Morphological Structure in Language Processing [Reprint 2011 ed.] 9783110910186, 9783110178920

Citation preview

Morphological Structure in Language Processing

W

Trends in Linguistics Studies and Monographs 151

Editors

Walter Bisang Hans Henrich Hock (main editor for this volume)

Werner Winter

Mouton de Gruyter Berlin · New York

Morphological Structure in Language Processing

Edited by

R. Harald Baayen Robert Schreuder

Mouton de Gruyter Berlin · New York

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

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

Library of Congress Cataloging-in-Publication

Data

Morphological structure in language processing / edited by R. Harald Baayen, Robert Schreuder. p. cm — (Trends in linguistics. Studies and monographs ; 151) Includes bibliographical references and index. ISBN 3-11-017892-3 (alk. paper) 1. Grammar, Comparative and general — Morphology — Psychological aspects. 2. Psycholinguistics. I. Baayen, R. Harald, 1958- II. Schreuder, Robert. III. Series. P241.M5986 2003 415'9-dc22 2003044270

ISBN 3-11-017892-3 Bibliographic information published by Die Deutsche Bibliothek Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data is available in the Internet at .

© Copyright 2003 by Walter de Gruyter GmbH & Co. KG, D-10785 Berlin All rights reserved, including those of translation into foreign languages. No part of this book may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording or any information storage and retrieval system, without permission in writing from the publisher. Cover design: Christopher Schneider, Berlin Printed in Germany.

Contents Preface Inflectional morphology and word meaning: Orthogonal or co-implicative cognitive domains? Aleksandar Kos tic, Tanja Markovic and Aleksandar Baucal

vii

1

Visual processing of Italian verbs and adjectives: The role of the inflectional family size Daniela Traficante and Cristina Burani

45

Morphological resonance in the mental Nivja H. de Jong, Robert Schreuder andlexicon R. Harald Baayen

65

Morphology and frequency: Contrasting methodologies Michael A. Ford, William D. Marslen-Wilson and Matthew H. Davis

89

Derivational morphology in the German mental lexicon: A dual mechanism account Harald Clahsen, Ingrid Sonnenstuhl and James P. Blevins

125

The interplay of root, suffix and whole-word frequency in processing derived words Cristina Burani and Anna M. Thornton 157 On the role of derivational affixes in recognizing complex words: Evidence from masked priming Helene Giraudo and Jonathan Grainger

209

Morphological facilitation: The role of semantic transparency and family size Laurie Beth Feldman and Matthew John Pastizzo

233

Recognition of spoken prefixed words: The role of early conditional root uniqueness points Lee H. Wurm and Joanna Aycock 259

vi Contents Lexical representation of morphologically complex words: Evidence from Polish Agnieszka Anna Reid and William DavidMarslen-Wilson

287

Identification spoken Fanny Meunierofand Juanprefixed Segui words in French

337

Frequency effects in regular inflectional morphology: Revisiting Dutch plurals R. Harald Baayen, James M. McQueen, Ton Dijkstra and Robert Schreuder

355

How does a child detect morphology? Evidence from production Wolfgang U. Dressier, Marianne Kilani-Schoch and Sabine Klampfer

3 91

Frequency effects in processing inflected Dutch nouns: A distributed connectionist account Matthew H. Davis, Maarten van Casteren and WilliamD. Marslen-Wilson

427

When word frequencies do not regress towards the mean R. Harald Baayen, Fermin Moscoso del Prado Martin, Robert Schreuder and Lee Wurm

463

Spelling errors with a view on the mental lexicon: Frequency and proximity effects in misspelling homophonous regular verb forms in Dutch and French Dominiek Sandra and Michel Fayol

485

List of contributors

515

Subject index

519

Preface This volume brings together a series of studies of morphological processing in Germanic (English, German, and Dutch), Romance (French and Italian), and Slavic (Polish, Serbian) languages. The question of how morphologically complex words are organized and processed in the mental lexicon is addressed from different theoretical perspectives (single and dual route models), for different modalities (auditory and visual comprehension, writing), and for language development. Experimental work is reported, as well as computational and statistical modeling. The aim of this volume is to provide an overview of the range of issues currently attracting research at the intersection of morphology and psycholinguistics. The role of morphological complexity for lexical processing has been studied most extensively for the visual domain. In this volume, three lines of current research are represented. The first line of research addresses the role of paradigmatic relations between complex words in the mental lexicon. The chapter by Kostic, Markovic, and Baucal introduces an information-theoretic approach to the inflectional paradigms in Serbian. The consequences of the size of inflectional paradigms in Italian for lexical processing is studied in the chapter by Traficante and Burani. Paradigmatic relations are also present outside the domain of inflection. While paradigmatic relations in inflection are often highly structured by features such as case, number, person, tense, and aspect, the paradigmatic relations in a morphological family brought about by derivation and compounding are only loosely structured. Nevertheless, the size of the morphological family has been observed to be a relevant predictor for lexical processing. The chapter by De Jong, Schreuder, and Baayen explores the way in which the family size effect is modulated by linguistic context. The chapter by Ford, Marslen-Wilson, and Davis studies the family size effect in English in its relation to a range of other frequency measures. Feldman and Pastizzo report an investigation of the relation between family size and semantic transparency. A second line of research in visual studies represented in this volume addresses the role of morphological decomposition in lexical processing. The chapter by Clahsen, Sonnenstuhl, and Blevins addresses the role of decomposition for German productive derivational suffixes, and reports frequency effects for the derived forms combined with full cross-modal priming for those forms. The chapter by Burani and Thornton studies the contribution of root and affix to morphological decomposition for high and low frequency complex words.

viii

Preface

A third line of research addresses the question of whether morphological regularities affect lexical processing during the early stages of visual identification. The chapter by Giraudo and Grainger reports experimental research suggesting that morphological effects arise only after visual identification has been completed. Four studies in this volume address aspects of the processing of complex words in the auditory modality. The chapers by Wurm and Aycock and Meunier and Segui investigate the processing of prefixed words, the chapter by Baayen, McQueen, Dijkstra, and Schreuder studies suffixed words. Wurm and Aycock provide further evidence for the relevance of the conditional root uniqueness point in English. Meunier and Sequi show for French that the stem of prefixed words is not processed as if there were no preceding prefix. Baayen et al. document frequency effects for regular inflected words in Dutch. The chapter by Reid and Marslen-Wilson reports a series of experiments in Polish documenting stem and affix priming, a suffix-suffix interference effect similar to that observed in earlier work for English, as well as the importance of semantic transparency. The final four chapters complement the preceding studies in visual and auditory lexical processing. The chapter by Dressier, Kilani-Schock, and Klampfer addresses the question how children acquire morphology. The chapter by Davis, Van Casteren, and Marslen-Wilson shows by means of distributed connectionist modeling that data previously argued to support a dual-route model can also be accounted for within a single-route framework. The issue of the reliability of corpus-derived frequency data for experimental designs using frequency measures as diagnostics for decomposition is discussed in the chapter by Baayen, Moscoso del Prado Martin, Wurm, and Schreuder, in reply to some of the issues raised in the chapter by Ford, Marslen-Wilson, and Davis. In the final chapter, Sandra and Fayol address the role of morphological structure in written language production, documenting for Dutch and French how frequency and proximity give rise to spelling errors for homophonous regular verb forms. The editors would like to thank Lanneke van Dreumel for editorial help.

Inflectional morphology and word meaning: Orthogonal or co-implicative cognitive domains? Aleksandar Kostic, Tanja Markovic and Aleksandar Baucal Typically, linguistics and psycholinguistics treat word meaning and inflectional morphology as two domains that do not interact. The aim of the present study is to evaluate this assumption empirically. The evaluation is based on a contrast between two types of information probabilities for inflected noun forms in Serbian. One, derived from all nouns of a particular gender (e.g., accusative singular in feminine gender) is generic across words and thus free of meaning. This is a "type" probability. The other is related to the probability of a particular word in a particular inflected form, and is a "token" probability (e.g., the accusative for the noun meaning "chair"). Because it is specific to a particular noun, "token" probability is related to its meaning. A pattern whereby the cognitive system is sensitive only to information of type probability of inflected forms, would be consistent with the claim that inflectional morphology and word meaning are cognitively independent domains. By contrast, an outcome whereby the cognitive system is sensitive to information of token probabilities would suggest that the two domains are not truly independent. In a lexical decision experiment, six inflected forms of feminine nouns were presented. Nouns were divided into two groups that differed significantly in the patterning of the amount of information carried by six inflected forms. This difference derived from differences in token probability only. Variability in the patterning of decision latencies to inflected forms was predicted by the average amount of information for both type and token probabilities. This finding indicates that the cognitive system is sensitive to "token" as well as type probabilities. Because the "token" probability distribution is word specific, this outcome suggests that inflectional morphology constitutes an intrinsic part of a word's meaning. Accordingly the domains of word meaning and inflectional morphology should be treated as coimplicative rather than independent cognitive domains.

1. Introduction Linguists and psycholinguists traditionally treat word meaning and inflectional morphology and syntax as two orthogonal aspects of language. Thus, for example, a particular syntactic structure can encompass an infinite number of grammatical sentences that differ in meaning. Likewise, the

2

Kostic, Markovic and Baucal

taxonomy and classification of grammatical forms as offered by descriptive linguistics derive from a set of morphological alternations within a particular word class, irrespective of the meaning of the individual words that belong to that class. The declension of a noun of a particular gender, for example, is specified in terms of the morphological properties of cases singular and plural, but not in terms of the meaning of the noun. Likewise, conjugation specifies morphological properties of a verb including person and grammatical number, irrespective of the meaning of the particular verb. In general, the taxonomy and classification of grammatical forms of different word types is based on generic morphological properties of a particular word class or subclass within a particular word type (e.g., feminine nouns). With the above considerations in mind, inflectional morphology and word meaning satisfy two formal requirements of separability. The first requirement is the absence of an interaction between the two domains. Inflectional alternations within a defined paradigm (say declension) do not alter word meaning. The second requirement is that a set of inflectional alternations should apply to all instances of a defined class. In the case of nouns this means that all nouns can appear in all seven cases and two grammatical numbers. In that respect, a formal description of inflectional morphology is, in principle, distinct from the meaning of individual words. Psycholinguists adopt a similar position. In addition to accepting a formal distinction between these two aspects of language, psycholinguists assume that inflectional morphology and word meaning are also distinct cognitive domains that do not interact. Morphologically complex words are described in terms of their base form and affixes. While the base form is idiosyncratic to a particular word and relates to its meaning, the repertoire of affixes is a generic property that pertains to a class of words. Stated generally, psycholinguistic models that describe how we process morphologically complex words embrace, with no exception, the distinction between inflectional morphology and word meaning (cf. Taft and Forster, 1975; 1976; Taft 1979,1981, 1994). The aim of the present study is to evaluate the orthogonal status of the two domains and reexamine the cognitive status of inflectional morphology as independent from word meaning. The basis for such an evaluation distinguishes between two kinds of probabilities for inflected noun forms: the probabilities of inflected forms per se, irrespective of word meaning (e.g. if word is a feminine noun, what is the probability that it will end with a suffix e), and the probabilities of inflected forms for an individual word (e.g.

