The Mental Lexicon: Core Perspectives

The Mental Lexicon: Core Perspectives This page intentionally left blank The Mental Lexicon: Core Perspectives EDI...

Author: Gary Libben; Gonia Jarema

52 downloads 1119 Views 2MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

The Mental Lexicon: Core Perspectives

This page intentionally left blank

The Mental Lexicon: Core Perspectives

EDITED BY Gonia Jarema Institut universitaire de gériatrie de Montréal, Canada Gary Libben Department of Linguistics, University of Alberta, Canada

Amsterdam – Boston – Heidelberg – London – New York – Oxford Paris – San Diego – San Francisco – Singapore – Sydney – Tokyo

Elsevier Linacre House, Jordan Hill, Oxford OX2 8DP, UK Radarweg 29, PO Box 211, 1000 AE Amsterdam, The Netherlands First edition 2007 Copyright © 2007 Elsevier Ltd. All rights reserved No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without the prior written permission of the publisher Permissions may be sought directly from Elsevier’s Science & Technology Rights Department in Oxford, UK: phone (+44) (0) 1865 843830; fax (+44) (0) 1865 853333; email: [email protected]. Alternatively you can submit your request online by visiting the Elsevier web site at http://elsevier.com/locate/permissions, and selecting Obtaining permission to use Elsevier material Notice No responsibility is assumed by the publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Library of Congress Cataloging-in-Publication Data A catalog record for this book is available from the Library of Congress ISBN: 978-0-08-045353-8 For information on all Elsevier publications visit our website at books.elsevier.com Printed and bound in The Netherlands 07 08 09 10 11

10 9 8 7 6 5 4 3 2 1

PREFACE AND ACKNOWLEDGEMENTS

Research on the Mental Lexicon has occupied a central role in the cognitive sciences for over three decades. This reflects a consensus that the investigation of words in the mind offers a unique opportunity to understand both human language ability and general human cognition. The goal of this volume is to bring together core perspectives on the representation and processing of words in the mind by providing an overview of key aspects of mental lexicon research and a discussion of the implications of this research for the functional architecture of human language and cognition. The creation of this volume has indeed been a team activity. We would like to thank Sarah Oates of Elsevier who helped us bring the project through the early stages of development and Ben Davie of Elsevier, who helped see it to completion. We would like to express our special thanks to our assistants in the editorial process, without whose careful and insightful attention to detail, the volume would not have come to fruition. Linda Bergeron and Elizabeth French scrutinized both the style and format of all chapters of the volume and worked effectively and tirelessly through the many stages required to mold manuscript text into a complete and integrated volume. They were assisted in their work by Heather Golberg, to whom we also express our gratitude. Together with the authors of this volume, we offer our great appreciation for the support that we have received from the Social Sciences and Humanities Research Council of Canada (SSHRC). The Council has been a critical factor in the inception, development, and completion of the project. It has funded the research collaborations that have inspired the collegial exchange of ideas and perspectives that are represented in this volume and also much research on the mental lexicon that is referred to herein. Of course, it is the contributors who are at the heart of the volume. They have presented us with syntheses of what has been learned so far, their characterizations of the key hurdles and opportunities in mental lexicon research, and their sense of how this research is likely to develop in the future. It has been our great pleasure to work with them and we are confident that readers will find their insights as informative and as thought provoking as we have. Gonia Jarema and Gary Libben

v

This page intentionally left blank

CONTRIBUTORS

Mark Aronoff, Department of Linguistics, SUNY at Stony Brook, Stony Brook, NY 117944376, USA [email protected] Harald Baayen, Max Planck Institute for Psycholinguistics, Wundtlaan 1, 6525 XD Nijmegen, The Netherlands [email protected] Dana M. Basnight-Brown, Department of Psychology, The University at Albany, SUNY, Social Sciences 369, 1400 Washington Avenue, Albany, NY 12222, USA [email protected] Wolfgang U. Dressler, Dept. of Linguistics, University of Vienna, Berggasse 11, A-1090 Vienna, Austria [email protected] Laurie Beth Feldman, Department of Psychology, The University at Albany, SUNY, Social Sciences 369, 1400 Washington Avenue, Albany, NY 12222, USA [email protected] Kenneth I. Forster, Department of Psychology, University of Arizona, 1503 E University Blvd., Building 68, Tucson, Arizona, 85721, USA [email protected] Mira Goral, Speech-Language-Hearing Sciences, Lehman College, 250 Bedford Park Blvd., Bronx, NY 10468, USA [email protected] Geoff Hollis, Department of Psychology, University of Alberta, P220 Biological Sciences Building, Edmonton, Alberta T6G 2E9, Canada [email protected]

vii

viii

Contributors

Gonia Jarema, Département de linguistique et de traduction, Université de Montréal, C.P. 6128, succursale Centre-ville, Montréal, Québec H3C 3J7, Canada [email protected] Gary Libben, Department of Linguistics, 4-32 Assiniboia Hall, University of Alberta, Edmonton, Alberta T6G 2E7, Canada [email protected] James Myers, Graduate Institute of Linguistics, National Chung Cheng University, Min-Hsiung, Chia-Yi 62102, Taiwan [email protected] Loraine K. Obler, Program in Speech and Hearing Sciences, CUNY Graduate Center, 365 Fifth Avenue, New York City, NY 10016, USA [email protected] Dominiek Sandra, Department of Psychology, University of Antwerp, Prinsstraat 13, 2000 Antwerp, Belgium [email protected] Chris F. Westbury, Department of Psychology, University of Alberta, P220 Biological Sciences Building, Edmonton, Alberta T6G 2E9, Canada [email protected]

CONTENTS

1.

Introduction: Matters of Definition and Core Perspectives Gonia Jarema and Gary Libben

2.

Putting Humpty Together Again: Synthetic Approaches to Nonlinear Variable Effects Underlying Lexical Access Chris F. Westbury and Geoff Hollis

1

7

3.

Visual Word Recognition: Problems and Issues Kenneth I. Forster

31

4.

Language: Between Words and Grammar Mark Aronoff

55

5.

Storage and Computation in the Mental Lexicon R.H. Baayen

81

6.

Generative Morphology as Psycholinguistics James Myers

105

7.

Origins of Cross-Language Differences in Word Recognition Laurie Beth Feldman and Dana M. Basnight-Brown

129

8.

Productivity in Word Formation Wolfgang Dressler

159

9.

Bilingual Lexica Loraine K. Obler and Mira Goral

185

10. Skills and Representations in Learning to Spell and in Experienced Spellers Dominiek Sandra

ix

207

This page intentionally left blank

1

INTRODUCTION: MATTERS OF DEFINITION AND CORE PERSPECTIVES

Gonia Jarema, University of Montréal, Montréal, Canada Gary Libben, University of Alberta, Edmonton, Canada

1. MATTERS OF DEFINITION In planning this volume, our goal was to facilitate the creation of a discussion that would not only present new knowledge concerning the mental lexicon, but would target, in a fundamental way, the question of the essential character of the enterprise, its dimensions, and its most probable trajectory. To do this most often requires that one step back from some of the details of research in order to consider what one has learned about the appropriateness of particular assumptions and methodological approaches and about the fundamental nature of the object of inquiry. In the case of the mental lexicon, characterizing the object of inquiry is not as straightforward as it might seem. The metaphor of a mental lexicon implies a thing – commonly referred to as the dictionary represented in the mind, which allows individual language users to engage in everyday processes of language comprehension and production. Yet, the vast majority of psycholinguistic research on the mental lexicon involves the investigation of lexical processing, from which lexical representation is inferred. Thus, mental lexicon research is in practice the study of lexical activity. This dual character of the mental lexicon brings it naturally to the intersection of psychology and linguistics and forces us to deal with the fact that the metaphors that we use can be at once tools and encumbrances. Similarly, definitions associated with such metaphors can both move us forward and shackle us to the ground. Perhaps for this reason, definitions of the mental lexicon are so rarely articulated. Indeed, as authors, as organizers of conferences on the mental lexicon, and as editors of a journal called The Mental Lexicon, we

1

2

The Mental Lexicon: Core Perspectives

do not recall any paper beginning or ending with a sentence like “The mental lexicon can be defined as…”. The most likely reason for this is that authors and presenters know that any attempt at a definition will likely be wrong or, at the very least, incomplete. We know that the mental lexicon is not really a thing. But, on the other hand, defining it solely in terms of processing loses the insight that people can be said to possess words, to acquire words, to use words, and to lose words. There is a fundamental way in which words seem to be entities that can be stored and counted. In our view, a second reason for the reluctance of authors to align themselves with a particular definition of the mental lexicon is that it seems to defy the assignment of boundaries. It is clear, for example, that a complete definition of the mental lexicon, which can also be seen as an adequate theory of the mental lexicon, will be in a co-dependence relation with an adequate theory of phonology, of morphology, of syntax, of semantics, and to an adequate theory of the psychological processes with which those linguistic concepts are associated. To truly know what the mental lexicon is, we would have to be in possession of some consensus on answers to questions such as “What is the meaning of meaning?” and perhaps more fundamentally “How do we assign labels to individual and cultural experience so that we can achieve and maintain contact between minds?” For this is indeed what the possession of a vocabulary allows us to do. Acknowledging all the problems and potential pitfalls noted above, let us nevertheless attempt a definition: “The mental lexicon is the cognitive system that constitutes the capacity for conscious and unconscious lexical activity.” It seems relatively uncontroversial to suggest that the mental lexicon is a cognitive system. The term system suggests a degree of functional integrity that we know to be characteristic of lexical activity. Words may or may not be represented in a similar manner in the mind/brain. There is no question, however, that words are linked to one another. The term cognitive system highlights this fact, while making no claims regarding the extent to which the mental lexicon is monolithic and the extent to which it is structurally or functionally encapsulated. Our intention in suggesting such a definition is to create focus but not blinders, and to pay particular attention to both the nature of the research enterprise and to the notion of a mental lexicon itself. A key component of the definition above is that it characterizes the mental lexicon not as that entity which enables lexical activity but, rather, as that entity which is lexical activity. Word comprehension, for example, is a lexical activity. It takes place in the mental lexicon. Saying and writing words are lexical activities. Their cognitive components also take place in the mental lexicon. It seems to us also that a focus on the mental lexicon as “capacity” has some advantages. First, it draws our attention to what humans can do with words, not only what

Introduction

3

they typically do do with words. Of all aspects of language, the lexical component changes the most over the lifespan, with the acquisition of new words extending throughout adulthood. In that sense, our mental lexicons are never fixed and never cease being linguistic capacities. This is most true in the case of morphological productivity, discussed in this book by Wolfgang Dressler. In our lexical activity, we not only encounter new words, we also create many new words. This capacity, in our view, deserves to have central status in any characterization of the mental lexicon. Our definition of the mental lexicon as “the cognitive system that constitutes the capacity for conscious and unconscious lexical activity” seems to contain a redundancy. Why specify “conscious and unconscious”? After all, if this is all that is possible, the terms do not seem to provide much useful information. Our intention here, however, is to foreground the claim that the psycholinguistic study of the mental lexicon should target the understanding of both conscious and unconscious lexical activities as part of core online lexical processing. Spelling, for example, as discussed by Dominiek Sandra in this volume, is a language production activity that typically occurs in a time frame that allows for considerable conscious intervention. Indeed, it is not only spelling, but many other aspects of lexical activity that seem to have a substantial conscious component. Native speakers of a language who have not had specialized linguistic training can typically say relatively little about their phonemic inventory or their grammatical constraints. They can, however, quite readily access and reflect upon many aspects of their lexical knowledge. But the bulk of psycholinguistic research on the mental lexicon has focused on the unconscious aspects of lexical activity. Here, it seems to us that the term lexical activity is preferable to formulations such as word access or word production. The term lexical activity highlights the covert processes of word composition and decomposition, of lexical priming, and, generally, of the activities in the first few hundred milliseconds of activation.

2. CORE PERSPECTIVES The definition we have proposed is very unrestrictive. It does not say much about what the mental lexicon is not. It does not seek to answer the core questions. Rather, it seeks to frame them. To our minds, the goal of mental lexicon research is indeed to understand the human capacity for conscious and unconscious lexical activity. But now, what is the best way to achieve this goal, what must we absolutely take into consideration? What kinds of variables do we need to consider and how should they best be modeled? The views that a researcher holds with respect to questions such as these will have substantial impact on how that researcher situates the enterprise of mental lexicon research and what he or she considers to be of key importance. In inviting authors in this book to present their core perspectives on the mental lexicon, we were not looking for consensus. Rather, our goal was to provide the structure within which insights could be gained from

4

The Mental Lexicon: Core Perspectives

distinct vantage points and, in so doing, generate a set of core perspectives that would advance our understanding by bringing key question to the foreground and by proposing the principles and possibilities that might serve to answer them. We discuss some of these below.

2.1. Can We Have a Common Architecture for All Languages? Mental lexicon research is unavoidably framed within the boundaries of specific languages, yet linguistic investigation has long demonstrated that, in terms of structure, languages vary significantly along multiple dimensions. One such dimension is morphological structure. As is evident from Wolfgang Dressler’s discussion of productivity and James Myers’ discussion of Chinese, words feature differing degrees of complexity and differing structural patterns across languages. Therefore a legitimate and important question to ask when attempting to define the nature of lexical activity is whether the cognitive system that constitutes the capacity for this activity is one and the same across languages. A corollary of this question, then, is whether cognitive models proposed on the basis of observations in one language do in fact generalize to other languages. Ultimately, what one would need to establish is whether behavioural differences across languages merely reflect structural variation or rather reveal fundamentally different processing patterns. Laurie Feldman’s chapter bears directly on this issue. Feldman argues that, in the domain of visual word recognition, differences observed across languages may be the result of non-standardized methodologies rather than structural variation per se. Indeed, the data presented in that chapter point to similarity rather than divergence across Indo-European and Semitic languages. It may thus be the case that once researchers adopt more comparable methodologies (i.e., common tasks and common sets of factors), common processing patterns will be revealed and a common architecture, rather than diverse architectures, will emerge.

2.2. Gaining Insight from the Non-Obvious In striving to achieve insights into the nature of lexical activity, researchers have mainly concentrated on evidence from word recognition and word production. There exist, however, language skills that have received much less attention. This may be due to methodological issues or considerations of ecological validity and generalizability. Yet converging evidence from varied empirical sources can shed light on phenomena that are not fully elucidated. One such understudied source, discussed in Dominiek Sandra’s chapter, is the capacity to spell words. In this domain, the following types of questions are asked: Do children memorize spelling patterns? Are they aware of orthographic regularities, i.e., of any “rules” that might govern spelling? What role does analogy play in their ability to spell? Importantly, these questions employ constructs that are also central in many other aspects of mental lexicon research. Sandra presents evidence that children rely on memorized patterns even when they are aware of rules, thus offering support to perspectives such as those advanced by Harald

Introduction

5

Baayen in his chapter. In addition, Sandra reports that children’s spelling ability is subserved by, and can be predicted from, their phonological and morphological awareness. This underscores the extent to which metalinguistic knowledge such as the conscious awareness of the linguistic concepts of phoneme and morpheme may be necessary to fully master particular linguistic skills. It has long been appreciated that the study of bilingual processing offers an important window into the nature of lexical activity. In their chapter, Loraine Obler and Mira Goral address key issues as well as methodological hurdles in the study of bilingual and multilingual lexical processing. They discuss new directions and also bring to our attention other less-studied and in a sense less-obvious phenomena such as idioms which, as they point out, are particularly relevant for the study of bilingual processing because they are typically acquired late by both monolinguals and multilinguals. We see in this chapter how the study of bilinguals enables us to probe cross-linguistic variables in a unique manner and to ensure that we acknowledge the fact that the assumption of “one individual, one language, one lexicon” does not accord with the reality of language use around the world – one in which the majority of people are bilingual. This seems to be a case in which the non-obvious should be completely obvious, but is not.

2.3. What is the Right Approach to Modeling the Mental Lexicon? In a fundamental way, the likelihood of significant advancement in the study of lexical activity is related to the likelihood that we are thinking about it in the right manner and from a profitable vantage point. In his chapter, Mark Aronoff draws our attention to the primacy of the word in the characterization of both structurally simple and structurally complex lexical representations. By focusing on the centrality of productivity in word formation, Wolfgang Dressler, on the other hand, brings our attention to the manner in which morphemic representations must be employed. The chapter by Chris Westbury and Geoff Hollis and the chapter by Harald Baayen present cases for a more mathematical approach to the modeling of the mental lexicon, and one in which new classes of variables and new computational entities are posited. Baayen presents a case for how Hierarchical Temporal Memory may provide the means by which computation and storage in the mental lexicon can be understood in a more integrated and adequate manner. Westbury and Hollis present a case for the use of Genetic Programming in order to understand the interaction of variables that may be considerably more ontologically realistic. The approaches in both these chapters represent renunciations of the traditional “divide and conquer” dictum in experimental psycholinguistics. They both advocate abandoning the assumption of linearity in the examination of psycholinguistic variables and Westbury and Hollis, in particular, highlight the advantages of synthetic rather than analytic approaches to lexical modeling.

6

The Mental Lexicon: Core Perspectives

A rather different perspective is presented by Kenneth Forster, who highlights the fact that being able to model (in the sense of reproduce computationally) a particular phenomenon is not at all the same as being able to explain it, or even understand it. He is considerably more skeptical about the profitability of examining large numbers of variables in regression approaches and thus stands in support of the “divide and conquer” approach. The types of perspectives outlined above are part of the interdisciplinary nature of mental lexicon research. Disciplines and theoretical approaches bring their own methodologies to the table. A researcher’s choice of method and level of abstraction will be substantially influenced by his or her disciplinary background and by his or her specific research goals. Perhaps, then, the success of multidisciplinarity in a domain such as mental lexicon research is to be judged neither in terms of methodological consensus nor in terms of metaphorical consensus. Rather, it seems to us that the real benefits of multidisciplinarity lie in the expansion of the body of information that must be acknowledged and considered in the formulation of adequate explanations and theories. James Myers presents an optimistic perspective on such interdisciplinarity and on the ways in which theoretical linguistics and experimental psycholinguistics can both benefit from the insights of the other. He documents the manner in which the disciplines have moved into much closer proximity and are each in a position to gain from the strengths of the other. Lexical processing in general and morphological processing in particular may represent the domains in which this is most likely to happen. Myers concludes in stating that “perhaps it isn’t overly idealistic to hope that a mutually respectful multiculturalism will someday become the norm in the language sciences as a whole”. It is our hope that readers will obtain a similar sense of mutually respectful multiculturalism across the chapters of this volume. The positions are strong and, in our view, they must be so if we are to achieve clarity regarding what about the mental lexicon is known, what is under dispute, and which directions are likely to be most fruitful.

2

PUTTING HUMPTY TOGETHER AGAIN: SYNTHETIC APPROACHES TO NONLINEAR VARIABLE EFFECTS UNDERLYING LEXICAL ACCESS

Chris F. Westbury and Geoff Hollis, University of Alberta, Edmonton, Canada “[The brain is an] enchanted loom, where millions of flashing shuttles weave a dissolving pattern, always a meaningful pattern though never an abiding one…” Charles Sherrington, Man and His Nature

1. INTRODUCTION In his charming book on the role of fictional entities in science and other cognitive enterprises, the philosopher Hans Vaihinger wrote that “Psychological conditions in particular are so intricate that, a priori, just those fictions are on the whole possible and conceivable which [sic] in the main emphasize only one point and neglect others in order thus to make the treatment more practicable” (Vaihinger, 1949, p. 21). The observation serves to remind us that experiments in psychology must always be a compromise. We experimental psychologists do our best to find the best dependent measure, the best experimental paradigm, the best model of the phenomenon under study, and the best control of nuisance variables that we can. However, we are always aware that a slight change in any of these might have given us better insight. Perhaps we controlled for a factor that is actually interacting in a relevant way with our independent variables. Perhaps another paradigm would give us a more sensitive measure of the effects of our independent variables. Perhaps our predictors would have been better at explaining variance in the dependent measure if we had transformed them, by taking their logarithm, their square root, or their inverse. Perhaps

7

8

The Mental Lexicon: Core Perspectives

the constructs guiding our choice of experiments are too simplistic or non-optimally defined in some other way. Perhaps a complex nonlinear regression would reveal that our predictors are together much better at accounting for variance than any linear model suggests they are. As Vaihinger suggests, there are too many possibilities for us to test all of them. Every experiment in psychology is a compromise that makes experimental treatment practical. In this chapter we consider in detail some implications of and ways of dealing with this necessary compromise in experimental psycholinguistics. Psycholinguistics abounds with well-defined independent variables that have been shown to impact on accessing the mental lexicon: frequency, orthographic and phonological neighbourhood, summed bigram frequency, morphological family size, and word entropy measures, to name just a few. In the first section of this chapter, we will consider an ontological question that has been pursued in psychology since its beginnings: In what sense are these variables “real”? Are they pure constructs – fictions, in Vaihinger’s sense? Is their apparent causal role an epiphenomenon of the psychological computation, or do they somehow actually have a causal role to play? Are they perhaps proxies for some correlated factor that really does play a role in the computation? What implications do the answers to these questions have for our conceptualization of the mental lexicon? In the second section we review some of the work that has been done recently in constructing new predictors of lexical access that are explicitly non-linear, and consider why this matters. In the penultimate section, we will focus in more detail on work we have done ourselves, that uses computational methods to synthesize non-linear combinations of predictors that maximize the amount of variance we can account for in behavioral measures of lexical access.

2. PSYCHO-ONTOLOGY: VARIABLES

ON

THE

NATURE

OF

PSYCHOLOGICAL

Discussions of the ontological status of psychological variables go back to the beginnings of experimental psychology (see Carnap, 1936; Tolman, 1938; Hull, 1943). The first scientific psychologists made a distinction between two kinds of variables: intervening variables and hypothetical constructs. A hypothetical construct is an entity for which an existence claim has been made. An example of a hypothetical construct is a subatomic particle. Some particles have not yet been observed, even indirectly, but it is claimed that they exist, with enough evidence to justify the construction of very expensive machines to try to detect them. Perhaps a clearer historical example is genes, which were for many decades a hypothetical construct. It was known that something had to exist to carry characters in the observed distributions, but until the discovery of DNA no one knew exactly what that thing was (Westbury and Dennett, 2000). An intervening variable is defined as any variable that is reducible to an empirical event or property, but for which no claim of existence has been made. An example of an intervening variable is temperature, which we know to be a proxy measure of the motion of

Putting Humpty Together Again

9

particles. In this sense, we could theoretically (however inconveniently) discard the construct of heat and instead speak directly of the measured motions of real particles. This kind of elimination is by definition always possible for intervening variables. They summarize or otherwise represent an observable phenomenon that could theoretically be directly observed and described without ever mentioning the intervening variable. A contemporary reviewer of the distinction between hypothetical constructs and intervening variables has noted that “a failure to distinguish these two leads to fundamental confusions” (Meehl and MacCorquodale, 1991, p. 262). The two must be treated in very different ways. To disprove hypothetical construct, one can only appeal to empirical evidence. Many hypothetical entities (for example: ether, the four humours, Cartesian hydraulics) have been proven not to exist, by finding observable entities that could do all the same work as the postulated non-observable entities. Disproving intervening variables is a very different matter, since an intervening variable is nothing more than a description of a state of affairs; a way of looking at things. For this reason, Meehl and MacCorquodale (1991) claimed that “the only consideration which can be raised with respect to a given proposed intervening variable . . . is the question of convenience” (p. 260; see also Meehl, 1977). Refuting an intervening variable would be like refuting a graph or painting. Assuming that there has been no outright fraud (that is, assuming that the representation does accord with the known facts), one cannot refute a graph or a painting. The best one can do is to suggest that a different representation would be better for some purpose. The reasons for offering such a suggestion are many, ranging from the purely aesthetic (you prefer the pristine simplicity of line graphs to the visual heaviness of bar graphs; you prefer impressionism to realism) to the highly pragmatic (the way data were graphed exaggerates what is in fact a very small difference; an architectural rendering specifies relevant details for constructing a building that a water color painting of the building does not). The bane of cognitive psychology (in particular) as a science is that the variables available for the study of cognitive processes are always intervening variables1. Assuming that materialism is true and dualism is false, the actual empirical entities that underlie cognitive variables are streams of charged particles careening at great speed between and within neurons. No one who understands the matter could believe that these streams could be observed directly in their entirety or that those streams are even constant within an individual over time or between individuals (see Quine, 1960; Rorty, 1970; Churchland, 1981, 1988 for sophisticated discussions of the eliminative materialist position that comes closest to making such a claim). For this reason psychology has to content itself with descriptions that may accord better or worse with the observed facts, but that ultimately are only convenient fictions. 1

Many would claim that it is not just psychology, but all of science that is limited to intervening variables. Reichenbach (1938) memorably referred to the existence claim for hypothetical constructs as “surplus meaning”.

10

The Mental Lexicon: Core Perspectives

2.1. Construct Extravagance: The Problem of Over-fitting We have found that the claim that all cognitive variables are intervening variables is met with strong resistance from many quarters. Some scientists simply do not like to be told that their work consists of constructing what the philosopher Ludwig Wittgenstein called “perspicuous representations”, those representations that produce “just that understanding which consists in ‘seeing connections’” (1953, S122). Others object for more pragmatic reasons, suggesting that the kind of construct profligacy the claim implies will open up issues that are not just philosophically distasteful but actually anti-scientific. In particular, we are often told that the claim that if all cognitive variables are intervening variables, then over-fitting our data is inevitable. Since this claim would be the end of cognitive science if it were true, it is worth considering in some detail. Over-fitting refers to the problem of handling degrees of freedom, the number of free parameters in a model. If a model contains too many degrees of freedom – in particular, at least as many degrees of freedom as the phenomenon being modeled – then it is possible to build models that are extremely accurate but trivial. If a model developed on a specific dataset is allowed to have too many degrees of freedom, the result will often be a model that is accurate for the data on which it was produced, but which generalizes very poorly to other data sets collected in similar circumstances. In essence, over-fit models of this type model a specific dataset related to a phenomenon, instead of modeling a general principle underlying that phenomenon in all its myriad manifestations. Modeling a general principle is science. Modeling a specific dataset is tautology. The argument against construct profligacy goes roughly as follows. If all of cognitive science’s variables are intervening variables, and if intervening variables are descriptions that can vary as we please, and if things than can vary as we please are degrees of freedom, then cognitive science has an infinite number of degrees of freedom and therefore all its theories are over-fit. If this argument were correct, all of cognitive science would be waste of time. Fortunately for cognitive scientists, it is false. The error in the argument stems from the equation of intervening variables with degrees of freedom. An intervening variable is not a degree of freedom, because an intervening variable is a description. Descriptions are not degrees of freedom. A reductio ad absurdum will make this clear. We may sometimes think about what kind of descriptive device we will use to present our data to others. A table? A bar graph? A line graph? A boxand-arrow graph? A pie chart? If descriptions of data were degrees of freedom, then each device that was considered for presenting a dataset would add one degree of freedom to that dataset. The absurd outcome would be that the more educated, creative, or thoughtful a scientist was, the more options he or she pondered for data description, the more likely it would be that his or her data was over-fit. Moreover, degrees of freedom would accrue

Putting Humpty Together Again

11

invisibly to others, since there could be no way of knowing by looking at the final data presentation how many different descriptions were considered before it was chosen. An intervening variable has exactly the same relation to a dataset as a graph: it describes the dataset in a certain way. If we don’t like one graph, we are free to choose another. Changes in how a dataset is described cannot affect the dataset itself. Although there is no relation between the number of intervening variables we use to describe a dataset and the degrees of freedom of that dataset, there are nevertheless constraints on the number of intervening variables that should be used in any description. However, the constraints are not mathematical or otherwise formal, nor are they written in stone. They are aesthetic and pragmatic. William of Ockham famously stated in the fourteenth century that “entities should not be multiplied beyond necessity”. He was not stating a law. He was giving sound advice. Most of us accept that accounts of data that have fewer variables are better than those with more variables. There may be exceptions, however. Perhaps a description with more variables might be more mathematically transparent, in which case we might choose to go against Ockham’s Razor and present the more accessible description. Perhaps (as we will see below) the variables that best describe the dataset may be mathematical constructs that have no obvious real-world counterpart, which we might prefer to keep.

3. NONLINEARITY: A NECESSARY EVIL Along with the fallacious argument about over-fitting, another reason for resistance to the idea that psychology is limited to coming up with coherent and useful descriptions is that very few psychologists go beyond categorical and linear analyses. Having just one good description at hand makes it easier to assume that the description is actually reality. The decision to stick to categorical and linear analyses is often made unconsciously – this is what we are trained to do in graduate school and we are rarely informed why we should do otherwise. Our reliance on linear analyses has the unfortunate consequence of valuing (short term) convenience at the expense of (the best approximation to) correctness. In this section we consider some empirical evidence that shows why non-linear, continuous descriptions are vital in psychology in general and psycholinguistics in particular. One important reason why psychologists should be interested in working with nonlinear models was framed by Minsky and Papert (1969). At the time this paper was published, a type of neural network known as a perceptron was garnering a great deal of interest from psychologists. Perceptrons are a type of neural network that have two parallel input nodes chained to a third output node. Banks of perceptrons are capable of psychologically relevant tasks such as pattern recognition and classification. However, Minsky and Papert provided a mathematical proof showing that traditional perceptrons are unable to solve a certain class of problems known as linearly non-separable problems. The definition of these problems requires understanding of some technicalities of representation

12

The Mental Lexicon: Core Perspectives

in neural networks that would take us outside the scope of this article. The standard example of a linearly non-separable problem is the exclusive-OR function, which outputs true if just one of its inputs is true, and outputs false otherwise. Minsky and Papert’s proof that this very simple problem could never be solved by a perceptron rendered perceptrons uninteresting in the context of complex psychological behavior. Since Minsky and Papert’s proof it has been realized that other types of neural networks are powerful enough to offer insights to psychology. Whole perceptrons can be chained together to provide more complex behavior. However, their utility is contingent on the nodes in each perceptron having nonlinear activation functions. Chains of perceptrons with nodes only employing linear activation functions can always be reduced to a single bank of perceptrons (Dawson, 2004, pp. 170-173) and, thus, are uninteresting by Minsky and Papert’s proof. The lesson is that computational power does not necessarily increase with structural complexity in any system that only performs linear transformations on its inputs. If a system is to be psychologically interesting – if it is to be more than merely the sum of its environment – the system must necessarily be a nonlinear one. Several researchers working more directly in language-related fields have found other evidence that suggests nonlinearity should be at the forefront of psycholinguistics. Van Orden et al. (2003) stress the importance of nonlinear dynamics in cognitive systems. Nonlinear, dynamical systems depend on reciprocal causality, as opposed to the domino causality that psychologists typically assume. In a domino world, one thing follows another in an ordered way, as one domino knocks down the next. In a system with nonlinear dynamics, processing does not unfold in a straightforward, sequential, orderly manner. Instead, each subcomponent or sub-process is fundamentally tied to every other one, in such a way that the manipulations or calculations of one can fundamentally augment the state or trajectory of all others. The end result is a system where the individual contribution of any one component cannot be worked out in the behavior of the whole. Such systems are fundamentally nonlinear and no amount of diligent linear analysis can adequately capture the complexity of the system. Such systems can be behaviorally complex, and even chaotic. However, they have a tendency to organize their behavior around attractors in their space of possible behavioral states. These attractors might be point attractors – singularities in behavioral state space (phase space). They might be cyclic attractors – unbreaking loops of behavior in phase space. Or they might be strange attractors. Strange attractors are volumes in a multidimensional phase space that the system converges to. However, once in that attractor, the system does not necessarily follow a recurrent, cyclic pattern as it does with cyclic attractors. Chaotic behavior can persist within the volume of the strange attractor, so long as the system’s behavior does not leave the volume of the strange attractor. Insofar as they organize around attractors, dynamic systems are said to be self-organizing – no matter what their original behavior, they tend to reorganize their behaviors into a specific region of space (for introductions to nonlinear dynamic systems in psychology, see Abraham and Shaw, 1992; Port and van Gelder, 1995; Ward, 2002).

Putting Humpty Together Again

13

This may seem like an abstract philosophical idea on which to found any theory of cognition, language, or language processing. However, some convincing empirical evidence supports the claim that the language access system is a complex dynamic system. One of the hallmarks of a dynamic, self-organizing system is the presence of pink noise. Pink noise shows up as an inverse (1/f) relationship between the frequency and amplitude of composite waves in a signal on log scales. Van Orden et al. (2003) consider a series of word naming tasks in which reaction time was recorded on each trial. They hypothesized that if this behavior was underlain by a self-organizing dynamic system, then reaction times should follow a pink noise pattern instead of the naively assumed white (random) noise. They demonstrated quite clearly that this is the case. The variations in word naming times are not random (normal). Rather, they vary precisely as we would expect them to if they were produced by a self-organizing system. If this sort of self organizing, reciprocally causal hypothesis is a hallmark of language processing systems, then nonlinear analyses must be a fundamental tool for understanding the pattern of interactivity within such systems. It is not going to be convenient to break down all of the massively interdependent components as linear effects. It is simply not tractable. Nonlinear analysis and description is required if the cognitive sciences are to progress. Another line of research implicating nonlinearity in linguistic processing is John Holden’s (2002) work on spelling and pronunciation (see also Van Orden et al., 2001). Holden asserts and provides evidence that ambiguity in orthography and phonology is best described as a fractal pattern. Fractals are a collection of self-similar, nested forms that have no characteristic scale. An example of the British coastline is often given to illustrate what is meant by this concept. The British coastline has many bays and peninsulas. If you were told to measure the length of that coastline, how would you go about doing this? You could use a meter stick to get a rough estimate of the length. But what about all the smaller peninsulas and bays embedded within the ones that your meter stick can easily measure? You could move down to a smaller stick – perhaps a centimeter stick – for finer resolution of these smaller jags. But what about the ones embedded in these finer features you still cannot pick up? True mathematical fractal patterns have infinitely many layers of self-similarity, like these nested bays and peninsulas on the British coastline. They also have no characteristic scale of measurement, since any scale we choose to use – whether it be meters, centimeters, millimeters, or whatever – provides us with a different estimate of length. Holden claims that, just like the bays and peninsulas embedded within each other along the British coastline, there are scales of ambiguity embedded within the sound/spelling of words. Consider the word form, lead. At the whole word (semantic) scale, it is ambiguous with one version of the word rhyming with bead and the other rhyming with head. At the grapheme scale, there is also ambiguity. Consider the _ea_ portion of our word form. This combination of letters maps to multiple phonemic forms – like the two different ones found in the words mead and bread. This sets the stage for multiple scales of ambiguity in word forms.

14

The Mental Lexicon: Core Perspectives

If a psychological phenomenon is best characterized as a fractal pattern, no linear descriptions suffice to capture it – it will take far too many parameters to characterize the layers upon layers of self-similarity within the phenomenon. We must necessarily turn to nonlinear methods of description and analysis. Fractals are fundamentally mathematical entities, so we know they can be described mathematically – however, doing so is beyond tractable with linear methods. Baayen (2005) took steps towards encouraging nonlinear analyses in language research. He used restricted cubic splines (a type of polynomial function with three degrees of freedom) to look at the nonlinear relationships between 13 predictors and lexical decision/word naming reaction times. Six of his predictors enter into nonlinear relationships with reaction times. Furthermore, these six predictors are quite varied conceptually: frequency, morphological family size, inflectional entropy, number of simplex synonym sets, word length, and neighbourhood size. Nine of the predictors he looked at had a nonlinear relationship with naming times. Some of these relationships are also non-monotonic: that is to say, continual increase of cause intensity does not lead to continual increase of effect intensity. For instance, Baayen found that word length is best described as having a u-shaped relationship with word naming times, with a minimum in the bend at a word length of 4 (he suggests this may reflect response optimization of our cognitive systems to the structure of language. The median word length for words in his data happened to be 4). Such u-shaped relationships are impossible to get at with conventional analytical techniques, such as twoway ANOVAs, that are most often used by psychologists. When a factor is looked at, it is typically sampled from the two extremes of its range. This completely overlooks the possibility of finding a nonlinear relationship. Westbury et al. (2003) performed work that is more closely related to our own research (discussed in the next section). Westbury et al. used evolutionary search techniques (genetic programming, described in more detail below; see Koza, 1992) to capture mathematically the nonlinear nature of various psychological phenomena. In one experiment, they mathematically modeled the effect of the interaction between orthographic neighbourhood size and orthographic frequency on lexical decision reaction times (LDRTs). These are two rigorously studied constructs in the psycholinguistic literature, whose relationship is characterized by a frequency-modulated orthographic neighborhood effect. LDRTs are faster for words with more neighbours, but only for low-frequency words. Westbury et al. were able to mathematically characterize this relationship, providing a more in-depth description of how the frequency-mediated neighbourhood effect works (see Figure 1; see also Hollis and Westbury, 2006). Specifically, convergence of the neighbourhood effect appears to happen around a frequency of 16 occurrences/million. The simple non-linear equation they found to describe the relationship (that was used to generate Figure 1) accounts for 23.2% of the variance in LDRTs, as opposed to just 4.8% that was accountable for with a linear regression. The effect is obviously highly nonlinear, and therefore not something accurately characterized by standard linear statistical tools.

Putting Humpty Together Again

15

Figure 1. Graph across a range of variation of the two best-evolved functions relating orthographic neighborhood (ON) to frequency. Equation1 was sensitive to changes across the range of ON; Equation 2 distinguished only between ON ≤ 2 and ON > 2. Reprinted from Westbury et al., 2003. Though relatively little work has been done on characterizing the nonlinear aspects of accessing the mental lexicon, the work reviewed above convincingly shows that such nonlinear description is valuable. The use of non-linear tools helps us more accurately characterize language phenomena, and in some cases it can provide characterizations of phenomena that simply cannot be made with our normal analytic tools (e.g., in the case of non-monotonic relations). The view that the variables of psycholinguistics are intervening variables that relate to behavioral observables in a nonlinear dynamic manner has important implications for the way we conceive of the mental lexicon. These implications are both positive and negative. One positive implication of an awareness that many of the representations we work with were selected as worthy of study for reasons that are primarily pragmatic or aesthetic is that we are less likely to reify those representations. This can allow us to be more creative and flexible in our thinking, which in turn may help us achieve what we see as the only viable scientific goal of psycholinguistics: the construction of the most perspicuous representations of the elements of language. The negative aspect of this is that we are forced to admit that we know less, and will always know less, than we would ideally like to know. Of course we would all like to know how the mental lexicon is “really” accessed. We believe that it is hubris and/or naïveté to believe that such a goal is achievable. Language is highly multi-faceted in its functional structure. That multi-faceted structure is represented in distressingly complex ways neurologically. The idea that somewhere there is a central place (the mental lexicon) where it all comes to together seems to us both highly unlikely and conceptually unnecessary. This

16

The Mental Lexicon: Core Perspectives

tempting view that there must be a central locus of neurological computation has been derided (in a different context) by the philosopher Daniel Dennett and the neurologist Marcel Kinsbourne as being the error of the Cartesian Theatre (Dennett, 1991; Dennett and Kinsbourne, 1992). When we think of lexical access in terms of non-linear dynamic processes that are described with intervening variables, we are freed from the grip of this tempting viewpoint. To us, the mental lexicon is itself an intervening variable: a short-hand and sometimes useful way of conceiving and discussing a system that is in fact dynamic and far too complex to be fully captured by any static representational metaphor. We believe that the claim that words have a single indexed representation that is stored and accessed as a computer stores and accesses indexed information in RAM is misleading. Lexical information is evoked for many reasons and in many ways. The fact that the final outcome seems the same (e.g., a subject produces the word “cat”) cannot necessarily be taken as evidence that the same processes have been evoked to access the same representation, anymore than the fact that we seem to ourselves to have unitary consciousness can necessarily be taken as evidence that we must therefore have a single locus of consciousness (the Cartesian Theatre) inside our heads. We now turn to some more practical demonstrations of the utility of thinking about the mental lexicon in non-linear terms.

4. PUTTING HUMPTY TOGETHER AGAIN: MATHEMATICAL SYNTHESIS IN PSYCHOLINGUISTICS Linear thinking imposes stifling restrictions on our scientific imagination and understanding. Since there is no limit to the number of descriptions we are allowed to ponder, why should we ponder only the single one that satisfies the assumptions of linearity? Why not pick the best one we can find among the infinite number of nonlinear descriptions? If we can describe an algorithmic way of recognizing a good description, then we can go and drink beer and argue about theoretical questions (because that is what we humans are good at) while our computers search through a very large portion of the infinite space of intervening variables for us (because that is what computers are good at). This is how we propose to put Humpty together again. Infinity is problematic. Searching a space that is infinitely large will take infinitely long, no matter how fast our computers search it. When we are faced with a very large space that we must describe or search, one method we often use is random sampling. We could ask our computers to make up random mathematical descriptions, to assess each one for value, and to present the best one it has found to us whenever we ask it to. This method might be mildly helpful: occasionally the computer might come up with a better description of a dataset than we had. However, since the space of possible descriptions (as well as the subspace of frankly stupid descriptions) is very much more massive than the subspace of good descriptions, the odds are naturally stacked against this method.

Putting Humpty Together Again

17

Fortunately, nature has come up with a much more intelligent way of searching infinitely large spaces than randomly: natural selection. Nature does not randomly create new organisms and see if they are viable. It only creates small variants of organisms it already knows are viable, and “keeps” the best variants. Natural selection thereby searches a very small portion of the infinitely large “possible organism space”. There are infinitely many creatures that will never evolve on earth that could have evolved here. However, the ones that did evolve here are (by definition) guaranteed to be good at the one thing they were selected for: producing viable offspring. Natural selection is in fact an algorithm for searching a space of possibilities in a clever way (for mathematical analysis, see Holland, 1992; for an extended non-mathematical discussion of this idea, see Dennett, 1995). Natural selection searches only those possibilities that are similar to possibilities that have previously been identified as being good ones. Instead of trying to randomly come up with a good idea, natural selection comes up with good ideas by tweaking good ideas it has already had. We have made use of this algorithm to get our computers to search the space of intervening variables, using a computational technique known as genetic programming (GP). The algorithm of natural selection needs three things: a way of producing variants, a way of choosing the good ones, and a way of tweaking those good ones to produce more variants. The production of variants is easy when we consider the space of mathematically defined models. Each variant is an equation that relates some measured set of independent variables to some dependent variable of interest. For example, if we wanted to have a mathematical model of how ON and orthographic frequency interact in terms of their affect on LDRT, we can easily tell the computer to make up equations that set some function of ON and frequency equal to RT. How do we distinguish a better predictor equation from a worse one? This is the problem of defining a fitness function. In genetic programming, fitness can be anything that we wish to maximize and that we can describe to a computer. In our work to date we have used a well-defined and well-motivated fitness function: the amount of variance that we can explain in some dependent measure (the square of the linear correlation). When we compare two mathematical models of a phenomenon to each other, we usually think that the one that fits the phenomenon most closely is the better one. The amount of variance explained is one way (though not the only way) of measuring how closely any mathematical model matches the data. This measure has the advantage of being both mathematically well defined (so we can tell the computer how to use it) and bounded (convenient for avoiding over-flow errors such as may occur if one used summed squared error as a fitness function, since error can be arbitrarily large). We have a way producing random variants and a way of assessing their goodness as predictors of some dependent variable. The third and final element we need for genetic programming is a way of tweaking the good variants so that we look for variants in the same general area of search space as the good ones we have already found. Since our variants are equations, this is very easy to do: we can simply combine random elements of good equations

18

The Mental Lexicon: Core Perspectives

to make up new equations. We represent our equations as trees (using reverse Polish notation, which always places the operator before its arguments), randomly choose two equations identified as good, and swap one random substring chosen from each of those equations. The “offspring” are similar to their parents in the sense that each one only contains elements that are also contained in its parents. However, they are different from their parents because they contain at least one element that is not contained in either parent.2 The new population is populated with these offspring, and the process of assessing fitness and allowing the fit to create offspring is repeated. Over time, equations evolve that are better and better at the task for which they have been selected: predicting variance in the dependent measure. This process is purely stochastic and there is no guarantee that any solution reached is the best one. We use many software techniques to increase the probability of finding a good solution. We will not discuss them in detail here (they are discussed in Westbury et al., 2003; Hollis and Westbury, 2006; Hollis et al., 2006). Most of them amount in one way or another to one of two methods. The first is checking to make sure that evolved equations are good at predicting variance in datasets other than the one the one on which they were evolved, to decrease the chance of over-fitting by finding a locally good but universally bad solution. The second method is repeating searches many times, to increase the probability of covering a sufficiently large portion of the search space. Hollis and Westbury (2006) demonstrated that solutions evolved independently are usually very highly correlated, which increases our confidence that these may indeed be the best solutions. They also conducted simulations showing that when there is a known perfect or very good solution to a prediction problem, GP usually finds it very quickly. We have released free platform-independent software that allows users to harness the power of GP for themselves very easily. The software is called Naturalistic University of Alberta Nonlinear Correlation Explorer (NUANCE) and is available from the Psychonomic Society website at http://psychonomic.org/archive or from the University of Alberta website at http://www.psych.ualberta.ca/~westburylab (see Hollis and Westbury, 2006; Hollis et al., 2006). In the remainder of this section, we discuss one application of that software, in which we used it to study the effects of 16 predictors of LDRTs and their pairwise interactions3. LDRTs are currently a particularly suitable dependent measure for synthetic work on the type we are advocating here, because a large collection of average LDRTs has already been compiled for the English Lexicon Project (Balota et al., 2002). One advantage of the synthetic approach over factorially-designed experiments is that it is possible to study the effects of predictors over a very large set of stimuli if a large set of dependent measures is available. In the work we report here, we looked at LDRTs for 4778 2

It is of course possible that the two randomly chosen substrings might by chance swap an identical substring, resulting in an offspring that is a clone of one of its parents. However, our software disallows this possibility, as well as disallowing duplicates in the population. 3 This work has been previously discussed in Hollis et al., 2006.

Putting Humpty Together Again

19

words. This was every word that was between three and six letters long for which we were able to obtain measures for all 16 of our predictors, The sixteen predictors are defined in Table 1. They consisted of five measures computed across both orthographic and phonological word representations: length (letter or phoneme count), frequency, neighbourhood size, neighbourhood frequency, and positioncontrolled and position-uncontrolled bigram/biphone frequencies. These were calculated using the CELEX database (Baayen et al., 1995). We used the same database to calculate initial and final controlled-trigram frequencies for all stimuli, as an estimate of orthographic head and body frequencies. The final two predictors were measures taken from a cooccurrence model of semantic organization (Shaoul and Westbury, 2006). These models depend on measures of the Euclidean distance between long vectors representing how often each word in the dictionary occurred within a small window (10 words) of each word. One measure, NN, was the number of co-occurrence neighbours (from our 150,000 word dictionary) that fell within a specified distance threshold from the target word. The other measure, Average Radius of Context or ARC, is the average distance from the target word of those co-occurrence neighbours. Each of these has been shown to be predictive of LDRTs (Shaoul and Westbury, 2006).

4.1. Method Our method was to use NUANCE to evolve equations using one or two predictors that accounted for the most variance in the LDRTs. We randomly divided our 4778 words into two sets. One set, the training set, was used to evolve the equations. The other set, the validation set, was set aside so that we could make sure those equations generalized to a new dataset. We ran every predictor singleton or pair through NUANCE twenty times in order to maximize the chance that we would find the best predictor equation. An important point to stress is that we report results for the best predictor equation across all of those runs, as it performed on the validation set. Very few users of linear regression ever set aside a data subset to use as a validation set. Failing to do so can almost guarantee that the reported r-values will be over-estimates. Any regression – linear or otherwise – computed across a particular dataset is able to account for both dataset-specific and dataset-general variance. Dataset-general variance is of scientific interest. However, dataset-specific variance is, by definition, a source of error when a regression equation is applied to a different set. Increasing error will decrease the amount of variance that can be accounted for in a new dataset. Many studies compound this error by computing their regression equations over a small dataset, which makes it easier to fit. It is safe to say that almost every reported correlation in the psycholinguistic literature is an over-estimate (see Westbury et al., 2003, for an empirical demonstration).

20

The Mental Lexicon: Core Perspectives

4.2. Results We had several goals in undertaking this study. One was to find mathematical models of variable-predictor relations that would allow us to better understand those relations. Table 1 The sixteen measures used to predict LDRTs Variable

Description

LETTERS

The word’s length, in letters.

PHONEMES

The word’s length, in phonemes.

OFREQ

The orthographic frequency (per million) of the word.

ON

The number of orthographic neighbours of the word.

ONFREQ

The average OFREQ of the word’s orthographic neighbours.

PFREQ

The phonological frequency (per million) of the word.

PN

The number of phonological neighbours of the word.

PNFREQ

The PFREQ of the word’s phonological neighbours.

CONBG

The summed frequency that any two letter-pairs in the word occur together, in the place they are in the current word, in same-length words.

UNBG

The summed frequency that any two letter-pairs in the word occur. Position in the word and word length do not matter.

CONBP

The summed frequency that all two phoneme-pairs in the word occur together in the place they are in for the current word, only in words with an equal number of phonemes.

UNBP

The summed frequency that any two phoneme-pair in the word occur. Position in the word and phoneme count do not matter.

FIRSTTRI

The frequency with which the first three letters of the word occur as the first three letters for all words of the same length.

LASTTRI

The frequency with which the last three letters of the word occur as the last three letters for all words of the same length.

ARC

The average distance between a word and its co-occurrence neighbours.

NN

The number of co-occurrence neighbours the word has.

We were particularly interested in understanding how much explained variance is “lost” under the assumption of linearity. The amount of variance in LDRTs accounted for in

Putting Humpty Together Again

21

the validation set by each individual predictor in displayed in Table 24. In that table, we compare the amount of variance in LDRTs that is accounted for by the best evolved equation to that accounted for by a linear regression of the same predictor, and by a linear regression of the same predictor after log transformation. Fifteen of the sixteen variables (all but ONFREQ, the average frequency of the target word’s orthographic neighbours) were reliably linearly correlated with LDRTs ( p < 0.05). All of them were reliably correlated after log transformation, which improved the correlation for every variable except UNBP (the lengthuncontrolled, place-uncontrolled summed biphone frequency of the target word) and LETTERS and PHONEMES (the number of letters and phonemes, small numbers for which log transformation makes little sense). We statistically tested the difference in the correlations between the NUANCE transformations, and their untransformed and logged counterparts (see Blalock, 1972, for methodological details). Eight of the sixteen NUANCE-transformed values correlated reliably better ( p < 0.05) with LDRTs than either their untransformed or logged values. Some of the cases (e.g., ONFREQ, PNFREQ, PFREQ, and CONBG) are of particular interest because the transformation changed a correlation that was very near 0 to a correlation that was much higher. The average of the untransformed r 2 values of ONFREQ, PNFREQ, PFREQ, and CONBG is a negligible 0.002. The average of their NUANCE-transformed r 2 values is over 41 times larger, 0.07. Although log transformation of the four variables reduces this difference substantially, the average r 2 value of the NUANCE-transformed variables is still 1.4 times larger than the average of the log-transformed variables (0.05). This comparison of r 2 values underscores one benefit of looking at correlations across a range, as opposed to factorial manipulations, of predictors: one is able to get an accurate idea of the effect size one is studying across that entire range. Almost everyone understands that a reliable effect in a factorial experiment is no guarantee that the independent variable accounts for a compelling amount of variance in the dependent variable. Although it is possible to calculate an effect size for any experimental effect, this can only give the size of the effect at the values where the factors were blocked, usually their extremes. None of the untransformed variables accounted for more than 5% of the variance in LDRTs. The average was 2.3%. When the predictors were logged, the average was 6.9%. The NUANCEtransformed variables accounted for an average of 8.3%. Simply averaging the r 2 values is problematic, for the reason we have given above. When we average r 2 values, some of the variance is being counted more than once because some of the variables are accounting for over-lapping variance. The average therefore overestimates how much variance is being accounted for. One of the benefits of the NUANCEtransformed variables is that they were evolved to have as close a linear relationship to the dependent variable as possible. This makes it appropriate to run a linear stepwise backwards regression on LDRTs of those transformed variables, despite our general suspicion of linear 4

Note that the summed variance accounted for is greater than 1. This reflects the fact that many of the variables are highly correlated, and account individually for the same variance.

