Tokens

Irregular forms


Non-past report hoozi-ru 9,890 hoozuru 2,120 37/86 hoozezu/hoozizu Negative not report hoozi-nai 888 hoozi-sase-ru 19 Causative make report Passive be reported hoozi-rare-ru 6,320 hoozerareru 127 Passive stative-affirmative is reported hooji-rare-te iru 9,140 Passive stative-negative is not reported hooji-rare-te inai 426 Honoroific non-past other reports o-hoozi ni naru 0 Honorific past other reported o-hooji ni natta 1 TABLE 20. Tokens of regular hooji-ru and irregular hoozuru recovered from internet using Google.

2.4.7. IRREGULAR FREE VERBAL-NOUNS. There are alternate forms of some Japanese verbal nouns, which are derived from the combination of a noun and the irregular verb suru. For example, the combination of ai 'love' and suru 'do' yields the verbal-noun ai (0) suru 'to love'. Since the accusative particle 0 is allowed to occur between ai and suru, Martin (1988: 870) calls these verbs free derived-verbs. In some cases, the combination of noun and irregular

suru is contracted to form a new consonant-stem verb. For some speakers (Martin 1988:301) verbal nouns such as ai suru 'love' and naku suru 'lose' have alternate forms, which conjugate like regular stem-final consonant verbs. The alternate form of

ai suru is ais-u and the alternate form of naku suru is nakus-u. A partial paradigm contrasting the alternating forms of ai suru and ais-u is given in Table 21.


Non-past Negative Past

Conjugation type




Irregular suru ai suru ai sinai ai sita ai sase-ru ai sare-ru ai sare-ru Regular consonant-final ais-u ais-anai aisita ais-ase-ru ais-eru ais-are-ru TABLE 21. Contrasting forms of irregular verbal nouns and contracted regular forms.

Martin, writing in 1988 (302) also points out that suppletive *ai dekiru is not used as the potential of verbs like ai suru. Rather, as shown in Table 21, the older potential form sare-ru appears to be preferred. Recall from the discussion in section 2.4.2 that the expected suppletive potential of suru is dekiru. However, an Internet search did recover 114 tokens of different forms of ai dekiru from various Japanese sources. It is not clear what kind of language Internet examples represent. It has been informally suggested that Internet communication includes language samples that may be jargonladen, both more and less formal than conversational speech, and perhaps often atypical in terms of linguistic innovation. It may be that the dekiru forms recovered here do not represent what native speakers would actually say. However, some ai deki-ru forms were recovered. It may be that at least for some speakers, there is a language change in progress since Martin reported, nearly 15 years ago, that ai dekiru forms were not used. The results of the internet search using Google are given in Table 22. Form


Number of tokens

60 Non-past ai deki-ru able to love Past ai deki-ta was able to love 16 Negative ai deki-nai not able to love 14 Gerund ai deki-te able to love and 10 Conditional ai deki-reba if able to love 7 2 Provisional ai deki-tara if/when able to love Informal-negative ai deki-zu not able to love 2 Past-negative ai deki-na-katta was not able to love 2 Representation ai deki-tari able to love and 1 Total 114 Table 22. Tokens of ai deki-ru forms recovered from the Internet using Google search.


It is clear that there is at least a two-way, and perhaps a three-way, alternation

between ai sare-ru, ais-eru, and ai deki-ru to express the meaning 'able to love'. Table 23 provides the results of another search for tokens of these forms. Note that the number of tokens reported for ai sare-ru includes both potential 'able to love' and passive 'be loved' interpretations. The point of this discussion is to demonstrate how three alternate forms of the potential of ai suru might represent irregularity within the inflectional paradigms of Japanese verbal morphology. The three different alternate forms represent three possible paradigms. If speakers use alternate forms, or mix paradigms, for example if a speaker uses the verbal noun form ai suru 'love', but also makes use of the contracted form ais-anai 'not love', then it is no longer possible to predict which other forms that speaker will use. And, if the form that a speaker will use is not predictable, then it is no longer regular. Form


Number of tokens recovered

able to love/be loved 121,000 Potential/passive ai sare-ru Potential ais-eru able to love 26,000 Potential ai deki-ru able to love 60 Table 23. Tokens of potential forms of ai suru and ais-u recovered from internet using Google search.

The significant conclusion for the present discussion is that while the core of Japanese verbal inflection can be characterized as regular, there is a wide periphery of alternating forms that also need to be considered as irregular. In fact, a typical Japanese speaker probably makes use of more irregular forms than a typical English speaker. There are only about 180 irregular verbs in English. In many cases that irregularity represents a pair of forms, run and ran, for example. There are also some verbs where there is a robust three-item paradigm (go/went/gone). However, there are many more inflected forms in a Japanese paradigm, a total of about twenty or so. Therefore, the two irregular verbs suru and kuru and the suppletive paradigm of dekiru alone represent about sixty irregular forms. Furthermore, while there are clearly many


high frequency English irregulars (eat/ate, drink/drank, sleep/slept), many of the 180 verbs on the irregular list have a much lower frequency of use. The fact that some irregulars have regularized alternate forms (weeped/wept) is usually interpreted as evidence of relatively low frequency of use. Japanese irregulars, on the other hand (kuru, suru, deki-ru, nai, nakatta, itta and others) tend to be forms with high token frequency. Therefore, the conventional assumption that compared to English, Japanese has few irregular verbs, needs to be reconsidered. There are few irregular verbs, but a considerable number of irregular forms.


A brief description of the native Japanese analysis is included here for two reasons. First, unlike the previously described linguistic analysis, which requires segmentation into unpronounceable consonant-final stems: kaer- 'return', mat- 'wait', hanas- 'speak', kak- 'write', the native analysis respects phonotactic constraints in order to achieve the same general categorization into two regular verb categories. Consonant-stem verbs of the linguistic analysis are equivalent to native go-dan 'five row' verbs and vowel-stem verbs are equivalent to native ichi-dan 'one-row' verbs. This native analysis makes use of an idiosyncratic segmentation. The result of this segmentation is an analysis with much variation in the shapes of both stems and suffixes. This analysis contrasts sharply with the invariant stem analysis of the linguistic analysis. Since the focus of the present research is on the representation and processing of inflection, rather than economical description, the possibility that the native analysis might be closer to what speakers actually do is considered. The second reason for consideration of the details of the native analysis is that there may be instructional or literacy effects for adult or literate speakers of Japanese.


The native analysis presented here is the analysis presented to all Japanese school children, and the only analysis that can be presented in the native orthography, Miller (1967:314) attempts to attribute the peculiarities of the native analysis to the native syllabic orthography. The traditional Japanese approach has limited itself to statements about syllables which can be written in the kana writing system, which allows no forms with syllable-final consonants other than the [mora] nasal. Hence it has been unable to arrive at a scientifically acceptable description of the language, which would require statements involving forms with final consonants.

Vance (1987) has a detailed description and evaluation of the native system, based on the model provided in a study aid (Nichieisha: 1952) intended for native students studying their own language. In his elaborated description and evaluation Vance (1987:181) makes the following statement, which includes a response to Miller. In the case of study aids, of course, it would probably just confuse students to talk in terms of components which cannot be written in the native orthography. But what we have here is not simply a pedagogically useful description that deviates from a "correct" or "scientific" description. The kana syllabary cannot represent syllables with final consonants other than the mora nasal and the mora obstruent (which Miller forgets), because such syllables violate the phonotactic pattern of Japanese. In other words, since no word can contain such syllables, there is no need to have letters for them.

The most compelling reason to consider the native analysis is for the reason given in the quote. The native analysis consists of segmentation into stems and suffixes that can be pronounced by native speakers. The linguistic analysis does not. Unger (1993:59) was concerned with Old Japanese, but has a comment relevant to processing as well.


Actually this system, developed independently of Western linguistic tradition, is remarkably efficient and ingenious. Its only drawback is that it treats the syllable rather than the phoneme as the basic unit of analysis (Gardner 1950:8ff, Yokoyama 1950:2ff, Miller 1967a: 319ff). This obscures the immediate constituent analysis since cuts between consonants and vowels are not readily possible.

Since the native analysis depends on and is reflected in the standard orthographic system, one might wonder if the instructional effect of literacy might be to reinforce processing strategies that depend on the association of meaning with completely inflected words in memory. It actually might be more economical (in terms of processing) for a Japanese speaker to store whole words, rather than stems and suffixes.

2.5.1. THE SIX-STEM SYSTEM. The native Japanese analysis was developed to deal with the verb forms of classical Japanese and recognizes six separate stems for each verb. As in the linguistic analysis, two distinct categories of regular verbal patterning are distinguished, but the basis of the distinction is quite different. In particular, the segmentation of stems is very different. The native analysis makes reference to the arrangement of items in the kana syllabary. A romanized version of the syllabary in its traditional right to left order appears in Table 24. N


ra ri ru re





ma mi mu me ~

ta ti tu te

sa si su se




na ni nu ne

ki ku ke

u e







ha hi


TABLE 24. Japanese syllabary showing arrangement of gyoo 'columns and dan 'rows'.


The syllabary is traditionally arranged into dan 'rows' and gyoo 'columns'. Depending upon the number of varying stern forms, verbs are classified as either ichidan 'one-row' or go-dan 'five-row' verbs.

Verbs like yom-u 'read' show stern-final vowel alternation across five of the six classical stern forms. The five alternating stern forms are complex sterns, each consisting of a root and a suffix. The five relevant stern-final syllables are in bold type in the preceding table. The alternating stern forms of yom-u are yo-ma-, yo-mi-, yo-mu, yo-me-, and yo-mo. These complex stern forms, along with the suffixes with which

they can occur are listed in Table 25. Because the final syllables of the complex sterns ma, mi, mu, me, and mo represent all five rows of the ma column of syllabary, verbs

like nom-u are called go-dan 'five-row' verbs. Table 25 is a translated and expanded version of a chart found at a web site posted by Yano (2003) and intended for reference use by native speakers, Japanese teachers, and advanced learners of Japanese. Table 25 illustrates the native analysis of a go-dan (five row) verb which is equivalent to a consonant-stern verb in a linguistic analysis. The native analysis of ichi-dan (one-row) verbs is based on the same notion of stemvowel alternation. Verb

Native term


Representative suffix

gokon yoRoot Indefinite stem 1 mizenkei yo-ma-nai/-seruJ-reru passive mizenkei yo-mo--o Indefinite stem 2 Conjunctional stem 1 renyookei yo-mi-masu Conjunctional stem 2 renyookei yo-n-da Final stem shusikei yo-mu -nal-mai Attributive stem rentaikei yo-mu (2) Hypothetical stem kateikei yo-me-ba Imperative stem meireikei yo-me (2) TABLE 25. Native segmentation with complex stems of go-dan 'five-row' verb yomu 'read'.


Only a single vowel is found in the six complex classical stems of verbs like faberu 'eat'. The alternating stem forms are segmented as fa-be-, fa-beru (not fa-beru), fa-be-re-, and fa-bero (not fa-be-ro). In this case, the native analysis looks at the

invariant vowel e of the repeating syllable be (written using one character from the kana syllabary) as representing only one row in the column under a in the syllabary.

For this reason, verbs like fabe-ru are called ichi-dan 'one-row' verbs. Table 26 reproduces the segmentation provided by Yano. Verb

Native term


Representative suffix

gokon taRoot Indefinite stem mizenkei ta-be-nai/-saseru/-yoo/-mai/-rareru potential/passive Conjunctional stem renyookei ta-be-ta/-masu Final stem shusikei ta-beru -na Attributive stem rentaikei ta-beru @ Hypothetical stem kateikei ta-bere- -ba Imperative stem meireikei ta-bero @ TABLE 26. Native segmentation with complex stems of ichi-dan 'one-row' verb taberu 'eat'.

The result of this analysis is that go-dan (five row) verbs in the school grammar are equivalent to consonant-stem verbs in the linguistic analysis and ichi-dan (one row) verbs in the school grammar are equivalent to vowel-stem verbs in the linguistic analysis. However, the linguistic analysis is clearly much more economical in terms of stem representation with only one stem versus the five distinct (complex) stems of the school grammar. As a result of the stem segmentation in Table 25 and 26, the suffixes of the school grammar sometimes have consistent shapes, but sometimes vary. The expectation that a native analysis that respected phonotactic constraints might also be able to eliminate some of the suffix allomorphy of the linguistic analysis cannot be fully realized. Some economy is gained in the native analysis where the negative, conditional, polite, and desiderative make use of invariant endings; however, elsewhere there are alternating forms. The native analysis, for example, makes use of the single negative suffix -nai 34

while the linguistic analysis requires -nai for vowel-stem verbs and -anai for consonant-stem verbs. Some representative suffixes from both analyses are compared in Table 27. Native analysis

Linguistic analysis

go-dan ichi-dan Consonant-stem Vowel-stem Suffix Negative -nai -nai -anai -nai Conditional -ba -ba -eba -ba Polite -masu -masu -imasu -masu Desiderative -fai -fai -itai -tai Past -to/-do/-ida -ta -to/-do/-ida -ta Volitional -0 -yoo -00 -yoo Causative -saseru -seru -aseru -saseru Passive -reru -rareru -areru -rareru TABLE 27. Contrasting suffixes resulting from native and linguistic analysis.

The real issue to be addressed here is whether morphemes are mentally represented in the same way that they are formally described. At best, it seems doubtful that economy in the number of forms in a descriptive analysis is the single most important consideration in the representation and processing of inflection. It seems more likely that the cognitive ability to detect and form an association between a linguistic form and meaning would be more cruciaL Yet, the relevance of formal description to mental representation may be informed by careful examination of the formal description. The formal description presented here is the one used in the Japanese school system and while there may be possible instructional effects for older children and adults, we will suggest that the limitations on representation reflect typological constraints that should be apply to first language learners as well. We consider next an example from a typical Japanese school grammar that suggests why native speakers might associate grammatical meaning with whole words rather than morphemes. The native analysis segments some verbs so that they are described as having complex stems, but no roots.



In Table 9 of section the traditional linguistic analysis of the vowel-stem verb ki-ru 'wear' and the consonant-stem verb kir-u 'cut' was presented along with contrasting and overlapping forms. These two verbs are segmented very differently in the native analysis. There is the peculiar result that the monosyllabic stems of vowelstem verbs in the linguistic analysis as well as the two irregular verbs kuru and suru end up with no root at all in the typical school grammar. This relatively minor peculiarity in the formal grammar is included here because it suggests how Japanese speakers might map meaning to fully inflected forms rather than to morphemes. The following tables give expanded examples from Yano (2003) that have been romanized for use here. Basic form

Root Indefinite



kiru wear 0 kikikiru suru do 0 si-, sa-, se- sisuru kuru come 0 kokikuru TABLE 28a. Japanese school grammar ichi-dan (vowel-stem)

Attributive Hypothetical Imperative kiru kirekiro/kiyo suru kiresiro, seyo kuru kurekoi verbs analyzed as no-root verbs.

Basic form Root Indefinite Continuative Final Attributive Hypothetical -e-e-eru -eru -erekaeru change kaireru put in i-re-re-reru -reru -rereokiru get up 0-ki-ki-kiru -kiru -kireTABLE 28b. Japanese school grammar of other ichi-dan (vowel-stem) verbs analyzed to

Imperative -ero/-eyo -rero/-reyo -kiro/-kiyo show a root.

Basic form Root Indefinite Continuative Final Attributive Hypothetical kiru cut ki-ra-/-ro-ri-/Q -ru -ru -renomu drink no- -ma-/-mo-mi-/-n-mu -mu -meTABLE 28c. Japanese school grammar go-dan (consonant-stem) verbs analyzed to show

Imperative -re -me a root.

Basic form Root Indefinite Continuative Final Attributive Hypothetical Imperative kiru wear [email protected] @ -ru -ru -re-ro/-yo Inflected form [email protected] [email protected] ki-ru ki-ru ki-re-ba ki-ro/ki-yo TABLE 28d. Proposed alternate segmentation of ichi-dan (vowel-stem) verbs analyzed to show a root.


Notice that in the native analysis of kiru 'wear' in Table 28a, every suffix begins with the syllable ki and consequently nothing remains to be designated as the root. This analysis contrasts with the segmentation of other ichi-dan verbs like kae-ru 'change', ire-ru 'put in' and oki-ru 'get up' where a root is analyzed out. The go-dan verbs are also segmented to show a root. The root ki is segmented out of the go-dan verb kir-u 'cut'. This analysis results in a general description of verbal inflection that is complex, asymmetrical, and apparently linguistically unmotivated. One wants to ask why the verbs in Table 28a are analyzed as having no roots. Vance (1988:181) makes the following suggestion. The only reason I can imagine for treating one-row verbs in this fashion is to avoid having to say that the indefinite and conjunctional stems have zero suffixes.

In other words, there is a choice between one asymmetrical analysis where some verbs have zero-roots and another asymmetrical analysis where some verbs have zeroindefinite and zero-continuative stems and some verbs do not. These two choices are illustrated in Tables 28e and Table 28f, where sample inflected forms for each complex stem have also been included. Basic form

Root Indefinite

Continuative Final


Hypothetical Imperative

kiru wear 0 kikikiru kiru kirekiro/kiyo Inflected form ki-nai kikiru kiru kire-ba kira/kiya TABLE 28e. Japanese school grammar ichi-dan (vowel-stem) verbs analyzed as no-root verbs.

Basic form Root Indefinite Continuative Final Attributive Hypothetical Imperative kiru wear ki0 0 -ru -ru -re-ro/-yo Inflected form ki-(i)-nai ki-(i) ki-ru ki-ru ki-re-ba ki-ro/ki-yo TABLE 28f. Proposed alternate segmentation of ichi-dan (vowel-stem) verbs analyzed to show a root.

Economy and symmetry are considerations in the linguistic analysis of morphology. For the native Japanese analysis, there are phonotactic constraints, which limit how much economy can be achieved. The native Japanese analysis also appears to be driven by meaning. The native grammar can afford to reject the root analysis of 37

short verbs because the bare roots it ends up with tend not to be very meaningful. Unlike the bare stems of English, which are free morphemes and can have a tensed interpretation: walk, eat, run, Japanese roots are bound morphemes. Even in the linguistic analysis, Japanese verbal roots ki-, kir-, and tabe- have no tense or aspect interpretation. The native analysis produces even less meaningful root forms. Some roots are not even distinctive. For example, on the native analysis, yo- is the root of both yom-u 'read' and yob-u 'call'. It is only in the complex stems that forms such as yo-ma- and yo-ba- emerge as distinctive. The native analysis cannot segment out yob-

and yom- and we have argued that native speakers probably cannot mentally represent these forms either. In the native analysis, the meaning of a verb emerges along with the shape of the complex stem. In general, the complex stem forms are much more meaningful than bare roots and much closer to being fully inflected words. Furthermore, as we have already shown, the native analysis of the short ichi-dan verbs actually does consider forms like kuru 'come' suru 'do' and kiru 'wear' to be basic rather than inflected forms. One way to evaluate this analysis is to see if it is consistent with speech errors. The formal linguistic analysis is decompositional and contrasts morphemes in the words su-ru and ku-ru against the native lexical analysis with suru and kuru. If Japanese

speakers can and do manipulate morphemes, we might expect to find speech errors such as *siru and *koru and *kiru. This is a question that is investigated in some detail in Chapter 6.



3.1. INTRODUCTION: The purpose of this chapter is to present a brief overview of descriptive models of morphology. These models make use of familiar mechanisms, rules, paradigms, and analogy. In Chapters 4 and 5, additional processing mechanisms and processing models are introduced. However, since the basic mechanisms are common to both descriptive and process models, we begin here. Our ultimate purpose is to consider whether the rules and processes presented in the various models are plausible in terms of the way that speakers actually process language. At the end of Chapter 6, consideration is also given to the specific predictions that models of processing make about inflectional productivity, the acquisition of inflection, and inflectional error.

3.2. MODELS AND THEORIES. Some of the characteristic details of Japanese inflectional morphology were described in Chapter 2. The descriptions presented were based on and included different conventions and assumptions about the possible ways language can be structured. It was demonstrated, for example, that both the native and the linguistic analysis assume that it is possible to segment fully inflected verbs into smaller meaningful parts. We also showed how the native analysis allowed some short verbs and the irregular verbs suru and kuru to have non-past forms, which were not segmented into morphemes. The linguistic and the native description were compared and contrasted and it was pointed out that both had limitations as possible models of representation and processing. In the sections that follow we want to present a more


specific definition of the term model and mention some of the criteria that have been used in the literature that have been used to evaluate morphological models. Models of language are often associated with the more fully developed notion of a theory of language. Martin (1995:45) contrasts these two terms and suggests one way in which they are different. The choice of the term reflects the view adopted toward language: language seen as a formal system has a theory, while language as a means of human interaction is presented in the form of models. The word model is used in this study to denote all those presentations which explicitly aim at some level of generality...

Good models, just like good theories, should also be falsifiable. There should be a way to disconfirm them. This is a widely accepted evaluation measure and consistent with what one finds in the literature. The following definition of a theory and description of a model are from Bates and MacWhinney (1987: 174). A theory is a set of inter-related hypotheses that can be directly tested and rejected by some line of evidence. -

A model has much less internal coherence, insofar as it reflects an open-ended or

'bottom-up' attempt in the strict sense; it can only be confirmed or disconfirmed in pieces. When a model undergoes too many ex post facto repair attempts, it finally becomes a patchwork of assumptions that has no architectural center - the model collapses.

In a classic 1954 paper Hockett uses the term "model of grammatical description" to outline the models of morphology that we consider next.

3.2.1. HOCKETT'S 1954 MODELS.

Hockett's paper mentions three general descriptive models: word and process (WP), item and arrangement (IA), and word and paradigm (WP). The defining


characteristics of each model represent a "frame of reference" and an approach to the analysis of morphology that are relevant to both descriptive as well as process models. For example, one of the most current process models of morphology, Pinker's (1999) Words and Rules model combines elements of what Hockett described as the WP model and the Item and Process IP model. Actually, Hockett mentions the WP model only briefly on the first page of his paper, even though the suggestion is that he may have favored the WP model. Most of the discussion in the original paper is concerned with a comparison of the IP and the IA models. Since there has been ample discussion of these three models in the literature, the brief presentation here is mainly concerned with their relevance to Japanese. ITEM AND PROCESS (IP).

The defining characteristic of an IP model is the establishment of a single underlying form from which all other allomorphic forms are derived by the application of feature-changing rules. Various forms of IP models, which include both structuralist versions and those within the field of generative phonology, are some of the most long-lived and widely applied morphological model types of the last 100 years. McCawley's 1968 generative treatment of Japanese inflection remains the best-known example of an IP analysis of Japanese. A typical derivation is illustrated in Table 29. 'returned' Underlying form Ilkaer + ta/I r -> obs lobs kaed + ta voicing assimilation kaet + ta Table 29. Derivation of stem allomorph in IP model of Japanese inflection.

The stem allomorphy described in Chapter 2 of this dissertation, where the non-past, consonant-stem verb kaer-u 'return' has the allomorphic past form kaet-ta 'returned' is treated in the following way. The form kaer is chosen as the underlying stem form and two 41

rules are required to convert the stem-final Irl to its surface realization as It!. The obstruent Irl is first converted to Idl and then a voicing assimilation rule applies.

An even longer abstract sequence of processes or rules is required to account for the past form of a verb like oyo-ida 'swam'. The steps proposed by McCawley are summarized in Table 30. 'swam' Rule Underlying form Iloyog + tall 1 Voicing assimilation oyog + da 2 Sprirantization oyoh + da 3 Epenthesis oyoh + ida oyow + ida 4 h -> W 5 w -> 0 oyo + ida Table 30. Derivation of stem and past allomorph in IP model of Japanese inflection.

Within the IP framework of analysis, more economical models, those with fewer deep structure forms and fewer rules, are valued over models with more forms and more rules. In terms of the number of forms employed, the native analysis presented in Chapter 2 with its six complex stems and overlapping forms is not nearly as economical as the linguistic model with its single underlying stem. ITEM AND ARRANGEMENT (IA). The native analysis of verbal inflection as well as the Bloch's linguistic analysis, as presented in Chapter 2, are essentially IA models of morphology. In an IA model there are no processes, only items, which represent morphemes, and their arrangements with respect to one another. There are no abstract underlying or intermediate forms, since there are no rules that apply sequentially. The question of allomorphy is addressed by allowing all allomorphic alternations to have equal status. No one form is considered to be more basic than any other form. In the native analysis, we saw how the indefinite and continuative stems of consonant-stem verbs were


allowed two different shapes. The alternate stems of the native analysis of kaer-u 'return' are given in Table 31. Verb

Native stem designation


Representative suffix

gokon kaRoot Indefinite stem 1 mizenkei ka-era-nai negative Indefinite stem 2 mizenkei ka-ero-0 volitional ka-eri-masu polite Conjunctional stem 1 renyookei Conjunctional stem 2 renyookei ka-et-ta past TABLE 31. IP native analysis of go-dan (consonant-stem) verb with allomorphic stem variants.

Notice that while the IA model does not necessarily make use of formal rules, it minimally requires either a prose statement or some type of tabular representation to show which stems combine with which suffixes and in which order. In Table 31, for example, there has to be some indication that the indefinite stem 1 combines with the suffix -nai and not with the suffix -0. Within a formal IA framework of analysis, economy in the number of morphemes is not a concern, although it is considered important to limit the number of types of morphemes. A strict IA analysis of Japanese would not be concerned with the fact that there are two possible stem shapes kaer- and kaet-, because both of these allomorphs are represented in inflected forms such as kaeranai 'not return' and kaetta 'returned'. A strict IA analysis would be concerned with the fact that the native Japanese analysis requires six different stem morphemes, two of which have identical shapes. Both the WP model and the IA model are formal models of description that are concerned with economy. However, as we have previously pointed out, the less economical native model divides inflected forms into constituents, which can actually be pronounced by native speakers. There are no abstract underlying forms.


The WP analysis dates back to classical descriptions of ancient Latin, Greek, and Sanskrit. It operates on the assumption that some generalizations can only be made on the basis of the whole word. Description consists of whole words organized into paradigms. Matthews (1991 :204) is an important reference. In the ancient model the primary insight is not that words can be split into roots and formatives, but that they can be located in paradigms. They are not wholes composed of simple parts, but are themselves the parts within a complex whole.

Formal descriptions of Japanese tend not to be formulated as WP because the agglutinating typology of Japanese usually makes it possible to do direct one-to-one mappings between morphemes and meanings. However, in the case of what has been called extended or overlapping exponence (Spencer 1991:51), where a one-to-one mapping cannot be achieved, a WP analysis is a powerful alternative. For example, In English the past participle is associated with both the suffix -en and the stem vowel change in the form written. It is not possible to uniquely assign the past participle meaning to a single morpheme. Both the vowel change in the stem and the inflectional ending en must be associated with the past participle. WRITE




A WP analysis obviates this problem by saying that fully inflected word written represents the past participle form in the paradigm of write. The native Japanese analysis of yomeru 'can read', entails a related problem of analysis. In the Japanese analysis it is not possible to isolate a single morpheme that can be associated with the potential inflection. In a linguistic analysis, yom-eru is 44

derived from the consonant-stem verb yom-u by addition of the potential morpheme e. The potential sense 'can' is directly mapped to the single vowel e. The form yom-u 'read' contrasts with the derived potential form yom-e-ru 'can read'. The derived stem yom-e- is conjugated as a vowel stem verb, just like tabe-ru 'eat'. The non-past suffix

is -ru for all vowel-stem verbs. In the formal native analysis of yomeru 'can read', the segmentation of the different complex stem forms are as follows: yo-me, yo-mere-, and yo-meru. However, in the native analysis, there is no way to separate e from the

consonant m in the syllable me. (The form is written using the three kana symbols yo, me, ru and there is no symbol that represents malone.) The formal native analysis thus

ends up attributing the potential meaning to the whole word yomeru in each of its inflected forms. Assuming that the same phonotactic constraints (syllables and morae rather than consonants and vowels) apply in the actual processing of inflection, it would not be surprising to find that native speakers also process certain inflected forms both in terms of words and in terms of paradigms. Analogy is the central process implicit in a WP model. Depending on the typology of the language being taught, WP paradigms that illustrate exemplary patterns have often been employed in second language instruction. Analogical extension from an illustrative paradigm to new lexical items need not involve making explicit the complexities of morphemic analysis. The direct mapping of meaning to a completely inflected word illustrates a kind of simplicity that may be more important to second language learners than the "economy" of IA and IP models. Pedagogical grammars intended for second language learners tend to be combinations of IP and WP. The widely used Jorden (1987, 1988, 1990) textbook series, intended for foreign language instruction of Japanese at the university level,


makes use of a basic IP approach for most inflected forms. In volume three, students are directed to use the following rule (Jorden, 1990:91) to form the provisional. For verbals, we add -reba to the root of vowel verbals (= imperfective minus -ru) and -eba to the root of consonant verbals (= imperfective minus -u).

However, in an earlier volume of the same series, the instruction given for creating the past form is quite different. The text explicitly and implicitly suggests that students learn the inflected past forms of representative verbs and then create the past of newly learned forms by analogy. Students are provided with an explicit general rule followed by advice about how to avoid complications with consonant-stem verbs by making use of analogy rather than the rule. The term analogy is not used, but the intent of the recommendation is clear. The following quote is from volume two in the Jorden (1987:239) series. The sentence in capitalized form is as it appears in the original text. The pattern for the perfective is Iverbal root + perfective ending -tal ... However, when we add -ta to consonant-verbal roots (which always end in consonants), except for mat-ta, the consonant combinations that result are combinations that do not occur in Japanese. Adjustments are made on basis of the particular consonant at the end of the root: THE SAME ADJUSTMENT IS MADE FOR ALL ROOTS ENDING WITH THAT CONSONANT. This means that as long as we know the perfective of one sample verbal for each consonant, we can handle the forms of all verbals in the language from now on!

The relevant portion of the table that follows this encouragement to make use of analogic extension is reproduced in Table 32. The table provides the details of the troublesome starred forms that result when the combinatory rule approach is followed for consonant-stem verbs.