Inflectional morphology and word meaning

3

what is the probability that the feminine noun χ will end with a suffix e) where the unique probability distribution presumably pertains to that word's idiosyncratic properties, meaning being the most plausible candidate. By implication, a cognitive system that is sensitive to the unique probability patterning of the inflected forms of an individual noun calls into question the status of inflectional morphology as orthogonal to word meaning. In the present paper, we examine the cognitive status of inflectional morphology in Serbian, which is a highly inflected language with the potential for extensive experimental investigation of inflected forms. In the forthcoming paragraphs, we first describe the morphological properties of Serbian and results of several studies. Our interpretation assumes that processing of inflected noun forms is determined by the amount of information carried by an inflected form. Because we specify information in the terms of information theory, we refer to it as an Informational Approach. We then elaborate on the general assumptions and our empirical evaluation of this approach in subsequent sections.

1.1. The noun system of Serbian Serbian is a highly inflected and, to a great extent, a free word language. Forms of nouns, adjectives, verbs and some pronouns include inflectional suffixes that specify the syntactic roles that a particular word type can have in a sentence context. Thus, for example, when an inflectional suffix is added to the base form of a Serbian noun it specifies its case, grammatical number and grammatical gender. This is also true for adjectives and some pronouns. When added to the base form of a verb, the suffix denotes person, grammatical number, tense and sometimes grammatical gender. Each Serbian noun can appear in seven cases in singular and in plural, each case being marked by an inflectional suffix (Table l).1 Inspection of Table 1 indicates that the declension of Serbian nouns is characterized by a number of homomorphs due to the fact that sometimes grammatical forms of a noun share a common inflectional suffix. While in natural language homomorphs are disambiguated by context or by stress, in the visual lexical decision task, when a noun is presented in isolation, some forms remain ambiguous with respect to case and sometimes grammatical number (e.g. the inflected form vod-e can be either a genitive singular or a nominative plural of noun vod-a (water).

4

Kostic, Markovic and Baucal

Table 1. Declension of masculine, feminine and neuter nouns Masculine Feminine Singular Plural Singular Plural Case vod-e konj-i vod-a b Nominative konj a a vod-a Genitive konj-a konj-a vod-e konj-u konj-ima vod-i vod-ama Dative vod-e Accusative konj-a konj-e vod-u vod-ama Instrumental konj-em konj-ima vod-om konj-ima vod-i Locative konj-u vod-ama a. konj - horse; b. voda - water, c. selo - village

Neuter Singular Plural sel-a sel-oc sel-a sel-a sel-u sel-ima sel-a sel-o sel-om sel-ima sel-ima sel-u

This ambiguity necessitates a clear distinction between grammatical form and inflected form. Although linguists do not make a distinction between the two, in the present study the term grammatical form is reserved for grammatical instances as, for example, a feminine noun in nominative singular, or a feminine noun in genitive plural, while the term inflected form will refer to a noun form with a particular inflectional suffix, irrespective of the cases encompassed by that suffix, i.e. irrespective of its grammatical status (e.g. vod-e) The inflected forms of Serbian feminine nouns and the cases they encompass are presented in Table 2. The inflected forms for masculine, feminine and neuter nouns and the cases they encompass are presented in Appendix B. The probability of Serbian inflected noun cases differ and, by the same token, there are probability differences for inflected noun forms as well. The probabilities of inflected cases and inflected forms for all three grammatical genders are given Appendices A and Β (from Kostic, D., 1965a). An inflectional suffix when attached to a noun specifies its case, grammatical number and grammatical gender. Of these three attributes, case is the most complex because it encompasses a number of syntactic functions and meanings that a noun can assume in a sentence context. Take, for example, the nominative case. It can denote a subject role, as in a sentence Petar je ucitelj. (Peter is a teacher.), but it can also denote a predicate role as in the same sentence: Petar je ucitelj. (Peter is a teacher). The accusative case mainly denotes an object role: Uzeo je svoju knjig-u. (He took his book.), but it can also denote time Dosao je u sred-u. (He came on Wednesday.), place Otisli su u Moskv-u. (They went to Moscow.), cause On je odgovoran za njihovu nesrec-u. (He is responsible for their tragedy.). Other cases likewise are characterized by a number of syntactic functions and meanings. The syntactic functions and meanings of a noun

Inflectional morphology and word meaning

5

can be specified either by contextual factors as well as by case alternation. While linguists agree that case is indeed characterized by the number of syntactic functions and meanings, there is neither consensus about the distinction between function and meaning, nor agreement about the taxonomy of case functions and meanings (cf. Hjelmslev 1937; de Groot 1956; Kurilowicz 1960; Benveniste 1962; Diver 1964; Blake 1994). The taxonomy of Serbian case functions and meanings, presented in Appendix C, was compiled from six standard grammar books in the early sixties (Kostic, D. 1965b). The absolute number of case functions and meanings, presented in Appendices A, Β and C, should be taken as tentative: What is more meaningful is the proportion of functions and meanings encompassed by a particular case, relative to other cases. Table 2. Morphologically unique forms of regular feminine nouns and cases they encompass Form Vod-aa Vod-e vod-i vod-u vod-om vod-ama a. voda - water

Case n. s. + g. p. g. s. + n. p. + a. p. d. s. +1. s. a. s. i. s. d. p. + 1. p. + i. p.

In light of the properties of Serbian nouns delineated above, we can ask how we process inflected noun forms. The basic spirit of two standard approaches, the full listing hypothesis (Manelis and Tharp 1977; Butterworth 1983) and decomposition account (Taft and Forster 1975, 1976; Taft 1979, 1981) is that processing latency for inflected noun forms should be some function of probability (frequency). While the full listing hypothesis makes no difference between monomorphemic and morphologically complex words, assuming that surface frequency of a word is the pivotal factor that affects processing latency, more recent versions of the decomposition model posit implicit morphological parsing where each morpheme representation contributes to the overall activation whenever there is an correspondence between letter-string and morpheme (Taft 1994; Taft and Zhu 1995). There are also models that incorporate aspects from both whole word and decomposed approaches as, for example, the Augmented Addressed Morphology Model (AAM) (Burani and Caramazza 1987;

6

Kostic, Markovic and Baucal

Caramazza, Laudana, and Romani 1988). While the role of frequency in the processing of morphologically complex words has not been challenged, a number of recent studies suggest a distinction between surface frequency (i.e., the frequency of a presented form) and cumulative base frequency (i.e., the sum of the frequencies of all affixed forms that share a particular base morpheme). Typically, processing latency for words of the same surface frequency vary as a function of their cumulative frequency: Items of higher cumulative frequency tend to be processed faster than those of lower cumulative frequency. In other words, to predict processing latency for a morphologically complex word not only its own frequency but also the frequency of all other words that share the same base morpheme must be considered (cf. Cole, Segui, and Taft 1997; Baayen, Dijkstra, and Schreuder 1997; Sereno and Jongman 1997). Many studies that investigate processing of inflectional morphology are done in English or Dutch, languages that are not characterized by a particularly rich inflectional morphology. As a consequence, in an experiment with a lexical decision task relatively few inflected endings can be presented to participants. The results are generally analyzed with an analysis of variance, with a focus on the ordinal relations among reaction times for different inflected forms that are presented within an experiment. This ordinal relation is usually interpreted in terms of the particular form's probabilities: the ordinal relationship between fast/slow on the latency continuum purportedly reflects the ordinal relationship between high/low on the frequency continuum. In other words, faster and slower reaction times parallel higher and lower probabilities. Because Serbian is a highly inflected language, where nouns of a particular gender can appear in five and sometimes more morphologically distinct forms, it allows for correlational analysis between the probability of a form and its processing latency. In effect, it is possible to make a direct test as to whether processing latency for inflected forms is a function of the frequency of their suffix. Furthermore, using regression analysis, it is possible to specify what proportion of variability in the processing of Serbian inflected noun forms can be attributed to the variation among their suffix probabilities. In a series of experiments that included all morphologically distinct forms of masculine, feminine and neuter nouns in Serbian it was demonstrated that the probability of an inflected forms in a particular grammatical gender does not correlate significantly with processing latency (Kostic, A. 2003, submitted). Such an outcome suggests that a predictor restricted to

Inflectional morphology and word meaning

7

the probability of an inflected form alone may not be adequately specified and that there might be yet another factor that contributes to the processing latency of inflected noun forms.

2. The informational approach to processing inflectional morphology The processing latency for an inflected form should, in principle, be related to its complexity relative to other forms of the same word. The more complex the form, the longer its processing latency. As traditionally conceived, there were two criteria of complexity. One criterion, captured by the decomposition approach, is based on structural properties of a word; Affixed words are more complex than non-affixed ones because they require morphological parsing (Taft and Forster, 1975, 1976; Taft 1979, 1981). The other criterion is grounded in processing time for the affix. The more frequent the affix, the less time is required to process the morphologically complex word on which it appears. Implicitly, complexity of an affixed word depends on properties of its affix such as its probability. Both, the decomposition approach and the full listing hypothesis assume that, other things being equal, the time to process an affixed word depends in part, on the affix probability. The outcome of experiments conducted with Serbian inflected nouns suggests that specifying complexity in terms of the probability of occurrence of an affix is misleading, if not wrong. When base morpheme is held constant, no systematic relationship between reaction times for inflected forms and their suffix probability has been observed and this outcome suggests that the cognitively relevant properties of inflected forms have yet to be identified. As noted earlier, an inflected noun form carries a number of syntactic functions and meanings only one of which can be realized in a sentence context. It can be assumed that an inflected form that carries a greater number of syntactic functions and meanings is linguistically more complex than one with fewer functions. If so, complexity of an inflected form indexes its grammatical complexity, quantitatively expressed in terms of its number of syntactic functions and meanings. Inspection of Appendix Β reveals great variability in the number of syntactic functions and meanings among inflected noun forms. In each of three grammatical genders, correlations between processing latency for inflected noun forms, and number of syntactic functions and meanings were not significant (Kostic, A. 2003, submitted). Evidently, the number of syntactic functions and meanings in

8

Kostic, Markovic and Baucal

itself is not an adequate descriptor of the cognitive complexity of inflected forms. An alternative account is that the probability of an inflected form and its number of syntactic functions/meanings is better treated as two descriptors whose effect on processing are inversely related. While the probability of a form is inversely proportional to processing latency, its number of syntactic functions/meanings is directly proportional to processing latency. The fact that neither of the two assumptions had been confirmed empirically could suggest that these two factors are in some mutually compensatory relationship. Accordingly, the observed processing latency for an inflected noun form could be a function of two inverse effects (i.e., as form probability and number of syntactic functions/meanings ratio [Equation 1])·

RTm=f

Eul

(1)

In Equation 1 RTm stands for the processing latency for an inflected form, Fm refers to the probability of an inflected form (i.e., inflectional suffix within a given paradigm, say, feminine nouns), while Rm stands for number of syntactic functions/meanings encompassed by that form. The equation states that the processing latency for an inflected noun form is some function of the average probability per syntactic function/meaning for a particular inflected form. This unit can be stated in term of the amount of information (viz., bits). When expressed as a proportion (relative to a sum of the average proportions per function/meaning for other noun forms for a particular gender 2 ) and logarithmically transformed, the average probability per syntactic function/meaning for a particular form is now specified in terms of bits of information (Equation 2). In Equation 2, Im stands for the amount of information carried by a particular inflected noun form, Fm refers to probability of a form, and Rn, stands for the number of syntactic functions/meanings encompassed by that form. Because the equation may appear to be rather complicated, it will be described more comprehensively.