22

The Mental Lexicon: Core Perspectives

modeling. In this case we know that the relation of the transformed values to the LDRTs is indeed roughly linear, because it was explicitly designed to be so. Table 2 Variance in LDRTs accounted for on the validation set by the 16 predictors, their log transformations, and their best-fit NUANCE transformation Variable

Untransformed Log Transform NUANCE

OFREQ

0.015***

0.331

0.363††

PFREQ

0.002**

0.121

0.141†

LASTTRI

0.003***

0.072

0.131†††

FIRSTTRI

0.004***

0.092

0.115††

ON

0.078***

0.093

0.093

NN

0.065***

0.096

0.085

ONFREQ

0.000

0.045

0.076†††

PN

0.054***

0.072

0.066

ARC

0.039***

0.047

0.059

LETTERS

0.053***

0.048

0.059

PHONEMES

0.042***

0.039

0.048

PNFREQ

0.001*

0.027

0.048†††

CONBG

0.004***

0.008

0.025†††

UNBP

0.011***

0.011

0.018†

UNBG

0.006***

0.007

0.006

CONBP

0.001*

0.005

0.003

Note. All log and NUANCE-transformed effects significant at p < 0.001. For untransformed variables, p-values of 0.05, 0.01 and 0.001 denoted by *, **, and ***, respectively. Differences in predictive power between NUANCE-derived fits and best maximum of the other two fits are marked: p-values of 0.05, 0.01, and 0.001 denoted by †, ††, and †††, respectively (for the methodology used to determine significance values for correlational differences, see Blalock, 1972). Reprinted from Hollis and Westbury, 2006.

When we did this regression, five of the variables did not enter into the model: PFREQ, CONBP, PN, PNFREQ, and ONFREQ. The first three have closely orthographic analogues in the model. The eleven remaining predictors together accounted for 41% of the variance in LDRTs. Just four of those eleven variables account for 96% of this variance (39% of total variance): OFREQ (orthographic word frequency), LETTERS (word length), ON

Putting Humpty Together Again

23

(orthographic neighbourhood), and LASTRI (the frequency of the last three letters of the word, an approximate measure of body frequency). All of these variables have been wellstudied and are well-known in the study of lexical access. 4.2.1. Interactions. Accounting for as much variance as possible in a relevant behavioral measure is one method of judging the goodness of psychological models. The use of a hard quantitative measure serves as an objective arbiter between models that provide such a measure. However, another desideratum of models of language is that they capture something about how different elements in the model are organized and interact. Information about which predictors interact with each other can potentially be used to constrain language models that are more fully specified, such as box-and-arrow models or neurological models. If we know that two or more predictors interact, then a good model should be able to give an account for why this interaction occurs. Because we have so many variables, we confined ourselves in this study to using NUANCE to look at all pair-wise interactions between our sixteen predictors, in the manner described above. Evolved nonlinear regression equations have one disadvantage compared to linear regression equations. Linear regression equations are designed to break up the variance into that portion attributable to each variable and that attributable to interactions. The equations evolved by NUANCE mix these two sources of variance up in arbitrarily complex ways, leaving no obvious way to decompose each function into its component main and interaction effects. We have addressed this problem using a somewhat imperfect approach that should nevertheless allow us to get a principled estimate of how much variance we can attribute to any interaction effect. Our method takes advantage, as we did above, of the fact that the output from NUANCE has been transformed to optimize its linear relationship to the dependent variable. We conducted two multiple linear regressions, one containing only terms for each variable alone, and the other containing an additional interaction term. By subtracting the variance accounted for by the first equation from the variance accounted for by the second, we obtain an estimate of the size of the interaction term in which we are interested. The approach is imperfect since we cannot be sure that the interaction function contains the same transformation of each predictor as the one that we obtained when we ran that predictor alone: that is, we cannot be sure that two independently-evolved equations are equivalent in terms of how they account for variance in LDRTs. However, we believe that as a rough heuristic this method is a reasonable way to estimate the size of an interaction effect when we cannot decompose the equations by hand. Twelve (10%) of the 120 possible interaction effects were large enough to be reliable after Bonferroni correction. An additional 14 (12%) were reliable ( p < 0.05) before the correction. These 26 interactions are presented in Table 3. Although most of the effects are small in themselves, together they account for 12.9% of the variance in LDRTs. Altogether our 16 variables together therefore can account for as much as 54% of the total variance in

24

The Mental Lexicon: Core Perspectives

LDRTs (or just 3.5% less if we exclude the interactions that did not meet the very conservative Bonferroni correction). This is probably a small overestimation because some of the interactions may be accounting for common variance, since they may share a variable or contain a predictor that is correlated with other predictors. A few observations may serve to guide models of lexical access. One is that several variables that seem to account for almost no variance on their own enter into reliable interactions. For example, ONFREQ (the frequency of the target word’s orthographic neighbours) did not enter into the backwards regression on LDRTs, but it was second only to PFREQ (phonological frequency) in entering into reliable interactions. Similarly, UNBP (the length-uncontrolled, place-uncontrolled summed biphone frequency of the target word) accounted for very little (1.8%) of the variance in LDRTs on its own but was the third most interactive variable. These results raise the possibility that some variables that have little direct effect on lexical access may nevertheless modulate the effects of other variables that do have direct effects on lexical access. Another interesting result from studying interactivity is that it can provide information about how similar or different two variables may be. Across all the words in the validation set, the NUANCE transformation of ON (orthographic neighbourhood size) and ONFREQ (the average frequency of the target word’s orthographic neighbours) are highly correlated (r = 0.74; p < 0.0001), suggesting that these two variables may bear a similar relation to LDRT and therefore perhaps to each other. This would explain why ONFREQ did not enter into the stepwise backward regression: because ON did. However, their pattern of interactivity is quite dissimilar. They both interact with just a single predictor, PFREQ (phonological frequency). ONFREQ interacts with four other variables (UNBP, PN, PHONEMES, LETTERS) while ON interacts with just one other, OFREQ (orthographic frequency). These variables also interact (r = 0.06; p < 0.005). All this suggests that the two variables do indeed probably play distinct roles in lexical access. A somewhat related analysis can be done on the PFREQ and OFREQ, phonological and orthographic frequency. These variables are also highly correlated (r = 0.51 across all 14,582 words of length 4, 5, and 6 in the CELEX database; p < 0.001). Nevertheless, interactions with PFREQ account for about three times as much variance as interactions with OFREQ. This finding invites further study through more traditional studies. We believe that other of the interactions in Table 3 may also reward further investigation, but we leave their identification as an exercise for the reader. 4.2.2. The Shape of Things. Another benefit of the synthetic approach is that it allows researchers to put on firmer ground certain practical decisions that have to be made in the course of psycholinguistic research. Here we consider as an example the convention of taking the logarithm of frequency-influenced variables when they are used as predictors. This transformation is driven by the fact that many psycholinguistic predictors have a non-linear distribution that does not map easily onto the linear range of behavioral measures. The convention of logging flattens the curve. As is clear from Table 2, this makes most log-

Putting Humpty Together Again

25

transformed variables better predictors of behavioral measures than their untransformed counterparts. Table 3 Reliable pair-wise interactions in the validation set among the 16 predictors of LDRT Var. 1

Var. 2

R2 0.015***

Var. 1

Var. 2

PHONEMES PFREQ

R2

LETTERS

PFREQ

0.004**

FIRSTTRI

LASTTRI 0.010***

ON

ONFREQ 0.004**

ON

PFREQ

0.009***

OFREQ

PN

0.003**

CONBP

UNBP

0.008***

PN

PNFREQ

0.003*

LETTERS

OFREQ

0.007***

UNBG

UNBP

0.003*

CONBG

UNBG

0.006***

PHONEMES ON

0.003*

ONFREQ

PFREQ

0.006***

LETTERS

CONBP

0.002*

PHONEMES ONFREQ

0.006***

PN

UNBP

0.002*

ONFREQ

UNBP

0.005***

PN

UNBG

0.002*

ONFREQ

PN

0.005***

ON

NN

0.002*

LETTERS

ONFREQ

0.005***

PFREQ

UNBP

0.002*

OFREQ

ON

0.005***

PNFREQ

UNBP

0.002*

PFREQ

NN

0.004**

PFREQ

PNFREQ

0.002*

ONFREQ

UNBG

0.004**

Note. P-values of p < 0.05/120, p < 0.01, and p < 0.05 denoted by ***, **, and *, respectively. Reprinted from Hollis and Westbury, 2006.

We are not aware of anyone who has ever considered that there may be a better transformation than logging, but Table 2 shows that NUANCE can almost always find a better transformation. We have examined these transformations and found a general pattern. All but two (UNBP and UNBG) of the NUANCE transformations in Table 2 are reciprocal transformations, which transform N into some function of 1/N. On average, those transformations account for 34.6% more variance in LDRTs than the logged variables do. This is “money for nothing”, in the sense that merely looking at the same data set in a different way allowed us to make substantial and well-validated gains in our ability to predict LDRTs. We believe that the reciprocal transformation is a better transformation because its properties are more biologically realistic than the log function. As illustrated in Figure 2, the log function is continually increasing as its argument increases, while the reciprocal function flattens out. Since there is a floor effect in human reaction times, a function that does not

26

The Mental Lexicon: Core Perspectives

continually increase as its argument increases is generally likely to be better fit to the human data. 2

1

0 0 -1

200

400

600

800

1000

1200

zLOG -zRECIPROCAL

-2

-3

-4

-5

-6

Figure 2. A comparison of the shape of the logarithmic and inverse transformations. Values have been normalized and the inverse curve flipped vertically in order to make the differences in their shape clear.

5. CONCLUSION We have had several related goals in this chapter. The first goal has been to champion a synthetic approach towards studies of lexical access. Although we do not believe that a synthetic approach should be the only approach, we do believe the approach is a necessary adjunct to standard factorial experimental manipulations. We provide several reasons for this. One is philosophical: only the synthetic approach can help us escape from the reification of predictors that so dangerously attractive in factorial studies. Just because a manipulation of a factor has a measurable behavioral effect does not necessarily mean that the factor actually enters into the neural computations underlying language. A synthetic approach opens our minds to the idea that alien combinations of well-defined variables that may ultimately prove to be more ontologically realistic than the more neatly defined variables of human imagination. We need to get used to the idea that the variables we define are almost

Putting Humpty Together Again

27

certainly descriptive proxies for what is really happening in our brains when we use language. They are shorthand descriptions for something much more complex. A second reason for our enthusiasm for synthetic approaches is that they allow us to consider the effects of variables across their entire range, rather than limiting us to snapshots of their effect at the extremes of that range as factorial experiments almost always do. When we look at variables of interest as continuous rather than categorical, we get an understanding of their effects that is better in many ways that what we get from factorial experiments: better because we can more easily estimate their effect size; we can see how their effects may change across their range; we can more easily understand how they interact with other variables; and we can more easily consider different ways of defining them that may be more scientifically helpful. A third reason for championing a synthetic approach, to which we have only alluded in this chapter, is purely pragmatic. As the number of well-defined variables that impact on language access increases, it becomes increasingly impractical to conduct all the experiments that would be necessary to understand how they interact with each other. The sixteen variables we considered here by no means exhaust the space of variables that might be worth considering, but they have 120 pair-wise interactions. Although some of those might be ruled out a priori as uninteresting, we are doubtful that researchers’ intuitions have a sufficiently comprehensive grounding to be trustworthy to make such rulings. However, to conduct 120 factorial experiments to determine which interactions might be worthy of consideration in our models of language is totally impractical: the tedious work required would take years and most of it would be dedicated to uncovering null effects. We need more powerful tools for searching variable space than experiments. In this chapter we have not even considered higher-order interactions such as the 560 possible three-way interactions among our variables. However, the synthetic program, by virtue of its principled skepticism about the ontological status of human-defined variables, gives reasons to believe that some of these interactions may be worth knowing about. A related goal to this main goal is to encourage the use of the free tool we have developed for undertaking synthetic studies without making assumptions about linearity: our platform-independent computer program, NUANCE. The last section of this chapter was dedicated to giving some idea of the kinds of information that can be gleaned using NUANCE. It has been designed to be as user-friendly as possible. More detailed information about NUANCE can be obtained from Hollis and Westbury (2006), and Hollis et at. (2006). Another goal of this paper that has been to motivate certain considerations for building models of lexical access/processing. One criterion more implicit than explicit is to encourage the development of algorithmically well-defined measures for comparing models of language processing. Box-and-arrow models of language processing are common, but are almost certain by their very nature to encourage reification and are almost impossible to compare empirically. In this paper we have concentrated on considering mathematical models that predict a behavioral measure. Although such models will never be the final endpoint for understanding language (leaving aside as they do a great many important issues, such as

28

The Mental Lexicon: Core Perspectives

neurological and algorithmic instantiation of the processes suggested by the model) we believe that they are an important and under-used stepping stone towards any final models. Abstract mathematical models can provide a great many constraints on building models that connect in more ways to empirical findings such as lesion data, imaging data, error data, and developmental data. The second consideration motivated by synthetic approaches is to take variable interactions more seriously. Interactions may alert us to computational/neurological cocomputations that may be occurring or to aspects of our understanding where human-defined variables may be too neat and too simple. Most of these goals would not be possible if they were not nested in the meta-goal of taking non-linearity seriously. We believe that the use of only linear methods in psychology and experimental linguistics conceals a great deal of the evidence we need in order to understand the processes we are studying. We may labor in vain for years to understand research puzzles using linear tools that might be trivial if we awoke ourselves from our dogmatic slumbers and allowed ourselves to use non-linear methods. Now that the most under-funded researcher is able to access computational resources that would have been unthinkable a few decades ago, it has become very easy to harness the power of nonlinearity. We titled our chapter “Putting Humpty Together Again”. Although they could not put Humpty together again, we can make some good guesses about what all the king’s men were able to do. Examining the pieces of poor Humpty, they were certainly able to obtain a great deal of locally accurate information about each piece of him. They would have been able to document the weight, shape, and curvature of every piece. What they were not able to do, as we are told, was put all the pieces together. Traditional experimental psycholinguistics is in a similar situation to the king’s men. It has taken us to a stage where we can say that this variable affects that behavior but not this one, whereas another variable has opposite effects, and a third affects both. This is local information about the pieces that may be parts of the much larger process of lexical access. What we cannot yet do is to put these pieces together to come up with an unambiguous, uncontroversial big picture of that process. In part this is because, unlike all the king’s men, we do not even know if we have all the pieces, or the right pieces, or even if the pieces we have are all relevant to our larger goal. We believe that we will only be able to know this if we try to work backwards at the same time that we work forwards, synthesizing back into language the pieces that we have first so carefully analyzed out.

Putting Humpty Together Again

29

REFERENCES Abraham, R. H. and C. D. Shaw (1992). Dynamics: The geometry of behavior. Addison-Wesley, Redwood City, CA. Baayen, R. H. (2005). Data mining at the intersection of psychology and linguistics. In: Twenty-first century psycholinguistics: Four cornerstones (A. Cutler, ed.), pp. 69-83. Lawrence Erlbaum Press, Hillsdale, NJ. Baayen, R. H., R. Piepenbrock and L. Gulikers (1995). The Celex lexical database. Release 2 (CDROM). Philadelphia, Pennsylvania: Linguistic Data Consortium, University of Pennsylvania. Balota, D. A., M. J. Cortese, K. A. Hutchison, J. H. Neely, D. Nelson, G. B. Simpson and R. Treiman (2002). The English Lexicon Project: A web-based repository of descriptive and behavioral measures for 40,481 English words and nonwords. http://elexicon.wustl.edu/, Washington University. Blalock, H. (1972). Social statistics. McGraw-Hill, NY. Carnap, R. (1936). Testability & Meaning, Parts I-III. Philosophy of Science, 3, 419-471. Churchland, P. M. (1981). Eliminative Materialism and the Propositional Attitudes. Journal of Philosophy, 78 (2), 67-90. Churchland, P. M. (1988). The Neurocomputational Perspective: The Nature of Mind and the Structure of Science. MIT Press, Cambridge, Mass. Dawson, M. (2004). Minds and machines: Connectionism and Psychological Modeling. Blackwell Publishing, Malden, MA. Dennett, D. (1991). Consciousness Explained. Little, Brown, And Co, Boston, MA. Dennett, D. (1995). Darwin's Dangerous Idea: Evolution and the Meanings of Life. Simon & Schuster, New York, NY. Dennett, D. and M. Kinsbourne (1992). Time and the observer: The where and when of consciousness in the brain. Behavioral and Brain Sciences, 15 (2), 183-247. Holden, J. G. (2002). Fractal Characteristics of Response Time Variability. Ecological Psychology, 14 (1&2), 53-86. Holland, J. (1992). Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence. MIT Press, Cambridge, MA. Hollis, G. and C. Westbury (2006). NUANCE: Naturalistic University of Alberta Nonlinear Correlation Explorer. Behavior Research Methods, 8 (1), 8-23. Hollis, G., C. Westbury and J. Peterson (2006). NUANCE 3.0: Using Genetic Programming to Model Variable Relationships. Behavior Research Methods, Instruments, and Computers, 38 (2), 218-228. Hull, C. L. (1943). Principles of Behavior. Appleton-Century, New York, NY. Koza, J. (1992). Genetic programming: On the programming of computers by means of natural selection. MIT Press, Cambridge, MA. Meehl, P. (1977). Construct validity in psychological tests. In: Psychodiagnosis: Selected papers (P. Meehl, ed.), pp. 3-31. W.W. Norton & Co, New York, NY. Meehl, P. and K. MacCorquodale (1991). On a distinction between hypothetical constructs and intervening variables. In: Paul E. Meehl: Selected Philosophical & Methodological Papers (C. Anderson and K. Gunderson, eds.), pp.249-263. University of Minnesota Press, Minneapolis, MN.

30

The Mental Lexicon: Core Perspectives

Minsky, M. and S. Papert (1969). Perceptrons: An Introduction to Computational Geometry. MIT Press, Cambridge, MA. Port, R.F. and T. van Gelder (1995). Mind as motion: Explorations in the dynamics of cognition. MIT Press, Cambridge, MA. Quine, W.V.O. (1960). Word and Object. MIT Press, Cambridge, MA. Reichenbach, H. (1938). Experience and prediction. University of Chicago Press, Chicago, IL. Rorty, R. (1970). In Defense of Eliminative Materialism. The Review of Metaphysics, 24,112-121. Shaoul, C. and C. Westbury (2006). Word frequency effects in high dimensional co-occurrence models: A new approach. Behavior Research Methods, 38 (2), 190-195. Sherrington, C. (1940). Man and His Nature. Cambridge University Press, Cambridge, UK. Tolman, E. C. (1938). The determiners of behavior at a choice point. Psychological Review, 45, 1-41. Vaihinger, H. (1949). The Philosophy of ‘As if’. Routledge & Kegan Ltd., London, UK. Van Orden, G., B. F. Pennington and G. O. Stone (2001). What do double dissociations prove? Cognitive Science, 25(1), 111-172. Van Orden, G., J. Holden and M. Turvey (2003). Self-organization of Cognitive Performance. Journal of Experimental Psychology: General, 132 (3), 331–350. Ward, L. (2002). Dynamical Cognitive Science. MIT Press, Cambridge, MA. Westbury, C. & Dennett, D. (2000). Mining the past to construct the future: Memory and belief as forms of knowledge. In: Memory, Brain, and Belief (D. L. Schacter and E. Scarry, eds.), pp. 11-32. Harvard University Press, Cambridge, MA. Westbury, C., L. Buchanan, M. Anderson, M. Rhemtulla and L. Phillips. (2003). Using genetic programming to discover nonlinear variable interactions. Behavior Research Methods, Instruments, and Computers, 28, 202–216. Wittgenstein, L. (1953). Philosophical Investigations. Blackwell, Oxford, England.

3 VISUAL WORD RECOGNITION: PROBLEMS AND ISSUES

Kenneth I. Forster, University of Arizona, Tucson, USA

The study of visual word recognition is about much more than just reading. It’s about one of the most fundamental properties of the brain – the ability to store and rapidly retrieve information about familiar visual patterns. Without this ability, no learning or adaptation to the environment could occur, since each stimulus would essentially be novel. Visual word recognition is also particularly appropriate as an arena to stage the battle between different models of pattern recognition, and it is no accident that much of the debate about neural network models of cognition centers on this topic. In what follows we will briefly explore what I take to be the central issues in the field. This is very much a personal view, much of which has been presented before, but it is to be hoped that a second airing of these comments will bear more fruit this time around.

1. COMPUTATIONAL MODELING A good place to start is the debate about the most appropriate way for theories of complex cognitive systems to be developed. One camp insists that the best (only?) method is to actually build a working model of the system, using computer simulation as the medium, rather than an actual physical model. Examples of this approach are the DRC model (Coltheart et al., 2001), the MROM model (Grainger and Jacobs, 1996), and the various PDP models (Seidenberg and McClelland, 1989; Plaut, 1997). The advantages of this approach are said to be that it is only when the theory is represented as a program that it is explicit, precise, and complete. Without this, one has to rely on imprecise and loosely worded verbal assumptions and arguments in order to decide whether a verbally stated theory actually makes the predictions that are claimed for it.

31

32

The Mental Lexicon: Core Perspectives

I have no doubt that this is correct, but there is a catch. Suppose we have a computationally explicit model that correctly predicts the outcome of an experiment. The question then arises, why did the model behave in this way? As far as providing an explanation of what happened, all we know is that there exists a model which provides output that is the same as that provided by humans, but we have no idea whether the computations in the human brain bear any similarity to the computations in the model. To even make a start on this problem, we need to develop a theory of how the model actually operates. In a PDP model of reading aloud such as the Seidenberg and McClelland (1989) model, this involves the task of deciding which hidden units are mostly responsible for controlling the inputoutput mapping. Is there a special set of hidden units that controls the pronunciation of irregular words, or are the same hidden units involved in both regular and irregular words? A high degree of overlap would lend support to the idea that only a single “route” is required. However, this assumes that there is only one set of weights that needs to be considered, when in fact there may be thousands of different sets of weights that are equivalent. The method used by Coltheart et al. (2001) with DRC is more straightforward, and involves switching off certain components of the model to see what role they play. Nevertheless, it is sometimes quite difficult to determine why the model does what it does. One then needs to resort to debugging techniques, in which the program is halted at various stages so that the state of the system can be carefully inspected. Cutting to the chase – here is the problem. Assuming that we do eventually come to an understanding of why the model behaved the way it did, how is this theory of the model’s behavior to be expressed? In practice, of course it is verbally stated, and it is hard to imagine how it could be otherwise. So we are back where we started. Our theory of the computation is expressed in precisely the loose, imprecise way that we sought to avoid. To see that this must be so, consider the possibility that two investigators could disagree about why the model behaves the way it does. One would then have to carry out experiments with the model to try to find out which view is correct. To avoid this problem, it is sometimes said that the theory is the computer program itself, and there is no point in asking how it works. I think this is unsatisfactory. The object of the exercise is to gain an understanding of how cognitive systems work. Understanding is not something that can be achieved just by inspecting a computer program. So how are other researchers to gain an understanding? They might try experimenting with the program, seeing what happens when this or that parameter is manipulated, trying to predict its behavior. But what is happening is that they are constructing their own informal theory of how the program works. My view is that understanding should come first. Initially, the researcher develops an informal hypothesis about how the human system works, which is expressed verbally. To check that this hypothesis would actually account for the data, a model of the system is developed, expressed as a computer program. During the development of this model, it will be discovered that some parts of the initial theory were incomplete, or inconsistent, and the verbal statement of the theory is then amended, and made more precise. In this way, the

Visual Word Recognition

33

simulation is just an aid for the theorist. It is a way of refining the original idea, of making it more formal. But the program itself does not become the theory.

2. EXPLANATORY VS. DESCRIPTIVE ADEQUACY This is an issue that traditionally engages linguists more than cognitive psychologists, but is vital in both fields. It relates to Chomsky’s distinction between descriptive adequacy (of a grammar) and explanatory adequacy. A grammar could be descriptively adequate if it generated all and only the sentences of a particular language, but the problem is that there are many different types of grammars that could do this equally well. Chomsky’s aim was to take things further, and to provide a theory which explained why the grammar had to be of one particular form. In particular, the grammar had to be something that a child could learn, given fairly limited data. Applying this approach to visual word recognition, a model that accurately simulates a wide range of experimental results is like a grammar that generates a wide range of grammatical constructions. It has descriptive adequacy. To go further, and to claim a degree of explanatory adequacy, we would need a set of general cognitive principles from which the properties of the model could be deduced. These principles would form a kind of metatheory. A possible example might be Modularity theory as proposed by Fodor (1983). So, for example, I assume that the lexical processor is itself a module, contained within a broader module, the language system. Accordingly, the lexical processor must be strictly informationally encapsulated, and hence it knows nothing about semantics or syntax, or about the real world. All it can do is find lexical entries that match up with the specified sensory input. A consequence of this is that this module operates in the same way whether it has context or not. So words in a sentence context must be accessed in the same way as isolated words, i.e., the lexical processor is autonomous. One could just construct a model in which there “happened” not to be any feedback connections between the semantic processor and the lexical processor. This would make the same predictions, but no explanation is given for why there are no such connections. Claiming that this design is what a modular system would dictate represents only a modest increase in explanatory adequacy. We would do far better if we could justify the assumption of modularity on other grounds, e.g., efficiency of design, biological plausibility, etc. An excellent example of this approach is Norris’s model of word recognition, termed the Bayesian Reader (Norris, 2006). In his discussion of the frequency effect in visual word recognition, Norris complains that existing accounts fail to explain why high-frequency words should be recognized faster than low-frequency words. In models such as DRC and MROM, the word units corresponding to high-frequency words are assumed to have higher resting levels of activation than the units corresponding to low-frequency words, and therefore they reach criterion levels of activation faster. But, says Norris, why not give the

34

The Mental Lexicon: Core Perspectives

low-frequency word units higher resting levels of activation as well? That would be more efficient, and the system would still work perfectly well. Thus it seems that the frequency effect is “explained” by making low-frequency words harder to access than they need to be. It would be different if one could show that a system like DRC was quite unable to discriminate accurately between words if all word units had the same resting level, but no attempt has been made to see whether this might be the case (and it seems unlikely that it would). Instead, one can only resort to some kind of “rust” factor, in which neural circuits that have not been used for some time have to overcome greater inertia in order to fire correctly. This at least is an attempt to provide an explanation, unlike the resting level account. The situation is not much better with PDP models, which explain all effects in terms of variations in the strength of connections between input units and hidden units, and between hidden units and output units. According to this approach, each time a word is perceived, its connections are strengthened. Thus the frequency effect is explained as a learning effect. This at least constitutes an explanation, albeit not a very plausible one, since it seems implausible that the slower response to a relatively low-frequency, but perfectly familiar, word like eagle should be the result of incomplete learning. The solution offered by Norris is that people behave, or try to behave, as if they are ideal observers. For Norris, both visual and auditory perception are extremely noisy processes, and therefore the stimulus is ambiguous. Under these conditions, the rational procedure is to take into account the a priori probability of occurrence of a word when evaluating the perceptual evidence, and the best way to do this is to use a Bayesian approach. Since frequency of occurrence maps directly onto a priori probability, one has a ready explanation of the frequency effect. An alternative approach appeals to a hardware limitation, namely that the input stimulus cannot be compared with all word units simultaneously, as assumed in parallel activation models. Instead, the input can be compared only with a limited subset of word units at the same time. In the extreme case, we have a serial search model (e.g., Forster, 1976), in which the input is compared with just one word unit at a time. In a less extreme case, the input is simultaneously compared with a limited subset of units (Forster and Hannagan, in preparation). In either case, the most efficient procedure is to make comparisons with the highest-frequency words first, and the lowest-frequency words last. This approach not only provides an explanation for the existence of a frequency effect, it also provides an explanation of its functional form (Murray and Forster, 2004). Essentially it says that access time should be a linear function of the rank of a word in a frequency-ordered list. Allowing for error, this prediction is essentially correct, although it has not been demonstrated that a rank function is superior to other possible functions (e.g., a power function). These examples serve to illustrate what is meant by explanatory adequacy. How one gets there involves a certain amount of bootstrapping. Initially, the metatheory might be totally speculative, but if it forces the theorist to make assumptions of one type rather than

Visual Word Recognition

35

another, and this eventually leads to correct predictions, then one has confidence that the metatheory might be an approximation of the true state of affairs. More likely, the theory will run into difficulties, and modifications to either the theory, or the metatheory itself will have to be made. In making such modifications, one must constantly ask the question: What is it about the metatheory that forces one to assume X rather than Y? If there is no good answer, then the theory says nothing about the phenomenon in question. Of course, one could make a shrewd guess that X is going to generate better predictions than Y, but if the metatheory is silent with respect to this choice, then such a move is tantamount to cheating. Perhaps the most spectacular example of theory construction driven by a metatheory is the PDP approach. In this case, we have a fairly limited, and precisely stated set of assumptions about how any cognitive system will operate. In the case of visual word recognition, the components (layers) of the system are also specified. The only room to manoeuvre is in the nature of the input units. Indeed, much of the improvement in these “triangle” models has come from improvements in the input coding. Whether these very general principles will serve to generate deeper insights is an open question at the moment.

3. THE AUTONOMOUS LEXICAL PROCESSOR In addition to the assumption of modularity, I also endorse another principle, namely bottomup processing. There are several reasons for making this choice. First, a non-interactive system will be much easier to understand than a fully interactive system. Second, strategically, it is more sensible to start out with a simpler system that is much less flexible, and then see where it breaks down. Third, the idea that top-down expectations can actually alter the input data itself seems counter-productive, which is how I interpret top-down feedback from word units to letter units. If we pursue the analogy between perception and hypothesis testing in science, then this is equivalent to allowing a partially supported hypothesis to alter the evidence to make it more compatible with the hypothesis. One may well decide to ignore disconfirming evidence (say, on the grounds that it is unreliable), but one should never discard that evidence, or worse, alter it to make it more consistent. The notion of an autonomous lexical processor is dictated by the assumption that the component systems of the language system are themselves modules. This means that their operations are domain specific (i.e., the lexical module only knows about word forms, the syntactic module only knows about syntax, and the semantic module only knows about semantics), and they are also informationally encapsulated, which means that they do not rely on information contained within other modules. This is not to say that the modules do not communicate with each other. Instead they communicate through structured interfaces. The lexical module outputs a lexical address which enables other modules to extract the properties of the current word being processed. But there is also a top-down communication system. If the lexical module has made an error, then there has to be some kind of system that instructs

36

The Mental Lexicon: Core Perspectives

the lexical module to go back and reanalyze the input. In the case of normal reading of text, this would typically involve commands to the eye movement system, and hence this system is likely to be a fairly high-level problem-solving system (not part of the language module), but one that receives output from the syntactic and semantic modules. So the important thing to note here is that the operation of the lexical module is not modulated by higher-level systems. It is not provided with contextual information that increases the speed or accuracy of access. A common response to this argument is to assert that everyone knows that words are easier to recognize when they occur in the context of a sentence (e.g., Ehrlich and Rayner, 1981). How can this fact be reconciled with an autonomous lexical processor? Here is where we see the value of a metatheoretical assumption. If one didn’t have a metatheory, one could simply admit the error, and insert the necessary arrow. But when a metatheoretical assumption is at stake, then one looks more carefully to see whether there is another way to interpret the data. I became suspicious of context effects very early on, when I found that the target word had to be degraded before you could get a context effect (Forster, 1976). The dependent measure in this experiment was naming latency. In an earlier time, before reaction times became popular again, it would always be the case that the target was degraded, because the dependent measure was accuracy of identification under (degraded) tachistoscopic conditions. But with reaction time methods, one could present the stimulus in clear, and then it was not so obvious that there was any effect of context. Of course, if you made the context so constraining that only one word would fit the context, then a clear effect was obtained – the expected word was responded to more rapidly than an unexpected word (e.g., Fischler and Bloom, 1979). But why wouldn’t the same thing happen when there were say, ten words that were potentially appropriate to the context? If context had the effect of reducing the number of candidates that needed to be considered, then it should still work in this case. A critical assumption in an interactive view of context is that the context should produce a benefit. However, this is not necessarily the case. Instead, the difference between appropriate and inappropriate words might be due to slower responses to the inappropriate words. Without a neutral baseline, one cannot tell. If the effect was purely inhibitory, then that would change the story entirely. Accordingly, I carried out an experiment using both lexical decision time and naming time as dependent measures, and including a neutral baseline (Forster, 1981). The context words were presented one word at a time at a slow rate, and the subject’s task was to respond as rapidly as possible to the final item in the sequence (preceded by a warning signal). The contexts were either a sentence fragment (e.g., It is a good idea to clean your teeth every ---), or a random sequence of words presented at the same rate as the sentence contexts (this was the neutral baseline). The targets were either highly predictable from context (e.g., day in the previous example), appropriate to the context (e.g., morning), or inappropriate to the context (e.g., month). The results are shown in Figure 1. The left panel shows the results for lexical decision. For predictable completions there is a substantial benefit from the sentence context relative to the list context. For appropriate completions, there was a slight inhibitory effect

Visual Word Recognition

37

(not significant), but for inappropriate completions there was a very substantial inhibitory component. The right panel displays the results for the naming task. Here the context effects are minimal, and tending towards inhibition, but only the inhibitory effect for inappropriate completions was significant. Except for the predictable completions in the lexical decision task, the effects of a sentence context are either null, or inhibitory. Could this inhibitory pattern indicate that the list context (random words) was somehow inappropriate, perhaps because subjects were not as attentive? But this would lead to slower responses in the List condition, which would enhance any facilitatory effect. To try to explain the inhibitory effects in terms of an inappropriate context, we would have to argue that responses in the list context were too fast. The only clear evidence for a facilitation effect comes from the highly predictable completions in the lexical decision task. It seems questionable whether this effect is due to faster lexical access since there is no trace of such an effect in the naming task. LEXICAL DECISION

NAMING 600 580

Sent

600 NAMING TIME (ms)

LEXICAL DECISION TIME (ms)

650

550

500

450

Sent List

400

560

List

540 520 500 480 460 440

Pred

App

Inapp

Pred

App

Inapp

Figure 1. The effects of predictable, appropriate and inappropriate sentence contexts on lexical decision time (left panel) and naming time (right panel) for the final word of the sentence. Adapted from Forster, K.I. (1981). Comments made by subjects in the lexical decision experiment gave a clue as to the right interpretation. These subjects complained that their initial impulse was to respond in terms of the appropriateness of the target word, not its lexical status. So when the target word was highly inappropriate, there was a strong bias to respond “No” (it’s not appropriate), whereas the correct response was “Yes” (it’s a word). However, when the target word was in fact the word they expected, their first impulse was to respond “Yes”, which was the correct response. Of course, no such bias was present in the naming task. It is possible that the absence of facilitation in the appropriate condition was due to the fact all contexts were constructed so that there was a highly predictable completion. This

38

The Mental Lexicon: Core Perspectives

would generate a strong expectancy, and when that expectancy was not fulfilled (as would be the case in the appropriate condition), the response would be slower than normal. Accordingly, a further lexical decision experiment contrasted appropriate and inappropriate completions to relatively unfocussed sentences (e.g., Judy enjoyed reading about ---, where horses was the appropriate completion, and crumbs was the inappropriate completion). The results were unchanged. There was no facilitation for the appropriate completions, but substantial inhibition for the inappropriate completions. It is difficult to see how an interactive model would explain a context effect that was only operative when the target word was highly predictable (and when the task was lexical decision). But that’s only half the problem, because in fact, no explanation has ever been offered as to how a context effect is supposed to work. The general notion is that context provides an additional source of activation, so that the resting levels of word units that correspond to likely continuations of the sentence are increased. This sounds reasonable, until one tries to implement such a system. For example, one might imagine that all the word units for edible objects are connected together, so that when a verb such as eat is encountered, all these words would receive activation. But obviously, this will depend on the subject of the verb, because alligators don’t eat the same kinds of things as humans. One doesn’t want to be in the position of predicting no difference in reaction time to the word hay when the context is The alligator ate the --- compared with The horse ate the ---. To get the right prediction here, one would have to create subcategories. The things that an alligator could plausibly eat would need to be connected together, so that only when the subject of the verb was alligator would these words receive activation. In addition, to explain the inhibitory effects observed for inappropriate completions, one would have to decrease the resting activation levels for all other words. But even this is insufficient. Consider what might happen to hay when the context is None of the horses would eat the ---. Would we expect faster responses to hay than say, cactus? A simpler solution is to assume that recognition of the target word is strictly bottom up (provided that it is not degraded), and that the context plays no role prior to access. However, it does play a role when the target word is integrated with the context, and inhibitory effects may well be produced when the target word is difficult to integrate semantically with the prior context in a plausible manner. So why should degradation matter? My assumption is that under these conditions, the bottom-up system either fails altogether, or is very slow. Under these conditions, the reader adopts a problem-solving approach. If some of the letters can be identified, the reader essentially switches to what could be called crossword mode. What’s a nine-letter word beginning with de- and contains -ss- and means “unhappy”? There is no question that readers possess such a skill, and there is no point doing experiments that do nothing more than demonstrating the existence of such a skill. For this reason, experiments with degraded targets are not particularly helpful. What is critical in this debate is when context plays a role – before the stimulus is recognized or after? Assuming that it plays a role before, amounts to requiring that the

Visual Word Recognition

39

language system is constantly attempting to predict the upcoming input. This is obviously computationally very expensive. However, recent experiments using event related brain potentials have provided convincing evidence that prediction is involved in sentence processing. Using a clever design, Van Berkum et al. (2005) showed an increased N400 effect for an adjective that differed in syntactic gender from that of an upcoming predictable noun, indicating clearly that an expectancy for that word had been formed. No effect was observed when the context was modified so that the noun was no longer predictable. This evidence is taken as a refutation of a purely bottom-up approach. However, it is not clear that this evidence is actually decisive. Obviously, the effect could only occur for highly predictable words (if there were several possibilities with different genders, then no gender mismatch could occur). Further, it is not clear that the context facilitated the recognition of the predicted noun. What it did was to establish the belief that such a word would occur. Much the same effect could presumably be obtained if the context was not so constraining, but subjects were told what the final noun in the sentence would be. One needs to be clear about what is meant by the term “autonomy”. To demonstrate a violation of this principle, one would have to show that the operation of the normal bottomup system (the one that is used when there is no constraint imposed by context) is fundamentally modified by the context. Preactivating the word units for all plausible continuations of the sentence would qualify as an example, but that is not what this experiment shows. Nor would it qualify to so degrade the input that the bottom-up system fails altogether, thereby forcing the utilization of a quite different top-down system. Another claimed refutation of autonomy is the work reported by Van den Brink et al. (2006). They demonstrated that the detection of a semantic anomaly in a spoken sentence (as indexed by the onset of an elevated N400) begins before the isolation point (the point at which there is sufficient information to identify the word without context), as indexed by a gating procedure. Once again, this must depend on how predictable the critical word is. To take an example of their materials, the context is This 14-yr old pianist obviously has a lot of ---, and the anomalous completion is climate instead of talent. The results show that the N400 begins before completion of the word climate, and well before it was clear what word was presented. Another way to put it, however, is to say that the N400 began as soon as it was clear that the sentence was ending with something other than the word talent. What does this show? It shows that at some point in the sentence (presumably somewhere around the word lot), subjects anticipated that the sentence would end with the word talent. When the input failed to confirm this expectancy, an N400 was generated. However, this does not necessarily show that feedback from the semantic processing system modified the configuration of the bottom-up lexical processor so that any input other than talent would generate an error signal. The discovery that the input was not what was expected may have been made by a pure bottom-up processor that operated entirely independently of context. In some respects, I fear that the issue of the autonomy of the lexical processor is just a war of words. There can be no doubt that we can anticipate particular lexical items in highly predictable environments, and there is also good evidence that we anticipate what kinds of

40

The Mental Lexicon: Core Perspectives

syntactic constructions are more likely in certain contexts (Lau et al., 2006). These predictive mechanisms will have behavioral consequences, but the crucial issue is whether we are dealing with one system or two. One way to gain information about external stimuli is to make predictions about what should occur given the context, and then to test whether those predictions are valid. Another way is to ignore the context, and to build up an interpretation based purely on the stimulus information. (The latter system is presumably the one that tests the validity of the predictions made by the former system.) If we insist that these two methods are really different aspects of the same system, then we must conclude that the lexical processor is not autonomous. But if we believe that they reflect properties of entirely separate systems, making use of different sources of information, then we can continue to maintain our belief in an autonomous system.

4. IS ACTIVATION INCREMENTAL? What is really important about computational modeling is that it forces the theorist to come up with concrete ideas about how information is processed. Consider the concept of association. Is it sufficient to say that a stimulus becomes associated with a response, or that a stimulus elicits a response? During the Behaviorist era, the only kind of mechanism that was thought necessary was that of a telephone switchboard (although strangely enough, there was no telephone operator). But as soon as we begin to consider how we might program a computer to form the same kinds of associations, we begin to realize that the term “association” is just a cover term for a very complex process. The concept of parallel activation, originally proposed by Morton (1970) and subsequently taken up by McClelland and Rumelhart (1981), provided an actual mechanism that would automatically select the stored representation that most closely matched the input. Essentially, these systems compare the input with every word in the mental lexicon simultaneously. A key assumption in these models is the notion of incremental activation – the matching word unit does not reach the maximum level of activation in one cycle, but rather over many cycles. The motivation for this assumption is that it allows the model to explain why some words are recognized faster than others, and it gives time for processes such as competition and feedback to influence the rate at which activation increases. The alternative to a completely parallel system is a serial search model (e.g., Becker, 1979; Forster, 1976; Paap et al., 1982). The difference is that comparisons are not made simultaneously with every word in the lexicon, but instead with a limited subset of words. At one extreme, the comparison is carried out one word at a time (e.g., Forster, 1976), but less extreme versions are possible. For example, it could be that the lexicon is divided up into a number of partitions, such that only the words within a given partition are compared with the input simultaneously. In either case, a serial search is involved – either word by word, or partition by partition. The chief advantage of this approach is that it provides a very natural

Visual Word Recognition

41

account of the frequency effect. In order to optimize the search process, the words that are needed most often are placed at the beginning of the search, and the words that are seldom needed are placed at the end. This leads to the prediction that access time should be a function not of absolute frequency, but rather relative frequency. More specifically, access time should be a linear function of the rank of a word in a frequency-ordered list. This prediction appears to be correct (Murray and Forster, 2004). Trying to devise an experimental test of whether a search is involved is not easy. However, there is one approach that looks promising. Parallel models such as DRC (Coltheart et al., 2001) and MROM (Grainger and Jacobs, 1996) include a special unit which sums activation across all word units. One purpose of this global activation index is to decide how long to wait before responding “No” in a lexical decision experiment. If the index is relatively high (i.e., there is great deal of activity in the network), then the decision is delayed in case the input turns out to be a low-frequency word. But if it is low (and no word unit has reached criterion), then the decision is not delayed any further. Another purpose for this index is to permit “fast guesses”. Because the global index is strongly influenced by the presence of a word unit that matches the input, it is possible to make better than chance decisions about the status of the input without waiting to see whether any word unit reaches criterion. So what happens in a lexical decision experiment if we impose a response deadline that is considerably shorter than the subject’s average reaction time? The error rate increases sharply, but there is a clear frequency effect. More accurate decisions are made for highfrequency words than for low-frequency words. This creates a problem for a search model. If a matching entry has not been found, then how could decisions ever be anything but random? No such problem exists for a parallel model with a global activation index, since the global activation curves for high-frequency words, low-frequency words, and nonwords all diverge from each other very early, and would therefore provide a basis for better than chance accuracy. The above argument creates problems for a search model only if the deadline imposed by the experimenter is set so early that there is insufficient time for the search to identify even very high-frequency words. If this is not the case, then at least some high-frequency words will be accessed, and hence reasonably accurate lexical decisions for these words would be expected. So it turns out that both types of models can explain better than chance performance for high-frequency words. But what about low-frequency words and nonwords? According to the search model, one would expect chance performance on both, since there is no way to decide whether the input is more likely to be a word or a nonword. But for a parallel model, one would expect very low global activation levels for nonwords, and hence it should be possible to distinguish between nonwords and low-frequency words reasonably well. Preliminary experimental tests of these hypotheses show that there are three types of subjects (Forster, 2006). There are some subjects who perform better than chance on all types of items. These subjects are not diagnostic, because although they support the parallel model’s prediction, they do not necessarily disconfirm the serial model’s prediction, because

42

The Mental Lexicon: Core Perspectives

for these subjects, the deadline might have been set late enough to enable successful access for all but the lowest frequency words. Then there are subjects who are at chance for all items. These subjects also are not helpful. Finally, there is a sizeable group who are above chance on high-frequency words, but at chance on both low-frequency words and nonwords. This is exactly what the search model predicts should happen, but a parallel model has difficulty explaining why there is no difference in global activation for low-frequency words and nonwords. To get around this awkward fact, one could assume that the activation functions for high- and low-frequency words have slightly different parameters, so that it takes longer for low-frequency words to become distinguishable from nonwords. In other words, for this third group of subjects, the global activation index is insensitive to the presence of a matching lowfrequency word unit. Without explaining why this should be the case, this account is less than satisfactory. But an even greater problem is presented by the fact that the decisions made by this third group are strongly influenced by the number of neighbors. Items with many neighbors are classified as words regardless of their lexical status, and items with few or no neighbors are classified as nonwords regardless of lexical status. The reason that this is problematic is that neighbors make only a small contribution to the global activation index relative to the contribution of a matching word unit. So the problem for a parallel account is to explain how the global activation index could ever be insensitive to lexicality, but sensitive to the presence of neighbors. If a straightforward, non-arbitrary account can be found, this will enhance the status of parallel models considerably. In the meantime, the advantage appears to lie with search models.

5. THE FORM OF THE FREQUENCY EFFECT Relatively little attention has been paid to the question of the precise relation between word frequency and lexical decision time. It is usually assumed that the relation is logarithmic, but no reason is offered as to why it should be this way. The reason for making this assumption is that at the low-frequency end of the spectrum, small changes in frequency have very large impact on reaction time, but at the high-frequency end, large changes in frequency have negligible effects. Although the log function provides a reasonably good description of the data, it seriously underestimates reaction time in the high frequency range (Balota et al., 2004). In part, this is because it assumes that reaction time will continue to decrease as frequency increases, whereas it appears that reaction time approaches a lower limit. As mentioned earlier, a search model provides a principled explanation of the form of the frequency effect. If the search is frequency-ordered, and the search speed is uniform, then lexical decision time should be a linear function of the rank of a word in a frequency-ordered list. Predictions derived from this model are quite accurate, and rival the predictions based on the log frequency account (Murray and Forster, 2004). Further, the search account explains

Visual Word Recognition

43

why lexical decision times should approach a lower limit. This is not because it is impossible for subjects to respond any faster, but instead is because once a word reaches a rank of 1, no further increase in frequency can have any effect on access time. The rank function, however, does not provide a better fit than the best-fitting power function. One might be inclined to accept this as a superior account on the grounds that a power function gives a good description of the effects of practice on the acquisition of a skill, and the amount of practice is obviously analogous to word frequency (e.g., Kirsner and Speelman, 1996). But two points need to be considered before adopting this view. First, it is not enough to say that a power function accounts for the data, one has to instead specify which power function accounts for the data, which requires the estimation of three parameters. Unless one has good theoretical reasons for choosing the parameter values of the function, one is simply engaged in a curve-fitting exercise. Indeed, using the parameters from one particular model that predicts a power function (Logan’s instance model; Logan, 1990) provides a rather poor fit to the data (Murray and Forster, 2004). This problem of parameter estimation does not arise to the same degree in the rank account. The essence of this account lies in the fact that it predicts that the slope of reaction time must be linear with rank. The actual slope itself is not critical, nor is the intercept. These are parameters that are determined by characteristics of the individual subjects, such as their search rate and decision-making speed. So in a sense, the rank function is parameter free. The second thing to consider is that a power function will provide an excellent account of any set of data in which improvement in performance over time is a constant proportion of the amount of possible improvement remaining (given by the lower bound on performance). This seems like a very sensible limitation, but is not particularly informative or interesting. Recently, Adelman et al. (2006) show that contextual diversity is a better predictor of lexical decision time than word frequency. This measure is derived from word frequency norms, but refers to the number of different texts that the word was found in, not the total number of times the word was encountered. This measure may in fact be a more reliable estimate of frequency at the lower end of the scale, since some words may get a spuriously high value by virtue of occurring very often in a particular text that happened to be sampled, but occurring very infrequently in other texts. Adelman et al. argue that this cannot account for the better prediction of lexical decision time, but it should be pointed out that the relevant causal variable is the frequency with which the average individual encounters a word, not the total number of times the word appears over all texts. Words that occur with a relatively low frequency but in a wide range of texts must be more likely to be encountered by the average individual. Adelman et al. interpret their findings in terms of a rational analysis of memory, which asserts that words with high contextual diversity are more likely to be needed in any new context. This approach appears to mesh quite well with the assumptions of a search model of lexical access. Words that have a high need should be placed earlier in the search path than words with low need. Nevertheless, one might still reject a search model out of hand because a word-byword search would obviously be far too slow to explain the relatively fast “No” responses to

44

The Mental Lexicon: Core Perspectives

nonwords in a lexical decision experiment. In the original version of the “bin” model of lexical access (Forster, 1976), this problem was overcome by subdividing the lexicon into bins based on orthographic form, so that the input letter array could be converted by a hash coding function to an approximate address (a bin number), and if the array was not present in that bin, a “No” decision could be reached after searching just one bin. However, there is a major problem with this account, namely that form priming effects can only occur in this model if both the prime and the target are located in the same bin. Given that the prime can differ from the target in many different ways and still produce priming, it seems virtually impossible to design a hash code that will always assign the same bin number to both the prime and the target. As mentioned earlier, a search does not necessarily have to be carried out on a wordby-word basis. Instead, the lexicon could be multiply partitioned, and the search could be carried out partition-by-partition, with each partition being activated in parallel. Under these conditions, the entire lexicon can be searched in a reasonable time, which avoids the problems raised by form-priming. If the partitions are activated in a fixed sequence, and the words that are needed most often are allocated to the first partition, and the words that are needed least often are allocated to the last partition, then it seems possible that the relevance of contextual diversity to lexical access could be explained.

6. HOW ARE SEMANTIC DECISIONS MADE? Recently there has been increased interest in using semantic categorization tasks, where the subject must classify a word according to whether it refers to an object that falls into a certain semantic category (e.g., is it an animal?). The advantages of this task are several. For one thing, some researchers are concerned that it may be possible to make a lexical decision without accessing the meaning, or without even uniquely identifying the word. This is obviously not the case with a semantic task. Another advantage is that it shares a common component with lexical decision (lexical access), but may not be affected by other variables that could affect lexical decision. For example, semantic decisions are unlikely to be affected by perceived familiarity, which may explain why there is no long-term repetition priming when the same words are presented first in a lexical decision task and then a second time in a semantic categorization task (Vriezen et al., 1995). Taft and van Graan (1998) found that irregular words did not take longer to classify than regular words in a semantic task, despite the fact that the same words differed in a naming task, suggesting that phonological activation was not a prerequisite for access to meaning. In similar vein, Forster and Shen (1996) used a semantic task to determine whether the effect of neighborhood size on lexical decision was really an access effect, but found no such effect (but see Sears et al., 1999, for a different result). In these investigations, a semantic task is used essentially as a control, the aim being to check whether an effect found

Visual Word Recognition

45

in another task is a true access effect. The previously mentioned findings of Vriezen et al. (1995) fits into this category. If long-term repetition priming was the result of faster lexical access on the second presentation, then an effect should have been observed when the repeated word was initially presented in a lexical decision task and subsequently in a semantic categorization task. But there are also cases where a semantic task reveals effects that were not apparent in a lexical decision task. For example, Bueno and Frenck-Mestre (2002) report strong semantic priming effects in a masked priming experiment using a semantic task. Such effects are not readily obtained with lexical decision (Perea and Gotor, 1997). Similarly, although there is no masked cross-language translation priming effect from the second language to the first when the task is lexical decision, a strong effect is observed if a semantic task is used (Finkbeiner et al., 2004). One important qualification here is that the effect occurs for exemplars of the category, not for nonexemplars, suggesting that the category itself provides some additional constraint. To understand the implications of results such as these, we need to develop a model of how the decisions are made. The most obvious candidate is that after lexical access is completed, and the semantic properties of the word have been extracted, these properties are then used to decide category membership. This might be a relatively trivial exercise (e.g., deciding that walrus is the name of an animal, or that fusion is not), or it may be more complex (e.g., deciding that a walrus is bigger than a brick). An alternative proposal is to adopt a feature monitoring approach, in which the decision system selects one or more semantic feature units to monitor, so that a “Yes” decision is triggered as soon as activation in those units exceeds some specified value (e.g., Carreiras et al., 1997). However, if none of these units reaches criterion before a deadline expires, a “No” decision is made. This procedure for making a “No” decision predicts no frequency effect for nonexemplars, since the same deadline applies for all words. However, it has been shown that high-frequency words are rejected faster than low-frequency words (Monsell et al., 1989; Forster and Shen, 1996), which implies that the decision system also monitors when access is completed. Another interesting problem is presented by the fact that the number of neighbors has no effect on semantic categorization times unless one of the neighbors happens to be an exemplar (Forster and Hector, 2002; Rodd, 2004). So if the category was animal, then a nonword such as turple would take longer to reject than a nonword such as cishop. But if the category was changed to profession, then cishop would take longer than turple. The puzzle here is to specify how the semantic properties of the neighbors can be established without testing each neighbor separately, which would of course take time. Feature monitoring provides a neat solution, because there is no need to consider the neighbors separately. In an animal categorization task, a nonword such as turple will activate the “animalness” feature to some degree, and hence force a delayed decision. This will not happen with cishop, and the fact that this even has a neighbor may never be registered. Thus, a feature monitoring approach seems like a good solution to the turple problem, provided that an appropriate feature is available to monitor, and that the presence of a frequency effect for nonexemplars can be explained.