Imperfective Perfective

mat-u mat-ta wakar-u *wakar-ta > wakat-ta *kaw-ta > kat-ta w ka(w)a-u *-mas-ta > -masi-ta s -masu-u k kak-u *kak-ta > kai-ta m nom-u *nom-ta > non-da b yob-u *yob-ta > yon-da TABLE 32. Summary of Part 1 Jorden (1987:240) chart. r

The last volume in the Jorden textbook series concludes with a more subtle encouragement to use the word and paradigm approach. In Appendix A (1990:220221) there is an extended two-page chart, showing 14 different inflected forms with a representative sample paradigm of each possible consonant-stern and vowel-stern type and also the two irregular verbs kuru and suru. The inflected forms are presented as completely inflected words, without showing any segmentation. The implication is that based on the 280 forms listed in the chart, all other forms of all other regular Japanese verbs could be produced by the single mechanism of analogic extension.

3.3. CONCLUSION: HOCKETT'S 5 EVALUATION CRITERIA. Hockett concludes his paper (1954:398) with a list of 5 criteria that can be used to evaluate models of grammatical description. First, he suggests that a model must be general: the model should apply to all types of languages. Second, he suggests that a model should be specific: the model should be so highly specified in its procedures that it does not allow arbitrary analyst choice. Third, he suggests that a model should be inclusive: the model should be able to account for most of the known and expected data. Fourth, he suggests that the model should be productive: the mechanisms of the model should be able to produce an "indefinite number of valid new utterances". Finally, and this is the criterion of which he is least certain, he suggests that a model should be efficient: the model should get by with the "minimum of machinery".


These basic criteria have been used to evaluate models of linguistic description from Bloomfield's time to the present. The question we now ask is whether they are sufficient to evaluate the processing of Japanese inflection. Do these models and measures help us understand what speakers do when they mentally process language? Can these models of formal description also account for the processing of language in a way that is psychologically real? In the next two chapters we consider specific mechanisms and models which move beyond description and attempt to characterize how inflection might actually be processed.




In addition to the three basic models of formal description, lA, IP, and WP there

are also models that are specifically designed as models of processing. Process models are designed to account for what speakers do or might do as they mentally represent and process inflection. The two basic mechanisms of the descriptive models, rules and analogy, are both elaborated in process models. In addition to rules and analogy, different process models also propose mechanisms such as competing forms, connections, and paradigm related mechanisms. Before we consider current process models in detail, we briefly review the mechanisms that characterize the process models.

4.2. RULES.

Rules arise from the notion of regularity. Matthews (1991: 130) presents a general common sense definition of regular inflection. To say that a form has a regular inflection is to say that it has the inflection one would expect unless one knew that it was different.

Pinker (1993:473) provides a specific English example and an illustration of how a rule might represent a productive psychological reality. Regular inflection, being fully predictable, is computed by a rule that concatenates the affix -d to the verb stem. This allows a speaker to inflect an unlimited number of new verbs, an ability seen both in adults, who easily create past forms for neologisms like faxed, and in preschoolers, who, given a novel verb like to rick in experiments, freely produced ricked.


Pinker is of course referring to Berko's (1958) wug test experiment in which children who could not have previously heard the past form ricked were assumed to be able to extract the past rule and apply it to produce the new form. On the basis of Berko's experiment, in addition to the descriptive economy of formal rules, it has also been widely assumed that rules represent a fundamental mental mechanism in the processing of verbal inflection. Some experimental researchers such as Pinker (1999), Marcus (2001), and Clahsen et al. (2001) have made the even stronger claim that there is a further demonstrable distinction between the manipulation of abstract rules in the processing of regular inflection and the processing of irregular forms by associative memory. Not all investigators would agree with this interpretation of the wug experiment. Hockett himself (1968), as well as the psycholinguists Ervin-Tripp (1966), and Slobin (1971), and more recently Givon (1984) have suggested that there is no direct evidence that speakers actually process language by making use of symbolic rule manipulation. There are a number of problems with a rule manipulation treatment of Japanese inflection. As noted in earlier sections, the bare stems of verbs do not appear in isolation in colloquial spoken Japanese. In English, bare stems occur with high frequency in constructions that highlight the least marked stem form. Verb learning in English requires stem identification, but not necessarily by means of the segmentation of verbs into stems and endings. If a listener waits long enough, sooner or later the stem form will appear in the input. For example, when an English-speaking parent says "Do you want to play with daddy now?" there is no need for the listening child to isolate the stem of the verb play. The stem is already there. English-speaking children need only


identify stem forms. In the case of regular verbs, this is not a difficult task because there usually are only three forms: play, played, and playing. If the same IP rule mechanism were applied in the acquisition of Japanese, a

Japanese child would need to isolate the stem, asob- from high frequency inflected forms such as asobitai 'want to play', asobanai 'not play' , asonde 'play and', asoboo 'let's play' and the past form asonda 'played'. While there are general references to affix-stripping, a specific and detailed proposal for how this might be accomplished in Japanese has not yet appeared in the relevant literature. What rule-based models of inflection do propose is that the basic mechanism of inflection is a combinatory rule, which combines a stem and an ending. The crucial assumption is that speakers are able to and do segment verbs into stems and affixes. But, how do first learners and adults go about achieving the necessary segmentation in the first place? A satisfactory detailed account of the mechanism of affix stripping has not been proposed for English, the language most often associated with rule based verbal inflection. A mechanism sufficient to account for affix stripping in Japanese would further need to account for the absence of bare stems and the complications of stem allomorphy. For example, what kind of and how much input is necessary for a Japanese child to recognize that ason-da 'played', a word with no b, belongs to the paradigm of a asob-u 'play' or that kaet-ta 'returned', a word with no r belongs to the paradigm of kaer-u 'return'? A final question that should be asked in this regard has to do with the motivation for segmenting and mentally representing Japanese verbs separately as stems and affixes. If the bare stems of consonant-stem verbs never appear in speech, why would speakers need to expend the energy of performing mental segmentation? The IP rules that we have previously discussed can be used to economically describe Japanese


verbal inflection in terms of mental processing. However, the economy of undoing inflected forms only so that obligatory rules can apply to recombine them would appear to be poorly motivated. An alternate mechanism is considered in the next section.


In the processing of Japanese inflection, analogy is a mechanism that could obviate the problematic segmentation, implausible storage of unpronounceable stems, and unmotivated recombination of morphemes required by IP rule mechanisms. The following definition of analogy is from MacWhinney (1978: 113). A procedure which generates a target form from a given form by locating two other known forms related to each other functionally as the given is related to the target and applying their differences to the target, ie., wug : X :: bug: bugs, where X -> wugs (ana ["back, again"] + logos ["word'])

A rule-based account of the processing of the Japanese verb kaer-u 'return' would presumably involve the segmentation of a given inflected form to isolate the unpronounceable root kaer-, which would be committed to storage, along with segmented affixes such as the non-past -u, the past -ta, the volitional


and all the

others. Such a model would be economical in the sense that it would not require the storage of fully inflected forms in memory. In the case of production of the inflected past form (kaet-ta), an additional phonological rule would need to be specified in order to adjust the unpronounceable form kaer#ta to the surface realization kaet-ta. Using the schema proposed by MacWhinney, an analogical account of the production of the same inflected past form of kaeru might start by assuming that fully inflected forms are stored in memory. A search for functionally related forms might take a number of paths. One possibility is that Japanese speakers make use of


phonological similarity based on word-final rhymes. If, for example, the inflected past form of kaeru were not immediately available from memory, a search could be performed for another non-past verb form with a similar word-final -ru. There are a number of candidates. The results of the dictionary type-frequency search reported in Appendix II show that the single most frequent verb type (approximate 29% of all modem verbs listed in the Daijirin dictionary) is the stem-final e verb that has a nonpast form ending with -ru. The high frequency verbs ne-ru 'sleep', i-ru 'be located', and tabe-ru 'eat' come to mind. The second portion of the functional relationship could be the non-past to past relationship. These verbs have the respective past forms ne-ta 'slept', i-ta 'was located', and tabe-ta 'ate'. Unfortunately, a proportional

analogy based on any of these non-past/past pairs yields an incorrect analogy: kaeru:X::taberu:tabeta, > *kaeta.

It is the second most frequent verb type (consonant-stem verbs with stem-final r) that is the appropriate category for the verb kaer-u. About 19% of all modem verbs listed in the Daijirin are members of this paradigm, which has non-past form. There are many high frequency examples: ar-u 'have', hair-u 'enter', kir-u 'cut'. The respective past forms are at-ta 'had', hait-ta 'entered', and kit-ta 'cut'. A proportional analogy based on any functionally related pair of words in this category (a non-past form with the word final syllable -ru and a known past form) yields the appropriate result: kaeru:X::kiru:kitta > kaetta. While this type of classical proportional analogy is widely recognized as an important diachronic mechanism of language change (Antilla 1974, Paul 1880), linguists willing to acknowledge any significant synchronic role for analogy in language acquisition and processing are in the minority. Skousen (1989, 1992), in an extreme position, has gone as far as to claim that everything in language is analogical.


His model would appear to be intended as a process model, also taking into account similarity, frequency, and recency, factors generally not considered in classical proportional analogy. The frequency effect is also referred to as the gang effect, which is the terminology used by Rumelhart and McClelland (1986, 1987) in the discussion of their connectionist model, where there is no mention of traditional rules at all. MacWhinney also makes use of the term gang effect and allows for an expanded role for analogy in acquisition and processing. In his 1978 monograph MacWhinney proposes three processes that would operate in first language acquisition. While his original proposal was intended as an account of language acquisition, there is no reason to believe that adult speakers do not make use of these same three processes: rote, analogy, and combination. The term combination is used by MacWhinney to make a distinction between descriptive rules, which do not claim to be psychologically real and the kinds of formalized operations that are used in actual mental processing. MacWhinney intends that combinations have psychological reality. MacWhinney's (1978:6) model also provides for a hierarchy of application and a method of distinguishing analogy from rote and combination. Moreover, the model holds that combination only applies when rote fails and that analogy only holds when both rote and combination fail. ... When the child produces correct real forms without problems but cannot generate nonce forms even when given maximum contextual support, there is evidence for rote. When the child produces erroneous real forms and regularized nonce forms, there is evidence for combination. Use of combination may also show somewhat longer response latencies. When the child produces correct real forms and nonce forms of many different shapes, there is evidence for analogy. The use of priming should increase the use of analogy.


In this rote> combination> analogy hierarchy, response latency may also give some hint about the kind of processing that is occurring. The use of combination (a combinatory rule which brings together a stem and a suffix) may require more time than the recall of forms from memory. Since analogy follows combination in this same hierarchy, MacWhinney seems to suggest that analogy may be an even slower alternative mechanism in the processing of inflected forms. This notion of slow and deliberate analogical processing contrasts with current assumptions about rule-based processing, which is generally assumed to be rapid and automatic. It is widely believed that Berko's wug experiments with children provided

conclusive evidence that children are able to extract and manipulate abstract morphological rules without explicit instruction. This is the conclusion that Berko (1958:171) herselfreached. The answers were not always right so far as English is concerned: but they were consistent and orderly answers, and they demonstrated that there can be no doubt that children in this age range operate with clearly delimited morphological rules.

However, closer inspection of Berko's report reveals that the children's answers were neither as rapid nor as automatic as a rule-based model of processing might predict. In fact, the children were often quite deliberate and even hesitant (1958: 171) in their responses. The children answered the questions; in some instances they pronounced the inflexional endings they had added with exaggerated care, so that it was obvious that they understood the problem and wanted no mistake made about their solution. Sometimes, they said "That's a hard one," and pondered a while themselves.


Berko does not report the details of individual responses, so we cannot know how often the four to seven year old children in her study hesitated to respond. What a careful reading of her paper suggests is that hesitation was not at all atypical. If rule-based processing is characterized as automatic, the lack of automaticity in

response may indicate that the ability to pass a wug test is not a conclusive demonstration of rule acquisition and manipulation after all. How do we know that a child, who is able to produce the past form ricked, when given the nonce form rick, is not making use of some simple analogy? It may be that speed of performance is not helpful in distinguishing analog and rule-based processing. Some have argued that in as much as both types of processing can yield the same result, it is not possible to distinguish them. Even Kiparsky (1975:189), who has argued most forcefully against analogy, has made this same observation. Generative phonologists argue that at the point at which the analogies begin to make the right generalizations, they are indistinguishable from rules.

However, Kiparsky goes on to present the argument that because there is a functional constraint on both rules and analogy, there cannot be an analogy which has not already been described as a rule. For example, in the following proportion, analogy provides an acceptable result: sister:sisters::brother:X;X = brothers. Kiparsky explains that analogy is possible here because there is an existing functional relationship in English whereby nouns are pluralized by the addition of s. Kiparsky regards this type of analogy as plausible, because it is already represented as a rule. On the other hand, he regards the following proportion as implausible:

ear:hear::eye::*heye. In other words, because there is no rule that produces a form like *heye, neither will there be an analogy. Zager (1980) and Bybee (1995) argue


against Kiparsky's position and extend the notion of analogy by making a crucial distinction between source-oriented and product-oriented analogies. Bybee's schema model is discussed in more detail in Chapter 5. For the present, we consider just two examples from Japanese to illustrate that there are productive analogic schemas in Japanese that cannot be expressed as generative rules. We look first at one example of diachronic change. Martin (1988:477) cites Shibata's (1965:34) report of a downtown Tokyo dialect that appears to have eliminated a complete verb class based on an analogic relationship. The verb sin-u 'die' is the only verb in modern colloquial Japanese with a stem-final n. According to Shibata, the downtown dialect has replaced the stem sin- with sim- throughout the paradigm. What is significant here is the direction of the change. The significant partial paradigms of three representative verb paradigms are given in Table 33. Non-past




Dialect fonn sim-u die sim-anai not die sin-da died sin-de die and Tokyo standard sin-u die sin-anai not die sin-da died sin-de die and Tokyo standard yom-u read yom-anai not read yon-da read yon-de read and Tokyo standard yob-u call yob-anai not call yon-da called yon-de call and TABLE 33. Loss of stem-final n verb class in an innovative Tokyo dialect (Shibata, 1965).

Standard Japanese stem-final n, m, and b verbs have the stem allomorph N in the past and gerund forms. The innovative dialect speakers in Tokyo recognize a relationship between forms such as sin-da and sin-de and the paradigm of verbs like yom-u. A representative example of this relationship is the proportion: yonda:yomu::sinda:X;X

= simu.

This analogy has the result that n of the standard

dialect becomes m in the paradigm of the innovative dialect. However, there is no phonological or morphological precedent for this analogy in the standard dialect. On a strict generative account, the underlying stem forms of verbs like yob- and yom- and sin- become yoN- and siN- in the inflected past and gerund forms. The direction of


change is from b, m, and n to N. There is no context free morphophonological rule that changes an n to an m. If inflected verbs are fashioned from the concatenation of stems and endings, the loss of the stem-final n paradigm in the innovative Tokyo dialect is exactly the kind of implausible analogy that Kiparsky predicts should not occur. The preceding example makes use of classical proportional analogy, however, Antilla (1989:91), who has literally written the book on analogy, points out that proportional analogy is just one type of analogy. In other types of analogy, for example, contamination, there may be no proportions. Antilla cites the examples *feets and *camed from the speech of children to show how an analogy is contaminated by forms that do not constitute part of a valid proportion: hand:hands::feet:X;X = *feets, walk:walked::came:X;X = *camed. In these two examples, the terms necessary to establish a valid proportion (foot and come) are not found. Phonologically and semantically related forms contaminate the analogy and the result in a blend (*feets and *camed). Examples from the speech of Japanese children (Klafehn, 2001 b) are *aranai, *aronai, and *arenai for the standard adult form nai 'not have, not be'. These examples are quite different from the usual pattern of overregularization in English, where it appears that a regular rule is extended to apply to an irregular verb, for example *holded in place of held or *singed in place of sang. In these English examples, what gets regularized is the ending. In the Japanese example, the anticipated regularization would have to apply to the shape of the irregular stem. In Japanese the verb ar-u 'have, be' has the irregular negative form nai 'not have, not be' where the form *aranai is expected. The irregular past form is na-katta 'didn't have, wasn't' where the form *ar-anakatta is expected. Contrastive partial paradigms of irregular ar-u and regular stem-final Irl verbs are given in Table 34.




Negative past

wakar-u understand wakar-anai not understand wakar-ana-katta didn't understand kaer-u return kaer-anai not return kaer-ana-katta didn't return ar-u have, is *ar-anai> nai not have, isn't *ar-anaktta > na-katta didn't have, wasn't TABLE 34. Partial paradigms of regular verbs and irregular aru.

The child error form *aranai is sometimes anecdotally reported and does appear to be a regularization of the stem in the irregular paradigm. However, Klafehn (2001b) reports that a careful search of archived files (Miyata 1992, 1993, 1995, 2000) in a longitudinal database shows that this child error form is extremely rare. Furthermore, the same child who produced *aranai, also produced *arenai and *aronai during approximately the same time period. There are inflected forms in the paradigm to which ar-u belongs where the stem-final consonant r is followed by the vowel


or by

the vowel e (ar-eba and ar-oo), but these lower frequency forms are not produced by the children in the database. Therefore, these error forms cannot be explained as examples of overregularization, where an existing rule is applied to a new lexical item. These error forms are better explained as blends or contaminations arising from analogic extension. Although there are many possible proportions for *aranai based on similarity with other stem-final r verbs: wakaru:wakaranai::aru:X;X

= *aranai,

there are no proportions for *arenai or *aronai. Klafehn (2001b) also reports child errors such as *neroo for the adult form ne-yoo 'let's sleep, let's go to bed', from the same database. As previously described, Japanese verbs with the non-past shape -eru and -iru are ambiguously either vowelstem or consonant-stem verbs. Contrastive partial paradigms are repeated in Table 35. Verb type




Consonant-stem kaer-u return kaer-oo let's return kaet-ta returned Vowel-stem ne-ru sleep ne-yoo let's sleep, ne-ta slept TABLE 35. Contrastive consonant-stem and vowel-stem verbs with ambiguous stem shapes.


The conventional rule description of the Japanese volitional form says that



added to the stem of consonant-stem verbs and -yoo to the stem of vowel-stem verbs. For example, the volitional form of the consonant-stem verb ik-u 'go' is ik-oo 'let's go'. Therefore, the child error form *neroo for *ne-yoo cannot be described as an overregularization in the usual sense of the term. First of all, the paradigm of the vowel-stem verb ne-ru is already perfectly regular. Furthermore, it is not possible to describe a productive rule for the Japanese volitional form that combines a stem and the suffix -roo. Such an analysis would only be valid for one arbitrary class of verb stems, namely the class that the standard analysis describes as stem-final r. In order to accommodate a combinatory rule analysis that would identify -roo as the volitional suffix of the verb kaer-u, it would be necessary to identify a different volitional suffix for each initial consonant of the volitional suffix. This very uneconomical alternate analysis of the volitional form is illustrated in Table 36. Volitional

Non-past ka-ku oyo-gu no-mu yo-bu kae-ru mo-tu tuk-au hana-su

write swim drink call for return hold use speak

ka-koo oyo-goo no-moo yo-boo kae-roo mo-too tuka-oo hana-soo

Volitional suffix -koo -goo -moo -boo -roo -too

let's write let's swim let's drink let's call for let's return let's hold let's use let's speak



TABLE 36. Alternate analysis of regular consonant-stem verbs showing segmentation.

Therefore, the child error form *neroo for *ne-yoo cannot be described as an overregularization in the usual sense of the term, where a more general rule is extended to include an irregular lexical item. Japanese does not have a more general combinatory rule that makes use of the suffix -roo. However, it is possible to identify any number of analogic proportions that would explain *neroo, for example kaeru:kaeroo::neru:X:X


= *neroo.

Once again, as with

the child error forms *aronai and *arenai, the error form *neroo is not easily explained as the implementation of an existing general rule in a new environment. In addition, Klafehn reports that there was no consistent pattern of volitional form error where *-roo was used in place of the consonant-stem suffix


and the vowel-stem

suffix -yoo. Therefore, these errors do not appear to be the result of productive rulebased processing. Rather, these errors are evidence that Japanese children make use of some type of analogic processing (based on a similarity of forms represented in memory) in the production of regular inflected forms. Furthermore, there is no reason to believe that analogic processing is not being used in the production of correct forms as well. This is a significant departure from the usual assumption about the acquisition of first language inflectional morphology. Generally it is assumed that productive rules are acquired to deal with regular forms and that analogy has a much more limited role associated with the processing of exceptional irregular forms. However, it appears that there may often be constraints on analogic processing that do not allow it to become fully productive. Analogic processing appears to be limited to the processing of forms of similar shapes. Pinker (1999) and others (Bybee and Moder 1983, Kim, Pinker et al. 1991, Prasada and Pinker 1993) have reported experimental findings that support the claim that irregular English verb pairs such catch and caught are semi-productive. Adult speakers were found to frequently extend

irregular patterns to nonce stems such as spling-splung. Pinker (1994:323) also admits that analogy has been responsible for some historic language change. Irregulars can be added to the language diachronically by analogy with existing forms (i.e., caught, cost, flung, knelt, quit, slung, stuck, strung; Jesperson 1942), a process that is especially obvious when many dialects are compared (e.g., bring-brung, drag-drug, climb-dumb, heat-het: see Mencken 1936).


However, even MacWhinney (1978), whose analogy model was described above, generally reports little evidence of analogy in experimental responses from children in his monograph study of eight different languages, which included English. Both Marcus (2001) and Pinker (1999:16) have specifically argued that full productivity is only characteristic of rule-based processing in English. The following quote is from Marcus (2001:70). A further bit of evidence that the operation of -ed is rule-like is that children seem to be able to apply it even to verb stems that are homophonous with irregular verbs. When they are told This is a ring, Now I'm ringing your finger. What did I just do?, adults (Kim, Pinker, Prince & Prasada,

1991) and even three-year-old children (Kim, Marcus, Pinker, Hollander & Coppola, 1994) respond You just ringed my finger not You just rang my finger.

Pinker 1999: 16) includes a reference to the original Berko experiment. And the freewheeling children in Berko Gleason's study were downright stodgy when it came to irregular forms: Only one out of eighty-six turned bing into bang and one other turned gling into glang.

These differences suggest a simple theory. Regular past-tense forms are predictable in sound and generated freely because they are the products of a rule that lives in the minds of children and adults: "The past tense of a verb may be formed from the verb followed by the suffix -ed."

This simple rule theory is compelling and the examples offered appear to confirm the claim that analogic processing is not productive for young children. However, both Marcus and Pinker seem to have overlooked one crucial fact in the original Berko paper. Berko (1958:161) reports that none of her four and five year old preschool subjects and only 25% of her five and a half to seven year old first grade subjects responded to the following prompt with rang. "This is a bell that can ring. It is ringing. It did the same thing yesterday. What did it do yesterday? Yesterday it -. "


If a three year old has not yet learned that the irregular past of ring is rang, then

there is no reason to expect an analogy based on this paradigm. However, even when there is clear evidence of learned forms in memory that could serve as the basis of analogic extension, there is no way to know when children will actually use those forms. Analogy may be a productive process, but it is not predictive in the same way that rules are predictive. A Strong Rule hypothesis of inflectional processing predicts that once a rule has been acquired that rule will apply in every environment when it can apply. Kiparsky (1975:189) has used this indeterminacy of analogy to argue against the analogic processing of inflection. It is clear that this schema predicts and (sic) indeterminacy in actual speech behavior which we in

fact don't find there. If anything like this was right, then speakers should form the plural of dog sometimes with [-s] on the analogy of cats, sometimes with [-z] on the analogy of birds, and sometimes with [-iz] on the analogy of fishes, depending maybe on which sort of plural happened to precede in the context. Just this indeterminacy is of course inherent in the notion of analogy, and it is essential, as in it lies the whole plausibility of analogy as a mechanism of linguistic change.

Kiparsky is correct about the significance of context. Today it is customary to refer to the effect of similar words in the immediate context as a recency effect or a priming effect. However, at the time when Kiparsky was writing not much experimental work had yet been done investigating recency or frequency effects, both of which are now considered significant in the processing of lexical items. It is not clear, as Kiparsky claims, that an analogic schema necessarily predicts that children would alternate their use of the plural suffix. Kiparsky makes the conventional assumption that English plurals arise from the combination of a singular stem and an abstract inflectional morpheme, which later receives some type of phonetic interpretation. It is just as possible, however, that children, as well as adults, could have lexical storage of both


the form dog and the inflected plural form /dog-z/, including its appropriate phonetic specification. Depending on recency and frequency, these fully inflected forms in memory would be more or less likely to be affected by similarity with other forms. Low frequency forms would be less stable than high frequency forms. A high frequency token such as dogs would therefore be much less likely to show the indeterminacy that Kiparsky predicts from analogic schemas. The significance of type and token frequency is absolutely crucial to this discussion and beyond the explanatory scope of either traditional proportional analogy or the contamination and blending of a non-proportional analogy. Neither traditional rules nor traditional analogy make use of frequency or recency considerations. The conventional assumption about analogy is that a single analogy produces one form, which is based on some perceived similarity of form and function. The conventional assumption about rules is that a single rule will apply any time it can apply. An expanded schema account of lexical processing, including type and token frequency considerations will be elaborated in following sections which deal with specific competition, connectionist and network models.


In the previous section we noted MacWhinney's conclusion that morphological production depended on three essential skills: production by rote, production by rules, and production by analogy. A further elaboration of MacWhinney's (1978) early distinction between rules and analogy was that analogy involves the comparison of just two similar items, while rules involve the comparison of more than two similar items. While a comparison of forms has always been implicit in the notion of analogy,


Bates and MacWhinney (1987:256) claim that the competition between forms, or more precisely, the competition between cues is what characterizes language processing. The status of rules in descriptions of children's language has often been called into question. For example Hockett (1968) and Givan (1984) have pointed out that there is no direct evidence that language users actually manipulate rules and rule systems in their heads in the same way that rules are processed in a linguist's grammar. Nor is there direct evidence that children pick up rules during actual conversational interactions. How is it then that children come to behave as if they knew the rules, even without learning rules as discrete entities? Does the mechanism underlying this learning resemble the rule itself or does it look like something very different? Our goal here is the articulation of an approach to morphological processing and structure that does not depend on rules. This approach views morphological form as arising from a competition between a large set of phonological, syntactic, and semantic cues.

The Competition Model places great emphasis on the association of form and function, and the concept of the cue is seen as most important in this form and function mapping. In a recent paper MacWhinney (2003a) provides a set of cue-related mental processes that the current version of the Competition Model assumes to be available for language learning. There is no reason to believe that adult native-speakers do not employ at least some of these same mechanisms in the routine processing of morphology. In the following section we examine three items from MacWhinney's list that have relevance to the processing of Japanese verbal inflection: frequency effects, chunking, and resonance. The competition model makes explicit claims about frequency effects. It predicts that more frequent items are learned before less frequent items. The term cue availability refers to the relative frequency of cues in the input. Because in Japanese, the formal style is normally employed in situations where the -masu ending is a


measure of social distance and we find that that young children produce the plain style non-da 'drank' earlier than the semantically equivalent formal style nom-imas-ita 'drank'. The term cue reliability is used to refer to how often a particular cue is associated with a particular category. In our discussion of Japanese verbal inflection in Chapter 2, we pointed out how a conventional rule treatment of Japanese does not address the issue of potentially ambiguous inflected forms. Even when there is a context that makes the meaning clear, paradigm identification is another matter. For example, the paradigm categorization of an isolated verb form like katta is a problem for native adults as well as learners. The word final sequence -tta can be considered an important cue for the marking of the past. However, the form katta could represent the past form of the stem-final r paradigm (kar-u 'cut'), the past form of the stem-final t paradigm (kat-u 'win') or the past form of the stem-final W paradigm (ka-u 'buy'). The Competition Model recognizes that that the word final sequence -tta does not have high cue reliability for paradigm categorization in Japanese. The term task frequency refers to how often a particular marking task occurs in discourse. Bates and MacWhinney (1987:260) offer the example of German where the task of marking the nominative case occurs much more frequently than that of marking the genitive case. The consequence of this frequency difference is that Germanspeaking children learn nominative case marking before genitive case marking. For English-speaking children, the task of marking the present has a relatively high frequency and is facilitated by the fact that the present form is also the stem form, which is believed to be stored in memory. For Japanese children (Clancy 1985), on the other hand, the past form is among the first to appear and the relative task frequency of marking the past is higher than the task frequency of marking the non-past. It appears


that English-speaking children first acquire the present form of a verb and then learn how to mark the past form. Japanese-speaking children acquire the past form of a verb then learn how to mark the non-past form. This concept of task frequency is also relevant to the investigation of adult processing of Japanese verbal morphology because there are so many possible inflected forms in Japanese. Not all of these inflected forms occur with the same frequency. The Words and Rules model of inflectional processing claims that combinatory rules are insensitive to task-frequency. The Competition Model, on the other hand, predicts that both productivity and accuracy depend to some extent on how often a certain inflected form must be produced. Therefore inflected forms that are not produced as often should be learned later and are more likely to be produced with errors. The contrasting inflectional typologies of English (few inflected forms) and Japanese (many inflected forms with varying type frequency) provide an excellent opportunity to test the predictions of these two models. Chapters 6 and 7 take up this issue. The notion of chunking is also relevant for both acquisition and processing. Acquisition researchers (MacWhinney 1978, Peters 1983, 1985, Slobin 1985 and Tomasello 1992) have repeatedly pointed out that the units of linguistic analysis employed by children and adults may not be the same as the units used by linguists. Adults as well as children clearly make use of memorized phrasal chunks longer than the word. MacWhinney reiterates the significance of longer chunks in attaining fluency in a second language. The use of fully inflected unanalyzed verb chunks could facilitate the processing of first language Japanese by native adults as well. Traditional rule-based approaches to the processing of inflection assume that the lexical storage of fully inflected forms is uneconomical in terms of memory requirements. This assumption may be well motivated for English. However, Japanese


represents a different morphological typology. The bare stems of vowel-stem verbs rarely occur in discourse and the bare stems of consonant-stem verbs never do so. A processing mechanism that requires analysis and segmentation of verbs into stems and grammatical morphemes for storage, only to reassemble them again during production, may be economical in terms of storage, but it is not economical in terms of processing. MacWhinney also makes brief mention of the mechanism of resonance and points out how it is historically related to the notion of spreading activation in psychology. (2003:7) Resonance occurs when several concepts are activated and their mutual excitation leads to the formation of a resonant pattern

McClelland and Rumelhart (1981) made use of the term spreading activation to describe how their model of word naming was able to recognize written words. A word was recognized when bottom-up activation from segments and letters spread and activated the correct lexical item more strongly than its nearby competitors. This sort of interactive model demonstrates an important property and mechanism of both competition and connectionist models. These models do not make a distinction between memory and processing. Spreading activation is relevant in the investigation of Japanese verbal morphology because it allows for the possibility that there may be several related and competing forms in the lexicon of Japanese speakers, any or all of which may resonate during recall and processing.