Inflectional morphology and word meaning

/m =

Rmm

- log k

(2)

F

Σ R~~~ j=1 m,

9

/ -

The obtained unit is the average probability per syntactic function/meaning for that inflected form. The same procedure is applied to all inflected forms within a given paradigm (say, for the six morphologically distinct forms of feminine nouns). In order to express the obtained probability (F/R) for a given inflected form in terms of the amount of information (viz., bit), the F/R ratio for each inflected form must be expressed as proportion. This is done by summing up the F/R ratios for all inflected forms, each form's F/R being divided by the sum. The obtained proportions are then logarithmically transformed and multiplied by - 1 in order to get positive values. The result is the amount of information derived from the average probability per syntactic function/meaning for a particular inflected noun form. Note that this descriptor is directly proportional to the complexity of an inflected noun form: the higher the I value, the higher the complexity of a form. To emphasize, the amount of information is not derived from the form probability, but rather, from the average probability per syntactic function/meaning for a particular inflected noun form. In series of experiments with Serbian inflected nouns it was demonstrated that almost all of the processing variability can be accounted for by the informational values derived from Equation 2. When reaction times for all seven inflected forms of masculine nouns were regressed on the informational values derived from Equation 2, 88% of the variance was accounted for (p Kind, Waggon+s —> Waggon). This is not the case for irregularly inflected words, which are based on frozen (undecomposed) entries from which the unmarked stem/root is not directly available, yielding reduced priming.

5.

Visual lexical decision of derived stems

We have also examined -ung nominalizations and diminutives in an (unprimed) visual lexical decision task. In unprimed lexical decision, a subject's task is to discriminate between existing words (that have been encountered before) and nonce words (that have never been encountered before). This means that the task is sensitive to any trace of a word left in memory. Thus, given that unprimed lexical decision encourages subjects to rely on memory, this task is likely to tap stored full-form representations (see e.g. Pinker 1999: 138f.). The following predictions were tested. If -ung nominalizations and diminutives are derived stems with stored full-form representations, we

142

Clahsen, Sonnenstuhl and Blevins

would expect to find corresponding frequency effects in this task, i.e., reaction times for high frequency word forms should be shorter than for low frequency ones. In contrast, if the derivational forms tested do not have stored full-form representations, we should not find word-form frequency effects.

5.1. Method Materials. For this experiment, we selected 20 -ung nominalizations and 20 diminutive forms with -chen. In order to investigate word-form frequency effects, we arranged the items pairwise so that each member had a similar stem frequency but a different word-form frequency. This resulted in the four experimental conditions shown in Table 5; see appendix Β for a complete list of experimental items. Table 5. Example stimulus set Word Form Frequency (mean)

Stimulus Example

Condition

Stem Frequency (mean)

-chen high -chen low

176 176

14 0.8

Kätzchen 'small cat' Pflänzchen 'small plant'

-ung high -ung low

245 245

140 35

Gründung 'foundation' Fälschung 'falsification'

To prevent the participants from developing expectations during the experiment, 80 morphologically complex nouns (e.g. Lügner 'liar') and 120 monomorphemic nouns (e.g. Bluse 'blouse') were added to the 40 test items as word-fillers so that half of the presented words were morphologically complex and half of them were monomorphemic. This list of 240 words was further supplemented by the same amount of pseudo-words, constructed by changing two or three letters of existing words. The pseudonouns had the same morphological structure as the real nouns, i.e., 60 of them were pseudo-derivational forms with -ung (e.g. Dömterung) and 60 of them were pseudo-diminutives (e.g. Mörkcheri). In order to eliminate undesired priming effects within the experimental list, the 480 stimuli were presented in a pseudo-randomized order making sure that no semantic associations existed between consecutive items and that not more than four

Derivational morphology in the German mental lexicon

143

words or pseudo-words occurred in sequence. Participants. 45 students (12 male and 33 female, mean age 25) participated in the experiment. Procedure. Each trial consisted of the presentation of a fixation point in the middle of a 17-inch computer monitor, followed after 600 ms by the stimulus in the same position. The stimuli were presented in Arial 24 point with white letters on a dark background. They stayed visible on the screen for 350 ms. The measuring of the reaction times began with the presentation of the target. The participants reacted by pressing a green button (for a word) or a red button (for a pseudo-word) on a dual box. After an intertrial time of 1,000 ms the next trial was initiated. Participants read a written instruction with a detailed description of the task, and performed a short practice phase before the actual experiment started. Two further breaks were provided during the experiment. The overall duration of an experimental session was approximately 35 minutes. Analysis. Errors, i.e. nonword-responses to existing words and wordresponses to pseudo-words, were removed from the data set before further statistical analyses. The error rate was higher for low frequency derivational forms (5.1 %) than for high frequency forms (2.2 %). This difference was significant for diminutives (t (44) = 3.43, p = .001) and for -ung derivations (t (44) = 3.79, ρ < .001) and was consistent with the pattern of recognition times to be reported below. Extreme reaction times exceeding more than 2 SDs from a participant's mean reaction time in each condition were removed. These data (2.2 %) did not show any significant differences across conditions. The remaining data for each derivation type were entered in two separate ANOVAs for subjects and items with the factor 'Word-Form Frequency'.

5.2. Results The mean lexical decision shown in Figure 2 demonstrate that for both derivational forms high frequency items produced shorter lexical decision times than low frequency ones. Moreover, -ung derivations yielded overall shorter reaction times than diminutives, which is probably due to the fact that -ung forms are more frequent than diminutives; see Table 5.

144

Clahsen, Sonnenstuhl and Blevins

Statistically, there were significant effects of 'Word-Form Frequency' in both the subject and the item analyses for diminutives (Fl (1,44) = 2630,ρ < .001, F2 (1,19) = 8.91, ρ = .008) as well as for -ung derivations (Fl (1,44) = 27.23,ρ < .001, F2 (1,19) = 6.46,/? = .020). Pairwise comparisons using matched /-tests confirmed these effects (589 vs. 548 ms, t (44)= 5.13, ρ < .001; 555 vs. 525 ms, t (44)= 5.22, ρ < .001). These results suggest that stored full-form representations are available for -ung nominalizations and -chen diminutives. Figure 2. Mean lexical decision times to visual targets presented in isolation (subjects). 600-1

589

580 560-

548

555 • low frequency • high frequency

540 520 500 Η 480

Diminutives

6.

-ung Derivations

Discussion

In this section, we compare the previously reported results on derived forms with results from corresponding experiments on inflected forms. It will be shown that the pattern of experimental effects is compatible with the postulation of a fundamental linguistic difference between derivational and inflectional processes. Recall that according to this view, derivational morphology creates stems with stored representations whereas regular productive inflection generates inflected word forms that are not listed at any level (Anderson 1982, 1992; Perlmutter 1988; Scalise 1988; among others). Moreover, fully productive and transparent derived stems differ from irregularly inflected forms in that the former are based on rules that

Derivational morphology in the German mental lexicon

145

contain variables and are therefore analyzable into roots/stems and exponents, whereas irregulars have undecomposed representations.

6.1. Comparing experimental findings on German inflection and derivation To see the different experimental effects between derivational and inflectional processes, Table 6 presents a summary of the results that were obtained from (cross-modal) priming and (visual) lexical decision tasks. The first column shows the assumed representation for the test items under study. Brackets indicate which parts of an item represent stems or roots; stems may have internal structure, as for example [[stift]ung] 'donation', [[kindjchen] 'small child', and [[tasche]n] 'pockets'. The second column shows whether or not cross-modal priming produced a full priming effect for the forms under study. Note that in all priming experiments derived or inflected forms such as those shown in Table 6 were used as primes, whereas targets were forms containing the corresponding unmarked stems or roots, e.g. Waggons —> Waggon 'wagon'. The third column shows whether or not the forms under study produced full-form frequency effects in (unprimed) lexical decision. Such effects were obtained in cases in which an inflected or derived form with a relatively high word-form frequency produced shorter response times than an item with a relatively low word-form frequency. Note that the base frequencies, i.e. the frequencies of the unmarked stem or root, were controlled in all experiments. As is clear from Table 6, regularly inflected word forms such as -t participles and -s plurals produced full priming and no word-form frequency effects. For irregularly inflected forms, such as -n participles, -er plurals, and (irregular) -n plurals the opposite pattern of results was obtained, reduced priming and full-form frequency effects. These results correspond to the linguistic representations shown in Table 6 for these kinds of items. Irregular participles and noun plurals are stored, undecomposed stems, hence the full-form frequency effect. They can only indirectly access their corresponding unmarked base and therefore produce reduced priming effects. Regular participles and noun plurals, on the other hand, do not have stored word-form representations, and hence the lack of a full-form frequency effect. Instead, they are decomposable into an

146

Clahsen, Sonnenstuhl and Β levins

unmarked stem/root plus the regular affix, and given this representation they can fully prime their corresponding base stem or root. Table 6. A summary of experimental findings on German inflection and derivation Full priming effect?

Full-form frequency effect?

-t participles: ge[kaufj-t

yes

no

Sonnenstuhl et al. (1999), Clahsen et al. (1997)

-s plurals: [waggon]-s

yes

no

Sonnenstuhl et al. (1999), Sonnenstuhl & Huth (2002), Clahsen et al. (1997)

-er plurals: [kinder]

no

yes

Sonnenstuhl et al. (1999), Sonnenstuhl & Huth (2002), Clahsen et al. (1997)

-n participles: [gelogen]

no

yes

Sonnenstuhl et al. (1999), Clahsen et al. (1997)

-η plurals II: [bauern]

no

yes

Sonnenstuhl & Huth (2002)

-ung nominalizations: [[stift]ung]

yes

yes

this study

diminutives: [[kind] chen]

yes

yes

this study

-n plurals I: [[tasche]n]

yes

yes

Sonnenstuhl & Huth (2002)

Representation

Source

The derivational forms tested in the present study yielded a pattern of experimental effects that differs from both the one for regular inflection and the one for irregulars. However, like in the case of the inflected words, the experimental effects obtained for derived forms correspond to their linguistic representations, illustrated in Table 6. The derived forms share with the irregulars the fact that their full forms represent stems, and hence the full-form frequency effects in the lexical decision task. Moreover, their internal structure is parallel to regularly inflected -t participles and -s plural forms in that -ung nominalizations, -chen and -lein diminutives are built forms, as for example in kind+chen. We assume that the recognition of such forms involves the activation of the underived stem form (e.g. kind),

Derivational morphology in the German mental lexicon

147

in addition to the associated derivational suffix. Thus the full priming effect obtained for such forms can be explained based on the shared stem entry of prime and target.