46

The Mental Lexicon: Core Perspectives

7. DESIGNING EXPERIMENTS The traditional approach to experimentation in the field has been to find two or more sets of words that differ on some variable of interest, and then see whether they differ on some measure of access time, holding all other relevant variables constant. There now seems to be fairly wide agreement that this strategy is in danger of running aground. There appear to be several reasons for this, the first and foremost being that it is very difficult (if not impossible) to find words that differ on one variable, but are equated on all other variables. For one thing, the investigator may not know what the relevant variables are. For example, one variable that evidently accounts for some variance is the ratio of written to spoken word frequency (Baayen et al., 2006). That this is a possibly relevant variable was only recently discovered, and hence it is unlikely that any previous investigations took this variable into account. Further, as any researcher knows all too well, reviewers of articles are sometimes able to find a variable that hasn’t been controlled , and this can lead to a rejection, even if the relevance of that variable is unknown. In approaching this problem, we need to distinguish between systematic confounds and accidental confounds. A systematic confound arises when there is some kind of causal connection between the two variables. For example, age of acquisition and frequency are systematically confounded because children tend to learn high frequency words first. An accidental confound occurs when the two sets of words under investigation happen to differ on both the independent variable and another variable, although there is little or no connection between them. An example of such a confound would be comparing high and low frequency words that happened to differ on the average number of neighbors. One way to cope with accidental confounds is to use very large (random) samples of words at each of several different levels of the independent variable. If the means on the dependent variable vary in a systematic way as a function of the independent variable, then an accidental confound is very unlikely, although a systematic confound might still be present. Attempting to eliminate a systematic confound by judicious item selection (e.g., by finding high-frequency words that are acquired late) is fraught with danger. The words that are eventually chosen are obviously not representative of the language, and are likely to have special properties. Furthermore, it might be the case that the words would no longer be matched if an independent set of norms were used. So, for example, an apparent late-acquired high-frequency word might result from errors of measurement – either the frequency or the age of acquisition may have been overestimated, or both. It is now being argued that multiple regression is the only way to get around these problems (Balota et al., 2004). The recommended procedure is to take a large number of words that have been measured on a large set of predictor variables, and to then apply multiple regression techniques. Baayen et al. (2006) point out a number of statistical problems with this approach, but nevertheless agree that this is the correct solution. What we discover in such an experiment is that a dependent variable such as lexical decision time is

Visual Word Recognition

47

correlated with just about every predictor variable, and that the predictors are themselves correlated. The task then is to determine whether a given variable exerts an independent effect on lexical decision time – an effect that cannot be explained by the correlation of the variable with some other variable that is also correlated with lexical decision time. An issue that is seldom discussed is whether multiple regression actually gives us the right information. This technique is principally designed to provide the most efficient method of predicting scores on some dependent variable. If two variables both correlate with the dependent variable to about the same degree, and they themselves are highly correlated, then multiple regression tells us that there is no point including both predictors. One of them can be discarded, and in practical terms, it doesn’t much matter which is discarded. But if we were trying to decide whether it was frequency or age of acquisition or both that affected lexical decision time, then it would matter a great deal which variable was discarded. Usually, this would be determined by which variable had the higher correlation with decision time, and that in turn might be controlled simply by the reliability of measurement. But there is a deeper problem. Consider the following partial correlation problem. We observe that males are on average able to run faster than females, so sex is correlated with running speed. But sex is also correlated with certain physiological variables, which themselves correlate with running speed. So the question is this: Would women be able to run as fast as men if the physiological variables were held constant? The answer is provided by the formula for the partial correlation of sex with speed, holding physiology constant. But to my mind, this is a purely hypothetical question, with a purely hypothetical answer, roughly akin to asking how many angels could dance on the head of a pin. Applied to the frequency vs. age of acquisition issue in word recognition, it seems equivalent to asking whether frequency would still affect lexical decision time if children did not tend to learn high frequency words first. Perhaps what we should say instead is that the value of the partial correlation coefficient tells us whether such a state of affairs is at least possible. That is, if the regression results show that the correlation of frequency with lexical decision time can be accounted for in terms of its correlation with age of acquisition, then in a world where children did not learn high frequency words first, there should be no correlation at all between frequency and lexical decision time. I find this kind of conclusion very unsatisfying. Could it be that our problem is that we are not actually doing enough experiments? Astronomers cannot do experiments on things like galaxies, so they are forced to make inferences from observations of the properties of galaxies. In a way, that is analogous to our approach to visual word recognition. We gather various observations of the properties of words (e.g., how fast they can be recognized, how many neighbors they have, etc.), and then we try to understand how these properties relate to each other. A more experimental approach would involve assigning words at random to one of two treatment conditions, and then applying a treatment, and observing its effects. The same approach applied to the frequency and age-of-acquisition issue might involve assigning high- and low-frequency words at random to two different conditions. In the treatment condition, a special group of children would be trained daily on both sets of words, and in the control condition, the same children

48

The Mental Lexicon: Core Perspectives

would receive no special training on the words assigned to this condition. After several years, the training would be stopped, and then when the children reached college age, they would be tested on these words in a lexical decision experiment. Obviously, this experiment would be impractical, not to mention unethical. So what kind of experiment can be used? Priming experiments are a good example, especially so because we hold the items constant and vary the primes. Similarly, experiments that hold the items constant while the task is varied can be used. So, for example, if the effect of neighborhood size is an access effect, then it ought to be present not only in lexical decision, but also in a semantic categorization task. Of course, this strategy is most effective when the two tasks share just the pre-access and access components, but have completely different post-access systems. Hopefully, they will also have different sensitivities to nuisance variables. As an example, consider the claim that masked repetition priming is really due to a change in the perceived familiarity of the target word. Since words are more familiar than nonwords, this increased familiarity would bias the subject towards a “Yes” decision in a lexical decision task, leading to faster responses for primed than unprimed words. However, familiarity is irrelevant in a semantic task. The fact that a word is perceived as being familiar does not bear on whether it refers to an animal or not. Thus, the presence of robust masked repetition priming in a semantic categorization task provides a strong rebuttal to this claim (e.g., Forster, 2004).

8. STRATEGIES AND MASKED PRIMING Finally, a few brief remarks about the role played by strategies in priming experiments. It would be helpful if we could give a reasonable definition of a strategy, but this is not easy. One might suggest that a strategy is any procedure consciously devised by the subject that takes advantage of the structure of the experimental materials and leads to changes in performance. The problem here is the emphasis on “conscious”. Jo Hector and I once ran a lexical decision experiment that used the same materials as in a previous masked priming experiment, only now the prime was perfectly visible. However, we forgot that there was a perfect correlation between the lexical status of the prime and the target. If the prime was a word, the target was a word. If the prime was a nonword, then the target was a nonword. Not surprisingly, many subjects (but not all) had mean reaction times that were ridiculously fast. Unfortunately, we didn’t become aware of this until the end of the day’s testing, and hence there was no opportunity to question the subjects more carefully. However, none of the subjects had offered any comments about the strange design of the items, and hence it is conceivable that they did not explicitly recognize the correlation and its significance for their decision. If this was indeed the case, would it mean that a strategy was not involved? Perhaps this is not a question that has to be answered. What we can say instead is that the subject’s responses were not influenced by the intended variable (which was orthographic overlap

Visual Word Recognition

49

between prime and target), but were instead influenced by the correlation between the lexical status of the prime and target. This could easily be shown by repeating the experiment with the correlation removed. So this may be a case of implicit learning. First, the correlation between the input events is detected, and this is then used to predict the correct response. Whether we call this a strategy or not is actually beside the point – what is critical is that the experimental design introduced unwanted and irrelevant influences. Recently, Bodner and Masson (2001) have raised the possibility that something similar might be happening even in masked priming. What they have reported is that masked repetition priming effects are stronger when there is a high proportion of trials in which the masked prime is the same as the target. They refer to this phenomenon as a prime validity effect, implying that in some (unspecified) fashion, information extracted from the prime is recruited to assist in the processing of the target. This could be taken as compatible with an activation account of priming, except that this would not explain the repetition proportion effect. Rather, Bodner and Masson want to argue that subjects are sensitive to the correlation between prime and target, and learn to take advantage of it, despite the fact that they are unaware of the prime’s existence. Obviously, this would be a remarkable and exciting finding, if it can be confirmed. But the down side is that Bodner and Mason see this as a demonstration that masked priming does not eliminate strategic influences, which in turn threatens to undermine the utility of this procedure for investigating the properties of the lexical processor. We have begun a series of experiments to examine this repetition proportion effect (the first results were reported in Kinoshita et al., 2004), and it is too early yet to reach firm conclusions. But we can at least take heart from two observations, one informal and the other formal. The informal fact is that when we ran the experiment with the perfect correlation between prime and target again with masked primes, the lexical decision times were perfectly normal. Either these subjects were unable to detect the correlation, or they were unable to take advantage of it. The second finding was reported in Forster (1998). This dealt with the effect of expectancies in a form-priming experiment. The related primes were word fragments, and the target was a predictable completion of the fragment (e.g., terrifTERRIFIC). The unrelated prime was also a word fragment, but unrelated to the target (e.g., pyram-TERRIFIC). The critical manipulation was whether the prime was masked or visible. As one might have expected, priming was much stronger in the visible prime condition, but not because of any difference in the related condition. Instead, it was the unrelated baseline condition that showed the difference. Responses were far slower when the prime was visible. This makes sense if we consider that seeing the prime pyram sets up a strong expectancy for PYRAMID, which leads to a slower than normal response when the target is an unexpected word like TERRIFIC. This is exactly the same as what happened in the sentence context experiments discussed earlier (Forster, 1981). But no such effect occurred when the prime was masked, suggesting that expectancies did not play any role in that condition. Of course, there was priming, but this can be explained purely in terms of the orthographic overlap

50

The Mental Lexicon: Core Perspectives

between terrif and TERRIFIC. That is, the masked presentation of terrif activated the lexical representation of terrific, but did not lead to the formation of an expectancy. This is at least consistent with the assumption that masked priming reflects processing in the lexical processor itself, and minimizes contributions from higher cognitive centers.

9. THE FUTURE Looking ahead, it seems clear that more attention needs to be given to providing detailed descriptions of actual processing mechanisms. In this regard, the development of computational models is a move in the right direction. However, this should be coupled with increased effort at testing core assumptions (e.g., incremental activation), rather than allowing these assumptions to settle into dogma. In addition, theorists need to consider the explanatory adequacy of their models more carefully. This will involve constructing a system of principles that will guide theorizing along certain lines. Most importantly, we desperately need to reach consensus about basic experimental findings. For example, it should not be so difficult to reach agreement about whether neighborhood size has a facilitatory, inhibitory, or no effect on lexical decision time. Yet a review of the findings is somewhat discouraging (Andrews, 1997). Related to the above point, journal policy needs to be modified so that serious attempts at replication which fail can nevertheless be accepted for publication. Currently, researchers tend to rely on word-of-mouth testimony about the reliability of certain findings. Finally, a new statistical procedure is required that assesses the generality of an effect across items (and for that matter, across subjects as well). Current procedures deal only with the means of distributions, and a significant effect in an item analysis can easily be generated by just a small subset of the items. If measures of generality were also included, it might turn out that much of the disagreement about experimental findings would eventually disappear.

Visual Word Recognition

51

REFERENCES Adelman, J. S., G. D. A. Brown and J. F. Quesada (2006). Contextual diversity not word frequency determines word naming and lexical decision times. Psychological Science, 17, 814-823. Andrews, S. (1997). The effect of orthographic similarity on lexical retrieval: Resolving neighborhood conflicts. Psychonomic Bulletin & Review, 4 (4), 439-461. Baayen, R. H., L. B. Feldman and R. Schreuder (2006). Morphological influences on the recognition of monosyllabic monomorphemic words. Journal of Memory and Language, 55, 290-313. Balota, D. A., M. J. Cortese, S. D. Sergent-Marshall, D. H. Spieler and M. Yap (2004). Visual word recognition of single-syllable words. Journal of Experimental Psychology: General, 133 (2), 283316. Becker, C. A. (1979). Semantic context and word frequency effects in visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 5, 252–259. Bodner, G. E. and M. E. J. Masson (2001). Prime validity affects masked repetition priming: Evidence for an episodic resource account of priming. Journal of Memory & Language, 45 (4), 616-647. Bueno, S. and C. Frenck-Mestre (2002). Rapid activation of the lexicon: A further investigation with behavioral and computational results. Brain & Language Special Issue: Mental lexicon II, 81 (1-3), 120-130 Carreiras, M., M. Perea and J. Grainger (1997). Effects of orthographic neighborhood in visual word recognition: Cross-task comparisons. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 857-871. Coltheart, M., K. Rastle, C. Perry, R. Langdon and J. Ziegler (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108 (1), 204-256. Ehrlich, S. F. and K. Rayner (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning & Verbal Behavior, 20, 641-655. Finkbeiner, M., K. Forster, J. Nicol and K. Nakamura (2004). The role of polysemy in masked semantic and translation priming. Journal of Memory & Language, 51 (1), 1-22. Fischler, I. and P. A. Bloom (1979). Automatic and attentional processes in the effects of sentence contexts on word recognition. Journal of Verbal Learning and Verbal Behavior, 18, 1-20. Fodor, J. A. (1983). The Modularity of Mind. MIT Press/Bradford Books, Cambridge, Mass. Forster, K. I. (1976). Accessing the mental lexicon. In: New approaches to language mechanisms (R. J. Wales and E. Walker, eds.), pp. 257–287. North-Holland, Amsterdam. Forster, K. I. (1981). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 33A (4), 465-495. Forster, K. I. (1998). The pros and cons of masked priming. Journal of Psycholinguistic Research, 27 (2), 203-233. Forster, K. I. (2004). Category size effects revisited: Frequency and masked priming effects in semantic categorization. Brain & Language, 90 (1), 276-286. Forster, K. I. and J. Hector (2002). Cascaded versus noncascaded models of lexical and semantic processing: The turple effect. Memory & Cognition, 30 (7), 1106-1116.

52

The Mental Lexicon: Core Perspectives

Forster, K. I. (2006). Five challenges for activation models. In: From Inkmarks to Ideas: Current Issues in Lexical Processing (S. Andrews, ed.), pp. 95-121. Hove: Psychology Press. Forster, K. I. and D. Shen (1996). No enemies in the neighborhood: Absence of inhibitory neighborhood effects in lexical decision and semantic categorization. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 696-713. Forster, K. I. and T. Hannagan (in preparation). PSM: A Parallel Search Model of Lexical Access. Grainger, J. and A. M. Jacobs (1996). Orthographic processing in visual word recognition: A multiple read-out model. Psychological Review, 103 (3), 518-565. Kinoshita, S., K. I. Forster and M. Mozer (2004). A non-magical account of the prime validity effect. Paper presented at the 45th Annual Meeting of the Psychonomic Society, Minneapolis. Kirsner, K. and C. Speelman (1996). Skill acquisition and repetition priming: One principle, many processes? Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 563–575. Logan, G. D. (1990). Repetition priming and automaticity: Common underlying mechanisms. Cognitive Psychology, 22, 1–35. Lau, E., C. Stroud, S. Plesch and C. Phillips (2006). The role of structural prediction in rapid syntactic analysis. Brain and Language, 98 (1), 74-88. McClelland, J. L. and D. E. Rumelhart (1981). An interactive activation model of context effects in letter perception: I. An account of basic findings. Psychological Review, 88, 375-407. Monsell, S., C. Doyle and P. Haggard (1989). Effects of frequency on visual word recognition tasks: Where are they? Journal of Experimental Psychology: General, 118, 43-71. Morton, J. (1970). A functional model of human memory. In: Models of human memory (D. A. Norman, ed.), pp. 203–254. Academic Press, New York. Murray, W. S. and K. I. Forster (2004). Serial Mechanisms in Lexical Access: The Rank Hypothesis. Psychological Review, 111 (3), 721-756. Norris, D. (2006). The Bayesian Reader: Explaining word recognition as an optimal Bayesian decision process. Psychological Review, 113, 327–357. Paap, K. R., S. L. Newsome, J. E. McDonald and R. W. Schvaneveldt (1982). An activation verification model for letter and word recognition: The word superiority effect. Psychological Review, 89, 573–594. Perea, M., and A. Gotor (1997). Associative and semantic priming effects occur at very short stimulus-onset asynchronies in lexical decision and naming. Cognition, 62 (2), 223-240. Plaut, D. C. (1997). Structure and function in the lexical system: Insights from distributed models of word reading and lexical decision. Language & Cognitive Processes, 12 (5-6), 765-805. Rodd, J. M. (2004). When do leotards get their spots? Semantic activation of lexical neighbors in visual word recognition. Psychonomic Bulletin & Review, 11 (3), 434-439. Sears, C., S. Lupker and Y. Hino (1999). Orthographic neighborhood effects in perceptual identification and semantic categorization tasks: A test of the Multiple Read-out Model. Perception & Psychophysics, 61, 1537-1554. Seidenberg, M. S. and J. L. McClelland (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96 (4), 523-568. Taft, M. and F. van Graan (1998). Lack of phonological mediation in a semantic categorization task. Journal of Memory and Language, 38 (2), 203-224.

Visual Word Recognition

53

Van Berkum, J. J. A., C. M. Brown, P. Zwitserlood, V. Kooijman and P. Hagoort (2005). Anticipating upcoming words in discourse: Evidence from ERPs and reading times. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31 (3), 443-467. Van den Brink, D. I., C. M. Brown and P. Hagoort (2006). The cascaded nature of lexical selection and integration in auditory sentence processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32 (2), 364-372. Vriezen, E. R., M. Moscovitch and S. A. Bellos (1995). Priming effects in semantic classification tasks. Journal of Experimental Psychology: Learning, Memory, & Cognition, 21 (4), 933-946.

This page intentionally left blank

4 LANGUAGE: BETWEEN WORDS AND GRAMMAR∗

Mark Aronoff, SUNY at Stony Brook, Stony Brook, USA

Linguists have followed two ways in the study of words. One seeks to accommodate the word, the other to obliterate it. I will defend here an approach to language that respects the autonomy of arbitrary individual words or lexemes and privileges the interaction between the idiosyncrasies of lexemes and a highly regular linguistic system. Putting this approach in the context of the history of our field, I will show that “Chaque mot a son histoire [Each word has its [own] history]” (Gilliéron; see Malkiel, 1967 p. 137) in “un système où tout se tient [a system where everything hangs together]” (Meillet 1903, p. 407), but instead of phonology, which was Gilliéron’s main concern, I will concentrate on syntax, semantics, and morphology. I am not claiming that the word is the only idiosyncratic element in language, but that it is the central such element.

1. COMPOSITIONALITY All linguists assume that natural languages are compositional, which we understand informally to mean that the meaning of a complex syntactic expression is determined by its structure and the meanings of its constituents. The main argument for compositionality of natural language is rooted in syntactic productivity and was stated succinctly by Frege: “The possibility of our understanding propositions which we have never heard before rests evidently on this, that we construct the sense of a proposition out of parts that correspond to the words” (Frege, 1914, p. 79).

∗

This article is derived from my Presidential address, delivered at the annual meeting of the LSA in January 2007. Much of the content presented here is the result of joint work with Frank Anshen, Robert Hoberman, Irit Meir, Carol Padden, Wendy Sandler, and Zheng Xu. This research was supported in part by the US-Israel Binational Science Foundation, the National Institutes of Health, and the Stony Brook Provost’s Office.

55

56

The Mental Lexicon: Core Perspectives

Exactly what these “parts that correspond to words” are in natural languages has been debated since the beginning of linguistics (cf. the Ancient Greek grammatical term meros logou “piece/part of speech”). Saussure (1915/1975) called them (simple) signs but he was not entirely clear about what linguistic units he took to be simple signs, though for the most part his examples are what we would call lexemes (Carstairs-McCarthy, 2005).1 Most linguists, though, following Baudouin de Courtenay (1972), believe that the “parts that correspond to words” in natural languages are not words but rather morphemes, the parts that make up complex words. A term like antidecompositionalist “opposed to the view that the meanings of words can be broken down into parts”, for example, can be analyzed into [[anti [[[de [com=pos-it]V]V ion]N al]A]A ist]A.2 The standard view is that the meanings of this word and others like it are compositionally derived from the meanings of these constituent morphemes, making morphemes, not words, the “parts that correspond to words” in natural languages. A bit confusing, but linguists like exoticizing the ordinary. I am antidecompositionalist by temperament, because I believe that what happens inside lexemes is qualitatively different from what happens outside them. My adherence to this classic version of the lexicalist hypothesis (Chomsky, 1970) is grounded in a simple love of words. As I wrote in my first publication (Aronoff, 1972), on the semantics of the English word growth and its relevance to the lexicalist hypothesis, “all a person had to do [in order to notice the peculiar differences in meaning between the verb grow and the noun growth] was to look the words up in a reputable dictionary, or think for a long time: grow – growth” (p. 161). Yet very few modern linguists either look words up in a reputable dictionary or think about them for a long time. Philosophers do, but linguists don’t often think much about the subtleties of word meanings, which is why most linguists are not antidecompositionalists. An important and often forgotten point about antidecompositionalism is that it is an empirical claim, not a theoretical hypothesis. Still, we antidecompositionalists do have theoretical obligations that our opponents do not, because we must account for the internal properties of lexemes in a different way from how we account for syntax (including inflection). That is what I have been trying to do for the last thirty years. The rest of my article is divided into three parts. First, I will discuss a language that appears to be completely compositional down to its smallest pieces, Al-Sayyid Bedouin Sign Language (ABSL). I will show that ABSL is an instantiation of Saussure’s picture of language, with little if any structure below the level of the lexeme. ABSL thus provides an unusual type of evidence for both compositionality and the lexicalist hypothesis, showing that a language can emerge on its own very quickly in which words are indeed “the parts that correspond to words”.

1

Psycholinguists, starting with Levelt (1989), use the term lemma where linguists have long used lexeme, for the abstract lexical entry, and use lexeme to designate either a phonological word or a grammatical word. I apologize for any confusion that the different usage may cause. 2 The form pos-it is the allomorph of the root pose that appears before the suffix -ion.

Language: Between Words and Grammar

57

I will then turn to well-known phenomena from English and present evidence against one common understanding of the special nature of words, that they diverge from compositionality only because they are stored in memory. I will review old data and present one new example to show that even productively coined new words can be noncompositional. I will close with some traditional linguistic analysis and show that Semitic roots, the original lexeme-internal linguistic units, have robust properties that bear no relation to meaning and little to phonology. Taken together, these three sections point to a particular picture of the role of the word in language.

2. AL-SAYYID BEDOUIN SIGN LANGUAGE For several years, I have been privileged to be a member of a team, along with Wendy Sandler, Irit Meir, and Carol Padden, working on Al-Sayyid Bedouin Sign Language (ABSL). ABSL has arisen in the last seventy years in an isolated endogamous community with a high incidence of nonsyndromic genetically recessive profound prelingual neurosensory deafness (Scott et al., 1995). What distinguishes ABSL from all other welldocumented new languages are the circumstances of its creation and use, which show neither discontinuity of social structure nor the influence of other languages. In the space of one generation from its inception, systematic word order has emerged in the language. This emergence cannot be attributed to influence from other languages, since the particular word orders that appear in ABSL differ from those found both in the ambient spoken languages in the community and in other sign languages in the area. The Al-Sayyid Bedouin group was founded almost two hundred years ago in the Negev region of present-day Israel. The group is now in its seventh generation and contains about 3500 members, all residing together in a single community exclusive of others. Consanguineous marriage has been the norm in the group since its third generation. Such marriage patterns are common in the area and lead to very strong group-internal bonds and group-external exclusion. Within the past three generations, over 100 individuals with congenital deafness have been born into the community, all of them descendants of two of the founders’ five sons. Thus, the time at which the language originated and the number of generations through which it has passed can be pinpointed. All deaf individuals show profound prelingual neurosensory hearing loss at all frequencies, have an otherwise normal phenotype, and are of normal intelligence. Scott et al. (1995) identify the deafness as (recessive) DFNB1 and show that it has a locus on chromosome 13q12 similar to the locus of several other forms of nonsyndromic deafness. The deaf members of the community are fully integrated into its social structure and are not shunned or stigmatized (Kisch, 2004). Both male and female deaf members of the community marry, always to hearing individuals. The deaf members of the community and a

58

The Mental Lexicon: Core Perspectives

significant fraction of its hearing members communicate by means of a sign language. Siblings and children of deaf individuals, and other members of a household (which may include several wives and their children) often become fluent signers. Members of the community generally recognize the sign language as a second language of the village. Hearing people in the village routinely assess their own proficiency, praising those with greater facility in the language. Those who have any familiarity with Israeli Sign Language, including those who have attended schools for the deaf outside the village, recognize that the two sign languages are distinct. Nor do Al-Sayyid signers readily understand the Jordanian sign language used in simultaneous interpreting on Jordanian television programs received in the area. Many of the signers in this community are hearing, a highly unusual linguistic situation, but one that is predicted to arise as a consequence of recessive deafness in a closed community (Lane et al., 2000). One result of the recessiveness is that there are a proportionately large number of deaf individuals distributed throughout the community (over 4 percent). This means that more hearing members of the community have daily contact with deaf members, and consequently signing is not restricted to deaf people. Furthermore, each new generation of signers is born into a native-like environment with numerous adult models of the language available to them. ABSL thus presents a unique opportunity to study a new language that has grown inside a stable community without obvious external influence. We have identified three generations of signers. The first generation in which deafness appeared in the community (the fifth since the founding of the community) included fewer than ten deaf individuals, all of whom are deceased. Information on their language is limited to reports that they did sign and one very short videotape record of one of these individuals. I report here only on the language of the second generation. The most noticeable structural feature of ABSL is the strictness of its word order. As reported in Sandler et al. (2005), we found a very robust Subject Object Verb (SOV) order within sentences and a similarly strict order of modifier elements within phrases (adjectives, negatives, and numerals) relative to their head nouns and verbs. In almost all instances in our data, the modifier follows its head. These word order regularities cannot be attributed to the ambient spoken language. The basic word order in the spoken Arabic dialect of the hearing members of the community, as well as in Hebrew, is Subject Verb Object (SVO). This generation of signers had little or no contact with Israeli Sign Language, whose word order appears to vary more widely in any case. Nor can the Head-Modifier order be ascribed to the ambient colloquial Arabic dialect of the community. In this dialect, and in Semitic languages generally, although adjectives do follow nouns, numerals precede nouns; and negative markers only rarely follow their heads. Hence the robust word-order pattern exhibited by the data is all the more striking, since it cannot be attributed to the influence of other languages; rather, this pattern should be regarded as an independent development within the language. This remarkable structural clarity breaks down entirely when we look inside words, where we find very little structure, either morphological or phonological. The language does have a fair number of compounds, but compounds are not morphological in the strict sense,

Language: Between Words and Grammar

59

since they are comprised of words. Sign languages are known for their robust inflectional morphology, especially their complex verb agreement systems, which arise very quickly in the history of individual languages. Our team was therefore surprised to discover that ABSL has no agreement morphology, indeed no inflectional morphology at all. Nor does it appear to have a phonological system of the kind we are familiar with in other sign languages. Phonology is an instantiation of what Hockett (1960) calls duality of patterning, or what others have called double articulation (Martinet, 1957). When we say that a sign language has phonology, we are saying that it has a set of discrete meaningless contrastive elements (handshapes, locations, and movements) that form their own system, with constraints on their combination, comprising a second articulation independent of the first articulation of meaningful elements. This system of contrastive elements makes up the signifiés of signs. It was the existence of such a system in American Sign Language that first led William Stokoe to proclaim its legitimacy as a language (Stokoe, 1960). ABSL appears not to have a system of discrete meaningless elements within words. Instead, each word has a conventionalized form, with tokens roughly organized around a prototype, but no internal structure. The variation found among signers in many words in our data lies precisely along the parameters that create contrasts in other sign languages. ABSL has been able to develop into a full-fledged linguistic system without benefit of phonology because of the visual medium of signing, which has many more dimensions than sound does and which allows for direct iconicity (Aronoff et al., 2005). As Hockett notes (1960): There is excellent reason to believe that duality of patterning was the last property to be developed, because one can find little if any reason why a communicative system should have this property unless it is highly complicated. If a vocal-auditory system [italics added] comes to have a larger and larger number of distinct meaningful elements, those elements inevitably come to be more and more similar to one another in sound. There is a practical limit, for any species or any machine, to the number of distinct stimuli that can be discriminated, especially when the discriminations typically have to be made in noisy conditions.(p. 95) It may be that ABSL will develop phonology as it matures, perhaps simply as a function of the size of its vocabulary, as Hockett suggests. Indeed, our team is currently investigating questions of just this sort. For the moment, though, ABSL is a language with a compositional syntax, but very little word-internal structure of any kind. It is, from a certain perspective, a perfect language, comprised of unanalyzeable words arranged in sentences.

3. THE MEANINGS OF MORPHOLOGICALLY COMPLEX WORDS The tenet that every natural language is a unitary compositional system of sentences whose basic building blocks are morphemes has an intellectual sister in the notion that the meaning and structure of morphologically complex words can not merely be represented in terms of

60

The Mental Lexicon: Core Perspectives

their sentential paraphrases but should be formally reduced to these paraphrases, so that the words themselves are reduced to sentences. To proponents of this method, words and sentences really are the same thing, because words are sentences in miniature, linguistic munchkins. This idea dates back at least to Panini and it has been popular among generativists almost from the beginning; not a surprise, since the most successful early generative analyses, notably those of Chomsky (1957), make no distinction between syntax and morphology, at least inflectional morphology.3 I will show that precisely those morphologically complex word types that have seemed most akin to sentences in their meanings, when studied in more detail, provide empirical support for just the opposite conclusion: morphologically complex words are not sentences and their meanings are arrived at in an entirely different fashion. Some of the evidence that I will review is over twenty-five years old and most of the field has successfully ignored it, so I am very much in the minority. Nonetheless, I will take advantage of the occasion to make a new plea for what, to my mind at least, have been very important findings. The founding generative work in the enterprise of reducing words to sentences is Lees (1960).4 The details of Lees’ analysis have been forgotten, but his idea has flourished on and off ever since, more on than off, largely unchanged except in notation. I offer two examples of Lees’ analysis. The first is what he calls action nominals. Here are two sample such nouns and their sentential sources: 1) The committee appoints John → John’s appointment by the committee 2) The committee objects to John → The committee’s objection to John These and others with different morphology are produced by means of a single action nominal transformation. The second example is nominal compounds. Here Lees discusses eight types, which include subject-predicate, subject-verb, verb-object, and others, all based on syntactic relations. His discussion of these types takes up fifty pages, so I can’t paraphrase it here, but a few samples are given below: 3) The plant assembles autos → auto assembly plant 4) The cap is white → white cap 5) The artist has a model → artist’s model 6) The sheep has a horn. The horn is like a prong → pronghorn 3

Syntactic Structures itself has a prescient short section (7.3) showing that the adjective interesting is not derived from a sentence containing the verb interest, for precisely the sorts of reasons that I will get to shortly. 4 One indicator of its popularity at the time is the fact that it was reprinted four times by 1966, the date of issue of my own copy, no mean feat when one takes into account the size of the community of linguists in the early 1960’s.

Language: Between Words and Grammar

61

Linguists are tempted to try to reduce the structures and meanings of complex words to those of sentences because they have a good set of tools for studying the syntax and semantics of sentences, which they apply to words. But words have much less apparent internal structure. This difference is both a blessing and a curse: we can transfer with ease what we know about the internal structure and meanings of sentences to words, but we have no simple way of finding out whether there is actually any isomorphism between the two domains. We may very well be deluding ourselves. A second problem is that, unlike sentences, words are not essentially ephemeral objects. As Paul Bloom has shown, individual people accumulate words at an astonishing rate throughout their lives (Bloom, 2000). They do not accumulate sentences. Because sentences are ephemeral, their meanings must be compositional sums of the meanings of their parts. This was Frege’s main argument for the compositionality of sentences in natural languages. By the same token, because words are not so ephemeral, and are more often than not retained in memory and society, complex words do not have to be compositional. Words that have been around for any time at all develop idiosyncrasies that are passed on to their new learners. There have been two lines of response to this obvious noncompositionality of some complex words. The more common response is to say that complex words are compositional at heart and that the departures from regularity accrete precisely because words are stored and used. The semantic differences between sentences and complex words are thus purely accidental on this view, the result of nonlinguistic or noncomputational factors. One example in support of the position that words aren’t born with noncompositional meaning but have it thrust upon them, is the trio of words cowboy, refugee, and skinner, which are curiously synonymous in one sense: “one of a band of loyalist guerillas and irregular cavalry that operated mostly in Westchester County, New York, during the American Revolution” (Webster’s Third New International Dictionary, 1987, henceforth WIII). Explaining why these three words came to be used to denote this particular group is a philological adventure of precisely the sort that gives linguists a bad taste for etymology, but it is obvious that that story has nothing to do with the compositional meaning of any of them. Instead, this peculiar meaning arose for all three words under certain very particular historical circumstances. The fact that this meaning of these words is not compositional is, linguistically speaking, an accident. The meaning itself is no more than a curiosity. This analysis allows us to preserve the intuition that words are compositional at birth, just like sentences, but that they lose their compositional meanings and acquire idiosyncratic meanings like this one just because they are stored and used. I followed this well-trodden road at first, firm in the faith that the meanings of productively-formed words would turn out to be compositional if we could ever catch the words at their moment of birth. But I found that road too easy and so I took the one less traveled by, and that has made all the difference. This less traveled road follows a Miesian method, taking the surface structure of complex words at face value and replacing the analogy to sentences with a very sparse non-sentential syntax that mirrors this simple

62

The Mental Lexicon: Core Perspectives

structure. The semantics of this sparse syntax for complex words is compositional, but only because it accounts for much less of the actual meaning of even newborn words than other semantic systems attempt to cover, leaving the rest to pragmatics. This bifurcating treatment rests on the assumption that words have complex meanings precisely because neither words nor their meanings are entirely linguistic objects, but rather the bastard offspring of language and the real or imagined world; it is this union of sparse linguistic resources with the vastness of the nonlinguistic universe that makes all words so rich from birth. Noncompositionality, on this account, is the product of the nature of words, not of their nurture. Of course, new complex words cannot be entirely arbitrary in meaning, because, except for Humpty Dumpy, we use them expecting our interlocutors to understand us.5 Some are more predictable than others, but new complex words are never entirely the product of language. As James McClelland has kindly pointed out to me (personal communication, January 7, 2006), the meanings of some words may be compositional, just in case the world that these words and their meanings intersect with is itself a logical world. A good example is chemical terminology. It is well known that this terminological system is compositional, but this is obviously because the system is itself entirely determined by the scientific principles of chemistry. Indeed, the name of any chemical compound is governed by the principles of the Interdivisional Committee on Terminology, Nomenclature and Symbols (ICTNS) of the International Union of Pure and Applied Chemistry. Not surprisingly, this terminology can yield very long words, the longest one attested in print, in English, being acetylseryltyrosylserylisoleucylthreonylserylprolylserylglutaminylphenylalanylvalylphenylalanylleucylserylserylvalyltryptophylalanylaspartylprolylisoleucylglutamylleucylleucylasparaginylvalylcysteinyl-threonylserylserylleucylglycylasparaginylglutaminylphenylalanylglutaminylthreonylglutaminylglutaminylalanylarginylthreonylthreonylglutaminylvalylglutaminylglutaminylphenylalanylserylglutaminylvalyltryptophyllysylprolylphenylalanylprol ylglutaminylserylthreonylvalylarginylphenylalanylprolylglycylaspartylvalyltyrosyllysylvalyltyrosylarginyltyrosylasparaginylalanylvalylleucylaspartylprolylleucylisoleucylthreonylalanyll eucylleucylglycylthreonylphenylalanylaspartylthreonylarginylasparaginylarginylisoleucylisoleucylglutamylvalylglutamylasparaginylglutaminylglutaminyserylprolylthreonylthreonylalanylglutamylthreonylleucylaspartylalanylthreonylarginylarginylvalylaspartylaspartylalanyl5

"There's glory for you!" "I don't know what you mean by 'glory,' " Alice said. Humpty Dumpty smiled contemptuously. "Of course you don't—till I tell you. I meant 'there's a nice knock-down argument for you!' " "But 'glory' doesn't mean 'a nice knock-down argument,' " Alice objected. "When I use a word," Humpty Dumpty said, in rather a scornful tone, "it means just what I choose it to mean—neither more nor less." "The question is, " said Alice, "whether you can make words mean so many different things." "The question is," said Humpty Dumpty. "which is to be master—that's all." Lewis Carroll, Through the Looking Glass

Language: Between Words and Grammar

63

threonylvalylalanylisoleucylarginylserylalanylasparaginylisoleucylasparaginylleucylvalylasparaginylglutamylleucylvalylarginylglycylthreonylglycylleucyltyrosylasparaginylglutaminylasparaginylthreonylphenylalanylglutamylserylmethionylserylglycylleucylvalyltryptophylthreonylserylalanylprolylalanylserine (a strain of tobacco mosaic virus). My point is that chemical terminology is compositional, not for linguistic reasons, but because the terminology mirrors chemical theory, which is a logical world. But when the world that the words reflect is not entirely logical, then neither are the meanings of the words. To show what motivates this bifurcating treatment of the meanings of words, I will return to one of our trio of Revolutionary words, skinner, the one that appears to have traveled furthest from its compositional roots to arrive at this peculiar meaning. I will not attempt to elucidate the meaning under consideration, whose origin is entirely opaque, but will look only at the meaning that seems to be most rooted in language and least tied to language-external context. We will see how quickly even that simple attempt comes to grief. A skinner is transparently “one who skins” and it is easy to derive the most common meaning of the noun skinner from that of the verb skin via just about any syntactic or semantic theory. So far, so good, but move one step further back to the meaning of the verb and the going gets tough. This verb skin is derived from the noun skin, whose initial [sk] cluster betrays its Old Norse origin and tells us it has been around for over a millennium. How are the two related semantically? Here are a few of the 26 senses of the verb listed in WIII. I have omitted all those that are apparent extensions of others: 7) a) To cover with or as if with skin b) To heal over with skin c) To strip, scrape, or rub off the skin, peel, rind, or other outer coverings of: to remove a surface layer from d) To remove (skin or other outer covering) from an object: pull or strip off e) To chip, cut, or damage the surface of <skinned his hand on the rough rock> f) To outdistance or defeat in a race or contest g) To equalize the thickness of adhesive on (a pasted or glued surface) by placing a sheet of wastepaper over it and rapidly rubbing or pressing h) To become covered with or as if by skin – usually used with over i) To climb or descend – used with up or down <skinned down inside ladders from the bridge deck –K. M. Dodgson> j) To pass with scant room to spare: traverse a narrow opening – used with through or by Rest assured I did not choose this word because of its many senses. Truthfully, I did not choose it at all, and the whole excursion that led to the curious incident of the

64

The Mental Lexicon: Core Perspectives

Westchester synonyms began quite by accident, during the course of a conversation about a group the country has since chosen to call (Katrina) evacuees. Such rich variety of senses is simply typical of zero-derived denominal verbs in English (henceforth zero-verbs). Clark and Clark (1979) provide a catalog of the major categories of sense types for zero-verbs, based on their paraphrases, which number in the dozens. A synopsis is given in Table 1. In the face of such a wealth of sense types, the syntactically-inclined explainer has no choice but to deny all the best Ockhamite tendencies and resort to a multiplicity of tree-types, each with the noun or root in a different configuration. Lees would have done so, as he did for compounds, and as have more recent scholars, notably Hale and Keyser (1993) and their followers.6 Even if we ignore the problem of unrecoverable deletion that plagues these analyses, the proliferation of syntactic and semantic types assembled under a single morphological construction is embarrassing. But how can one unite all these types of senses? Jespersen is prescient, as he often is. In the volume on morphology of A Modern English Grammar on Historical Principles (1943), he has the following to say about the meanings of compounds: Compounds express a relation between two objects or notions, but say nothing of the way in which the relation is to be understood. That must be inferred from the context or otherwise. Theoretically, this leaves room for a large number of different interpretations of one and the same compound... The analysis of the possible sense-relations can never be exhaustive. (pp. 137-138) The only method that has been used successfully to simultaneously avoid and explain this choice among an infinity of senses, removes both the complexity and the variance from the linguistic to the pragmatic realm, as Jespersen suggests and as I am advocating. Pamela Downing pioneered it in her 1977 article on English nominal compounds. As Downing puts it, echoing Jespersen: Indeed, because of the important differences in the functions served by compounds, as opposed to the sentence structures which more or less accurately paraphrase them, attempts to characterize compounds as derived from a limited set of such structures can only be considered misguided. A paraphrase relationship need not imply a derivational one. (pp. 840-841) She concludes that there are “no constraints on the N + N compounding process itself” (p. 841). The constraints lie instead in how people categorize and refer. 6

Incidentally, an example like skin demonstrates on its face the wrongheadedness of approaches to conversion pairs based on roots or underspecification, in which the meanings of both the verb and the noun are derived from some third category that is neither a verb nor a noun: the direction of the semantic relation is clearly one-way, from noun to verb. The different senses of the verb are all readily traceable to the sense of the noun, but there is no inverse relation for any sense of the verb, so that the verb must be derived from the noun. If we start from some third point, we can’t account for the directionality. The opposite derivation, from verb to noun, holds for instance for result nouns like hit.

Language: Between Words and Grammar

65

Table 1 Categories of zero-verbs from Clark and Clark (1979) LOCATUM VERBS

GOAL AND SOURCE VERBS

LOCATUM VERBS

GOAL VERBS

on/not-on

grease

Human roles

widow

in/not-in

spice

Groups

gang up

at, to

poison

Masses

heap

around

frame

Shapes

loop

along

hedge

Pieces

quarter

over

bridge

Products

nest

through

tunnel

Miscellaneous

cream

with

trustee

SOURCE VERBS

word

LOCATION AND DURATION VERBS

INSTRUMENT VERBS

LOCATION VERBS

go

bicycle

on/not-on

ground

fasten

bail

in/not in

lodge

clean

mop

at, to

dock summer

hit

hammer

cut, stab

harpoon

destroy

grenade

catch

trap

block

dam

follow

track

Musical instruments

fiddle

Kitchen utensils

ladle

Places

farm

Body parts

eyeball

Simple tools

wedge

Complex tools Miscellaneous

mill ransom

DURATION VERBS

AGENT AND EXPERIENCER VERBS

MISCELLANEOUS VERBS

AGENT VERBS

Meals

lunch

Occupations

butcher

Crops

hay

Special roles

referee

Parts

wing

Animals EXPERIENCER VERBS

parrot witness

Elements Other

snow house/s/

66

The Mental Lexicon: Core Perspectives

English zero-verbs show a similarly rich and varied set of interpretations, as we have seen in the case of skin. Clark and Clark (1979) come close to an account like Downing’s for these verbs, but they confine themselves to innovations, which they call contextuals. In my response to the Clarks (Aronoff, 1980), I suggested that it is possible to account uniformly for all zero-verbs, not just the innovations, by a conversion rule of the simple form N → V and that the meaning of the innovative verb always comprises what I call an evaluative domain of the noun’s denotation (essentially a dimension along which the denotation of the noun can be evaluated; a knife is good if it cuts well, a mother is good if she does well what mothers do. A club is good for clubbing, etc.). For most nouns or other lexical items, there is no fixed evaluative domain, so that what the meaning of the novel zero-verb will be depends on the context of its use. Of course, any word’s meaning will become fixed lexically with enough use and time, but that fixing should be of no interest to a linguist. This story holds most remarkably for verbs like boycott and lynch, which are derived from proper nouns and whose meanings are traceable to very specific incidents in which the named person, here Lynch or Boycott, played an important role; William Lynch was an 18th Century Virginian who led a vigilance committee during the American Revolutionary War, while Charles Boycott was a land-agent in 19th Century Ireland who was ostracized by the Irish Land League. There is nothing else to say about the semantics of zero-verbs. Even the notion of an evaluative domain is superfluous, since Gricean principles dictate that the verb have something to do with the noun, no more and no less than what needs to be said to account for the range of data. On this analysis, then, all that the grammar of English contains is the nounto-verb conversion. The beauty of this kind of account is that it leaves the words largely untouched, freeing them to vary as much as speakers need or want them to. Critics have replied that this variance is no beauty and that people like me who advocate such a sparse account, few though we may be, are simply irresponsible and should be purged from the field. These critics, though, ignore the most important point about Downing’s conclusion, which is that it is not theoretical but empirical. When we look at the actual meanings that speakers attribute to novel compound nouns under experimental conditions, as she did, we find that they vary in ways that are formless and void. “The constraints on N + N compounds in English cannot be characterized in terms of absolute limitations in the semantic or syntactic structures from which they are derived” (Downing, 1977, p. 840). Words have idiosyncratic meanings not just because they are preserved in memory and society but because words categorize and refer outside language. I offer one more English zero-verb to buttress my point. The verb is friend, which has been around for at least a century, but has never been used much. The one definition given in WIII is “to act as the friend of” and WIII cites a line from A Shropshire Lad: . In the last couple of years, though, friend has emerged as a common term among the myriads who use the websites friendster.com and facebook.com, both of which crucially involve individual members maintaining lists of “friends”. Though the term seems to have originated with the

Language: Between Words and Grammar

67

earlier Friendster, I will concentrate on Facebook, which was in November 2005, by its own definition, “an online directory that connects people through social networks at school”, because I have the best ethnographic sources of data for it (three children of the right age and their friends). Facebook was inaugurated only in late 2003, but the community of members is quite large and active and has been the subject of articles in major media outlets. To friend someone within the Facebook community is not to act as a friend to that person, but to invite someone to be your Facebook friend, which you do by trying to add that person to your list of Facebook friends. Here is how that works, according to the November 2005 Facebook help page:7 You can invite anyone that you can see on the network to be your friend. Just use the "Search" page to find people you know and then click on the "Add to Friends" button on the right side of the screen. A friend request will be sent to that person. Once they confirm that they actually are friends with you, they will show up in your friends list. Here are some uses of friend in this sense: 8 ) a) EWW! This guy from my London seminar who's a total ASS just friended me! b) All these random people from high school have been friending me this week! c) Should I friend that dude we were talking to at 1020 last night? I don't want him to think I'm a stalker. The invitation is crucial to the meaning of friendv, because the reciprocity of the friend relation in Facebook is enforced by the system: your invitee is not added to your list of friends unless they accept the invitation. Oddly, if the invitee does not accept, (quoting again from the November 2005 Facebook website) “the [inviter] will not be notified. They also will not be able to send you [the invitee] another friend request for some amount of time, so to them, it will just seem as if you haven't confirmed their friendship yet.” The creators of Facebook may see this as a polite method of rejection. I find it odd, but I am not twenty years old, so I guess I just don’t understand. The point of all this is that the meaning of friend x can not just be “act as a friend of x”, as it clearly is in the Housman citation (WIII definition of friendv; note the use of the modal may there) but the invitation, which we may think of as conative aspect, must be built into the meaning. To friend x in this world is “to try to become a friend of x” or “to ask x to be one’s friend”, in the special sense of the noun friend that applies in this world. Of course, we know why the extra predicate is part of the meaning of the verb: it is built into the program and therefore into the social community that surrounds the program! One could try to amend the Clarks’ list or its syntactic equivalent to allow for multi-predicate or aspectual meanings, but that would avoid the underlying problem, that the meaning depends on 7

The verb friend never appears in official Facebook text, only among users.

68

The Mental Lexicon: Core Perspectives

extralinguistic factors from the very first moment of the coining of the word, which is precisely Downing’s empirical point.8

4. LEXICAL ROOTS The lexeme-centric view of word meaning, where lexical semantic information is at least partly nonlinguistic and does not reside in morphemes, has implications for the meanings of roots. If I am right about how lexical semantics works, then two words can have the same root and not share much lexical meaning. There is little relation in meaning, for example, between the name Bork and the verb bork, yet they are as closely related morphologically as any two words can be and must share a root, if the term is to have any content. What does one say then about lexical roots, which are supposed to be the atomic meaningful units of language, if two instances of the same root can share so little meaning? The simplest ploy is to deny the linguistic reality of roots entirely, what my colleague Robert Hoberman calls the anti-rootarian position. But there is a middle ground, where words have morphological structure even when they are not compositionally derived, and where roots are morphologically important entities, though not particularly characterized by lexical meaning. In the first piece of morphological analysis I ever did (published in Aronoff, 1976), I showed that certain roots in English verbs can be active morphologically even when they are obviously meaningless: each root conditions a certain set of affixes and alternations that is peculiar to that root alone. Some examples are given in Table 2. These Latinate roots in English are the historical reflexes, through a complex borrowing process, of the corresponding verb roots in Latin. Even for Classical Latin, the alternations exhibited by at least some of these verb roots were not phonologically motivated. But in Latin, these same roots were morphologically active in other ways too, ways that are independent of syntax or semantics. Perhaps most intriguing is the role of roots in determining whether a given verb was deponent. The traditional Latin grammatical term deponent is the present participle of the Latin verb deponere “set aside”. Deponent verbs have set aside their normal active forms and instead use the corresponding passive forms (except for the present participle) in active syntactic contexts. The verb admetior “measure out”, for example, is transitive and takes an accusative object, but passive in form, because it is deponent. Similarly for obliviscor “forget” (oblitus sum omnia “I have forgotten everything” [Plautus]), scrutor “examine”, and several hundred other Latin verbs. There are even semi-deponent verbs, which are deponent only in forms based on the perfect stem. 8

I am not denying that non-causative individual verbs can have multi-predicate meaning. The verb propose “to make an offer of marriage” (WIII) is similar to friend in involving an invitation. The question is whether there is a systematic multi-predicate syntactic relation between nouns and their derivative zero-verbs involving conative or similar aspect.

Language: Between Words and Grammar

69

Table 2 Some Latinate Roots in English with their alternations Word-Final Root

Sample verb

Root +ion

Root +ive

sume

Resume

resumption

resumptive

mit

Permit

permission

permissive

pel

Repel

repulsion

repulsive

ceive

Receive

reception

receptive

duce

Deduce

deduction

deductive

scribe

Prescribe

prescription

prescriptive

pete/peat

Compete

competition

competitive

cur

Recur

recursion

recursive

For some time, Zheng Xu, Frank Anshen, and I have been working on a comprehensive study of Latin deponent verbs, using a database of all the main-entry deponent verbs in the Oxford Latin Dictionary, which covers the period from the first Latin writings through the end of the second century CE (Xu et al., in press). I will report here on only one small part of our research, that involving deponent roots. We show elsewhere, based on an exhaustive analysis of all the senses of all deponent verbs, that neither syntax nor semantics is the best predictor of deponency, though both are factors. Our database contains 287 deponent verbs not derived from another lexical category (about half the total number of deponents), the great majority of them consisting of either a bare root or a root and a single prefix. The senses of all deponent verbs are about evenly divided between transitive and intransitive values and a given root may appear in both transitive and intransitive verbs. Of these 287 verbs not derived from another lexical category, 85% (244) have deponent roots, which we define as a verb root that occurs only in deponent verbs. There are 52 deponent roots, 22 of which occur in four or more distinct verbs. Table 3 is a list of these roots. Thus, whether a given verb will be deponent is determined to a very large extent by its root, independent of the meaning of either.

4.1. Hebrew Roots I now turn to the role of the root in the morphology of Modern Hebrew verbs. For more than a millennium, traditional grammarians have characterized Semitic languages as being morphologically grounded in meaningful roots, each root consisting usually of three consonants, with some number of lexemes (verbs, nouns, and adjectives) built on each root.