4.5. CONNECTIONS. Connectionist models of language are based on the assumption that items are connected in networks of association where the strengths of these connections are adjustable. Connectionist models as well as the Competition Model in its different


applications can be characterized as models of language that do not depend on formal rules. Connectionist models depend upon associative memory. Some of the best known examples in the literature are those that have been used to model the learning of the English past tense (Jaeger et al. 1996:454). The central claim is that the memory system is able to generalize across items so efficiently that the regular stems are mapped directly to their appropriate past tense forms in the same way that irregular stems are mapped directly to their appropriate past tense stems.

Rather than rules, connectionist models depend on associative networks where connection weights can be adjusted through experience. In general Rumelhart and McClelland's (1986) model of the English past tense performed quite well and responded with the correct past form of the 460 verbs it was taught. However, the network did not do so well with some lower frequency or novel verbs that were tested, offering *membled as the past tense of mail, for example. There are two very important consequences of associative memory mechanisms: networks do not make use of combinatory rules, and networks do not make use of verb stems. For Pinker (1999: 115) the connectionist model's inability to recognize verb stems is the fatal flaw in the architecture of all network models. It is not just a quibble; it explains an embarrassing lapse in the performance of Rumelhart and

McClelland's model. The model was mute when asked for the past tenses of simple but somewhat unusual sounding words, like jump, pump, warm, and trail. And it garbled several others, turning squat into squakt, tour into toureder, and mail into membled. The lapses are puzzling to us because intuitively nothing could be simpler than copying a stem over to the past-tense form before adding ed. But a pattern associator memory has no placeholder called "stem" that can be copied, and no operation to do the copying.


Pinker has argued most forcefully in a number of recent articles and books that the human mind is innately organized to categorize and store and process verbal morphology in terms of stems and affixes. The basic mechanism of his Words and Rules model is the combination of stem and verbal suffix. However, a number of experimental investigators (de Chene 1982, Yokomizo 1990, Vance 1991, Batchelder 1999, Burch 2002) report results that are consistent with the claim that native adult speakers of Japanese do not have a clear notion of verb stem. For example, some of de Chene's subjects thought that the negative form of the nonce verb kizu would be *kizoranai or *kizuranai rather than the expected form kizanai. Some of Yokomizo's

subjects thought that the actual loan verb toireru 'go to the toilet' had the negative form *toirenai rather than the expected form toireranai. Some of Vance's subjects thought

that the conditional form of the nonce verb hoku was *hokureba or *hokereba rather than the expected form hokeba. Some of Batchelder's subjects (in an elicited production test) thought that the nonce form verb ganoru had the past form *ganosita rather than the expected form ganotta. Some of Burch's (2002) subjects (in a paradigm completion task) thought that the nonce verb karatu had negative form *karawanai, or *karataranai, or *karatarenai rather than the expected form karatanai. In each of these

experimental studies native speakers of the language, just like the pattern associator model, inserted or deleted segments in a way that was inconsistent with the pattern predictable from the shape of an abstracted stem. These results suggest that in addition to not being able to identify the stems of verbs, Japanese speakers may not make much use of paradigmatic cues. Pinker (1999), Marcus (2001) and others (Smith & Tsimpli 1995, Smith 2000) have argued that connectionist models are inappropriate models of human verbal processing because they are too powerful and are not constrained by a knowledge of stem shape. It


is argued that connectionist models rely on a statistical processing capacity that is not typical of human language processing. As we have noted in the case of English verbs, Rumelhart and McClelland's 1986 connectionist model predicted that *membled was the past tense of mail. The error form *membled could be interpreted as the result of detailed statistical averaging across the shapes of various input patterns. This type of statistical averaging is generally not considered a possible mechanism for human speakers. On the other hand, experimental evidence from various nonce-probe experiments with Japanese begins to suggest that the knowledge of stem shape that English speakers appear to make use of may in fact be a typological artifact of the isolating morphology of English, rather than some type of innate processing mechanism. This possibility is investigated in Chapters 6 and 7.




In the following three subsections we consider briefly three recent models that have been proposed for the processing of verbal inflection. We have already discussed some of the distinctive mechanisms employed by each model. Now we consider how each of these three models might deal with the specific details of Japanese verbal inflection. It is important to point out that these three models should not be thought of as formal models of linguistic competence distinct from linguistic performance, but rather as "processing models" that are intended to deal with the psychological status of language processing. Even the computer simulations within the connectionist framework are intended to model at least some aspects of how processing might be handled by actual speakers. This focus on processing is consistent with the theme of this dissertation. For example, in Chapter 6 when we examine examples of verbal errors in the speech of Japanese children, we attempt to look for error patterns, but we have neither interest in nor method for distinguishing performance error from competence error. In Chapter 7 when we examine the patterns of verbal productivity of adult native speakers and learners, we consider experimental evidence based either on elicited forms or forced choice testing, but do not make use of grammaticality judgments. The distinction between rule-based processing versus non-rule based processing is one of the central issues in this regard. Pinker's Words and Rules model makes the strong claim that morphology is processed by means of abstract formal rules. The various connectionist models and Bybee's network model do not make use of formal rules. We consider the implications of this distinction in each of the following model summaries.


5.2. WORDS AND RULES MODEL (PINKER 1999). Pinker and colleagues (Pinker & Prince 1988, Pinker 1991, Marcus et al. 1992) reacted strongly against the notion that the single processing mechanism of Rumelhart and McClelland's (1986) connectionist model could account for verb inflection in English. Pinker's (1984) own model of processing is based on his early work relating to the significance of paradigms in the leamability of language and on various objections to connectionist models. His Words and Rules Model (1999) is proposed as a general model of morphological acquisition and processing and is sometimes referred to as a dual-mechanism model. One mechanism is dedicated to the rule processing of regular forms and another distinct mechanism is required for the processing of irregular word forms. The stems of regular verbs are economically stored separately from their various inflectional endings, and appropriately combined by use of abstract rules. This is an extremely elegant mechanism for languages without much inflectional morphology, English for example. It is not clear, however, exactly how this mechanism would work for agglutinating languages like Turkish and Japanese. English allows only one suffix per verbal stem. The combination of walk and -ed is a terminal operation. Japanese allows up to four suffixes: tabe-sase-rare-na-katta (stem-causative-passive-negativepast) '(I) was not made to eat'. How is it that speakers know that -rare- must precede -

na, but -na may also follow the stem when there is no -rare- morpheme in the word? The mechanism for this is not specified in the Words and Rules model. The Words and Rules model does specify a mechanism for processing irregular forms, which are assumed to be stored in memory. Since irregular past forms such as English ran and held and hit are not predictable by the general rule for the past, which combines a stem and the -ed ending, they are entered into memory along with the present forms run, hold, and hit. Interestingly, although Pinker rejects the single


mechanism hypothesis by itself, he (1999:117) suggests that some type of unspecified associative memory mechanism which may actually work much like the connectionist model may be employed to relate the various sub-types of irregular patterning, for example the semi-productive string/strung pattern. Our proposal is simply the traditional words-and-rules theory with a twist. Regular verbs are computed by a rule that combines a symbol for a verb stem with a symbol for the suffix. Irregular verbs are pairs of words retrieved from the mental dictionary, a part of memory. Here is the twist: Memory is not a list of unrelated slots, like RAM in a computer, but is associative, a bit like the Rumelhart-McClelland pattern associator memory. Not only are words linked to words, but bits of words are linked to bits of words.

In this model irregular forms entered into memory serve another function in the processing of inflection. When a past form of an English verb is required in speech, the presence of the memorized form blocks the application of the regular rule. This mechanism is sometimes disrupted and production errors do result. Occasionally, even native-speaker adults are distracted for some reason and do not immediately access a stored irregular form, the operation of the regular rule is not blocked and an irregular form is overregularized. This failure of blocking results in overregularized forms such as *knowed and *blowed instead of knew and blew. This blocking function of irregular memorized forms has been used to explain what has been called the pattern of V-shaped learning (Bloom 1993) in first language acquisition. For example, young English-speaking children are reported to pass through an early developmental stage where both regular verb forms and irregular verb forms are used appropriately. This early period is followed by a stage of development when irregular forms may be overregularized. Non-adult forms such as *goed and *breaked are produced. In a final developmental stage children stop overregularizing and return


to adult-like performance. The Words and Rules model proposes that the blocking mechanism whereby memorized irregular forms block regular rule application explains the transition between the period of overregularization and adult-like performance in the final stage of child language development. The Words and Rules model is intended as model of how the human mind is structured and therefore should not depend on language typology. It is not clear, however, just how it would apply in the case of an agglutinating languages where there are no bare stem forms as independent morphemes. It appears that for Japanese, in addition to the distinction between regular and irregular verbs, there must also be the assumption of some additional mechanism to segment stems from the fully inflected forms that appear in speech. Segmentation of the most economical stem form would therefore entail the mental storage of unpronounceable consonant-final forms such as ik- 'go', mat- 'wait', and oyog- 'swim'. The stem of the verb ki-ru 'wear' would be

stored as ki-, but the stem of the verb kir-u 'wear' would be stored in memory as kir-. The storage of both vowel-final and consonant-final stems would be consistent with the abstract symbolic knowledge assumed in the Words and Rules model. Since combinatory rules in the model operate on the basis of paradigmatic distinctions, native speakers would be expected to be able to establish the paradigm identity of homophonous forms on the basis of form comparison. When confronted with the nonce-forms riru and ritai it would seem a simple task for a native-speaker to segment the forms, identify the paradigm, and predict the past form: 1) the forms are segmented ri-ru and ri-tai, 2) the verb ri-ru is a vowel-stem verb, 3) the past form should be ri-ta.

A different result should be obtained from the nonce-forms riru and riritai: 1) the forms are segmented rir-u and rir-itai, 2) the form rir-ru is a consonant-stem verb, 3) the past


form should be rit-ta. We will review the experimental results of just such a test in Chapter 6. The Words and Rules model makes the strong claim that productivity is not linked to verb type-frequency. It predicts that there will be default regularization, but that the default pattern need not be the most frequent pattern. In German, for example, it has been argued (Clahsen et al. 1992, Marcus et al. 1995, and Pinker 1999) that the default past participle pattern is the -t affix, which has a lower type frequency than the -en frequency. This claim has been disputed. Both Bybee (1995) and Dabrowska (2001) have pointed out that the method used to determine the verb type-frequency in the German studies seriously underrepresents the number of weak verb types. They have both argued that a more accurate count shows that productivity is directly related to type frequency.

In another German study, Kopcke (1988) reported that German speakers did not make use of a single default plural when inflecting novel nouns. The properties of the noun stem determined which plural marker was used. Speakers used -e with masculine monosyllabic nouns, but -n with monosyllabic feminines and feminines ending in schwa. In yet another challenge to the single default hypothesis, Dabrowska (2001) has claimed that the highly irregular system of the Polish genitive is acquired relatively error free by young children. Furthermore, Dabrowska claims that there is no evidence of a single default pattern operating in Polish. It is the allomorphic variation in the regular Japanese verbs that might be expected

to be problematic for young Japanese children. We might therefore expect Japanese children acquiring the language to show some pattern of error. Furthermore, even though Marcus et al. (1992) have argued that the rate of child overregularization is actually quite low, it would seem that as long as there is a distinction between regular


and irregular verbs in a language, the Words and Rules model actually predicts a Vshaped pattern of acquisition. Pinker has argued that failure of the blocking mechanism in acquisition, when the memory trace of irregular forms has not yet been sufficiently established is responsible for the V-shaped pattern learning in English. Since Japanese children, just like their English-speaking counterparts are not expected to exhibit perfect memory, a strong version of the Words and Rules model would appear to predict and actually require a pattern of V-shaped learning in Japanese verb acquisition. A preliminary test of the default error pattern is made in Chapter 6.

5.3. CONNECTIONIST MODELS. Mention of the connectionist models is made here again because these computer simulations of language processing, unlike the Words and Rules model, do not make explicit use of rules. Rather than formal rules for processing verbal morphology, connectionist models make use of adjustable connection weights between input units, output units and hidden units. No distinction is made between the processing of regular and irregular verbs and therefore connectionist models are also referred to as singlemechanism models. Single mechanism models are of interest in the present dissertation because in addition to offering an alternative to formal rules, they also provide a tool for the investigation of human inflectional error. The errors as well as the correct forms that the model produces are of interest and can be compared with the error patterns of real speakers. The original Rumelhart and McClelland (1986) network model was able to produce output similar to a child learning English, including overregularizations. Later models by Plunkett and Marchman (1991) and MacWhinney and Leinbach (1991) responded to criticism that the input used in the original model was biased in a way that


did not reflect the typical input available to young children: a relatively small set of irregular verbs followed by a much larger set of regulars. Plunkett and Marchman (1991) also showed that learning was actually improved if the model was introduced to semi-productive patterns such as the sing, sang and ring, rang pattern. MacWhinney and Leinbach's (1991) later model was also able to distinguish homophonous pairs such as wring, wrung and ring, rang. Consideration of verb type and token frequency, as well as meaning, are clearly important in the modeling of English verb inflection. Unfortunately, the typology of English has reduced much of connectionist modeling to what Elman (2003) has called "the great past tense debate". That debate has focused on the appropriateness of models that attempt to output an English past form based on the shape of present forms that have been input during training. For languages such as Japanese there are two specific objections to such a modeling approach. First of all, the inflectional morphology of Japanese is much richer that English with many more inflected forms, Second, there is evidence in the first language acquisition of Japanese (Clancy 1985) that the past form (-ta) is generally acquired before the non-past form (-(r)u). Therefore, modeling the

past form on the basis of the present may be only a beginning (and perhaps not an appropriate beginning, at that). Hayden (2002) attempted such a beginning with a very basic network model and trained it on a set of actually occurring Japanese non-past and past verb forms, for example kak-u 'write' and ka-ita 'wrote'. Hayden and Klafehn (2003) were later able to use another version of Hayden's network to test nonce verb forms that the network had not been exposed to during training. The input set that this second model was trained on was adjusted to the relative frequencies of 13 different Japanese verb types. The verb type frequency that was used was determined by doing a count of verb types used by


the child Ryo in the Miyata (1993) database. After training, this second model was tested on the same nonce verb forms that Vance (1991) used with adult native speakers of Japanese. The model was able to produce appropriate past forms for some of the nonce verbs that it had never encountered before. It correctly produced hat-ta as the past of hat-u, mut-ta as the past of mur-u, and kat-ta as the past of kap-u. A summary of results appears in Table 37. The input forms and corresponding past forms have been segmented to help the reader in identifying the verb type. In the actual model, of course, the network was presented with unsegmented forms. With the exception of *honta for honda, all of the errors that the network produced duplicated errors made by native speakers in the Vance nonce verb study. A good number of subjects (44%) in the Vance study thought that *hotta was the past of hok-u and a majority (70%) thought that *rita was the past of rir-u. Input non-past fonn

Correct past form

Network output

hat-ta hatta hon-da *honta ho-ita *hotta hok-u mut-ta mutta mur-u kat-ta katta kap-u met-ta *meta mer-u rir-u rit-ta *rita TABLE 37. Hayden (2002) model of Japanese past fonns for selected nonce verbs. hat-tu hom-u

However, there were other errors forms produced by the native speakers that the network did not produce. The errors that native-speaker adults make on nonce-probe tests as well as the errors that Japanese children make in first language acquisition tend to be errors in the identification of stems and paradigms. Just like English speakers, Japanese speakers appear to be primed by similar sounding lexical items in the same and different paradigms. Some of Batchelder's (1999) experimental subjects thought that the nonce verb ganor-u had the past form *ganosita. This type of error suggests that native speakers of Japanese generally do not mentally organize verbs on the basis 79

of paradigms. Table 38 provides a sample contrastive paradigm of the nonce-verb ganor-u and the two actually occurring verbs suru 'do' and hanas-u 'speak'. Verb type

Nonce consonant-stem Occurring consonant-stem

Occurring irregular

ganor-u hanas-u talk suru do Non-past ganot-ta hanas-ita talked sita did Past Polite past ganor-imasita hanas-imasita talked simasita did TABLE 38. Partial contrastive paradigms of nonce and actually occurring verbs.

There is no s stem allomorph expected in the paradigm of ganor-u, although the word final sequence -sita does appear in the polite past form ganor-imas-ita. The error *ganosita might have come about through contamination from ganorimashita > *ganosita or from the high frequency irregular suru: suru:sita::ganoru:ganosita.

Connectionist modeling is based on the assumption that the type and token frequency of all derived forms is important. A network model of Japanese, which considers only two inflected forms, the non-past and the past, and dismisses related and similar sounding polite past forms ending in -masita, which also includes the past morpheme -sita, seriously underestimates the frequency of the sequence -sita. We know that Japanese speakers are influenced by similarity in phonological shape. Therefore, to only consider two forms in a Japanese model of inflection, when other items in the paradigm influence production, would not appropriately model what real speakers do in processing. The association of a single past form and the present stem form may be sufficient for modeling English, but it is not representative of the task that confronts speakers of Japanese. A better network model needs to consider typology and recognize that for many languages, more than two forms may be morphologically and phonologically (and semantically) related Thyme et al. (1994) addressed just this issue with a network model of Finnish nominal inflection. There are approximately 16 cases in Finnish with both singular and plural distinctions and a number of declensional classes with considerable allomorphic 80

variation. Not all forms within a paradigm are diagnostic for class identification, but many are. Thyme et al. chose four noun classes and four case inflections: nominative, genitive, partitive, and inessive to be modeled in their network. The lexical items used were artificial language forms patterned after actual Finnish forms. After training on simplified real forms, when the network was presented with an unknown inflected noun, it was able to output an appropriately inflected requested form. For example, when presented with the class one, nominative form satus and requested to produce the inessive form, the network appropriately output the correctly inflected form satusis. Thyme et al. (1994:464) came to the following conclusion. The model appears to have achieved some representation of paradigm classes and cell diagnosticity. It induced appropriate class knowledge without being explicitly "told" what form belongs to a

particular class. It did so by performing a task of linking similar patterns. This knowledge is subsequently used in performing the task with novel data.

This connectionist model is of interest because it appears able to perform inflection without the use of formal rules by making use of paradigmatic associations. Rather than considering only two forms as the Hayden and Klafehn (2003) model did, a better connectionist model of Japanese would need to follow this approach. A model could be trained using four or more inflected forms, for example non-past, past, negative and volitional. It would be of interest to see if the model could distinguish the paradigm of the consonant-stem verb kir-u 'cut', with forms kir-u, kit-ta, kir-anai, and kir-oo, from the paradigm of the vowel-stem verb ki-ru 'wear' with the inflected forms ki-ru, ki-ta,

ki-nai, and ki-yoo. It would be of particular interest to see if error forms bore resemblance to the errors of native speakers. The Japanese child Aki in the Miyata database (1995) has the error form *neroo for the adult form ne-yoo 'let's go to bed' of the vowel-stem verb ne-ru 'sleep, go to bed'. The child makes use of what looks like a


consonant-stem volitional suffix


with the vowel-stem verb ne-ru. Since the child

produces other inflected forms of the same verb correctly, this error does not appear to be the result of categorical rule-based processing. Assuming that the child had correctly concluded that ne-ru was a vowel-stem verb and was correctly producing ne-ta 'slept' and ne-nai 'not sleep' by the use of combinatory rule, the error form *neroo makes no sense. Where did the r in the error form come from? However, the word final sequence -roo does occur with consonant-stem verbs ending in r. For example, the verb kaer-u 'return' has the volitional form kaer-oo 'let's return'. It has been suggested that in addition to producing correct inflections, network models might also be able to model this type of error as well. The early Rumelhart and McClelland model of the English past was apparently able to make use of subtle pattern similarities and frequency considerations when it calculated *membled as the past form of mail. Rather than some type of rule-based error, this type of error, and inflectional error in Japanese as well, may emerge from the interactions of verb type and token frequencies during processing. The significance of verb type and token frequency in the representation and processing of morphology is developed even further in Bybee's (1985) Schema Model of morphology, which is considered next.


Both Langacker (1987, 2000) and Bybee (1985, 1988, 1995, 2001) have proposed models of grammar that are based on usage. In the present section we review briefly to the portion of Bybee's schema model that is most relevant to the processing of Japanese inflectional morphology. The defining characteristic of this type of network approach is the notion of the schema (Bybee 1995:430).


Sets of words having similar patterns of semantic and phonological connections reinforce one another and create emergent generalizations describable as schemas.

Bybee (1995:428) also makes it clear that her approach is different than more traditional models because it takes into account the use that speakers make of forms. Use determines both how forms are represented and how schemas emerge from representations. The basic proposal is that morphological properties of words, paradigms, and morphological patterns once described as rules emerge from associations made among related words in lexical representation.

This approach is based on the assumption that both fully inflected regular forms as well as irregular forms are represented in memory. Unlike rule models, this provision would very conveniently accommodate the typology of Japanese verbal morphology, where bare sterns do not occur in speech. Fully inflected regular forms such as non-da 'drank', and kaer-oo 'let's go back' would not need to be segmented into the unpronounceable consonant-final sterns kaer and nom. In the schema model, words entered into the lexicon are not segmented into constituent morphemes, but rather morphological properties emerge from associations among lexical items. An illustrative set of possible lexical connections for four Japanese verbs is given in Figure 1. 'ended' ///ta 'eat'





~~ ta be





FIGURE 1. Sets oflexical connections showing word-internal associations


'met' 'ate'

The lines connecting the four inflected forms in the figure (tabe-ru, tabe-ta, owatta, and at-ta) represent phonological and semantic relations. A thin line represents either a phonological or a semantic relation. For example, the thin line between the syllable wa in the form owat-ta and the vowel a in at-ta represents only a phonological relation. The lines between the syllables ta and be in the two different inflected forms (tabe-ru and tabe-ta) of the verb 'eat' are thick because they represent both a semantic and a phonological relation (two thin lines have become one thick line). When parallel sets of phonological and semantic relations occur across sets of words, higher order morphological relations result. Thick connecting lines represent the paradigmatic relationship between the word tabe-ru and the word tabe-ta. The semantic relationship between the three verbs, owat-ta 'ended', at-ta 'met' and tabe-ta 'ate', namely that they are all past forms, is shown by the heavy lines (semantic and phonological similarity) connecting the word final ta sequence in all three forms. Because the network analysis does not segment verbs into stems and endings it is able to capture significant generalizations that a generative model of morphology misses. For example, the thick line in the figure connecting the first t in owat-ta and the first t in at-ta indicates that there is both a semantic (past) and a phonological similarity. The t in these two words represents the mora obstruent allomorph. In a traditional analysis, the mora obstruent is a kind of automatic phonological patch up that occurs when a combinatory inflectional rule brings together two consonants. For example, owar + ta > owat-ta or aw + ta > at-ta. In a generative analysis, the t in owatta and the t

in atta are mere phonological similarities. The network analysis on the other hand

shows that there may be a semantic similarity as well as a phonological similarity. The mora obstruent in these words is also associated with the past meaning. There is anecdotal support for this analysis in the first language acquisition data where it appears


that young Japanese children think there are three past markers: -ta, da, and -tta. In Chapter 6 a review of past errors found in the Miyata data files show that *detta was produced in place of deki-ta 'was able to do' and *kowatta for koware-ta 'it broke'. In the schema model, the notion of lexical strength, based primarily on token frequency is seen as determining how words enter into associations and undergo change. Words with high lexical strength are easier to access and serve as bases in morphological relations. Irregular words usually show high token frequency. Irregular words with weaker lexical strength are more likely to undergo change and be regularized. This would explain why the very high frequency suppletive form dekiru 'able to do' doesn't look anything like other forms in the paradigm of suru 'do'. Irregular words like deki-ru can maintain their autonomy if their lexical strength is sufficient. This lexical strength explanation should also hold for irregular forms such as itta 'went' instead of regularized *i-ita 'went' and nai 'not have, not exist' instead of regularized *ar-anai. These forms should be supported by high token frequency. In this network analysis, productivity arises from schemas. Two different types of schemas are described. Source-oriented schemas are generalizations over pairs of basic and derived forms and are similar to generative rules that specify how to get from one form to another. Product-oriented schemas are generalizations over sets of derived forms and do not specify what operations need to apply in order to produce a form. An example of an English source-oriented schema, would be the generalization over the words walk and walked. An example of an English product-oriented schema would be the semi-productive generalization over derived words strung, stung,flung, and hung resulting when English speakers are asked to produce a past form of the nonce word spling.


In her most recent book Bybee (2001:129) has suggested that most schemas in English are probably product-oriented and then raises the question whether sourceoriented schemas exist at all. One might ask the same question about Japanese schemas. It would seem very probable that all Japanese schemas are product-oriented, working

toward some target rather than starting from some source. Recall that Batchelder's (1999) experimental subjects thought that *ganosita and not gano-tta was the past of the nonce verb ganor-u. The erroneous form *ganosita was not a random error, but was clearly influenced by some actual Japanese words ending in the sequence -sita. The Batchelder subjects who chose *ganosita over ganot-ta were apparently more influenced by product words ending in -sita than by product words ending in -tta. In this example, we could say that the -sita schema was more productive than the -tta schema.

In the Bybee model, productivity depends on two factors: the defining

properties of the schema and strength (product type frequency). Schemas that are more open and are not as restrictive in their membership are more productive. The English past-tense -ed has no phonological or lexical restrictions, can be applied to any verb, and is very productive. The sting and stung past-tense schema, on the other hand, depends on a very restricted product model with the product words ending in some type of velar: strung, flung, snuck. Highly restricted schemas like this tend to be less productive. The second consideration in productivity is the strength of the schema and Bybee relates this property directly to the type frequency of the product. In general, high type frequency is likely to be associated with greater productivity. In English it happens that the past-tense -ed schema also has high type frequency,(it is used on many different verbs) so it is not surprising that this schema is easily extended to new forms in the language (jax-ed) and used by children and adults in wug tests (rick-ed).


Finally, the schema model described here differs from rule-based models in one final way. Bybee (1995:452) makes a clear distinction between productivity, default patterns of inflection, and the definition of regularity. Bybee defines the productivity of a pattern as the ability to be applied to new items and argues that productivity is a graded property. Depending on the structure of a language there may be more than one pattern of productivity, with some patterns being more productive than others. This is clearly the case in Japanese, where the verb suru is productive for novel verb forms, but the paradigm of consonant-final r also appears to be prolific in youth slang: sutabar-u 'patronize Starbucks', makur-u 'go to McDonalds'. The past of makur-u would be

makut-ta 'went to McDonalds' . The default pattern of inflection is defined as the most open pattern of inflection with regard to lexical and phonological restrictions. This definition provides a very convenient explanation for why the most productive inflectional pattern in Japanese, contrary to the claims of the strong rule hypothesis, makes use of an irregular rather than a regular verb. The unique two-way stem vowel alternation suru 'do', sinai 'not do' is irregular because it is unlike the pattern in any other verb. (The only other irregular verb kuru 'come', ko-nai 'not come' ki-ta 'came' shows a three-way stem vowel alternation.) However, any word that is phonologically permitted in Japanese, including nouns, foreign loan-words and onomatopoeic words can form compounds with suru: benkyoo suru 'study', taipu suru 'type', betabeta suru 'be sticky'. Finally, Bybee defines regular as the pattern of inflection as with the least allomorphy in both stem and affix. On this definition the regular pattern in Japanese would have to be the vowel-stem verb pattern. This definition differs from most rule models of inflection. The Words and Rules Model (Pinker 1999), for example, claims that in a given language, the default pattern and the productive pattern of inflection will always


converge on a regular pattern of inflection. The typology of Japanese presents a challenge to this claim that can be tested in acquisition and nonce-probe experiments.

5.5. JAPANESE INFLECTION AND MODEL PREDICTIONS. In this final section we consider briefly what predictions the models we have reviewed make about productivity, acquisition, and errors in the processing of Japanese inflection. We are especially interested in investigating what this dissertation will refer to as the Strong Rule Hypothesis (SRH). The SRH is the claim that the representation and processing of verbal inflection requires the use of some type of mental rules. We will compare the Words and Rules version of the SRH with the Bybee Network model, which does not make use of conventional rules, because these two models make such dramatically contrasting predictions about inflectional processing. In the next two chapters the following three predictions are considered: 1. PRODUCTIVITY. The Rules model predicts that there will be a single

productive default rule that will produce regularized forms in acquisition and on nonce-probe tests. The Network model predicts that more than one pattern may be productive. 2. ACQUISITION. The Rules model assumes that the acquisition of inflection is determined by the learning of rules not words. The prediction that follows from this assumption is that rules are productive and automatically apply to all verbs in a category. (The Rules model also assumes that it is possible to distinguish the productive and nonproductive use of inflection.) The Japanese child who knows that the volitional form of tabe-ru 'eat' is tabe-yoo 'let's eat' should also know that the volitional form of ne-ru 'go to bed' is ne-yoo 'let's go to bed'. The network model assumes that the acquisition of inflection is determined by the learning of individual


words (and longer chunks) and that schemas emerge from generalizations across words (and chunks), not categorical rules. The prediction that follows from this assumption is that a Japanese child who produces tabe-yoo 'let's eat' may not be able to produce ne-yoo 'let's sleep'.

3. ERRORS. The Rules Model predicts overregularization in acquisition and that errors are more likely to occur with irregular forms than with regular forms. The Network Model predicts that regulars are just as likely to be involved in errors as irregulars.



6.1. INTRODUCTION: In this chapter we begin with a brief comparison of English and Japanese acquisition orders. This comparison suggests that typology is important in how verbal inflection is processed and results in differences in acquisition. In following sections the results from detailed database searches of the longitudinal records of three young Japanese children is reported. Each database was searched for evidence of verb errors and overregularization. A comparison of regular and irregular verb errors is reported. In the final section, error forms inflected for the volitional (-00) are evaluated for evidence of default rule application.