6.2. German -n plurals Consider, in the light of the account given for derived forms, the experimental results from Sonnenstuhl and Huth (2002) on -n plurals in German. Note that -n plurals do not represent a homogeneous class in German. There are (at least) two types, with different degrees of productivity and predictability. While plural formation with -n is fully predictable for feminine nouns with a stem-final Schwa, e.g. (die) Tasche - Taschen 'the pocket - pockets', plural formation with -n is not predictable for nonfeminine nouns without a stem-final Schwa, e.g. der Bauer -Bauern 'the farmer - farmers'. In Table 6, the latter are labelled class II -n plurals and the former class I -n plurals. Given the fact that there is no single exception to -n plurals of class I, i.e., all feminine nouns with a stem-final Schwa take the -n plural, we might hypothesize that class I -n plurals are rule-based, i.e., constitute built forms, whereas all other -n plurals are irregular, i.e., stored forms. On the other hand, however, all kinds of -n plurals (irrespective of their gender or stem properties) are in principle available for further word-formation processes yielding forms such as taschen-los 'pocket-less', or Bauernschaft 'farming community'. Thus, class I -n plurals such as Tasche - Taschen differ from regular -s plurals in that -n plurals may feed further lexeme formation, whereas -s plurals never do; compare taschen-los with *autos-los 'car-less'. These properties are reflected in the representations posited for the two types of -n plurals in Table 6, according to which all -n plurals are stems and in which -n plurals of class I (but not of class II) are built forms yielding internally structured root+affix representations. The experimental results on class II -n plurals obtained by Sonnenstuhl and Huth (2002) correspond to their linguistic representation. Class I -n plurals form stems, hence the full-form frequency effect in the lexical decision task. On the other hand, these -n plurals are rule-based and decomposable, hence the full priming effect found in the cross-modal priming task.

148

Clahsen, Sonnenstuhl and Β levins

6.3. Alternative accounts Finally, we will briefly discuss how previous psycholinguistic accounts of the similarities and differences between inflection and derivation might explain the results summarized in Table 6. Consider first Raveh and Rueckl's (2000) proposal that the processing of derivation and inflection is not fundamentally different and that factors other than their linguistic representation (e.g. statistical variables, orthographic and visual properties of the items involved) should determine how they are processed. This account is not supported by our findings. If, for example, orthographic similarity determined morphological priming, one would expect that the plural form kinder 'children' primes kind more efficiently than the diminutive kindchen. Yet, we found the opposite pattern, full priming for diminutives and reduced priming for (irregular) plurals. Another example comes from -s versus -er plurals, which are statistically similar, in that both plural forms with -J and with -er have low (type) frequencies in the German language; see Sonnenstuhl (2003: 83) for relevant frequency counts. Yet, in our experiments they behaved very differently; see Sonnenstuhl, Eisenbeiss and Clahsen (1999) for further discussion. Another suggestion comes from Laudanna et al. (1992) who argued for a stem representation hypothesis, according to which (regularly) inflected words have lexical representations that are decomposed into stems plus inflectional affixes, whereas derivational forms are not decomposed into roots plus derivational affixes. Our finding that both productive derivation and regular inflection (but not irregular inflection) produce full priming effects in the cross-modal priming task, whereas irregular inflection does not, challenges this hypothesis. A third proposal comes from several proponents of the Dual Mechanism Model (Alegre and Gordon 1999, Pinker 1999, Hagiwara et al. 1999, Sugioka et al. 2001, Ullman 2001) who argued for a general opposition between combinatorial operations and lexical lookup which should apply to both inflectional and derivational morphology. It is predicted here that productively defined inflected and derived forms should produce the same experimental effects which should uniformly contrast with irregular formations, since they are both based on combinatorial operations. Our findings provide only partial support for this view. Table 6 shows that the priming results are indeed parallel for the derivational and the regularly inflected forms we tested, and that they contrast with the results on irregular inflection in that only the former yielded full priming effects. These results

Derivational morphology in the German mental lexicon

149

are compatible with the view that both (productive) derivation and (regular) inflection involve combinatorial rules. In unprimed lexical decision, however, we found that the same derivational forms produce fullform frequency effects and hence pattern with irregular forms and unlike regularly inflected forms. The idea of a general opposition between combinatorial operations and stored forms does not explain why (productive) derivational forms pattern differently from both regular and irregular inflection.

7.

Conclusion

Investigating fully transparent and productive derivational forms with the suffixes -ung, -lein, and -chert, we found that in cross-modal priming tasks these forms were effective primes of corresponding underived forms. We argue that these effects correspond to the lexical representations of these items, i.e., derived forms with -ung, -lein, and -chen are built (rule-based) forms from which the base stems are directly available for priming. In unprimed lexical decision, on the other hand, we found that diminutives and -ung nominalizations produced full-form frequency effects. We argue that these effects reflect the fact that derived forms are stems with stored full-form representations that are picked up by the lexical decision task. To account for these findings, we suggest a refinement of the Dual Mechanism Model that treats productive inflection and derivation both as the result of combinatorial operations but associates productive derivation (like irregularly inflected items) with stored entries. This interpretation of our experimental results is compatible with models of morphology that distinguish inflection from derivation and particularly with realizationbased models of morphology that express this split in terms of a contrast between entry-defining derivation and form-defining inflection. We conclude that the refined Dual Mechanism Model provides the best account for our experimental findings; alternative accounts of the derivation/ inflection contrast were found to be less successful.

Acknowledgements The research reported in this paper is supported by a German Research Council grant (SFB 282/C7) to HC. We thank Meike Hadler, Rebecca Groß, Peter Priifert,

150

Clahsen, Sonnenstuhl andBlevins

and Axel Huth for assistance in administering the experiments. We are also grateful to the ESRC Research Seminar series on 'Models and Methods in Morphology' that stimulated this collaboration. Acknowledgements for helpful comments are due to Harald Baayen, Iris Berent, Geert Booij, Andrew Carstairs-McCarty, Grev Corbett, Wolfgang Dressler, Sonja Eisenbeiss, Jen Hay, Andrew Spencer, Greg Stump, Michael Ullman, Richard Wiese and Dieter Wunderlich.

Appendices Appendix A.

Cross-modal priming: -ung nominalizations

The following three lists present the prime-target pairs used in the test condition of this experiment. CELEX (Baayen et al. 1993) word frequencies of the items used as visual targets in the experiment are shown in brackets. List 1: mean target frequency 299 Lackierung 'varnish' - lackieren 'to varnish' (0); Heizung 'heating' - heizen 'to heat' (34); Dichtung 'poetry' - dichten 'to write poetry' (53); Bremsung 'braking' bremsen 'to brake' (83); Lenkung 'steering system' - lenken 'to steer' (212); Sammlung 'collection' - sammeln 'to collect' (291); Verfolgung 'pursuit' verfolgen 'to pursue' (469); Entdeckung 'discovery' - entdecken 'to discover' (472); Hebung 'lifting' - heben 'to lift' (685); Teilung 'division' - teilen 'to divide' (687) List 2:

mean target frequency 298

Fahndung 'search' - fahnden 'to search for' (13); Pflanzung 'planting' - pflanzen 'to plant' (40); Bohrung 'drilling' ? bohren 'to drill' (48); Schaltung 'circuitry'? schalten 'to switch' (66); Verwaltung 'administration' ? verwalten 'to administer' (96); Gründung 'foundation' ? gründen 'to found' (370); Rettung 'rescue' ? retten 'to rescue' (427); Prüfung 'examination' ? prüfen 'to examine' (460); Öffnung 'opening' ? öffiien 'to open' (617); Erzählung 'telling' ? erzählen 'to tell' (846) List 3:

mean target frequency 298

Lüftung 'airing' - lüften 'to air' (22); Siedlung 'settlement'- siedeln 'to settle' (28); Kassierung 'collection' - kassieren 'to collect' (48); Ernährung 'nutrition' ernähren 'to nourish' (64); Täuschung 'deception' - täuschen 'to decieve' (101); Störung 'disturbance'- stören 'to disturb' (320); Zeichnung 'painting' - zeichnen 'to paint' (336); Nutzung 'utilization' - nutzen 'to use' (442); Zahlung 'payment' zahlen 'to pay' (504); Lesung 'reading' - lesen 'to read' (1113)

Derivational morphology in the German mental lexicon

Appendix Β

151

Lexical decision: Diminutives and -ung nominalizations

The following lists present the items used in the experimental conditions. CELEX frequencies (Baayen et al. 1993) are shown in brackets, noun or verb stem frequencies on the left and word-form frequencies on the right. Low frequency diminutives (mean noun stem frequency: 176, mean word-form frequency: 0,8) Nüßchen 'little nut' (24 -0); Mäulchen 'little mouth' (39-1); Kälbchen 'little calf (43-1); Schwänchen 'little swan' (48-0); Äpfelchen 'little apple' (74-0); Zwiebelchen 'little onion' (77-0); Täfelchen 'small board' (83-0); Kränzchen 'little garland' (97-2); Schränkchen 'small cupboard' (131-0); Heftchen 'booklet' (140-0); Hüttchen 'little hut' (165-0); Tänzchen 'little dance' (167-5); Treppchen 'small stairs' (168-1); Späßchen 'little joke' (174-1); Stämmchen 'small trunk' (209-0); Täschchen 'small bag' (216-1); Kärtchen 'small card' (283-2); Öhrchen 'little ear' (292-0); Höfchen 'little yard' (474-0); Lichtchen 'little light' (621-1) High frequency diminutives (mean noun stem frequency: 176, mean word-form frequency: 14) Kästchen 'little box* (9-8), Glöckchen 'little bell' (40-10); Hähnchen 'small rooster' (41-17); Tütchen 'small paper bag' (43-32); Würstchen 'small sausage' (63-20); Hühnchen 'little chicken' (81-12); Schäfchen 'lamb' (83-10); Bändchen 'little ribbon' (96-8); KöfFerchen 'small suitcase' (123-14); Stöckchen 'little Stick' (123-15); Körbchen 'little basket' (129-14); Mäntelchen 'small coat' (157-9); Pärchen 'couple' (158-15); Fähnchen 'little flag' (166-10); Weilchen 'little while' (179-19); Fläschchen 'small bottle' (216-23); Liedchen 'little song' (289-13); Sternchen 'little star' (422-15); Brüderchen 'little brother' (528-10); Töchterchen 'little daughter' (576-15) Low frequency -ung derivations (mean verb stem frequency: 245, mean word-form frequency: 35) Peinigung 'torture' (10-0); Wucherung 'proliferation' (17-0); Programmierung 'programming' (25-6); Fälschung 'forging' (25-6); Schulung 'training' (28-49); Schlichtung 'arbitration' (33-14); Bohrung 'drilling' (48-6); Dichtung 'poetry' (53-204); Bewerbung 'application' (54-98); Erbauung 'construction' (64-3); Schaltung 'circuitry' (66-12); Schiebung 'pushing' (109-1); Unternehmung 'enterprise' (131-9); Wanderung 'hike' (164-28); Lesung 'reading' (415-46); Vertretung 'representation' (581-184); Kennung 'identification' (650-0); Nennung 'naming' (716-12); Gewinnung 'production' (752-26); Schreibung 'spelling' (964-1)

152

Clahsen, Sonnenstuhl and Β levins

High frequency -ung derivations (mean verb stem frequency: 245, mean word-form frequency: 140) Züchtung 'breeding' (25-10); Siedlung 'settlement' (28-88); Schöpfung 'creation' (30-93); Forschung 'research' (44-523); Streichung 'cancellation' (54-16); Zündung 'ignition' (64-15); Sendung 'transmission' (85-278); Leugnung 'denying' (100-3); Mischung 'mixture' (113-75); Bewunderung 'admiration' (136-50); Überwindung 'effort' (192-125); Lenkung 'steering' (212-56); Begleitung 'company' (237-76); Vermietung 'renting' (266-30); Sammlung 'collection' (291-106); Gründung 'foundation' (370-274); Prüfung 'examination' (460-254); Beratung 'discussion' (503-396); Öffiiung 'opening' (617-39); Rechnung 'calculation' (1068-294)

Notes 1.