70

The Mental Lexicon: Core Perspectives

Table 3 Latin deponent verb roots Latin deponent root

Number of verbs with this root

Example verb

Gloss

gradi

22

gradior

proceed

lab

16

labor

glide

sequ

15

sequor

follow

min

11

comminor

threaten

loqu

10

loquor

talk

nasc

10

nascor

be born

moli

9

molior

build up

mori

9

morior

die

ori

8

orior

rise

tue

8

tueor

look at

f(or)

7

affor

address

hor(t)

7

hortor

encourage

luct

7

luctor

wrestle

meti

7

metior

measure

fate

5

fateor

concede

prec

5

precor

ask for

quer

5

queror

regret

apisc

4

apiscor

grasp

fru

4

fruor

enjoy

fung

4

fungor

perform

medit

4

meditor

contemplate

spic

4

conspicor

see

ut

4

utor

make use of

Language: Between Words and Grammar

71

But not every Modern Hebrew root has a constant meaning.9 Table 4 contains a single root with an apparently constant meaning.10 The meanings of the individual noun, verb, and adjective can be predicted reasonably well if each morphological pattern is associated with its own compositional semantic function. This is what we expect from the traditional account. Table 5, though, quickly brings such an enterprise to grief. What do pickles (kvu∫im) have to do with highways (kvi∫im)? The story of how these two words are related is quite simple. Paved roads in early modern Palestine were macadamized: made of layers of compacted broken stone bound together with tar or asphalt. Modern roads are still built by pressing layers, but in a more sophisticated manner. Traditional pickling also involves pressing: whatever is to be pickled is immersed in brine and pressed down with a weight, but the container should not be sealed. If it is, it may explode (as I know from personal experience). But pressing alone without brine does not constitute pickling. Every cook worth his or her salt knows that beef brisket should be pressed after it is cooked and before it is sliced, regardless of whether it is pickled (corned) or not. Modern industrial pickling does away with pressing in various ways and not every paved road in Israel is called a kvi∫, only a highway is, so while it is clear why pickles and highways both originate in pressing, there is nothing left of pressing in the meanings of these Hebrew words today. One who points out their related etymologies to a native speaker of the language is rewarded with the tolerant smile reserved for pedants, as I also know from personal experience. The reflex response to problems like this is to posit an “underspecified” core meaning for the root, which is supplemented idiosyncratically in each lexical entry. It is logically impossible to show that underspecification is wrong, but trying to find a common meaning shared by pickles and highways brings one close to empirical emptiness and this methodological danger recurs frequently in any Semitic language. In any case, there is no need to find a common meaning in order to relate the two words morphologically, as I will show. In fact, Hebrew verb roots can be identified on the basis of alternations remarkably similar in type to those that operate on Latinate roots in English, without reference to meaning or regular phonology. This observation is nothing new; root classes are the bane of students of Hebrew as a second language.11

9

Earlier stages of Hebrew or related Semitic languages do not show much more regularity in this regard. 10 Even these examples are a little forced, since some of the lexical items are very infrequent or even obsolete. 11 I will confine my discussion to verbs, both because there are fewer verb patterns than noun patterns and because it is easier to believe that verb roots are related semantically, so if we can talk about verb roots without recourse to meaning, then all the more so for nouns and adjectives, whose meanings show much greater variety.

72

The Mental Lexicon: Core Perspectives

Table 4 Lexemes formed on the Modern Hebrew root ZLP Root = zlp

‘sprinkle, spray, drip’

Nouns zelef

‘sprinkling fluid (perfume)’

zalfan

‘sprinkler’

mazlef

‘watering can, sprayer’

hazlafa

‘sprinkling’

ziluf

‘sprinkling’

zlifa

‘sprinkling’

Verbs zalaf

‘to pour, spray, sprinkle’

zilef

‘to drip’

hizlif

‘to sprinkle’

zulaf

‘to be sprinkled’

huzlaf

‘to be sprinkled’

Adjective mezulaf

‘sprinkled’

First, we must know a few basics of Hebrew verb morphology. Hebrew has a set of what are traditionally called binyanim, seven in number, often called conjugations in the recent theoretical literature, a term I will adopt here. Each conjugation rigidly assigns to each cell of the verb paradigm a stem pattern containing a prosodic shape complete with vowels. A stem pattern may also include a prefix. The hif `il conjugation, for example, has a prefix hior ha-, and the pattern CCiC (e.g., higdil “grew”). The pi`el conjugation has the stem pattern CiCeC in the past, meCaCeC in the present, and CaCeC in the future (e.g., megadel “grow”). Traditionally, each conjugation (like each root) is said to have a constant meaning or syntactic structure. Some modern scholars have questioned that claim, but whether conjugations have meaning is orthogonal to the matter at hand, which is roots.

Language: Between Words and Grammar

73

Table 5 Lexemes formed on the Modern Hebrew root KB ∫ Root = kb∫

‘press’

keve∫

‘gangway, step, degree, pickled fruit’

kvi∫

‘paved road, highway’

kvi∫a

‘compression’

kiv∫on

‘furnace, kiln’

maxbe∫

‘press, road roller’

mixba∫a

‘pickling shop'

Verbs kava∫

‘to conquer, subdue, press, pave, pickle, preserve, store, hide’

kibe∫

‘to conquer, subdue, press, pave, pickle, preserve’

hixbi∫

‘subdue, subjugate’

Adjectives kavu∫

‘subdued, conquered, preserved, pressed, paved’

kvu∫im

‘conserves, preserves’

mexuba∫

‘pressed, full’

Besides being determined by conjugations, verb forms depend on the class of their root. Root classes, except for the default full class, are traditionally identified in terms of one of the consonants that occupy the three canonical root consonant positions, which I will call R 1 , R 2 , and R 3.12 For example, the root class R 3 h is so named because the third consonant of the root is h; in a root of class R 3 n the third consonant is n. These classes, named in this way, figure prominently in traditional Hebrew grammar. At some point in the early history of the language, the differences in verb forms among root classes were undoubtedly predictable from the phonology of the consonant that defined each type. The passage of time, however, has made these phonological conditions opaque, so that verbs with identical consonants in the same position can show different alternation patterns. This differential patterning proves that the classes are no longer predictable phonologically, but are more like verb classes of the sort 12

There are roots with more than three consonants, but these are not germane to our conversation.

74

The Mental Lexicon: Core Perspectives

found, for instance, in Germanic languages, where the classes are defined by distinctions in ablaut and verbs must be marked for membership in a given strong class.13 In Hebrew, similar sorts of idiosyncratic alternations are determined by the class of a verb’s root (which also appears to be true in Germanic), a fact that, together with conjugations, gives roots their linguistic reality. The root classes and conjugations of Semitic constitute two dimensions of a fourdimensional matrix, the others of which are person/number/gender combinations (ten are possible), and tense (in Modern Hebrew: past, present, future, and imperative, though the last is becoming increasingly rare). The regularities of any given individual verb form are thus exhausted by four specifications: its class, its conjugation, its person/number/gender value, and its tense. Lest it be thought that this matrix is trivial, a reasonably complete traditional table of verbs (Tarmon and Uval, 1991) lists 235 distinct combinations of class and conjugation alone, what they call verb types, of which only 9 contain just a single root. Multiplying 235 by the ten person/number/gender values in all of the tenses (a total of 26), yields a little more than six thousand distinct form types into which any verb form must fall. In general, there is a contrast among all R x z roots between those that pattern like regular roots and those that fall into a marked class. For example, there are two morphologically distinct sets of R 1 n roots and they cannot be distinguished by their phonological makeup, which means that the R 1 n roots in one set must be marked as belonging to a strong/marked class. Because one set patterns exactly like regular full verbs, the other set constitutes the marked class. The two morphologically distinct sets of R 1 n roots are those in which the n is deleted between the vowel of a prefix and R 2, which are traditionally termed missing R 1 n roots, and those in which the n is not deleted. The latter pattern exactly as do completely regular roots, and thus are full roots, not members of any marked root class. Originally, the difference in patterning was predictable from a given root’s phonological makeup, but this phonology has been unproductive in recorded history, so that whether R 1 n is missing is now a property of individual verb roots: npl is a missing R 1 n root; ngd is not.14 We even find etymologically distinct homophonous roots, one of which is a missing R 1 n root, and the other not. Compare hibit “look”, in which the n is missing, with hinbit “sprout”, in which it is not. Importantly, individual lexemes belonging to the same root do not vary. Examples of missing R 1 n roots are given in Table 6; full R 1 n roots are exemplified in Table 7. Forms in which the n is missing are italicized; those in which it is phonologically eligible to be missing but isn’t, are boldfaced. Although I do not have space to show it here, it can easily be shown that R 1 y roots also fall into a number of different classes, one of which is the missing R 1 n class (Aronoff, in press). 13

Not even the most abstract of phonological representations could save the system. In the case of R 3 n, for instance, one would have to posit n 1 and n 2 , without any phonetic way to distinguish them beyond the fact that they pattern differently. 14 The root ngd is ancient, but its use in verbs is not, which shows that what exempts a verb root from assimilation of n is not the age of the root, but rather the point at which it was first used as a verb.

Language: Between Words and Grammar

75

Table 6 Missing R 1 n roots Root ntp

Conjugation pa`al

Past nataf

Future yi-tof

Infinitive li-tof

ntp ncl

hif`il nif`al

hi-tif ni-cal* (cf. nixtav)

ya-tif yi-nacel

Gloss drip (intr.) drip (tr.) le-ha-tif le-hi-nacel arrive

npl npl

pa`al hif`il

nafal hi-pil

yi-pol ya-pil

li-pol le-ha-pil

npl

huf`al

hu-pal

ju-pal

n.a.

nbt

hif`il

hi-bit*

ya-bit*

le-ha-bit*

fall cause to fall be thrown down look

ntn

pa`al

natan

yi-ten

la-tet*

give

ntn

nif`al

nitan* (cf. nixtav)

yi-naten

le-hi-naten be given

lkx

pa`al

lakax

yi-kax

la-kaxat

take

npc

pa`al

nafac

yi-poc

linpoc

disperse, smash

npk

pa`al

nafak

yi-pok

linpok

ngd

hif`il

hi-gid

ya-gid

le-ha-gid

go/come out, result tell

Remarks n drops before R2 (t) *would be nincal if n were not assimilated

*b would be v if n were not assimilated *infinitive is irregular, should be liten *would be nintan if n were not assimilated only example of initial l that behaves like missing R 1 n n does not assimilate in infinitive n does not assimilate in infinitive

Nothing of what I have said depends on these roots having any constant meaning. Except that the marked classes each contain a subset of the roots that begin in coronal sonorants, none of it depends on phonology either. Instead, root classes are identified by their paradigms, a claim that should be amenable to experimental verification, although most previous psycholinguistic work on Hebrew roots has concentrated on demonstrating the psycholinguistic reality of individual roots (Deutsch and Frost, 2003; Berent and Shimron, 2003; Shimron, 2006). Modern Hebrew verb roots, like those of English and Latin, thus

76

The Mental Lexicon: Core Perspectives

furnish yet another example of morphology by itself, leaving lexical meaning to reside where it belongs, not in roots, which are purely grammatical objects, but in lexemes, where language meets the world. It is very difficult to do linguistics while respecting the integrity of words, but I hope that at least some small fraction of my readers now believe that this kind of work is not only possible but also necessary if we are to make progress in understanding language. Table 7 Full R 1 n roots Root

Conjugation

Past

Future

Infinitive

Gloss

npk

hif`il

hi-npik

ya-npik

le-hanpik

issue

ngd

pa`al

Nagad

yi-ngod

li-ngod

oppose

nbt

hif`il

hi-nbit

ya-nbit

le-ha-nbit

sprout

ngb

pa`al

Nagav

yi-ngov, yigov

li-ngov

become dry

ngb

pi`el

Nigev

ye-nagev

le-nagev

dry

ngb

hif`il

hi-ngiv

ya-ngiv

le-ha-ngiv go south

ncx

hif`il

hi-ntsíax

yantsíax

le-hantsíax

Remarks

n drops variably in the future

memorialize, eternalize

5. IMPLICATIONS Under extreme conditions like those of a brand-new language, words emerge as the basic building blocks of syntactic structure. At the other extreme, a close examination of the coining of brand-new words like the verb friend in English supports a view of language in which words reside at the intersection between the linguistic system and extra-linguistic knowledge. Finally, the residue in Modern Hebrew of ancient linguistic structure, most likely close to ten thousand years old, remains as a purely formal structure within words, showing that, though words may be syntactic units, they are not unstructured units. The overall picture that results is one in which words or lexemes or lemmas comprise what oldfashioned structuralist linguists (Harris, 1951) would have called a linguistic level, a basic dimension of language along which important generalizations must be stated. The general model of the mental lexicon that emerges from this work is quite conventional. The lexicon is organized around lexemes or lemmas (depending on what terminology you wish to follow), these lexemes are not purely linguistic entities, but reflect properties of the cognitive and

Language: Between Words and Grammar

77

social system in which the speaker and his or her language are embedded, and the lexemes are not unstructured, but obey the sometimes very intricate and arbitrary principles of the language in which they are also embedded. The mental lexicon thus lives, like the words within it, at the intersection of language with cognition and culture. It is quintessentially human.

78

The Mental Lexicon: Core Perspectives

REFERENCES Aronoff, M. (1972). The semantics of growth. Quarterly Progress Report of the Research Laboratory of Electronics, 113, 159-162. Aronoff, M. (1976). Word Formation in Generative Grammar. MIT Press, Cambridge, MA. Aronoff, M. (1980). Contextuals. Language, 56, 744-758. Aronoff, M. (in press). In the beginning was the word. Language 87. Aronoff, M., I. Meir, and W. Sandler (2005). The paradox of sign language morphology. Language, 85, 301-344. Baudouin de Courtenay, J. (1972). A Baudouin de Courtenay Anthology (E. Stankiewicz, ed.). Indiana University Press, Bloomington. Berent, I. and J. Shimron (2003). What is a root? In: Language Processing and Acquisition in Language of Semitic, Root-based, Morphology (J. Shimron, ed.), pp. 201-222. Lawrence Erlbaum, Mahwah, New Jersey. Bloom, P. (2000). How children learn the meanings of words. MIT Press, Cambridge, MA. Carstairs-McCarthy, A. (2005). Basic terminology. In: Handbook of Word-Formation (P. Stekauer and R. Lieber, eds.), pp. 5-23. Dordrecht, Springer. Chomsky, N. (1957). Syntactic Structures. Janua linguarum, no. 4. ’s-Gravenhage, Mouton. Chomsky, N. (1970). Remarks on nominalization. In: Readings in Transformational Grammar (A. Jacobs and P. Rosenbaum, eds.), pp. 184-221.Ginn and Col, Waltham, MA. Clark, E. V. and H. H. Clark (1979). When nouns surface as verbs. Language, 55, 767-811. Deutsch, A. and R. Frost (2003). Lexical organization and lexical access in a non-concatenated morphology. In: Language Processing and Acquisition in Language of Semitic, Root-based, Morphology (J. Shimron, ed.), pp.165-186. Joseph Erlbaum, Mahwah, New Jersey. Downing, P. (1977). On the creation and use of English compound nouns. Language, 53, 810-842. Frege, G. (1914). Letter to Jourdain. In: Philosophical and Mathematical Correspondence (G. Gabriel., H. Hermes, F. Kambartel, C. Thiel, and A. Veraart, eds., abridged for the English edition by B. McGuinness), pp. 78–80. Basil Blackwell, Oxford. Hale, K. and S. J. Keyser (1993). On argument structure and lexical expression of syntactic relations. In: The View from Building 20 (K. Hale and S. J. Keyser, eds.), pp. 53-109. MIT Press, Cambridge, MA. Harris, Z. (1951). Methods in Structural Linguistics. University of Chicago Press, Chicago. Hockett, C. F. (1960). The origin of speech. Scientific American, 203 (3), 88-96. Jespersen, O. (1943). A Modern English Grammar on Historical Principles, Part VI: Morphology. George Allen and Unwin, London. Kisch, S. (2004). Negotiating deafness in a Bedouin community. In: Genetics, disability, and deafness (J. Van Cleve, ed.), pp. 148-73. Gallaudet University Press, Washington. Lane, H., R. Piller and M. French (2000). Origins of the American Deaf-world: Assimilating and differentiating societies and their relation to genetic pattern. In: The signs of language revisited (K. Emmorey and H. Lane, eds.), pp. 77 - 100. Lawrence Erlbaum, Mahwah, NJ. Lees, R. B. (1960). The grammar of English nominalizations. Indiana University Research Center in Anthropology, Folklore, and Linguistics. Publication 12. Bloomington, Indiana.

Language: Between Words and Grammar

79

Levelt, W. (1989). Speaking: From intention to articulation. MIT Press, Cambridge, MA. Malkiel, J. (1967). Each word has a history of its own. Glossa, 1, 137-149. Martinet, A. (1957). Arbitraire linguistique et double articulation. Cahiers Ferdinand de Saussure, 15, 105-116. Reprinted (1966) in: Readings in Linguistics II (E. Hamp et al., eds.), pp. 371-378. University of Chicago Press, Chicago. Meillet, A. (1903). Introduction á l’étude comparative des langues indo-européennes. Hachette, Paris. Sandler, W., I. Meir, C. Padden and M. Aronoff (2005). The emergence of grammar: systematic structure in a new sign language. Proceedings of the National Academy of Sciences, 102, 26612665. Saussure, F. de (1975). Cours de linguistique générale [Course in general linguistics]. Edited by Charles Bally and Albert Sechehaye, with the collaboration of Albert Riedlinger; critical edition prepared by Tullio de Mauro. Payot, Paris. Scott, D., R. Karmi, K. Eldebour, G. Duyk, E. Stone, E. and V. Sheffield (1995). Nonsyndromic autosomal recessive deafness is linked to the DFNB1 locus in a large inbred Bedouin family from Israel. American Journal of Human Genetics, 57, 965-968. Shimron, J. (2006). Reading Hebrew. Lawrence Erlbaum, Mahwah, New Jersey. Stokoe, W. (1960). Sign language structure: an outline of the visual communication systems of the American Deaf. Buffalo, Dept. of Anthropology and Linguistics, Studies in Linguistics occasional papers, 8. Tarmon, A. and E. Uval (1991). Hebrew verb tables. Tamir, Jerusalem. Webster’s third new international dictionary of the English language unabridged (1986). P. B. Gove and the Merriam-Webster editorial staff, eds. Merriam-Webster, Springfield, MA. Xu, Z., M. Aronoff and F. Anshen (in press). Deponency in Latin. In: Deponency (M. Baerman and G. Corbett, eds.). Oxford University Press, Oxford.

This page intentionally left blank

5 STORAGE AND COMPUTATION IN THE MENTAL LEXICON

R. H. Baayen, Radboud University and Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands

1. INTRODUCTION In the seventies of the previous century, the mathematical properties of formal languages have provided a key source of inspiration to morphological theory. Models such as developed by Lieber (1980) and Selkirk (1984) viewed the lexicon as a calculus, a formal system combining a repository of morphemes with rules for combining these morphological atomic units into complex words. This approach to the lexicon was driven by two fundamental assumptions. First, the lexicon was assumed to be a compositional derivational system. Complex words were believed to be generated from simpler forms (Bloch, 1947; Chomsky and Halle, 1968). Second, following Bloomfield (1933), the set of atomic elements was assumed to comprise any word or formative that is not predictable by rule. Rule-governed combinations of these atomic units, the regular complex words, were assumed not to be available as units in the lexicon, as storage would introduce unnecessary redundancy in the model. Instead of being listed (i.e., stored without analysis or substructure), regular complex words were generated (produced or parsed) by rule. Unsurprisingly, the goal of morphological theory was seen as accounting for which words belong to the set of possible words in the languages of the world. The question of whether a regular complex word exists in a language was regarded as a question addressing performance rather than competence, and hence irrelevant for morphological theory. Although many other formalisms have been developed to replace sequences of rules (Halle and Marantz, 1993; McCarthy and Prince, 1993), these formalisms did not challenge these fundamental assumptions of generative morphology. In optimality theory, for instance,

81

82

The Mental Lexicon: Core Perspectives

forms still enter into derivational relations, even though the algorithm that relates underlying forms to surface forms is not based on a sequence of rules but on constraint satisfaction. This type of theory of the lexicon is to a surprising extent equally adequate as a competence theory for how a pocket calculator works. A pocket calculator has a set of atomic elements, the symbols on its keys. Its chip is endowed with a small set of arithmetic rules that, when supplied with a legal string, compositionally evaluate this string. Whenever a pocket calculator is requested to evaluate a string such as “2 + 3”, it computes the outcome. It has no memory that holds the output of previous evaluations of the same string. It never learns from past experience. The balance of storage and computation is shifted totally to the maximization of computation and the minimization of storage. The first goal of this chapter is to show that the pocket calculator provides a fundamentally flawed metaphor for understanding morphological structure and processing in the mental lexicon. To this end, we first survey evidence from experimental studies of lexical processing, and then consider another source of information, the fine phonetic detail that is present in the acoustic signal. We then address the second goal of this chapter, to provide an indication of the kind of formal mathematical model that may help us to better understand process and representation in the mental lexicon.

2. EXPERIMENTAL EVIDENCE Over the last twenty-five years, the regular and irregular past tense forms in English and related languages have provided a rich testing ground for theories of morphological processing. Whereas English regular verbs have a past tense form in the dental suffix -ed (e.g., walked, claimed), irregular verbs have past tense forms that range from suppletion (go/went) to invariance (cut/cut) and from pure vocalic alternation (give/gave) to combinations of vocalic alternation and a variant of the dental suffix (sell/sold). Bybee and Slobin (1982) and Bybee and Moder (1983) called attention to the many kinds of subregularities that characterize the irregular verbs, which older structuralists had already characterized as semi-productive (e.g., Van Haeringen, 1940). Most researchers understand regular past tense forms as being derived from their present tense stems (Bloch, 1947; Chomsky and Halle, 1968) by a simple rule adding the dental suffix. Although irregular past tense forms might also be analyzed as governed by various unproductive rules, such rule-based descriptions tend to be baroque, fairly arbitrary and uninsightful. For understanding the semi-regularities of the irregular past tense, connectionist models offered an alternative that obviated the need for a series of ad hoc unproductive rules (Rumelhart and McClelland, 1986; McClelland and Patterson, 2002b). The response of generative linguists (Pinker and Prince, 1988, 1994; Pinker, 1991) was to defend the Bloomfieldian model by claiming that regular and irregular morphology belong to two completely independent cognitive systems, the dual mechanisms of rule (for regulars)

Storage and Computation in the Mental Lexicon

83

and rote (for irregulars). Irregulars would be stored in an associative memory, regulars would not be stored at all but always derived by means of morphological rules (Pinker, 1991, 1997). The theory of speech production developed by Levelt et al. (1999) and its computational implementation in the WEAVER model (Roelofs, 1996, 1997a,b) provide a psycholinguistic formalization of the generative approach to storage and computation. The WEAVER model embodies a fully decompositional theory of morphological processing in production. Conceptual representation for words like handbooks and walked are linked to lemma representations, which specify their syntactic and inflectional properties (handbooks is a plural noun, walked a verb in the past tense). These lemma representations activate the form representations (lexemes) of their constituent morphemes, hand, book and s for handbooks, and walk and -ed for walked. In this model, complex words do not receive their own lexemes, they are assumed always to be produced through their constituents. In the absence of immediate constituents, irregular verbs are assigned their own lexemes. In short, the WEAVER model, as well as the general approach of Pinker and colleagues, take as their point of departure that morphology is a simple formal system similar in essence to the computational system implemented in a standard pocket calculator. Although a model like WEAVER is attractive for its simplicity, economy, and the broad range of experimental data that it accounts for, it is becoming increasingly clear that its generative design leads to a series of subtle predictions that turn out to be demonstrably incorrect.

2.1. Storage is Ubiquitous First, storage is not restricted to irregular words. Fully regular complex words also leave traces in lexical memory, as shown by Taft (1979) and Sereno and Jongman (1997) for comprehension in English, Baayen et al. (1997, 2002) for comprehension in Dutch, and Bien et al. (2005) for production in Dutch. All these studies observed that the frequency of a complex word itself was predictive for processing latencies, independently of the frequencies of its constituents. Such a frequency effect is widely regarded as a proof of the existence of a separate independent representation for a complex word. It has been argued that in English regular complex words would not leave traces in lexical memory when their frequencies fall below a threshold of 6 per million (Alegre and Gordon, 1999). However, Wurm and Baayen (2007) observed a word frequency effect for English regular inflected words well below this threshold in both visual and auditory lexical decision. Interestingly, the presence of a word frequency effect went hand in hand with the absence of a stem frequency effect in the data of Wurm and colleagues. Traditionally, word frequency and stem (or root) frequency effects have been interpreted as the hallmarks of whole word based processing versus decompositional processing respectively. For very lowfrequency words, however, it is highly unlikely that morphological structure is completely irrelevant. Therefore, Wurm and colleagues suggest that word frequency be reinterpreted as a

84

The Mental Lexicon: Core Perspectives

joint probability, the probability of the co-occurrence of the immediate constituents. If this interpretation is correct, lexical frequency effects may derive from two sources: memory for phonological sequences (morphs), and memory for sequences of such sequences, i.e., sequences of morphs. For (regular) complex words, both kinds of memory are probably involved simultaneously. In more recent work, Pinker and colleagues acknowledge that regulars can be stored. Nevertheless, their argument is that regulars may perhaps be stored, but crucially they need not be stored. In normal language use, regulars would be processed by rule, and only under extreme experimental conditions would one begin to see that regulars have their own, albeit normally superfluous, representations (Pinker and Ullman, 2002a,b). However, why would supposedly redundant storage of regulars take place at all? In fact, we can infer from its mere existence in experiments that it must be advantageous for the brain to keep track of detailed combinatorial probabilities, contrary to what the metaphor of the pocket calculator would lead one to expect. Interestingly, De Vaan et al. (2007) discuss preliminary evidence that regular complex forms may already leave a trace in memory after just a single exposure. Crucially, their evidence is not restricted to visual lexical decision, but extends to self-paced reading, a task with a much higher degree of ecological validity. One of the advantages that storage offers for lexical processing is probability-driven elimination of unlikely but possible segmentations in comprehension. Spurious segmentations are ubiquitous not only in syntax but also in morphology, and knowledge about the likelihood of substructures is crucial for efficient selection of the most likely morphological parse (Baayen and Schreuder, 2000; Bod, 2006). 2.2. Processing is not Derivational Given that regular inflected words seem to have their own traces in memory, it is no longer necessary to assume that a past tense form is derived from its stem. The idea that complex forms are, in some real sense, constructed on-line from their parts, i.e., that a past-tense form like walked is derived in real time from its stem walk, is one of the few assumptions that many connectionist models (Rumelhart and McClelland, 1986; Plunkett and Marchman, 1991; MacWhinney and Leinbach, 1991) share with symbolic theories. Evidence is accumulating, however, that this very common assumption is wrong. One source of evidence stems from research on speech errors. Stemberger and Middleton (2003) presented irregular verbs in the progressive form (is falling), and asked subjects to respond with the simple past (fell). When the vowel of the past tense is more frequent in the language in general compared to the vowel of the present tense, overregularization errors (falled for fell) decreased. When it is less frequent, overtensing errors (felled for fall) were more likely. These data suggest that the present and past tense forms are in competition, and that this competition is modulated by the a priori probabilities of the vowels in these verb forms (see also Stemberger, 2004).

Storage and Computation in the Mental Lexicon

85

Tabak et al. (2005b) obtained further evidence for competition between the past and present tense forms using a task in which subjects were shown a present or a past tense form, and were asked to say out loud the corresponding past or present tense form. In this task, henceforth cross-tense naming, the ratio of the frequency of the form seen to that of the form said was inhibitory. This inhibitory effect was the same for both irregular and regular verbs. It also did not vary with the direction of naming, from present to past or from past to present. In cross-tense naming, the form seen (the cue) apparently inhibits the form to be said (the target). The presence of this effect for irregulars is expected in the light of the results of Stemberger and Middleton (2003). Its presence also for regulars shows that the present and past tense forms of regulars exist in the mental lexicon just as the present and past tense forms of irregulars. Interestingly, the ratio of the two inflectional variants was not predictive at all in straightforward word naming nor in picture naming of the same forms. This shows that in normal situations, the two inflectional variants are not considered jointly, which would be necessary if one form is to be derived from the other. Instead, the targeted form is retrieved from memory without noticeable interference from its counterpart in the opposite tense.

2.3. Paradigmatic Structure Affects Processing We have seen that probabilistic information about individual inflectional variants is available in lexical memory. It is well-known that inflectional variants are organized in paradigms (see, e.g., Matthews, 1974). From the syntagmatic perspective of standard decompositional approaches, paradigms are enigmatic oddities with little more status than educationally useful ways of displaying inflectional variants. After all, to the extent that an inflectional variant is decomposable, its structure can be accounted for by a syntagmatic rule. However, paradigmatic structure and its complexity is emerging from recent experimental studies as a genuine independent factor in lexical processing. Moscoso del Prado Martín et al. (2004b) developed, as part of an overall informationtheoretic measure of processing complexity, a measure of paradigmatic inflectional complexity based on the entropy measure of Shannon and Weaver (1949),

H = – ∑ pi log2(pi),

(1)

where i ranges over inflectional variants, and pi is the probability of a variant in its paradigm (estimated by its relative frequency in the paradigm). Baayen et al. (2006) considered this measure (henceforth inflectional entropy) by itself in a regression study of English monomorphemic words, applying (1) in its simplest form by allowing i to range over all inflectional variants of a given word. They observed a negative correlation of inflectional entropy with response latencies in visual lexical decision, and a positive correlation for

86

The Mental Lexicon: Core Perspectives

subjective frequency estimates. Words with a more complex, informationally rich inflectional paradigm have more connections to other words in the mental lexicon, and this superior lexicality allows faster lexical decision responses and gives rise to higher subjective frequency estimates. Tabak et al. (2005a) also observed a (non-linear) effect of inflectional entropy for Dutch verbs in visual lexical decision. The information structure of more complex paradigms has also been shown to affect morphological processing. Kostić et al. (2003) investigated Serbian nominal paradigms with visual lexical decision. Their key predictor was the average probability of a syntactic function of a given case inflection expressed in bits of information. This information measure can be calculated across all relevant nouns in the language, or it can be calculated for each noun separately. For instance, Serbian plural feminine nouns take the ending -ama for the dative and the instrumental. We can base our estimate of a syntactic function such as recipient either across all feminine plurals, or for a specific noun, say ženama. Kostić and colleagues observed that both the general and the item-specific measures were predictive, with forms carrying a higher information load giving rise to prolonged response latencies. Current work on Dutch and Spanish using the picture naming paradigm suggests that inflectional structure is also predictive for speech production. Tabak et al. (2006) observed a facilitatory effect for inflectional entropy for Dutch past-tense forms, for both regular and irregular verbs. For Spanish verbs, (Van Buren et al., 2007) observed facilitation for the inflectional family size, i.e., for the number of nonzero inflectional variants realized in a corpus, over and above an effect of the word’s lemma frequency. Note that the effects of inflectional entropy and the inflectional family size bear further witness, albeit indirectly, to the presence of memory traces for regular inflected words. Paradigmatic effects in morphological processing are not restricted to inflection. The derived words and compounds in which a word occurs, its morphological family, codetermines lexical processing (Moscoso del Prado Martín et al., 2004a). Furthermore, morphological family members sharing the same structural position have been found to constitute a domain of analogical generalization (Krott et al., 2001). Analogical generalization challenges the high level of abstraction that is part and parcel of classical syntagmatic approaches — a syntagmatic rule by definition is blind to the properties of individual words and has access only to a selection of abstract general features. Not surprisingly, the syntagmatic design of decompositional models renders them unable to account for the many analogical, graded effects in morphology and morphological processing (Seidenberg and Gonnerman, 2000; Ernestus and Baayen, 2003; McClelland and Patterson, 2002b). Importantly, whereas Pinker (1999), assumed that analogical similarity would be restricted to irregulars, as irregulars and only irregulars would be stored in an associative memory, Albright and Hayes (2003) have shown that regulars are also subject to analogical similarity (islands of reliability in their terminology) just as are irregulars (see also Ernestus and Baayen 2003).

Storage and Computation in the Mental Lexicon

87

2.4. Form and Meaning Interact The decompositional approach of Levelt et al. (1999) implements a high degree of modularity and encapsulation. Conceptualization leads to the selection of a lemma. Once a lemma has been selected, a new process activating the relevant constituent morphemes is started up. This process is completely independent of prior conceptualization processes. Once a morpheme (lexeme) has been activated, it in turn activates its constituent phonemes, again fully independently of any preceding processes. The hypothesis underlying encapsulated modeling is that the rules and regularities at the level of word-form operate independently from rules and regularities at the level of word meaning. Thus, Pinker regards semantics as irrelevant for the past tense (Pinker, 1999). In connectionist approaches to morphology, the modularity assumption is dropped, and form and meaning are allowed to interact, as for instance in the triangle model (Joanisse and Seidenberg, 1999; Seidenberg and Gonnerman, 2000). Patterson et al. (2001a,b) argue that irregulars come to depend more on semantic similarity compared to regulars due to reduced similarity in their phonological form. They call attention to the cooccurrence of semantic deficits with degraded performance on irregular verbs (but see Tyler et al., 2004). Furthermore, there is independent distributional evidence that irregular verbs tend to have more semantic neighbors than do regular verbs (Baayen and Moscoso del Prado Martín, 2005). Experimental evidence that this difference in semantic density between regulars and irregulars may affect lexical processing is reported by Tabak et al. (2005a). They observed that in visual lexical decision, semantic density (gauged by means of the count of synonym sets in WordNet) was more facilitatory for irregular past tense forms than for regular past tense forms. It is precisely the forms of irregular verbs that carry the irregularity for which we find that the greater semantic density of irregular verbs boosts lexical processing. Differences in embodiment (Barsalou, 2003; Feldman, 2006) may also be involved. Table 1 lists a number of basic verbs for position and movement in English and Dutch. In both languages, irregular verbs are in the majority, suggesting that irregular verbs have a greater degree of embodiment than regulars. Work in progress (Tabak et al., 2006) provides further evidence for this hypothesis. We had artists make photographs of verbal actions, covering some 180 verbs. The artists who made the photographs of verbal actions reported that acting out irregulars was much easier than acting out regulars. This subjective impression was supported by two observations. First, the byte size of the jpeg files of photographs for irregulars were significantly smaller than those for regulars. Second, if we look at the names elicited from subjects for the pictures, we see that the uncertainty about the pictures’ names (measured by means of entropies over the frequency distribution of different verbs elicited for a given picture) was much smaller for irregulars than for regulars. Possibly, the greater degree of embodiment of irregular verbs, in combination with their greater semantic density, may not only render them more easy to understand and conceptualize, but may also contribute to their remarkable resistance against regularization.

88

The Mental Lexicon: Core Perspectives

Table 1 Basic verbs for position and motion in English and Dutch Verbs of Position LIE

SIT

STAND

LEAN

HANG

float

hover

LIG

ZIT

STA

leun

HANG

DRIJF

zweef

Verbs of Motion walk

crawl

jump

RUN

SWIM

SINK

DIVE

LOOP

KRUIP

SPRING

ren

ZWEM

ZINK

DUIK

RIDE

FLY

climb

ascend

descend

FALL

RIJD

VLIEG

KLIM

STIJG

daal

VAL

Note. Irregular verbs are shown in upper case, regular verbs in lower case.

We have seen that storage is ubiquitous and not restricted to regulars, that morphological processing is not derivational in the sense that more complex forms are derived on-line from less complex forms, that the paradigmatic relations between complex words across inflection and word formation co-determine lexical processing and generalization, and that form and meaning interact. All these observations run counter to the predictions of fully decompositional generative models.

3. PHONETIC EVIDENCE The evidence considered thus far was gathered from domains that traditionally are at the heart of psycholinguistic investigation. In this section, we consider the consequences of morphological structure for the fine phonetic detail of complex words. Evidence is accumulating that the speech signal itself constitutes an additional source of information about the architecture of the mental lexicon.

3.1. Lexical Competition and the Computation of Fine Phonetic Detail According to the model of Levelt et al. (1999), the process of word form encoding is initiated once a word’s lexeme has been activated. The lexeme sequentially activates the word’s phonemes from first to last. Phonemes are grouped into syllables, and syllables are looked up in a syllabary, which provides gestural scores driving articulation.

Storage and Computation in the Mental Lexicon

89

The WEAVER model implements key insights of mainstream generative phonology. It embodies an important intuition, an axiom driving its design, namely, that once we know which word we need for expressing a given concept, we can select the form of that word without interference from the forms of other irrelevant words. In WEAVER, this intuition is formalized by the restrictions that the only way a lexeme can be activated is through its own lemma, and that only the lemma selected from the set of candidate lemmas will activate its lexeme. As a consequence, lexemes are viewed as highly encapsulated representations that would not enter into competition with each other. Studies investigating speech errors (e.g., Sevald et al., 1995; Dell et al., 1999) have long suggested that word forms enter into a process of lexical competition during speech production. More recently, advances in laboratory phonology and phonetics have provided further evidence for lexical competition during word form encoding. Recall that in auditory comprehension, the cohort of a word’s lexical competitors is gradually reduced as acoustic information unfolds over time (Marslen-Wilson, 1996; Marslen-Wilson and Welsh, 1978). Several phonetic studies provide evidence that a similar competition process characterizes speech production. Van Son and Pols (2003) observed that the fine phonetic detail of a given segment in the word reflects the information load of that segment. In their study of Dutch, segments that contributed more to reducing the cohort were pronounced with longer acoustic durations and with increased articulatory effort, quantified by means of the spectral center of gravity (see also Van Son and Van Santen, 2005). The measure that Van Son and Pols used for gauging a segment’s information load, I L , is the negative log to base 2 of a ratio of two probabilities. The first probability, p+, estimates the joint probability (by means of token counts) of all words that begin with the same sequence of segments up to and including the target segment. For the [I] in sit, all words beginning with [sI] are taken into account. The second probability, p– , estimates, again by means of token counts, the joint probability of all words beginning with the same sequence of segments up to but not including the target segment. For the [I] in sit, this second probability considers all words that begin with [s]. Thus, a segment’s information load is defined as

I L = – log 2 (p + /p – ).

(2)

What makes I L interesting is that it gauges the extent to which the [I] in sit contributes to reducing the uncertainty in the cohort. Before the [I] is considered, the amount of information in the cohort is – log (p – ) (Shannon and Weaver, 1949). Once the [I] is considered, however, the amount of information in the cohort increases to – log (p + ). Since – log (a/b) = log (b) – log (a), it is easy to see that I L quantifies the extent to which the set of competitors before the [I] comes in (characterized by the larger and hence less information rich joint probability p–) is reduced by the [I] (resulting in the smaller more informative joint probability p + ).

90

The Mental Lexicon: Core Perspectives

Van Son and Pols (2003) actually used a more complex estimate of I L that was weighted for the contextual likelihood of the word containing the target segment, i.e., the word for which duration and spectral center of gravity were measured (sit). However, how to optimally estimate the cohort probabilities p + and p– requires further research. For instance, Kuperman et al. (2007), following up on Van Son and Pols (2003), observed improved correlations when estimating p + and p– not with token-based counts but instead with type counts. Furthermore, this study investigated the predictivity of I L as defined in (2) for the divergence of phonemes from their mean durations, rather than the raw durations of these phonemes. Although the two studies are not directly comparable, their results are consistent: It is clear that a segment with greater lexical information tends to have a longer duration. Furthermore, the acoustic signal as produced by the speaker is advantageous for the listener, as segments which are more important for distinguishing the target word from its competitors in the auditory cohort are more prominently present in the speech signal. This finding raises the question of whether the speaker is modulating the fine phonetic detail of a word’s segments explicitly with the purpose of facilitating comprehension for the listener, in line with Lindblom’s hypo- and hyper-articulation theory (Lindblom, 1990). Although it makes sense to assume that the communicative efficiency of speech is enhanced by the hyperarticulation of informationally more salient segments and the hypoarticulation of informationally more redundant segments (Aylett and Turk, 2004; Van Son and Van Santen, 2005), it seems unlikely that this efficiency is due to some conscious or even unconscious effort on the part of the speaker to accommodate to the listener. Speakers can adjust their speech depending on their audience, the acoustics of their environment, whether or not they are using a telephone, etc. But purposeful modulation of phonetic detail at the fine-grained level that is at issue here seems highly unlikely. For instance, Bard et al. (2000) observed that clarity of spontaneous speech was predictable from the speaker’s knowledge, and not from the listener’s knowledge. Furthermore, Kuperman et al. (2007) observed the effect of informational redundancy on the details of acoustic durations within a single register, readaloud speech from the library for the blind as available in the spoken Dutch corpus. Although the speakers sampled in this corpus produced carefully articulated speech modulated to fit the needs of their intended audience, informational redundancy still emerges as an independent predictor for phonetic detail within this register. Trón (2006) likewise points out that it is unlikely that the modulation of acoustic duration by previous mentions of a word in the discourse involves adaptation of the speaker to the needs of the listener. Although it is logically possible that speakers purposefully adjust to their listeners, it is logically equally possible that fine phonetic detail is the straightforward consequence of the organization of the mental lexicon. Instead of repeating traditional explanations based on the interaction between speaker and listener, we therefore explore the viability of explanations based on what we know about lexical access. We know from research on auditory word recognition that the incoming speech signal is matched incrementally against a pool of lexical candidates that is winnowed down as more acoustic information becomes available. According to the theory of Levelt et al. (1999), word

Storage and Computation in the Mental Lexicon

91

form encoding in speech production would require a fundamentally different architecture, with word form selection being driven by semantics, and proceeding without lexical competition. However, consider the possibility that word form encoding in production makes use of the same phonological memory that is addressed in auditory comprehension, and that accessing this phonological memory always involves a probabilistic process of lexical competition during which the target word is singled out from its phonological neighbors. In comprehension, the greater amount of fine phonetic detail in the acoustic signal (in the sense of a strengthened articulatory realization) that is present for more discriminative segments allows the listener to distinguish the carrier words of these segments more rapidly and reliably from their lexical competitors. In production, by contrast, the speaker has to retrieve a representation from phonological memory. During this retrieval process, lexical competitors are co-activated. The co-activation of these competitors seems to come with two benefits. First, as suggested by naming experiments reported by Vitevitch (2002), phonological neighbors appear to gang up to facilitate production in a morphologically simple language like English (but see Vitevitch and Stamer, 2006, for the opposite effect in a morphologically rich language, Spanish). Second, a greater neighborhood density also has been observed to correlate in English with strengthened phonetic detail. Wright (2004) and Munson and Solomon (2004) reported that words with a high frequency and low neighborhood density (easy words) were articulated with more centralized vowels than words with a low frequency and a high neighborhood density (difficult words). Scarborough (2004) observed that words with a low frequency and a high neighborhood density were characterized by higher degrees of coarticulation. One type of coarticulation that she studied was nasal coarticulation, which concerns the extent to which the vowel in a word like band is nasalized in anticipation of the following nasal. Scarborough also measured vowel-to-vowel coarticulation, i.e., the extent to which the first (or second) vowel affects the location in F1 – F2 acoustic space of the second (or first) vowel. What she found was that more coarticulation takes place for words with more neighbors. Apparently, once in the course of lexical competition the phonetic characteristics of a segment have been highly activated, these characteristics are not easily de-activated and may spill over to neighboring segments with which they are compatible. Interestingly, the effect of neighborhood density on coarticulation also emerged in Scarborough’s experiments for nonwords. This indicates that these effects do not hinge on phonetically rich stored representations, but emerge during lexical competition. These studies suggest that hyperarticulation is part and parcel of increased lexical competition. The more intense the process of lexical competition is, the more the unique properties of a word become relevant for distinguishing it from its competitors. As a consequence, greater lexical competition results in superior and more detailed lexical activation. In short, the corollary of increased competition is greater articulatory precision. The results obtained by Van Son and Pols (2003) and Kuperman et al. (2007) add a temporal perspective to the consequences of neighborhood structure for articulation reported by Wright (2004) and Munson and Solomon (2004). Standard definitions of a word’s

92

The Mental Lexicon: Core Perspectives

neighbors are string-based, and consider those words as competitors that differ with respect to a single segment, unweighted for its position in the string. The measure of lexical information studied by Van Son and Pols, by contrast, taps into the temporal dynamics of lexical competition by gauging the extent to which a given segment succeeds in disqualifying irrelevant competitors that up till then were viable alternatives. Further evidence for sequentiality in speech production is provided by Sevald and Dell (1994), who observed slowed production for sequences of words with discrepant initial segments (initial neighbors) compared to words with discrepant final segments (final neighbors). Their results suggest that the position of the segment that is exchanged to obtain a neighbor is crucial for understanding word form encoding in speech production. (For discussion of the vulnerability of initial segments against the backdrop of the phenomenon of prosodic strengthening, see Keating, 2006.) Further evidence for the relevance of the position at which words differ from their neighbors, i.e., the position at which the competition is focused, pertains to morphologically complex words. Bien et al. (2006) calculated separate counts of the numbers of neighbors at the initial, the second, and the third position of the stems of derived and inflected words. In parallel, Bien also considered the entropy of the relative frequencies of the cohort of lexical competitors at these three positions. She studied these predictors in a position-response association task (cf. Cholin et al., 2004; Bien et al., 2005), a naming task that seeks to minimize the effect of comprehension processes in production experiments. Bien observed an inhibitory effect of the neighborhood count for the initial position, and a facilitatory eﬀect of the cohort entropy at the second position. Her results confirm that lexical competition at the initial position slows word form encoding, and add the new finding that lexical competition at the second position facilitates word form encoding. Tabak et al. (2006) observed a similar pattern of results with the standard word naming paradigm, using monomorphemic Dutch verbs. Again the positional neighborhood count at the initial position of the word was inhibitory, whereas the positional count at the second position, and also the summed count of neighbors for later positions in the word, were both facilitatory. The inhibition observed in naming latencies for initial neighbors is consistent with the results reported by Sevald and Dell (1994) for rapid sequence naming. The facilitation at later positions may be the driving factor behind the facilitation reported by Vitevitch (2002) for a non-positional neighborhood count. Assuming that these results are robust and replicable, the hypothesis suggests itself that positional densities might be predictive for the duration with which the corresponding target segments are produced. Preliminary results indicate that indeed a greater density at the initial phoneme gives rise to prolonged acoustic duration of this phoneme. In summary, rich phonetic detail seems to be the by-product, or perhaps even the goal, of lexical competition in speech production. Assuming that replication studies will consolidate these findings, we may speculate that a word’s phonological form is not a static representation (as a string in computer memory) nor a simple piece of code that sequentially

Storage and Computation in the Mental Lexicon

93

triggers the activation of an otherwise static sequence of segments, as in the WEAVER model. Instead, a word’s phonological form may be the outcome of a dynamic competition process that is biased either by acoustic input (in comprehension) or by meaning (in production). In other words, the morphs that, from a morphologist’s perspective, seem to be the basic units stored in memory are themselves the resultant of a dynamic selection process. 3.2. Syllabification and Fine Phonetic Detail Syllabification is a well-studied phonological process that affects morphologically complex words, where it may assign stem-final consonants to the onset of a new syllable containing a vowel-initial affix as rime. The ensuing changes in fine phonetic detail have surprising consequences for the listener. When the comparative suffix -er is added to an adjectival base, our orthographic conventions suggest that the comparative form is simply a longer continuation of the adjectival base, and that morphological information becomes available to the listener once the first phoneme of the suffix has been heard. However, syllabification of warmer as war-mer has far-reaching consequences for the fine phonetic detail of the stem (Lehiste, 1972). Kemps et al. (2005a) and Kemps et al. (2005b) showed that listeners are highly sensitive to the durational differences between a base word by itself (warm) and the base word as it occurs in an inflected or derived word (warmer). In a word like warmer, the vowel and the coda of warm tend to be articulated with shorter durations than when warm is a word by itself. Even though there are tremendous differences in speech rate both between and within speakers, listeners nevertheless have been found to be highly sensitive to these durational differences. For instance, when the Dutch singular kant (side) is spliced out of its plural kanten and presented to Dutch listeners in a number decision task, response latencies to the spliced-out singular were longer than for normal singulars. Moreover, the shorter the spliced-out singular was compared to its normal counterpart, the longer response latencies were found to be. This prosodic mismatch effect was observed both for words and (phonotactically legal) pseudowords, which shows that we are dealing with general inferential processes that are not driven primarily by word-specific articulatory information in memory. A key challenge in this line of research is to establish whether these findings generalize from laboratory speech to spontaneous conversational speech. The mere fact that listeners are able to make use of these subtle cues already suggests that they must be functional as well in normal language use. Work reviewed by Hawkins (2003) points in the same direction.

94

The Mental Lexicon: Core Perspectives

3.3. Morphological Effects on Fine Phonetic Detail Fine phonetic detail is predictable not only from the dynamics of lexical competition and from general syllabification processes, but also from a word’s morphological properties. We first consider syntagmatic properties, and then discuss paradigmatic properties. According to Levelt et al. (1999), the word frequency effect arises at the level of the lexeme. The higher the frequency of a morpheme is, the faster its lexeme is assumed to initiate activation of its phonemes. This model predicts that frequency effects for complex words do not arise. We have already reviewed ample chronometric evidence that falsifies this prediction. Further evidence arguing against this decompositional approach to speech production is provided by a detailed examination of the fine phonetic detail of complex words. Pluymaekers et al. (2005b) studied the acoustic durations of four Dutch derivational affixes in spontaneous conversations. For three out of four affixes, Pluymaekers documented that the acoustic duration of the derivational affix tended to decrease with increasing word frequency. In a laboratory study eliciting complex words with these same derivational affixes across three speech rates, Pluymaekers et al. (2006) observed the very same negative correlation between frequency and acoustic duration, which now was robust for all four affixes studied. These results, which are in line with the data reported by Jurafsky et al. (2001) for monomorphemic words, show that stored word-specific information co-determines a word’s acoustic realization. That the amount of effort invested in articulation is inversely related to the frequency of the complex words has also been demonstrated for assimilation in compounds by Ernestus et al. (2006). Higher-frequency compounds tended to undergo more assimilation at the constituent boundary (e.g., tb in wet + boek, “law book”, assimilating to db) than lowfrequency compounds. Paradigmatic relations have also been shown to be predictive for fine phonetic detail. Hay (2001) observed that t-deletion in words like swiftly is more likely than in words like softly. The likelihood of deletion turns out to be positively correlated with the ratio of the frequency of the complex word to that of its base. The greater the extent to which the complex word is independent of its stem, the lesser the functionality of the low-probability diphone tl becomes for morphological segmentation, and hence the greater the likelihood that this cluster can be simplified without loss of comprehension. A second example of the reflection of paradigmatic structure in fine phonetic detail concerns the duration of the interfix in Dutch compounds. Krott et al. (2001) showed that the probability distribution of the interfixes in the set of compounds sharing the left immediate constituent is crucial for understanding the otherwise enigmatic selection of the interfix in Dutch novel and existing (Krott et al., 2004) compounds. The greater the likelihood of a given interfix in this probability distribution (henceforth its bias), the greater the likelihood is that it is used, and the shorter its required processing time. Kuperman et al. (2007) measured the acoustic duration of the interfixes -s and -en in the spoken Dutch corpus (Oostdijk, 2002;

Storage and Computation in the Mental Lexicon

95

Oostdijk et al., 2002), using the read aloud speech from the subcorpus “library of the blind”. Kuperman considered many other variables along with this paradigmatic bias in the statistical analysis. One of these control variables was the abovementioned I L measure of Van Son and Pols (2003), which showed that if the interfix conveyed more information with respect to its acoustic cohort, it was realized with longer durations. Interestingly, Kuperman observed that independently of the other predictors and independently of the I L measure, interfixes with a greater bias were pronounced with greater acoustic durations. Apparently, a greater bias is not a measure of greater informational redundancy, but a measure of the amount of paradigmatic support for an interfix: Interfixes with greater paradigmatic support are articulated more robustly. A final example of the consequences of paradigmatic structure for morphological processing, but now for comprehension, concerns the phenomenon known as final devoicing in Dutch. For a subset of Dutch stems ending in an obstruent, this obstruent alternates between voiceless (when syllable-final) and voiced (when syllable-initial), compare raaf (raven), plural ra-ven. Ernestus and Baayen (2006b) calculated the probability in a word’s inflectional and derivational paradigms that its obstruent is voiced, its paradigmatic likelihood of voicing. They presented word forms in which the final obstruent is voiceless or nearly voiceless in an auditory lexical decision experiment. Words with a high paradigmatic likelihood of voicing, i.e., words that are predominantly used with inflectional forms that have the obstruent voiced, elicited longer reaction times. This shows that the distribution of voicing within the paradigm codetermines the listener’s expectations. When these expectations are violated, responses are slowed. Interestingly, forms realized with residual voicing elicited longer latencies than words with completely voiceless final obstruents. The fine phonetic detail of residual voicing in the acoustic signal (which itself probably arises due to phonological paradigmatic analogy; Ernestus and Baayen, 2003, 2006a) is detected by the listener, and decreases the listener’s estimate that a voiceless variant is being heard. As a consequence, the response to the voiceless variant is slowed. We have seen that lexical competition among monomorphemic words gives rise to enhanced articulatory detail. This suggests that a word’s canonical form is dynamically computed during the lexical competition process, and that this competition process causes a word’s phonetic form to be optimally distinct from its nearest phonological and morphological neighbors. Dynamic computation likewise takes place at the paradigmatic level, across sets of morphological neighbors instead of across sets of phonological neighbors. For both monomorphemic and complex words, computations are involved that crucially involve a word’s own specific neighbors. Since decompositional models only have abstract rules expressing global syntagmatic generalizations across the lexicon at their disposal, they are severely challenged by what is now known about the consequences of phonological and morphological neighborhoods for the articulation of fine phonetic detail. In addition to dynamic computation, word-specific biases for phonetic detail may be at work. Jurafsky et al. (2002) argue that English of is realized differently in the Switchboard corpus according to whether it expresses the genitive, the partitive, or a complement. Kemps

96

The Mental Lexicon: Core Perspectives

et al. (2005b) discuss the possibility that the odds of encountering the base versus a bisyllabic derivative may further help listeners to optimize their responses in the number decision task. Gahl (2006) reports evidence that homophones such as thyme and time differ systematically in duration, with the more intensively used alternative receiving, on average, the shorter acoustic realization. Especially with respect to extremely reduced words, word-specific variation can be quite extensive. Dutch natuurlijk (nature-like, i.e., of course), for instance, is attested with reduced forms ranging from ntuurlijk, tuurlijk, ntuuk, tuuk to tk (Ernestus, 2000). The choice of a given reduced form depends in part on the syntactic and discourse context (Plug, 2005), and in part on social and geographical variables. Keune et al. (2005) document, for instance, that severe reduction in colloquial Dutch of high-frequency words such as natuurlijk and eigenlijk (own-like, i.e., actually) to tuuk and eigk is more common in the Netherlands than in Flanders. The syntagmatic and paradigmatic lexical forces affecting the fine phonetic detail of a word are themselves part of a larger set of forces that co-determine the details of articulation, such as the probability of a word given the preceding or following word (Bell et al., 2003; Gregory et al., 1999; Jurafsky et al., 2001), the probability of the syntactic construction in which a word is used (Gahl and Garnsey, 2004, 2006), and the recency with which a word has been heard in (conversational) discourse (Fowler and Housum, 1987; Fowler, 1988; Hawkins and Warren, 1994; Bard et al., 2000; Pluymaekers et al., 2005a; Trón, 2006). From this overall perspective, it seems that fine phonetic detail at lexical and sublexical levels realizes a kind of informational prosody that reflects probabilistic generalizations at all levels of linguistic structure, complementary to (or possibly subsuming) classical prosodic structure and its consequences for articulatory realization (see, e.g., Keating, 2006).