There is a long tradition and an ample literature of acquisition studies dealing with English. Most non-specialists are more likely to be familiar with general assumptions about acquisition that are based on these studies involving English-speaking children than with research dealing with different language typologies. The purpose of this dissertation is not to perform a contrastive analysis of the acquisition of English and Japanese verbal morphology. However, as a starting point it is of interest to compare just how much more inflectional morphology must be acquired by Japanese-speaking children than by English speaking children. Much of what is accomplished by syntax in English (modals and auxiliaries, for example) gets taken care of by inflectional suffixation in Japanese. A SRH claims that inflectional typology is not significant in processing and acquisition. Regardless of typology, rules are acquired. In this chapter we will demonstrate that this claim is almost certainly not true. Typology matters. The


following short section will suggest how the agglutinating verbal typology of Japanese might require processing mechanisms other than those represented as formal rules.


The relative order of the acquisition of English verbal inflection reported by Brown (1973:274) is often referred to in first and second language acquisition studies. 1. Present progressive

is playing, is singing

2. Past irregular 3. Past regular 4. Third person regular 5. Third person irregular 6. Uncontractible auxiliary 7. Contractible auxiliary

came, went, fell, saw, hurt played, washed, wanted sees, wants, washes does, has She was eating He's crying

TABLE 39. Order of acquisition by English speaking children.

Compared to Japanese, the most remarkable aspect of the English acquisition order is how little verbal inflection there is to be learned. There are only seven items on the list. Furthermore, the last two items on the list, uncontractible and contractible auxiliaries, do not involve additional inflectional morphemes beyond what is necessary for items one through five. However, since four out of seven items on the list do depend on the regular/irregular distinction (the progressive suffix ing is the same for both regulars and irregulars) it is clear why the notion of regularity has historically been so important in the description of English verbal morphology and in the debate over models of processing. Items six and seven in the table also demonstrate why a rule-based account of processing is so compelling for English. The difference between an uncontracted and a contracted auxiliary form is very neatly represented by a contraction rule, which derives He's crying from He is crying. No abstract underlying representation is required. Both forms actually occur in speech and the transparent relationship between 91

the base form and the derived form appears obvious. This kind of rule which deletes a single element from one form to derive another form, is a very elegant and economical way to describe the inflectional typology of English.

6.2.2. JAPANESE ACQUISITION ORDER. Detailed acquisition orders for Japanese have also been reported (Clancy, 1985), but have received much less attention in the literature concerned with the processing of inflection. Clancy divides grammatical development, including the acquisition of inflected forms, into three stages. Table 40 is based on the text description of acquisition order that appears in Clancy (1985:381-382). The order in which the suffixes are listed within each stage does not represent the order of acquisition within that stage. As Clancy (1985:399) herself points out, compared with many other languages, the Japanese inflectional system is not exceptionally complex and not many errors are reported during acquisition. However, the contrast between the three English inflectional morphemes (-ing, -ed, -s,) and the eighteen Japanese inflectional morphemes that appear in Table 40 is striking. Also striking is the fact that Japanese acquisition researchers, who may be more interested in when and what children do rather than how they do it, do not find the distinction between regular and irregular verbs to be a salient feature in child acquisition. Notice that no mention of regular and irregular forms is included in Table 40. Clancy (1985:427) is almost dismissive of this categorization. Furthermore, all verbs are inflected regularly in Japanese, even the most common ones, such as suru 'do' or the copula da. Stem changes do occur with the different inflections, and for some

verbs these changes are irregular, but the inflectional ending does not change.


Forms that appear sometime during the first stage of grammatical development (before age 2) Imperative -te tabete eat it Past -ta tabeta ate Forms that appear sometime during the second stage of grammatical development (about age 2) Present-progressive/resultative in -teru tabeteru eating/is eaten Non-past -ru taberu eat Completed past in -tyatta tabetyatta finished eating, ate all Non-past negative -nai tabenai not eat Desiderative -tai tabetai want to eat Forms that appear sometime during the third stage of grammatical development (about age 3-3;6) Non-past -tyau tabetyau will finish eating, will eat all Volitional-yoo tabeyoo let's eat Past progressive/resultative -teta tabeteta was eating Potentials in -eru tabereru able to eat tabemasu eat Polite form -masu tabemasyoo lets eat Polite form -masyoo Passive -rareru taberareru be eaten Causative -saseru tabesaseru make/let eat Obligation -nakya tabenakya must eat During -nagara tabenagara while eating Provisional -tara tabetara if eat TABLE 40. Stages of Japanese grammatical development based on Clancy (1985).

The distinction between regulars and irregulars has figured so prominently in the debate over models of processing and acquisition (including V-shaped learning) because this distinction accounts for errors in such a transparent way. It is claimed that English speaking children overregularize, making regular forms out of what should be irregular forms. Children say *goed instead of went. The same type of overregularization error either does not occur in Japanese, or else the Japanese acquisition researchers have not yet reported it. A SRH claims such errors should occur. Both Pinker (1991, 1999) and Marcus et al. (1992,2001) have claimed that the existence of innate symbolic rules that depend on the distinction between regular and irregular are independent of language typology because they reflect "how the mind works". In the sections that follow we report on a detailed database search for examples of inflectional error in Japanese, including overregularization.



In the following sections we investigate the early acquisition of Japanese verbal morphology by making use of the archived longitudinal data files of three Japanese children between the ages of one-and-a half and three years. Miyata (1992, 1993, 1995, 2000) obtained the spontaneous play speech of the three children, their mothers and occasionally other family members during weekly recording sessions. Miyata has transcribed each utterance and added a tier of complete morphological analysis that can be computer searched by using the CHILDES system (MacWhinney 1995, Oshima-Takane & MacWhinney 1998). Since the morphological tier categorizes each verb according to type and inflectional ending and includes a consistent method of marking child errors, a very detailed analysis of each child's production is possible. The data files for the child Ryo consist of transcriptions of about thirty hours of recorded speech. The transcriptions in the data files for the child Aki and the child Tai each represent about fifty hours of recorded speech.


The English verb go has the irregular past form went. The fact that English speaking children pass through a stage of linguistic development where they produce non-adult *go-ed either in place of, or alternating with adult went has been used to argue in favor of a SRH in first language acquisition. This hypothesis claims that overregularized forms like *go-ed, and less frequently *went-ed, result because children are making productive use of the rule: past

= stem + ed.

According to

Pinker's (1984, Pinker and Prince 1994) version of this hypothesis, when an irregular past form like went is retrieved from memory, application of the regular rule is blocked. However, when there is a failure to retrieve the irregular form from memory,


the rule applies and produces an overregularized error form go + ed > *go-ed or went

+ ed > *went-ed. As noted earlier, a very interesting coincidence occurs in Japanese. The regular Japanese verb ik-u 'go' also has an irregular past form. The past form is iUa 'went' and not *i-ita, the form predicted by the productive rule for the past. If the English past rule is past could be past

= stem + ed,

the Japanese past rule, for verbs with a stem-final k,

= stem minus k + ita. The past form of ik-u predicted by this past rule

should therefore be ik minus k> i + ita> *i-ita. A partial paradigm of some stem-final k verbs appears in Table 41. Non-past


kak-u write ka-ita wrote kik-u ask, listen ki-ita asked, listened tuk-u be attached tu-ita was attached ik-u go *i-ita > itta went TABLE 41. Partial paradigms of regular stem-final k verbs with irregular form itta.

The form predicted by the rule (*i-ita) does not actually occur in the paradigm of ik-u. However, we might consider *i-ita to be the Japanese equivalent of English *goed, a rule generated, overregularized form.

Marcus et al. (1992:145-147) working with data from Kuczaj (1976, 1977) and Brown (1973) report 75 overregularized tokens of *goed and *wented and 328 tokens of went in the archived files of the English speaking children Adam, Eve, Sarah and Abe. A summary of the English data is presented in Table 42a. A search of the Miyata Japanese files for the children Aki, Ryo, and Tai retrieved 147 tokens of the correct irregular past form itta , but no tokens of *i-ita, the overregularized form predicted by the rule for the past. A summary of the Japanese data is presented in Table 42b.


Child Age Form Tokens Errors Tokens Error rate 125 none 0 0 Adam 2;6-5;0 went Eve 1;6-2;3 went 20 *goed 5 .25 Sarah 2;3-5;1 went 66 *goed 6 .09 117 *goed (60), *wented (4) 64 .55 Abe 2;6-5;0 went Total 328 75 .23 TABLE 42a. Overregularization of English go (data from Marcus et al. 1992). Child Age Form Tokens Errors Tokens Error Rate Aki 1;3-3;0 itta 75 none 0 0 31 none 0 0 Ryo 1;5-3;0 itta Tai 1;5-3;1 itta 41 *ikita 1 .02 Total 147 1 .0068 TABLE 42b. Overregularization of Japanese iku 'go' (data from Miyata 1992, 1993, 1995, 2000).

A parallel further search for related irregular forms in the paradigm of ik-u was conducted. As noted by Clancy (1985:383) one of the first inflected forms to become productive for Japanese children is the non-tensed gerund form ending in -teo This form has early use as an imperative when used alone itte '(please) go', and gets high frequency use in a number of V-te V constructions. There is early use of the progressive/resultative construction itte kuru 'go and come (back)'. The conditional ittara 'if/when go' and the representational form it-tari 'do things like go' are also

irregular forms of ik-u. If ik-u patterned like all other regular Japanese verbs, the gerund form would be *ii-te, the conditional would be *ii-tara, and the representational form would be *ii-tari. The search results for all these forms appear in Table 43. No overregularized error forms were recovered. Child


Tokens Errors

Error Rate

itte 62 0 0 Aki itte 48 0 0 Ryo itte 279 0 0 Tai Total 389 TABLE 43. Gerund forms of iku 'go' (data from Miyata 1992, 1993, 1995,2000).


The overregularization predicted by a SRH was not found in the Japanese database. Thus, there is no Japanese equivalent of English *goed found in the files for Aki, Ryo, and Tai.


English speaking children sometimes produce the overregularized form *do-ed in place of did and the overregularized forms *com-ed, or more rarely *cam-ed, in place of the adult form came. Since the two Japanese verbs kuru 'come' and suru 'do' are also irregular, we might expect them to be overregularized as well. A SRH hypothesis, which also requires stem computation, admits the possibility of stem and ending mismatches and predicts that we might find a Japanese *do-ed, which would have the form *su-ta rather than the adult form si-ta 'did'. We might also find a Japanese equivalent of *com-ed, which might have the form *ko-ta or *ku-ta rather than the adult form ki-ta 'came'. Table 44 provides data from Marcus et al. (1995) showing that even for English the overregularization rates of do and come typically tend to be low, although there can be considerable variation from child to child. Child

did doed






0 51 .01 .01 1 Adam 93 1 .22 0 .15 23 4 Eve 7 2 1 15 .21 127 0 3 Sarah 0 1 .09 .01 Totals 227 3 89 8 TABLE 44. Correct and error tokens of do and come (Marcus et al. 1995).

Table 45 provides a summary of all irregular verb form tokens and errors found in the Miyata files. The 257 tokens recorded under suru include all inflected forms of that verb including, for example, the imperative site 'do', the past sita 'did', and the continuative /resultative construction siteru 'is doing/is done'.


kuru Errors Rate Child suru Errors Rate .04 2 Aki 257 136 .02 10 0 40 0 88 0 0 Ryo .02 Tai 627 316 3 .009 9 .02 492 19 .01 5 Totals 972 TABLE 45. Total correct tokens and error tokens of irregular suru, kuru.

Table 46 demonstrates that the unpredictable stem-vowel alternation in the paradigm of suru accounts for only 9 of 19 error tokens. Child

Stem-vowel error tokens


sen (1) for sun, siru (5) for suru 6 Aki Ryo none found Tai *sinande (2) for sende, *siru (1) for suru 3 Total 9 TABLE 46. Stem-vowel error tokens in paradigm of suru.

A few notes about child and dialect verb forms are in order here. The three children in Miyata's study are from Nagoya and speak a local Nagoya dialect. The number of dialect verb forms produced by the children is very small, although a few forms do appear. Miyata is careful in her transcription to indicate the appropriate dialect target whenever it appears that a dialect form was the child's intention. For example, the standard construction suru no 'do' is contracted to the form sun and the standard form sinai 'not do' has the alternate dialect form sen. The standard construction sinai de 'without doing' therefore becomes sende in Nagoya (for both adults and children). These three error tokens (sinande (2) and sen

(1» may reflect

dialect contamination. If that is the case, rather than an indication of confusion over the alternation in the standard dialect stem vowel, the errors in Table 46 should be interpreted as lexical rather that morphological. No tokens of the predicted overregularized forms *su-ta, *ku-ta, or *ko-ta were recovered. Out of a total of 492 verb tokens in the paradigm of kuru, there was only one error token that could be characterized as a stem vowel error: *koreba for kureba 'if come' does get the stem vowel wrong. 98

The five tokens of *siru, where the adult target would be su-ru 'do', could also be interpreted as a stem-vowel error, where the expected stem-vowel i is regularized to u. Is this error predictable? The SRH says only that patterns of overregularization do not depend on token frequency. The schema model on the other hand makes strong claims for the importance of token frequency. Table 47 provides a list of all inflected forms of the verb suru produced by the child Aki. Stem vowel Most frequent tokens

Total number of tokens

sita (42), siteru (42), siyoo (39), site (30), sitara (9), *siru (5) 204 62 u suru (62) 1 e sen (1) (dialect form) 267 Total TABLE 47. Stem type frequency based on tokens of suru paradigm produced by Aki.

The number of inflected forms with stem vowel i are more than three times the number of forms with stem-vowel u. The schema prediction that frequency determines productivity is confirmed. The total number of verb tokens considered in the English data was 404 including 68 overregularized forms yielding an overregularization rate of .16. Data from the notorious overregularizer Abe has been the source of continued controversy in the acquisition literature. If the Abe data is excluded from consideration, the other three children alone produced a total of 328 tokens including 12 error tokens. Adam and Eve and Sarah have an overregularization rate of .037 with the two verbs do and come. In the much larger Japanese database the total of 1,488 suru and kuru tokens include only 8 error tokens comparable to English overregularization. The resulting .005 rate of overregularization is not a compelling argument in favor of a model that predicts overregularization. There is no consistent pattern of error and there is no Japanese equivalent to English *doed and *comed.



Irregular forms also occur in the paradigm of ar-u 'have, be'. Ar-u has the irregular negative form nai 'not have, not be' . Contrastive partial paradigms of regular stem-final r verbs and irregular nai are given in Table 48. Non-past



wakar-ru understand wakar-anai not understand wakar-ana-katta kaer-u return kaer-anai not return kaer-ana-katta ar-u have, be nai not have, not be nakatta TABLE 48. Partial paradigms of regular stem-final r verbs and irregular forms

didn't understand didn't return didn't have, wasn't of aru.

The irregular form nai 'not have, not be' is an exception to the negative rule for consonant-stem verbs: negative =stem + -anai. The expected negative form of ar-u is *ar-anai, which does not occur. A strong rule hypothesis predicts that children would produce the overregularized form *ar-anai and also the overregularized past form *arana-katta 'didn't have, wasn't'. A summary of correct and error tokens retrieved from the files is presented in Table 49. Child Form Tokens Aki nai Tai nai Ryo nai Totals

323 385 434 1,142

Errors none none *aranai, *arenai, *aronai

Tokens Rate 0 0 0 0 3 .0069 3 .0026

Form nakatta nakatta nakatta

Tokens Errors 3 9 2 14

none none none

Rate 0 0 0 0

TABLE 49. Correct and error tokens of irregular nai and nakatta.

Aki, Tai, and Ryo produced 14 correct tokens of irregular nakatta 'didn't have, wasn't' but no tokens of overregularized *ar-anakatta. They produced 1,142 tokens of irregular nai but only five error tokens. Ryo produced one overregularized *ar-anai and the two innovative forms *arenai, and *aronai. The innovative error forms are of interest because they raise the issue of how Japanese error forms should be analyzed. The question is whether these forms should be considered to include a correct stem form or not. There are only two variants of the non-past negative ending. Vowel-stem verbs add -nai to the stem: tabe- + nai > tabe-nai 'not eat'. Consonant-stem verbs add 100

-anai to the stem: wakar- + anai > wakar-anai 'not understand'. Do the error forms *arenai and *aronai represent the correct stem analysis with incorrect endings: ar + *-onai and ar + *-enai, or incorrect stem analysis and correct endings: *aro + -nai and *are + -nai?

The fact that the child Ryo produced all three of these variations presents a problem for a SRH. A SRH predicts that a single productive rule will operate to produce a single default pattern when an irregular form cannot be accessed in memory. Even when two different outputs result in English overregularization, for example *goed and *wented, it is argued that a single rule is sufficient to account for the process. It appears transparent that the source of *goed is the word go and the source of *wented is the word went. However, in the case of Japanese error forms *aranai, *aronai, and *arenai, while there are the forms are-ba and ar-oo, these forms have a much more limited distribution and frequency and do not appear in the database as words produced by the children themselves. Therefore, there is no legitimate single input for a generative rule (a source-oriented schema). However, if one believes that fully inflected regular forms are located in the lexicon, the variation in error form can be interpreted as the result of a product-oriented schema. A productoriented schema would emerge from the part of the lexicon with the strongest lexical representation, the inflectional portion of the verb, the -nai part. Many verbs, all consonant-stem verbs, share a word final negative -anai sequence. There are many other verbs, (all stem-final e verbs) that share a word-final negative -enai sequence. Even the form *aronai satisfies the minimal product condition that the form of ar-u that cannot be located in memory be a negative form (ending in the sequence -nai). The fact that three alternate error forms result in moving from some lexical form of aru to the specific form nai is consistent with the claim that schema strength depends


on task frequency. For a young child, a low frequency task, moving from some other form of ar-u (at-te, at-ta, ar-imasen, ar-imasu) to nai when nai has not yet developed sufficient lexical strength is just as likely to arrive at the product *arenai, as *aronai or *aranai. There is no pattern. There is no regularity. There is no evidence of a rule. For Aki, Ryo and Tai, 1,172 attempts at irregular forms in the paradigm of aru (nai, nakatta, naku, and nakutya) produced a total of three errors. Thus, 99.8% of the time, attempts at these irregular forms yielded the appropriate adult form. These results can be combined with the results from Table 49 to provide a comprehensive summary of all irregular verbs and irregular forms. Aki, Ryo and Tai made a total of 4,388 attempts at all irregular forms and verbs and produced only 35 error forms. (99.9% of attempts produced the appropriate irregular form). Japanese irregular forms do not present the distinctive acquisition problem that is predicted by a SRH. The fact that so few errors are discovered is consistent with the schema claim that fully inflected forms are stored in memory. Although a few errors are found, Japanese children do not overregularize irregular nai.


There is a pattern of error associated with the irregular verb form nai not found in the Miyata files, but well documented elsewhere. A number of investigators (Clancy 1985, Fujiwara 1977, Ito 1976, Sano 1995) report that some children have difficulty with negation. Kanagy (1991) has reported a similar error pattern among some adult second language learners. These children and learners attempt to make the negative form of verbs by adding the suffix -nai to already inflected verbs. Examples reported in the literature appear in Table 50. It could be argued that these examples are counterexamples to the claim made in this dissertation that Japanese children do not make use


of combinatory rules. In order to address this challenge it is necessary to clarify the definition of a rule and distinguish overregularization from overgeneralization. Error form reported


*tabe-ru nai tabe-nai I won't eat *deki-ta nai deki-nakatta I couldn't do it *omo-u nai omow-anai I don't think so *it-te nai ir-anai I don't need it *ar-u nai nai I don't haveitJIt isn't here/there *at-ta nai na-katta I didn't have it/It wasn't here/there TABLE 50. Use of nai to form the negative of inflected verbs.

Matthews (1991: 130) relates the notion of regularity to predictability. To say that a form has regular inflection is to say that it has the inflection one would expect unless one knew that it was different.

Implicit in this definition is the notion that one knows where one is starting from and also has a highly specified procedure (a rule) for getting to the goal. Pinker (1999:472) specifies how a single -ed past pattern operates as the productive pattern for English. Regular inflection, being fully predictable, is computed by a rule that concatenates the affix -d to the verb stem.

This particular formulation of the past rule is consistent with the traditional generative version of a rule where both the structural description of an input (a stern) and the structural change (add -ed) must be clearly specified. Overregularization is generally defined as the application of a regular rule to an irregular form where the rule does not normally apply. In English *go-ed is an overregularized form of go. In a previous section we suggested that the Japanese equivalent of an overregularized form would be *ii-ta in place of irregular itta. Overgeneralization is quite different. Overgeneralization is defined as extending the meaning of a word too broadly. A small English-speaking child who calls any man papa has overgeneralized the meaning of the word. It is also recognized (Clancy


1985, Rispoli 1981) that some Japanese children overgeneralize the meaning of the gerund form -te early in acquisition. In addition to using the -te form as an imperative,

it appears that these children use the non-tensed -te form in place of other tensed forms. Rispoli (1981) reports that 20% of the -te forms produced by one 18 to 20 month old child he studied were not requests. Clancy (1985:384) reports that one of her subject children (age 2;2) responded to a question about what was inside a noisemaking toy with the single word hait-te. The full adult response would make use of the tensed auxiliary expression haitte iru 'it is inside' or the contracted form haitteru. Rather than overregularizations, the error examples in Table 50 are better described as overgeneralizations. The narrow meaning of the word nai 'not have or not be located in a place' is extended too broadly to include the notion of 'not'. Furthermore, the operation that combines nai with something else cannot be described as a rule since it does not specify a structural description. In some cases nai combines with a non-past form, in other cases nai combines with a past form, and in one case

nai combines with a -te form. The first four examples also fail as examples of overregularization because the verbs tabe-ru 'eat', deki-ru 'able to do', omo-u 'think', and ir-u 'need' are all perfectly regular verbs by any formal definition of regularity. All of the forms combined with nai are legitimate word forms by themselves, not the stem forms that Pinker would require in a rule. Rather than rule-generated forms these error forms would seem to be the result of a very basic combinatory strategy required because bare stems are not available in the input. Furthermore, as we pointed out in previous sections, the bare stems omow-, ar-, and ir- are not pronounceable forms. An alternative explanation for the error forms is that fully inflected forms are available in memory and that the highest frequency forms, the non-past, the past, and the -te form have the strongest lexical representation and are therefore the easiest to


access in lexical schemas. When all else fails children will make use of stored whole words rather than default rules.


A SRH account of inflection requires that the stems of regular verbs be isolated from inflectional endings. This dissertation assumes that segmentation is not easily performed in acquisition because the bare stems of Japanese consonant verbs do not appear in isolation and are not pronounceable. Stems would need to be computed across paradigms where there is much variation. To give some idea of the difficulty of the task, the partial paradigms of some unsegmented consonant-stem verb forms and their stems are given in Table 51. Non-past Negative



kaeru nomu kaku oyogu matu au hanasu

kaetta nonda kaita oyoida matta atta hanasita


kaeranai nomanai kakanai oyoganai matanai awanai hanasanai

return, go back drink write swim wait meet speak

TABLE 51. Partial paradigms of representative consonant-stem verbs.

English speaking children hear the forms walking, walked, and the bare stem walk. Japanese children hear aruite iru > aruiteru 'walking' and aruita 'walked', but never the bare stem *aruk. Japanese has a strict phonotactic constraint against consonantfinal syllables and it seems unlikely that native speakers segment and store the stems of verbs in shapes that they cannot pronounce (Hale 1973). Japanese children don't hear bare stems and they don't produce them. With the exception of the mora nasal, no example of a consonant-final bare stem was recovered from 19,171 verb tokens produced by Aki, Ryo, and Tai in the database.


In order to determine some measure of the relative difficulty of segmentation into stem and ending, a search was made of all coded verb errors in the database. As previously noted, it is often difficult to determine whether a given error represents a stem error, an ending error, or both. In this analysis of verb errors, the following criterion was used. If an error token included the minimal correct stem of the target form, the stem was judged to be correct, even if additional segments appeared before legitimate affixes. Each error token was compared to the intended adult form as indicated on the morphological tier of the coded database. For example, the child error form *mii for the adult form mi-ru 'see' was judged to be a correct stem form because it included the minimal stem form mi-. The child error form *neroo for the adult form ne-yoo 'let's go to bed' was judged to contain the minimal correct stem ne-. Verbs like ne-ru 'go to bed, sleep' are of particular interest because the high frequency non-past

form has potential analysis as either a vowel-stem or as a consonant-stem verb. The form neru might represent either ne-ru or ner-u. On this basis, the child error form *neroo might represent the stem misanalysis *ner-oo. However, the generous minimal

stem criterion used here did not count error tokens like *neroo as examples of stem error. The results of the regular verb error analysis are recorded in Table 52. For each child there were more than twice as many errors when the target was a consonant-stem verb than when the target was a vowel-stem verb. Child Aki Ryo Tai Total

Error Tokens 219 42 249 510

Stem-errors Consonant-stem errors 164 21 148 333

Vowel -stem errors

112 16 112 240

52 5 36 93

TABLE 52. Total number of verbal error tokens and errors by stem type.

Aki produced a total of 219 error tokens where the adult target was a regular verb form. Of these, 164 tokens or 74% were stem errors. Ryo produced 42 error tokens 106

where the adult target was a regular verb form. Of these, 21 tokens or 50% were stem errors. Tai produced 249 error tokens where the adult target was a regular verb form. Of these, 148 tokens or 59% were stem errors. The conclusion from this analysis is that young children find the regular verb allomorphy at the stem boundary to be opaque. While the children produce most forms correctly, the errors that do occur are not the regularization errors found in English. There does not appear to be a rule-like pattern of error. Instead, errors seem to be segmentation errors at opaque boundaries. Table 53 provides lists of the highest frequency error tokens for each child. Notice that there are relatively few high frequency items. Errors tend to look like blends of different forms in the same lemma or phonologically similar forms. Tokens Error

Adult target

*detta deki-ta *hakonnai hakob-anai kabur-anai 4 *kabunai naos-ite 4 *naotte tore-nai 4 *tonnai minai 4 *miinai yaranai 3 *yainai kowareta 3 *kowatta neyoo 3 *neroo asonderu 2 *asonde Most frequent error tokens: Aki.

23 5

Tokens Error was able 2 *aiku not carry 2 *hakobote not wear 2 *nigisite repair 2 *iremasu not get 2 *warete not see 1 *akita not do 1 *asobokunai broke 1 *hairoreteru let's sleep 1 *irimimasen is playing 1 *kaettekonai Most frequent error tokens: Ryo.

Adult target Tokens Error tukaw-anai not use 11 *tukanai kutttuk-u stick to 4 *kuttsuu 4 *tunanai tunagar-anai not connect 4 *tunaranai tunagar-anai not connect 3 *kattematta katte morat-ta had one buy please follow 3 *tuitekore tuitekur-e 3 *kattematta kattemorat-ta had one buy kuttuk-anai not stick 3 *kutinai nor-u become 3 *noriru 3 *tunaganai tunagar-anai not connect Most frequent error tokens: Tai TABLE 53. Most frequent error tokens for each child.


Adult target ak-u hakon-de nigas-ite ir-imasu wat-te a-ita

open carry let loose need break opened asob-itakunai don't want to play hait-teru is inside ir-imasen don't need kaettekure-nai doesn't return


A SRH predicts that regular verbs should be less problematic for learners than irregular verbs. Once a rule is learned it should automatically apply whenever it can. Furthermore, while it has been argued that speakers may sometimes forget irregular forms, no one has suggested that speakers forget rules. Irregular verbs on the other hand, depending on token frequency might be more difficult to learn, more easily forgotten, and more likely to be overregularized. In order to test this hypothesis, a count was made of all regular and irregular verb tokens in the database. Regular and irregular error tokens were also retrieved and counted. All three children produced more tokens of regular than irregular verbs. A summary of this analysis is provided in Table 54. Irregular tokens

Regular tokens

Total Irregular errors

3,489 4,349 13 Aki 860 2,885 3,530 4 Ryo 645 1,691 9,601 11,292 14 Tai 15,975 19,171 31 Total 3,196 TABLE 54. Total number of correct regular and irregular tokens

Rate Regular errors .015 224 46 .006 .008 256 .0097 526 and error tokens.

Rate .064 .016 .026 .027

Once again, the verbs and verb forms considered to be irregular included all forms of irregular nai 'not have, not be', (na-katta 'didn't have, wasn't') all irregular forms of ik-u 'go' (it-ta 'went', it-tara 'if go'), and all forms in the paradigms of irregular suru 'do' and irregular kuru 'come'. The verb deki-ru 'able to do, finish', which can be considered a suppletive form of suru, was coded by Miyata as a regular vowel-stem verb and that is the way it was counted here. All forms of the copula desu were excluded from the count. The Japanese copula is verb-like and it does exhibit some interesting alternations of form, da and de and na, but it lacks the complete paradigm exhibited by other verbs. The copula has no imperative, for example. Furthermore, since there is a unique copula paradigm the copula would have to be considered


irregular. A preliminary count of the copula and copula error tokens revealed an extremely low error rate. The three children produced 3,323 correct tokens of the copula, but only 10 error tokens. Inclusion of the copula data would have skewed the regular/irregular error frequency even more. For Ryo the error rate of regulars was more than twice the error rate of irregulars. For Tai the error rate of regulars was more than three times the error rate of regulars. For Aki the error rate of regulars was more than four times the error rate of irregulars. The predictable regular verb forms were clearly more problematic than irregular forms

6.3.7. ERROR PATTERNS. It is difficult to find a pattern of verb error in the data. Tables 55, 56, and 57

present partial lists of error types from the database. The tables are arranged by word type frequency. The complete list of each child's error types is much longer because most error types occurred only once. For example, in the table showing Aki's error forms, only the five most frequent types are shown. There was also a small number of errors that occurred three or two times. Of Aki's 224 regular verb error tokens, 114 tokens, or 51 %, were errors that occurred only once. Of Ryo' s 44 regular error tokens, 27 tokens, or 61 % were errors that occurred only once. Of Tai' s 252 regular verb error tokens, 140 tokens, or 56% were errors that occurred only once. The adult targets that appear are those provided by Miyata in the morphological tier of the database. A question mark indicates an uncertain target. There are few higher frequency error tokens. Aki does have 23 tokens of *detta for

deki-ta 'was able to do, finished', but this type of error appears to be a highly lexicalized, idiolectal form, rather than a pattern across verbs of a similar type. The error form *detta alternated with the correct form deki-ta. Aki persisted with *detta,


even while correctly producing tokens of similarly conjugating vowel-stem verbs: mita 'saw', i-ta 'was, was located', oti-ta 'fell', and si-ta 'did'. Frequency Tokens Child error *detta 1 23 *mii 2 13 *miita 3 9 *miita 4 5 *hakkonai 4 5 *miinai 4 5 *ason 4 5 4 *kabunai 5 *naotte 4 5 4 *tonnai 5 TABLE 55. Regular verb errors: Aki.