2.

Although we dispute the existence of 'inherently' inflected items, we fully accept Booij's central claim that the patterns in question do not in any event support the distinction between pre-syntactic derivation and post-?syntactic inflection expressed by the Split Morphology Hypothesis of Perlmutter (1988). The schematic rule in (3 b) abstracts away from inessential issues involving the choice of verb roots or the semantic effect of nominalization. Since this is normally regarded as a lexeme-?creating operation, the output has a new lexeme index, which is represented informally here by concatenating ING to the lexeme index of the input.

References Alegre, Maria, and Peter Gordon 1999 Rule-based versus associative processes in derivational morphology. Brain and Language 68: 347-354. Anderson, Stephen R. 1982 Where's morphology? Linguistic Inquiry 13: 571-612. Anderson, Stephen R. 1992 A-Morphous morphology. Cambridge: Cambridge University Press. Aronoff, Mark, and S. N. Sridhar 1988 Prefixation in Kannada. In Theoretical morphology, Michael Hammond, and Michael Noonan (eds.), 179-191. San Diego: Academic Press. Aronoff, Mark 1994 Morphology by itself. Cambridge, MA: MIT Press.

Derivational morphology in the German mental lexicon

153

Baayen, R. Harald, Richard Piepenbrock, and Hedderik van Rijn 1993 The CELEX lexical database (CD ROM). Philadelphia, PE: Linguistic Data Consortium, University of Pennsylvania. Baumann, Herbert, John Nagengast, and Gerd Klaas 1993 New experimental setup (NESU). Ms. Nijmegen: Max-PlanckInstitute for Psycholinguistics. Bertram, Raymond, R. Harald Baayen, and Robert Schreuder 1999 Effects of family size for complex words. Journal of Memory and Language 42: 390-405. Bertram, Raymond, Robert Schreuder, and R. Harald Baayen 2000 The balance of storage and computation in morphological processing: The role of word formation type, affixal homonymy, and productivity. Journal of Experimental Psychology: Memory, Learning, and Cognition 26: 419-511. Blevins, James P. 2001a Paradigmatic derivation. Transactions of the Philological Society 99: 211-222. Blevins, James P. 2001b Stems and paradigms. Ms. Cambridge: University of Cambridge, Research Center for English and Applied Linguistics. Booij, Geert 1994 Against split morphology. In Yearbook of morphology 1993, Geert Booij, and Jaap van Marie (eds.), 27-49. Dordrecht: Kluwer. Booij, Geert 1996 Inherent versus contextual inflection and the split morphology hypothesis. In Yearbook of morphology 1995, Geert Booij, and Jaap van Marie (eds.), 1-16. Dordrecht: Kluwer. Booij, Geert 2002 The morphology of Dutch. Oxford: Oxford University Press. Burani, Cristina, and Alessandro Laudanna 1992 Units of representation for derived words in the lexicon. In Orthography, phonology, morphology, and meaning, Ram Frost and Leonard Katz (eds.), 361-376. Amsterdam: North-Holland. Clahsen, Harald 1999 Lexical entries and rules of language: A multi-disciplinary study of German inflection. Behavioral and Brain Sciences 22: 991-1060. Clahsen, Harald, Sonja Eisenbeiss, and Ingrid Sonnenstuhl 1997 Morphological structure and the processing of inflected words. Theoretical Linguistics 23: 201-249. Fabb, Nigel 1988 English suffixation is constrained only by selectional restrictions. Natural Language and Linguistic Theory 6: 527-539.

154

Clahsen, Sonnenstuhl and Blevins

Feldman, Laurie Β. 1994 Beyond orthography and phonology: Differences between inflections and derivations. Journal of Memory and Language 33: 442-470. Fowler, Carol Α., Shirley E. Napps, and Laurie B. Feldman 1985 Relations among regular and irregular morphologically related words in the lexicon as revealed by repetition priming. Memory and Cognition 13: 241-255. Friederici, Angela D., Herbert J. Schriefers, and Patty A. M. Graetz 1989 Abruf und Repräsentation morphologisch komplexer Wörter verschiedener Wortklassen. In Experimentelle Studien zur deutschen Flexionsmorphologie, Hartmut Günther (ed.), 41-70. Hamburg: Buske. Garrett, Merrill F. 1980 Levels of processing in sentence production. In Language production, Brian Butterworth (ed.), Vol. 1, 177-220. London: Academic Press. Hagiwara, Hiroko, Yoko Sugioka, Takane Ito, Mitsuru Kawamura, and Jun-Ichi Shiota 1999 Neurolinguistic evidence for rule-based nominal suffixation. Language 75: 739-763. Janssen, Dirk 1999 Producing past and plural inflections. Ph. D. diss., Max-PlanckInstitute for Psycholinguistics, Nijmegen. Kiparsky, Paul 1982 From cyclic phonology to lexical phonology. In The structure of phonological representations, Harry van der Hulst, and Norval Smith (eds.), Part 1, 131-175. Dordrecht: Foris. Laudanna, Alessandro, Alfonso Caramazza, and William Badecker 1992 Processing inflectional and derivational morphology. Journal of Memory and Language 31: 333-348. Marslen-Wilson, William, Lorraine K. Tyler, Rachelle Waksler, and Lianne Older 1994 Morphology and meaning in the English mental lexicon. Psychological Review 101: 3-33. Matthews, Peter H. 1991 Morphology. 2nd ed. Cambridge: Cambridge University Press. McQueen, James M., and Anne Cutler 1998 Morphology in word recognition. In The handbook of morphology, Andrew Spencer, and Arnold M. Zwicky (eds.), 406-427. Oxford: Blackwell. Palmer, Frank R. 1974 The English verb. London: Longman.

Derivational morphology in the German mental lexicon

155

Perlmutter, David M. 1988 The split morphology hypothesis. In Theoretical morphology, Michael Hammond, and Michael Noonan (eds.), 79-100. San Diego, CA: Academic Press. Pinker, Steven 1999 Words and rules: The ingredients of language. New York: Basic Books. Raveh, Michal, and Jay G. Rueckl 2000 Equivalent effects of inflected and derived primes: Long-term morphological priming in fragment completion and lexical decision. Journal of Memory and Language 42: 103-119. Scalise, Sergio 1988 Inflection and derivation. Linguistics 26: 561-581. Sonnenstuhl, Ingrid, Sonja Eisenbeiss, and Harald Clahsen 1999 Morphological priming in the German mental lexicon. Cognition 72: 203-236. Sonnenstuhl, Ingrid, and Axel Huth 2002 Processing and representation of German -n plurals: A dual mechanism approach. Brain and Language 81: 276-290. Sonnenstuhl, Ingrid 2003 Deutsche Plurale im mentalen Lexikon: Experimentelle Untersuchungen zum Verhältnis von Speicherung und Dekomposition. Tübingen: Niemeyer. Spencer, Andrew 1991 Morphological theory: an introduction to word structure in generative grammar. Oxford: Blackwell. Stanners, Robert F., James J. Neiser, William P. Hernon, and Roger Hall 1979 Memory representation for morphologically related words. Journal of Verbal Learning and Verbal Behavior 18: 399-412. Stump, Gregory T. 2001 Inflectional morphology: a theory of paradigm structure. Cambridge: Cambridge University Press. Sugioka, Yoko, Takane Ito, and Hiroko Hagiwara 2001 Computation vs. memory in Japanese causative formation: Evidence from agrammatic aphasics. Cognitive Studies 8: 37-62. Ullman, Michael 2001 A neurocognitive perspective on language: the declarative/ procedural model. Nature Reviews Neuroscience 2: 717-727. Wiese, Richard 1995 The phonology of German. Oxford: Oxford University Press.

The interplay of root, suffix and whole-word frequency in processing derived words* Cristina Burani and Anna M. Thornton In three lexical decision experiments we investigated whether the relative frequency of root, derivational suffix and whole-word affects processing of Italian printed derived stimuli. Experiment 1 considered pseudowords made up of pseudoroots combined with either high-,medium-, or low-frequency suffixes. Only pseudowords with high-frequency suffixes resulted in increased decision times and higher error rates relative to nonsuffixed pseudowords. Experiments 2 and 3 dealt with suffixed derived words. In Experiment 2, low-frequency words with high-or low-frequency roots and with high or low-frequency suffixes were orthogonally contrasted. Lexical decision latencies were a function of the frequency of both the root and the suffix. However, post-hoc comparisons showed an effect of wholeword familiarity. In Experiment 3, low-frequency derived words with orthogonal variation of root and suffix frequency, and equal whole-word familiarity, were investigated, and were contrasted with low-frequency nonderived words. Words with high-frequency roots showed quicker and more accurate lexical decision responses, irrespective of suffix frequency. By contrast, words with low-frequency roots, irrespectively of suffix frequency, did not differ from nonderived words. These results are interpreted within Schreuder and Baayen's (1995) parallel dualroute model for morphological processing, as evidence for both benefits and costs of morphemic access, due to the balancing of the quantitative characteristics of root, suffix and whole word.