4. TOWARDS A NEW CLASS OF THEORIES The results reviewed in the previous sections show that the mathematics of formal languages do not provide an adequate metaphor for understanding the mental lexicon. One of the key issues for current research is what kind of formal frameworks might then be worth pursuing instead. An alternative approach that has been studied intensively but with hotly debated success is connectionist modeling (e.g., Rumelhart and McClelland, 1986; Seidenberg and Gonnerman, 2000; Pinker and Ullman, 2002a,b; McClelland and Patterson, 2002a,b). Connectionist networks have the advantage that they can account for graded, probabilistic phenomena. But they also have their share of disadvantages. One such a disadvantage is the merging of rules and representations. This might seem a step forward compared to the standard Von Neumann computer architecture that is the source of inspiration for symbolic models. However, rules and representations might have distinguishable neural substrates, as argued by Ullman (2001, 2004). According to Ullman, symbolic rules reside in a procedural

Storage and Computation in the Mental Lexicon

97

memory system, and monomorphemic words and formatives in a declarative memory system. Evidence for such a division of labor for regular and irregular verbs comes from studies such as Jaeger et al. (1996; positron emission tomography) and Beretta et al. (2003; functional magnetic resonance imaging). Interpretation of the reported differential activation of brain regions for regulars and irregulars is not straightforward, unfortunately. On the one hand, functional differentiation can take place in neural networks. Furthermore, present-to-past naming as used by Jaeger et al. (1996) is a task that, as discussed above, induces coactivation of inflectional variants that does not take place in normal speech. In addition, the experimental materials of regulars and irregulars are not appropriately controlled for semantic density and paradigm complexity (Baayen and Moscoso del Prado Martín, 2005). Nevertheless, the distinction between procedural and declarative knowledge is an important one and the available evidence reviewed by Ullman seems compelling. Given the data reviewed in the preceding sections, it is clear that Ullman’s straightforward traditional divide between regulars (processed by procedural memory) and irregulars (stored in declarative memory) must be too simplistic. Several modifications suggest themselves. Recall that the word frequency effect for very low-frequency complex words suggests that combinatorial probabilities may be at issue, probably in combination with phonological memory traces. It is conceivable that such combinatorial probabilities are evaluated in procedural memory as memories of previous assembly and decomposition allowing subsequent faster assembly and decomposition, whereas the phonological memory traces reside in declarative memory. Another possibility is that procedural memory is responsible for analogical generalization over lexical exemplars in declarative memory. Yet another alternative interpretation is that the weaker embodiment of regular verbs compared to irregulars goes hand in hand with greater multimodal computational demands across memory systems in the brain for conceptual interpretation, and that it is these additional costs that show up in brain imaging studies. A second disadvantage is that the connectionist networks in common use are neurologically implausible. This is a key criticism made by Hawkins and Blakeslee (2004), in a study that researches the consequences of the biological structure of the neocortex for intelligent computation (see also Miller, 2006). Interestingly, Hawkins and George (2006) provide an outline of a newly implemented technology, Hierarchical Temporal Memory (HTM) that is designed to replicate the general structural and algorithmic properties of the neocortex. HTMS consist of a hierarchy of memory nodes. Each node is itself a network that learns causes from its child nodes and forms beliefs that it propagates to its parent nodes. Going from low-level sensory nodes to higher-level nodes, the HTM performs as a classifier, coalescing a series of input patterns into a relatively stable output pattern. Conversely, a stable pattern at the top of the hierarchy can unfold into a complex temporal pattern at the bottom of the hierarchy. HTM memory is not designed specifically for language, but it promises to be a great tool for the computational modeling of the mental lexicon. HTM as described by Hawkins and George (2006) is probably best seen as a computational model of declarative memory,

98

The Mental Lexicon: Core Perspectives

albeit a memory system that is intrinsically predictive. In this sense, it is fundamentally different from a declarative memory conceived of as a static store of bits and bytes in the memory registers on a Von Neumann computer. Therefore, HTM bears the promise of being able to deal in a natural way with graded linguistic phenomena such as fine phonetic detail and the semi-morphology of phonaesthemes (Bergen, 2004). At the same time, rules and representations are not merged a priori. Nodes in the HTM memory can represent individuals and effectively function as symbols, albeit as symbols in a system with “rules” that are inherently analogical and probabilistic in nature. HTM memory may also help resolve the problem of positional encoding that is rampant in analogical (Skousen, 1989) and machine learning (Daelemans and Van den Bosch, 2005) models as well as in connectionist networks. The alternative developed by Albright and Hayes (2003) is attractive in that it does not depend on positional encoding, but the price paid is a highly complex system of discrete rules. Since HTM memory is designed explicitly to match the spatial and temporal hierarchical structure of the real word, it may be able to detect structure in time without depending on predefined slots for the constituents of a linguistic unit. Whereas the mathematics of formal languages has been a key source of inspiration for morphological theory and models of the mental lexicon, I expect new advances at the intersection of statistics, information science and the neurosciences such as Hierarchical Temporal Memory (and models using related techniques such as Dynamic Bayesian Networks) to constitute an important source of inspiration for research on the mental lexicon during the coming years. As a consequence, not only the controversy between connectionist and symbolic approaches to the mental lexicon, but also the controversy between abstractionist and exemplar-based approaches may well be resolved in harmony. Any current exemplar-based machine-learning algorithm must make use of smart economical storage; otherwise the system will grind to a halt when trying to survey all exemplars in its memory (see, e.g., the IG-tree technology used by TiMBL; Daelemans and Van den Bosch, 2005). On the other hand, the abovementioned frequency effects for fully regular complex words and the effects of inflectional and paradigmatic entropy bear witness to the remarkable sensitivity of lexical memory to very item-specific probabilities. In memory systems such as HTM, such itemspecific probabilities are bound to be captured. On the other hand, such memory systems do not require as a matter of principle that all exemplars, such as the inflectional variants of a Spanish or Georgian verb, or all of a word’s phonetic variants, are represented by individual nodes. Instead, HTM-like lexical memory systems promise to be fully compatible with the dynamic analogy-driven computation of morphologically complex words and their fine phonetic detail.

Storage and Computation in the Mental Lexicon

99

ACKNOWLEDGEMENTS I am indebted to Gonia Jarema, Gary Libben, Mark Pluymaekers, Mirjam Ernestus, Rachel Smith and Susanne Gahl for comments and discussion.

REFERENCES Albright, A. and B. Hayes (2003). Rules vs. analogy in English past tenses: A computational/experimental study. Cognition, 90, 119–161. Alegre, M. and P. Gordon (1999). Frequency eﬀects and the representational status of regular inflections. Journal of Memory and Language, 40, 41–61. Aylett, M. and A. Turk (2004). The smooth signal redundancy hypothesis: a functional explanation for relationships between redundancy, prosodic prominence, and duration in spontaneous speech. Language and Speech, 47, 31–56. Baayen, R., L. Feldman and R. Schreuder (2006). Morphological influences on the recognition of monosyllabic monomorphemic words. Journal of Memory and Language, 53, 496–512. Baayen, R. H., T. Dijkstra and R. Schreuder (1997). Singulars and plurals in Dutch: Evidence for a parallel dual route model. Journal of Memory and Language, 36, 94–117. Baayen, R. H. and F. Moscoso del Prado Martín (2005). Semantic density and past-tense formation in three Germanic languages. Language, 81, 666–698. Baayen, R. H. and R. Schreuder (2000). Towards a psycholinguistic computational model for morphological parsing. Philosophical Transactions of the Royal Society (Series A: Mathematical, Physical and Engineering Sciences), 358, 1–13. Baayen, R. H., R. Schreuder, N. H. De Jong and A. Krott (2002). Dutch inflection: the rules that prove the exception. In: Storage and Computation in the Language Faculty (S. Nooteboom, F. Weerman and F. Wijnen, eds.), pp. 61–92. Kluwer Academic Publishers, Dordrecht. Bard, E., A. Anderson, C. Sotillo, M. Aylett, G. Doherty-Sneddon and A. Newlands (2000). Controlling the Intelligibility of Referring Expressions in Dialogue. Journal of Memory and Language, 42, 1–22. Barsalou, L. W. (2003). Situated simulation in the human conceptual system. Language and Cognitive Processes, 18, 513–562. Bell, A., D. Jurafsky, E. Fosler-Lussier, C. Girand, M. Gregory and D. Gildea (2003). Eﬀects of disfluencies, predictability, and utterance position on word form variation in English conversation. Journal of the Acoustical Society of America, 113, 1001–1024. Beretta, A., C. Campbell, T. Carr, J. Huang, L. M. Schmitt, K. Christianson and Y. Cao (2003). An ER-fMRI investigation of morphological inflection in German reveals that the brain makes a distinction between regular and irregular forms. Brain and Language, 85, 67–92. Bergen, B. K. (2004). The psychological reality of phonaesthemes. Language, 80, 290–311. Bien, H., W. Levelt and R. H. Baayen (2005). Frequency eﬀects in compound production. Proceedings of the National Academy of Sciences of the USA, 102, 17876–17881. Bien, H., W. Levelt and R. H. Baayen (2006). Frequency effects in the production of derivations and inflections. Manuscript in preparation, Max Planck Institute for Psycholinguistics, Nijmegen. Bloch, B. (1947). English verb inflection. Language, 23, 399–418.

100

The Mental Lexicon: Core Perspectives

Bloomfield, L. (1933). Language. Allen and Unwin, London. Bod, R. (2006). Exemplar-based syntax: How to get productivity from examples. The Linguistic Review, 23, in press. Bybee, J. L. and C. L. Moder (1983). Morphological classes as natural categories. Language, 59, 251– 270. Bybee, J. L. and D. I. Slobin (1982). Rules and schemas in the development and use of the english past tense. Language, 58, 265–289. Cholin, J., N. O. Schiller and W. J. M. Levelt (2004). The preparation of syllables in speech production. Journal of Memory and Language, 20, 47–61. Chomsky, N. and M. Halle (1968). The sound pattern of English. Harper and Row, New York. Daelemans, W. and A. Van den Bosch (2005). Memory-based language processing. Cambridge University Press, Cambridge. De Vaan, L., R. Schreuder and R. H. Baayen (2007). Regular morphologically complex neologisms leave detectable traces in the mental lexicon. The Mental Lexicon, 2, in press. Dell, G., F. Chang and Z. Griﬃn (1999). Connectionist Models of Language Production: Lexical Access and Grammatical Encoding. Cognitive Science, 23, 517–542. Ernestus, M. (2000). Voice assimilation and segment reduction in casual Dutch. LOT, Utrecht. Ernestus, M. and R. H. Baayen (2006). The functionality of incomplete neutralization in Dutch. The case of past-tense formation. Laboratory Phonology, 8, 27–49. Ernestus, M. and R. H. Baayen (2007). Paradigmatic eﬀects in auditory word recognition: The case of alternating voice in Dutch. Language and Cognitive Processes, 22, 1-24. Ernestus, M. and R. H. Baayen (2003). Predicting the unpredictable: Interpreting neutralized segments in Dutch. Language, 79, 5–38. Ernestus, M., M. Lahey, F. Verhees and R. H. Baayen (2006). Lexical frequency and voice assimilation. Journal of the Acoustical Society of America, 120, 1040–1051. Feldman, J. (Ed.). (2006). From molecule to metaphor. A neural theory of language. The MIT Press, Cambridge, MA. Fowler, C. (1988). Differential shortening of repeated content words produced in various communicative contexts. Language and Speech, 31, 307–317. Fowler, C. and J. Housum (1987). Talkers’ Signalling of “New” and ”Old” Words in Speech and Listeners’ Perception and Use of the Distinction. Journal of Memory and Language, 26, 489–504. Gahl, S. (2006). Is frequency a property of phonological forms? Evidence from spontaneous speech. Paper presented at the 19th Annual CUNY Conference on Human Sentence Processing, New York City. Gahl, S. and S. Garnsey (2004). Knowledge of grammar, knowledge of usage: Syntactic probabilities affect pronunciation variation. Language, 80, 748–774. Gahl, S. and S. Garnsey (2006). Knowledge of grammar includes knowledge of syntactic probabilities. Language, 82, 405–410. Gregory, M., W. Raymond, A. Bell, E. Fosler-Lussier and D. Jurafsky (1999). The effects of collocational strength and contextual predictability in lexical production. CLS, 35, 151–166. Halle, M. and A. Marantz (1993). Distributed morphology and the pieces of inflection. In: The View from Building 20: Essays in Linguistics in Honor of Sylvain Bromberger (K. Hale and S. J. Keyser, eds.), Vol. 24 of Current Studies in Linguistics, pp. 111–176. MIT Press, Cambridge, Mass. Hawkins, J. and S. Blakeslee (2004). On intelligence. Henry Holt and Company, New York.

Storage and Computation in the Mental Lexicon

101

Hawkins, J. and D. George (2006). Hierarchical temporal memory. Concepts, theory and terminology. Numenta Technology, http://www.numenta.com/technology.php. Hawkins, S. (2003). Roles and representations of systematic fine phonetic detail in speech understanding. Journal of Phonetics, 31, 373–405. Hawkins, S. and P. Warren (1994). Phonetic influences on the intelligibility of conversational speech. Journal of Phonetics, 22, 493–511. Hay, J. (2001). Lexical frequency in morphology: Is everything relative? Linguistics, 39, 1041–1070. Jaeger, J. J., A. H. Lockwood, D. L. Kemmerrer, R. D. Van Valin and B. W. Murphy (1996). A positron emission tomographic study of regular and irregular verb morphology in English. Language, 72, 451–497. Joanisse, M. F. and M. S. Seidenberg (1999). Impairments in verb morphology after brain injury: a connectionist model. Proceedings of the National Academy of Sciences, 96, 7592–7597. Jurafsky, D., A. Bell, M. Gregory and W. Raymond (2001). Probabilistic relations between words: Evidence from reduction in lexical production. In: Frequency and the emergence of linguistic structure (J. Bybee and P. Hopper, eds.), pp. 229–254. Benjamins, Amsterdam. Jurafsky, D., A. Bell and C. Gyrand (2002). The role of the lemma in form variation. In: Papers in Laboratory Phonology VII (C. Gussenhoven and N. Warner, eds.), pp. 1–34. Mouton de Gruyter, Berlin/New York. Keating, P. A. (2006). Phonetic encoding of prosodic structure. In: Speech production: Models, phonetic processes, and techniques (J. Harrington and M. Tabain, eds.), pp. 167– 186. Psychology Press, New York and Hove. Kemps, R., M. Ernestus, R. Schreuder and R. H. Baayen (2005a). Prosodic cues for morphological complexity: The case of Dutch noun plurals. Memory and Cognition, 33, 430–446. Kemps, R., L. Wurm, M. Ernestus, R. Schreuder and R. H. Baayen (2005b). Prosodic cues for morphological complexity in Dutch and English. Language and Cognitive Processes, 20, 43–73. Keune, K., M. Ernestus, R. Van Hout and R. H. Baayen (2005). Social, geographical, and register variation in Dutch: From written ‘mogelijk’ to spoken ‘mok’. Corpus Linguistics and Linguistic Theory, 1, 183–223. Kostić, A., T. Marković and A. Baucal (2003). Inflectional morphology and word meaning: Orthogonal or co-implicative domains? In: Morphological Structure in Language Processing (R. H. Baayen and R. Schreuder, eds.), pp. 1–44. Mouton de Gruyter, Berlin. Krott, A., R. H. Baayen and R. Schreuder (2001). Analogy in morphology: modeling the choice of linking morphemes in Dutch. Linguistics, 39, 51–93. Krott, A., P. Hagoort and R. H. Baayen (2004). Sublexical units and supralexical combinatorics in the processing of interfixed Dutch compounds. Language and Cognitive Processes, 19, 453–471. Kuperman, V., M. Pluymaekers, M. Ernestus and R. H. Baayen (2007). Morphological predictability and acoustic salience of interfixes in Dutch compounds. Acoustical Society of America, in press. Lehiste, I. (1972). The timing of utterances and linguistic boundaries. JASA, 51, 2018– 2024. Levelt, W. J. M., A. Roelofs and A. S. Meyer (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 1–38. Lieber, R. (1980). On the organization of the lexicon. Ph. D. thesis, MIT, Cambridge. Lindblom, B. (1990). Explaining phonetic variation: A sketch of the H & H theory. In: Speech production and speech modeling (W. Hardcastle and A. Marchal, eds.), pp. 403-440. Kluwer, Dordrecht.

102

The Mental Lexicon: Core Perspectives

MacWhinney, B. and J. Leinbach (1991). Implementations are not conceptualizations: revising the verb learning model. Cognition, 40, 121–157. Marslen-Wilson, W. D. (1996). Function and process in spoken word recognition. In: Attention and performance: Control of language processes. Vol. X. pp. 125–150. Lawrence Erlbaum Associates, Hillsdale, NJ. Marslen-Wilson, W. D. and A. Welsh (1978). Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology, 10, 29–63. Matthews, P. H. (1974). Morphology. An Introduction to the Theory of Word Structure. Cambridge University Press, London. McCarthy, J. J. and A. Prince (1993). Generalized alignment. In: Yearbook of Morphology (G. E. Booij and J. van Marle, eds.), pp. 79–154. Kluwer Academic Publishers, Dordrecht. McClelland, J. L. and K. Patterson (2002a). Rules or connections in past-tense inflections: what does the evidence rule out?. Trends in Cognitive Sciences, 6, 465–472. McClelland, J. L. and K. Patterson (2002b). ‘Words or Rules’ cannot exploit the regularity in exceptions: Reply to Pinker and Ullman. Trends in Cognitive Sciences, 6, 464– 465. Miller, G. (2006). An enterprising approach to brain science. Science, 324, 76–77. Moscoso del Prado Martìn, F., R. Bertram, T. Höikiö, R. Schreuder and R. H. Baayen (2004a). Morphological family size in a morphologically rich language: The case of Finnish compared to Dutch and Hebrew. Journal of Experimental Psychology: Learning, Memory and Cognition, 30, 1271–1278. Moscoso del Prado Martín, F., A. Kostić and R. H. Baayen (2004b). Putting the bits together: An information theoretical perspective on morphological processing. Cognition, 94, 1–18. Munson, B. and N. P. Solomon (2004). The effects of phonological neighborhood density on vowel articulation. Journal of Speech, Language, and Hearing Research, 47, 1048–1058. Oostdijk, N. (2002). The design of the Spoken Dutch Corpus. In: New Frontiers of Corpus Research (P. Peters, P. Collins and A. Smith, eds.), pp. 105–112. Rodopi, Amsterdam. Oostdijk, N., W. Goedertier, F. Van Eynde, L. Boves, J. Martens, M. Moortgat and R. H. Baayen (2002). Experiences from the Spoken Dutch Corpus Project. In: Proceedings of the third International Conference on Language Resources and Evaluation (M. Gonz ez Rodriguez and C. Paz Su ez Araujo, eds.), pp. 340–347. ELRA. Patterson, K., M. Lambon Ralph, J. Hodges and J. McClelland (2001a). Deficits in irregular past-tense verb morphology associated with degraded semantic knowledge. Neuropsychologia, 39, 709–724. Patterson, K., M. A. Lambon Ralph, J. R. Hodges and J. L. McClelland (2001b). Deficits in irregular past-tense verb morphology associated with degraded semantic knowledge. Neuropsycologia, 39, 709–724. Pinker, S. (1991). Rules of language. Science, 153, 530–535. Pinker, S. (1997). Words and rules in the human brain. Nature, 387, 547–548. Pinker, S. (1999). Words and Rules: The Ingredients of Language. Weidenfeld and Nicolson, London. Pinker, S. and A. Prince (1988). On language and connectionism. Cognition, 28, 73–193. Pinker, S. and A. Prince (1994). Regular and irregular morphology and the psychological status of rules of grammar. In: The Reality of Linguistic Rules (S. Lima, R. Corrigan and G. Iverson, eds.), pp. 353–388. John Benjamins, Amsterdam. Pinker, S. and M. Ullman (2002a). Combination and structure, not gradedness, is the issue. Trends in Cognitive Sciences, 6, 472–474.

Storage and Computation in the Mental Lexicon

103

Pinker, S. and M. Ullman (2002b). The past and future of the past tense. Trends in the Cognitive Sciences, 6, 456–462. Plug, L. (2005). From words to actions: The phonetics of ’eigenlijk’ in two communicative contexts. Phonetica, 62, 131–145. Plunkett, K. and V. Marchman (1991). U-shaped learning and frequency effects in a multilayered perceptron: implications for child language acquisition. Cognition, 38, 1–60. Pluymaekers, M., M. Ernestus and R. H. Baayen (2005a). Articulatory planning is continuous and sensitive to informational redundancy. Phonetica, 62, 146–159. Pluymaekers, M., M. Ernestus and R. H. Baayen (2005b). Frequency and acoustic length: the case of derivational affixes in Dutch. Journal of the Acoustical Society of America, 118, 2561–2569. Pluymaekers, M., M. Ernestus and R. H. Baayen (2006). Effects of word frequency on articulatory durations. Proceedings of Interspeech 2006, ICSLP, pp. 953-956. Roelofs, A. (1996). Serial order in planning the production of successive morphemes of a word. Journal of Memory and Language, 35, 854–876. Roelofs, A. (1997a). Morpheme frequency in speech production: Testing WEAVER. In: Yearbook of Morphology 1996 (G. E. Booij and J. van Marle, eds.), pp. 135–154. Kluwer, Dordrecht. Roelofs, A. (1997b). The WEAVER model of word-form encoding in speech production. Cognition, 64, 249–284. Rumelhart, D. E. and J. L. McClelland (1986). On learning the past tenses of English verbs. In: Parallel Distributed Processing. Explorations in the Microstructure of Cognition. Vol. 2: Psychological and Biological Models. (J. L. McClelland and D. E. Rumelhart, eds.), pp. 216–271. The MIT Press, Cambridge, Mass. Scarborough, R. A. (2004). Coarticulation and the structure of the lexicon. UCLA dissertation. Seidenberg, M. S. and L. M. Gonnerman (2000). Explaining derivational morphology as the convergence of codes. Trends in Cognitive Sciences, 4, 353–361. Selkirk, E. (1984). Phonology and Syntax. The MIT Press, Cambridge. Sereno, J. and A. Jongman (1997). Processing of English inflectional morphology. Memory and Cognition, 25, 425–437. Sevald, A. C. and G. S. Dell (1994). The sequential cuing effect in speech production. Cognition, 53, 91–127. Sevald, A. C., G. S. Dell and J. S. Cole (1995). Syllable Structure in Speech Production: Are Syllables Chunks or Schemas? Journal of Memory and Language, 34, 807–820. Shannon, C. E. and W. Weaver (1949). The Mathematical Theory of Communication. The University of Illinois Press, Urbana. Skousen, R. (1989). Analogical Modeling of Language. Kluwer, Dordrecht. Stemberger, J. P. (2004). Phonological priming and irregular past. Journal of Memory and Language, 50, 82–95. Stemberger, J. P. and C. Middleton (2003). Vowel dominance and morphological processing. Language and Cognitive Processes, 18, 369–404. Tabak, W., R. Schreuder, R. H. Baayen (2005a). Lexical statistics and lexical processing: semantic density, information complexity, sex, and irregularity in Dutch. In: Linguistic Evidence — Empirical, Theoretical, and Computational Perspectives (S. Kepser and M. Reis, eds.), pp. 529– 555. Mouton de Gruyter, Berlin.

104

The Mental Lexicon: Core Perspectives

Tabak, W., R. Schreuder and R. H. Baayen (2005b). The processing of regular and irregular verbs. In: Proceedings of the Interdisciplinary Workshop on the Identification and Representation of Verb Features and Verb Classes, Saarbrücken, pp. 121–126. Tabak, W., R. Schreuder and R. H. Baayen (2006). Nonderivational inflection. Manuscript, Max Planck Institute for Psycholinguistics. Taft, M. (1979). Recognition of affixed words and the word frequency effect. Memory and Cognition, 7, 263–272. Trón, V. (2006). Corpus evidence for a priming account of durational reduction. Paper presented at the 2nd Annual Edinburgh Psycholinguistics postgraduate conference, 2006. Tyler, L., E. Stamatakis, R. Jones, P. Bright, K. Acres and W. Marslen-Wilson (2004). Deficits for semantics and the irregular past tense: A causal relationship? Journal of Cognitive Neuroscience, 16, 1159–1172. Ullman, M. (2001). The declarative/procedural model of lexicon and grammar. Journal of Psycholinguistic Research, 30, 37–69. Ullman, M. (2004). Contributions of memory circuits to language: the declarative/procedural model. Cognition, 92, 231–270. Van Buren, H., W. Tabak, M. Carreiras and R. H. Baayen (2007). Morphological effects in picture naming of Spanish verbs by L1 and L2 speakers. Manuscript in preparation. Van Haeringen, C. B. (1940). De taaie levenskracht van het sterke werkwoord. De Nieuwe Taalgids, 34, 241–255. Van Son, R. and L. Pols (2003). Information Structure and Efficiency in Speech Production. Proceedings of Eurospeech-2003. Geneva, Switzerland, 769–772. Van Son, R. and J. Van Santen (2005). Duration and spectral balance of intervocalic consonants: A case for efficient communication. Speech Communication, 47, 100–123. Vitevitch, M. S. (2002). The influence of phonological similarity neighborhoods on speech production. Journal of Experimental Psychology: Learning, Memory and Cognition, 28, 735–747. Vitevitch, M. S. and M. K. Stamer (2006). The curious case of competition in Spanish speech production. Language and cognitive processes, 21, 760–770. Wright, R. (2004). Factors of lexical competition in vowel articulation. In: Papers in Laboratory Phonology 6 (J. Local and R. Ogden, eds.), pp. 75–87. Cambridge University Press, Cambridge. Wurm, L. H., J. Aycock and R. H. Baayen (2007). Lexical dynamics for low-frequency complex words: A regression study across tasks and modalities. Submitted.

6 GENERATIVE MORPHOLOGY AS PSYCHOLINGUISTICS James Myers, National Chung Cheng University, Min-Hsiung, Taiwan

1. INTRODUCTION The title of this chapter is, of course, deliberately provocative. How can generative linguistics be psycholinguistics? Both sciences test mentalistic hypotheses about language, but their differences go beyond the cliché that linguists study competence while psycholinguists study performance. Instead they represent two different cultures, with linguistics more deductive and rationalist, psychology more inductive and empiricist. Psychologists, who are concerned with “cause and effect” (as Miller, 1990, p. 321, observes), run carefully designed experiments, conduct quantitative analyses, and write papers with explicit method sections. Generative linguists, who (Miller says) prefer “simplifications as explanations,” don’t. And what is “generative morphology”? Isn’t morphology lexical knowledge, and isn’t grammar what permits one to go beyond rote lexical memory? Chomsky (1957) put regular interactions between word and sentence structure into the syntax, not the lexicon, and Chomsky and Halle (1968) put regular interactions between word and sound structure into the phonology; even today, most generative linguists are syntacticians and most of the rest are phonologists. As for those few who specialize in morphology itself, Marantz (1997, p. 202) complains that “when morphologists talk, linguists nap.” Yet my title is no oxymoron. Generative morphology, like the rest of generative linguistics, is indeed psycholinguistics, albeit methodologically sloppy psycholinguistics. Moreover, the lack of canonical pronouncements on what generative morphology is supposed to be is actually fortuitous, since it forces us to think hard about the notion of “generative grammar,” and this is essential if we are to see grammar as psychological and psycholinguistics as grammar-oriented. I start in section 2 by introducing the central notion of competence-performance linking models and explaining their relevance to morphology. Special focus is given to the theory of Distributed Morphology (Halle and Marantz, 1993; Harley and Noyer, 2003), not

105

106

The Mental Lexicon: Core Perspectives

because I have any special fondness or animosity towards it, but because it is arguably the “most generative” of morphological theories. In section 3, I apply competence-performance linking models to a specific morphological question: Is there any difference between affixation and compounding in Chinese, and more to the point, how could we tell? I then analyze three sources of new evidence bearing on the question, each requiring its own linking model: two traditionally “linguistic” (native-speaker judgments and corpus analyses) and one traditionally “psycholinguistic” (reaction times in lexical decision tasks).

2. COMPETENCE AND PERFORMANCE IN MORPHOLOGY Far from defining the border between linguistics and psychology as often assumed, the competence-performance distinction of Chomsky (1965) actually provides the conceptual basis for fitting the study of mental grammar comfortably into the broader study of language in the mind. In this section I show how, focusing on morphology.

2.1. Competence-Performance Linking Models Generative linguistics is given its most succinct definition on page 4 of Chomsky (1965). There Chomsky states that “linguistic theory is mentalistic, since it is concerned with discovering a mental reality underlying actual behavior,” that is, “a description of the ideal speaker-hearer’s intrinsic competence.” Competence is defined as “the speaker-hearer’s knowledge of his language,” that is, mental grammar, as opposed to performance, which is “the actual use of language in concrete situations.” A linguist’s grammar is called “generative” if it is “perfectly explicit — in other words, if it does not rely on the intelligence of the understanding reader but rather provides an explicit analysis of his contribution.” This passage has had a rather controversial history, but if we translate the linguistese into psychologistese, the concepts actually transfer quite well. The most obvious cognate is mentalism; nobody wants to return to the bad old behaviorist days, when the mind was dismissed as unscientific. Less obviously, the “ideal speaker-hearer” (particularly notorious among sociolinguists) also plays a starring role in psycholinguistics: S/he hovers in the Platonic statistical space over the heads of the actual experimental participants, the population mu to their sample x-bars. Similarly, psychologists believe in “universal psychology” just as much as generative linguists believe in “universal grammar”; how else could a study on Chinese lexical access have any bearing on theories of lexical access more generally? Moreover, the goal of “discovering a mental reality underlying actual behavior,” which is at the core of the competence-performance distinction, is what cognitive psychologists aim for every day. Psychologists know not to confuse overt behavior with the underlying mental operations themselves. In the same way, syntacticians know that acceptability judgments, their favorite data source, are partly contaminated by parsing effects (Phillips and Lasnik, 2003), and phonologists know that dictionaries, their favorite data

Generative Morphology as Psycholinguistics

107

source, are partly contaminated by historical accidents (Kenstowicz and Kisseberth, 1979). As Penke and Rosenbach (2004, p. 492) emphasize, “there is no such thing as competence data,” only performance data, and performance data include not just judgments and dictionaries, but also traditionally “psycholinguistic” evidence like “rapidity, correctness, and uniformity of recall and recognition” (Chomsky, 1965, p. 10). It may not always be obvious where to draw the line between competence and performance, but similar demarcation problems plague psychology as well, as with prelexical vs. postlexical stages in lexical access (see 2.3). But what really makes generative linguistics psychology is the notion of generativity itself. Chomsky (1980, p. 48) suggests that (generative) linguistics can be taken as “the study of the computation in language” in Marr’s sense of “computation” (e.g., Marr, 1982): an explicit description of the engineering problem to be solved by the cognitive system. Computation is key because the modern cognitive revolution (as it emerged in the second half of the twentieth century) is defined by it. Without computation, the behaviorists had every right to assert that the black-box nature of the mind prevents scientific study. With it, a tool becomes available for linking the observable world to the hidden contents of the black box. Computation is the “mind-stuff” out of which we can construct inference chains to probe as deeply into the mind as we care to go. I’ll call a cognitive inference chain of this sort a linking model. A particularly elegant type is the additive model of Sternberg (1998), with which one can make justifiable inferences about the order, duration, and modularity of mental processes simply by measuring reaction times in different behavioral tasks. (Of course, there’s no reason to share Sternberg’s assumption that linking models must be linear; see Westbury and Hollis, this volume.) An independently justified linking model gives us just as much confidence about the reality of hidden mental entities as we have about any other natural phenomena that can only be observed indirectly. Linking models need not be fully explicit in order to be useful. The usual strategy in cognitive psychology is to describe whatever is systematic in behavioral evidence, and leave the rest for later. This strategy is necessitated by the complexity of behavior, and it is made rigorous by statistics, another essential tool of the modern cognitive sciences (e.g., in Sternberg’s additive model, modularity is ruled out by a significant interaction). Despite its empiricist connotations, a statistical linking model still permits the rationalist view of grammar as static knowledge, because inferential statistics boils down to correlation, and correlation, as every statistics student knows, isn’t the same as causation. Thus there is no obligation for a linking model to treat grammar as having “real time” influences on behavior. That is, it doesn’t have to be an algorithm, the next level in Marr’s hierarchy, as long as some aspects of behavior are predicted by it. Therefore, generative linguistics is psycholinguistics. At least it is in principle. Due to the rationalist nature of their culture, linguists generally fail to take seriously the empirical challenge of testing mental claims with behavioral data. On the one hand, linguists often treat performance as transparent reflections of competence; even Penke and Rosenbach (2004, p.

108

The Mental Lexicon: Core Perspectives

492) include “grammaticality judgments” (rather than acceptability judgments, i.e., performance data) on their list of “direct evidence” (as if metalinguistic musing is what human grammar is for). On the other hand, when linguists do invoke performance, they use it primarily as a dumping ground for counterevidence to competence claims, though this move can be valid if the performance-based explanation is independently motivated (see 2.2). The failure to respect competence-performance linking models has widespread consequences in generative linguistics. Schütze (1996) surveys the serious difficulties with interpreting informally collected syntactic judgments, and Cowart (1997) shows how proper experimental protocols resolve them. Ohala (1986) diagnoses and advises on similar problems in phonology. This chapter attempts to perform a similar methodological service for the study of morphological competence.

2.2. Morphological Competence What makes a morphological theory generative? If we take “generative” as a synonym for “computational” in Marr’s sense, and then add the rationalist preference for simplicity, the question splits into two: What is the design problem that morphology exists to solve, and what is the simplest way to describe it? The obvious answer to the first question is compositionality: The central morphological fact is that words show sound-meaning regularities. Compositionality demands analysis at the computational level whether decomposition is obligatory at the algorithmic level or merely simulated (see Baayen, this volume). If we then consider the second question, the “most generative” answer must be a theory that is maximally compositional. Of course, several morphologists in the generative tradition have given empirical arguments against extreme compositionality. Aronoff (1976), for instance, points out that speakers know things about words that they cannot have derived from morphological structure alone; thus in English, transmission can refer to a specific bit of automotive hardware, not merely the act of transmitting. Extending the argument, Becker (1993) notes that not only are the meanings of seaman and airman not derived transparently from those of sea, air, and man, but they also share a complex of other notions, such as “travel”: Wordlevel semantic idiosyncrasies can be productive (a point reiterated by Aronoff, this volume). Yet such arguments do not really falsify extreme compositionality at the computational level. Grammar needn’t be given the power to memorize arbitrary meanings for morphologically complex words, since this ability is already sufficiently well explained by the extra-grammatical fact that brains are voracious memorizers: If transmission is treated as a unit at any stage of processing (as what Packard, 2000, calls a psycholinguistic word), it is a potential target for semantic barnacles. Not even syntactic phrases are immune, as idioms demonstrate. Similarly, we need not ascribe the productivity of the seaman schema to

Generative Morphology as Psycholinguistics

109

grammar if we admit that humans also possess the extra-grammatical ability to form analogies (as discussed further in 2.3). For reasons like these, Di Sciullo and Williams (1987) dismiss word-level phenomena as mere performance effects; they call memorized forms “listemes” to give a name to what they don’t care about. This may sound radical to a psycholinguist, but it is justifiable: A linguistic form is not automatically “grammatical” merely by virtue of existing, since it may be an exception, and an unattested form isn’t automatically “ungrammatical,” since it may be an accidental gap. Yet the theory of Di Sciullo and Williams (1987), like those of Aronoff (1976), Anderson (1992), and most other “generative” morphologists, is still not positioned at the extreme of the compositionality scale, since it assumes that word-internal and word-external composition are fundamentally distinct. Much closer to the extreme is Distributed Morphology (DM; Halle and Marantz, 1993; Harley and Noyer, 2003), so named because it distributes the jobs traditionally given to the “word level” among different components of the mind. Word-level semantic idiosyncrasies (like the special meaning of transmission) are extracted from grammar and placed in what DM calls the Encyclopedia. Underlying phonological forms and morphosyntactic conditioning environments are bundled into what DM calls Vocabulary Items. These are then combined with universal morphosyntactic features by essentially syntactic operations, and have their forms readjusted and otherwise transformed by essentially phonological operations. In terms introduced by Harley and Noyer (2000), “f-morphemes” are the functor-like morphosyntactic features, like PLURAL, that license “l-morphemes” like DOG (the parallels, if any, between the l-/f-morpheme dichotomy and the traditional root/affix dichotomy are the focus of section 3). Readers unfamiliar with DM but who know a little generative history may wonder how it deals with Chomsky (1970). This is the paper that introduced the lexicalist hypothesis, which is what led virtually every generative morphologist for over twenty years to assume a strict division of labor between morphology and syntax. Moreover, linguistic folklore credits Chomsky (1970) with killing the theory of Generative Semantics, which, like DM, brought syntax down inside words and morphemes. Marantz (1997) diffuses the apparent challenge to DM very simply: He shows that Chomsky (1970) has exactly the opposite implications from those generally ascribed to it, and in fact its fatal blow was aimed at the lexicalist hypothesis itself. Chomsky was not criticizing syntactic compositionality below the word level, since that’s precisely what he means by the “lexicalist hypothesis.” Rather, he was only criticizing the use of Generative-Semantics-style transformations to convert forms like destroy the city into forms like destruction of the city. Though not transformationally related, such forms, Chomsky argues, do share syntactically parallel structures. Crucially (as Marantz emphasizes), in order for Chomsky’s argument to go through, the root DESTROY shared by destroy and destruction must be visible to the syntax, thereby falsifying the “lexicalist hypothesis” as it is usually (mis)understood. Chomsky supports this claim with, among other examples, the contrast between destroy/destruction and grow/growth. In DM terms, the l-morpheme DESTROY, whether

110

The Mental Lexicon: Core Perspectives

inside a noun or a verb, expects to go with a patient, and that’s why both (1a) and (1b) below are acceptable. By contrast, the l-morpheme GROW describes an internally caused change; its default verbal use is intransitive, as in (2a). A transitive usage like (2b) only works with a causative meaning, represented in the syntax by the silent f-morpheme CAUSE, which is only permitted in verb phrases. The result is that the noun growth cannot have a causative implication: (3a) is acceptable, but not (3b) (the star indicates unacceptability, not necessarily ungrammaticality). Since the syntactic difference between the pairs destroy/destruction and grow/growth is predictable from the semantic difference between the l-morphemes DESTROY and GROW, these roots must be visible to syntax. (1)

a. b.

the enemy destroyed the city the enemy’s destruction of the city

(2)

a. b.

the tomatoes grew John grew the tomatoes

(3)

a. b.

the growth of tomatoes *John’s growth of tomatoes

Importantly in the context of this chapter, Harley and Noyer (2000, pp. 364-366) later invoked the competence-performance distinction to deal with a gap in the DESTROY/GROW argument. The problem involves l-morphemes like SEPARATE, which describe an internally caused change, as in (4a), yet unlike GROW do permit causative noun phrases, as in (4b). (4)

a. b.

Jim and Tammy Faye separated the teacher’s separation of the children

Harley and Noyer argue that such observations do not undermine Chomsky’s argument because knowledge of the difference between GROW and SEPARATE belongs in the grammar-external Encyclopedia. They do not consider this performance-based explanation as special pleading because they provide independent evidence for it: The acceptability of forms like (4b) depends on how causer-like the subject is perceived as being, which in turn depends on real-world knowledge. Thus, they claim, some speakers balk at examples like (5). (5)

the Cold War’s separation of E. and W. Germany

As Harley and Noyer admit, however, not all speakers reject examples like (5). They therefore predict that cross-speaker judgment variation must derive from cross-speaker variation in thresholds for the perception of causer-hood. Unfortunately this prediction

Generative Morphology as Psycholinguistics

111

remains rather speculative at this point, since they don’t explain how it could be tested against the alternative hypothesis that Chomsky’s (1970) argument is essentially mistaken. Given its extreme compositionality, DM is quite “generative” indeed, and its use of performance as a dumping ground for apparent counterevidence (word-level idiosyncrasies, GROW vs. SEPARATE) is also quite typical of argumentation in the generative tradition. Before we can consider the validity of this latter move, however, we need to take a closer look at the nature of morphological performance.

2.3. Morphological Performance Syntax tends to stick with judgment data, and phonology relies primarily on corpus (dictionary) data, but morphology, the middle child, uses both types, in roughly equal proportions. There are good reasons for this. Morphology is syntax-like in dealing with rather productive forms. The more productive a system is, the less likely one is to find theoretically relevant examples in a random sample; this observation is at the core of Chomsky’s (1957) argument against corpusbased syntactic analysis. Productivity in (lexical) phonology is limited by the small number of phonological units, so it is relatively easy for phonologists to distinguish systematic from accidental gaps, even in small corpora, but morphologists and syntacticians aren’t so lucky. This is why syntacticians rely on judgments, which can be elicited by the experimenter in unlimited supply. As we saw above, judgments also play an important role in DM argumentation. At the same time, however, morphology is like (lexical) phonology in dealing with word knowledge; syntactic productivity must be greater than morphological productivity because sentences are made of words (even in polysynthetic languages, where morphological productivity reaches syntactic levels). Thus morphology is actually in an intermediate position between phonology and syntax with regard to the usefulness of corpora, which is why dictionaries have been a major data source in generative morphology since the beginning. Some generative morphologists have also begun to use corpora of fluent discourse, most often as a tool for finding examples that sound right but are hard to think up spontaneously; Marantz (2005) applies this method to argumentation within a DM framework. Judgments and corpora provide complementary performance windows into competence: Judgments reflect both grammar and comprehension processes, while corpora reflect both grammar and production processes. Studying both makes it easier to factor out competence from performance. Moreover, the proper interpretation of judgments about listeme-like forms depends crucially on understanding corpus patterns. The reason is analogy. Most linguists assume they can eliminate lexical effects on judgments simply by testing novel forms that could not be memorized, but this doesn’t block speakers from analogizing, that is, generalizing superficially (extra-grammatically) from memorized exemplars. Even with

112

The Mental Lexicon: Core Perspectives

morphosyntactic judgments, Chomsky (1970, pp. 27-29) shows how speakers can accept ungrammatical forms “by analogy” (his term) with similar-sounding grammatical forms. Since both grammar and analogy involve generalizing across memorized tokens (the former via biologically-constrained acquisition processes, the latter haphazardly on the fly), judgments of novel forms are, to some extent, a sort of intuitive corpus analysis. For (lexical) phonological judgments, the influence of analogy has been precisely quantified. Bailey and Hahn (2001) had English speakers make wordlikeness judgments on nonword syllables, and calculated the proportions of judgment variation accounted for by phonotactic probability (e.g., the likelihood of /v/ being word-final) and by neighborhood density (e.g., the number of real words differing from the target in at most one phoneme). Both variables had significant effects on judgments (Myers and Tsay, 2005, replicated this in Mandarin Chinese). Since phonotactic effects in spoken word recognition are apparently more “prelexical”, representing listener expectations, and neighborhood effects more “postlexical”, involving comparison with lexical neighbors (Vitevitch and Luce, 1999), the latter are arguably more like analogy than the former. The good news from this research is that (phonological) judgment patterns are not solely analogy. The bad news is that after analogy is extracted, what’s left (including phonotactic probability) cannot really be identified with “grammar”. Phonotactic probability and neighborhood density are highly correlated, and each may be defined in many different ways, making it difficult to figure out where the line should be drawn. Moreover, reactiontime measures reveal that they interact with each other (Luce and Large, 2001), so according to the additive model (Sternberg, 1998) they are not associated with distinct processing modules. Similarly, the “phonotactic” magnetoencephalography (MEG) component identified by Pylkkänen et al. (2002) is also sensitive to lexical frequency (Embick et al., 2001). Thus there may be no fully “prelexical” component in lexical access at all, suggesting that grammar, which is supposed to be independent of the lexicon, plays no real-time role in the processing of lexical phonology (though this does not necessarily rule out an essential role for it at the computational level). Note how this discussion of the place of traditional linguistic data sources in competence-performance models has led naturally to data sources that are traditionally “psycholinguistic”. This is a simple consequence of the principle that answering new types of questions requires collecting new types of data. Unfortunately, this principle has had very little influence on linguistic practice, where data always involve acceptability judgments or corpora of productions, never any other type of linguistic behavior. Myers (2006b) illustrates this by comparing the histories of generative research on lexical phonology and psycholinguistic research on spoken word production between 1970 and 2000. The psycholinguists began by studying only corpus data (natural speech errors), then added experimentally-induced speech errors, then other types of word-production experiments, then brain-imaging studies, along the way sharpening theories by implementing them on computers. During the same period, however, the vast majority of generative phonologists made no changes in their data sources at all (dictionaries). This self-imposed limitation is

Generative Morphology as Psycholinguistics

113

deeply strange: Why should the science of competence be defined in terms of its preferred performance types? Fortunately, some linguists do occasionally take the extra effort to apply new data sources to grammatical questions. A case in point is Stockall and Marantz (2006), who test deep compositionality (which happens to be a central claim of DM) at Marr’s most fundamental level, that of neural implementation, by looking for MEG evidence of root priming in irregularly inflected English verbs. If the test had utterly failed (it didn’t), this wouldn’t have falsified DM since there are many ways to implement a computational goal, and a successful test doesn’t automatically “prove” DM, not only because DM makes many other claims, but more importantly because linking performance data of any kind to competence claims requires numerous auxiliary assumptions that must eventually be tested themselves. Nevertheless, studies like this do add to our understanding of grammar because they falsify alternative hypotheses that traditional data sources cannot. In particular, as we have seen, it is difficult to rule out an analogical analysis of word-internal patterns solely from corpus data (which are memorized) or judgments (which are partly influenced by analogy). Yet if we are safe in assuming that analogical effects are relatively slow (though perhaps not literally “postlexical”), then a root priming effect in an early MEG component cannot be dismissed as mere analogy. (Another application of “psycholinguistic” data to DM claims is Barner and Bale, 2002.) In short, there are many sources of performance data bearing on competence claims: The ones traditionally used in linguistics can be useful, the same holds for non-traditional ones, and most importantly, every data source is associated with a different competenceperformance linking model that must be understood, at least in rough outline, before data can be interpreted properly.

3. CASE STUDY: AFFIXATION VS. COMPOUNDING IN CHINESE In keeping with the back-to-basics approach of this chapter, I’ve chosen to illustrate morphological competence-performance linking models with a fundamental yet understudied issue: Does grammar really make a distinction between affixation and compounding? Rather than testing it empirically, this distinction is simply assumed by all of the morphological theories I am aware of: Affixes are assumed to be functor-like, whereas roots, the concatenation of which defines compounding, are assumed to be argument-like. In DM this contrast, instantiated in terms of f-morpheme and l-morphemes, is absolutely central to its syntactic approach. Inconveniently, however, languages don’t always draw a very sharp distinction between affixes and roots, and Chinese, as I will show, seems to draw no real distinction at all. Does this mean that something is fundamentally wrong with theories of morphological competence, DM in particular? Or is it another one of those cases where the competenceperformance distinction can (validly) come to the rescue?

114

The Mental Lexicon: Core Perspectives

Obviously I can’t really resolve such enormous issues in the space available, and anyway, the point is the journey, not the destination. What I do instead is sketch out the basic facts in Chinese, list some options available in the DM framework, and then see what of relevance can be extracted from data sources of various types.

3.1. The Basic Facts Chinese is the prototypical compounding language, and some readers may be surprised to learn that it has anything affix-like at all, but it does (Li and Thompson, 1981; Packard, 2000). One of the clearest examples of a (Mandarin) Chinese affix is zi (here and elsewhere transcriptions are in Hanyu Pinyin), found in nouns like háizi “child”, chēzi “car”, and shuāzi “brush”. This morpheme is affix-like because it can never be used in isolation (though the character used to write it is also used for the root morpheme zǐ “child”), it is phonologically reduced (toneless, indicated in Pinyin by the lack of a tone diacritic), it always appears in a fixed location (immediately after a root morpheme), it indicates the syntactic category of noun (even if the base morpheme is a verb, like shuā “to brush”), and it is semantically “bleached” (in fact it seems to mean nothing). Chinese has been argued to have several other affixes as well. A few could be called inflectional, like the human plural suffix men (háizimen “children”) and verbal aspect markers like zhe (xuézhe “studying”). Others seem derivational, like the human nominal suffix zhě (xuézhě “scholar”). The rest are native roots that are also used to fill in for the affixes in loan translations; an example we will meet again below is jiā “expert”, used to translate -ist (xīnlǐxuéjiā “psychologist”, literally “heart reason study expert”). Yet something is not quite right with this picture. First, unlike English, the great majority of morphologically complex words in Chinese are composed entirely out of roots. Perhaps that’s why the affix list has to be padded out with loan translations. Should “affixes” like jiā be on the list at all? Morphological status in the donor languages is synchronically irrelevant. Why not consider them to be roots that have acquired somewhat bleached semantics? Second, boundedness cannot be used as a reliable diagnostic for affixhood in Chinese, which has a large number of bound roots. Bound-root compounds in Western languages tend to be opaque curiosities (e.g., helico + pter), but Chinese has many fully transparent examples like xiàozhǎng “school principal”, where both xiào “school” and zhǎng “chief” are bound (as free words, “school” and “chief” are respectively xuéxiào “study school” and shǒuzhǎng “head chief”). Third, it’s not clear if all of the “affixes” are really word bound. Chinese orthography doesn’t help clarify the issue, since it fails to mark word boundaries at all. The aspect markers seem more like clitics (though clitic groups can be treated as listemes by Chinese readers; Myers et al., 2006). Even the “derivational” human marker zhě can be attached to things that

Generative Morphology as Psycholinguistics

115

look a lot like phrases (e.g., wèi liáng tǐwēn zhě “people who have not measured their body temperature,” ubiquitous on public signs in Taiwan during the SARS scare). Finally and most importantly, Chinese “affixes” don’t really behave like functors. Even noun-final zi, the most affix-like morpheme, is not really a nominalizer, since it’s most often used with roots that are already nouns (chēzi “car”). Pairs like shuā “to brush” - shuāzi “a brush” are misleading because conversion in Chinese requires no (overt) affixation; shuā itself can be a noun (Tai, 1997). A related observation is that derivational “affixation” in Chinese can’t feed output into other “affixation” processes (Packard, 2000), quite unlike languages like English (cf. nationalization). Though Chinese blurs the affix-root distinction to an extreme, all languages have their controversial cases. In English we have able doing double duty as suffix (lovable) and root word (able to love), ful/full (why would a true affix permit spoonsful?), tele (prefix or bound root?), and so on. The question is whether this sort of fuzziness in the lexicon should be a fatal problem for theories like DM that assume a sharp distinction in the grammar.