Adult fonn Frequency Tokens Child error deki-ta 1 *kuttukutta 3 mi-ru 2 2 *aiku mie-ta 2 2 *attearu mi-ta 2 2 *hakobote 2 hakob-anai 2 *nigisite mie-nai 2 2 *irimasu ason-de 2 2 *warete kabur-anai 1 3 *akita naos-ite 1 3 *asobokonai tore-nai 1 3 *hairoreteru TABLE 56. Regular verb errors: Ryo.

Tokens Child error 11 *tukanai *simarimasita 9 *yarerenai 5 *tunagenai 5 4 *kuttuu 4 *noriru 3 5 *tunaganai 3 5 *kattematta 3 5 *kutinai 3 5 *kazoo 3 5 TABLE 57. Regular verb errors: Tai.

Adultfonn tukaw-anai sima-imasita yar-anai? tunagere-nai kuttuk-u nore-ru?/nos-eru? tunagar-anai kattemorat-ta kuttuk-anai katazuke-yoo

Frequency 1 2 3 3

Adultfonn kuttu-ita ak-u atte-ru hakon-de nugas-ite ire-masu wat-te ai-ta asob-itakunai hait-teru

There are few high frequency error tokens. There is no persistent pattern of canonical error similar to what happens in English where the no change verbs (hit, put, sit, hold) are replaced by overregularized forms (*hitted, *putted, *sitted, *holded).

Aki's other higher frequency error, *mii for the adult form mi-ru 'look at, see, watch' looks like an early phonological simplification, however, as with his use of *detta, Aki persisted with this idiolectal form even when he was clearly articulating

similarly inflecting vowel-stem forms like i-ru 'be, be located'. There are three possible explanations. Perhaps Aki is just using his personal variant of a lexical item as young children often do. He can pronounce the word form 110

deki-ta, but chooses not to use it. The second possibility is that the pattern for the past is not yet productive and that similar forms that are produced correctly are simply memorized forms. A third possibility is the one suggested by Tomasello (1992:23) that the acquisition of verb patterns is not determined by the acquisition of rules. I call this the Verb Island hypothesis. It can be stated more specifically as follows: Until proved otherwise, we should assume that young children's early verbs and relational terms are individual islands of organization in an otherwise unorganized grammatical system. In the early stages the child learns about arguments and syntactic marking on a verb-by-verb basis, and ordering patterns and morphological markers learned for one verb do not immediately generalize to other verbs. The reason for this is that nascent language learners do not have any adult-like syntactic categories or rules, nor do they have any kind of word class of verbs that would support generalizations across verbs.

What young children do have and what they appear to make use of is memory. The relatively few error forms that are produced look more like approximations at related forms of similar shape and meaning, i.e. product-oriented schemas, rather than regularized forms. This raises an interesting question. It has been argued that relatively young English-speaking children pass through a stage of development where memory is sufficient to distinguish regulars and irregulars. The youngest children in the earliest developmental stage are able to produce go and went appropriately by the use of memory alone. It is only later with the implementation of rule-based processing that errors begin. If this is true, one has to ask why very young Japanese children who can apparently manage a much more complicated system of inflectional paradigms from memory would ever want to change to rule-based processing? Young Japanese-speaking children, just like Tomasello's young English-speaking children do not appear to make use of the same morphological categories that adult


linguists use to describe the language. Japanese children do not demonstrate error patterns consistent with implicit knowledge of verb stems and verb suffixes. In Chapter 7 we will demonstrate that the performance of Japanese adults on nonce probe tests is also consistent with product-oriented processing that does not require the traditional category of stem and suffix.


A SRH argues that regularity and productivity will converge on a default pattern of inflection and that a single default pattern should be apparent in inflectional error. In this section we look for evidence of a Japanese default form. In previous sections we attempted to show that the evaluation of Japanese children's error forms is problematic because it is not always possible to determine which portion of an error form represents a stem and which portion represents an inflection. However, the distinction between Japanese vowel-stem verbs and consonant-stem verbs is clearly marked in the volitional form because there are only two alternating forms. Consonant stem verbs add


to the stem; the volitional form of kaer-u

'return' is kaer-oo 'let's return'. Vowel-stem verbs add -yoo to the stem; the volitional form of tabe-ru 'eat' is tabe-yoo 'let's eat'. All three children produced a considerable number of correctly inflected volitional forms. Table 58 provides the total number of correct volitional verb types and tokens produced by each child along with the three most frequent verb types.



Tokens Types


Three most frequent correct verb types with number of tokens

ik-oo 'lets go' (39), si-yoo 'let's do' (38), yar-oo 'lets do' (29) 251 43 15 Aki yar-oo 'let's do' (18), asob-oo 'let's play' (6), ik-oo (5) 'let's go' Ryo 45 15 0 si-yoo 'let's do' (151), yar-oo 'let's do' (115), ik-oo (74) 'let's go' Tai 641 118 18 Total 937 176 33 TABLE 58. Number of volitional tokens with three most frequent verb types.

Tokens Error *neroo 3 *kaiyoo 2 *kaeyoo 2 *daseyoo 1 *kakusiyoo 1 *oitaroo 1 *osiyoo I *toyoo 1 *miitoo 1 *mitoo 1 *tukeroo 1 15

Target ne-yoo ka-oo kaer-oo das-oo kakus-oo ok-oo os-oo tor-oo mi-yoo mi-yoo tuke-yoo

Error type V>C C>V C>V C>V C>V C>*C C>V C>V V>C V>C V>C

TABLE 59a. All volitional errors: Aki.

Tokens Error Target *kazoo katazuke-yoo 3 *kattematteyoo kattemora-oo 2 *yattoo yar-oo 2 *kakusiyoo kakus-oo 1 *hawoo hair-oo I 1 *noboo nobor-oo 1 *oitokooko oiteok-oo tukur-oo 1 *tukuraoo *kuroo ko-yoo 1 I *tunagoroo tunage-yoo *hamoo hame-yoo 1 *mitoo mi-yoo 1 *kaiyoo not clear 1 *kazueru 1 katazuke-yoo 18 TABLE 59b. All volitional errors: Tai.

Error type V>C C>V C>*C C>V C>*C C>*C C>? C>*C V>C V>C V>C1 V>C V>0

Tables 59a and 59b show that the distinction between vowel-stem verbs and consonant-stem verbs was not clear to Aki and Tai. No volitional form error tokens were recovered from the much smaller Ryo files. In this analysis of volitional form errors, the verb endings or word final phonological sequences that the children produced were compared with the word final sequences of the adult target forms. The direction of error is indicated in the error type column. The code V>C indicates that the consonant-stem ending


was used

inappropriately on a vowel-stem verb. The code C>V indicates that the vowel-stem ending -yoo was used inappropriately on a consonant-stem verb. The code C>*C means that an inappropriate innovative ending, for example -roo was used on a consonant-stem verb. 113

It is clear that these children were not operating with the same categories that are

normally used to describe Japanese verbal inflection, namely the vowel-stem and consonant-stem distinction. Furthermore, it does not seem likely that children or even adults do or could process verbs at the level of abstraction required by rules that require all inflected forms to be generated from a single verb stem. Sometimes Aki and Tai used consonant-stem endings on vowel-stem verbs (V>C). The productivity of verbs with non-past shapes ending in -iru or -eru might be problematic in acquisition because such forms can be analyzed as either consonantstem verbs or as vowel-stem verbs. For example, a form like neru could be analyzed as either the vowel-stem form ne-ru or the consonant-stem form ner-u. The vowelstem form would have the volitional form ne-yoo and the consonant-stem form would have the volitional form ner-oo. The actual verb ne-ru 'sleep, go to bed' has the volitional form ne-yoo 'let's sleep, go to bed', but Aki persisted with the error form *neroo, even after repeated correct production of other forms in the paradigm: ne-ru, ne-nai 'not sleep', and ne-ta 'slept'. A similar pattern was observed with the vowelstem verb mi-ru 'see, look at, watch'. Aki correctly produced mi-ru, mi-nai 'not see, not look at, not watch', and mi-ta 'saw, looked at, watched', which should have provided sufficient evidence to determine the shape of the verb stem. However, Aki made use of the consonant-stem volitional ending


and produced *mitoo and

*miitoo in place of mi-yoo 'let's see, let's look'. The child Tai also used consonantstem endings on some of these (potentially ambiguous) vowel-stem verbs. The verb katazuke-ru 'tidy up' is a vowel-stem verb, but Tai had *kazoo instead of katazukeyoo 'let's tidy up'. The verb hame-ru 'slide it in' is a vowel-stem verb, but Tai produced *hamoo for hame-yoo 'let's slide it in'.


Sometimes Aki and Tai used vowel-stem endings on consonant-stem verbs (C>V). Some of these errors may also be related to the potentially ambiguous -iru and -eru verbs. Aki makes use of the inappropriate vowel-stem ending -yoo with the consonantstem verb kaer-u 'go back, return'. Other errors reflect other paradigmatic complications. The consonant-stem verb ka(w)u 'buy' has the volitional form ka-oo 'let's buy' but Aki makes use of the vowel-stem volitional ending -yoo and produces the error form *kaiyoo. Consonant-stem verbs with stem-final w might be problematic for both first and second learners because the stem-final w has been lost in most modern inflected forms. The non-past forms are ka-u 'buy' and the polite ka-imasu 'buy'. The stem-final w appears only in the negative (kaw-anai 'not buy'), and some derived forms (the causative kaw-aseru 'make buy, let buy', and the passive kawareru 'be bought'). However, both the shape of the past form kat-ta 'bought' with the word final sequence -tta and the shape of the negative form kaw-anai 'not buy' are clear evidence that ka-u is consonant-stem verb. A similar explanation might be made for Tai's difficulty with the volitional form of mora(w)-u 'receive, get'. Tai makes use of the vowel-stem ending and produces *katte matteyoo in place of katte mora-oo 'let's have him buy it for us'. Sometimes consonant-stem verbs were assigned the wrong consonant-stem ending (C>*C). Aki had *oitaroo for ok-oo 'let's put it here/there'. It is possible that oitaroo might be a casual Nagoya dialect form of the benefactive construction oite yar-oo or oite age-yoo 'let's put it here/there (as a favor)'. However, the morphological tier of the transcription does code this form as an error with ok-oo as the intended target. The few errors that do occur in forms of the volitional do not support a single default rule for Japanese. Together Aki and Tai use inappropriate consonant-stem endings on vowel-stem verbs 13 times, inappropriate vowel-stem endings on


consonant-stem verbs 11 times, and innovative consonant-stem endings six times. Rather than a single default rule, it appears that children make use of a number of strategies based on type and token frequency. It has been informally suggested that the semi-productive paradigm of stem-final r

verbs might serve as the default paradigm, since it is the only paradigm, other than suru, that accommodates new verb forms in Japanese. However, the findings from the

volitional error analysis, where there was no clear default, do not support this hypothesis. Another way of testing this hypothesis would be to check for past form errors. The paradigm of stem-final r has a past form with the distinctive word final sequence -tta. However, a very preliminary investigation was not able to distinguish a clear pattern of -tta past errors by Aki, Ryo, and Tai. Consideration of a default pattern is also discussed in Chapter 7, where there is evidence that the stem-final r paradigm was not a very attractive response for adult native-speakers on a forced choice nonce test. Rather than a default pattern of error, verb type and token frequency may provide a better explanation of volitional form errors. If there is a sufficiently high type frequency of stem-final -r verbs (in a child's production), it might explain why children tend to segment the entire final sequence -roo from words like yar-oo 'let's do it' as the marker of the volitional form. The verb type and inflected form yaroo clearly has high type frequency for Aki, Ryo, and Tai. The verb-type frequency in the Japanese language as well, as far as can be determined by the dictionary count in Appendix I, also supports this conclusion. Consonant-final r verbs are the second most frequent category 0,368 members) following vowel-final e verbs (2,054 members) out of 7,060 verbs in the Daijirin Japanese dictionary. There is no single dominant verb-type and as a result productivity is graded as predicted by the Schema Model.



7.1. INTRODUCTION. Very few experimental studies dealing with the processing of Japanese inflection have been reported in the literature. De Chene (1985) and more recently Batchelder (1999) have reported on the experimental elicitation of inflected nonce verbs. De Chene found that adult native speakers had great difficulty producing appropriately inflected nonce verbs. Batchelder concluded that her experimental subjects were not using rule-based processing because they could not reliably perform a paradigm matching task. Vance (1987, 1991) used a written forced choice format and reported that adult native subjects had considerable difficulty choosing the analogically correct forms of nonce verbs. He also concluded that fully inflected regular Japanese verbs were stored in the lexicon. Klafehn (2001a) adapted Vance's test questionnaire for use with L2 instructed learners and reported that his non-native speaker subjects outperformed Vance's native speakers. The non-native speaker group chose 592 of 800 correct responses (74%) compared with 506 of 800 correct responses (63%) by the native speaker group. Section 7.3 reports on a new replication of the Vance experiment that again compares results obtained from L2 adult learners with results from adult native subjects. Revisions in the answer choices allowed for a more detailed analysis of subject responses. Finally, Section 7.4 describes an attempt to test the ability of young Japanese children to appropriately inflect novel verbs, the first Japanese wug-test.


7.2. PREVIOUS JAPANESE NONCE-PROBE EXPERIMENTS. 7.2.1. ORAL NONCE-PROBE EXPERIMENTS SENTENCE COMPLETION (DE CHENE 1982). De Chene (1982) used an oral protocol, whereby ten native speakers were asked to respond to a cue form of the nonce verb being tested with another form of the same verb. The subjects were trained to do a sentence completion task. For example, when they heard a sentence that made use of the volitional form (-00), followed by the cue phrase iia, moshi 'no if, the subjects were supposed to complete the response with the conditional form (-ba) form of the same verb. The following training example makes use of the real verb oyog-u 'swim'. Cue: Asita issyo ni oyog-oo? Iia, mosi

Response: oyog-eba. benkyoo dekinai.

'Tomorrow lets go swimming together.

No, if I swim, I can't study.'

De Chene tested fifteen different consonant-stem nonce-verbs. However, only five of the nonce verbs tested had the shape of actual verbs. These were all nonce verbs with stem-final -b (yub-u, neb-u, etc.). It was expected that these nonce verbs would be inflected with the pattern of actual Japanese verbs such as yob-u 'call, summon'. The other ten verbs tested were all stem-final z verbs (baz-u, kiz-u, etc.). The conjugation of these forms is problematic because as pointed out in Section 2.4.7. there are some speakers who make use of the alternate form kanzuru in place of the standard form kanji-ru 'feel'. It appears that the existence of this alternation may have complicated the conjugation task for de Chene's subjects. Since there are no standard verbs with stem-final z, de Chene expected the verb kiz-u to have the negative form kiz-anai. Only 46% of subject responses were consistent with this paradigm. Table 60,

is from Sato's (1985: 153) summary and critique of de Chene's results.


Category Cue Response Number of Responses % of total o kizoo kizanai 347 (46.3%) la kizoo kijinai 22 (2.9%) lb kizoo kizuranai 41 (5.5%) 2a kizoo kizonai 65 (8.7%) kizoo kizoranai 175 (23.3%) 2b kizoo kinai 15 (15.0%) 3a 3b kizoo kiranai 59 (7.9%) 4 kizoo other 26 (3.4%) Total 750 (100.0%) TABLE 60. Past forms of nonce verb kizu produced by native speakers.

Since most of the nonce verbs tested did not have shapes that resembled actual Japanese verbs, the validity of this experiment is questionable. There are other problems with the design of the experiment. For example, no attempt was made to associate any meaning with the nonce verbs other than the test frame sentence itself. Furthermore, the experimental protocol was complicated and required subjects to remember five different cue phrases, one for each inflected form tested. For example, in addition to the already mentioned cue iia moshi, the cue nn, demo itu 'yes, but when' was supposed to elicit the non-past form. The cue nn zehi 'yes, by all means' was supposed to elicit the hortative form and so on. Although the subjects did not produce a single pattern of responses consistent with the predictions of a Strong Rule hypothesis, this result is suspect due to the design of the experiment. PARADIGM MATCHING (BATCHELDER 1999).

More recently, Batchelder (1999) used a very carefully designed set of three experiments to test nonce verbs with the shapes of actual Japanese verbs. Twenty-six adult speakers were tested to see whether they could correctly identify the paradigm of sixteen nonce verbs. In each experiment, subjects listened to a short recorded conversation in which two different forms of the target verb were included. Each verb form was heard twice. In experiment one, after listening, subjects were asked only to 119

repeat one of the two verbs forms heard in the conversation. In experiments two and three, subjects were asked to either repeat a form or produce a third form that had not been heard. The nonce-verb types tested in each of the three experiments were varied in order to test various paradigm contrasts. An example follows.

In the recorded conversation between a clerk and a customer in a shop selling stereo components, the customer confirms with the clerk that the verb bimiru used to describe a stereo component means 'produce a rough sound at high volume'. The verb

bimiru was the nonce-verb being tested and the subject heard both this non-past form bimiru and the negative form bimiranai 'not produce a rough sound at high volume' two times each. In each conversation, the two forms of the nonce verb provided were logically sufficient to determine the paradigm of the test verb. Representative alternative paradigms for bimiru showing the two contrasting segmentations are given in Table 61. Verb-type



Vowel-stem (stem-final i) bimi-ru bimi-nai Consonant-stem (stem-final r) bimir-u bimir-anai TABLE 61. Alternate paradigms for nonce verb bimiru.


bimi-ta bimit-ta

In the example above, the verb bimiru with the negative form bimiranai must have the past form bimitta.

In addition to the contrast between stem-final i verbs and stem-final r verbs there are a number of other inflected verb shapes that are potentially ambiguous. The paradigm of the vowel-stem nonce verb ganosi-ru overlaps with the compound verb

gano suru, which is formed using the irregular verb suru. The past form ganosita might represent either the irregular compound form gano sita or the vowel-stem form

ganosi-ta. Batchelder was interested in discovering if her subjects could make these appropriate paradigm matches. A partial list of the Batchelder (1999:3) nonce verbs follows in Table 62, where the overlapping forms in each column have been bolded. 120

Non-past napir-u napi-ru napi suru napisur-u napisi-ru ganor-u ganosur-u gano suru ganosi-ru ganos-u TABLE 62.

Past Negative Similar paradigm real verb Verb type napit-ta napir-anai tigir-u 'vow' consonant-stem, stem-final r napi-ta napi-nai deki-ru 'able to do' vowel-stem, stem-final i napi sita napi sinai suru 'do' irregular suru + X napisut-ta napisur-anai gomasur-u 'flatter' consonant-stem, stem-final r napi-sita napi-sinai oti-ru 'fall' vowel-stem, stem-final i ganot-ta ganor-anai tadoru 'trudge' consonant-stem, stem-final r ganosut-ta ganosur-anai kosur-u 'rub' consonant-stem, stem-final r gano sita gano-sinai suru 'do' irregular suru + X ganosi-ta ganosi-nai oti-ru 'fall' vowel-stem, stem-final i ganos-ita ganos-anai kobos-u 'spill' consonant-stem, stem-final s Nonce-verbs with overlapping paradigms (Batchelder 1999).

Eight different verbs were tested in each experiment. Subjects in the first experiment were asked only to repeat one of the inflected forms heard in the conversation (either the negative or the past). In experiments two and three the subjects were asked to repeat a form after hearing a conversation the first time. Later during the second round, the same conversation was heard again, but this time the subjects were asked to produce a third form (either the negative or the past form) that had not been heard. Using this design Batchelder was able evaluate the effect of three different independent variables, verb paradigm, the heard versus unheard condition, and the first versus the second round condition. The results were quite different than might be expected if the subjects were making use of rules based on some type of paradigm representation. A Strong Rule Hypothesis predicts performance close to 100%. That prediction was not confirmed. In general the subjects made an average of one error for every eight forms tested. Depending on the paradigm contrast condition in experiments two and three, subjects correctly identified the paradigm of vowel-stem, stem-final e verbs only between 40% and 59% of the time. They correctly identified vowel-stem, stem-final i verbs between 73% and 81 % of the time. Subjects did best with the irregular suru verb, which they identified between 72% and 92% of the time.


These experimental results do not support the notion that the mental lexicon of Japanese speakers is organized on the basis of paradigms. Batchelder and Ohta (2000:47) make the following comment. Anecdotally, we have the impression that many Japanese adults are not consciously aware of the existence of two -ru paradigms, and when they are confronted by both paradigms in this manner, they become confused and often resort to random assignment of rapidly shifting choice strategies.

Batchelder and Ohta's conclusion (2000:48) is consistent with the argument presented in the present dissertation. A rule-governed process should yield highly deterministic results, but the data presented here show that Japanese verb inflection performance is not very deterministic. Perhaps this is the 'product oriented' process of Zager (1980) (cited in Bybee & Moder 1983, and Kopcke 1988), a choice of one of several abstract schemas which reside in the mental lexicon.


Batchelder comments that her subjects often could not repeat the form of a verb heard less than a minute before. About 28% of the time subjects could not reproduce the form correctly. For example, after hearing the negative form bimir-anai in the recorded conversation, a subject might report the form as bimi-nai. The difficulty of introducing a novel form into memory appears to be a major problem with Japanese oral nonce-probe tests. This situation constitutes a strong typological contrast with English. Berko's six and seven year old children demonstrated no difficulty remembering the nonce verbs rick and zib. The children were easily able to productively inflect these monosyllabic verbs that they had just heard. However, when one considers the complications involved in processing the verbal inflections of


Japanese, the ability of English speaking children to produce ricked from rick is a much less compelling argument in favor of universal rule-based processing. One alternative to oral production is the use of written tests. The objection to a written protocol is that subjects may be able to make use of general problem solving abilities that do not represent linguistic competence alone. If that is the case, one might expect better performance on written nonce tests. However, even when provided with the opportunity for more reflection and the chance to consciously compare each new response with a previous response, Yokomizo's (1990) subjects did not perform very well on a written paradigm identity task. Yokomizo was interested in the sociolinguistic aspects of loan-verbs and tested the ability of native speakers to conjugate newly established loanwords belonging to the semi-productive consonant-final-r paradigm. Yokomizo asked his subjects to write19 different inflected forms of three loanwords. In order to model the task, but not be too obvious, he presented two side-by-side sets of examples. The paradigm of the consonant-stem verb kir-u 'cut, was the model that Yokomizo expected his subjects to follow, but examples of the verb ki-ru 'wear' were also provided. One objection to the use of nonce verbs in experimental testing is that speakers do not manipulate nonce words in the same way as real words because nonce words lack meaning. The verbs Yokomizo tested were taken from current periodicals and were not nonce verbs. The subjects claimed that they knew or could guess the meaning of most or all of the verbs tested. For example the one loan-verb tested, dabur-u 'be doubled, left over, extra' , from English double is in common use. Yet, not one of the thirteen native-speakers tested produced only consonant-stem forms following the model of the consonant-stem verb kir-u 'cut'. The three verbs Yokomizo asked his subjects to conjugate were daburu, suusaidar-u 'commit suicide' and toirer-u 'go to


the toilet'. A translated version of Yokomizo's written questionnaire is given in Table 63. Notice that the distinction between vowel-stern verbs and consonant-stern verbs is neutralized in the non-past, the provisional, the passive, and the past-passive forms. To aid the reader, these forms have been bolded and all forms have been segmented to show appropriate sterns. Conjugate the foreign loan-verbs daburu, suusaidaru, toireru according to either example 1 or 2. Example I Example 2 'wear' 'cut' Form , kir-u ki-ru 'do... non-past ki-ta kit-ta 'did' past it! , 'do kir-e ki-yo imperative 'let's' kir-oo ki-yoo presumptive ki-taroo past-presumptive 'must have' kit-taroo 'if do' kir-eba ki-reba provisional 'if do' kit-tara ki-tara conditional ki-tari kit-tari 'doing' alternative kir-i ki 'do and' infinitive 'do' kit-te ki-te gerund 'don't' ki-nai kir-anai negative 'didn't' kir-nakatta ki-nakatta past-negative kir-areru ki-rareru 'be done' passive ki-rareta 'was done' kir-areta past-passive ki-saseru kir-aseru 'make do' causative past-causative 'made do' kir-aseta ki-saseta desiderative 'want to' kir-tai ki-tai past-desiderative 'wanted to' kir-itakatta ki-takatta formal 'do' kir-imasu ki-masu TABLE 63. Translated questionnaire form from Yokomizo (1990). English glosses and example segmentation have been added.

Many subjects were very close to 100 percent, but all of the subjects produced some forms like the vowel-stern model ki-ru 'wear'. Performance was most variable with the verb toirer-u 'go to the toilet'. One speaker produced only 39% consonantstern forms. Another produced only 44%. Two more subjects produced only 50% consonant-stern forms. Yokomizo suggests that these subjects produced vowel-stern forms like *toire-nai 'go to the toilet' and *toire-saseru 'cause to go to the toilet' 124

rather than the expected consonant-stem forms toirer-anai and toirer-aseru to avoid an "unnatural repetition of an r segment". Even when side-by-side models were provided, these subjects produced forms that were not paradigmatically consistent, and in this case, not even consistent with the rest of their own production. Sometimes these subjects used a vowel-final stem, and other times they made use of a consonant-final stem. Findings from this written format are consistent with the findings from oral testing. As in the Batchelder experiments, it does not appear that Yokomizo' s subjects were able to compute paradigm identity, even though there were none of the time constraints that might be associated with on-line oral production. FORCED CHOICE NONCE TEST (VANCE 1991). Vance (1987: 199-200) provides a very concise statement of motivation for his experimental study of nonce verb inflection by translating a quote from Akinaga. Akinaga (1966:45) makes the following remarks about Japanese verb forms (Vance's translation). "We conjugate verbs we have never used before without conscious thought, and we seldom make mistakes. This is because we have in our heads something like the grammatical rules for how the different groups of verbs conjugate, and we analogize from the verbs we know without conscious thought."

In the translated passage Akinaga mentions both rules and analogy, but Vance (1987:200) sidesteps the issue of a specific processing mechanism to focus on the question of whether speakers can actually perform the task of conjugation with novel verbs. In designing the experiment described here, my intention was to test the claim that Japanese speakers can extrapolate from the verb forms they know and correctly conjugate verbs that they


have never heard before. I limited the experiment to consonant verbs, since such extrapolation is clearly much easier for vowel verbs.

The results of the first experiment are reported in Vance (1987). In a revised and improved second experiment (1991), Vance tested four nonce verbs, which were given meanings and sample sentence contexts. A written questionnaire provided the non-past and the desiderative forms of the nonce verbs that were tested (homulhomitai,

hokulhokitai, murulmuritai, and kapu/kapitai). Table 64 provides a comparison of these nonce verbs and actual Japanese verbs that have similar shapes. Nonce verb forms provided

Similar shape actual verbs

Non-past Hortative Meaning yom-u yom-itai read kak-u kak-itai write nemur-u nemur-itai sleep kas-u kas-itai lend kat-u kat-itai win TABLE 64. Nonce verbs tested and actual verbs of similar shape. There is no actual stem-final p verb and two possible alternate models are provided. Non-past hom-u hok-u mur-u kap-u

Hortative hom-itai hok-itai mur-itai kap-itai

Meaning provided eat watch eat sleep

Subjects were asked to choose appropriate conditional (-eba), volitional (-00), negative (-anai), and past (-ta) forms of each verb from three answer choices. The response choices offered were determined on the basis of a pilot study, which Vance did not describe. Three of the four verbs tested had the shape of actual verbs, but the one verb kapu did not. The verb kap-u has the unpronounceable stem kap- and Vance suggested that if subjects could conjugate kap-u as well as the other three verbs, it would be an indication of the ability to "apply rules that involve reference to phonotactically inadmissible morphs". It was expected that the nonce verb kap-u would be conjugated like the actual verb kas-u 'lend' or kat-u 'win'. Once again, a Strong Rule hypothesis predicts performance close to 100%, especially since subjects were only required to choose a correct form. Unlike the previously reported de Chene, Yokomizo, and Batchelder experiments, where subjects 126

were asked to actually produce a form, the Vance subjects were asked only to choose a form. The Vance subjects did not do very well on the test. Only 63% of responses chosen (506 of 800) were paradigmatically consistent with the paradigms of real Japanese verbs. Performance on kap-u, the verb unlike any actual verb, was the worst ofthe four verbs tested. Vance (1991:156-157) reported the following conclusions. The results of the experiment reported in 2 are quite similar to the results of the 1987 experiment. This similarity strongly suggests that choosing analogically correct forms of made-up Japanese verbs is a task that involves some kind of intrinsic difficulty. As I noted, this difficulty is consistent with the claim that even morphologically regular Japanese verb forms are stored in the lexicon...

The full details of the Vance experiment are not provided here because they appear in the next section where the findings from a new replication are reported.


There were four specific reasons for the replication of the Vance experiment that is described next. First, since Vance (1991) reported his findings in the Journal of Japanese Linguistics, the debate over the details of the acquisition and processing of verbal inflection has assumed a more prominent position of interest in many areas of linguistics. Experimental findings that clearly do not support a Strong Rule hypothesis inform this debate. Vance's findings are not well known and no other experiments of this type with Japanese have been described in the literature. It was assumed that the experimental design was sound and the experiment was repeated to see if the Vance findings could be independently replicated.


Second, Vance reported only the total number of correct and the total number of incorrect responses for each answer choice on his test. Therefore, there was no indication of how many subjects chose 16 of 16 correct answer choices, how many chose 15 of 16 and so on. Furthermore, this method of reporting left no way to confirm any pattern of responses from individual subjects since each subject's individual responses were simply included in the total number of responses. The replication provided the opportunity to obtain completed test forms from each subject so that a more detailed analysis of response patterns could be reported. The third reason for the replication was to test non-native instructed learners of Japanese and to compare these results with the results from the native speakers. The completed forms from both sets of subjects allowed for a contrastive analysis of response patterns. Finally, the fourth reason for the replication was to make a slight, but significant modification to each set of answer choices on the original questionnaire form. In an endnote Vance (1991: 165) suggested that future nonce experiments include answer choices with the shape non-past + ending. This suggestion was motivated by the acknowledgement that there is a semi-productive process in Japanese whereby words (usually truncated forms of loan words) are treated as roots. The English word demonstration, for example, shortened to demo becomes the Japanese verb demoru 'to

demonstrate' and has a stem-final r consonant conjugation, demor-anai 'not demonstrate', demot-ta 'demonstrated'. Vance suggested that some of his subjects who chose the incorrect conditional form *kapureba and the incorrect volitional form *kapuroo of kapu might be making use of just such a strategy, namely treating kapu as

a root. The combination of kapu and ru would result in a verb with the conditional form *kapur-eba and the volitional form *kapur-oo. This rule-like process and the fact


that it does show some productivity in Japanese is also consistent with the notion of product-oriented schemas in the Schema ModeL To accommodate this modification in the replication a change was made to each set of three answer choices. In each set, one wrong choice was offered that had the shape of the non-past nonce verb plus an inflectional ending. For example, the nonce verb hat-u that appears on the first page of the questionnaire form has the past form

hat-ta. In the question that asks for the appropriate past form, the three choices offered are hata, hatuta, and hatta. The wrong choice hatuta is a combination of the already inflected form hat-u and the past ending -ta.