1. Introduction In most models of morphological processing, it is assumed that the probability of accessing morphological constituents of words is conditioned, at the different processing stages, by many properties of the morphologically complex words. Among these properties, the frequency of morphological constituents relative to the frequency of the complex word as a whole form can play a major role. Evidence for reliance on morphological structure in accessing printed complex words comes from low-frequency words which include higher frequency constituents (see, e.g., Andrews 1986; Burani and Caramazza 1987; Meunier and Segui 1999). In parallel dual-route models of lexical

15 8

Burani and Thornton

access (see, e.g., Burani and Laudanna 1992; Chialant and Caramazza 1995; Frauenfelder and Schreuder 1992; Schreuder and Baayen 1995), words composed of more than one morpheme may activate in parallel two types of access units, namely units corresponding to the whole word and units corresponding to the morphemes included in the stimulus. In these models, the relative frequency of the whole word and of the constituent morphemes affect the relative time-course of activation of the different units in the different components. Hence, frequency is the major determinant of the relative probability that lexical access is either whole-wordbased or morpheme-based. The assumption underlying these models is that the higher the frequency of a given lexical unit, be it a word, a root or an affix, the greater the likelihood that this unit is quickly activated and processed in the different processing components. What is crucial, in determining the probability that lexical access is provided by either whole-word or morpheme processing, is the complex balance existing between the frequency of the whole word and the frequency of its constituent morphemes, both roots and affixes, i.e., it is relative frequency, rather than absolute frequency (for a similar proposal, see Hay 2000; 2001). Hence, it might be predicted that a transparent derived word which has low-frequency in the language, like Italian bassezza ('lowness'), but is composed of a very frequent root (i.e., bass-, 'low') and a very frequent suffix (i.e., -ezza, '-ness') is likely to be accessed via activation of its morphemic constituents, rather than via the unit corresponding to the wholeword, which is supposed to be very scarcely activated. This prediction implies that the frequencies of both the root and the suffix are capable of affecting processing, and calls for evidence concerning the roles of both root and suffix frequency. However, and surprisingly, only the frequency of roots has been considered so far, while the frequency of affixes has been usually neglected. No study on lexical access to derived words has systematically varied the quantitative values of derivational affixes. By contrast, these values have been investigated in the context of pseudoword processing (see below).

1.1. Studies on words Several studies conducted in different languages, including English, Italian, French and Dutch, have shown that access times and accuracy to suffixed

The interplay of root, suffix and whole-wordfrequency

159

derived words are significantly affected by root frequency (see, e.g., Beauvillain 1996; Bradley 1979; Burani and Caramazza 1987; Cole, Beauvillain and Segui 1989; Holmes and O'Regan 1992; Schreuder, Burani, and Baayen 2002). Lexical decisions were faster and more accurate when a suffixed word, usually of low frequency, included a root of high frequency. The facilitatory effect of high-frequency root morphemes was found both when calculation of root frequency included the frequency of the base word and its inflected forms only (Burani and Caramazza 1987), and when it was extended to include the frequencies of all the derived word-forms sharing the same root (Cole, Beauvillain and Segui 1989). The root frequency effect has been found in the context of suffixed derived words that were both orthographically and phonologically transparent with respect to their base root, and were usually transparent for meaning with respect to the meaning of their base. However, there has been evidence of root frequency effects also for derived words that included bound roots, or could be rated as semantically opaque with respect to their base (see, e.g., Holmes and O'Regan 1992; Schreuder, Burani and Baayen 2002). In the studies in which root frequency effects have been found, suffixes were usually productive and had high frequency. Suffix frequency was not directly investigated per se, but it was usually kept constant across categories by including the same suffixes in the high-frequency and lowfrequency root sets.

1.2. Studies on pseudowords An investigation of suffix frequency per se was recently made, by adopting pseudoword contexts made up of illegal root-suffix combinations. Burani et al. (1997) submitted, to both lexical decision and naming, pseudowords that were made up of real roots combined with derivational suffixes not compatible with the root. In order to demonstrate that the probability of access through activation of morphemic units corresponding to suffixes is constrained by their frequency values (see also Laudanna and Burani 1995), Burani et al. (1997) made use of suffixes belonging to two distinct frequency ranges. In one experimental set, roots were combined with highfrequency suffixes, and the resulting pseudowords were contrasted with pseudowords in which the same roots were combined with control sequences that had analogous orthographic frequency in final position of Italian words, but were not suffixes. In the second set, a comparison was

160

Burani and Thornton

made between pseudowords composed of roots plus low-frequency suffixes, and the same roots combined with control low-frequency orthographic final sequences. Roots were in both cases of medium frequency. In order to control for asymmetries in the possibility to assign meaning to suffixed pseudowords of the two kinds (i.e., with high- and low-frequency suffixes, respectively), suffixed pseudowords in the two frequency sets were matched for mean interpretability values derived from participants' empirical ratings. The lexical decision results by Burani et al. (1997) showed that the interference effect which is usually found on pseudowords that include real affixes (see, e.g., Caramazza, Laudanna and Romani 1988; Taft and Forster 1975; see also, for derivational suffixes, Jarvella and Wennstedt 1993) is conditioned by the frequency of the embedded suffixes: Longer reaction times and higher error rates were found, with respect to control pseudowords, only when pseudowords included high-frequency suffixes. By contrast, pseudowords with low-frequency suffixes took no longer to be rejected than control pseudowords. From these results on suffixed pseudowords, Burani et al. (1997) concluded that the probability that suffixes will affect processing is conditioned by their frequency (see also Laine 1996, for Finnish productive derivational suffixes causing interference effects on pseudoword lexical decision).

1.3. Suffix frequency, suffix numerosity, and productivity In considering the frequency of suffixes, two main quantitative measures can be adopted. On the one hand, frequency in the proper sense is calculated on word tokens, by summing up the cumulative frequency in a given corpus of all the word tokens in which a given suffix occurs. On the other hand, suffix frequency can be measured by calculating the number of word types in which a given suffix occurs in a given language. This second measure could be named numerosity of the suffix (as proposed by Burani et al. 1995). There could be reasons for considering numerosity (i.e., suffix typefrequency) as a better quantitative characterization for suffixes and a stronger predictor of performance in access tasks. Suffix numerosity is closely related to suffix productivity thus allowing the suffix to "emerge" as a separate processing unit (see, e.g., Baayen 1989; 1992; Bybee 1995a). However, there is a strong link and a complex interplay among suffix type

The interplay of root, suffix and whole-wordfrequency

161

and token frequency, productivity and probability of morphemic parsing (Hay and Baayen 2002). In the study by Burani et al. (1997), suffix numerosity and suffix frequency were not disentangled because, after inspection of frequency distributions, it was found that suffix token-frequency and suffix numerosity tended to be highly correlated. Suffix frequency was used in a broad sense to subsume the two quantitative measures that could affect processing. Consequently, suffixes were either high or low on both dimensions, frequency and numerosity, calculated on a corpus of Italian written language (Istituto di Linguistica Computazionale CNR 1989).

1.4. Other properties of suffixes relevant for word processing Some recent research has investigated the role of properties of derivational suffixes in lexical access to words (see Bertram, Laine and Karvinen 1999, for Finnish; Bertram, Schreuder and Baayen 2000, for Dutch). For both Finnish and Dutch, properties like suffix productivity and suffix homonymy (i.e., suffix ambiguity in serving more than one semantic function) were found to affect processing, with words including productive and nonhomonymous suffixes being more likely to induce morpheme-based processing (see also, for Japanese, Hagiwara et al. 1999). In the studies by Bertram, Laine and Karvinen (1999) and Bertram, Schreuder and Baayen (2000), no information was given on the frequency values of the productive vs. unproductive suffixes, and the issue of suffix frequency/numerosity was not assessed directly. A variation in productivity usually corresponds to a variation in frequency/numerosity. However, although very related to suffix productivity, suffix frequency and numerosity do not necessarily correspond to productivity. There may be differences in suffix numerosity that do not correspond to differences in suffix productivity. At the same time, it is not always the case that differences in productivity correspond to differences in suffix frequency or numerosity.2 Thus there are reasons for investigating suffix frequency/numerosity in derived words, without identifying these quantitative measures with productivity.

162

Burani and Thornton

2. The present study While there is evidence that root frequency affects access to printed suffixed derived words, evidence for a role of quantitative properties of suffixes in visual processing comes almost exclusively from pseudowords. The present study aimed at assessing simultaneously the roles of both root and suffix frequency in Italian derived words, by testing different combinations of roots and suffixes with differing frequency. If the role of quantitative properties of morphemes have to be assessed per se, derived words with morphemic constituents of different frequencies should be matched for a number of factors, including orthographic/phonological transparency, semantic transparency, and whole-word frequency. Italian derivation occurs mostly through agglutination of suffixes to roots which are not occurring words themselves (see Peperkamp 1995). Orthographic/phonological transparency of derived forms with respect to their roots is quite common, wide-spreading across different frequency ranges of both words and morphemes, and can be easily controlled for. Semantic transparency can also be kept under control, while varying suffix frequency/numerosity. Suffix numerosity, i.e., the number of word types in which a given affix occurs, is one determinant of semantic transparency, but it does not identify with it (see Bybee 1995a). Moreover, in intramodal tasks there might be reasons for expecting effects of morphological constituency also in derived words that are less transparent for meaning or in semantically opaque words (see, e.g., Bentin and Feldman 1990; Feldman and Soltano 1999; Plaut and Gonnerman 2000; Schreuder, Burani and Baayen 2002; Stolz and Feldman 1995; Vannest and Boland 1999; but see also, for contrasting evidence in cross-modal tasks, Marslen-Wilson et al. 1994). Given the basic prediction that low-frequency words with two highfrequency constituents should be the best candidates for access through morphemes, which predictions could be made for low-frequency words that include only one high-frequency constituent (either the root or the affix)? Would lexical access be equally sensitive to the higher frequency constituent? Would it be differentially sensitive to the frequency of the root and the affix, respectively? In some studies the assumption has been made that root frequency effects should manifest themselves in low-frequency derived words with frequent and productive suffixes. To our knowledge, there were no investigations of whether root frequency effects would show up in the context of low-frequency suffixes. Similarly, no study has inves-

The interplay of root, suffix and whole-wordfrequency

163

tigated whether high-frequency suffixes may affect the probability of morpheme-based access in the context of low-frequency roots. In synthesis, no study has addressed the issue of whether low-frequency derived words that include either one or both low-frequency constituents were likely to activate morphemic units at all. Our predictions were developed in the framework of the model proposed first by Schreuder and Baayen (1995) (see, for recent updates, Baayen and Schreuder 1999; 2000; Baayen, Schreuder and Sproat 2000). This is a race model for the recognition of morphologically complex words in which there are two parallel access routes, one based on whole-form information, and the other based on morphemic decomposition. One assumption of the model is that, for the visual modality, the complete input is available from the start. Thus in principle, for low-frequency derived words, both root and suffix frequency should affect processing at the stages in which morphemic access representations are activated over time by the sensory input. In the framework of this race model, it is crucial to assess the complex balance between access through storage (whole-word activation) and access through computation (morphemic activation). An open issue is how processing proceeds for low-frequency words which include low-frequency constituent morphemes. This implies assessing the relation existing, in terms of processing costs and benefits, between components in which morphemic units are segmented and activated, and subsequent components in which they are re-combined in order to derive meaning. So far, the probability of faster access through whole-word activation has been suggested for high-frequency derived words that tend to be highly lexicalized (Baayen and Neijt 1997; Burani and Laudanna 1992; Bybee 1995b; Chialant and Caramazza 1995; Frauenfelder and Schreuder 1992; Schreuder and Baayen 1995).3 Additionally, whole-word storage has been proposed for English derived words which include non-neutral affixes and whose stems tend to cluster around recurring patterns thus constituting "gangs" (Alegre and Gordon 1999b), or for derived words which include unproductive suffixes (Bertram, Laine and Karvinen 1999; Bertram, Schreuder and Baayen 2000; Hagiwara et al. 1999). However, the possibility should be conceived that also low-frequency derived words which include low-frequency morphemic constituents even if phonologically transparent - are more likely to be accessed as whole forms through direct whole-word access. For these words, the reduced probability of access through morphemic decomposition would de-

164

Burani and Thornton

rive from the fact that the slight difference between the frequency of morphemes and whole-word frequency is not large enough for morphological processing to result in benefits relative to whole-word based lexical access. The following visual lexical decision experiments addressed the latter issues by combining evidence from pseudoword and word processing. Experiment 1 was conducted on pseudowords, whereas both Experiments 2 and 3 involved words. Experiment 1 on pseudowords aimed at replicating and extending results on the role of suffix quantitative properties in nonlexical contexts. It addressed issues that should help in interpreting results from the two following experiments on words. By including suffixes in pseudoword contexts in which the initial orthographic sequence did not correspond to an existing root, the role of suffix frequency/numerosity could be differentiated from its consequences on the semantic transparency or interpretability of a newly derived form with respect to its base. At the same time, by investigating stimulus contexts in which the lexical morphemic unit (the suffix) occurred in the rightmost part of the stimulus in the absence of a morphemic unit on its left side, we aimed at providing evidence for a role of morphemic units which is independent of sequential left-to-right processing. Experiments 2 and 3 addressed the issue of the interplay of root, suffix and whole-word frequency in access to derived words, by orthogonally varying high- and low-frequency roots and suffixes in low-frequency transparent derived Italian words. In order to specify the balance between the processing routes based on whole-word and morphemic units, respectively, the derived words were contrasted with nonderived words of analogously low frequency (Experiment 3).