3.2. What Would DM Do? Even separate from the affix-root question, root compounding has attracted far less attention in the generative literature than affixation (Aronoff, 1976, doesn’t discuss compounding at all), and DM is no exception. Xue (2001) uses DM to argue that whether a string of Chinese roots is a word or a phrase is contextually conditioned, but his focus isn’t on compounding per se. Harley (2004) restricts her attention to compounds like X + verb + er/ing (e.g., scriptwriter, script-writing), which, conveniently for DM’s syntactic approach, contain overt affixes. About affixless compounds DM seems to have had nothing to say so far, not even well-behaved modifier-noun compounds like teapot or bookstore (but see Zhang, 2007). Nevertheless, DM does make it clear that the notions of f-morpheme and traditional affix are not identical. In particular, boundedness is morphosyntactically irrelevant; free function words like the and will are f-morphemes. Some f-morphemes even have competing free and bound Vocabulary Items, like COMPARE in English being realized as more or -er. Heidi Harley (personal communication, June 5, 2006) believes that able may perhaps be amenable to a similar analysis, taking advantage of syntactic devices to flip the argument roles in able to love vs. lovable (= “able to be loved”). Other properties of f-morphemes do seem to have relatively invariant correlates in performance, however. The simplest of these is semantic bleaching; f-morphemes are assumed to have extremely general, universal semantics. This fits well with one of the arguments in Packard (2000, pp. 72-73) for analyzing zhě (which he translates “one who does/is X”) as an affix, but yuán (“person whose job/position is X”) as a root. Since the semantics of the latter entails that of the former, zhě has the more basic meaning; it is more bleached. Semantic bleaching can have many causes other than grammar, a fact that weakens an argument made by Marantz in unpublished work (summarized in Harley and Noyer, 2000,

116

The Mental Lexicon: Core Perspectives

pp. 369-371) for treating suppletive pairs (go-went) as competing realizations of semantically bleached f-morphemes (GO); both suppletion and semantic bleaching are correlated with high frequency, so neither has to cause the other. Nevertheless, on average, f-morphemes must still be more bleached than l-morphemes. To quantify semantic bleaching in Chinese morphemes (for experiments described in 3.3.3), Wang and Myers (2004) collected semantic relatedness judgments (1 = very unrelated, 6 = very related) for morpheme-word pairs, as in (6), and for word-word pairs, as in (7). As the mean ratings indicate, Packard’s semantic contrast between zhě and yuán was borne out. (6)

(7)

a.

zhě

ruòzhě weak-zhě “weakling”

2.0

b.

yuán

fǎngyuán inquire-yuán “reporter”

3.6

a.

ruòzhě weak-zhě “weakling”

yìzhě translate-zhě “translator”

1.4

b.

guǎnyuán diànyuán house-yuán store-yuán “librarian” “salesclerk” (guǎn = túshūguǎn “library”)

4.0

As Packard (2000) points out, zhě has a second characteristic consistent with affixhood: It is functor-like (converting “X” into “one who does/is X”). Like bleaching, functorhood is partly in the eye of the beholder, but as functors, f-morphemes have two very specific properties. First, there are selectional restrictions on what they may license; this is why growth can’t have a causative implication, because CAUSE is restricted to verbal environments. Second, unlike l-morphemes, the Vocabulary Items of f-morphemes are in competition with each other for the same syntactic slots (Halle and Marantz, 1993). Since each f-morpheme defines its own syntactic frame, different f-morphemes compete as well. In the following exploration, then, I take surface evidence of semantic bleaching, (morphosyntactic) selectional restrictions, and competition as indicative of underlying f-hood.

Generative Morphology as Psycholinguistics

117

3.3. Three Kinds of Evidence This section examines three distinct data sources for evidence of f-hood. Two are considered traditionally “linguistic”, since they involve, respectively, native-speaker judgments and corpus data. The third data source is traditionally “psycholinguistic”, since it involves reaction times in lexical decision experiments. 3.3.1. Judgments. Searching for a judgment-based diagnostic for f-hood, my first thought was to find a Chinese parallel to the English judgment contrast in (8b,c) claimed by Di Sciullo and Williams (1987, p. 33). These examples were intended to show that affixes like -er are transparent to the assignment of semantic roles (bread is the patient of bake), whereas roots like man are not. (8)

a. b. c.

to bake bread a baker of bread *a bake-man of bread

Unfortunately, this turns out to be yet another example of what happens when linguists neglect careful methodology. The claim involves a relationship between two elements, -er/man and bread, and therefore the judgment experiment should involve two factors, not just -er vs. man, but also bread vs. no bread. Adding the [±bread] factor results in the two additional forms in (9), where (9b) is clearly worse than (9a). (9)

a. b.

a baker *a bake-man

There is, thus, a fatal confound in (8), making the pattern there irrelevant. As Spencer (1991, p. 333) points out, (9b) is no fluke; man generally doesn’t attach to verbs, especially not transitive ones. It’s not surprising, then, that my search for a Chinese parallel to this pattern didn’t turn up anything useful either. Then I stumbled upon an interesting empirical claim made by He (2004) regarding the interaction, in zhě forms, of the affixation of the plural marker men and the order of verb and object. Affixation of men is permitted with object-verb strings, as in (10a), but not with verbobject strings, as in (10b). Crucially, both forms are acceptable without men, as in (10c,d); thus, unlike Di Sciullo and Williams, He provides a complete experimental design. (10)

a.

OV-zhě-men yáoyán zhìzào zhě rumor make zhě “rumor-mongers”

men PLURAL

b.

VO-zhě-men *zhìzào yáoyán zhě men

118

The Mental Lexicon: Core Perspectives

c.

OV-zhě

yáoyán zhìzào zhě “rumor-monger”

d.

VO-zhě

zhìzào yáoyán zhě “rumor-monger”

The object-verb zhě forms in (10a,c) impress any English-speaking linguist as being strikingly similar to the scriptwriter-type compounds analyzed by Harley (2004). Perhaps in (10a,c) zhě licenses the adjacent verb, whereas in (10b,d) it cliticizes to an entire phrase, and this structural difference then affects whether or not men affixation is permitted (this is He’s own analysis, in broad outline). Note that this interpretation assumes that zhě has selectional restrictions: It behaves differently with a verbal vs. phrasal sister in the syntactic tree. Thus if zhě is a “true affix” but yuán is not, as Packard (2000) argues, we would not expect yuán to share this selectional restriction. He’s paper is not focused exclusively on zhě, and in fact, many of his examples involve root morphemes, including yuán. If the judgments are subtle enough, however, He may have simply missed some difference between zhě and yuán. Indeed, when I informally checked He’s paradigm examples with Chinese-speaking colleagues and students, I merely found that some didn’t like men at all (unlike the plural in English, men is used rather sparingly), and others didn’t like the verb-object order at all. These two biases are irrelevant to He’s generalization, which concerns the interaction between the two factors, but they are so powerful they overwhelm any informal search for such an interaction. Hence I decided to conduct a more careful judgment experiment. To do so, I took advantage of MiniJudge (Myers, 2006c), a free online software tool for designing, running, and analyzing linguistic judgment experiments in accordance with the general principles advocated in Cowart (1997). MiniJudge makes things easier for the novice experimenter by automating the generation of materials and using powerful statistics (generalized linear mixed effect modeling, a sort of repeated-measures logistic regression; Agresti et al., 2000) run in the free statistics package R (R Development Core Team, 2005) to extract the maximum amount of information from the minimal amount of data: a small number of speakers making binary good/bad judgments on a small number of items. One of the innovations of MiniJudge is the treatment of the (random) order of items as a continuous variable in the statistics to remove lingering order confounds, and if the experimenter so chooses, interactions with order can be factored out as well; among other things, this helps reduce the influence of cross-item analogizing. The materials were designed primarily with He’s two binary factors in mind ([±VO] and [±men]), but the head morphemes also varied across the three types shown in (11). Surveys with 48 randomly ordered items modeled on those in (10) were distributed by email to students and faculty in my linguistics department (in southern Taiwan) who didn’t know the purpose of the experiment. I received 18 completed surveys.

Generative Morphology as Psycholinguistics (11)

a. b. c.

semantically bleached, bound: semantically rich, bound: semantically rich, free:

119

zhě, jiā (-ist) yuán, shī (teacher) rén (person)

The default MiniJudge analysis revealed significant main effects of [VO] (verb-object strings lowered acceptability) and [men] (men suffixation lowered acceptability), reconfirming the irrelevant biases I had already noted in the informal judgments. As in the informal test, the [VO] × [men] interaction predicted by He was not detected. However, there was also a main effect of order, indicating a (performance-based) change in judgment strategy over the course of the experiment. Hence I used the option in MiniJudge to test for order interactions, and this revealed a significant interaction between order and [VO] × [men]. More importantly, factoring out this three-way interaction allowed the previously hidden [VO] × [men] interaction to emerge as a significant effect itself. He’s generalization was thus confirmed: For [+VO] forms, [-men] forms were judged better than [+men] forms almost four-to-one (10d vs. 10b), about twice as high as the ratio for [-VO] forms (10c vs. 10a). The patterns associated with the three-way contrast among morpheme heads (tested directly in R) were more complex. The simplest prediction is that He’s generalization should only hold for “true affixes” like zhě and jiā. Thus if we look only at bound morphemes, there should be an interaction between semantic bleaching (e.g., zhě vs. yuán) and the [VO] × [men] interaction. Unfortunately no such interaction was found, no matter how I played with the data. Another interesting pattern did emerge, however: Speakers preferred the head morpheme to be as affixlike as possible (zhě and jiā were better than yuán and shī, which were in turn better than rén). Unfortunately this cannot be taken as evidence for fmorphemes, since bound roots were also preferred over free roots; bleaching is a diagnostic for f-hood, but boundedness is not. Though this judgment experiment confirmed He’s generalization, it failed to confirm any special status for “true affixes”. As a null result, it can hardly be considered conclusive. Still hoping that my extension of He’s generalization might pan out, I made another attempt from a different angle, as described next. 3.3.2. Corpus Analysis. Another argument Packard (2000) gives for treating zhě as an affix and yuán as a root is the much greater productivity of the former, as demonstrated by a dictionary analysis. The notion of productivity is inextricably linked with generativity, not only because of the similarity of the concepts “produce” and “generate” (about which Chomsky, 1965, p. 9, expresses some regret), but also because productivity indicates that speakers are going beyond rote memory (though they may do this by analogy). In corpus linguistics, however, it’s usually applied to comparisons among morphemes of the same type, typically affixes (e.g., Baayen and Renouf, 1996). When comparing affixes and roots, other

120

The Mental Lexicon: Core Perspectives

factors complicate the interpretation of corpus data. In particular, the semantically richer a morpheme is, the more specific its information content, and thus the more restricted its usefulness in discourse; productivity may diagnose semantic bleaching rather than affixhood directly. Nevertheless, as long as we maintain respect for the extra-grammatical forces influencing corpus frequencies, it seems reasonable to assume that grammatical status should have some measurable affect on the rate at which linguistic forms are coined and used. I decided to apply this logic in a second test of the generalization of He (2004) and my extension of it. My data source was the ten-million-word Academia Sinica Balanced Corpus of Modern Chinese (Chen et al., 1996), which is parsed by words (no mean feat given the lack of word boundaries in Chinese orthography) and tagged for syntactic category. I searched for examples of each of the twenty classes implied by the design of the judgment experiment ([±VO], [+men], zhě/jiā/yuán/shī/rén), ensuring that I only chose transitive verbs and common nouns that were both disyllabic (He, 2004, observes that different principles apply with monosyllabic roots). I then tabulated type frequencies as a measure of coinage rate. The type frequencies shown in (12) yet again revealed clear biases against men (5 types with men vs. 927 without) and against verb-object strings (214 verb-object types vs. 718 object-verb types). Unfortunately, there wasn’t even the slightest hint of an interaction; He’s generalization was not confirmed. There’s thus no point looking for further interactions with affixhood. (12)

a. b. c. d.

[+VO, +men] [–VO, +men] [+VO, –men] [–VO, –men]

0 5 214 713

The preference found in the judgments for more affixlike morphemes was only partially replicated, as suggested by the ranking of type frequencies by head morpheme shown in (13). (13)

a. b. c. d. e.

zhě rén jiā shī yuán

766 101 49 5 1

The surprisingly high type frequency of rén forms may simply be due to its own high token frequency, which is far higher than that of the other five morphemes combined, including zhě, which has the second-lowest token frequency after yuán. Setting rén aside, the corpus results actually provide better evidence than the judgments for the special status of

Generative Morphology as Psycholinguistics

121

semantic bleaching (as opposed to boundedness), since the “true affixes” zhě and jiā are preferred in coinages far more than the “roots” shī and yuán. With such a small sample of types, however, this could merely be a statistical accident. Thus the little that came out of the corpus analyses ended up conflicting, partly, with patterns detected in the judgments. Since each data source has its pros and cons, we can’t take one as inherently more reliable than the other. In any case, my extension of He’s generalization did not prove very useful as an f-hood diagnostic. 3.3.3. Context and Morpheme. Frequency Effects on Lexical Decision. The final data source I consider here is a set of reaction-time experiments reported in Wang and Myers (2004). Our original motivations were traditionally “psycholinguistic”, in that we were testing hypotheses about real-time lexical access, not grammatical knowledge. Nevertheless, as we will see, even traditionally “psycholinguistic” experiments must be interpreted in the context of competence-performance linking models, and doing so may even reveal new insights into competence itself. Our experiments were based on Andrews (1986), who used morpheme frequency effects as a diagnostic for morphological decomposition in visual lexical decision tasks in English. Such studies generally find that decisions about complex words are faster if the words contain higher-frequency morphemes, even with whole-word frequency matched. Since frequency effects indicate lexical access, such results suggest decomposition of complex words into morphemes. However, Andrews suspected that decomposition is not automatic, but depends on the contextual influence of inherently more parsable forms on the access of inherently less parsable forms. Consistent with her suspicions, morpheme frequency effects were found with compounds presented alone (inherently easy to parse), but not with suffixed words presented alone (inherently hard to parse), and when she mixed both word types together, both showed morpheme frequency effects (context effects). Our interest in these experiments had nothing to do with decomposition, automatic or otherwise; in Chinese there’s nothing to debate, given the morpheme-based nature of its orthography (Myers, 2006a). We were instead attracted by the possibility that the Andrews paradigm could provide a diagnostic for distinguishing between affixation and compounding in Chinese. We ran one experiment with “suffixed” words only (i.e., bimorphemic words ending with semantically bleached morphemes according to our pretests), one with compound words only, and one with mixed morphological types. The results replicated the English results almost exactly, suggesting that at least according to this diagnostic, Chinese has “affixes” just as English does. Yet in the context of this chapter it’s important to consider what sort of competenceperformance linking model is presumed by the Andrews diagnostic. Clearly the effect of mixing word types must be a performance effect; grammar can’t change itself depending on the processing context. What about the finding that it’s inherently easier to parse roots out of compounds than out of suffixed words? Could we posit that the licensing relationship between the affix (f-morpheme) and root (l-morpheme) forces the parser to treat the root-affix

122

The Mental Lexicon: Core Perspectives

string as a whole listeme? Perhaps not; Taft and Forster (1975) and follow-up work found that roots are as easy to parse out of prefixed words as out of compounds. Aren’t prefixes just as much f-morphemes as suffixes are? My current suspicion is that the Wang and Myers (2004) results don’t bear on the fhood issue at all, but actually reveal nothing more than a secondary performance difference. The asymmetry between prefixed words and compounds on the one hand, and suffixed words on the other, can be at least partly explained as arising from the directionality of time itself, which causes a left-to-right bias in lexical access generally, even in monomorphemic words (Marslen-Wilson, 1987). A more productive research strategy might be to focus on affix competition in primed lexical decision experiments. The basic insight derives from the observations that root priming is robust when both prime and target are compounds (teacup - teapot; Zhou and Marslen-Wilson, 2000) but not when prime and target are suffixed words (confession confessor). The latter observation comes from an English study by Marslen-Wilson et al. (1994), and they explain it as resulting from competition of different suffixes for the same root. Recall from 3.2 that in DM, f-morphemes define competing syntactic frames; thus competing is just what we would expect suffixes to do. The most obvious wrinkle with using competition as an f-hood diagnostic is that yet again, prefixes behave differently; MarslenWilson et al. found that prefixed words do prime each other (unfasten - refasten). This time, however, ironing out such wrinkles may be worth the effort, since the competition notion also fits quite neatly with recent research on relation priming in modifiernoun compounds (see Gagné and Spalding, 2004, for a review). For example, student vote (head BY modifier) is primed by student accusation (head BY modifier) relative to student car (head FOR modifier). What’s particularly interesting about this research is that it points towards a possible DM analysis of root compounding: Perhaps the psychologically active compound-internal relation is expressed by a covert f-morpheme, similar to the overt affix -er in scriptwriter. Thus the compounds in (14a) would have the abstract structures in (14b). (14)

a.

student vote student accusation student car

b.

student-BY vote student-BY accusation student-FOR car

In favor of this hypothesis, note that the relations used in compounds are semantically bleached and universal: Li and Thompson (1981) list the same relations for Chinese compounds as those found in English, and Ji and Gagné (2004) have replicated the relation priming effect in Chinese compounds. Moreover, priming of student vote by student accusation relative to student car means that the BY and FOR relations compete (indeed,

Generative Morphology as Psycholinguistics

123

Gagné calls her theory “Competition Among Relations in Nominals”), just like the competition between overt suffixes found by Marslen-Wilson et al. (1994). This is why I treat BY and FOR as suffixes in (14b), on the assumption that the prefix/suffix priming asymmetry has a competence-based rather than performance-based explanation. This assumption is further supported by an asymmetry first discovered by Gagné (2001), whereby relation competition is found if compound and target compounds have the same modifier, as in (15a), but not if they have the same head, as in (15b) (Gagné, 2002, found that this competition also occurs with semantically related modifiers, e.g., scholar and student). (15)

a. b.

student-BY vote student-BY vote

student-FOR car reform-FOR vote

Gagné et al. (2006) have recently confirmed the positional restriction on relation priming by showing that it does not occur when compounds merely share roots and relations; thus there is no priming of (16a) by (16b) relative to (16c). (16)

a. b. c.

reading-FOR lamp lamp-FOR shade lamp-PRODUCED_BY light

This pattern of results makes sense if the covert relation-marking f-morpheme in modifier-head compounds is indeed suffixed to the modifier, so that the f-morphemes in (15a) compete with each other, just as -ion and -or compete in confession-confessor. If this interpretation of relation priming holds up, it would represent a striking example of how the study of performance, even when conducted solely for its own sake, can provide useful insights into the nature of competence. Indeed, it is not obvious how silent f-morphemes like BY or FOR, or their precise position inside a compound, could ever be established except via traditionally “psycholinguistic” evidence like priming effects. What does all this mean for Chinese? Though Ji and Gagné (2004) found relation priming in Chinese compounds, they failed to find any modifier-head asymmetry. This failure is consistent with those noted in 3.1, 3.3.1, and 3.3.2, suggesting that Chinese word formation may not, in fact, require the use of functor-like f-morphemes (relation priming in Chinese compounds would instead involve some sort of word-level semantic priming). If further exploration continues to uncover nothing but null results like these, Chinese may indeed pose a genuine challenge to theories like DM, if only a typological one.

124

The Mental Lexicon: Core Perspectives

4. WHAT NEXT? Generative morphology, like generative linguistics as a whole, is psycholinguistics: The two disciplines do not work on qualitatively different issues or in qualitatively different ways. Humans being what they are, maintaining the division between the cultures of linguistics and psychology assumes an illusory importance that does more harm than good. If only linguists would learn from psychologists about how to collect and analyze data in ways that respect the noisy channel between behavior and mind. If only psychologists would learn from linguists that interpreting performance data always depends on testable competence assumptions. Interdisciplinary interaction can only be improved through mutual respect and education. This truism applies to morphology as much as to any other aspect of language, but here the immediate future looks brighter. The undeniable importance of rote memory in word knowledge has long obliged generative morphologists to keep abreast of psycholinguistic findings, and (de)composition has long played a central role in the study of lexical access. This very book is testimony to the relative ease with which psychologists and linguists exchange views on the nature of words. Perhaps it isn’t overly idealistic to hope that a mutually respectful multiculturalism will someday become the norm in the language sciences as a whole.

Generative Morphology as Psycholinguistics

125

ACKNOWLEDGEMENTS Portions of the work described in section 3 were supported by National Science Council (Taiwan) grant NSC 94-2411-H-194-018, and MiniJudge is co-copyrighted by National Chung Cheng University. I have benefited from comments from Edith Aldridge, Christina L. Gagné, Heidi Harley, and Niina Zhang, as well as the editors; Ko Yuguang helped with the judgment experiment, and Hong Jiafei helped with the corpus analysis. Any mistakes are all my fault.

REFERENCES Agresti, A., J. G. Booth, J. P. Hobert and B. Caffo (2000). Random-effects modeling of categorical response data. Sociological Methodology, 30, 27-80. Anderson, S. R. (1992). A-morphous Morphology. Cambridge University Press, Cambridge. Andrews, S. (1986). Morphological influence on lexical access: Lexical or nonlexical effects? Journal of Memory and Language, 25, 726-740. Aronoff, M. (1976). Word Formation in Generative Grammar. MIT Press, Cambridge, MA. Aronoff, M. (this volume). Language: Between words and grammar. Baayen, R. H. (this volume). Storage and computation in the mental lexicon. Baayen, R. H. and A. Renouf (1996). Chronicaling the Times: Productive lexical innovations in an English newspaper. Language, 72, 69-96. Bailey, T. M. and U. Hahn (2001). Determinants of wordlikeness: Phonotactics or lexical neighborhoods? Journal of Memory and Language, 44, 569-591. Barner, D. and A. Bale (2002). No nouns, no verbs: Psycholinguistic arguments in favor of lexical underspecification. Lingua, 112, 771-791. Becker, T. (1993). Back-formation, cross-formation, and “bracketing paradoxes” in paradigmatic morphology. In: Yearbook of Morphology 1993 (G. Booij and J. van Marle, eds.), pp. 1-25. Kluwer, Dordrecht. Chen, K.-J., C.-R. Huang, L.-P. Chang and H.-L. Hsu (1996). SINICA CORPUS: design methodology for balanced corpora. Language, Information and Computation, 11, 167-176. Chomsky, N. (1957). Syntactic Structures. Mouton, The Hague. Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press, Cambridge, MA. Chomsky, N. (1970). Remarks on nominalization. Reprinted 1972 in Studies on Semantics in Generative Grammar (pp. 11-61). Mouton, The Hague. Chomsky, N. (1980). Rules and representations. Behavioral and Brain Sciences, 3, 1-15, 42-75. Chomsky, N. and M. Halle (1968). The Sound Pattern of English. Harper and Row, New York. Reprinted 1990, MIT Press, Cambridge, MA. Cowart, W. (1997). Experimental Syntax: Applying Objective Methods to Sentence Judgments. Sage Publications, London. Di Sciullo, A. M. and E. Williams (1987). On the Definition of Word. MIT Press, Cambridge, MA. Embick, D., M. Hackl, J. Schaeffer, M. Kelepir and A. Marantz (2001). A magnetoencephalographic component whose latency reflects lexical frequency. Cognitive Brain Research, 10 (3), 345-348.

126

The Mental Lexicon: Core Perspectives

Gagné, C. L. (2001). Relation and lexical priming during the interpretation of noun-noun combinations. Journal of Experimental Psychology: Learning, Memory, and Cognition, 27 (1), 236-254. Gagné, C. L. (2002). Lexical and relational influences on the processing of novel compounds. Brain and Language, 81, 723–735. Gagné, C. L., L. Figueredo and T. L. Spalding (2006, October). Does snow man prime plastic snow? The effect of position in accessing relational information during conceptual combination. Poster presented at the Fifth International Conference on the Mental Lexicon. Montreal, Canada. Gagné, C. L. and T. L. Spalding (2004). Effect of relation availability on the interpretation and access of familiar noun-noun compounds. Brain and Language, 90, 478-486. Halle, M. and A. Marantz (1993). Distributed morphology and the pieces of inflection. In: The View From Building 20: Essays in Linguistics in Honor of Sylvain Bromberger (K. Hale and S. J. Keyser, eds.), pp. 111-176. MIT Press, Cambridge, MA. Harley, H. (2004). Merge, conflation, and head movement: The First Sister Principle revisited. In: Proceedings of Northeast Linguistics Society 34 (K. Moulton and M. Wolf, eds.). University of Massachusetts, GLSA, Amherst. Harley, H. and R. Noyer (2000). Formal versus encyclopedic properties of vocabulary: Evidence from nominalisations. In: The Lexicon-Encyclopedia Interface (B. Peeters, ed.), pp. 349-374. Elsevier, Oxford. Harley, H. and R. Noyer (2003). Distributed morphology. In: The Second GLOT International Stateof-the-article Book: The Latest in Linguistics (L. Cheng and R. Sybesma, eds.), pp. 463-496. Studies in Generative Grammar 61. Mouton de Gruyter, Berlin. He, Yuanjian (2004). The words-and-rules theory: Evidence from Chinese morphology. Taiwan Journal of Linguistics, 2 (2), 1-26. Ji, H. and C. L. Gagné (2004, July). Lexical and relational influences on the processing of Chinese modifier-noun compounds. Presented at the Fourth International Conference on the Mental Lexicon, Windsor, Canada. Kenstowicz, M. and C. Kisseberth (1979). Generative Phonology: Description and Theory. Academic Press, New York. Li, C. N. and S.A. Thompson (1981). Mandarin Chinese: A Functional Reference Grammar. University of California Press, Berkeley, CA. Luce, P. A. and N. R. Large (2001). Phonotactics, density, and entropy in spoken word recognition. Language and Cognitive Processes, 16 (5/6), 565-581. Marantz, A. (1997). No escape from syntax: Don’t try morphological analysis in the privacy of your own lexicon. In: University of Pennsylvania Working Papers in Linguistics, (A. Dimitriadis, L. Siegel, C. Surek-Clark, and A. Williams, Eds.), Vol. 4.2, pp. 201-225. University of Pennsylvania, Pennsylvania. Marantz, A. (2005, December). Rederived generalizations. Handout for talk presented at Tsinghua University, Hsinchu, Taiwan. Marr, D. (1982). Vision. W. H. Freeman, San Francisco. Marslen-Wilson, W. (1987). Functional parallelism in spoken word recognition. Cognition, 25, 71102. Marslen-Wilson, W., L. K. Tyler, R. Waksler and L. Older (1994). Morphology and meaning in the English mental lexicon. Psychological Review, 101 (1), 3-33.

Generative Morphology as Psycholinguistics

127

Miller, G. A. (1990). Linguists, psychologists and the cognitive sciences. Language, 66, 317-322. Myers, J. (2006a). Processing Chinese compounds: A survey of the literature. In: The representation and processing of compound words (G. Libben and G. Jarema, eds.), pp. 169-196. Oxford University Press, Oxford. Myers, J. (2006b, May). Linguistics as cognitive psychology. Pre-Conference proceedings of the 14th Annual Conference of the International Association of Chinese Linguistics and 10th International Symposium on Chinese Languages and Linguistics Joint Meeting (pp. 150-174). Taipei, Taiwan. Myers, J. (2006c). MiniJudgeJS (Version 0.9.9) [Computer software]. URL: http://www.ccunix.ccu.edu.tw/~lngproc/MiniJudgeJS.htm Myers, J. and J. Tsay (2005, May). The processing of phonological acceptability judgments. Proceedings of Symposium on 90-92 NSC Projects (pp. 26-45). Taipei, Taiwan. Myers, J., Y.-C. Huang and W. Wang (2006). Frequency effects in the processing of Chinese inflection. Journal of Memory and Language, 54, 300-323. Ohala, J. J. (1986). Consumer’s guide to evidence in phonology. Phonology Yearbook, 3, 3-26. Packard, J. L. (2000). The Morphology of Chinese: A Linguistic and Cognitive Approach. Cambridge University Press, Cambridge, UK. Penke, M. and A. Rosenbach (2004). What counts as evidence in linguistics? An introduction. Studies in Language, 28 (3), 480-526. Phillips, C. and H. Lasnik (2003). Linguistics and empirical evidence: Reply to Edelman and Christiansen. Trends in Cognitive Science, 7 (2), 61-62. Pylkkänen, L., A. Stringfellow and A. Marantz (2002). Neuromagnetic evidence for the timing of lexical activation: An MEG component sensitive to phonotactic probability but not to neighborhood density. Brain and Language, 81, 666-678. R Development Core Team (2005). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URL http://www.Rproject.org. Schütze, C. T. (1996). The Empirical Base of Linguistics: Grammaticality Judgments and Linguistic Methodology. University of Chicago Press, Chicago. Spencer, A. (1991). Morphological Theory: An Introduction to Word Structure in Generative Grammar. Basil Blackwell, Oxford. Sternberg, S. (1998). Discovering mental processing stages: The method of additive factors. In: An Invitation to Cognitive Science, Vol. 4: Methods, Models, and Conceptual Issues (D. Scarborough and S. Sternberg, eds.), pp. 703-863. MIT Press, Cambridge, MA. Stockall, L. and A. Marantz (2006). A singe route, full decomposition model of morphological complexity. The Mental Lexicon, 1 (1), 85-123. Taft, M. and K. Forster (1975). Lexical storage and retrieval of prefixed words. Journal of Verbal Learning and Verbal Behavior, 14, 638-647. Tai, J. H-Y. (1997). Category shifts and word-formation redundancy rules in Chinese. Chinese Language and Linguistics III: Morphology and Lexicon. Symposium Series of the Institute of History and Philology, No.2, Academia Sinica, pp.435-468. Vitevitch, M. S. and P. A. Luce (1999). Probabilistic phonotactics and neighborhood activation in spoken word recognition. Journal of Memory and Language, 40, 374-408. Wang, W. and J. Myers (2004, July). The processing of affixation and compounding in Chinese. Presented at Fourth International Conference on the Mental Lexicon, Windsor, Canada.

128

The Mental Lexicon: Core Perspectives

Westbury, C. F. and G. Hollis (this volume). Putting Humpty together again: Synthetic approaches to nonlinear variable effects underlying lexical access. Xue, N. (2001). Defining and automatically identifying words in Chinese. Unpublished doctoral thesis, University of Delaware, Newark, DE. Zhang, N. (in press). Root merger in Chinese compounds. National Chung Cheng University. Zhou, X. and W. Marslen-Wilson (2000). Lexical representation of compound words: Cross-linguistic evidence. Psychologia, 43 (1), 47-66.

7 ORIGINS OF CROSS-LANGUAGE DIFFERENCES IN WORD RECOGNITION

Laurie Beth Feldman and Dana M. Basnight-Brown, The University at Albany, SUNY, Albany, USA and Haskins Laboratories, New Haven, USA

1. INTRODUCTION Claims that structural differences across language introduce variation in visual word recognition are not new. Comparisons across languages reveal variety with respect to orthographic, phonological and morphological structure. In the orthographic – phonological domain, languages differ with respect to whether written units correspond to syllables as in Chinese and Japanese kana, or to phonemes, as in alphabetic scripts such as English and Serbian. In this domain, when differences across languages in visual word recognition emerge, typically they are associated with orthographic depth and the complexity of the correspondences between spelling and sound. For example, latencies in the naming task are faster than in the lexical decision task in English and in Serbian. By contrast in Hebrew without its diacritics, where phonology is not fully specified, the pattern tends to reverse (Frost et al., 1987). In the morphological domain, differing outcomes across language are attributed to factors associated with the prominence of combinations of base morphemes with (combinations of) affixes, either infixes, prefixes or suffixes and, consequently, to the systematicity with which words with similar form tend to be similar in meaning. Because of the tendency for base morphemes in Hebrew to appear in combination with other morphemes to form many words relative to base morphemes in English, some have suggested that the lexicon in morphologically rich languages, such as in Hebrew, is structured morphologically whereas in more impoverished languages (meaning English) it is structured orthographically (Frost et al., 2005; Plaut and Frost, 2001). In traditional dual route accounts of phonological as well as morphological aspects of word recognition, recognition proceeds either by whole words or by analysis of constituents where options are mutually exclusive at early stages of processing. Consequently, the failure to

129

130

The Mental Lexicon: Core Perspectives

observe effects of constituent (phoneme or morpheme) structure in word recognition tasks is interpreted as evidence in support of the whole word processing option. Within a less localist and more connectionist framework (Harm and Seidenberg, 1999, 2004; Seidenberg and McClelland, 1989), accounts of word recognition emphasize statistical learning of mappings between orthographic, phonological, and semantic codes. Options include mappings from spelling to meaning, or mappings from spelling to phonology and then to meaning. Instead of a single processing option determined by word regularity or a language’s orthographic depth, routes work cooperatively and interdependently with a division of labor. To anticipate, as one departs from the dichotomous processing framework based on routes that work competitively, and effects that either are present or absent, the argument for differing modes of processing across languages is not as persuasive as once believed. Part of the complexity when assessing evidence for cross language variation, is that different tasks, presentation formats, word structure and list composition often accompany changes in language. Consequently, methodological factors may contribute to or even create the illusion of cross-language variation in word recognition. In the present chapter we selectively review evidence from visual word recognition for language specific variation in the domains of orthographic, semantic, and morphological processing. We focus primarily on the primed visual lexical decision task with short presentations of the prime with (Forster and Davis, 1984) and without a forward mask (Feldman, 2000) because a large proportion of the experimental literature has focused on this very early stage of recognition. Early recognition is defined by processing time for the prime that is limited both by a preceding pattern mask and by the target that follows, or only by the target that follows. These are presentation conditions that purportedly minimize contamination from strategic and less automatic processes. To anticipate, we argue that in visual word recognition, similarities across languages predominate over differences. Claims for cross-language variation in early stages of processing appear less credible when methodologies are standardized and interpretations emphasize graded effects as opposed to dichotomized outcomes where significant effects are the focus and null effects are ignored even when the two do not differ statistically.

2. EFFECTS OF ORTHOGRAPHIC SIMILARITY: LANGUAGE UNIVERSAL OR LANGUAGE VARYING? It has been asserted that the organizational principle of the Hebrew lexicon is morpheme-based and therefore differs from that in the lexicon of speakers of Indo-European languages and that as a consequence, form effects fail to arise in Hebrew (Frost et al., 2005), whereas form priming is robust in Indo-European languages (Forster and Azuma, 2000; Rastle et al., 2000, Experiment 1). Nonetheless, a systematic examination of the orthographic priming literature restricted to forward masked prime presentations at prime durations less than 60 ms reveals

Origins of Cross-Language Differences in Word Recognition

131

that inhibitory or null effects within orthographically similar word-word prime-target pairs predominate across many languages. Although this finding has been replicated many times, careful comparison across studies reveals several experimental factors that may have contributed to the impression that facilitatory effects are common in English but not in non Indo-European languages (Frost et al., 2005). These factors include not only task and presentation conditions for the prime, but also neighborhood size, length and frequency of target, as well as definition of prime-target similarity, lexicality of the prime, and composition of filler items. We begin by delineating some properties of the target and then properties of the prime and its relation to the target that tend to produce graded inhibition in recognition tasks. These lead us to call into question the claim for reliable differences between languages with respect to processing of orthographic similarity.

2.1. Orthographic Neighborhood Size A word’s orthographic neighbors (N) [Coltheart et al., 1977] are those words that differ from it by a single letter (e.g., BOLD, HOLD, FOLD, MOLD, SOLD, GOLD, TOLL are some of the neighbors of TOLD). Effects of target N arise under both the single word and primed presentation formats. The typical N effect (Andrews, 1989, 1992; Forster and Shen, 1996) is that unprimed words with many neighbors (large neighborhoods) are recognized faster in the lexical decision task than those with fewer neighbors (small neighborhoods). Likewise, N or neighborhood density effects have been demonstrated in Spanish (Perea and Rosa, 2000) and in French (Ferrand and Grainger, 1992). For nonwords targets, large N tends to slow decision latencies. In the naming task, by contrast, increases in N decrease both word and nonword latencies in English (Andrews, 1989; Sears et al., 1995) and in French (Peereman and Content, 1995). It is possible that purported differences across languages with respect to the influence of form similarity between prime and target reflect orthographic characteristics of words and that any processing variation across languages is a consequence of the orthographic structure of its words. Specifically, target neighborhood size is not independent of target length such that longer words tend to have fewer neighbors (Baayen et al., 2006; Barca et al., 2002) and, perusal across languages reveals that average word length does vary. Therefore apparent differences among languages with respect to the role of orthographic similarity in visual word recognition may derive from differences in neighborhood size and average word length. Some of the original work on orthographic similarity in word recognition directly incorporated manipulations of target length and results are consistent with the claim that differing word length may contribute to or even underlie language-varying effects of form similarity. Typically in the range of 1-11 neighbors, orthographically related primes with many letters facilitate target processing (Forster et al., 1987), while those with fewer produce inhibition (Segui and Grainger, 1990). Thus, pairs such as ALTITUDE-ATTITUDE are more

132

The Mental Lexicon: Core Perspectives

likely to produce facilitation than pairs like FACT-FACE. In the aggregate, words that are long have fewer neighbors and, the existence of many neighbors tends to increase activation in the lexicon and therefore biases a “word” response. However, with high levels of activation overall, the benefit of a related as contrasted with an unrelated prime may diminish. This account is consistent with the findings of Forster and Veres (1998) who observed facilitation following orthographically related nonwords whose magnitude depended on their structure. When the nonwords did not look like real English words, or had few neighbors, there was facilitation; when they did look like real words and had a large N, there was none.

2.2. Alternative Measures of Form Similarity While number of neighbors or neighborhood size is the most thoroughly investigated, it is not the exclusive measure of a word’s orthographic similarity to other words. Also related to neighborhood size is neighborhood distribution (ND), the number of different letter positions in a target word where replacement of a single letter forms a neighbor. Greater inhibition with higher neighborhood distribution has been documented in French (Mathey et al., 2004) and in English (Johnson and Pugh, 1994). Twin neighbors of a target arise by multiple replacements of a letter in a constant position (e.g., changes in the 4th letter position of PROBE produce twin neighbors – PROVE, PROSE). By contrast, single neighbors, like the word IMPART for example, would have a distribution of 2, because substitutions in the 4th and 5th positions can produce neighbors IMPORT and IMPACT (Mathey et al., 2004, p. 535). Manipulating ND, while controlling N size and N frequency (all prime words were the highest frequency neighbor of the target) revealed that in French, ND interacted with forward masked priming such that single neighbors produced a significant inhibitory effect, while twin neighbors revealed the absence of an effect. The inhibitory influence of shared neighbors on forward masked priming has been replicated recently in English (Davis and Lupker, 2006). To reiterate, it is targets with large neighborhoods that are more likely to have neighbors that are distributed with respect to letter position as well as to have two or more neighbors in one position. When word targets are preceded by a prime that is orthographically similar and forward masked, word-word prime-target pairs in Dutch that do not share neighbors produce more facilitation than do pairs that do share neighbors (Van Heuven et al., 2001). Similar effects have been reported in English such that inhibition increases as does the number of shared neighbors between prime and target (Davis and Lupker, 2006). Once again, it is targets with large neighborhoods that are more likely share neighbors with the prime. Originally, measures of orthographic similarity were constrained by length so that words of one length were assumed to activate only orthographic neighbors with the same number of letters. However, recent findings indicate that words can activate words not only of an equivalent length, but also those that share orthographic overlap and are of different lengths (DeMoor and Brysbaert, 2000). It is also the case that counts of orthographic neighbors are

Origins of Cross-Language Differences in Word Recognition

133

generally based on corpora that fail to include inflected word forms so, for example, LAWS would not enter into the neighbor count for LAWN. Nonetheless, there is evidence from Polish (Feldman et al., 2004a) that for low frequency targets matched on length and surface frequency, decision latencies differed reliably as the number of orthographic neighbors increased, and that the sum of inflected as well as uninflected forms proved to be a better predictor of decision latencies than was the number of uninflected neighbors alone. One final alternative definition of orthographic similarity derives from the number of body neighbors (BN), or number of words that share the orthographic body of a word but not necessarily its length. When N size is controlled, word recognition is typically faster (resulting in significant facilitation) for those French words that have many body neighbors (Ziegler and Perry, 1998). For example, DRIVE has many body neighbors, LIVE, THRIVE, HIVE, DIVE and STRIVE, but only one orthographic neighbor, DROVE. When Janack et al. (2004) included this factor by pairing primes with targets that were mismatched in either the initial (mast-CAST), final (cash-CAST) or partial neighbor condition (lash-CAST), however, for targets with N > 2, they observed significant inhibition across all three positions of mismatch, suggesting that body neighbors were no different from other measures in that all positions of mismatch produced inhibition.

2.3. Effects of Prime Frequency and Lexicality on Form Priming Not only properties of the prime but also properties of the target, and the relation between prime and target, influence the magnitude and sometimes direction of any orthographic effect. In priming tasks, several variants of orthographic similarity tend to reduce or even eliminate the benefit of a large neighborhood. In addition to letters shared by prime and target, lower frequency prime words are more likely to show facilitation, whereas higher frequency prime words are more likely to inhibit the processing of orthographically similar target words. Similar findings have been reported in French, English, Dutch, and Italian (Arduino and Burani, 2004; Davis and Lupker, 2006; DeMoor and Brysbaert, 2000; Grainger and Ferrand, 1994). Curiously, the effect of relative frequency diminishes with the introduction of pseudohomophone filler trials, at least in French (Grainger and Ferrand, 1994). For example, JOIE and JOIS are pronounced similarly but only the first is a real word and the presentation of JOIS type items attenuates the tendency for low frequency primes to facilitate recognition of the target. Finally, lexicality of the prime word modulates the outcome of orthographic similarity. As discussed previously, significant inhibition tends to arise after word primes, while combinations of word and nonword primes produce a null effect (Drew and Zwitserlood, 1995; Janack et al., 2004), or inhibition for word primes in conjunction with facilitation for nonword primes (Davis and Lupker, 2006; Forster and Veres, 1998).

134

The Mental Lexicon: Core Perspectives

2.4. Absence of Form Effects: Evidence of Cross-item or Cross-language Variation? Across several languages, inhibition due to shared form tends to be more common than facilitation yet in some recent data of our own (Feldman and Basnight-Brown, submitted), we have observed that when targets are matched on surface frequency (94 [SD = 118] vs. 119 [SD = 194]), facilitatory as well as inhibitory patterns of forward masked facilitation arose but depended on the neighborhood size of the target. In the context of identity filler trials, we examined decision latencies to words that followed a morphologically (M), orthographically (O) or semantically (S) related prime that was forward masked. Importantly, when targets differed with respect to neighborhood size but the same targets appeared with morphologically (e.g., artist-ART), orthographically (e.g., artery-ART), and semantically related primes (e.g., craft-ART), results revealed significant orthographic inhibition for targets from small neighborhoods (7 [SD = 3.6]), but nonsignificant facilitation for targets from large orthographic neighborhoods (17 [SD = 3.4]). Targets after morphologically similar primes appeared to be less affected by the manipulation of neighborhood size of the target than were orthographic primes and those after semantic primes were not affected at all (see Table 1). Table 1. Magnitudes of facilitation for targets with large and small orthographic neighborhoods FACILITATION ORTHO

MORPH

SEM

Small N

-32 (118)

42 (81)

19

Large N

12 (102)

56 (110)

18 (100)

(97)

Hebrew words typically are composed of a root and a word pattern comprised mainly of vowels. However, in many contexts, printed Hebrew words are written without vowels so spelling consists of a three letter root as well as the one or sometimes two letters of the word pattern that are not vowels. As a result, the typical printed Hebrew word is three to five letters in length. If estimates of orthographic neighborhood size are calculated on a word’s written form, then Hebrew words will tend to be composed of fewer letters and will have more neighbors than do English words. Consistent with this convention, in the long-term variant of the priming task, magnitudes of facilitation do not differ when letters from the word pattern interrupt the root morpheme and when they do not (Feldman and Bentin, 1994). It has been documented in English (Feldman and Basnight-Brown, submitted; Pollatsek et al., 1999) and in Spanish (Perea and Rosa, 2000) that the presence of many neighbors may facilitate word performance overall in the lexical decision task but it tends to reduce form facilitation. Accordingly in Hebrew, high levels of target activation based on the root may be responsible

Origins of Cross-Language Differences in Word Recognition

135

for the failure to detect form priming (Frost et al., 2005). In essence, similarity with many other words and high levels of orthographic activation overall may render targets less susceptible to influences of the prime context in which they appear. In conclusion, apparently different outcomes across languages in word recognition tasks may reflect overall levels of orthographic activation as a consequence of the extent to which words within a language resemble few or many other words. Effects of N sometimes are interpreted as evidence that word level knowledge can feed back to activation at the letter level. Forster and Veres (1998), described recognition as a two-part combination of facilitation and inhibition. When the orthographic prime (neighbor) is processed, it activates letters that are shared with the target, thereby facilitating lexical access. However, at the word level after access, competition arises (especially if the prime is a word with a frequency that is higher than that of the target), thereby creating an inhibitory effect. In this framework, competition due to factors such as shared neighbors between prime and target may be responsible for the absence of facilitation in many studies. This analysis is particularly appropriate for Hebrew where many written words are three or four letters in length, have multiple neighbors and presumably, many shared neighbors between prime and target.

2.5. Unmasked Comparisons Across Languages Several experiments conducted across diverse languages have asked whether the brief (less than 60 ms) visual presentation of unmasked but orthographically similar word primes, not matched in length to the target (e.g., VOWEL-VOW), facilitate the processing of target words. Typically, orthographic controls in morphological studies are defined so as to preserve the initial portion of the word and length is allowed to vary. Collectively, outcomes range from inhibitory to null to facilitatory effects (see Feldman, 2000; Zwitserlood, 1996). For example, inhibitory effects under unmasked but brief visual presentation conditions have been documented in German and in Dutch (Zwitserlood, 1996) for pairs such as KERST (Christmas) - KERS (cherry), for Serbian pairs such as STANICA (station) - STAN (apartment) (Feldman and Moskovljević, 1987) and nonsignificant facilitation for English pairs such as VOWEL-VOW (Feldman, 2000). In conclusion, forward masked orthographic facilitation does not appear to be a reliable finding in English or in any other language and results vary little when the forward mask is eliminated. Experimental factors such as number of neighbors of the target, prime frequency relative to target, and number of shared neighbors between prime and target, seem to attenuate the magnitude of the inhibitory effect of orthographic similarity in all languages. In addition to these graded effects, prime lexicality can reverse the direction of the effect of similarity so as to render it facilitatory. Further, increases in target frequency seem to augment inhibitory effects (Davis and Lupker, 2006). Some or all of these factors in combination, but especially number of neighbors in conjunction with target length, are likely responsible for purportedly

136

The Mental Lexicon: Core Perspectives

differing effects of orthographic similarity that have been reported across languages. By implication, it is essential to consider these factors and anticipate their graded effects when designing future studies and when making assertions about putative differences in lexical organization or processes of word recognition across languages.

3. EFFECTS OF FREQUENCY AND SEMANTIC RELATEDNESS: LANGUAGE UNIVERSAL OR LANGUAGE VARYING? Languages with alphabetic writing systems differ with respect to their orthographic depth and the adequacy of predicting the phonology requisite for naming from orthographic form. In large part, a language’s orthographic depth is related to the prevalence of irregular words where the mapping is more complex. In contrast to regular words whose phonology can be computed from its orthographic form without consideration of its semantics, the traditional account for irregular words is that an accurate pronunciation must be associated with, but cannot be computed from, an orthographic form. Stated in more connectionist terms, a word’s semantic properties become more important when the mapping between its orthographic and phonological form is ambiguous and not fully systematic. In a connectionist framework, degree of systematicity of the mapping between orthographic and phonological form is crucial and the same basic characterization applies both across languages and among words within one language. The pronunciation of those English words whose mappings are irregular (e.g., SWORD, ANSWER, TONGUE) is unpredictable so that the phonology necessary for naming a word aloud cannot be computed independently of lexical and semantic knowledge about the word. Spelling in the Serbian language was reformed in the last century so that by comparison with English, an accurate pronunciation (with the exception of stress) always can be predicted from a word’s spelling. Italian is similar to Serbian in the regular correspondence between spelling and pronunciation, whereas Hebrew is closer to English in that conventional written forms in isolation fail to specify many vowels. An analog of phonological regularity, more precisely phonological specificity, exists in Hebrew and in Persian written forms with and without the inclusion of vowel diacritics. In earlier work promulgated as evidence of cross-language variation, effects on naming latencies of word frequency, a property of the whole word and a marker of lexical processing, varied with a language’s orthographic depth. Differential naming latencies for words of high and low frequency were more reliable in the “deep” writing systems of English and in Hebrew without vowel diacritics (Frost, 1994) than in the “shallow” writing systems of Serbian (Frost et al., 1987; Frost and Katz, 1992), Italian (Arduino and Burani, 2004), or Hebrew with diacritics (Frost, 1994). Results suggested to some that lexical or whole word processes were more salient when phonological specification was inaccurate or otherwise incomplete. However, others claimed it was processing time until onset of articulation, rather than

Origins of Cross-Language Differences in Word Recognition

137

orthographic depth per se, that determined the salience of lexical markers such as frequency. In support of a time criterion account, list composition can enhance or diminish differences in latencies that correspond to manipulations of frequency (Lupker et al., 1997). Similar outcomes have been documented for high and low frequency words in Korean Hangul as a function of the inclusion of phonographic or non phonographic filler trials (Simpson and Kang, 1994), and in Persian where words that are phonologically transparent and regular show a smaller frequency effect. Collectively, the salience of whole word processing as indexed by effects of a word’s surface frequency has been documented in many languages. Lexical effects are not independent of overall naming latencies, however. Finally across languages, lower frequency words tend to be more amenable to modulation by list composition and other manipulations than are higher frequency words. Similar to lexical specification of a word’s pronunciation as a gauge of lexical processing, is the benefit to recognition of presenting a target word in the context of a semantically related word. One interpretation of speeded recognition or facilitation after a semantically related word in the naming task for English, but not for Serbian, was differing influences of a word’s semantics, also an aspect of lexical knowledge (Katz and Feldman, 1983). Analogously, semantic facilitation in the naming task is greater in English than in Italian (Tabossi and Laghi, 1992) or in Persian, where words that are phonologically opaque, due to the deletion of letters for particular vowels, show greater magnitudes of semantic facilitation than do phonologically transparent words (Baluch and Besner, 1991). Stated generally, across languages, the phonology that underlies naming a word is subject to lexical and semantic influences (Katz and Feldman, 1983). Semantic effects tend to be robust when orthographic depth tends to be greater and the mapping between grapheme and phoneme is more complex. Effects of imageability, a semantic property of the target, as distinguished from the prime-target pair, likewise have been documented and effects are more reliable for phonologically irregular than for regular targets. In particular, for low frequency phonologically irregular words in English, naming latencies are faster and more accurate for words that are high (e.g., SIEVE) relative to low (e.g., COUTH) in imageability (Shibahara et al., 2003; Strain et al., 1995). The outcome suggests that pronunciation for irregular words is more affected by semantics than for regular words. Naming latencies for two character Kanji words of Japanese (written in a non alphabetic and non phonemic script) reveal an analogous effect of imageability for low but not higher frequency words. Effects of phonological complexity also can interact with semantic properties in the lexical decision task. For example in Serbian, strings that are phonologically ambiguous because they consist of letters that exist in both the Cyrillic and Roman alphabets but have different phonemic interpretations in each (e.g., PATAK can be read as /ratak/ in Cyrillic but as /patak/ in Roman) show greater magnitudes of semantic facilitation when targets are written in their phonologically ambiguous form than when the same targets are written in a phonologically unambiguous form (e.g., ΠATAK can be read only as /patak/).

138

The Mental Lexicon: Core Perspectives

In summary, lexical involvement in word recognition, and especially naming, does vary across languages but effects seem to be accountable in terms of overall recognition latencies and time-limited processing rather than disparities across languages with respect to their overall orthographic depth.

4. EFFECTS OF MORPHOLOGICAL RELATEDNESS: LANGUAGE UNIVERSAL OR LANGUAGE VARYING? The study of morphological processing and the way in which prime-target pairs that share morphemes (e.g., ALLOWANCE-ALLOW) are stored in the lexicon also has been the subject of extensive cross linguistic comparisons, because languages vary considerably in the manner and the extent to which morphemes combine to form words. A final domain in which processing potentially differs cross-linguistically originates with morphological structure. Based on the prevalence of words formed by appending one or more prefixes or suffixes to a base morpheme, English morphology is impoverished compared with languages such as Serbian or Hebrew. Consequently, systematicities in the mappings from form to meaning, defined by the degree to which words that look alike have similar meanings, are weaker in English than in Serbian or Hebrew. The term morphological richness reflects the tendency, within a language, for morphemes to combine to form more words, but the term is used inconsistently as there is no consensus as to whether definitions of morphological richness should include derivational as well as inflectional word formations or should be restricted to (either inflectional or) derivational formations. Relatedly, the term morphological family size reflects the number of words (derivations and usually also compounds) formed from the base morpheme. In the case of exclusively inflectional formation, in comparison with other languages, Hebrew no longer may qualify as a morphologically rich language. Nonetheless, some have suggested that processing in morphologically rich languages, referring to Hebrew, generally may be biased toward analysis of constituents and away from whole word processing. Two issues of long standing debate in morphological processing with potential for cross language variation are whether regular and irregular inflections are processed in the same manner or make use of separate mechanisms, and whether effects of morphological relatedness can be captured in terms of conjoint effects of similar form and similar meaning without invoking explicitly morphological representations. For the first, irregular inflection can entail form changes to the stem (e.g., HOPPED vs. RAN), substitutions of a less typical inflectional affix (e.g., TOASTED vs. BURNT) or alternations with a dissimilar full form (e.g., IS vs. WAS). For the issue of conjoint effects, not only morphological relatives formed by inflection (e.g., ALLOWED-ALLOW), but also those formed by derivation (e.g., ALLOWANCEALLOW) are of interest because, analogous to effects of form change, semantic transparency of the base morpheme can vary dramatically across words formed from the same base

Origins of Cross-Language Differences in Word Recognition

139

morpheme. Consequently, the degree of semantic similarity between pairs of morphologically related words can vary. The second source of potential cross-language variation in the morphological domain pertains to the role of semantics in early stages of morphological processing. Corresponding to the literature on orthographic priming, several characteristics of the prime word, target word or experimental conditions can influence the magnitude of morphological effects. The aim of the following section is to identify those factors that allegedly produce differences across languages in the single word and in the short-term forward masked and unmasked morphological priming tasks, and to evaluate those claims. Some factors are methodological and parallel those in orthographic priming but first we address those that seem to originate from differing word formation processes across languages.