7.3.2. QUESTIONNAIRE FORM WITH SAMPLE RESPONSES. An English translation of the first page explanation and model responses follows. For each nonce verb tested, both a meaning and model sentences are clearly indicated. Even though the paradigm identity of each nonce verb is clear from the non-past form provided, the second sample sentence allows another method of confirming the paradigm. The two forms hatu and hatitai confirm that the verb hatu has the stem hat-. On each of the following four pages the conjugation of a nonce verb was tested and three answer choices were provided for each inflected form tested. The four nonce verbs tested were hom-u, hok-u, mur-u, and kap-u. The inflected forms tested were the conditional (-eba), the volitional (-00), the negative (-anai), and the past (-ta). Only the nonce verb kap-u has shape unlike any actual occurring verb. There are stem-final

m, k, and r verbs, but no stem-final p verbs in Japanese.


Young Japanese people are good at making up and using trendy words. This research investigates the origin and use of trendy words. Some of these words are names and some are verbs. This research looks at how verbs are conjugated. One such word is introduced and conjugated below. The meaning of the word is given in parentheses. In each example sentence below there is a blank and three answer choices. In each case a circle has been placed around the most appropriate response that completes the sentence. On the following pages there are choices like this. Circle the one choice you think is most appropriate. There are no correct answers. meaning



examples Sugu hatu n da tte sa. Hatitai naa.

'Says he'll start soon!' 'I wanna start already!'

(hazimeru) (hazimetai)

'start' 'want to start'

[1] Kinoo _ _ yokatta no ni.

'If only I'd started yesterday.'


'if start'


'Let's start'


'not start'



trendy word




[2] Saa, _ _ ka.


'Shall we start?'

haturoo~ 'Don't start!'

[3] _ _ de kure yo.



[4] Moo _ _ noka.




'He already started.' hatuta


7.3.3. SCHEMA AND STRONG RULE PREDICTIONS. A SRH predicts that native speakers should perform close to 100% on the nonce verb task. Furthermore, since a SRH assumes the ability to decompose inflected verbs in the input into appropriate stems which are symbolically represented in memory, the unpronounceable stem of the nonce verb kap-u should be no more difficult to process than the nonce verb mur-u, or hok-u, or hom-u. A SRH also predicts that a single default rule will account for any pattern of error. Since the learner group is much less likely to have control of the rules necessary to perform inflection, a SRH also predicts that the native-speaker subjects should outperform the learner subjects. Both as a 130

group and individually, the native-speakers should choose a larger number of correct forms. Schema Model predictions about performance on the nonce test are based on assumptions about the way schemas emerge from the interaction of verb type and token frequency. The strongest prediction is that performance on the four test verbs will vary, not according to verb, but according to the specific target form requested in each question. Productivity will be graded for each target form. The Schema Model assumes that productivity is the result of product-oriented schemas that work to approximate a product that is represented in memory. Therefore, both natives and learners will tend to choose forms similar to actual occurring forms. Recall that all consonant-stem verbs have a negative form with the shape stem-anai. All vowel-stem verbs are either stem-final i or stem-final e. There are no existing verbs that have a negative form where nai is proceeded by the vowel u. The SM predicts that both learners and natives will avoid responses like *murunai because they will have no negative forms of this general shape represented in memory. On the other hand, the WRM assumes that rules determine productivity. If it is true that there is a single default rule that is called into play when all else fails, the process that treats the form muru as a stem and combines it with the negative suffix -nai to yield murunai should apply. The WRM predicts that *murunai should be a very attractive answer choice.


The nonce forms were distributed to Japanese students in an American culture class at Tsukuba University in Ibaraki prefecture outside Tokyo and 97 completed


forms were returned. The 40 forms that indicated a place of origin other than eastern Japan were eliminated in order to minimize the possibility of western dialect interference. Three more forms were discarded because some responses had been left blank. Four more forms were randomly selected and discarded. All 50 remaining forms reported Japanese as a first language. All of these also indicated varying degrees of English fluency. The longest stay abroad (America) reported was eight months. The average age reported was nineteen years. INSTRUCTED LEARNER SUBJECTS. The 50 instructed learner subjects were all college students enrolled in either second or third year Japanese language classes at the University of Hawai'i at Manoa. Three of the 53 forms that were returned were randomly discarded. Of the fifty remaining forms 40 indicated English as a first language. Other first languages were self-reported as Cantonese (6), Chinese (2), Polish (1), and Korean (1). Since the purpose of the study was to investigate possible instructional effects and since the test form itself was written in Japanese, these learners were considered valid subjects. The average number of years of Japanese study reported was five. Ten subjects reported living in Japan with the average length of stay being 22 months. Their average age was twenty years.

7.3.5. RESULTS. Each subject made a total of 16 responses. In each group there were 50 subjects making a total of 800 responses. The learner responses are compared with the native responses on each item in Table 65. The learners made a larger number of responses that were paradigmatically consistent with the patterns of actually occurring verbs.


From this point on these paradigmatically consistent responses will be referred to as correct responses. Native

Learner 1




murureba 13 Conditional homureba 3 0 39 mureba P 40 homeba P 28 43 10 murereba homereba 7 7 9 Volitional 10 muroo P 42 homeyoo 5 39 Volitional muruyoo 2 homuroo 19 0 5 mureyoo P 26 38 homoo 8 6 Negative 39 murunai 1 homanai P 35 2 Negative muranai P 37 1 homunai 7 43 10 murenai 12 homenai 8 5 muruta 1 6 9 Past 3 Past hota P 39 43 34 3 mutta homutta 38 mureta 10 4 honda P 10 38 Conditional kapureba 29 hokeba P 26 6 Conditional 3 kapereba 14 19 15 hokureba kapeba P 7 hokereba 9 29 5 42 Volitional kapeyoo hokoo P 34 19 Volitional 9 kapoo P hokuyoo 3 9 35 3 13 5 kapuroo 22 hokeyoo 6 Negative kapunai 14 10 9 honai 6 Negative kapanai P 27 hokanai P 30 36 35 4 kapenai 14 hokunai 9 6 14 Past katta P 9 22 hota 12 Past kapita P 20 30 4 17 hokutta 32 kaputa 21 11 P hoita 8 TABLE 65. Total number of responses for each choice on modified Vance test. Paradigmatically consistent choices, consistent with the paradigms of actual Japanese verbs, are indicated with a P. Since there are no actual Japanese verbs with stem final -p, there are two paradigmatically consistent choices for the past form of the nonce verb *kap-u. Possible analogies are kas-u > kas-ita 'lend' and kat-u > kat-ta 'win' with *kap-u > *kap-ita or *kat-ta.

The learners chose 608 correct forms out of 800 possible choices (76%). The natives chose only 427 correct forms out of 800 possible choices (53%). The number correct for the natives is comparable with the 506 correct forms out of 800 (63%) reported by Vance (1991) with the same number of native subjects and the same correct answer choices.


Another way to compare the two groups in the present study is to look at the number of subjects in each group who obtained the same score. There were 50 subjects in each group and 16 possible correct responses. The average learner score was 12 out of 16 correct responses while the native-speakers had an average of only 8.5 out of 16. This comparison appears in Table 66. Total correct responses 16 15 14 13 12 11 10




6 5


Total Average

0 11 4 2 5 6 4 9 4 0 4 0 1 Number of learners 50 Number of natives 0 0 0 2 3 6 8 7 5 3 10 5 1 50 TABLE 66. Number of subjects in each group with same number of correct responses.

12/16 8.5/16

It is not a simple task for a native speaker to produce or recognize a new form of a

novel verb. Perhaps more striking is the contrast between the number of subjects with perfect scores. No natives, but 11 learners chose all 16 of the correct responses. At the other end of the scale, ten native-speakers scored only six out of 16 correct. With one exception, one out of every three answer choices on the questionnaire was a correct response. A score of ten out of 16 approaches the performance one might expect from guessing. However, later analysis will show that performance was not random. It is also possible to compare the relative degree of difficulty of all four verbs

based on the average number of correct responses for three different inflected forms. Since there were two possible correct responses for the past form of kap-u we compare only responses for the conditional, volitional, and negative forms. Correct responses

Incorrect responses

Mean correct per subject


Native Learner Native Learner Native Learner Native (Vance 1991) muru 119 125 31 25 2.38 2.50 2.16 hoku 90 116 60 34 1.80 2.32 1.98 homu 89 116 61 34 1.78 2.32 2.38 kapu 43 99 107 51 0.86 1.98 1.30 TABLE 67. Comparison of mean number correct per subject (three inflected forms of four verbs).

Native speaker performance with kap-u was the worst, inconsistent with the Strong Rule hypothesis, which assumes the uniform processing of all regular forms in the 134

same way. Compared to the natives, the learners demonstrated a much more even performance. This too is inconsistent with the Strong Rule assumption that acquisition is characterized by the incomplete mastery of inflectional rules.


A rule analysis of performance on a nonce test yields a very short report. Either the subjects used the appropriate rules or they didn't. Furthermore, as discussed in previous sections, neither Japanese children nor adults demonstrate much awareness that Japanese verbs can be organized on the basis of formal paradigms. A Strong Rule Hypothesis denies that there are frequency effects for regular forms. Therefore, the brief review of results in the proceeding section, which simply compares performance on different verbs (paradigms), cannot provide much insight into how or why the subjects actually chose the forms they did. A Schema Model, which claims that productivity is related to type and token frequency makes use of different tools. DEFAULT RULE VERSUS TYPE FREQUENCY.

The questionnaire form used in this replication was revised to test for the operation of a possible Japanese default rule. Each set of three answer choices on the questionnaire included one choice that consisted of a combination of the non-past form and an appropriate word-final sequence from the consonant-final r paradigm. These wrong answer choices are listed in Table 68 along with the number of subjects who chose them. From this point on these forms will be referred to as default forms. The segmentation in the table shows how each default form would be an innovative combination of the already inflected non-past form (stem) used as a root, and an ending.


No detailed analysis is required to see that while some speakers did choose the default choice, the other two choices offered on the questionnaire were generally chosen by a wide margin. The learners were especially reluctant to choose the default form. That is not surprising since, as pointed out in Chapter 3, instructed learners (at least those outside of Japan) have texts that emphasize the importance of the paradigm. Learner responses

Native responses

This choice Other two choices This choice Other two choices 1 49 37 homu-reba 13 2 48 31 homu-roo 19 49 43 1 homu-nai 7 3 47 34 16 homu-tta 47 31 3 hoku-reba 19 47 47 3 hoku-yoo 3 4 46 44 hoku-nai 6 4 46 20 hoku-tta 30 47 0 50 muru-reba 3 45 5 0 50 muru-yoo 2 48 49 muru-nai 1 47 49 3 1 muru-ta 44 6 21 kapu-reba 29 44 6 22 28 kapu-roo 44 6 41 kapu-nai 9 11 39 29 kapu-ta 21 740 = 800 217 + 583 = 800 60 Total + TABLE 68. Number of subjects choosing the default rule form.

The learners seemed to know that there is no paradigm in Japanese that allows an already inflected form (hom-u) to serve as a root (hom-u-reba). Furthermore, no textbook introduces the product-oriented process that makes a trendy verb from a noun. No text teaches students to truncate the loan-noun makudonarudo 'McDonalds' to maku and add -ru to get the slang verb makur-u 'go to McDonalds'. If, as Bybee claims, lexical representation and schemas emerge from usage, the fact that learners are probably unaware of the default rule and do not use it is also reflected in their lack of preference for the default forms in Table 68.


It is also important to note that the single default answer choice was always the wrong answer and any subject who chose only default forms would have scored zero out of 16. No native-speaker scored zero and even the 24 native subjects who scored 50% or worse (8/16, 7/16, 6/16, 5/16, 4/16) showed no pattern of preference for the default forms. The conclusion from these results is that neither learners nor native-speakers showed a preference for the most likely default form. Both sets of subjects avoided the default form. Either there is no single default form (and rule which produces it) or the inflected stem + ru paradigm is not the default paradigm. TYPE FREQUENCYAND PRODUCTIVITY.

A SRH maintains a strict separation between rules and the lexicon and makes no distinction between rules that generate many forms and rules that generate only a few. Productivity has no effect on the rules themselves. The Schema Model on the other hand allows for the interaction of the schema and the actual forms that are represented in memory. Schemas that involve many forms are expected to be stronger and more productive than schemas that involve only a few forms. In English the schema that emerges from thousands of verbs with the past tense form -ed is much more productive than any other irregular past tense pattern. In a Schema Model, the basic determinant of productivity is type frequency. High type frequency is associated with high productivity and low type frequency is associated with low productivity. The sing /sang pattern of vowel change that marks the past of only a few English verbs is not

normally extended to include other verbs by children in acquisition, nor by subjects in nonce-probe experiments.

137 LOW TYPE FREQUENCYAND LOW PRODUCTIVITY. If verbs with low type frequency have low productivity, then the answer choices

on the nonce-verb test with shapes unlike any actual Japanese verbs should not have been chosen with any great frequency. Verbs with no frequency (non-occurring types) should have low productivity. This is a testable hypothesis that is confirmed by the data in Table 69. None of the word-final sequences in Table 69 occur in real Japanese verbs. These sequences do not represent phonotactic violations. They are all valid word forms, however, they do not occur in any real verbs. Japanese vowel-stem verbs are either stem-final i or stem-final e. Therefore, there are no inflected forms where the inflections nai, ta, and yoo are preceded by u or o. With few exceptions these eleven non-occurring forms were among the least chosen responses by both natives and learners. Natives


7 1 homu-nai ho-ta (hom-u) 6 9 hoku-yoo 3 3 ho-nai 14 4 hoku-nai 6 4 ho-ta (hok-u) 12 14 muru-yoo 0 5 muru-nai 1 2 muru-ta 1 3 kapu-nai 9 6 kapu-ta 21 11 TABLE 69. Number of subjects (50 total) choosing verbs with non-occurring shapes. HIGH TYPE FREQUENCYAND HIGH PRODUCTIVITY. A SRH argues that productivity is categorical and independent of verb type frequency. The Schema model predicts that productivity basically depends on type frequency and that productivity will be graded. The computer search of the Daijirin


dictionary reported in Appendix I, shows that stem final -r verbs have a higher type frequency (19.4% of all verbs listed) than stem-final k verbs (7.8%) and that stem-final

k verbs have a higher type frequency than stem-final m verbs (6.3%). Higher type frequency means that there are more members (verbs) in the group. More members in the group means that the schema that emerges from the group should be stronger. This relative type frequency is reflected in native performance on the nonce test. If correct answer choices are taken as indications of productivity, there is a good match between dictionary type frequency and subject productivity. Dictionary verb type frequency and native subject performance on the nonce test are compared in Table 70. Verb type



FrequencyNonce verb

Correct responses

.194 mur-u stem-final r kaer-u return 1,368 79% (1581200) .063 hom-u 50% (991200) stem-final m yom-u read 447 hok-u .078 stem-final k kak-u write 552 49% (981200) o 36% (72/200) stem-final p none 0 kap-u Total verbs in Daijirin dictionary 7,060 TABLE 70. Dictionary type-frequency and nonce verb performance by native subjects.

The relatively high type frequency reported for stem-final r verbs does not distinguish different sub-types with different vowels before the stem-final r. There are five different possibilities, wakar-u 'understand', hair-u 'enter', kaer-u 'return', nor-u 'ride', and tukur-u 'make'. The relative frequency of the tukur-u sub-type, the sub-type that is most phonologically similar to the nonce verb mur-u is not clear. However, verbs with similar shapes (kabur-u 'wear a hat', ur-u 'sell', yabur-u 'tear', nemur-u 'sleep', hur-u 'rain, snow') are found in all three of the data files for Aki, Ryo, and Tai. Therefore, it is reasonable to assume that it is well represented in the mental lexicon of most adults as well.

139 LOW TYPE FRQUENCY AND LOW PRODUCTIVITY (K STEMS) The Schema Model predicts that lower type frequency should be associated with lower productivity. This prediction helps explain why so few native speakers chose the correct form hoita as the past form of hoku. (A similar argument accounts for why so few native subjects chose the correct past form honda of homu.) Results from the nonce test are repeated in Table 71. Answer choices

hota 12 hokutta 30 hoita 8 TABLE 71. Number of native responses on nonce verb test (50 total).

The fact that only eight subjects chose the correct choice hoita might at first seem surprising since the actual verb kak-u, has the same k stem and paradigm. One might think that the form hoita would be identified with the form ka-ita 'wrote'. However, the intuitive assumption that k stem verbs must have both a high type and token frequency is neither supported by dictionary type frequencies nor by searches of the Aki, Ryo, and Tai database. (The CHILDES database was used because appropriate adult databases are not readily available.) Table 72 repeats the significant portion of the Daijirin dictionary search findings from Appendix I. The stem-final k verb type has only about 550 members and ranks only number four in overall type frequency. Rank Verb type

Verb shape Example


-eru tabe-ru eat' 2,054 1 regular vowel-stem -ru kaer-u return' 1,368 2 regular consonant-stem -su hanas-u speak' 1,230 3 regular consonant-stem -ku kak-u write' 4 regular consonant-stem 552 Total number of verbs 7,060 TABLE 72. Relative frequency of verb types found in Daijirin dictionary.

Frequency .291 .194 .174 .078

The three small children Aki, Ryo, and Tai make good use of the kak-u lemma because in addition to 'write', kak-u also means 'draw'. In the largest data file, the child Tai uses some inflected form of kak-u 79 times. However, because the past form 140

ka-ita represents only 18 tokens of that total, it does not have high token frequency.

This particular statistic illustrates another typological characteristic of Japanese. Not all inflected forms of a verb will occur with the same frequency. For these three children the type frequency of consonant-final k verbs inflected for the past is quite low. If this type frequency is low, then the expectation is that the pattern of inflection would not be readily generalizable and not as likely to be extended to new forms. The low frequency of kaita found in the CHILDES files is reported in Table 73. Child Most frequent past type Tokens Rank

kaita 'wrote, drew' 12 naita 'cried' 8 kaita 'wrote, drew' 6 221 1 kaita 'wrote, drew' 18 Tai atta 'had, was' TABLE 73. Most frequent past token and most frequent past stem-final k tokens. Aki Ryo

atta atta

'had, was' 'had, was'


Most frequent k verb past Tokens Rank


1 1

14 28 37


Of course, it might be expected that older school children and adults would have a much higher token frequency of kaita, however, there is no reason to believe the relative type distribution of verb-types (consonant-stem and vowel-stem verbs) would be significantly different for adults. The fact that few adults chose the correct nonce form haita confirms the Schema Model prediction that low productivity is associated with low type frequency. DIFFERENTIAL PERFORMANCE: NATIVES AND LEARNERS. Both natives and learners more often chose inflected verb forms on the nonce test that had shapes similar to actual Japanese verbs. Best performance by both groups was on the verb mur-u, which might be considered either a truncated form of the actual verb nemur-u 'sleep' or some type of extension of ur-u 'sell'. There are two likely reasons that the instructed learner group outperformed the native-speaker group. In their instruction, learners are exposed to a paradigmatic organization of verbal


inflection. This training means that learners would be much more likely to cognitively process verbs in terms of paradigms and analogies. A Schema Model predicts that task frequency underlies productivity. An instructed learner survives classroom instruction by learning to produce one inflected form on the basis of another. Inflection of new forms is a high frequency cognitive task. New verbs introduced one week are tested on the next week's quiz. The productivity of native speakers, however, must emerge from usage. For adult native speakers, the task frequency of inflecting novel verbs must be quite low and any kind of product-oriented schema that emerges to produce a new form will be determined by thousands of other verb forms represented in memory. A SRH argues that the most efficient way to process inflection must be by the application of abstract symbolic rules. Results from the nonce test are not consistent with this hypothesis.

7.4. ORAL NONCE PROBE TEST WITH NATIVE CHILDREN. 7.4.1. PURPOSE OF THE EXPERIMENT AND RESEARCH QUESTIONS. The purpose of this experiment was to see if young Japanese children could appropriately inflect Japanese nonce verbs. In Berko's (1958) original wug test experiment the subjects were first grade and preschool students. Eight of her 56 subjects were only five years of age and six were even younger (from 4 to 4;6). The intent of the Japanese experiment was to test young Japanese speakers of approximately the same age. There were three research questions:

Q1. Reports from previous Japanese nonce experiments indicate subject difficulty remembering novel forms. Since young children are in a developmental stage where vocabulary learning is a high frequency learning task, it seemed possible


that children might not experience the same difficulty as adults. Is it possible to get young children to remember novel verb forms?

Q2. Is it possible to get young Japanese children to productively inflect a novel verb for three different inflected forms (the hortative


the past -ta, the negative

-nai)? If all three inflections are productive, is anyone form more problematic than the others?

Q3. Adult Japanese speakers experience difficulty with paradigm identification. Again, assuming young children are in a developmental stage where verb learning is high frequency task, they might be more likely to make use of paradigm strategies. Can young Japanese children do paradigm matching? Given the nonce forms rir-u and rir-anai, can children produce the appropriate consonant-stem past form rit-ta rather than the vowel stem form ri-ta? Two hypotheses were considered.

STRONG RULE HYPOTHESIS. A SRH predicts that the processing of verbal inflection is independent of both typology and verb type frequency. If the rules required for the inflection of Japanese are productive, the children should be able to correctly inflect all regular novel verbs, just like the English-speaking children in the Berko wug test. By the age of three years the three children in the CHILDES database were producing hundreds of different inflected (non-copula) verb types (Aki 877 types, Ryo 635 types, Tai 2,448 types). Since the children tested in the experiment described in this dissertation were two or three years older than Aki, Ryo, and Tai, the assumption was that verb inflection was productive for the experimental subjects (five and six years of age).



The Schema Model predicts that the processing of verbal inflection makes use of lexical schemas. The productivity of schemas basically depends on the type frequency of fully inflected forms represented in memory. One possible nonce test result is that Japanese children will perform like Japanese adults and show graded productivity of forms. Differences in performance between Japanese adults and children should be attributable to differences in the type frequency of verbs used by children and adults. A second possibility is that the typology of Japanese (many different inflected forms and no bare stems) may require a much denser lexical representation in order to become productive. It appears that Japanese verbs are not mentally organized on the basis of paradigms, and simple phonological similarity may be more important in generalization across various inflected forms. It may be that there is a gang effect that operates across mentally represented inflected forms of the same verb. Bybee (2001:123) suggests that two or three similar lexical items are not sufficient to form a gang. In another study Bybee and Pardo (1981) suggest that it may require six items to constitute a gang. It may be that Japanese verbs are subject to gang effects that require more than two or three items in memory. Since there are so many different inflected forms possible in Japanese, it may require repeated exposure to many more inflected forms before a novel verb is entered into memory. Very specifically, the two different forms of the nonce verb provided in the experiment might not be sufficient to establish a lexical representation that could enter into a productive product-oriented schema. The children may not be able respond.


7.4.2. SUBJECTS.

The subjects in this experiment were all kindergarten students at Ushigomi Nakano Kindergarten in Shinjuku, Japan. The experiment was conducted over three days and all children in the class (Sakura Gumi) who were in attendance during that time were tested. One of the 29 students tested was a native speaker of Korean, a recent arrival in Japan who had little Japanese proficiency. One student may have been bilingual in Chinese. Perhaps several others were bilingual in both Japanese and Korean. This was an unanticipated complication that could not be confirmed because the background information form used did not request information about first language spoken or knowledge of other languages. Sometime after the experiment was completed it was also discovered that some background forms were incomplete and did not specify exact date and place of birth. Both the lead teacher at the school and the teacher in charge of the Sakura Gumi class were aware of the research purpose of the experiment. (The class teacher also took the nonce test herself.) With the exception of the one student who was a recent immigrant, the two teachers gave no indication that they believed the performance of any other student in the experiment would be influenced by a lack of Japanese ability. Background information forms and parental consent forms were obtained from all subjects. Background forms indicated the subjects had an average age of five years and three months. The Protection of Human Subjects Assurance Identification number for this project was M-1217 and the IRB number was 01.



A Macintosh 03 laptop computer was used to project the eight test frames used in the experiment. At the same time that a test frame appeared on screen, a recorded Japanese cue sentence was also played from the same computer.


On the first day of the experiment the investigator was introduced to the class of 29 students by the school principal and spent a few minutes talking with the group. Once the experiment started each child was brought to a separate room where the experiment was conducted. Another investigator and a project coordinator were seated behind a low screen behind the child in the same room. After a brief introduction and conversation to put each child at ease, the experiment started. Each child was asked if he or she liked Pokemon cartoon characters. All replied yes. Next, the experimenter explained that they were going to playa word game that involved watching Pokemon characters on the screen of a laptop computer. A small lapel microphone was attached to the child's clothing to record all responses. As the test frames automatically appeared on the computer screen, a recorded oral explanation and a set of cues was played by the same machine. The time required to test one subject was approximately eight minutes.


In order to keep the length of time required for testing reasonably short, yet allow for the testing of several different word shapes, two different test versions were prepared. 15 subjects took Test A and 14 subjects took Test B. The verb forms


provided and the forms tested in Test A and Test B are given in Table 74. The forms tested were the inflected forms that the test attempted to elicit from subjects. TestA


Forms provided Forms tested for production Forms provided hok-u, hok-itai, ho-ita, hok-oo, hok-anai mur-u, mur-itai ba-u, ba-itai bat-ta, ba-oo, baw-anai me-ru, me-tai ri-ru, ri-tai ri-ta, ri-yoo, ri-nai rir-u, rir-itai TABLE 74. Inflected nonce forms provided and forms tested.

Forms tested for production mut-ta, mur-oo, mur-anai me-ta, me-yoo, me-nai rit-ta, rir-oo, rir-anai

Each test version consisted of 12 picture frames. The first frame of each set of four pictures was a model frame. In the first model frame a picture of the sleeping Pokemon character Kabigon appeared. To encourage oral production, the recorded instructions first asked the subject to repeat the character's name and then to repeat the real verb forms ne-ru 'sleep' and ne-tai 'want to sleep' in an appropriate context. Finally, in the cue sentence the subject was encouraged to produce the past form ne-ta 'slept'. A translated version of the transcription for the model frame follows. This is Kabigon. Please say Kabigon. _ _ Kabigon says he wants to sleep. Please say want to sleep. _ _ Kabigon loves to sleep. Please say sleep. _ _ Kabigon will sleep today. What about yesterday? Yesterday too he slept. Please say slept. _ _ Let's meet with some more Pokemon characters. This time please answer like you just did.

In each of the following three frames, a different Pokemon character and a different nonce verb were introduced. In each case an attempt was made to associate the nonce verb with a real characteristic or ability of the actual Pokemon character. For example, the Pokemon character Pukurin is known for its ability to frighten adversaries by swelling up and becoming large. The accompanying recorded text says


that Pukurin "likes to get big and hok-u" in an attempt to suggest that hok-u means something like 'blow up like a balloon'. The second set of four frames was used to model and test the production of the hortative (-00 'let's'). A translated version of the transcription for the model frame follows. Oh, look! Its Kabigon. Kabigon says he wants to sleep. Please say want to sleep. _ _ Do you think Kabigon will sleep more? Please say sleep. _ _ I wonder if it's fun to sleep. Let's try it together. Together with Kabigon Let's sleep. Please say, let's sleep. _ _

The third and last set of four frames modeled and tested production of the negative (-nai). This set of frames was different than the first two sets. In this set of frames two

pictures of each character were used. In the picture at the top of the frame Kabigon 's eyes are closed and he is clearly sleeping. In the picture in the lower part of the frame Kabigon 's eyes are open and his hand is raised in greeting. A translated version of the

complete transcription for the first frame follows. Oh, look! It's Kabigon. Kabigon says he wants to sleep. Please say want to sleep. _ _ I wonder if Kabigon will sleep tomorrow too. Please say sleep. _ _ Today he will sleep, but what about tomorrow. No, tomorrow he won't sleep. Please say won't sleep. _ _

In each of the following three frames the character action depicted in the upper portion of the frame contrasts with the lack of that action in the lower part of the frame. By the time a subject reached the third and last hok-u test frame, that subject had heard the form hok-itai six times and the form hok-u twelve times and had repeated each form three times.


7.4.6. RESULTS. Due to an equipment error on the first day of testing, the data for that day (eight subjects) was lost. A later error in labeling the minidisk used on the third day of testing resulted in recording over data (13 subjects) from the third day of testing. The results reported are from eight students (and one parent) recorded on the second day of testing. The children and the one parent found inflection of nonce verbs to be a daunting task. A summary follows in Table 75. Subject Test


Repetions Test forms produced

none none produced ne-nai 'won't sleep' before cue perfect performance on negatives hok-anai, baw-anai, ri-nai unintelligible (ri)?nai produced ne-nai 'won't sleep' before cue none *nai? for mur-anai mur-anai? riritai repeated as *ritai *ri-nai for rir-anai paradigm matching error ho-ita, *bauta/*bata for bat-ta, rita *hok-u for hokoo, ba-oo, *howani for hok-anai, baw-anai, *rinai for rir-anai TABLE 75. Subject performance on nonce test: forms produced.

Child Tl Child T2 Child T3 Child T4 Child T5 Child T6 Child T8 Child T9 AdultA7


all cues all cues all cues all cues all cues all cues all cues all cues all cues

7.4.7. ANALYSIS. The results reported here look nothing like the productive use of the past (-ed) and progressive (-ing) by English-speaking children reported by Berko (1958). There is no indication that any of the Berko subjects failed to respond. In the Japanese experiment failures to respond were typical. It should be noted that the automated design used in the Japanese test precluded slow and thoughtful responses of the type described by Berko. This is a possible flaw in the test design. Berko reports that all of her subjects understood what was being asked. They knew what they where supposed to do when they heard the completion cue phrase at the end of each test frame (Berko, 1958:2)


This is a man who knows how to rick. He is ricking. He did the same thing yesterday. What did he do yesterday? Yesterday he _ _.