3. Experiment 1 In Experiment 1, we aimed at replicating, with three sets of suffixes, the effect of suffix frequency in lexical decision to pseudowords found by Burani et al. (1997). In that study, suffixes were combined with real roots to form pseudowords. In the present experiment, the pseudowords were obtained by combining suffixes of varying frequencies with orthographic sequences that did not correspond to real roots. If suffix frequency effects were to occur in lexical decision to pseudowords of this sort, strong evidence for the role of suffixes in visual processing would be provided. If activation of the morphemic units comprising a stimulus occurs irrespective of their sequential positions within the stimulus, provided they are

The interplay of root, suffix and whole-wordfrequency

165

frequent enough, the prediction could be made that high-frequency suffixes are activated and play some role also when affixed after nonroots. The expected result has twofold implications. On the one hand, if highfrequency suffixes delay lexical decisions to pseudowords even when they occur in stimuli that do not contain real roots, it could be concluded that the effects of frequency/numerosity of suffixes occur at a processing stage in which affix morphemes are available independently of their semantic content. When suffixes are combined with nonroots, no interpretability of the combination should be expected, because of the absence of a meaningful component in first position (for the effects of interpretability of new root-suffix combinations in lexical access, see Burani et al. 1999; see also, for the effects of interpretability on novel Dutch compounds, Coolen, van Jaarsveldt and Schreuder 1991; van Jaarsveldt, Coolen and Schreuder 1994). On the other hand, if we were to show a frequency effect induced by a morphemic unit located in the rightmost position of the stimulus, within an orthographic context in which no lexical or morphemic unit occurs in left position, we would challenge a sequential search model, which predicts that the frequency of the second constituent should not affect lexical processing (Taft and Forster 1976). Evidence against this sort of model has been provided by studies which used both compound nonwords (Lima and Pollatsek 1983), and real compound words (Andrews 1986; Andrews and Davies 1999; Pollatsek, Hyönä and Bertram 2000). The main finding of these studies, which mainly employed lexical decision, but also the recording of eye movements in sentence reading (Pollatsek, Hyönä and Bertram 2000), was an effect of the lexical status or of the frequency of the second constituent. Evidence for a frequency effect of the second constituent when this is a suffix is still lacking. However, even in a theoretical framework which incorporates principles of interactive activation (Taft 1994), it is still assumed that, while inflectional endings would be stripped off in word processing, derivational suffixes would not, because of their different role in processing. Within the latter framework, all the pseudowords that are tested in our experiment, provided they are equated in their leftmost nonlexical part on purely orthographic grounds, should be rejected equally fast, irrespective of the presence of a suffix on their rightmost side. In contrast with these predictions, if interference effects on nonword lexical decision do arise when the stimulus includes a highfrequency suffix in combination with a nonroot orthographic sequence, a

166

Burani and Thornton

model in which the processing system activates frequent morphemic units, irrespective of their relative locations within the word, would be supported In Experiment 1, three sets of Italian derivational suffixes were selected. The three sets, matched for length in letters and phonemes, and for orthographic/phonological structure, differed only for frequency, calculated both on word tokens and on word types. Suffixes in the three sets could be considered of high, medium and low frequency, respectively, by considering the overall distribution of frequency values of Italian suffixes of the same length. The main prediction was that high-frequency suffixes should cause more interference on nonword decision when included in pseudoword contexts, relative to low-frequency suffixes. Suffixes of medium frequency might either not constitute sufficiently activated processing units for interference to occur, or they might show interference effects of a smaller size than high-frequency suffixes.

3.1. Method 3.1.1. Materials and design Nine suffixes were selected, equally subdivided in three experimental sets, of high, medium and low frequency, respectively. Frequencies, in this experiment and in all the following experiments, were derived from a corpus of Italian written language of 1.5 million tokens (Istituto di Linguistica Computazionale CNR 1989). The mean suffix frequencies in the three sets, calculated on word tokens, were 1,060 per 1.5 million (range: 639-1,557); 68.3 (range: 55-90), and 12.3 (range: 7-18), for the three sets, respectively. Differences in mean suffix numerosity between the three sets, calculated on word types in the corpus, paralleled differences in frequency. The mean number of word types in which suffixes occurred were 165 (range: 145187), 20.6 (range: 11-39), and 6.3 (range: 2-10) for the three sets, respectively. There were both nominal and adjectival suffixes. No suffix was homonymous with another Italian suffix. All suffixes were four-letter long and were matched across sets for length in phonemes and for syllabic structure. For each suffix, a control sequence with similar orthographic and syllabic structure was selected. The sequences corresponding to a suffix and the control sequences were matched for orthographic frequency in word

The interplay of root, suffix and whole-wordfrequency

167

final position in each set. The mean frequencies of control sequences in word final position were: 1,538 per 1.5 million for the first set; 472 for the second set; 64 for the third set. These values were matched to the mean frequencies of the orthographic strings corresponding to the selected suffixes, calculated in word final position and including both real suffixes and pseudosuffixes: 1,335 per 1.5 million for the high-frequency set; 411 for the medium-frequency set, and 62 for the low-frequency set, respectively. The suffixes were combined with orthographically legal letter sequences that did not correspond to any existing root. Each suffix was combined with four different pseudoroots, for a total of twelve pseudowords in each set. The length of pseudowords fell within the length range of Italian words including the same suffix, and respected as much as possible the distribution of word length for each suffix in the Italian language (calculations were based on Ratti et al. 1988). Mean lengths in letters of the pseudowords were 9.1, 8.6, and 8.7 for the three sets, respectively. Each suffixed pseudoword was matched with a control pseudoword that included the same pseudoroot in combination with the orthographic sequence that constituted the control sequence for the suffix. Thus pseudowords in each suffixed-control pair had the same length, the same syllabic structure, similar orthographic/phonological structure, and similar orthographic frequency of the final part, either corresponding to a suffix or to a nonsuffix. Pseudowords including suffixes and control sequences were also matched for bigram frequency. Mean bigram frequencies, calculated on the base of the natural logarithm, were: 10.49, 10.57, and 10.60 for pseudowords in the three suffixed sets; they were 10.80, 10.55, and 10.44 for pseudowords in control sets. In combining initial letter strings with suffixes and control letter sequences, we avoided the presence in the pseudowords of embedded real words. Pseudowords in the six sets were also matched for their overall degree of orthographic similarity to a real word, i.e., for the number of orthographic neighbors. Adopting the N-count measure (Coltheart et al. 1977), i.e., the total number of words that can be obtained from each pseudoword by replacing one letter at a time with another letter, while preserving the other letters' positions, we determined that the great majority of pseudowords had a null N-count (with a few exceptions of N-count = 1, balanced across sets), i.e., we obtained pseudowords that were equally dissimilar from existing words, according to the N-count metric. In synthesis, there were six sets of pseudowords, arranged in a 2x3 design, in which the main factors were the presence vs. absence of a suffix,

168

Burani and Thornton

and the high vs. medium vs. low frequency of the orthographic sequence corresponding either to a suffix or to a nonsuffix. In each of the six sets, there were 12 pseudowords (4 for each suffix or control sequence), for a total of 72 experimental stimuli, 36 suffixed and 36 controls. The experimental items, with the mean RT and percent error for each item, are reported in Appendix A. In order to avoid presenting the same pseudoroot to the same participant both in the suffixed and in the nonsuffixed control condition, each experimental set was split in two subsets of 6 items each. Each participant was presented with 36 experimental pseudowords, 18 suffixed and 18 pseudosuffixed, in which no pseudoroot was repeated. In each subset there were two instances of the same suffix or final control sequence. For each set of suffixed and control pseudowords, the entire set of single scores was provided by two participants presented with two complementary sublists. In each sublist, the 36 experimental pseudowords were presented together with 66 filler pseudowords and 102 filler words. Each participant was presented with a total of 204 stimuli. Filler stimuli were the same in each of the two sublists: Words included medium/low frequency singular nouns and adjectives, either derivationally suffixed or nonsuffixed, in a proportion that reflected the composition of the Italian basic dictionary in the medium/low frequency range (see Thornton, Iacobini and Burani 1994; 1997). Each suffix and each control final sequence that occurred in experimental pseudowords was also included in the same number of filler words. Filler pseudowords were drawn from words analogous to the filler words by changing one or two letters in different positions. Mean length was the same for words and pseudowords (range: 6-11 letters). The list was presented to participants in a single experimental session, arranged in three randomized blocks of 68 items each. For each block, participants were assigned to one of two different randomizations of items. Each experimental list was preceded by a practice list of 50 items, 25 words and 25 pseudowords, assigned in the same proportion to two randomized blocks.

3.1.2. Procedure Participants were tested individually in a soundproof experimental booth. They received standard lexical decision printed instructions in which they were asked to decide as quickly and as accurately as possible whether a

The interplay of root, suffix and whole-wordfrequency

169

presented letter string was an Italian word or not. If it was a word (YES response), they had to press the right one of two response keys, otherwise (NO response) the left one. For left-handed participants, the order of the response buttons was reversed. Each trial started with the presentation of a fixation mark (a cross) in the center of the screen for 400 ms, followed after 300 ms by the stimulus centered at the same position. Stimuli were presented on a monitor in white uppercase letters on a dark background and remained on the screen until the participant pressed one of the two response buttons. They disappeared after a time period of 1,500 milliseconds if no response was given. A new trial began 1,200 ms after responding or time-out. If a participant responded more slowly than the preset limit of 1.5 sec, the words FUORI TEMPO ('out of time') appeared on the screen. If the participant gave the wrong response, the word ERRORE ('error') appeared on the screen. This signal was displayed for 500 ms. The interval between the disappearance of the feedback and the next warning signal was 1,200 ms. There was a pause after each block of stimuli. The total duration of the experimental session was approximately 20 minutes.