4.1. Modulations of Morphological Processing in Single Words One class of measures that plays a prominent role in morphological processing for derivations as well as inflections and thus contributes to single word decision latencies is a target’s morphological family size (i.e., the number of different words formed from a base morpheme thereby producing morphologically related words). Typically, with appropriate controls, base morphemes with large families are recognized faster than those with small families (Schreuder and Baayen, 1997). For example, response latencies to a word like AUTHOR which appears in AUTHORESS, AUTHORITY, AUTHORITARIAN, AUTHORIZE and many other derived words, is faster than for CATCH, which appears in very few other words (e.g., CATCHER). Despite variation in morphological richness and average family size across languages, a facilitatory effect of large family size has been documented in a range of languages including, Dutch, German, English, and Finnish (Baayen et al., 1997; Ford et al., 2003; Lüdeling and De Jong, 2002; Moscoso del Prado Martín et al., 2004a). Entropy-based measures are newer and capture more structure within the morphological family by incorporating the frequencies of individual family members as well as the number of members. Derivational entropy focuses on the derived forms that constitute a family and analogously, inflectional entropy focuses on the inflected forms within a paradigm where it uses the probability of a particular syntactic function of a particular inflected form expressed in information theoretic terms to predict decision latencies (Moscoso del Prado Martin et al., 2004c). Where large family size typically facilitates word processing, entropy measures provide for the interaction among family members where differences in surface frequency across individual members provides a correction to predictions based on family size alone (Kostić et al., 2003; Moscoso del Prado Martin et al., 2004c). Stated simplistically, the inflectional and derivational entropy measures are important because they encompass both word-specific and base-morpheme frequency and therefore do not lend themselves to a dual route account where whole word and constituent (morpheme) processes for a word are

140

The Mental Lexicon: Core Perspectives

mutually exclusive. The role of inflectional entropy has been investigated in English (Baayen et al., 2006) as well as in the morphologically rich language of Serbian. Another frequency measure, affix frequency or numerosity, also influences processing in unprimed contexts in Dutch (Bertram et al., 2000b), Finnish (Bertram et al., 2000a; Bertram et al., 1999), Italian (Burani and Thornton, 2003; Laudanna and Burani, 1995), as well as English (Feldman et al., 1995). Revealingly, an affix’s potential for homonymity or multiple functions (e.g., as in agentive ER [SWEEPER] or comparative ER [NEATER]) seems to enhance the effect of surface frequency (Baayen et al., 1997; Bertram et al., 2000a,b) and, at the same time, make it more difficult to detect an effect of base morpheme frequency. Insofar as it is a property of the affix, namely its multiple functions that seems to render whole word processes more salient, the effect is hard to reconcile with an account based on independent base morpheme analysis and whole word processing. A picture emerges where many researchers try to oppose various frequency measures and interpret effects as reflections of either constituent or whole word processing. Most typical is to assume that the routes work competitively so that the faster route wins, that the other option has no effect, and that there is no overall cost to operating two routes in parallel. Accordingly, evidence of whole word and analytic processes appeared to trade off so that the contribution of base morpheme frequency becomes more dominant as surface frequency decreases. For example, effects of surface frequency for high frequency regularly inflected forms suggested that they were stored as whole forms while the absence of an analogous effect for low frequency regular words suggested that they were likely to be decomposed (Alegre and Gordon, 1999). Limitations exist, however, because morpheme analysis can entail costs as well as benefits or null effects. For example in Finnish, decision latencies are slowed when processing morphologically complex nouns (i.e., inflected nouns), as compared to monomorphemic nouns (Laine et al., 1999). Further, Italian stems that combine with many inflectional affixes elicit slower decision latencies than those that combine with fewer (Traficante and Burani, 2003). Similarly, in some nonword contexts, latencies to lower frequency English words with more frequent base morphemes are harder to recognize than are words with less frequency base morphemes (Taft, 2004). Finally, competition between present and past tense inflected forms of the same verb has been documented for regulars as well as for irregulars (see Baayen, this volume). In fact, with more refined analytic techniques, frequency effects can be detected for regularly inflected forms over a full range of frequencies thereby invalidating claims for a switch in processing options as surface frequency increases (Wurm and Baayen, 2005). In contrast to traditional accounts with processing options that reveal either effects of surface or of morpheme frequency, these findings reveal conjoint effects of both measures and, therefore, are more consistent with models that permit complex and varied interactions between whole word and analytic processes.

Origins of Cross-Language Differences in Word Recognition

141

4.2. Modulation of Facilitation due to Frequency of Derivational Relatives and their Components Like effects of surface and base morpheme frequency for single words, patterns of facilitation to target decision latencies when prime and target are morphologically related provide a window on morphological processing. Once again, the absence of effects of morphemic structure is interpreted as evidence in support of the whole word processing option, whereas effects of constituent morphemic structure typically reflect analytic processing. In contrast to high frequency semantically related primes that tend to produce greater target facilitation than do lower frequency primes, higher frequency morphologically related words generally produce less facilitation than do low (Meunier and Segui, 1999, 2001; Raveh, 2002). Whereas attenuated facilitation could be interpreted as support for whole word processing of the prime, differences across prime types in forward masked facilitation fail to arise when targets are very high (600 per million) in frequency (Giraudo and Grainger, 2000). Stated simply, target surface frequency can modulate the activation that occurs within the lexicon and limit the priming potential of any particular prime, either a semantic associate or a member of the target’s morphological family. Although whole word processing of the prime is one account of attenuated facilitation, accounts that focus on magnitudes of facilitation without consideration of baselines lose much of their foothold when different prime-target relations are examined on different sets of prime-target pairs. In addition to properties of the base morpheme and target surface frequency, however, properties of the affix also influence morphological facilitation in a priming task and, consistent with a traditional processing framework, these effects are less amenable to an interpretation that entails a switch between independent processing alternatives. At a 250 ms stimulus onset asynchrony (SOA), (Feldman, submitted), target (HUNT) decision latencies were faster when specified by underlining in a low frequency morphologically complex form related by derivation (HUNTRESS), than in a higher frequency one (HUNTER). In addition, productivity of the affix influenced decision latencies such that decision latencies for targets after morphologically complex source words became more like the stem identity (HUNT) condition (i.e., smaller difference) as affix numerosity increased. Accordingly, the difference between target decision latencies for ACTIVE (low affix frequency) relative to ACT is larger than the difference between TEACHER (high affix frequency) and TEACH. Effects of affix numerosity are most salient in lower frequency words, which suggests that a property of the affix component is easier to detect in the context of whole words that are low in frequency. Here, routes fail to operate independently. Therefore, this pattern again calls into question the validity of interpreting the absence of a morphological effect (e.g., the failure to detect differential morphological facilitation when target frequency is high and recognition latencies are fast) as evidence of whole word processing. In Hebrew, where words are composed of a root and a word pattern, repetition of verbal (Deutsch et al., 1998), but not nominal (Frost et al., 1997) word patterns produces facilitation.

142

The Mental Lexicon: Core Perspectives

One interpretation draws on insights about affix homonymity, specifically evidence that repetition of homographic affixes failed to produce forward masked facilitation such that homogeneous French pairs such as BRISEUR-AMUSEUR with agentive EUR showed facilitation whereas non homogeneous pairs such as LARGEUR-AMUSEUR did not (Kazanina and Tremblay, 2006). Because the Hebrew materials were not included, it is not clear whether the nominal patterns in Frost et al. (1997) were semantically ambiguous and performed two or more functions. For example, when the -a-a- word pattern is infixed into a CC-C root, it can serve either nominally (GaMaD, dwarf; BdaVaR, thing; GaMaL, camel) as an adjective (KaTaN, small) or as a verb (KaTaV, he wrote). Historically, the vowels of the word patterns were distinct, but some differences have disappeared in modern Hebrew. The implication is that affix homonymity may have contributed to the failure to detect noun pattern facilitation in Hebrew (see also Feldman et al., 1995). Further, repetition of different but homonymous word patterns in prime and target may be a factor when morphologically unrelated, but orthographically similar pairs in Hebrew fail to influence recognition (Frost et al., 2005). To our knowledge, effects on morphological facilitation of affix homonymity in general, and effects of word pattern homonymity in particular, have not yet been systematically investigated in Hebrew. At this point, therefore, it seems premature to interpret an absence of morphological facilitation that, potentially, is associated with homonymity as evidence of a cross language difference with respect to where morphological processing does and does not occur.

4.3. Facilitation Due to Regular vs. Irregular Inflectional Relatedness: One Mechanism or Two? Formulation of irregularly inflected forms in English typically entails stem change (HOPPED vs. RAN), but can also entail irregular inflection (TOASTED vs. BURNT). Stem change may be less characteristic of irregularity in languages such as Hebrew (Berent et al., 2002) and Greek (Jarema et al., 2004), however. One of the most frequently investigated themes in the domain of morphology is whether morphologically related primes formed by regular inflection are recognized by the same processing mechanism as are irregulars, or whether irregular forms require a whole word processing mechanism. In priming tasks, magnitudes of facilitation that differ significantly from the baseline after regular primes in conjunction with nonsignificant facilitation after irregular primes, often are interpreted as evidence of competing processes. Thus, regular forms are decomposed into stem + affix and reactivation of the base morpheme in prime and then target produces facilitation. By contrast irregular forms, including irregular inflection, are represented as wholes in associative memory and activation between entries produces weaker facilitation. Notably, differences of differences in facilitation typically are not tested statistically. For example, research conducted on Hebrew with its infixing structure, as compared to the affixing structure of English, reports forward masked morphological

Origins of Cross-Language Differences in Word Recognition

143

facilitation for regular three consonant roots, but not for irregular forms (i.e., weak roots) that infix the word pattern into a two consonant root. Statistical comparisons of the two were never reported, however (Velan et al., 2005). One factor that complicates cross language comparisons of processing for regular and irregular inflectionally related pairs is the construction of baselines to calculate the magnitude of morphological facilitation. Some studies determine the size of the effect by comparing morphologically related word pairs to an (orthographically and semantically) unrelated baseline condition (e.g., Marslen-Wilson et al., 1994), whereas others compute the effect by comparing related items to an orthographically related (and semantically unrelated) condition (e.g., Forster and Davis, 1984; Frost et al., 2000a; Frost et al., 1997). As delineated above, orthographic similarity of the prime can inhibit the processing of the target. Therefore, it is possible that comparing latencies after a morphologically related prime to latencies after an orthographically related prime artificially magnifies the difference by encompassing a combination of both morphological facilitation and orthographic inhibition. On the other hand, regular and irregularly inflected forms tend to differ in their similarity to the base morpheme and that variation should be controlled rather than neglected as is the case with an unrelated baseline. Pastizzo and Feldman (2002) compared forward masked English target decision latencies after unrelated, orthographically similar and morphologically related primes for regular and irregular targets matched on frequency and number of neighbors. Over all target types, decision latencies after orthographically similar primes and unrelated forward masked primes did not differ significantly. Planned comparisons indicated that morphologically related primes significantly reduced response latencies for regular prime-target pairs such as HATCHED-HATCH pairs. However, due to nonsignificant orthographic inhibition, facilitation was more reliable when assessed relative to the orthographic baseline than an unrelated baseline. Choice of orthographic vs. unrelated baseline did not alter the assessment of morphological facilitation for irregular and formally dissimilar TAUGHT-TEACH type items. As occurs for irregular verbs in Hebrew, neither comparison was significant. Crucially, for native speakers of English the presence of significant facilitation relative to an orthographic baseline failed to differentiate regularly from irregularly inflected but formally similar verb pairs such that FELL-FALL type pairs as well as HATCHED-HATCH type pairs produced facilitation. Note also, however, that morphological facilitation for the same irregular and formally similar FELL-FALL type items was significant when assessed relative to the orthographic (FILL-FALL) but not relative to the unrelated (PAIR-FALL) baseline. In a context where magnitudes of facilitation vary in a graded rather than an all-or-none manner, a focus restricted to the presence or absence of significant facilitation generally constitutes an inadequate portrayal of morphological processing. Obviously, comparisons become more sensitive and the potential to detect differences in the magnitude of facilitation is enhanced if regular and irregular morphologically related primes can precede the same target (Feldman and Fowler, 1987; Fowler et al., 1985). Alternatively, differences of differences can

144

The Mental Lexicon: Core Perspectives

be tested statistically where reports of a significant interaction between verb type (regular vs. irregular) and prime type (morphological vs. unrelated) provide support for different processing mechanisms for regular and irregular inflection. In sum, the absence of forward masked facilitation for a subset of irregulars with attenuated form overlap arises both in morphologically impoverished English, and in morphologically rich Hebrew. Moreover in English, the ubiquity of item-specific (surface frequency) as well as base morpheme frequency contributions and the comparability of magnitudes of cross modal facilitation for irregulars (GROWN-GROW) matched to regulars (GUIDED-GUIDE) on degree of form overlap (Basnight-Brown et al., submitted) suggest that for the recognition of regular and irregular verbs, similarities predominate over differences. In our view, neither well-designed studies that contrast morphological facilitation over languages nor studies that contrast morphological facilitation over regularity provide convincing evidence of a processing dichotomy.

4.4. Under Additive Effects of Form and Meaning vs. Morphological Relatedness Further insights into morphological processing derive from comparing orthographic and semantic priming effects with morphological effects within the same study, and interpretations focus on whether morphological facilitation is greater than the sum of the magnitudes of semantic and form-based similarity. Across many languages, experiments have included comparisons of morphological with either semantic or orthographic similarity and outcomes within a common methodology fail to diverge cross languages. In one such study (Feldman, 2000), critical pairs shared morphological (e.g., VOWEDVOW) as well as semantic (e.g., PLEDGE-VOW) and orthographic (e.g., VOWEL-VOW) relatedness. Morphological and semantic pairs were equated by ratings for semantic similarity, and morphological and orthographic pairs were equated for letter overlap between prime and target. Crucially, the same targets were counterbalanced over all types of primes. In the lexical decision task, when there was no mask and both prime and target appeared visually, morphological facilitation was evident over a range of prime-target SOAs and effects of orthographic similarity (when primes had lower frequencies than their targets) tended to be numerically facilitatory or absent at SOAs of 32 and 66 ms, and became inhibitory at longer SOAs. Although semantic relatedness produced less facilitation than did morphological relatedness, both dimensions showed a similar increase with SOA. Most important, the magnitude of morphological facilitation was significantly greater than the sum of the effects of shared form and shared semantics in isolation. Some theorists take under-additive effects of form and meaning relative to morphological relatedness to mean that morphological structure is represented independently of form and meaning (Longtin et al., 2003; Rastle et al., 2004), whereas for others it means that morphological effects arise from a more complex convergence of form and semantics

Origins of Cross-Language Differences in Word Recognition

145

(Plaut et al., 1996; Rueckl et al., 1997). This is another domain where numerical interactions often fail to reach statistical significance but graded differences proliferate both in priming outcomes and in degree of form or semantic similarity between prime and target. Insofar as form and semantic effects fail to differ across languages when task and target properties are controlled, there is little justification to anticipate that their underadditivity varies across languages.

4.5. Does Word-Internal Patterning of Morphemes Influence Morphological Processing? Accounts of morphological processing tend to focus on the base morpheme or root, but there is variation in how base morphemes combine with affixes to form a word and these combinatorial processes also may influence processing. In Hebrew, base morphemes generally occur with a word pattern and it is possible, therefore, that its distinctive word structure, with infixing of one morpheme within another, serves to alter aspects of morphological processing. It can be argued that apart from infixing, word structure in Hebrew is similar to that in languages such as Spanish and Serbian, insofar as words other than masculine singular forms (that have no affix), necessarily append an affix to a stem (or base morpheme) so that the typical stem seldom appears in isolation. The implication is that noun and verb targets in Hebrew, but also Spanish and Serbian, always are composed of at least two morphemes and contrast with targets in English that can be monomorphemic. To the extent that these situations are analogous, one can ask whether its base plus word pattern structure renders recognition in Hebrew different from languages where base morphemes are free and can appear without an affix. One clue to processing of stem-plus-affix combinations derives from studies with homographic stem morphemes. If morphemic stems are processed orthographically but not semantically at early stages of recognition, then prime target pairs that share unrelated but homographic stems should show facilitation. Homographic stem inhibition has been documented in Italian for PORTARE (carry) - PORTE (door) type pairs when processing time for the prime is relatively unconstrained (Laudanna et al., 1992), as well as in Spanish (Allen and Badecker, 2002). However the Italian results do not generalize from verbs to nouns (Laudanna et al., 2002) and the Spanish results have been difficult to replicate (see also Carreiras et al., 2006. In fact, after forward masked homographic stems (e.g., SER in SERIO [serious man], SERIA [serious woman] vs. SERIE [series, serial]) at an SOA of 66 ms, magnitudes of facilitation were 33 ms greater after morphological than orthographic primes (Duñabeitia et al., in press). Here, stems were homographic so that form similarity was fully matched. Nevertheless, even under forward masked presentation conditions, stems were not processed in isolation from their affix. Accordingly, differences in forward masked facilitation across stem plus affix combinations can be interpreted as early effects of mismatching semantics between formally similar prime and target.

146

The Mental Lexicon: Core Perspectives

Further, when the same base morpheme appears in two words so that the words truly are morphologically related, words can differ in the degree of semantics they share with the base morpheme even though only affixes differ. For example, the stem FIX appears to share more meaning with its morphological family member and relative FIXABLE (where semantic meaning is transparent), than with its relative FIXTURE (where the semantic meaning is opaque). Similarly in Hebrew, the same morpheme (SH-L-X) underlies SHaLaX (to send), NiSHKaX (to be extended) and SHaLiaX (emissary) although some assert that semantic diversity within a family tends to be greater in Hebrew than in other languages. Manipulations of semantic transparency between members of a morphological family have been the focus of much recent work on morphological priming. Generally, the results reveal larger magnitudes of facilitation for transparent as compared to opaque items, although this difference often is nonsignificant when prime presentation is very brief and processing time is limited. Failure to detect an effect of semantic transparency among morphological relatives, under forward masked conditions with short SOA durations has been replicated in several languages: English (Feldman et al., 2004b; Forster and Azuma, 2000; Rastle et al., 2000, 2004), Hebrew (Frost et al., 1997, 2005), Bulgarian (Nikolova and Jarema, 2002), French (Longtin et al., 2003) and Dutch (Diependaele et al., 2005). Nonetheless, across other presentation formats that are not necessarily restricted to early phases of processing, outcomes suggest that morphological processing can be sensitive to semantic similarity, even in languages that are highly inflected or morphologically rich in derivation, such as Serbian and Hebrew (Bentin and Feldman, 1990; Feldman et al., 2002; Frost et al., 2000b; Frost et al., 1997). Crucial in investigations of the interaction between semantics and morphology is the time course of morphological processing based on manipulations of SOA within a task (Boudelaa and Marslen-Wilson, 2005; Feldman et al., 2004b; Rastle et al., 2000). A second factor that influences semantic transparency, discussed in the next section, focuses on list composition. The results from time course studies in English (Feldman et al., 2004) reveal significant morphological facilitation for transparent items that is generally consistent across all SOA durations. In contrast, facilitation after opaque primes occurs only at the shortest SOA and likely reflects similarity of form in the absence of elaborated semantics. Revealingly, semantic facilitation likewise only occurs at the longest SOA, when presentation durations make primes visible. Collectively, the pattern invites a distinction between morphologicalform effects that arise early, and morphological-semantic effects that tend to emerge later in the process. The purported distinction based on time course between early form contributions and later semantic contributions is not specific to English. In fact, it is characteristic of Arabic, another Semitic language (Boudelaa and Marslen-Wilson, 2005) as well as Serbian, another highly inflected language (Feldman et al., 2002). Effects of homography and of semantic transparency of the base morpheme implicate a role for the affix in morphological processing. To elaborate, the particular combinational probability of a stem plus affix combination (Baayen, this volume) guides processing and affects patterns of facilitation in a manner that is not easily accommodated by traditional

Origins of Cross-Language Differences in Word Recognition

147

accounts, where form-based analysis of the base morpheme and semantically influenced whole word processing are independent options. Undeniably, the finding that transparent and opaque facilitation typically do not differ significantly under highly constrained conditions (forward masked with the shortest SOA length), has profound implications for claims about the way in which morphology is represented within a language and perhaps invites comparisons across languages with infixing and affixing structures. Consistent with Forster and Azuma’s (2000) characterization of the forward masked lexical decision task, Rastle et al. (2004) have suggested that the similar outcome for semantically transparent and opaque words occurs because the earliest stage of word recognition is more dependent on form, and less dependent on semantics (see also Longtin et al., 2003). By this account, morphological processing does not reflect conjoint effects of form and meaning and influences of semantics on early morphological processing are not anticipated.

4.6. The Influence of List Composition on Semantic Processing In addition to properties of the target and its morphological relatives, several experimental factors pertaining to list composition and the construction of word and nonword filler trials can modulate the magnitude of morphological, as well as semantic, forward masked facilitation. In the same way, the impact of these manipulations has implications for our understanding of early morphological processing. The relatedness proportion (RP), or proportion of related word trials relative to all word-word pairs, is a factor that can influence the magnitude of facilitation in semantic priming tasks (see Neely et al., 1989). Similarly, in the morphological priming task, an increase in RP from .50 to .75 enhanced morphological facilitation for Dutch words (Drews and Zwitserlood, 1995). The number and composition of filler items also are relevant. Specifically, the inclusion of identity trials (e.g., ANT-ANT) appears to increase overall semantic activation (Bodner and Masson, 2003; Feldman and Basnight-Brown, submitted). As described above, it is rare to observe effects of semantic similarity in the forward masked priming task; however, the inclusion of pseudohomophones (e.g., CERE, DAIR) or identity fillers (Bodner and Masson, 2003; Joordens and Becker, 1997) do appear to increase semantic activation at early stages of processing. In support of this claim, when proportion of orthographically similar word-word and word-nonword pairs was matched so that orthographic similarity was not a valid predictor of lexicality, Feldman and Basnight-Brown (submitted) observed a significant increase in semantic facilitation (a change from -4 ms to +23 ms) for the same set of targets when identity filler word pairs replaced filler items that were a mix of orthographically, semantically, and morphologically related, and RP was held constant. More importantly, the change in list composition significantly increased the magnitude of morphological facilitation (+25 ms to +53 ms).

148

The Mental Lexicon: Core Perspectives

Of more relevance, the influence of semantics early in morphological processing was examined recently in word pairs where the degree of semantic transparency and target neighborhood size were manipulated (Feldman et al., submitted). Half of all word nonword pairs consisted of pseudohomophones (e.g., DAIRY-DAIR; GROCER-GROCE). Each target (e.g., FIX) was paired with a transparent morphological relative (FIXABLE), a partially transparent morphological relative (FIXTURE), an orthographically unrelated word (FITNESS), and an unrelated word (EVASIVE). Large and small N targets were matched on frequency (60 [SD = 74] vs. 73 [SD = 74] per million) but differed on neighborhood size and length such that small neighborhoods had an average size of 1 (SD = 1.1) and a length of 5.7 (SD = .9), and large neighborhoods of 11 (SD = 3.9) and a length of 4.1 (SD = .6). The semantic relationship between the transparent, partially transparent, and opaque items differed significantly both by ratings of semantic relatedness and by LSA values; differences were matched across large and small N targets (Landauer et al., 1998). Table 2 Forward masked decision latencies for targets with small and large orthographic neighborhoods after morphologically transparent, partially transparent, orthographically similar and unrelated primes PRIME TYPE TRANS

PARTIAL

ORTHO

UR

Small N

601

615

623

632

Large N

597

592

628

611

The results revealed a significant interaction between neighborhood size (and length) and prime type. In the context of pseudohomophone fillers as well as identity trials that emphasized a word’s orthographic structure, latencies to targets with large neighborhoods were significantly slower after orthographic as compared to the unrelated primes, and latencies after transparent and partially transparent primes did not differ (see Table 2). These large N results are reminiscent of findings in Hebrew; there was no effect of semantic transparency among morphological relatives and no form facilitation. Most notable was that for targets from small neighborhoods, response latencies after transparent and partially transparent primes significantly differed one from another. In addition to effects of transparency that are more salient when neighborhood size is small, one further implication of these findings is that it is premature to interpret differences in morphological processing between Hebrew and other languages, with respect to the absence of an effect of semantic transparency, as a reflection of the unique root plus word pattern structure of Hebrew where base morphemes for nouns and verbs fail to appear in isolation.

Origins of Cross-Language Differences in Word Recognition

149

Additional, albeit indirect, support for the non-distinctiveness of Hebrew morphological processing derives from a study in English, where primes as well as targets were affixed (Feldman et al., 2004b). Magnitudes of facilitation for affixed targets relative to an unrelated baseline failed to differ significantly for transparent (ACCORDING-ACCORDANCE) and opaque (ACCORDION-ACCORDANCE) word pairs both at an unmasked SOA of 48 ms (14 and 20 ms, respectively) and at a forward masked SOA of 83 ms (22 ms and 15 ms, respectively). Moreover, similar structures exist in Spanish where affixes are concatenated to stems to form morphologically complex words and there magnitudes of forward masked morphological facilitation at a 66 ms SOA were greater for transparent (46 ms) than for opaque (19 ms) prime target pairs (Carreiras et al., 2006). Curiously in Basque, where words tend to be longer due to agglutinating word formation processes, target facilitation after opaque and transparent prime pairs failed to differ under the same conditions (Carreiras et al., 2006). Collectively, morphological facilitation in forward masked priming is a robust finding, where the size of the effect depends on several factors including target frequency, family size, neighborhood size and length, SOA, RP, type and proportion of filler items, and structure of baseline items. Introducing different targets for each type of relatedness necessarily compromises the potential to detect an effect of opaque vs. transparent prime type. Effects of a target’s morphological family size and of its orthographic neighborhood size have been documented across several languages and are consistent with accounts that emphasize strength of mappings between form (orthographic, phonological) and semantic codes and interpret morphological processing in terms of the systematicity with which similar form is predictive of similar meaning. Likewise, effects of a target’s orthographic neighborhood have been documented where effects likely reflect competition between orthographically similar but semantically dissimilar forms. Stated generally, a large morphological family size as well as a large orthographic neighborhood render effects of morphological prime type more difficult to detect. In sum, as one moves away from the dichotomous processing framework based on analytic form and whole word semantic routes that work independently, and considers seriously graded magnitudes of facilitation as a function of target type and degree of similarity between prime and target, the case for differing modes of processing across languages is not as compelling as once asserted. Specifically, magnitudes of forward masked morphological facilitation vary with degree of form similarity between prime and target and, in contexts that foster semantic analysis, they can be influenced by the semantic transparency of the shared base morpheme. Outcomes seem to depend more on methodological variation than on language, however.

150

The Mental Lexicon: Core Perspectives

5. CONCLUSIONS We have asked whether the prevalence of whole word and analytic processing as revealed by the forward masked priming task varies across languages that differ with respect to the mapping of orthographic with either phonological or morphological structure. In the orthographic domain, as a rule, primes that are similar in form to the target that follows tend to slow decision latencies; this occurs across a variety of languages and definitions of orthographic similarity. Factors that influence the effect of orthographic similarity between prime and target under forward masked presentation conditions include number of neighbors, length and frequency of target, lexicality of the prime, composition of filler items, as well as task and presentation conditions for the prime. Our assessment is that variation in the number of neighbors and length of targets across languages may have created the illusion of systematic differences between Indo-European and Semitic languages with respect to the role of form similarity in visual word recognition. Among languages with alphabetic writing systems, orthographic depth and the accuracy of predicting the phonology from orthographic form likewise can vary. Word frequency and semantic relatedness provide measures of lexical processing and results in naming studies sometimes diverge across languages with differing orthographic depth. Here, variation across word types potentially provides insights into cross language variation. By one account, the presence of a whole word frequency effect is a marker for lexical or whole word processes and a frequency effect that is attenuated or absent, when the orthographic specification of phonology is accurate and otherwise complete, attests to variability in whole word recognition processes across languages. However, a more universal approach also considers the specifics of the task and any limits on processing that may influence production of a response. Collectively, without consideration of general processing constraints within a language and a task, speculation about potential language specific patterning due to orthographic depth or other factors is of limited value. Item specific as well as base morpheme frequency effects have been documented for regularly and irregularly inflected single words. Thus, effects of various morphologically derived frequency measures should not be interpreted as a reliable indicator that opposing whole word processing is absent. Further, magnitudes of morphological facilitation vary with degree of semantic and of form similarity between prime and target as well as with a target’s density (or number of neighbors) and its number of morphological relatives. Although form (dis)similarity and irregularity tend to covary, when degree of form similarity is high, as for GROWN-GROW type pairs, similarities continue to predominate over differences in priming by regulars and irregulars (Basnight-Brown et al., submitted). Finally, the finding that morphological facilitation in English generally is greater than the combined effects of orthographic and semantic similarity (Feldman, 2000; Rastle et al., 2000), demonstrates that morphological processing is not a simple byproduct of semantic and orthographic processing.

Origins of Cross-Language Differences in Word Recognition

151

Nevertheless, it remains plausible that morphological effects reflect a more complex interaction between form and meaning. Languages differ in their tendency to form new words by creating new combinations of morphemes and base morphemes can vary in a graded manner with respect to their degree of transparency within morphologically complex forms. Semantic influences on morphological facilitation due to transparency are robust and easy to detect at SOAs of 250 or longer in Hebrew (e.g., Bentin and Feldman, 1990; Frost et al., 2000b) and in Serbian (Feldman et al., 2002), two morphologically rich languages, as well as in English (Feldman and BasnightBrown, submitted; Marslen-Wilson et al., 1994), although SOA seems to be a limiting factor (Feldman and Soltano, 1999; Raveh, 2002; Raveh and Rueckl, 2000). Nonetheless, high proportions of related prime-target pairs (RP) and of identity trials, as well as nonword trials with a particularly word-like structure due to DAIRY-DAIR type homophony, each serve to augment magnitudes of forward masked morphological facilitation, thereby permitting effects of semantic transparency among morphologically related pairs to emerge. Evidently, in IndoEuropean, as well as in Semitic languages, early aspects of morphological processing are not limited to semantically similar prime-target pairs. More importantly, benefits of semantic similarity for morphologically related pairs do arise under forward masked presentation conditions and call into question a purely ortho-morphological stage of processing. In conclusion, graded effects of semantic transparency emerge across as well as within languages and therefore fail to provide convincing evidence of cross language variation. Future research likely will continue to examine orthographic, semantic and morphological processing in diverse languages, both alphabetic and non-alphabetic. Attention to graded effects invites a focus on whether various weightings of a common set of methodological factors and word properties including those identified in the current chapter (i.e., neighborhood size, family size, systematic manipulations of semantic transparency between items, and list composition) are adequate to account for any variation in word recognition processes across languages. It is only once a common set of factors is identified and empirically investigated in languages with diverse structures that one can accurately assess whether and how cross language differences in word recognition emerges. Based on the current discussion and on the empirical data to date, it appears that the same factors function throughout and that processing is more similar than different across Indo-European and Semitic languages.

152

The Mental Lexicon: Core Perspectives

REFERENCES Alegre, M. and P. Gordon (1999). Rule based versus associative processes in derivational morphology. Brain and Language, 68, 347-354. Allen, M. and W. Badecker (2002). Inflectional regularity: Probing the nature of lexical representation in a cross-modal priming task. Journal of Memory and Language, 46, 705-722. Andrews, S. (1989). Frequency and neighborhood effects on lexical access: Activation or search? Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 802-815. Andrews, S. (1992). Frequency and neighborhood effects on lexical access: Lexical similarity or orthographic redundancy? Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 234-254. Arduino, L. S. and C. Burani (2004). Neighborhood effects on nonword visual processing in a language with a shallow orthography. Journal of Psycholinguisitc Research, 33, 75-95. Baayen, R. H. (this volume). Storage and computation in the mental lexicon. Baayen, R. H., T. Dijkstra and R. Schreuder (1997). Morphological influences on the recognition of monosyllabic monomorphemic words. Journal of Memory and Language, 37, 94-117. Baayen, R. H., L. B. Feldman and R. Schreuder (2006). Morphological influences on the recognition of monosyllabic monomorphemic words. Journal of Memory and Language, 53, 496-117. Baluch, B. and D. Besner (1991). Visual word recognition: Evidence for strategic control of lexical and nonlexical routines in oral reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 644-652. Barca, L., C. Burani and L. S. Arduino (2002). Word naming times and psycholinguistic norms for Italian nouns. Behavior Research Methods, Instruments & Computers, 34, 424-434. Basnight-Brown, D., H. Chen, H. Shu, A. Kostić and L. B. Feldman (submitted). Monolingual and Bilingual Recognition of Regular and Irregular English Verbs: Does Sensitivity to Word Form Vary with Language Experience? Bentin, S. and L. B. Feldman (1990). The contribution of morphological and semantic relatedness to repetition priming at long and short lags: Evidence from Hebrew. Quarterly Journal of Experimental Psychology, 42A, 693-711 Berent, I., S. Pinker, and J. Shimron (2002). The nature of regularity and irregularity: evidence from Hebrew nominal inflection. Journal of Psycholinguistic Research, 31, 459-502. Bertram, R., M. Laine and K. Karvinen (1999). The interplay of word formation type, affixal homonymy and productivity in lexical processing: Evidence from a morphologically rich Language. Journal of Psycholinguistic Research, 28, 213-226. Bertram, R., M. Laine, R. Baayen, H. Schreuder and J. Hyona (2000a). Affixal homonymy triggers fullform storage, even with inflected words, even in a morphologically rich language. Cognition, 74, 1325. Bertram, R., R. Schreuder and R. H. Baayen (2000b). The balance of storage and computation in morphological processing: the role of word formation type affixal homonymy, and productivity. Journal of Experimental Psychology: Learning, Memory and Cognition, 26, 419-511. Bodner, G. E. and M. E. J. Masson (2003). Beyond spreading activation: An influence of relatedness proportion on masked semantic priming. Psychonomic Bulletin and Review, 10, 645-652.

Origins of Cross-Language Differences in Word Recognition

153

Boudelaa, S. and W. D. Marslen-Wilson (2005). Discontinuous morphology in time: Incremental masked priming in Arabic. Language and Cognitive Processes, 20, 207-260. Burani, C. and A. M. Thornton (2003). The interplay of root, suffix and whole-word frequency in processing derived words. In: Morphological Structure in Language Processing (R. H. Baayen and R. Schreuder, Eds.), pp. 157-207. Mouton de Gruyter, Berlin. Carreiras, M., J. A. Duñabeitia and M. Perea (2006, August). Effects of transparency on morphological decomposition. Poster presented at AMLaP 2006, Twelfth Annual Conference on Architectures and Mechanisms for Language Processing, Nijmegen, NL. Coltheart, M., E. Davelaar, J. T. Jonasson and D. Besner (1977). Access to the internal lexicon. In Attention and performance VI (S. Dornic, ed.), pp. 535-555. Academic Press, New York. Davis, C.J. and S. J. Lupker (2006). Masked inhibitory priming in English: Evidence for lexical inhibition. Journal of Experimental Psychology: Human Perception and Performance, 32, 668-687. Deutsch, A., R. Frost and K. I. Forster (1998). Verbs and nouns are organized and accessed differently in the mental lexicon: Evidence from Hebrew. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24(5), 1238-1255. De Moor, W. and M. Brysbaert (2000). Neighborhood frequency effects when primes and targets are of different lengths. Psychological Research, 63, 159-162. Diependaele, K., D. Sandra and J. Grainger (2005). Masked cross-modal morphological priming: Unraveling morpho-orthographic and morpho-semantic influences in early word recognition. Language and Cognitive Processes, 20, 75-114. Drews, E. and P. Zwitserlood (1995). Morphological and orthographic similarity in visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 21, 10981116. Duñabeitia, J.A., M. Perea, E. Gutiérrez, Y. Mena and M. Carreiras (in press). Priming morfológico: algo más que priming ortográfico. Anuario de Psicología. Feldman, L. B. (submitted). Lexical influences on morphological processing: Effects of low surface frequency on morphological family size. Feldman, L.B. (2000). Are morphological effects distinguishable from the effects of shared meaning and shared form? Journal of Experimental Psychology: Learning, Memory, and Language, 26 (6), 1431-1444. Feldman, L.B., D. Barac-Cikoja and A. Kostić (2002). Semantic aspects of morphological processing: Transparency effects in Serbian. Memory and Cognition, 30, 629-636. Feldman, L. B., D. M. Basnight-Brown and Y. Kanai (submitted). Parallels between forward masked semantic and morphological facilitation: Influence of list context invites a computational account of morphological processing. Feldman, L.B. and D. M. Basnight-Brown (submitted). Effects of semantic transparency in the forward masked morphological priming task depend on orthographic neighborhood size. Feldman, L. B. and S. Bentin (1994). Morphological analysis of disrupted morphemes: Evidence from Hebrew. Quarterly Journal of Experimental Psychology, 47A, 407-435. Feldman, L. B. and C. A. Fowler (1987). The inflected noun system in Serbo-Croatian: Lexical representation of morphological structure. Memory and Cognition, 15, 1-12. Feldman, L. B., R. Frost and T. Pnini (1995). Decomposing words into their constituent morphemes: Evidence from English and Hebrew. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 947-960.

154

The Mental Lexicon: Core Perspectives

Feldman, L. B. and J. Moskovljević (1987). Repetition priming is not purely episodic in origin. Journal of Experimental Psychology: Learning, Memory and Cognition, 13, 573-581. Feldman, L. B., G. Jarema, D. Perlak and M. J. Pastizzo (2004a). Defining orthographic neighbors in a highly inflected language: Do neighborhoods include inflected forms? In: Mental Lexicon Working Papers (G. Libben and K. Nault, eds.), pp. 109-123. University of Alberta, Edmonton, Canada. Feldman, L. B. and M. J. Pastizzo (2003). Morphological facilitation: The role of semantic transparency and family size. In: Morphological Structure in Language Processing (R. H. Baayen and R. Schreuder, eds.). Mouton de Gruyter, Berlin, Germany. Feldman, L. B. and E. G. Soltano (1999). Morphological priming: The role of prime duration, semantic transparency and affix position. Brain and Language, 60, 33-39. Feldman, L. B., E.G. Soltano, M. Pastizzo and S. E. Francis (2004b). What do graded effects of semantic transparency reveal about morphological processing? Brain and Language, 90, 17-30. Ferrand, L. and J. Grainger (1992). Phonology and orthography in visual word recognition: Evidence from masked nonword priming. The Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 45A, 353-372. Ford, M., W. D. Marslen-Wilson and M. H. Davis (2003). Morphology and frequency: contrasting methodologies. In: Morphological Structure in Language Processing (R. H. Baayen, and R. Schreuder, eds.), pp. 89-124. Mouton de Gruyter, Berlin. Forster, K.I. and T. Azuma (2000). Masked priming for prefixed words with bound stems: Does submit prime permit? Language and Cognitive Processes, 15, 539-561. Forster, K.I. and C. Davis (1984). Repetition priming and frequency attenuation in lexical access. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10, 680-698. Forster, K.I., C. Davis, C. Schoknecht and R. Carter (1987). Masked priming with graphemically related forms: Repetition or partial activation? Quarterly Journal of Experimental Psychology, 39A, 211-251. Forster, K. I. and D. Shen (1996). No enemies in the neighborhood: Absence of inhibitory neighborhood effects in lexical decision and semantic categorization. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 696–713. Forster, K.I. and C. Veres (1998). The prime lexicality effect: Form-priming as a function of prime awareness, lexical status, and discrimination difficulty. Journal of Experimental Psychology: Learning, Memory and Cognition, 24, 498-514. Frost, R., (1994). Prelexical and postlexical strategies in reading: Evidence from a deep and a shallow orthography. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 116-129. Frost, R., A. Deutsch and K.I. Forster (2000a). Decomposing morphologically complex words in a nonlinear morphology. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(3), 751-765. Frost, R., A. Deutsch, O.Gilboa, M. Tannenbaum and W. D.Marslen-Wilson (2000b). Morphological priming: Dissociation of phonological, semantic, and morphological factors. Memory and Cognition, 28, 1277-1288. Frost, R., K.I. Forster and A. Deutsch (1997). What can we learn from the morphology of Hebrew? A masked priming investigation of morphological representation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23 (4), 829-256. Frost, R. and L. Katz (1992). Orthography, Phonology, Morphology and Meaning. North Holland, Oxford, England

Origins of Cross-Language Differences in Word Recognition

155

Frost, R., L. Katz, and S. Bentin (1987). Strategies for visual word recognition and orthographical depth: A multilingual comparison. Journal of Experimental Psychology: Human Perception and Performance, 13, 104-115. Frost, R., T. Kugler, A. Deutsch and K. I. Forster (2005). Orthographic structure versus morphological structure: Principles of lexical organization in a given language. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31, 1293-1326. Fowler, C. A., S. E. Napps and L. 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. Giraudo, H. and J. Grainger (2000). Effects of prime word frequency and cumulative root frequency in masked morphological priming. Language and Cognitive Processes, 15, 421-444. Grainger, J. and L. Ferrand (1994). Phonology and orthography in visual word recogntion: Effects of masked homophone primes. Journal of Memory and Language, 33, 218-233. Harm, M. W. and M. S. Seidenberg (1999). Phonology, reading acquisition, and dyslexia: Insights from connectionist models. Psychological Review, 106, 491-528. Harm, M. W. and M. S. Seidenberg (2004). Computing the meanings of words in reading: Cooperative division of labor between visual and phonological processes. Psychological Review, 111, 662-720. Janack, T., M. J. Pastizzo and L B. Feldman (2004). When orthographic neighbors fail to facilitate. Brain and Language, 90, 441-452. Jarema, G., L. B. Feldman and D. Perlak (2004). Defining regularity: Does degree of phonological and orthographic similarity among Polish relatives influence morphological processing? In: Mental Lexicon Working Papers (G. Libben and K. Nault, eds.), pp. 91-108. University of Alberta, Edmonton, Canada. Johnson, N.F. and K. R. Pugh (1994). A cohort model of visual word recognition. Cognitive Pyschology, 26, 240-346. Joordens, S. and S. Becker (1997). The long and short of semantic priming effects in lexical decision. Journal of Experimental Psychology: Learning, Memory, & Cognition, 23, 1083-1105. Katz, L. and L. B. Feldman (1983). Relation between pronunciation and recognition of printed words in deep and shallow orthographies. Journal of Experimental Psychology: Learning, Memory, and Cognition, 9, 157-166. Kazanina, N. and M.C. Tremblay (2006, October). Representation of homonym bound affixes in the mental lexicon. Paper presented at the Fifth International Conference of the Mental Lexicon. Montreal, Canada. Kostić A., T. Marković and A. Baucal (2003). Inflectional morphology and word meaning: orthogonal or co-implicative cognitive domains? In: Morphological Structure in Language Processing (H. Baayen & R. Schreuder, eds.), pp.1- 45. Mouton de Gruyter, Berlin. Landauer, T.K., P. W. Foltz and D. Laham (1998). Introduction to latent semantic analysis. Discourse Processes, 25, 259-284. Laudanna, A., W. Badecker and A. Caramazza (1992). Processing inflectional and derivational morphology. Journal of Memory & Langauge, 31, 333-348. Laudanna, A. and C. Burani (1995). Distributional properties of derivational affixes: implications for processing. In: Morphological Aspects of Language Processing (L. B. Feldman, ed.), pp. 345-364. Erlbaum, Hove.

156

The Mental Lexicon: Core Perspectives

Laudanna, A., M. Voghera and S. Gazzellini (2002). Lexical representations in written nouns and verbs in Italian. Brain and Language, 81, 250-263. Longtin, C.-M., J. Segui and P. A. Halle (2003). Morphological priming without morphological relationship. Language & Cognitive Processes, 18(3), 313-334. Lüdeling, A. and N. De Jong (2002). German particle verbs and word formation. In: Explorations in Verb-Particle Constructions (N. Dehé, R. Jackendoff, A. McIntyre and S. Urban, eds.), pp. 315-333. Mouton de Gruyter, Berlin. Lupker S.J., P. Brown and L. Colombo (1997). Strategic control in a naming task: Changing routes or changing deadlines? Journal of Experimental Psychology: Learning, Memory and Cognition, 23, 570-590. Marslen-Wilson, W., L. Tyler, R. Waksler and L. Older (1994). Morphology and meaning in the English mental lexicon. Psychological Review, 101, 3-33. Mathey, S., C. Robert and D. Zagar (2004). Neighborhood distribution interacts with orthographic priming in the lexical decision task. Language and Cognitive Processes, 19, 533-559. Meunier, F. and J. Segui (2001). Cross-Modal Morphological Priming in French. Brain and Language, 68, 89-102. Meunier, F. and J. Segui (1999). Morphological Priming Effect: The role of surface frequency. Brain and Language, 68, 54-60. Moscoso del Prado Martín, F., R. Bertram, T. Häikiö, R. Schreuder and R. H. Baayen (2004a). Morphologically family size in a morphologically rich language: the case of Finnish compared to Dutch and Hebrew. Journal of Experimental Psychology: Learning, Memory, and Cognition., 30, 1271-1278. Moscoso del Prado Martin, F., M. Ernestus and R. H. Baayen (2004b). Do type and token effects reflect different mechanisms? Connectionist modelling of Dutch past-tense formation and final devoicing. Brain and Language, 90, 287-298. Moscoso del Prado Martín, F., A. Kostić and R. H. Baayen (2004c). Putting the bits together: An information theoretical perspective on morphological processing. Cognition, 94, 1–18. Neely, J.H., D. E. Keefe and K. L. Ross (1989). Semantic priming in the lexical decision task: Roles of prospective prime-generated expectancies and retrospective semantic matching. Journal of Experimental Psychology: Learning, Memory and Cognition, 15, 1003-1019. Nikolova, R. and G. Jarema (2002). Interaction of morphological structure and prefix transparency in the processing of Bulgarian verbs. Brain and Langauge, 81, 649-665. Pastizzo, M. J. and L. B. Feldman (2002). Discrepancies between orthographic and unrelated baselines in masked priming undermine a decompositional account of morphological facilitation. Journal of Experimental Psychology: Learning, Memory and Cognition, 28, 244-249. Peereman, R. and A. Content (1995). Neighborhood size effect in naming: Lexical activation and sublexical correspondences? Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 409-421. Perea, M. and E. Rosa (2000). The effects of orthographic neighborhood in reading and laboratory word identification tasks: A review. Psicologica International Journal of Methodology and Experimental Psychology, 21, 327-340. Plaut, D. C. and R. Frost (2001, November). Does morphological structure in Hebrew reduce to surface structure? Paper presented at the 42nd annual meeting of the Psychonomic Society: Orlando. USA.

Origins of Cross-Language Differences in Word Recognition

157

Plaut, D. C., J. L. McClelland, M. Seidenberg and K. E. Patterson (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56–115. Pollatsek, A., M. Perea and K. S. Binder (1999). The effects of neighborhood size in reading and lexical decision. Journal of Experimental Psychology: Human Perception and Performance, 25, 1142-1158. Rastle, K., M. H.Davis, W. D. Marslen-Wilson and L. K. Tyler (2000). Morphological and semantic effects in visual word recognition: A time course study. Language and Cognitive Processes, 15, 507-537. Rastle, K., M. H. Davis and B. New (2004). The broth in my brother's brothel: Morpho-orthographic segmentation in visual word recognition. Psychonomic Bulletin & Review, 11(6), 1090-1098. Raveh, M. (2002). The contribution of frequency and semantic similarity to morphological processing. Brain and Language, 3, 12-325. Raveh, M. and J. G. Rueckl (2000). Equivalent effects of inflected and derived primes: Long-term morphological priming in fragment completion and lexical decision. Journal of Language and Memory, 42, 103-119. Rueckl, J.G., M. Mikolinski, M. Raveh, C. S. Miner and F. Mars (1997). Morphological priming, fragment completion, and connectionist networks. Journal of Memory and Language, 36, 382-405. Schreuder, R. and R. H. Baayen (1997). How complex simplex words can be. Journal of Memory and Language, 37, 118–139. Sears, C. R., Y. Hino and J. Lupker (1995). Neighbourhood size and neighbourhood frequency effects in word recognition. Journal of Experimental Psychology: Human Perception & Performance, 21, 876-900. Segui, J. and J. Grainger (1990). Priming word recognition with orthographic neighbors: Effects of relative prime target frequency. Journal of Experimental Psychology: Human Perception and Performance, 16, 65-76. Seidenberg, M. S. and J. L. McClelland (1989). A distributed, developmental model of visual word recognition. Psychological Review, 96, 523-568. Shibahara, N., M. Zorzi, M. P. Hill, T. Wydell and B. Butterworth (2003). Semantic effects in word naming: Evidence from English and Japanese Kanji. The quarterly journal of experimental psychology: Human Experimental Psychology, 56A, 263-286. Simpson, G.B. and H. Kang (1994). The flexible use of phonological information in word recognition in Korean, Journal of Memory and Language, 33, 319-331. Strain, E., K. Patterson and M. S. Seidenberg (1995). Semantic effects in single word naming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1140-1154. Tabossi, P. and L. Laghi (1992). Semantic priming in the pronunciation of words in two writing systems: Italian and English. Memory and Cognition, 20, 303-313 Taft, M. (2004). Morphological decomposition and the reverse base frequency effect. Quarterly Journal of Experimental Psychology, 57A, 745-765. Traficante, D. and C. Burani (2003). Visual processing of Italian verbs and adjectives: The role of the inflectional family size. In: Morphological Structure in Language Processing (R. H. Baayen and R. Schreuder, eds.), pp. 45-64. Mouton de Gruyter, Berlin. Van Heuven, W., T. Dijkstra, J. Grainger and H. Schriefers (2001). Shared neighborhood effects in masked orthographic priming. Psychonomic Bulletin & Review, 8, 96-101.

158

The Mental Lexicon: Core Perspectives

Velan, H., R. Frost, A. Deutsch and D. Plaut (2005). The processing of root morphemes in Hebrew: Contrasting localist and distributed accounts. Language and Cognitive Processes, 20, 169-206. Wurm, L. H. and R. H. Baayen (2005, June). Surface frequency effects, even for suffixed inflections, below the magic number six. Paper presented at Perspectives on Morphological Processing, Cambridge, UK. Ziegler, J.C. and C. Perry (1998). No more problems in Coltheart’s neighborhood: Resolving neighborhood conflicts in the lexical decision task. Cognition, 68, 53-62. Zwitserlood, P. (1996). Form priming. Language and Cognitive Processes, 11, 589-596.

8 PRODUCTIVITY IN WORD FORMATION∗

Wolfgang U. Dressler, University of Vienna, Vienna, Austria

1. INTRODUCTION The lexical inventory of a language is always changing. Words fall into disuse, and new words are created to replace them or to meet new lexical demands. This chapter focuses on wordformation productivity, which involves the creation of new words, using existing morphemes of the language. The phenomenon of productivity in word formation (WF) has generated a flurry of linguistic publications in recent years, including monographs such as Bauer (2001), Plag (1999), Bolozky (1999) and collections of articles, such as Dal (2003a) and Aronoff and Gaeta (2003). In addition, recent overviews have been provided by Bauer (2005) and Rainer (2005). My goal in this chapter is to present a clarification of the core terms referring to concepts of productivity and to treat the lacunae in the above-mentioned overviews that are most relevant to core perspectives on the mental lexicon. We might begin, then, with a consideration of how notions of productivity are relevant for basic questions of the mental lexicon. First and foremost, issues of productivity interface with questions of whether it is appropriate to consider lexical innovation as involving symbolic rules (e.g., Pinker, 1999; Clahsen, 1999; Clahsen et al., 2003). If it is, then the productivity of a morphological pattern can be considered as a major motif for assuming lexical computation in the form of such symbolic rules. ∗

I have to thank, for using still unpublished material from child language, Marijan Palmović (Croatian); Péter Bodor and Virág Barcza (Hungarian); Anna de Marco, Sabrina Noccetti and Livia Tonelli (Italian); and Ineta Savickiene (Lithuanian); and from aphasia, Jackie Stark, Christiane Pons, Coralie Riedler and Katharina Korecky-Kröll (German) and Sándor Kovács and Ferenc Kiefer (Hungarian).