There is some reason to believe that the sentence completion format might not be well motivated for eliciting nonce forms in Japanese. In a an investigation of mother and child interaction during book reading with young children, Katoo (2001) reported the striking comparison that English speaking mothers often use sentence completion questions, whereas Japanese speaking mothers never use sentence completion. It may be that the subjects in the nonce experiment did not really understand the completion task. Furthermore, the subjects were never directly asked if they understood what they were supposed to do. This many explain why there were so few attempts to respond to the (-00) and (-ta) picture frame sets. Perhaps the children just did not understand the sentence completion task. However, there is very clear evidence that two of the subjects understood very well what they were supposed to do in the third set of frames where the production of the negative form nai was tested. While listening to the model frame (the first frame in the set) for the negative form, these subjects provided the negative form nenai 'won't sleep' before they actually heard the word in the recorded model. Yet, in the following three frames, the same subjects did not produce inflected negative forms of the nonce verbs.

7.4.8. CONCLUSION. It appears that typology is involved in how verbal inflection is represented and

processed. Japanese speakers must deal with fully inflected forms, unpronounceable bare consonant-stems, opaque allomorphy, and many different inflectional paradigms with a wide range of type and token frequency. Japanese five and six year oIds do not


appear to be able to use rules to inflect nonce forms productively. This finding raises the question of how new verbs are learned by Japanese speakers. What does it take to enter a new form in memory? The Schema Model suggests that repetition is very important in establishing lexical representation. However, experimental studies show that Japanese speakers typically do not demonstrate the ability to segment stems or even to determine paradigm identity on the basis of one or even two diagnostic forms. It may be that Japanese speakers require repeated exposure to many occasions of use

of many different inflected forms in order for lexical schemas to emerge. This is a possibility that could be investigated further by careful longitudinal studies. A detailed and carefully conducted series of priming experiments involving repetition and lexical decision tasks would also be helpful in determining what kind of lexical schemas are typical for both children and adult native speakers of Japanese.




In this study a preliminary review and evaluation of mechanisms and models of inflectional processing was followed by an examination of Japanese children's inflectional errors and three experimental nonce probe experiments. The first goal was to determine what kind of inflectional errors occur during first language acquisition of Japanese. The second goal was to experimentally determine the ability of Japanese children, adults, and instructed L2 learners of Japanese to use novel verb forms productively. The third goal was to determine a model of inflectional processing that would be most consistent with the acquisition and experimental data. The two models considered were Pinker's (1999) Words and Rules model and Bybee's (2001) Schema Model. In Chapter 2, the details of the traditional native Japanese account of verbal inflection were compared with conventional linguistic rule-based models. The native analysis of inflection appears to be constrained by orthographic convention. For example, there is no way to graphically represent the stem of the verb yom-u 'read) as yom- because there is no way to write a syllable-final m in the native script (kana).

Therefore, the orthographically constrained native analysis makes use of six different complex stem forms rather than the more economical single stem of the linguistic analysis. Furthermore, the native analysis results in a situation where some verbs (in a linguistic analysis they would be described as monomoraic stems) have no roots at all. The native analysis is also unable to assign the potential meaning to a unique morpheme for consonant-stem verbs. Unlike the linguistic analysis, which derives yom-e-ru 'able to read' from yom-u 'read', (attributing the potential meaning to the


vowel e) the native analysis contrasts yomu 'read' and yomeru 'able to read' as two distinct verbs (two lemmas). Therefore, the formal native analysis is consistent with the notion that native speakers may not mentally segment inflected verbs into stems and morphemes and that it is fully inflected forms which are represented in memory. In Chapter 3, a brief review of Hockett's (1954) classical description of three basic models of morphology, Item and Process (lP), Item and Arrangement (lA) and Word and Paradigm (WP) concluded that the native Japanese treatment of verbal inflection made use of IA and WP, but not IP methods of analysis. It also concluded that while most instructed (L2) learners of Japanese probably do receive explicit instruction in the use of formal rules, they also receive strong implicit encouragement to make use of WP strategies such as analogy (based on knowledge of formal paradigms). This knowledge of formal paradigms, which is not demonstrated by native speakers, is used in Chapter 7 to account for the differential performance of instructed learners and natives on nonce probe tests. In Chapter 4, the details of four processing mechanisms (rules, analogy, competing cues and connections) were reviewed. It was demonstrated how analogic processing of Japanese verbs (making use of fully inflected forms) could obviate the problematic stem segmentation required by rule-based mechanisms. Consideration was also given to Antilla's (1989) notion of non-proportional analogy. Analogy was used to explain the Japanese child error of neroo for ne-yoo 'let's sleep, let's go to bed' (kaeru:kaeroo::neru:*neroo). The competing cue mechanism and the related notion of

cue reliability (Bates and MacWhinney 1987) were shown to be relevant to the class of Japanese verbs (with non-past shape -iru or -eru). Finally, it was pointed out that the adjustment of connection weights, rather than a combination of stem and suffix can adequately model the inflection of the English past in a connectionist model.


In Chapter 5, three models of processing were considered: Pinker's (1999) Words and Rules, various connectionist models, and Bybee's (2001) Schema Model (SM). Special attention was given to the distinction Bybee makes between source-oriented schemas (equivalent to traditional generative rules) and product-oriented schemas. Product-oriented schemas do not specify a structural input and are basically driven by the type-frequency of the product. At the end of the chapter a distinction was made between any model of processing that requires the use of rules, a Strong Rule Hypothesis (SRH), and Bybee's SM, which does not. SRH and SM predictions with regard to productivity, acquisition, and errors were specified. In Chapter 6, the results of computer searches of the Miyata (1992, 1993, 1995, 2002) database were analyzed and it was found that: 1) The three young Japanese children in the study did not overregularize either irregular verbs (suru 'do' and kuru 'come') or irregular verb forms (verbs with a few irregular forms in the paradigm) as predicted by a SRH. 2) Most verb errors that did occur could be categorized as stem errors. 3) There were many more errors with regular verbs than with irregular verbs and verb forms. 4) There was no pattern of error consistent with a SRH claim that there is a default error pattern. It was concluded that these findings were consistent with the SM assumption that fully inflected forms are represented in memory. In Chapter 7, a review of the relevant literature showed that previous experimental studies with Japanese adults did not result in findings consistent with a SRH. In both written and oral production and recognition tasks native speakers did not achieve the near 100% productivity predicted by a SRH. In addition, Batchelder (1999) also demonstrated that natives do not do well on paradigm recognition tasks, suggesting that Japanese verbs are not mentally organized on the basis of paradigms. A new replication of Vance's (1991) nonce probe test compared an instructed learner group


and a group of native speakers. The 50 learners outperformed the 50 native speakers in choosing paradigmatically consistent inflected forms of nonce verbs. The learner group chose 608 correct forms out of 800 possible choices (76%). The native group chose only 427 correct forms out of 800 possible choices (53%). A modification of Vance's original test format also provided a test of the hypothesis that native speakers might be using the stem-final r paradigm as a default paradigm. Neither the learners nor the native-speakers showed a preference for forms derived from the stem-final r paradigm, leading to the conclusion that there is no Japanese default paradigm. Finally, an oral response, nonce probe test (Pokemon test) with eight Japanese children (five and six years of age) was not able to confirm that the children could make use of either rules or paradigmatic knowledge to productively inflect nonce verb forms. With very few exceptions, quite unlike the English-speaking children of approximately the same age in the Berko (1958) experiment, the Japanese children failed to respond when asked to produce inflected forms of novel verbs. The failure of both native adults (modified Vance test) and children (Pokemon test) to show full productivity in these two experimental tests was interpreted as inconsistent with the basic assumptions of a Strong Rule Hypothesis. Japanese speakers do not decompose inflected verbs into mentally represented stems and inflectional endings, which can be freely used in the production of novel forms.


The three Japanese children in the database study (Aki, Ryo, and Tai) did not make very many verb errors and the errors they did make tended to be with regular and not with irregular verb forms. These rates are not directly comparable with


English overregularization rates, because the verbal typologies of the two languages are so different. (English has bare stems, few inflections, and contrasts a single regular inflection against several irregular patterns. Japanese has no bare stems, many inflections, and contrasts several different patterns of irregularity against two irregular verbs.) Furthermore, in a very through database study, Marcus et al. (1992) concluded that even for English speaking children, only about 2.5% of irregular verbs are overregularized. Rather than ask why children overregularize, the more interesting crosslinguistic question may be to ask why children make so few errors at all? One hypothesis that addresses this issue characterizes children as conservative language learners. Perhaps children do not make use of novel forms, but wait until they have heard adults use forms before attempting to use them themselves. MacWhinney (2003b) rejects this hypothesis because he argues that a strong version of conservatism means that children would never say *goed or *unsqueezed, however, we know that children do produce these forms. On the other hand, a weak version of conservatism (children avoiding, and therefore only sometimes producing novel forms) is supported by anecdotal evidence from the Miyata data. The child Ryo made unusual use (three times) of the inappropriately polite form ne-masyoo 'let's go to bed, let's sleep' when addressing his mother. The expected

informal form (used among family members) ne-yoo 'let's go to bed, let's sleep' is also the (morphologically) problematic adult form that both Aki and Tai mistakenly inflected as *neroo. Ryo avoids the problematic informal form by substituting a pragmatically inappropriate polite form. The polite suffix -masyoo 'let's... ' is not normally used by children to address other family members. What is remarkable about Ryo's use of this polite suffix is that there are no other instances of its use by Ryo. There are also no uses of the appropriate informal form ne-yoo. However, Ryo does


make use of the informal form with another verb: si-yoo 'lets do it'. This anecdotal example suggests how children might make much use of early item learning rather than rule generalization. Ryo was clearly able to take advantage of Japanese typology (many inflected forms, some with overlapping meaning) to avoid a problematic form. He substituted a semantically equivalent form (ne-masyoo) in place of the pragmatically appropriate ne-yoo. Further crosslinguistic studies, where the effects of different verbal typologies (including pragmatics) are considered, may yet confirm that conservatism (avoidance of problematic forms) has been underestimated as an important strategy employed by children in the developmental process. At the same time, the avoidance of problematic forms may not represent a conscious choice or strategy at all. Rather, the use of an alternate inflected form may be an indication that the child's morphology is powerfully constrained by semantics (children go for meaning first) and by what Plunkett and Elman (1997) call the importance of starting small. Very specifically, a connectionist model, or a model of child language acquisition as well, can avoid overly complex computational tasks (and resultant error) if it does not attempt too much too early. For Japanese children, one or two or even three or four inflected forms of a single verb represented in memory may not be enough of a basis to form a generalization about the shape of a verb stem. The relatively error free acquisition of Japanese inflectional morphology may be the result of starting small (not generalizing too early across forms in memory). Furthermore, the experimental findings in the present study suggest that Japanese speakers start small and stay small.



In Chapter 7, it was reported that the frequency assumption of Bybee's Schema Model provided an explanation for the graduated performance on the modified Vance test. The Schema Model assumes that more frequent verb types are more productive and that productivity is therefore graded, depending on verb type. Native-speaker subjects chose the highest number of correctly inflected nonce forms (79%) of the consonant-stem verb mur-u. Verbs of this category (stem-final r) comprise about 19% of all verbs listed in the Daijirin dictionary. Subject performance with the nonce verbs hom-u (50%) and hok-u (49%) was considerably worse and is consistent with the fact

that these two verb types represent much smaller categories in the Daijirin (about 6.3% and 7.8% of all verbs respectively). Worst performance was with the nonce verb kap-u. There are no stem-final p verbs and this gap was reflected in the performance

data. Subjects chose only 36% of the correctly inflected forms. Subjects were apparently able to see a similarity between the paradigm of more frequent real verbs like ur-u 'sell' and the nonce verb mur-u. For the less frequent verb types (stem-final m and stem-final k verbs), the similarity between the nonce verbs hom-u and hok-u and real verbs such as yom-u 'read' and kak-u 'write' was less

apparent. However, subjects were also able to choose at least some correctly inflected forms of the nonce verb kap-u, even though there is no actual Japanese verb with stem-final p. If verb type-frequency determines productivity, and there is zero type frequency (no stem-final p verbs), what was the similarity that the subjects perceived when they chose the inflected responses of this test item? In an experimental study of English priming effects, Gonnerman (1999) describes a similarity metric used to evaluate semantic and phonological similarity in a systematic manner. Future studies


that attempt to account for the processing of Japanese inflection on the basis of frequency and similarity will also need to consider such measures.


The results of the present study suggest a number of future research questions. If productivity is determined by verb type-frequency, in addition to (difficult to obtain) detailed analysis of caretaker input frequencies; it would be of interest to report in detail the sequence of verbs that are produced by Japanese children. Verb type and token frequencies could be used to evaluate developmental and error patterns. Tomasello (1992) has suggested a verb island hypothesis for English, whereby children acquire the combinatorial rules of grammar verb-by-verb. It seems possible that a similar pattern of item-by-item development might be demonstrated for Japanese. An automated search of the Miyata database could confirm which inflected form of each new verb appears first and if there is a consistent sequence of appearance of inflected forms. Pragmatic and semantic effects could also be investigated. Developmental patterns could then also be related to errors. It could be determined whether error forms are new forms or old forms and what types of forms are more problematic. There is at least some suggestion that the loss of stem-final k in some inflected forms was a problem for Aki, Ryo, and Tai. The stem-final k that appears in the non-past form is lost in the past and in the gerund form (kak-u 'write', ka-ita 'wrote', ka-ite 'write and ...'). Aki produced *kaketa for kai-ta 'drew' ,. Ryo produced *kuttukutta for kuttu-ita 'stuck to, adhered' , and Tai produced *karete for ka-ite 'draw

and...'. These errors suggest that the interacting effects of type and token frequency (as well as phonological and semantic similarity) need to be investigated and evaluated in


order to present a more complete picture of how inflected forms are represented and processed.


The general reaction of both Japanese-speaking adults and children to nonce-probe testing suggests that the use of made-up words may not be the best method of testing productivity. Obviously, both children and even adults must be able to learn new verbs. However, some adults anecdotally report that the nonce-probe task is unnatural and difficult because the nonce verbs have no meaning. This is a somewhat surprising reaction because the modified Vance Test used in Chapter 7 provides a standard equivalent vocabulary verb for each nonce verb tested along with two sample sentences. The child subjects in the Pokemon Test also demonstrated uncertainty about the meaning of some nonce test verbs. When shown two contrasting pictures of the monster Pukurin (one normal size and one showing an inflated, balloon-like Pukurin) and asked to repeat the nonce verb hoku from the sample sentence Pukurin likes to get big and hoku, one little boy was heard to quietly say to himself hoku tte? 'What does hoku mean?' Berko (1958) does not report that any of her English-speaking children

asked what a wug was or what rick meant. It may be that inflectional typology (many inflected forms) and the pragmatics of language use do not predispose Japanese speakers to the kind of meta-linguistic manipulation of language that may be more typical of speakers of English. Bybee (2001) has claimed that all knowledge of language is procedural and emerges from use. This claim is consistent with the experimental findings of the present study where Japanese subjects experienced difficulty performing abstract operations on novel forms that had not been previously used for communicative purposes. Therefore, rather than the problematic task of


asking subjects to produce or recognize appropriately inflected novel forms, a more pragmatically well-motivated task may need to make use of response latencies. The use of priming to inhibit or facilitate lexical decision has been increasingly used to investigate lexical access (representation and processing). A suggestion about how response latency, the experimental measurement of reaction time, could inform future research on Japanese lexical access is considered in the following discussion.


Some early experimental research supported the claim that there was a distinction between the lexical access of regular and irregular forms. Both Prasada et al. (1990) and Seidenberg and Bruck (1990) independently reported that subjects had a longer response latency when asked to produce the past forms of irregular verbs than when asked to produce the past forms of regular verbs. While both studies also reported a frequency effect for the verb stems of both regulars and irregulars, only irregular verbs showed a past tense frequency effect. However, Bybee (2000) did find a frequency effect for English word-final tid deletion. She considered three word categories: monomorphemic words (old), irregular verbs (told) and regularly inflected verbs (walked). She found that high token-frequency items in all three categories were more likely to undergo tid deletion than low frequency items, suggesting that regularly inflected verbs were accessed in the same manner as irregulars and monomorphemic lexical items. Other studies (Bybee 1985, Hare and Elman 1995) have also shown that high frequency tends to protect items from error, suggesting that high frequency items (regardless of regularity) are fully stored in memory. These suggestions are consistent with the longitudinal and experimental findings of Chapter 6 and Chapter 7 of the present


study, where verb type-frequency was related to productivity. Japanese children demonstrated relatively error-free acquisition of high frequency irregular forms. Japanese adults made fewer errors on the nonce test when test items were similar to actually occurring higher frequency verb types. Very recently, in two carefully designed experiments, Hare et al. (2001) report frequency effects for both irregular and regular English verbs. In the first experiment subjects were asked to write a sentence using a word that was aurally presented. The aurally presented words were monomorphemic homophones. Three frequency conditions were considered: 1) Past tense frequency greater (irregular past made/maid or regular past allowed/aloud, 2) Homophone frequency greater (irregular ate/eight or regular fined/find), 3) Approximately equal frequency (ducked/duct and heard/herd). The expectation that the most frequent homophone would be accessed first was confirmed for the regulars as well as for the irregulars. The second experiment made use of a cross-modal priming task. Immediately after the aural presentation of a prime, subjects saw a target word on a computer screen and were asked to make a yes or no lexical decision. Regular and irregular primes were used under two conditions, homophone and unambiguous. For example, the regular homophone prime paced/paste was followed by the target pace. The regular unambiguous prime jumped was followed by the target jump. A third set of primes had no morphological, semantic, or phonological relation to their target verbs. As in the first experiment, the frequency condition of the homophones was controlled and a clear frequency effect for the regular verbs was found. When the frequency of the regular past tense verb was greater than the homophone, response to the decision task was facilitated and when the frequency of the homophone was greater, response was delayed.


It should be possible to experimentally investigate Japanese frequency effects in the same way. A major obstacle has been the lack of an accessible Japanese database with comprehensive and detailed word frequency counts. One sees frequent references in the literature to Francis and Kucera (1982) and the massive Lancaster-Oslo/Bergen (LOB) (Johansson and Hofland 1989) English databases, but no comparable Japanese database was available until recently. Now, a computer searchable version of the NTT (Nippon Telephone and Telegraph) Japanese database (Amano and Kondo, 2000) has been released. Database frequency counts for specific Japanese verb forms could be used to predict response latencies in both production and lexical decision tasks. If lexical access depends on frequency, regular non-past forms and past forms should both show the same type of frequency effect reported by Hare et al (2001) for English. More frequent Japanese regular forms should be produced more rapidly than less frequent regular forms. The competition effect of Japanese homophones could also be tested in production and lexical decision tasks. Depending on the prime/target relationship, it would be expected that the response latency of an unambiguous inflected verb form would be different than the response latency of an ambiguous inflected form in a lexical decision task. Response to the unambiguous past prime tabe-ru 'eat' with the target tabe-nai 'not eat' might contrast in response latency to the ambiguous prime i-

ru 'be located' /ir-u 'need', where there are two possible targets (either i-nai 'is not located' or the target ir-anai 'don't need'). Specific predictions about experimental outcomes depend on the availability of verb type and token frequency counts and the development of measures of similarity evaluation (Gonnerman 1999). These predictions and the details of experimental design remain part of the challenge of future research.



VERB TYPE-FREQUENCY INTRODUCTION. Native Japanese grammars, as well as western descriptive accounts (Bloch 1946, Martin 1952), and the two widely used pedagogical grammars (Jorden 1987, Tsukuba 1991) intended for second language learners, agree that there are only two irregular verbs in modem spoken Japanese. The paradigms of the verbs suru 'do' and kuru 'come' show unpredictable stem vowel alternations not found in the paradigms of other verbs. The various inflected shapes of all other verbs in Japanese are considered to be morphologically and phonologically predictable and therefore "regular". There is, however, considerable allomorphic variation at the verb stem boundary and it is this predictable variation which produces several different contrasting patterns of regularity. Unlike English and many languages which have been characterized as contrasting a single regular rule with a number of semi-productive irregular patterns (Pinker 1999), Japanese could be characterized as contrasting two irregular verbs against several different patterns of regularity. For English, it has been determined that there are about 180 irregular verbs (Pinker and Prince 1988). All other English verbs are considered to be regular. For Japanese, on the other hand, while there are two irregular verbs, the relative number of verbs in each of the other regular categories has not been determined. The purpose of this section is to approximate the relative type-frequency of Japanese verbs based on entries found in a representative and comprehensive Japanese dictionary. In the discussion that follows, type-frequency will be used to refer to the relative number of verbs in each type of regular and irregular paradigm. Corpus text searches in English (Francis and Kucera 1982, Johansson and Hofland 1989) have


been used to determine verb token-frequency in spoken and written language. More recently, large data base corpora of Japanese have also become available (Japan Electronic Dictionary Institute 1995, Amano and Kondo 2000), but at the time of the present investigation, there is no published report in the literature on the relative typefrequency of Japanese verbs. Assuming that verb type-frequency and token-frequency are factors involved in the acquisition, representation and processing of Japanese verbal inflection, it is of interest to determine the relative type-frequency of Japanese verbs. It is not possible to evaluate the significance of frequency effects if relative verb type frequencies are not known. One obvious corpus that could be used to determine a verb type-frequency of the language is a native Japanese dictionary. However, hand counts of verbs in large single-volume Japanese language dictionaries like the Koojien, (1998) published by Iwanami, and the Daijirin (1995), published by Sanseido, each with more than 100,000 entries, are impractical. Recently, however, CD and on-line versions of some of the major Japanese language dictionaries have become available. These dictionary resources allow for rapid computer assisted searches that return data much more efficiently than hand counting. The Asahi newspaper in Japan maintains a dictionary web site that is a particularly useful research tool. At the Asahi web site it is possible make exhaustive searches of an on-line 1995 edition of the Daijirin dictionary. Unlike many other online and CD dictionaries, a search in this on-line version of the Daijirin is not limited to the first ten, or twenty, or even one hundred items. It is possible to enter a search that will retrieve all items that match the entered search criteria. A brief explanation at the Asahi web site says that there are 233,000 entries in the dictionary including classical (koodai) to contemporary (gendai) forms. Because this on-line version of the


Daijirin is comprehensive and allows for exhaustive and efficient searches, it was

chosen as the corpus for this first attempt at a determination of Japanese verb typefrequency. PURPOSE OF THE TYPE-FREQUENCY DETERMINATION.

The general purpose of the dictionary search was to determine the relative typefrequency of all verbs listed in the Asahi on-line version of the 1995 Daijirin dictionary. There were five specific research questions:

Ql. What is the total number of verb entries in the Daijirin dictionary? Q2. Of the total number of verb entries, how many are modem forms? Q3. What is the number and relative frequency of modem irregular verbs? (The Daijirin lists both suru and kuru and some compounds of these two irregular verbs

as independent lexical items.) Q4. What is the number and relative frequency of the two types of modem vowel-

stem verbs: stem-final -i verbs and stem final -e verbs? Q5. What is the number and relative frequency of each of the nine modem stem-

final consonant verbs: -r, -s, -k, -m, -(w) u, -t, -g, -b, and -n?

CLASSICAL AND MODERN VERB CLASSIFICATIONS. If the ultimate purpose of determining a verb type-frequency is to relate type-

frequency and the processing of modem spoken Japanese, then it is desirable to exclude from consideration the classical Japanese verb forms that are listed in the Daijirin. The classical verbs are either no longer in use or have changed their

classification. A verb like sinu 'die' for example, was conjugated as an irregular verb in classical Japanese, but is now conjugated as a regular godan verb. The godan verbs


are equivalent to consonant-stem verbs in a linguistic analysis. The verb paradigms of classical and modem spoken Japanese are quite different, and although the Daijirin lists both modem and classical verb forms, it is possible to exclude the classical forms because the entry classifications for classical and modem forms are also different. The native Japanese description of modem verbs makes use of five classifications that have evolved from the nine classifications of classical Japanese (Ikeda 1975). Table 1 provides a summary of the classical and modem classifications with an example of a representative verb from each classification. The native classification system distinguishes two types of regular two-step and one-row verbs. The kamiichi verbs are stem-final -i verbs and the shimoichi verbs are the stem-final -e verbs. Nine classical verb classifications

Five modern verb classifications

Classification Classification Example irregular (sahen) Irregular (sahen) su 'do' irregular (kahen) Irregular (kahen) kuru 'come' sinu 'die' Irregular (nahen) Irregular (rahen) ari 'have' regular five-row (godan) Regular four-row (yodan) katu 'win' Regular two-step (kamini) otu 'fall' Regular two-step (shimoni) koyu 'grow fat' regular one-row (kamiichl) Regular one-row (kamiicht) iru 'shoot' Regular one-row (shimoichi) keru 'kick' regular one-row (shimoichi) TABLE 76. Classical and modern Japanese dictionary verb classifications.

Example suru'do' kuru 'come'

kat-u 'win'

ochi-ru 'fall' koe-ru 'get fat'

In the native classification system, both classical and modem forms are written using the native kana syllabary, which restricts a single character to the representation of a vowel syllable, a combination consonant and vowel syllable, or a mora unit. In Chapter 2, it was pointed out that the native analysis of verbal conjugation is also limited by the phonological constraints, which are reflected in this orthography. In this system of orthographic representation, it is not possible to classify a verb like yomu 'drink' as a consonant-stem verb yom-u, because it is not possible to analyze the single syllabary character mu as anything smaller than the syllable mu. It is not possible to 167

graphically represent the single consonant 1m!. In modern Japanese the verb yom-u is considered to be a five-row verb and it is classified in the Daijirin dictionary as a doo-

ma-go verb (verb-ma-five). The ma in the classification provides the same information as a western linguistic analysis, which analyzes yomu as a consonant-stem, stem-final

-m verb. Since the Daijirin dictionary clearly specifies a classification for each verb entry, it was possible to perform a search to identify, count, and exclude all classical verbs (and conjugations) which are no longer in use. Furthermore, by making use of the same classification system, it was possible to identify and count the number of verbs in each modern sub-type.

CLASSICAL VERBS: RESULTS. The nine classes of classical verbs yielded the following search results. Table 77 gives the results and some examples for the lowest frequency verbs. Verb classification



Irregular na-verbs (nahen) inu 'go', sinu 'die' Irregular ra-verbs (rahen) ari 'have, be', ori 'be' ikanari 'why', keri 'come' Total TABLE 77. Classical irregular na-verbs and ra-verbs from Daijirin.

2 18 20

A summary of the more frequent two-row vowel verbs is provided in Table 78. Classical two-step verbs that are still in use have become modern one-row verbs and have double entry in the Daijirin. Verb classification

a ka ga sa za

Upper two-row (kamini) Lower two-row (shimoni)


fa da na ha ba ma ya

ra wa Total

6 7 0 8 194 66 141

2 6 9 0 13 77 12 2 9 2 145 3 81 53 26 181 36202 83 267 12 1,353 1,498 TABLE 78. Classical regular upper and lower two-step (vowel-stem) verbs, in each category.

Classical verbs that were upper two-row are equivalent to modern vowel-stem, stemfinal -i verbs. Classical two-row verbs that were lower two-row verbs are equivalent to


modern vowel-stem, stem-final -e verbs. Both two-row and one-row verbs are considered to be regular verbs. The summary of the most frequent consonant-stem verbs in Table 79 shows that the total number of classical two-row vowel-stem verbs (1,498) is approximately equal to the total number of all four-row consonant verbs (1,765). Verb classification Four-row (yodan) 1m ga sa fa da na ha ba ma ra 311 42 303 104 1 256 37 162 548 Number TABLE 79. Classical regular four-row (consonant-stem) verbs, in each category.

Total 1,765

A summary of all classical verb forms to be excluded from the count of modern verb forms is presented in Table 80. There are a total of 3,283 verbs that were excluded. Verb classification


Irregular na-verbs (nahen) Irregular ra-verbs (rahen) Regular two-step upper (lmmini) (-i stem) Regular two-step lower (shimon!) (-e stem) Regular four-row (consonant-stem) Total TABLE 80. Summary of classical verb types.

2 18 145 1,353 1,765 3,283


An exhaustive search of the Daijirin retrieved 458 entries classified as irregular kuru (kahen) verbs and irregular suru (sahen) verbs. In the case of kuru verbs, all and

only verbs with the modern Japanese shape -kuru were counted. In Table six, the entry count of five for kuru includes modern forms such as yatte kuru 'come along, appear, show up'. Strictly speaking the grammatical construction consisting of the two verbs yatte (untensed gerund form) and the main verb kuru 'come' does not constitute another form of kuru, however, this is how it is entered in the Daijirin. A similar problem occurs with the use of irregular suru 'do', where hundreds of entries are 169

included in the Daijirin. The complete list of kuru entries appears in Table 81 along with some sample entries of suru 'do' verbs. Verb classification

Sample entries


Irregular (kahen) kuru 'come' kuru, sekikuru, segurikuru, yattekuru Irregular (sahen) suru 'do' suru, kansuru, koisuru, wasuru, dokusuru Total TABLE 81. Modern and classical irregular kuru and suru verbs.