3.1.3. Participants Forty-eight participants, mostly University students, were paid to participate in the experiment. All were native speakers of Italian.

3.2. Results and discussion The data of four participants, whose mean reaction times for correct responses or whose error rates were more extreme than 2 s.d from the mean of all participants, were excluded from further analysis. Using the remaining forty-four participants, the mean reaction times and error rates for all items were obtained and one pair of items in the medium-frequency set was removed because the number of errors for one of the two members of the pair (crofusso) was more than 2.5 s.d. above the mean. When means for length, bigram frequency and N-count were recalculated after removing the two paired items, the sets were still balanced. The remaining observations were used to calculate participants' and items' mean reaction times and error scores. Mean reaction times by items and percentages of errors for

170

Burani and Thornton

the three experimental categories and their respective controls are shown in Table 1. Table 1. Experiment 1. Mean reaction times by items in ms. and % error. Suffixed and control (nonsuffixed) pseudowords, with high-frequency (HF), medium-frequency (MF), and low-frequency (LF) final sequences Suffixed

Control

Difference

HF

Mean RT % Error

739 13.6

700 6.1

+39 + 7.5

MF

Mean RT %Error

701 4.1

701 7.4

0 -3.3

LF

Mean RT %Error

680 4.6

679 4.2

+1 + 0.4

Results were submitted to a mixed three-way analysis of variance with two within-participants factors: Suffixedness (suffixed vs. nonsuffixed pseudowords) and frequency of final sequence, both suffix and control (high vs. medium vs. low). The third between-participants factor was list (first vs. second sublist, each administered to one half of participants). The ANOVAs were performed both by participants and by items and showed interaction between suffixedness and frequency on both reaction times (Fl(2,84) = 7.59, p.l in both cases). Percent errors on suffixed pseudowords were 3.3 less and 0.4 more in the medium-frequency and in the low-frequency sets, respectively. These differences were not significant (p>.l in both cases). An effect of frequency was found on both reaction times (Fl(2,84) = 19.78, p.l, MSE= 72.23; F2(l,58) = 1.27, p>.l, MSE= 1.48). As revealed by the strong interaction between suffixedness and frequency of final sequence, and by post-hoc comparisons, response latencies and percentages of errors to suffixed pseudowords were higher, relative to their controls, only when the pseudowords included a high-frequency suffix. By contrast, pseudowords that included either medium-frequency or low-frequency suffixes did not reveal longer reaction times nor lower accuracy with respect to matched orthographic controls. These results confirm those obtained by Burani et al. (1997): High-frequency suffixes activate corresponding morphemic access units in pseudoword contexts. By contrast, no access unit seems to be available for suffixes that are either medium- or low-frequency, at least not in pseudoword contexts and within the time required to perform lexical decision. The present results allow us to build on the findings by Burani et al. (1997). In the present experiment, the interference effect caused by highfrequency suffixes occurred with suffixes that were combined with nonexisting roots. Hence, activation of morphemic lexical units corresponding to suffixes occurred in the absence of a real root on their left side. This finding hardly seems compatible with sequential search accounts (Taft and Forster 1976), and with recent reformulations (Tafit 1994), which predict that the frequency of the second constituent, the derivational suffix, should not affect lexical processing in the absence of a lexical unit as first constituent.

4. Experiment 2 Experiment 1 provided evidence for a role in processing of suffix frequency, with high-frequency suffixes significantly affecting rejection latencies in visual lexical decisions to pseudowords. In Experiment 1 there was no evidence that suffixes of medium/low frequency which extended up to a frequency of 90 per 1.5 million constituted effective processing units: No interference arose in lexical decision, when a suffix was either medium- or low-frequency.

172

Burani and Thornton

The role of suffix frequency in lexical decision to real words was assessed in Experiment 2, by varying both root and suffix frequency in transparent derived words of low surface frequency. Derived words included suffixes belonging to two sets of differing frequencies. Suffixes of high frequency were contrasted with suffixes that were of medium/low frequency. In Experiment 1 there was no evidence for differences between medium- and low-frequency suffixes. Hence, suffixes from both the latter frequency ranges were pooled together in a single set. For simplicity, hereafter we will refer to medium/low- frequency suffixes as low-frequency suffixes. All low-frequency derived words should in principle be accessed through constituent morphemes - even when both the root and the suffix are low-frequency, the derived word is nevertheless lower in frequency than its constituent morphemes. Hence, for low-frequency derived words that are equated for all the relevant properties except for frequency of the two constituent morphemes, either high or low, predictions were that both reaction times and error rates should not be function of whole-word frequency, but should rather reflect differences in the frequency of morphemic constituents. Words including higher-frequency morphemes were expected to be accessed more quickly and more accurately than words including lowerfrequency morphemes, with words including both root and suffix of high frequency being the fastest and the most accurate, and words with lowfrequency root and suffix being the slowest and the least accurate. Derived words in which only one morpheme, either the root or the suffix, is of high frequency, were expected to show intermediate reaction times and error rates. If for printed stimuli simultaneous parallel activation of both root and affix is assumed, irrespectively of their relative positions within the word, words in which the high-frequency constituent is either the root or the suffix were not expected to differ in activation times.

4.1. Method 4.1.1. Materials and design Four sets of equally low-frequency suffixed derived words were selected. In the four sets, root and suffix frequency varied orthogonally: The first set included high-frequency roots and high-frequency suffixes (HH); the sec-

The interplay of root, suffix and whole-wordfrequency

173

ond set included low-frequency roots and high-frequency suffixes (LH); the third set included high-frequency roots and low-frequency suffixes (HL); the fourth set included low-frequency roots and low-frequency suffixes (LL). Suffixes were either high- or low-frequency on both tokens and types, i.e., on both frequency tout court and numerosity. The root frequency measure included the cumulative frequency of both the inflected and the derived forms of the base. Thirteen words (nouns and adjectives) were included in each set, for a total of nine different suffixes in each set. No suffix was homonymous with a different Italian suffix. The same suffixes were included in the two highfrequency suffix sets and in the two low-frequency suffix sets, respectively. Suffixes were three to five letters long. High-frequency suffixes and low-frequency suffixes were matched for length and syllabic structure. Roots were different in the four sets. Root length was balanced across sets. The roots belonged to different grammatical categories (i.e., nouns, adjectives and verbs) that were balanced across sets. All the derived words were presented in singular citation form. They were orthographically and phonologically transparent with respect to their bases, i.e., there was no orthographic/phonological assimilation at the boundary between root and suffix. Across the four sets, words were matched for surface frequency, length, syllable structure and bigram frequency. The 52 experimental words were presented together with 108 filler words and 160 filler pseudowords, for a total of 320 stimuli. Any suffix that occurred in experimental words occurred also in the same number of filler pseudowords. Filler pseudowords were drawn from words analogous to the filler words by changing one or two letters in different positions in the word. Filler words included medium/low-frequency singular nouns and adjectives, either morphologically complex or simple, in a proportion that reflected the composition of the Italian basic dictionary in the medium/low frequency range (Thornton, Iacobini and Burani 1994; 1997). Mean length was the same for words and pseudowords (range: 6-11 letters). The list was presented to participants in a single experimental session, arranged in four randomized blocks of 80 items each. Each participant was presented with a different block randomization and with a different randomization of items within each block. Each experimental list was preceded by a practice list of 40 items, 20 words and 20 pseudowords, assigned in the same proportion to two randomized blocks.

174

Burani and Thornton

4.1.2. Procedure The procedure was the same as in Experiment 1. The experimental session lasted about 30 minutes.

4.1.3. Participants Forty-five participants, mostly University students, were paid to participate in the experiment. All were native speakers of Italian.

4.2. Results and discussion The data of ten participants, who made more than 15 percent errors on the experimental words, were excluded from further analysis. Using the remaining thirty-five participants, the mean reaction times and error rates for all items were obtained. We removed four experimental words that showed error rates exceeding 40% from the data set. One word (tenerume) was removed in set HL, two words (aratore and larvale) in set LH, and one word (ameboide) in set LL. One item (rimanenza) was removed in set HH because it was the only word which included a prefixed bound root of an irregular verb. Removal of these items did not affect the matching of the four sets for the relevant variables. In Table 2 the mean values with standard deviations for the variables in each experimental set are reported. The list of the experimental items, with root frequency, suffix frequency, word frequency, mean RT and percent error for each item, are reported in Appendix B. The remaining observations were used to calculate participant and item mean reaction times and error scores. Mean reaction times by items and percentages of errors for the four experimental sets are shown in Table 3.

The interplay of root, suffix and whole-wordfrequency

175

Table 2. Experiment 2. Mean values and standard deviations (s.d.) for the relevant variables. HH = Derived words with high-frequency root and high-frequency suffix HL = Derived words with high-frequency root and low-frequency suffix LH = Derived words with low-frequency root and high-frequency suffix LL = Derived words with low-frequency root and low-frequency suffix HH

HL

LH

LL

Mean

s.d.

Mean

s.d.

Mean

s.d.

Mean

s.d.

Root frequency

501

412.1

507

409.8

33.5

19.1

33.3

21.9

Family size Suffix frequency Suffix numerosity Semantic relatedness Familiarity Bignun frequency Word length in letters Root length in letters Suffix length in letters Word frequency

15.7 1,859 246 3.77 6.57 10.79 8.2 4.4 3.8

8.05 U15 123.6 0.43 0.78 0.4 1.5 1.4 0.6

10.20 57 17.3 3.47 6.05 10.69 8.2 4.3 3.9

3.82 33.7 10.8 0.89 1.03 0.3 1.1 1.1 0.5

4.4 1,636 217 3.66 6.15 10.56 8.4 4.5 3.9

2.5 1260 10139 0.68 0.9 0.3 1.0 1.0 0J

4.2 58 17.7 3.45 5.75 10.60 8.2 4.4 3.8

1.7 32J 10.6 0.95 1.01 0.3 0.9 0.8 0.6

3.1

2.8

3.5

3.2

2.2

2.3

1.7

1.2

Table 3. Experiment 2. Mean reaction times by items in ms and % error. Suffixed derived words with high-frequency root and high-frequency suffix (HH); high-frequency root and low-frequency suffix (HL); lowfrequency root and high-frequency suffix (LH); low-frequency root and low-frequency suffix (LL). HH

HL

Mean Reaction Time % Error

597 2.5 LH

624 5.9 LL

Mean Reaction Time % Error

634 6.7

670 12.2

Results were submitted to two-way analyses of variance, with root frequency (high vi. low) and suffix frequency (high vs. low) as the two factors. There were main effects of both root frequency and suffix frequency on both reaction times and error rates. For root frequency, Fl(l,34)=77.83, pc.001, MSE= 771.3; F2(l,43)= 7.96, p.l). A similar pattern was found on errors, with significant differences between HH words on the one hand, and LH, LL, and ND words, on the other (always p