5 453 458

The verb suru is very productive. The count in the present study includes all and only entries with the shape -suru or -zuru. There are many verbs classified as irregular sahen which have alternative shapes ending in -su and these double entries were excluded. The irregular verb aisuru 'love', for example, has the alternate regular form aisu, which is classified by the Daijirin as a doosago (verb-sa-five), or regular consonant verb. These alternate forms, which have double entries in the Daijirin, were excluded by hand count from the total count of suru verbs. It is important to note that the Daijirin, just like large one-volume Japanese to English dictionaries, such as

Kenkyusha's Japanese-English Dictionary (1974), do not separately classify compound verbs like benkyoo suru 'study'. In both dictionaries the lexical item

benkyoo 'study' is classified as a noun, and the note is made that benkyoo may be used with suru, but there is no separate entry for the compound verb benkyoo suru 'study'. The highly productive process that combines nouns and suru is therefore not reflected in the total entry count for suru. Nearly all of the verbs that would be expected to have relatively high token-frequency in modem spoken Japanese are not found in the dictionary as independent entries. Compound verbs like kekkon suru 'get married',

yoyaku suru 'make a reservation' and ryoori suru 'cook' are not counted as separate lexemes. There are many more regular vowel-stem verbs (2,314) than irregular verbs (458). A summary of vowel-stem verbs, based on dictionary classification and the total count 170

is given in Table 82. A summary of consonant-stem verbs, based on the same dictionary classification system is given in Table 83. a ka ga sa za

Verb classification One-row (kamiichi) i-stem One-row (shimoichi) e-stem

ta da na ha ba ma ya

16 14 15 3 88 24 0 3 251 444 196 196 11 155 44 45

4 37 39 8 24274

ra wa

1 10 0406



260 6 1 2,054 2,314

TABLE 82. Modern and classical regular one-row (vowel-stem) verbs, in each category. Verb classification a ka ga sa ta na ba ma ra 3 552 90 1,230 143 4 53 447 1,368 Five-row (godan) TABLE 83. Modern five-row (consonant-stem) verbs, in each category.

wa 398

Total 4,288

Finally, a summary of all results, based on stem-shape, is presented in Table 84. The presentation in Table 84 is based on shape of the verb in the dictionary citation entry and the conventional linguistic analysis, which distinguishes regular vowel-stem verbs, regular consonant-stem verbs and two types of irregular verb. It is clear that there are many more regular verbs than irregular verbs. There are 6,602 regular verbs, but only 458 irregular verbs. Verb type

Verb shape (dictionary citation form)



'come' kuru 'do' suru -iru -eru -Vu, -ku, -gu, -su, -tu, -nu, -bu, -mu, -ru

5 .0007 453 .0642 260 .0368 2,054 .2909 4,288 .6074 7,060 TABLE 84. Summary of modern Japanese verb type-frequency in Daijirin on-line dictionary.

Irregular verb Irregular verb Regular vowel-stem -i verbs Regular vowel-stem -e verbs Regular consonant-stem verbs

However, unlike English regular verbs, Japanese regular verbs are not all regular in the same way. Table 85 provides a type-frequency summary for these thirteen Japanese verb-types. With the exception of irregularity in the paradigms of ar-u 'have, be', ik-u 'go, and the five polite honorific verbs (irassyar-u 'be', kudasar-u 'give', nasar-u 'do', ossyar-u 'say', and gozar-u 'be, have') the thirteen verb types listed in

Table 85 are exhaustive.


Verb type


-e Regular vowel-stem Regular consonant-stem -r Regular consonant-stem -s Regular consonant-stem -k Irregular -i/-u Regular consonant-stem -m Regular consonant-stem -w Regular vowel-stem -i Regular consonant-stem -t Regular consonant-stem -g Regular consonant-stem -b Irregular -i/-o/-u Regular consonant-stem -n

Example tabe-ru kaer-u hanas-u kak-u su-ru yom-u ka-u mi-ru mat-u oyog-u yob-u ku-ru sin-u

eat return, go back speak write do read buy watch, see wait swim call, summon come die



2,054 1,368 1,230 552 453 447 401 260 143 90 53

.291 .194 .174 .078 .064 .063 .057 .039 .020 .013 .008 .001 .001

5 4

7,060 TABLE 85. Relative type frequency of modem Japanese verbs by sub-paradigm.


At the beginning of this Appendix, five research questions were raised. What follows is a summary of the research findings with regard to those five questions.

Ql. Out of a total 233,000 entries claimed for the on-line Daijirin dictionary, a total of 10,343 verb entries were recovered. This total number of verbs includes classical as well as modem verbs.

Q2. It is possible to separate classical verb paradigms and modem spoken verb paradigms by using the entry classification system used in the Daijirin. Of 10,343 verb entries, 3,283 entries belonged to exclusively classical categories. If these same verbs exist today in modem Japanese, they are no longer conjugated in the same way. Of a total of 10,343 verb entries, 7,060 entries are conjugated using the modem paradigms. While it is possible to separate classical and modem verbal paradigms, it is not possible to separate classical verbs and modem spoken verbs as lexical items. There 172

are four overlapping verb paradigms that are used in both classical and modern Japanese: kamiichidan (upper one-row), shimoichidan (lower one-row), kahenkaku (irregular ka), and sahenkaku (irregular sa). The four overlapping verb paradigms, as well as the modern godan (five-row) paradigm, include many verbs unfamiliar to even educated native speakers. The computer search conducted here considers a modern verb to be a verb that is conjugated using the modern verb paradigms. Q3. Of 7,060 modern verb entries, 458 verbs or about 6.5% are irregular verbs in

the suru 'do' and kuru 'come' paradigms, while 6,602 verbs or about 93.5% are regular verbs. Q4. Of the 6,602 regular verb entries, there are 2,054 vowel-stem, stem-final -e verbs, but only 260 vowel-stem, stem-final -i verbs. Of all regular verbs, stemfinal -e verbs are about 31.1 % (2,054/6,602) of the total and stem-final -i verbs are about 4% (260/6602) of the total. Q5. While there are about twice as many consonant-stem verbs (4,288) as vowel-

stem verbs (2,314), no single consonant-stem paradigm has nearly as many entries as the vowel-stem paradigm with stem-final -e (2,054). The consonantstem/vowel-stem distinction obscures the fact that there are about twice as many stem-final -e verbs as the next most frequent type of consonant-stem paradigm. There are 2,054 stem-final -e verbs, but only 1,368 stem-final -r verbs, only 1,230 stem-final-s verbs, and only 552 stem-final -k verbs.


APPENDIX II: POKEMON TEST FRAMES (ENGLISH TRANSLATION) TEST A, FRAME 1 This is Kabigon. Please say Kabigon. _ _ Kabigon says he wants to sleep. Please say want to sleep. _ _ Kabigon loves to sleep. Please say sleep. _ _ Kabigon will sleep to day. What about yesterday? Yesterday too he slept. Please say slept. _ _ Let's meet with some more Pokemon characters. This time please answer like you just did. TEST A, FRAME 2 This is Pukurin. Please say Pukurin. Pukurin says he wants to hoku. Please say want to hoku. _ _ Pukurin loves to blow up and get big and hoku. Please say hoku. _ _ Pukurin will hoku today too. What about yesterday? Yesterday too he _ _ TEST A, FRAME 3 This is Ratta. Please say Ratta. _ _ Ratta says he wants to bau. Please say bau. _ _ Ratta loves to lift his whiskers up and bau. Please say bau. _ _ Ratta will bau today too. What about yesterday? Yesterday too he _ _


TEST A, FRAME 4 This is Mariru. Please say Mariru. _ _ Mariru says he wants to ritai. Please say ritai. _ _ Mariru loves to shake his tail and riru. Please say riru. _ _ Mariru will riru today too. What about yesterday? Yesterday too he _ _ TEST A, FRAME 5 Oh, look! Its Kabigon. Kabigon says he want to sleep. Please say want to sleep. _ _ Do you think Kabigon will sleep more? Please say sleep. _ _ I wonder it's fun to sleep. Let's try it together. Together with Kabigon Let's sleep. Please say, let's sleep. _ _ TEST A, FRAME 6 Oh, look! It's Pukurin. Pukurin says he wants to hoku. Please say want to hoku. _ _ Do you think Pukurin will get big again and hoku? I wonder if it's fun to hoku. Let's try it together. Together with Pukurin _ _ TEST A, FRAME 7 Oh, look! It's Ratta. Ratta says he wants to bau. Please say want to bau. _ _ Do you think Ratta will raise his whiskers again and bau? I wonder if it's fun to bau. Let's try it together. Together with Ratta _ _ TEST A FRAME 8 Oh, look! It's Mariru. Mariru says he wants to riru. Please say want to riru _ _ Do you think Mariru will raise his whiskers again and riru? I wonder if it's fun to riru. Let's try it together. Together with Mariru _ _ 175

TEST A FRAME 9 Oh, look! It's Kabigon. Kabigon says he wants to sleep. Please say want to sleep. _ _ I wonder if Kabigon will sleep tomorrow too. Please say sleep. _ _ Today he will sleep, but what about tomorrow? No, tomorrow he won't sleep. Please say won't sleep. _ _ TEST A FRAME 10 Oh, look! It's Pukurin. Pukurin says he wants to hoku. Please say want to hoku. _ _ I wonder if Pukurin will hoku tomorrow too. Please say hoku. _ _ Today he will hoku, but what about tomorrow? No, tomorrow _ _ TEST A FRAME 11 Oh, look! It's Ratta. Ratta says he wants to bau. Please say want to bau. _ _ I wonder if Ratta will bau tomorrow too. Please say bau. _ _ Today he will bau, but what about tomorrow? No, tomorrow _ _ TEST A FRAME 12 Oh, look! It's Mariru. Mariru says he wants to riru. Please say want to riru. _ _ I wonder if Mariru will riru tomorrow too. Please say riru. _ _ Today he will riru, but what about tomorrow? No, tomorrow _ _


REFERENCES AKlNAGA, KATSUE. 1966. Kyootsuugo no akusento [Accent of the common language]. Nihon Hoosoo Kyookai [Japan broadcasting association]. 1966.45-90. AKusu-Koc, AYHAN, and DANIEL SLOBIN. 1985. The acquisition of Turkish. The crosslinguistic study of language acquisition, vol. 1, ed. by Daniel Slobin, 839878. Hillsdale, NJ: Lawrence Earlbaum. AMANO, SHIGENARI, and KIMIHISA KONDO. 2000. Goi-tokusei (Lexical properties of Japanese) Vol.7. Tokyo: Sanseido. ANTILLA, RAIMO. 1974. Allomorfien semiotiikaa, ViriWija. 78.331-337. - . 1989. Historical and comparative linguistics. Amsterdam/Philadelphia: John Benjamins. ASAill.COM. 2003. Daijirin. ASHWORTH, DAVID. and PETER LINCOLN. 1973. Loanwords and the morphology of the Japanese verb. Papers in Japanese Linguistics 2.2.1-16. BATCHELDER, ELEANOR. 1999. Rule or rote? Native-speaker knowledge of Japanese verb inflection. Proceedings of The Second International Conference on Cognitive Science and The 16th Annual Meeting of the Japanese Cognitive Science Society Joint Conference (ICCS/JCSS99), 141-146. Tokyo, July 2730, 1999. - , and ATSUKO OHTA. 2000. Rule vs. rote in Japanese verb inflection. LACUS 26. ed. by Alan Melby and Arle Lommel 55-66. Fullerton, CA: LACUS. BATES, ELIZABETH, and BRIAN MACWHINNEY. 1987. Competition, variation, and language learning. Mechanisms of language acquisition. ed. by Brian MacWhinney. New Jersey: Earlbaum. BERKO, JEAN. 1958. The child's learning of English morphology. Word 14.150-177. BLOCH, BERNARD. 1946. Studies in colloquial Japanese I: Inflection. Bernard Block on Japanese, ed. by Roy Miller, 1-24. New Haven: Yale University Press. BLOOM, PAUL. 1993. Overview: Controversies in language acquisition. Language acquisition core readings ed. by Paul Bloom, 5-48. Cambridge: MIT Press. BROWN, ROGER. 1973. A first language: The early stages. Cambridge, MA: Harvard University Press. BURCH, ALFRED. 2002. Verb inflectional morphology: A study of paradigmatic understanding in native and non-native speakers. Unpublished ms. University of Hawai'i, Department of East Asian Languages and Literatures.


BYBEE, JOAN. 1985. Morphology: A study of the relation between meaning and form. Amsterdam: Benjamins. - . 1988. Morphology as lexical organization. Theoretical morphology: approaches in modem linguistics, ed. by Michael Hammond and Michael Noonan, 119141. San Diego: Academic Press. - . 1995. Regular morphology and the lexicon. Language and Cognitive Processes 10.425-455. - . 2000. Phonology of the lexicon. Usage-based models of language, ed. by Michael. Barlow and Suzanne Kemmer, 65-85. Stanford: CSLI. - . 2001. Phonology and language use. Cambridge, UK: Cambridge University Press. - , and CAROL MODER 1983. Morphological classes as natural categories. Language 59.251-271. and ELLY PARDO. 1981. On lexical and morphological conditioning of alternations: a nonce-probe experiment with Spanish verbs. Linguistics 19.937-968. CAMPBELL, LYLE. 1974. Theoretical implications of Kekchi phonology. International Journal of American Linguistics 40.269-278. CHOMSKY, NOAM. 1957. Syntactic structures. The Hague: Mouton. CHANDLER, STEVE. 1993. Are rules and modules really necessary for explaining language? Journal of Psychological Research 22.593-606. CLAHSEN, HARALD; M. ROTHWEILER; ANDREAS WOEST; and GARY MARCUS. 1992. Regular and irregular inflection in the acquisition of German noun plurals. Cognition 45. 225-255. - ; SONJA EISENBEISS; MEIKE HADLER; and INGRID SONNENSTUHL. 2001. The mental representation of inflected words: an experimental study of adjectives and verbs in German. Language 77.3.510-543. CLANCY, PATRICIA. 1985. The acquisition of Japanese. The crosslinguistic study of language acquisition, vol. 1, ed. by Dan I. Slobin, 373-524. Hillsdale, NJ: Lawrence Erlbaum. CORRIGAN, ROBERTA. and SUSAN LIMA. 1994. Introduction. The reality of linguistic rules, ed. by Susan. Lima; Roberta Corrigan; and Gregory. Iverson, xii-xxiii. Philadelphia: Benjamins. DABROWSKA, EWA. 2001. Learning a morphological system without a default: the Polish genitive. Journal of Child Language 28.545-574.


DE CHENE, BRENT. 1982. The segmentation of Japanese verbs: Experimental evidence. Papers in Japanese Linguistics 10.170-207. - . 1985. -r epenthesis. Papers in Japanese Linguistics 10.170-207. DERWING, BRUCE, and ROYAL SKOUSEN. 1994. Productivity and the English past tense: Testing Skousen's analogy model. The reality of linguistic rules, ed. by Susan. Lima; Roberta Corrigan; and Gregory. Iverson, 193-218. Philadelphia: Benjamins. ELMAN, JEFFREY. 2003. The great past tense debate: a very select bibliography. retrieved 2/27 2003 http://crI.ucsd.edul-elman/Courses/cog202/Papers/pasttense-bibliography.pdf. ERVIN-TRIP, SUSAN. 1966. Language development. Review of child development research ed by L. Hoffman. NewYork: Russell Sage Foundation 2.55-105. FRANCIS, W. NELSON, and HENRY KUCERA. 1982. Frequency analysis of English usage: Lexicon and grammar. Boston: Houghton Mifflin. FUJIWARA, YOOICHI. 1977. Yooji no gengo hyoogen nooryoku no hattatsu. Hiroshima: Bunka Hyooron Publishing Co. GARDNER, ELIZABETH. 1950. The inflections of modern literary Japanese. Language dissertation no 4. Supplement to Language. Baltimore: Linguistic Society of America. GIVON, TALMY. 1984. Syntax. A functional typological introduction. Vol 1. Amsterdam: John Benjamins. GONNERMAN, LAURA, 1999. Morphology and the lexicon: Exploring the semanticsphonology interface. Los Angeles, CA: University of Southern California dissertation. GREENBERG, JOSEPH. ed. 1963. Universals of language. Cambridge, Mass.:MIT Press. HALE, KENNETH. 1973. Deep-surface canonical disparities in relation to analysis and change: An Australian example. Current trends in Linguistics, vol. 11, ed. by Thomas. A. Sebeok. 401-458. The Hague: Mouton. HANKAMER, JORGE. 1992. Morphological Parsing and the lexicon. Lexical representation and process, ed. by William. Marslen-Wilson, 392-408. Cambridge, MA: MIT Press. HARE, MARY, and JEFFREY ELMAN. 1995. Learning and morphological change. Cognition 56: 61-98. - ; MICHAEL FORD; and WILLIAM MARSLEN-WILSON. 2001. Ambiguity and frequency effects in regular verb inflection, Frequency and the emergence of linguistic structure, ed. by Joan Bybee and Paul Hopper, 181-200. Philadelphia: Benjamins. 179

HA YDEN, JEFFREY. 2002. Network model of Japanese past. Unpublished ms. University of Hawai'i, Department of East Asian Languages and Literatures. - , and TERRY KLAFEHN. 2003. Network model of Japanese past. Unpublished ms. University of Hawai'i, Department of East Asian Languages and Literatures. HOCKETI, CHARLES. 1954. Two models of grammatical description. Word. 10.2-3.210234. - . 1968. The state of the art. The Hague: Mouton. IKEDA,TADASHI. 1975. Classical Japanese grammar illustrated with texts. Tokyo: Toho Gakkai [Institute of Eastern Culture]. ITOO, K. 1976. Yooji ni tottte 'ayamari' to wa nani ka. Development in verbal and nonverbal behavior: Hattatsu to shuutoku ni okeru gengo koodoo. ed. by F. C. Pengo Hiroshima: Bunka Hyooron Publishing Co. IWANAMI SHOTEN. 1998. Joojien, ed. by Izuru Shinmura. Tokyo: Iwanami Shoten. JAPAN ELECTRONIC RESEARCH DICTIONARY INSTITUTE, LTD. 1995. The EDR dictionaries. Tokyo: The Japan Research Dictionary Institute. JAEGER, JERI; ALAN LOCKWOOD; DAVID KEMMERER; ROBERT VAN VALIN; BRIAN MURPHY; and HANIF KHALAK. 1996. A positron emission tomographic study of regular and irregular verb morphology. Language. 72.451-497. JESPERSEN, OTTO. 1942. A modem English grammar on historical principles, VI. reprinted 1961: London: George Allen & Unwin Ltd. JOHANSSON, STIG, and KNUT HOFLAND. 1986. Frequency analysis of English vocabulary and grammar. Oxford: Clarendon Press. Joos, MARTIN. 1957. Readings in linguistics. Washingon: American Council of Learned Societies. JORDEN, ELIZABETH, and MARl NODA. 1987. Japanese: The spoken language: Part 1. New Haven: Yale University Press. - . 1988. Japanese: The spoken language: Part 2. New Haven: Yale University Press. -.1990. Japanese: The spoken language: Part 3. New Haven: Yale University Press. KANAGY, RUTH. 1991. Developmental Sequences in the Acquisition of Japanese as a Foreign Language; The Case of Negation. Ann Arbor: Univ. Microfilms International. KENKYUUSHA. 1974. Kenkyuusha's new Japanese English dictionary, ed. by Koh Masuda. Tokyo: Kenkyuusha.


KIM, JOHN; STEVEN PINKER, ALAN PRINCE, and SANDEEP PRASADA. 1991. Why no mere mortal has ever flown out to center field. Cognitive Science 15.173218. KATOO, EIKO. 2001. Mother child interactions during book reading. Paper presented at third annual conference of the Japanese society for language sciences, June 24,2001. Japan Women's University, Tokyo, Japan. KIM, JOHN; GARY MARCUS, STEVEN PINKER, MICHELLE HOLLANDE; and MARIE COPPOLA. 1994. Sensitivity of children's inflection to grammatical structure. Journal of Child Language, 21.173-209. KIPARSKY, PAUL. 1975. Remarks on analogical change. Historical linguistics vol. 2. ed. by J.M. Anderson and C. Jones. Amsterdam: North-Holland. KLAFEHN, TERRY. 2001a. Japanese verbal morphology: Rote, rule, or analogy? Comparison of native and learner performance on a nonce form recognition test. Working Papers in Linguistics. Department of Linguistics, University of Hawai'i at Manoa. 200012001. 32.1-25. - . 2001b. Japanese verbal inflection: Words and rules or words and words? The proceedings of the third annual conference of the Japanese society for language sciences. ed. by Mineharu Nakayama. Tokyo: Kuroshio, to appear. KOPCKE, KLAUS-MICHAEL. 1988. Schemas in German plural formation. Lingua. 74.303-335 KUCZAJ, STAN. 1976. "-ing," "-s,", & "-ed': A study of the acquisition of certain verb inflections. Unpublished doctoral dissertation, University of Minnesota, Department of Psychology. - . 1977. The acquisition of regular and irregular past tense forms. Journal of Verbal Learning and Verbal Behavior. 16.589-600. LANGACKER, RONALD. 1987. Foundations of cognitive grammar, volume 1: Theoretical prerequisites. Stanford, CA: Stanford University Press. - . 2000. A dynamic usage-based model. Usage-based models of language, ed. by Michael. Barlow and Suzanne Kemmer, 1-63. Stanford, CA: CSLI. LEHMANN, WINFRED. 1983. Language: An introduction. New York: Random House. MACWHINNEY, BRIAN. 1978. The acquisition of morphophonology . Mongraphs of the Society for Research in Child Development. serial number 174, volume 43. - . 1995. The CHILDES project: Tools for analyzing talk. Hillsdale, NJ: Lawrence Erlbaum.


- . 2003a. Extending the competition model. Bilingual sentence processing. ed. by Roberto Heredia and Jeanette Altaririba. Amsterdam: North-Holland. to appear. - . 2003b. Rethinking the logical problem of language acquisition. to appear. - , and JARED LEINBACH. 1991. Implementations are not conceptualizations: Revisiting the verb learning model. Cognition 40.121-147. MADDIESON, IAN. 1984. Patterns of sounds. Cambridge: Cambridge University Press. MARCUS, GARY; STEVEN PINKER; MICHAEL ULLMAN; MICHELLE HOLLANDE; T. JOHN ROSEN; and FEI XU. 1992. Overregularization in language acquisition. Monographs of the Society for Research in Child Development 57.4. - ; URSULA BRINKMANN: HARALD CLAHSEN; RICHARD WIESE; and STEVEN PINKER. 1995. German inflection: the exception that proves the rule. Cognitive Psychology 29.189-256. - . 2001. The algebraic mind: Integrating connectionism and cognitive science. Cambridge. Mass: The MIT Press. MARTIN, MAISA. 1995. The map and the rope: Finnish nominal inflection as a learning target. Jyvaskyla, Finland: Jyvaskyla University Printing House. MARTIN, SAMUEL. 1952. Morphophonemics of standard colloquial Japanese. Language dissertation no 47. Supplement to Language. Baltimore: Linguistic Society of America. -.1988. A Reference Grammar of Japanese. Rutland, VT: Tuttle. MATTHEWS. PETER. 1991. Morphology. Cambridge: Cambridge University Press. MCCAWLEY, JAMES. 1968. The phonological component of a grammar of Japanese. The Hague: Mouton. MCCLAIN, YOKO. 1981. Handbook of modem Japanese Grammar. Tokyo: The Hokusei Press. MCCLELLAND, JAMES, and DAVID RUMELHART. 1981. An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review. 88.375-407. MENCKEN, HENRY. 1936. The American language. New York: Knopf. MILLER, Roy. 1967. The Japanese language. Chicago: University of Chicago Press. MIYATA, SUSANNE. 1992. Wh-questions of the third kind: The strange use of waquestions in Japanese children. Bulletin of Aichi Shukutoku Junior College, no 31.151-155.


- . 1993. Japanische Kinderfragen: Zum Erwerb von Form-Inhalt-Funktion von Frageausdruecken. Hamburg: OAG. - . 1995. The Aki corpus: Longitudinal speech data of a Japanese boy aged 1.6-2.12. Bulletin of Aichi Shukutoku Junior College. 34.183-191. - . 2000. The TAl corpus: Longitudinal speech data of a Japanese boy aged 1;5.203;1.1. Bulletin of Shukutoku Junior College, 39.77-85. NICHIEISHA. 1952. Kokubunpoo yookoo. Tokyo: Nichieisha. NHK (Nippon Hoosoo Kyookai [Japanese Broadcasting Company]). 1992. Nihongo hatsuon akusento jiten (Japanese pronunciation accent dictionary). Tokyo: Nihon Hoosoo [Japan Broadcasting]. OSHIMA-TAKANE, YURIKO. and BRIAN MACWHINNEY. eds. 1998. CHILDES manual for Japanese. Montreal: McGill University / Nagoya: Chukyo University. PAUL, HERMANN. 1880 (1960). Prinsipien der Sprachgeschicte. TbingeUn: Max Niemeyer. PETERS, ANN. 1983. The units of language acquisition. New York: Cambridge University Press. 1985. Language segmentation: Operating principles for the analysis and perception of language. The crosslinguistic study of language acquisition, vol. 2, ed. by Dan Issac Slobin, 1029-1067. Hillsdale, NJ: Lawrence Erlbaum. PINKER, STEVEN. 1984. Language leamability and language development. Cambridge, MA: Harvard University Press. - . 1989. Leamability and cognition: The acquisition of argument structure. Cambridge, MA: The MIT Press. - . 1991. Rules of language. Science. 253.530-535. - . 1993. Rules of language. Language acquisition core readings, ed. by Paul Bloom, 472-484. Cambridge: MIT Press. -.1994. The language instinct: How the mind creates language. New York: William Morrow. - . 1999. Words and rules: The ingredients of language. New Yark: Basic Books. - , and ALAN PRINCE. 1988. On language and connectionism: Analysis of a parallel distributed processing model oflanguage acquisition. Cognition, 28.73-193 - . 1994. Regular and irregular morphology and the psychological status of rules of grammar. The reality of linguistic rules, ed. by Susan. Lima; Roberta Corrigan; and Gregory. Iverson, 321-351. Philadelphia: Benjamins. 183

PLUNKETT, KIM, and VIRGINIA MARCHMAN. 1991. U-shaped learning and frequency effects in a multi-layered perceptron: Implications for child language acquisition. Cognition 38.43-102. - , and JEFFREY ELMAN. 1997. Exercises in rethinking innateness: a handbook for connectionist simulations. Cambridge, MA: The MIT Press. PRASADA, SANDEEP, STEPHEN PINKER, and WILLIAM SNYDER. 1990. Some evidence that irregular forms are retrieved from memory but regular forms are rule generated. Paper presented at the 31 st meeting of the Psychonomic Society, New Orleans. - , and STEVEN PINKER. 1993. Generalization of regular and irregular morphological patterns. Language and Cognitive Processes 8.1-56. RISPOLI, MATTHEW. 1981. The emergence of verb and adjective tense-aspect inflections in Japanese. Unpublished master's thesis, University of Pennsylvania. RUMELHART, DAVID, and JAMES MCCLELLAND, and the PDP Research Group 1986. Parallel distributed processing: Explorations in the microstructure of cognition 1: Foundations, USA: The MIT Press. 45-76. 1987. Learning the past tenses of English verbs: Implicit rules or parallel distributed processing. Mechanisms of language acquisition, ed. by Brian MacWhinney, 195-248. New Jersey: Earlbaum. SANDNESS, KAREN. 1999. The evolution of the Japanese past and perfective suffixes. Ann Arbor, MI: Center for Japanese Studies, the University of Michigan. SANO, TETSUYA. 1995. Roots in Language Acquisition: A Comparative Study of Japanese and European Languages, Doctoral Dissertation, University of California, Los Angeles. SATO, PAUL. 1985. Denominal verbs with -r: A response to de Chene. Papers in Japanese Linguistics 10.172-207. SANSEIDO. 1995. Daijirin, ed. by Akira Matsumura. Tokyo: Sanseido Co. SEIDENBERG, MARK, and M. BRUCK. 1990. Consistency effects in the generation of past tense morphology. Paper presented at the 31 st meeting of the Psychonomic Society, New Orleans. SELLS, PETER. 1985. Lectures on contemporary syntactic structures. An introduction to government-binding theory, generalized phrase structure grammar, and lexical-functional grammar. Stanford: CSLI. SHIBATA, TAKESHI, 1965. lkite iru hoogen. Green Series 63. Tokyo: Chikuma-Shoobo. SKOUSEN, ROYAL. 1989. Analogical modeling of language. Dordrecht: Kluwer Academic. 184

- . 1992. Analogy and structure. Dordrecht: Kluwer Academic. SLOBIN, DANIEL. 1971. Psycholinguistics. Glenview, IL: Scott, Foresman. - . ed. 1985. The crosslinguistic study of language acquisition. Vol. 1. The data. N.J: Erlbaum. SMITH, NEIL. 2000. Don't only connect. [Review of the book Words and rules: The ingredients oflanguage]. Times Literary Supplement. February 18,2000.10. SMITH, NEIL, and IANTHI TSIMPLI. 1995. The mind of a savant: Language learning and modularity. Cambridge, MA: Blackwell. SPENCER, ANDREW. 1991. Morphological theory: An introduction to word structure in generative grammar. Cambridge, MA: Blackwell. STEMBERGER, JOSEPH, and BRIAN MACWHINNEY. 1988. Are inflected forms stored in the lexicon? Theoretical approaches to morphology, ed. by Michael Hammond and Michael Noonan, 101-116. San Diego: Academic Press. THYME, ANN; FARRELL ACKERMAN; and JEFF ELMAN. 1994. Finnish nominal inflection: Paradigmatic patterns and token analogy. The reality of linguistic rules. ed. by Susan. D. Lima; Roberta L. Corrigan; and Gregory. K. Iverson, 445-466. Philadelphia: Benjamins. TOMASELLO, MICHAEL. 1992. First verbs: A case study of early grammatical development. Cambridge: Cambridge University Press. TSUKUBA LANGUAGE GROUP. 1991. Situational functional Japanese. Drills and Notes, Volumes 1-3. Tokyo: Bonjinsha Co., LTD. UNGER, JAMES. 1993. Studies in early Japanese morphophonemics. Bloomington, IN: Indiana University Linguistics Club Publications. VANCE, TIMOTHY. 1987. An introduction to Japanese phonology. New York: State University of New York Press. - . 1991. A new experimental study of Japanese verbal morphology. Journal of Japanese Linguistics 13.145-156. WELLS, RULON. 1949. Automatic alternation. Language 25.99-116. YANO, YOKO. 2003. Chotto shita hitorigoto; Furoku [appendix]. Retrieved 3/19, 2003, YOKOMIZO, SHINICHIRO. 1990. The characteristics of newly established loanverbs in correlation with social factors, Kagoshima Junshin Tankidaigaku Kenkyuu Kiyoo 20.65-87


YOKOYAMA, MASAKO. 1950. The inflection of eighth century Japanese. Language dissertation no 45. Supplement to Language. Baltimore: Linguistic Society of America. ZAGER, DAVID. 1980. A real time process model of morphological change. Buffalo, NY: SUNY dissertation.



university of hawai'! library - ScholarSpace


6MB Sizes 0 Downloads 0 Views

Recommend Documents

No documents