159

160

The Mental Lexicon: Core Perspectives

Productivity can also serve as a testing ground for the examination of cases in which modelling by rule is psycholinguistically inappropriate. Unproductive morphological patterns are often assumed not to be amenable to modelling by rules, but rather to be simply stored in the mental lexicon as simple and complex words, even if the respective complex word can be technically derived by a rule from the corresponding simple word. Examples of such cases would include the English words warmth, strength and warm, strong. The forms are linguistically related, and can be described in a rule-like manner. However, it has been claimed that these rules are not used at present to create new English words that follow this pattern. In some dual route models (Pinker, 1999; Clahsen, 1999) the distribution between computation or (de)composition and lexical access to stored items is strictly complementary, with some words being stored and others computed. As I will discuss below, this dichotomy becomes problematic, particularly if one assumes a gradable concept of productivity, as will be done in sections 2, 3, and 5 below. In models where (de)composibility and storage overlap, high productivity may increase the probability of (de)composition, e.g., in the race models of Baayen and Schreuder (1999; cf. Frauenfelder and Schreuder, 1992) and Pinker and Ullman (2002). This role of productivity in the balance of storage and computation in morphological processing has been shown explicitly for weakly inflecting languages such as Dutch, by Bertram, Schreuder and Baayen (2000) and for the strongly agglutinating language, Finnish, by Bertram et al. (1999; cf. Järvikivi, 2003, pp. 48-50). It has even been claimed by Vannest et al. (2002) that an insufficient degree of productivity results in the barring of decomposition in processing. Note that, in rule models, the possibility of computation by rule is not limited to transparent concatenative word formation rules, but also holds for morphological patterns where the simple word base and/or the affix is modified, as by velar softening in the English word pairs electric–electric-ity or admit–admission, as well as for conversions where overt morphotactic change is missing, as in the English to cut vs. a cut. Thus, the isolation of a base and an affix as distinct symbolic variables (cf. Berent et al., 2002) is not a necessary prerequisite to computation by rule. Structuralist models and earlier models of historical linguistics have typically focussed on the notion of type frequency of a morphological pattern as a sign of its productivity, a view partially taken up again by Bybee (1988) and Plag (1999). But the number of words formed according to a specific pattern or word-formation rule (henceforth WFR) may be an historic accident (see Bauer, 2001, pp. 20-21, 2005, pp. 328-329). An early critic within structuralist approaches was Marchand (1969) who relegated frequency to the traditional stock of accepted word formations, called word-formedness (German: Wortgebildetheit) and focussed any synchronic study onto active, productive word formation (German: Wortbildung). With this move, Marchand (see also Kastovsky, 2005) was a direct forerunner of Schultink (1961) and Aronoff (1976) who anchor productivity in the quasi-infinite number of possible formations (i.e., potential words) that can be generated by a productive word formation rule.

Productivity in Word Formation

161

Within a rule approach, another relevant notion, regularity can be distinguished from productivity as homogeneity of both input and output, which is a property of all productive WFRs and also many of the unproductive WFRs (cf. Dressler and Ladányi, 2000a, p. 105; Dressler, 2003, p. 36; Bauer, 2001, pp. 54-56, 2005, p. 329). In the dual-route approach to inflection (extended to word formation by Clahsen et al., 2003) among competing rules, the default rule (cf. Bauer, 2005, pp. 328-329), which is often also the most or only productive rule, is considered the regular one, whereas the competing patterns are considered to be irregular and thus not manipulated by rule mechanism but stored in their output forms. This approach fails to achieve adequacy for the competing productive inflection classes of strongly inflecting languages (e.g., the Slavic and Baltic languages, cf., Dressler, 2003) and is even less adequate for word formation with its frequent rivalry among competing productive WFRs. One instance is the use of German denominal adjective formation rules with the competing, productive suffixes -ig, -isch, -lich (as in wind-ig “wind-y”, stürm-isch “storm-y”, freund-lich “friend-ly”), none of which represents the default (cf. Pounder, 2000). The same holds for English denominal verb formation (cf. Plag, 1999), where neither conversion (e.g., to scapegoat) nor suffixations of -ize, -ify, -ate represent the default. A further relevant notion, creativity, can be either identified as a hyperonym of productivity (as in Chomsky’s rule-governed vs. rule-changing/violating creativity) or as referring to conscious, often not rule-governed, creation of new words. In this contribution I will follow this latter tradition (cf. Bauer, 2005, pp. 329-331). The theoretical linguistic model that I will use is semiotically-based Natural Morphology (cf. Dressler et al., 1987; Kilani-Schoch and Dressler, 2005; Dressler, 2005). The use of this framework is of importance for the following discussion only insofar as symbolic WFRs are assumed, which handle motivation of the morphosemantics and morphotactics of derived words at the same time. For example, the root and the suffix of compos-ition motivate both meaning and form of the derivative. Morphosemantic motivation refers to rule-predicted word-formation meaning (German: Wortbildungsbedeutung, cf. Corbin’s, 1987, notion of sens construit), whereas the exact, lexicalised word meaning (German: Wortbedeutung) of actual accepted words is stored in the lexicon. Thus the WF meaning of compos-ition is “act/result of composing”, which is materialized in a set of closely related senses of word meaning, which are motivated but not fully predicted by the WF meaning, e.g., the senses of “mental constitution, compromise”. Note that word meaning differs from word formation meaning much less in the related derivative de-compos-ition.

2. GRAMMATICAL PRODUCTIVITY: SYNTAGMATIC DIMENSION Grammatical productivity in WF can still be defined according to Schultink (1961; translated by van Marle, 1985, p. 45) as “the possibility for language users to coin, unintentionally, a number of formations which are in principle uncountable”. Unintentional does not mean that

162

The Mental Lexicon: Core Perspectives

intentional coinages have to be a priori excluded (cf. Bauer, 2001, pp. 66-70; Dal, 2003b, pp. 7-8) and uncountable does not mean infinite in the mathematical sense, but just without definite upper limit. This concept of WF productivity belongs to the level of grammatical WF competence that accounts for potential words (in the sense of Aronoff, 1976). This level is close to Coseriu's (1975) notion of language as potential system and to Chomsky's (1986) notion of internal language. This level has to be sharply distinguished, at least analytically, from external language, or more precisely, language as an institutional system of norms (see section 5). This level contains actual and accepted words, and grammatical productivity of the first level translates on this second level to actual type frequency of complex words derived via the same WFR. In my model of Natural Morphology (Dressler, 2003, 2005; Kilani-Schoch and Dressler, 2005), the first level represents dynamic morphology, with productive WFRs as its core. The second level represents static morphology consisting of stored morphologically derived words. These words are lexicalised. But if they are formed by a productive WFR, they belong to dynamic morphology as well. And, at least marginally, the domain of dynamic morphology may be extended to the application of unproductive rules (cf. Aronoff’s, 1976, notion of redundancy rules). Even on the morphosemantic side, WF meaning can also be accounted for by rules in the case of unproductive WFRs1. As is the case with syntax, phonology and inflection, I assume that rules of WF are also prototypically productive. This fits with the WFR function of lexical enrichment, i.e., of forming potential words, whose number is in principle non-finite. The property of being productive or unproductive is thus a basic property of a WFR. Therefore, if a WFR appears to be productive in one domain, but unproductive in another one, then we have an argument for postulating that this is not a unique rule, but two different rules (cf. Fradin et al., 2003). Thus, for example, Berrendonner and Clavier (1997) describe two French types of suffixation involving -age, one being productive, the other being unproductive but still regular. In German, denominal adjective formation via the suffix -ig is productive, whereas deverbal -ig adjective formation is not. This invalidates certain counter-examples against Aronoff’s (1976, p. 48) unitary base constraint (as in Plag, 1999, pp. 47-48, 2004), i.e., the assumption that prototypically WFRs apply to a homogeneous class of base words. Productivity may be graded, because gradedness is inherent to all models that include prototypes. Our proposal (Dressler, 2003; Dressler and Ladányi, 2000a) has been to grade degrees of grammatical productivity according to the relative severity of structural obstacles a rule has to overcome in order to perpetrate its productivity. The five thresholds of this scale of productivity are discussed below.

1

Extragrammatical operations of so-called expressive morphology, e.g., abbreviations of various kinds, blends (e.g., English smoke & fog smog) or echo-word formation (e.g., English zigzag, ticktock, cf. Dressler, 2000) are not subject to grammatical competence and behave differently from WFRs in many ways, and thus remain outside our discussion.

Productivity in Word Formation

163

2.1. Criterion (a): Morphological Integration of Loan Words with Unfitting Properties The most severe obstacles to rule application show up in the morphological integration of loan words with unfitting properties, which have to be fitted (accommodated) to the system adequacy of the loaning language. This criterion is the most important one, because a rule must have maximum productivity in order to overcome the two obstacles of foreignness and unfitting properties. Foreign words that enter the target language are stranger than indigenous words and therefore are more liable to shun morphological integration by application of a WFR. Newly loaned words are also easier to identify as new, and thus not traditional, as compared with new indigenous words. To this obstacle to rule application a new obstacle is added, if these new loan words have a structural property which does not fit the structural description of the WFR or lack a structural property which is an obligatory part of its structural description. These unfitting properties have to be changed in order to apply the WFR. A case in point is the adaptation of the Russian word [šo’se] from French chaussée “road”. It is phonologically non-integrated with its unstressed [o] (instead of [a], which would show the application of the phonological rule of akanje and its word-final stressed [e] and morphologically non-integrated, because it is undeclinable (as all similar loaned neuters). Thus this loan word has three unfitting properties which represent an obstacle to the application of morphological rules, because phonologically unintegrated loan words can hardly undergo the application of a morphological rule (cf. Dressler and Ladányi, 2000a, pp. 119-120) and the prohibition of application of an inflectional rule sets a precedence for other morphological rules. The two phonological unfitting properties are adapted in the derived adjective [ša’sej]nyj, with normal akanje of the unstressed [o] (obligatory in indigenous words) and addition of the glide [j] to the root-final stressed [e]. Akanje applies also in the denominal verb [ša’s]-irov-at’. These adaptations as well as the application of the suffixation WFR vouches for the highest degree of grammatical productivity of both WFRs. In a stratified lexicon, adaptations of unfitting properties may apply also to a marked stratum, such as to the Latinate stratum of English vocabulary, as in Plag’s (1999, pp. 126-128, 157, 162-164, 175, 186, 272, 279) examples (unless they were intentional, sophisticated coinages): (1)

féminìne fémin-ìze pátinà pátin-ìze

where the WFR of -ize suffixation does not apply concatenatively but substitutes the rhyme of the last syllable in order to arrive at a trisyllabic structure with a primary accent at the first syllable and a secondary accent at the last.

164

The Mental Lexicon: Core Perspectives

2.2. Criterion (b): Morphological Integration of Loan Words with Fitting Properties The obstacle to rule application is smaller, if a foreign word has fitting properties. Here the application of a WFR must overcome only the obstacle of foreignness. Therefore this is a less certain criterion for full productivity. The above-mentioned German WFRs of denominal adjective formation via the suffixes -ig, -isch, -lich fulfill this criterion, as in their application to English loans in: (2)

trend-ig film-isch sport-lich mainstream-ig (Profil Magazine, 21.8.2006)

Several German noun-compound formation rules fulfill this criterion as well (cf. Dressler et al., 2005), i.e., simple concatenation and concatenation with interfixation of the interfix /n/ after word-final schwa (written <e>) of the first (feminine or masculine) member, of the interfix /en/ after feminines not ending in schwa and substituting final /a/, as in: (3)

Laser#drucker “laser printer” Haupt#computer “main computer” Gruppe-n#bildung “group formation” (< French groupe) Coyote-n#fell “hide of a coyote” Farm-en#kauf “acquisition of a farm” Pizz-en#verkauf “pizza sale” Datsch-en#bau “construction of a datcha” (< Russian dača).

2.3. Criterion (c): The Application of Word Formation Rules to Abbreviations A lower criterion of productivity is the application of WFRs to abbreviations. Since these are formed by extragrammatical means (as discussed above), they are less system-adequate than traditional indigenous words. This partial strangeness makes them an obstacle to morphological integration, albeit less so than foreign words. Examples from Russian are: (4)

emgeuš-nik “student of the MGU” ( = Moscow State University) vuz-ov-ec “student of a college” (VUZ with interfix -ov-)

Productivity in Word Formation

165

2.4. Criterion (d): Shift from One Word Formation Rule to Another Hierarchically still lower as a productivity criterion is rule shift, i.e., shift of a derived word from one WFR to another one, typically from a less productive to a more productive WFR (although productivity of both WFRs may be very slight). In German deadjectival quality nouns, the unproductive suffix -e has often been substituted with one of the suffix allomorphs -heit/-keit/-igkeit, especially if the WF in -e was homophonous with the feminine base form of the adjective (see Osman, 1994), as in: (5)

Fein-e > Fein-heit “fine-ness” Heiter-e > Heiter-keit “cheerful-ness” Leicht-e > Leicht-igkeit “easi-ness” Schön-e > Schön-heit “beauty”

Here the obstacle lies in the necessity to substitute an existing, lexicalised and thus stored word (pace Sánchez Miret, 2006; who wants to downgrade this criterion).

2.5. Criterion (e): Novel Application of Word Formation Rules to Indigenous Words The last and hierarchically lowest criterion is the novel application of WFRs to traditional indigenous words (e.g., for German -ig suffixation see Peschel, 2002, pp. 151-165). Here the only obstacle is that a new word must be created and added to the lexicon, i.e., listed, accepted and thus institutionalized (cf. Hohenaus, 2005). Clearly not yet generally accepted new words are more relevant than already well-accepted neologisms, because usually the latter have been formed much earlier than the former ones. Furthermore, neologisms directly derived from well-accepted words are less relevant than neologisms derived via intermediate potential but non-actual words, i.e., via intermediate false steps (Rainer, 1997), as in the Russian example reported by Zemskaja (1996, p. 108): (6)

tret’emir-iz-acija “process of transforming a country of the second world, such as Russia in the 90s, into a country of the third world”  tret’ij mir “third world” via the merely potential verb tret’emir-iz-at’.

For further subdivisions see Dressler and Ladányi (2000a, pp. 124-127) and below. These criteria enable the establishment of a scale of productivity, from full productivity (if the highest possible criterion is fulfilled) down to a minimal degree of slight productivity, if only the last criterion (e) is fulfilled. Note that, by chance, no foreign words might be loaned whose properties do not fit the structural description of a WFR or no abbreviations may be conceivable bases of a WFR, which makes criteria (a) or (c) inapplicable.

166

The Mental Lexicon: Core Perspectives

Scales of productivity of WFRs (cf. Bauer, 2001; pp. 15-20, 125-162, 2005, pp. 330331) are usually established for the levels of norms or performance. Thus the above scale, first established for Italian and Polish inflection in 1994 (cf. Dressler, 2003) and modified for WF in Dressler and Ladányi (2000a), has been one of the rare attempts to postulate scalarity at the level of the potential grammatical system of WF. An earlier claim by Booij (1977, p. 5) that “the qualitative productivity of a word formation rule is inversely proportional to the number of competence restrictions on that rule” has been first criticised by Baayen (1989, p. 12). In the same vein, Dressler and Ladányi (2000a, pp. 111-116) have argued that ordinal number formation is highly productive in several languages (but not in Russian), because applicable to invented numbers, as in nth, xth and, potentially, mth, pth, etc., although the domain is extremely restricted, i.e., to numbers and their combinations. A similar argument has been made by Plag (1999, p. 15) for the productive derivation of English nouns in -ment derived from verbs with the prefix en/em-, which is now an unproductive prefix. One type of case where Booij’s claim may hold is the restriction of rule application to a marked domain within a stratified lexicon, as in the case of those English suffixes which are restricted to the Latinate vocabulary (with the caveats in Plag, 1999). Thus it is reasonable to say that English -ity is less productive than English -ness, cf. opac-ity and opaque-ness in contrast to wet-ness vs. ungrammatical *wet-ity (cf. Dressler and Ladányi, 2000a, pp. 131132). This dimension of grammatical productivity is orthogonal to the previous one. A consequence of different degrees of productivity is the principle of affix stacking that more productive affixes have a more peripheral position than less productive ones (cf. Hay, 2002; Rainer, 2005, pp. 339-341), including cases of hypercharacterization, as in Dressler (2004). Our system of grading productivity clearly shows that productivity is not just a matter of language use, at least in a model that differentiates in some way between competence and performance (cf. Bauer, 2001, pp. 29-32).

3. GRAMMATICAL PRODUCTIVITY: PARADIGMATIC DIMENSION The syntagmatic dimension of grammatical productivity discussed above focuses on an individual WFR in its input-output relation. However, both inputs (base words) and outputs (derived words) are related to other existing or potential words of the lexicon. Additionally, WFRs maybe interrelated, especially if they compete for the same inputs. This represents the paradigmatic dimension of rule application and therefore also of productivity (cf. van Marle, 1985; Plag, 1999, pp. 93-241; Bauer, 2001, pp. 71-97). If we consider the domain of application of a WFR, we must think not only of the actual, extrinsic domain of actually existing words formed by this rule, but also of the potential intrinsic domain of a rule, i.e., at the level of language as potential system, not of language as

Productivity in Word Formation

167

norm. This potential domain may be broader and more systematic. Here an overlap of the domains of two morphological rules is possible, particularly in WF because of the greater role of rule competition (rivalry; cf. Bauer, 2001, pp. 136-139, 177-199; Rainer, 2005). The absence or presence of other WFRs of the same or different degrees of productivity changes the probability (cf. Bauer, 2005, pp. 331-332) of applicability of each WFR (cf. Libben et al., 2002; Laaha et al., 2006). This allows for an array of different constellations of WFRs. Main types of such constellations are listed in the sections below.

3.1 Constellation (a): Fully Productive WFRs that do not Compete with Other Rules One polar case is the constellation of one fully productive WFR which does not compete with other rules. This leaves the degree of syntagmatic grammatical productivity, discussed in section 2 above, intact. Examples of fully productive rules are the English WFRs of ordinal formation with the suffix -th and of noun compounding by simple concatenation.

3.2. Constellation (b): Unproductive WFRs Another pole is represented by the constellation of an unproductive WFR, irrespective of whether it competes with other WFRs or not. Such rules have still the function of morphosemantic and morphotactic motivation or of redundancy rules, although not the function of lexical enrichment which productive WFRs have.

3.3. Constellation (c): A Productive WFR that Competes with an Unproductive WFR Very close to constellation (a) is the one where a productive WFR competes with an unproductive one. Here the existing words motivated by the unproductive WFR may block the application of the competing productive WFR to the same base word. However, an application of the productive WFR may supersede the application of the unproductive one and lead to the diachronic change of rule shift. This means that blocking holds just for the established norms of WF of static morphology, but not for the potential system of dynamic morphology, and this holds for the blocking by existing words for the other constellations as well.

3.4. Constellation (d): Competition Between More and Less Productive WFRs Another relevant constellation is represented by competition between a more productive and a less productive WFR, as between English verb formation by the suffix -ize and (the regular

168

The Mental Lexicon: Core Perspectives

subpart of) -ate (Plag, 1999). This should limit the probability of rule application more for the less productive than for the more productive WFR. A case in point is the competition between the fully productive rule of German nominal compound formation via simple concatenation of nouns with the less productive rule of replacing word-final -a of Latinate words with the interfix -en-. Thus the variant Pizza#verkauf “pizza sale” is more probable than the variant Pizz-en#verkauf, both derived from the relatively recent Italian loan pizza. However, Firmen#verkauf “sale of a firm” is much more probable than the potential variant Firma#verkauf, because Firma is a much older loan word and has already a family of compounds (see section 5), starting with the interfixed form Firm-en-, whereas there are no actual compounds starting with Firma-. Also in the case of existing doublets, such as English naïv-ity and naïve-ness (Plag, 2003, p. 67; Szymanek, 2005, pp. 441-442), higher token frequency of one variant does not necessarily go on a par with higher productivity of the corresponding WFR.

3.5. Constellation (e): Competition Between Equally Productive WFRs Another important constellation consists in competition (rivalry) of equally productive WFRs. Here other factors discussed below will decide the probable result of rule rivalry. As a result we may assume the following general scale of probability of rule application: a – c – d – e – (d from the point of view of the less productive WFR) – b. The issue of probability will be taken up again in section 5.

3.6. The Role of Analogy Analogical formation of neologisms belongs clearly to the paradigmatic dimension of WF. I cannot go into the various meanings and uses of this term (see Bauer, 2001, pp. 75-97; Plag, 1999, pp. 17-22) but will limit myself to the following differentiations (a-f ) below: 3.6.1. Type (a). If analogy is said to occur according to a well-defined pattern, then it is simply another term for formation by rule. And according to Schultink’s (1961) conception agreed with in section 2, this should be a productive WFR in case of unintentional coinages. 3.6.2. Type (b). If analogy means that the coinage of a new word is triggered by a specific existing word (called local analogy by Plag, 1999, pp. 203, 206), then this may still be compatible with formation by rule. In this case, the respective WFR represents the deeper cause and the specific model word the triggering occasion (similar to the historical separation between the deeper causes of the First World War and the triggering occasion of crown-prince Franz Ferdinand’s assassination). For example, in the case of chaindrink, morphotactics and word-formation meaning are fully motivated by the WFR of compounding, whereas the

Productivity in Word Formation

169

lexicalised, metaphoric word meaning is due to the analogical model chainsmoke. This means that not every new coinage by a productive WFR must be morphosemantically transparent (cf. Plag, 1999, pp. 38, 107, 111 for the general problem). Very often it is possible to identify a triggering model word or phrase, which may be much less closely associated to the new coinage than the afore-mentioned example, within the same written or oral text or discourse (cf. Matussek, 1994; Peschel, 2002). 3.6.3. Type (c). A WFR cannot be involved in the coinage of a new word if this is formed by surface analogy (Motsch, 1981) to a non-rule-derived word. A German example is the “prefixation” of aber in the long-existing idiomatic phrases hunderte und aber-hunderte, tausende und aber-tausende “hundreds and hundreds, thousands and thousands”, where aber is a cranberry morph (not identical with the homophonous conjunction aber “but”). In surfaceanalogy to this morphotactically transparent but totally unproductive formation occasionally Millionen und aber-millionen “millions and millions”, etc. have been and can be formed. Typically such surface analogies are not unintentional but sophisticated coinages, and their creators or users can clearly identify the model of a specific word, which is rarely the case with unintentional coinages according to a productive WFR. 3.6.4. Type (d). Also new coinages according to unproductive WFRs are predicted to be intentional (in agreement with Schultink’s (1961) conception of productivity, as cited in section 2) and very often due to surface analogy. A French example is parasynthetic denominal formation of verbs of the -ir class, which lost its productivity in the 19th century (cf. KilaniSchoch and Dressler, 2005, p. 132). New verbs formed later have aterrir “to land” (from terre “land”) as their model. First, after the invention of airplanes, technicians invented amerrir “to water” formed from mer “sea”: the orthographic double -rr- of amerrir can only be explained by the model of the rhyme-word aterrir. When expeditions to the moon were planned, the word alunir was formed from lune “moon”, and recently amarsir has been added. 3.6.5. Type (e). On the other hand, surface analogies may start a series of new words which may reanalysed as rule-derived. Well-known examples are the reanalyses of hamburg-er as ham-burger and of alcohol-ic as alc(-)o-holic (with interfix -o-), which has given rise to an increasing series of compounds such as cheese-burger and derived adjectives, such as worko/a-holic (cf. Bauer, 1983, p. 236; Szymanek, 2005, p. 436). 3.6.6. Type (f ). What poses a problem to our approach is the claim that new words can be coined productively according to an imprecise schema formed around a prototype (cf. Tuggy, 2005; Bybee, 1988; Plag, 2004, p. 217). Nearly all the possibilities of novel word formations given can be translated into one of the types (a-f ). For example, Tuggy (2005, pp. 249, 263) refers to vagueness of word-formation meaning and to additional different ambiguous or polyvalent word meanings referring to it: this is clearly an instance of Type (b) analogy. Plag’s

170

The Mental Lexicon: Core Perspectives

(1999, pp. 143-145, 2004) arguments hinge on the assumption that all complex words sharing the same (most peripheral) affix belong to a single schema, irrespective of whether the different subsets of the patterns which participate in the schema, are very or only slightly productive or unproductive. My approach to such instances is to split “schemas” into different WFRs according to the degree of productivity of each subset (see section 2 above). What cannot be accommodated into the rule format, are new words formed by operations of extragrammatical morphology, as in echo-word formation (Tuggy, 2005, pp. 261, 263-264), and this is predicted by the exclusion of such extragrammatical operations from the realm of WFRs (see section 2 above), as Bauer (2001, p. 84) has argued for phonaesthemes. Degrees of grammatical productivity are thus deduced from the superposition of the productivity gradient of the paradigmatic dimension on the scale of productivity of the syntagmatic dimension.

4. ON GRAMMATICAL PRODUCTIVITY AND LANGUAGE TYPOLOGY Studies on morphological typology are much more developed for inflectional morphology than for WF (cf. Haspelmath et al., 2001; Gaeta, 2005). This holds particularly for whole-system typologies which deal with the typological characteristics of whole language components or modules in reference to language types (for my specific approach see Dressler et al., 1987; elaborating on Skalička, 1979). Therefore just a few notes are possible which are meant to indicate that studies on the mental lexicon should take typological differences into account. Several authors, such as Bolozky (1999) and Werner (1983), have described the great productivity of Hebrew WFRs. What is typical for this introflecting language as well as for introflecting Arabic is the great amount of words derived by productive WFRs and the frequently great distance between word-formation meaning predicted by these rules and the actual word meaning, as in the following Aramaic examples of agentives derived from verbal roots (Rubba, 1993, pp. 412-413 ): (7)

šlx “undress” šalaxa “ex-priest” rkw “ride” rakawa “cheat” pθx “open” paθaxa “fortune-teller” ?tw “sit” matwana “landlord” ?xl “eat” maxlana “host”

This means that despite high productivity, morphosemantic transparency is rather low. Also morphotactic transparency is rather low, since many productive WFRs present multiple ablaut (or transfixes in an alternative analysis). This is a further argument against the derivation of productivity from transparency (cf. Bauer, 2005, p. 329).

Productivity in Word Formation

171

Also the agglutinating type has more productive WF than the inflecting-fusional type (as represented in English, German, Romance and Slavic derivational morphology). This comes, first of all, from the fact that many grammatical categories, such as causatives, possibilitives, reflexives, negation, Hungarian noun incorporation into verbs, which typically belong to syntax in the inflecting-fusional type, are expressed by WFRs in agglutinating languages (cf. Pöchtrager et al., 1998; Dressler and Ladányi, 2000a, pp. 136-137, 2000b; Ülkü, 1980; Gatauvoa, 2006; for Dravidian languages see Steever, 1998; Krishnamurti, 2003). Thus agglutinating languages have more WF productivity in the syntagmatic dimension, similar to inflectional morphology. Comparable WFRs have also been found to be more productive in agglutinating than in inflecting-fusional languages. For example, Dressler and Ladányi (2000b, pp. 62-64), in comparing Hungarian and German denominal adjective formation, have shown that the respective Hungarian WFRs may apply to abbreviations, which is not true for their German counterparts. Globally, agglutinating languages have a greater number of productive and a smaller number of unproductive WFRs than inflecting-fusional languages, similar to inflectional morphology. There is less rule competition and therefore less paradigmatic restrictions on the productivity of a single WFR. The reason for this is that, more than any other language type, the agglutinating type follows the morphological biuniqueness preference, i.e., the preference for expressing one grammatical meaning by one and the same morphological form, and vice versa. For consequences in language acquisition and aphasia see sections 7 and 8. There is less rule competition and thus more productivity in the paradigmatic dimension, because the universal preference for morphological biuniqueness is followed more than in the inflecting-fusional type. There is more morphotactic transparency than in the inflecting-fusional type. Compounding is less productive in the agglutinating than in the inflecting-fusional type, conversions are marginal, whereas strongly-inflecting languages, such as the Slavic languages, Latin and Greek have much stem- and root-based conversion, weakly inflecting languages, such as English and other Germanic languages, much word-based conversion (cf. Manova and Dressler, 2005). Within morphonology, vowel harmony is typical for many agglutinating languages, whereas umlaut (metaphony) is typical for many inflecting-fusional languages, and vowel harmony and umlaut do not exist or are very marginal in the opposite language type. Superficially, both processes are similar, insofar as they are, at least in their diachronic origins, vocalic distance assimilations and opacify morphotactically both inflection and derivation, e.g., in Hungarian denominal adjective formation with the suffix -es/-os/-as/-ös: (8)

folt “patch” foltos “patched” kerék “wheel” kerek-es “supplied with wheels” hó “snow” hav-as “snowy”

172

The Mental Lexicon: Core Perspectives

For German cf. palatalizing umlaut triggered by the feminine-motion suffix -in “-ess”: (9)

Hund “dog” Wolf “wolf”

Hünd-in Wölf-in

But there is one important difference: vowel harmony is usually fully productive (in Hungarian with the exception of the unproductive /a/ variant, if it is competing with an /o/ variant), whereas umlaut is either unproductive or slightly productive. For example the feminine of the German word Herzog “duke” is Herzog-in without umlaut, and native speakers of German are uncertain whether the feminine of Maulwurf “mole” should be Maulwurf-in or Maulwürf-in. We will see the effects of this difference in aphasia (section 8).

5. ON PRODUCTIVITY ON THE LEVEL OF NORMS This level refers to the type frequency of neologisms, i.e., of new words accepted within the norms of a speech community or a part thereof, as in the case of terminology in languages for special purposes. Here I will limit myself to brief indications, because my views do not differ much from ample and brilliant investigations available, such as Plag (1999) and Bolozky (1999). The source of high productivity on this level, is beyond high grammatical productivity on the level of the potential system (sections 2 and 3), the domain or scope of application, is the inverse of Booij’s (1977, p. 5) competence restrictions and is called profitability by Corbin (1987, p. 177). In English denominal verb formation, for example, this is highest, for conversion, and next highest for -ize suffixation (Plag, 1999). Many structural restrictions on productivity discussed, most recently in Rainer (2005) and Bauer (2005, p. 331; cf. Plag, 2003, pp. 61-68) are simply restrictions on rule application itself and not directly connected with grammatical productivity as discussed in section 2 (but several are of a paradigmatic nature, as in section 3). A next set of factors are pragmatic ones (cf. Dressler and Ladányi, 2000a, pp. 108-109) which link societal with linguistic norms. This starts with the usefulness of a WFR in a speech community for labelling new concepts for which a need is felt (Plag, 1999, p. 39, 2003, pp. 5960; Bauer, 2001, pp. 41-43, 159, 170-171, 208), nameability, i.e., pragmatic (in)compatibilities of a referential nature (Bauer, 2001, p. 32; Plag, 1999, p. 40, 2003, p. 60), changing fashions as in the use of the Latinate (including Greek) prefixes super-, mega-, etc. (Plag, 1999, p. 39, 2003, p. 60) and which may result in many trend words sharing the same WF pattern (cf. Loskant, 1998), emotional attitudes, as in the amount of diminutives used (Wierzbicka, 1991; Dressler and Merlini Barbaresi, 1994, pp. 409-414, 490; Plag, 1999, pp. 40-41). Here one may add the discourse function of WF of recategorizing and summarizing in a condensed way

Productivity in Word Formation

173

(Dressler, 1985; Plag, 1999, p. 39; Peschel, 2002, pp. 62, 75) and stylistic restrictions (Bauer, 2001, p. 135; Rainer, 1993, p. 125; Cowie, 2003). Greater pragmatic needs may have structural consequences. For example, the greater pragmatic needs for using diminutives than augmentatives (cf. Dressler and Merlini Barbaresi, 1994, pp. 411, 490) has a first effect of greater productivity of diminutives on the level of norms, i.e., of having more diminutive than augmentative neologisms (cf. Sabatini and Coletti, 1997), a second effect of having a greater number of productive diminutive than of augmentative rules (as in the Romance and Slavic languages), third in the cross-linguistic implication that if the morphology of a language possesses augmentative formation, it also possesses diminutive formation, but not vice versa (cf. Dressler and Merlini Barbaresi, 1994, pp. 430-431). The above-mentioned pragmatic factors influence the probability of new coinages so much that counting the absolute number of neologisms attested in a given time period delivers a very superficial account of productivity. Thus we must briefly indicate other measures of WF productivity proposed and tested. Great attention has been aroused by measures based on the relative frequency of hapax legomena, i.e., of new words occurring only once, in electronic corpora (Baayen and Lieber, 1991; Baayen, 1992; Baayen and Renouf, 1996; Baayen, 2001): the term comes from ancient Greek Homeric philology which attempted to explain words which occur in Homer only once. These are usually poetic occasionalisms, opaque formations of various types, but much less frequently new coinages produced by productive WFRs. Although the occurrence of the pertinent last type is much higher in the electronic corpora investigated by Baayen and his research associates, the set of hapax legomena is still a mixed bag. For other critiques and modifications see Plag (1999, pp. 26-33, 100 footnote 5, 111-117, 2003, pp. 54-59), Bauer (2001, pp. 148-155), Dal (2003b, pp. 11-20), Gaeta and Ricca (2006), who discuss other statistical measures as well. Such measures have often been used for determining the probability of application of a productive WFR (cf. Bauer, 2005, pp. 331-332; Rainer, 2005, p. 340 for overviews). Different linguistic and psycholinguistic models have attempted to derive productivity from morphotactic (or phonological) and/or morphosemantic (or semantic) transparency: the more transparent a WFR is, the more its chance of being used productively. However, since unproductive WFRs may also be, in most of their derivations, morphosemantically or morphotactically transparent, the connection can be only feeble and derives also from productivity towards transparency, i.e., in the reverse direction: the more productive a WFR is, the more transparent is the number of its derivatives, due to neologisms which are generally more transparent than older derivatives (cf. Dressler and Merlini Barbaresi, 1994, p. 96; Bauer, 2001, pp. 54, 59-60, 2005, pp. 310, 321, 329). Especially for compounds, the respective size of the morphological families sharing the same first or second compound member in the stored lexicon have been shown to

174

The Mental Lexicon: Core Perspectives

paradigmatically influence productivity (Bertram et al., 2000; de Jong, 2002; Krott et al., 2007). Another frequency-derived factor of lexical storage with a positive impact on degree of productivity is Hay’s (2001; cf. Hay and Baayen, 2003, pp. 102-104) tenet that derivations which are less token-frequent than their bases, favour parsing (decomposition) and thus rule application. Another factor, highlighted by Hay (2001; cf. Hay and Baayen, 2003, pp. 105-115) which favours decomposition are phonotactic consonant clusters which imply an intermediate morpheme boundary. Such situations, however, occur also with unproductive WFRs, as in the English words leng-th, streng-th, bread-th, which have little chance of being decomposed, due to opacifying umlaut, although, again, the base words have higher token frequencies than their derivatives. Such processing factors, inherent in the institutionalised lexical norms of a language, influence the probability of rule application on the level of performance and thus modify the effects of degrees of grammatical productivity. The claim of this chapter is then that the scale of grammatical productivity of section 2 modified by the effects of rule rivalry of the paradigmatic dimension (section 3) and by factors of language norms (section 5) determines the probability of computation by rule vs. lexical access to fully stored morphologically complex words.

6. GRAMMATICAL PRODUCTIVITY IN PSYCHOLINGUISTIC EXPERIMENTS In contrast to inductive productivity tests devised for searching which patterns are productive (cf. e.g., Bolozky, 1999), deductively devised tests have been done on German compounding in order to test concepts of grammatical productivity, as explicated in sections 2 and 3 above. The syntagmatic dimension of grammatical productivity (section 2) has been tested by Dressler et al. (2001) in an off-line composition test and in an on-line decomposition test. In both, the most productive WFRs of purely concatenative compounding, of -n- interfixation after first members ending in schwa and of -s- interfixation after masculine first members have come out best. Grammatical productivity as modified in the paradigmatic dimension (section 3) by contrasting productive WFRs with and without rule competition has been probed in an on-line composition test by Libben et al. (2002) and has come out again as significantly facilitating compounding, although this may be difficult to distinguish from the effects of family size (cf. Krott et al., 2007; Bertram, Baayen and Schreuder, 2000). Thus more specific on-line tests should be devised in order to probe predictions derivable from the productivity hierarchy of this contribution.

Productivity in Word Formation

175

7. GRAMMATICAL PRODUCTIVITY IN EARLY FIRST LANGUAGE ACQUISITION In many languages, the earliest WF pattern acquired is diminutive formation. Therefore, within the “cross-linguistic project on pre- and protomorphology in language acquisition”, the emergence and development of evaluative formation has been investigated in longitudinal corpus studies (Gillis, 1998; Savickiene and Dressler, 2007; Savickiene et al., 2006). First of all, whichever language has both diminutive and augmentative evaluatives (Croatian, Italian, Russian, Spanish), invariably diminutives have emerged before augmentatives and have been acquired first. This may be either only due to greater productivity and frequency of diminutives or also due the general implication that the presence of augmentatives implies that of diminutives (see section 5). In general, many diminutive WFRs are productive and belong to productive inflectional classes. Thus, often the switch from the simplex base to the corresponding diminutive form implies a switch from an unproductive to a productive inflection class, e.g., Italian feminine parte, plural parti “part” diminutive part-ic-ella, plural part-ic-elle. The Croatian, Hungarian, Italian and Lithuanian children investigated in Savickiene and Dressler (2007) have preferred the longer diminutive forms to the shorter simplex base forms in a statistically significant way, when this involved the switch from an unproductive to a productive class, which proves the relevance of morphological productivity for early language acquisition (cf. Nicoladis, 2005, pp. 100-103, 119-122 for compounds). Language typology (in the sense of section 4) plays only a minor rule, insofar as agglutinating Hungarian has only two fully productive diminutive suffixes (-ka and -czka), whereas the inflecting-fusional Slavic, Romance languages (with the exception of the less inflecting and more isolating language French), Greek and Lithuanian have more. This is an instance of less rule competition within the agglutinating language type (cf. section 4). Moreover, the two Hungarian suffixation rules have largely complementary domains of rule application, with the effect that children can follow the preference for morphological biuniqueness and make very few errors in acquisition (Bodor and Barcza, 2007). Evidently the question arises how young children may become aware of the degree of productivity of WFRs in child-directed speech, where our main criteria for high grammatical productivity, i.e., application to new loan words and abbreviations, rarely occur. Several factors play a role (cf. Dressler, Drążyk et al., 1996, pp. 188-189): High type frequency coupled with low token frequency (cf. Bybee, 2001, pp. 118-124) is a necessary but insufficient indication, further factors are morphosemantic transparency of novel forms, occurrence in playful variation and the phenomenon that novel words coined at the spot are produced in a more salient way by caregivers. Thus Bird et al.’s (1996) experiments showed that mothers produced in child-directed speech words they deemed novel for their infants (at the age brackets of 1;4 and 2;4) with longer duration and higher intensity. Infants are clearly sensitive to such prosodic cues (cf. Dominey and Dodane, 2004, pp. 126-132).

176

The Mental Lexicon: Core Perspectives

Since MacWhinney’s (1978) arguments it is generally assumed in acquisition models that distinguish analogy and rule application, that analogies precede the formation of rules in the first-language acquisition process. This means that, in the development of the mental lexicon, the detection of morphology by young children (cf. Dressler et al., 2003) leads first to the construction of schemas and only afterwards to the extraction of rules from such schemas. In this rule extraction the relevance of productivity is confirmed by the typical acquisition path of inflectional morphology: first, mostly, productive rules are overgeneralized (e.g., the formation of weak past tense formation in English and German), only later and temporarily, unproductive rules are overgeneralized to a significant extent (e.g., past participle English brung). This means presumably that children, at first, consider any new rule they extract as productive. This has not yet been investigated for the acquisition of WFRs.

8. GRAMMATICAL PRODUCTIVITY IN APHASIA In contrast to often studied effects of relative transparency on the number of errors in aphasia, specific effects of degrees of WF productivity have never been investigated systematically in aphasiology. We started to do this in conjunction with the effects of typological differences between an agglutinating (Hungarian) and an inflecting-fusional language (German), first with two Broca’s aphasics each in a pilot study (Dressler, Stark et al., 1996), then with a larger group of 5 Austrian aphasics (2 Broca, 2 Wernicke, 1 conduction aphasic) and 5 Hungarian aphasics (Dressler and Kiefer, in preparation). In addition to plural formation, two areas of WF were tested: denominal adjective formation and diminutive formation. Testing consisted of a naming and a lexical decision test. Some preliminary results are:

Table 1 Patient H.R.’s naming test performance from Dressler and Kiefer (in preparation). Base Word

English Gloss

Ärger Ehre Gefahr Quadrat Chaos Verrat Kitsch Gewalt Spott Kritik

anger honour danger square chaos treason kitsch power mockery criticism

Correct Adjective Derivation ärger-lich ehr-lich gefähr-lich quadrat-isch chaot-isch verrät-er-isch kitsch-ig gewalt-ig spött-isch krit-isch

Erroneous Substitution ärger-isch ehr-ig gefähr-ig quadrat-lich chaos-ig verrät-ig kitsch-lich gewalt-lich spött-lich kritik-lich

Productivity in Word Formation

177

Selection between the productive German suffixes -ig, -isch, -lich has been difficult, note, for example, the frequent errors by (female) Broca patient H.R. in the naming test shown in Table 1. Nevertheless H.R. produced 69% of -ig derivatives (highest productivity) correctly, 41% of the productive -isch and -lich derivatives, but only 20% of the unproductive derivatives. When there was a competition between two productive derivatives, she always chose one of the two correct ones. Similar error rates were found with the other Austrian aphasics. The respective percentages were for the Hungarian aphasics 89% correct productive derivation vs. 75% correct unproductive derivations with the suffix -s and its allomorphs, and 69% correct productive derivations vs. 52% unproductive ones for the rival suffix -i. Thus Hungarian aphasics committed fewer errors than the Austrian ones. This is in line with previous findings for Hungarian by Kertesz and Osmán-Sagi (2001) and for agglutinating Finnish by Helasvuo et al. (2001). But there are clear cases of the replacement of unproductive derivations by productive patterns (and rarely the reverse), as in the examples shown in Table 2 (note that an accent on a vowel, signals vowel length).

Table 2 The replacement of unproductive derivations by productive patterns in Hungarian aphasic errors (from Dressler and Kiefer, in preparation).

Base Word

English Gloss

hó tó bokor cukor

snow lake bush sugar

Correct Adjective Derivation hav-as tav-as bokr-os cukr-os

Erroneous Substitution hó-s tó-s bokor-os cukor-os

The lower error rates in Hungarian aphasia may be due to less rule competition in Hungarian than in German. As predicted (section 4), Hungarian vowel harmony and German umlaut fared differently in aphasia: Austrian aphasics committed many umlaut errors in both directions (in lexical decision tests, they preferred the umlaut variant, presumably as the more salient variant), Hungarian aphasics committed no errors with productive vowel harmony (cf. Kertesz and Osmán-Sagi, 2001; cf. Peuser and Fittschen, 1977 for agglutinating Turkish), but often replaced unproductive back harmony of the /a/ variant with productive back harmony of the /o/ variant, as shown in Table 3.

178

The Mental Lexicon: Core Perspectives

Table 3 The replacement of unproductive back harmony of the /a/ variant with productive back harmony of the /o/ variant in Hungarian aphasic errors (from Dressler and Kiefer, in preparation). Base Word

English Gloss

vona kanál

line spoon

Correct Adjective Derivation vonal-as kanal-as

Erroneous Substitution vonal-os kanal-os

9. CONCLUSION This contribution has aimed at demonstrating the complexity of the phenomenon of productivity in word formation by starting with the concept of grammatical productivity and at showing consequences for investigations of the mental lexicon in processing experiments, in early language acquisition and in aphasic impairments. I have claimed that grammatical productivity is basic for the notion of morphological rule, that it is graded and needs a definition both on the syntagmatic axis (section 2) in the rule-governed relations between complex words and their bases and on the paradigmatic axis in various constellations of rule competition/rivalry (section 3). But the role of productivity differs in different language types (section 4), also with respect to morphosemantic and morphotactic transparency. On the level of linguistic norms (section 5), mediated by various pragmatic factors, grammatical productivity affects type frequency. Degree of productivity is then claimed to determine the probability of computation by rule vs. lexical access to fully stored morphologically complex words. Evidence for these claims has come from the typological distribution of productivity, mainly in languages approaching the inflecting and the agglutinating type (section 4); from processing tests with German compounds (section 6); early phases of first language acquisition, particularly of diminutives (section 7); and from aphasic impairments of German and Hungarian word formation, which approaches the inflecting and the agglutinating type, respectively (section 8). As a consequence of this discussion, it is suggested that the proposed factors and variables of productivity should be considered in designing and interpreting experiments, especially in differentiating degrees of productivity, of language awareness and of proximity of the word formation of a given language to the ideal morphological language types. In addition to the general lack of such investigations, this contribution has not been able to report on acquisition studies of older children, on aphasic impairments in a wider typological variety of languages and, in general, on incorporating languages.

Productivity in Word Formation

179

REFERENCES Aronoff, M. (1976). Word Formation in Generative Grammar. MIT Press, Cambridge, MA. Aronoff, M. and L. Gaeta (2003). Morphological productivity. Italian Journal of Linguistics, 15, 1196. Baayen, R. H. (1989). A Corpus-Based Approach to Morphological Productivity. PhD thesis, Vrije Universiteit te Amsterdam. Baayen, R. H. (1992). Quantitative aspects of morphological productivity. Yearbook of Morphology 1991, 109-149. Baayen, R. H. (2001). Word-Frequency Distributions. Kluwer, Dordrecht. Baayen, R. H. and R. Lieber (1991). Productivity and English word-formation: a corpus-based study. Linguistics, 29, 801-843. Baayen, R. H. and A. Renouf (1996). Chronicling the Times: productive lexical innovations in an English newspaper. Language, 72, 69-96. Baayen, R. H. and R. Schreuder (1999). War and peace: morphemes and full forms in a non-interactive activation parallel route model. Brain and Language, 68, 27-32. Bauer, L. (1983). English Word-formation. Cambridge University Press, Cambridge. Bauer, L. (2001). Morphological Productivity. Cambridge University Press, Cambridge. Bauer, L. (2005). Productivity: theories. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 315-334. Springer, Dordrecht. Berent, I., S. Pinker and J. Shimron (2002). The nature of regularity and irregularity: evidence from Hebrew nominal inflection. Journal of Psycholinguistic Research, 31, 459-502. Berrendonner, A. and V. Clavier. (1997). Examen d'une série morphologique dite "improductive" en français: les noms dénominaux en -age. In : (D. Corbin, ed.) Mots possibles et mots existants. Silexicales 1 pp. 35-44. Université de Lille, Lille, France Bertram, R, R. H. Baayen and R. Schreuder (2000). Effects of family size for complex words. Journal of Memory and Language, 42, 390-405. Bertram, R., M. Laine and K. Karvinen (1999). The interplay of word formation type, affixal homonymy, and productivity in lexical processing: evidence from a morphologically rich language. Journal of Psycholinguistic Research, 28, 213-226. Bertram, R., R. Schreuder and R. H. 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: Learning, Memory and Cognition, 26, 489-511. Bird, E., K.-R. Robin, S. Chapman and P. Cleave (1996). Acoustic analysis of novel word productions. Poster presented at the 7th International Congress for the Study of Child Language, Istanbul. Bodor, P. and V. Barcza. (2007). Acquisition of diminutives in Hungarian. In: The Acquisition of Diminutives: Cross-linguistic Perspective. (I. Savickiene and W. U. Dressler eds.), pp. 231-262. Benjamins, Amsterdam. Bolozky, S. (1999). Measuring Productivity in Word Formation. Brill, Leiden. Booij, G. (1977). Dutch Morphology. Foris, Dordrecht.

180

The Mental Lexicon: Core Perspectives

Bybee, J. L. (1988). Morphology as lexical organization. In: Theoretical Morphology (M. Hammond and M. Noonan, eds.), pp. 119-141. Academic Press, London. Bybee, J. L. (2001). Phonology and Language Use. Cambridge University Press, Cambridge. Chomsky, N. (1986). Knowledge of Language. Praeger, New York. Clahsen, H. (1999).Lexical entries and rules of language: a multidisciplinary study of German inflection. Behavioral and Brain Sciences, 22, 991-1060. Clahsen, H., I. Sonnenstuhl and J. P. Blevins (2003). Derivational morphology in the German mental lexicon: a dual mechanism account. In: Morphological Structure in Language Processing (R.H. Baayen and R. Schreuder, eds.), pp. 125-155. Mouton de Gruyter, Berlin. Corbin, D. (1987) Morphologie dérivationelle et structuration du lexique. Niemeyer, Tübingen. Coseriu, E. (1975). System, Norm und Rede. In: Sprachtheorie und allgemeine Sprachwissenschaft (E. Coseriu, ed.), pp.11-101. Fink, München. Cowie, C. (2003). «Uncommon terminations»: proscription and morphological productivity. Italian Journal of Linguistics, 15, 17-30. Dal, G. (Ed.). (2003a). La productivité morphologique en questions et en expérimentations. Langue Française, 140, 3-23. Dal, G. (2003b). Productivité morphologique: définitions et notions connexes. Langue Française, 140, 3-23. De Jong, N. H. (2002). Morphological Families in the Mental Lexicon. PhD thesis, Catholic University of Nijmegen. Dominey, P. F. and C. Dodane (2004). Indeterminacy in language acquisition: the role of child directed speech and joint attention. Journal of Neurolinguistics, 17, 121-145. Dressler, W. U. (1985). Morphology. In: Handbook of Discourse Analysis II (T. van der Dijk, ed.), pp. 77-86. Academic Press, London. Dressler, W. U. (2000). Extragrammatical vs. marginal morphology. In: Extragrammatical and Marginal Morphology (U. Doleschal and A.M. Thornton, eds.), pp. 1-10. Lincom, München. Dressler, W. U. (2003). Degrees of grammatical productivity in inflectional morphology. Italian Journal of Linguistics, 15, 31-62. Dressler, W. U. (2004). Hypercharacterisation and productivity in inflectional morphology. In: Homini Universali Dicata Analecta (T. Kriesch et al., eds.), pp. 515-524. Heinz, Stuttgart. Dressler, W. U. (2005). Word-formation in Natural Morphology. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 267-284. Springer, Dordrecht. Dressler, W. U., R. Drążyk, D. Drążyk, K. Dziubalska-Kołaczyk and E. Jagła (1996). On the earliest stages of acquisition of Polish declension. In: Proceedings of the Groningen Assembly on Language Acquisition (C. Koster and F. Wijnen, eds.), pp. 185-196. Centre for Language and Cognition, Groningen. Dressler, W. U. and F. Kiefer (Eds.). (in preparation). Morphological and Syntactic Impairments in German and Hungarian aphasia. Verlag der Österreichischen Akademie der Wissenschaften, Wien. Dressler, W. U., M. Kilani-Schoch and S. Klampfer (2003). How does a child detect bound morphology? In: Morphological Structure in Language Processing (H. Baayen and R. Schreuder, eds.), pp.391-425. Mouton de Gruyter, Berlin. Dressler, W. U. and M. Ladányi (2000a). Productivity in word formation (WF): a morphological approach. Acta Linguistica Hungarica, 47, 103-144.

Productivity in Word Formation

181

Dressler, W. U. and M. Ladányi (2000b). On contrastive word formation: German and Hungarian denominal adjective formation. In: Words: Structure, Meaning, Function: A Festschriftg for D. Kastovsky (C. Dalton-Puffer and N. Ritt, eds.), pp. 59-74. Mouton de Gruyter, Berlin. Dressler, W. U., G. Libben, J. Stark, C. Pons and G. Jarema (2001). The processing of German interfixed compounds. Yearbook of Morphology 1999. 185-220. Dressler, W. U., G. Libben, G. Jarema, J. Stark and C. Pons (2005). Produttività nel processamento di composti: esempi tedeschi con e senza interfissi. In: La Formazione delle Parole, Proceedings of the 37th SLI congress (M. Grossmann and A.M. Thornton, eds.), pp. 153-162. Bulzoni, Roma. Dressler, W. U., W. Mayerthaler, O. Panagl and W. U. Wurzel (1987). Leitmotifs in Natural Morphology. Benjamins, Amsterdam. Dressler, W.U. and L. Merlini Barbaresi (1994). Morphopragmatics. Mouton de Gruyter, Berlin. Dressler, W. U., J. Stark, C. Pons, F. Kiefer, K. Kiss and É. Mészáros (1996). Cross-Language Analysis of German- and Hungarian-Speaking Broca’s Aphasics’ Processing of Selected Morphonological and Morphological Features. Hungarian Academy of Sciences, Working Papers in the Theory of Grammar 3, 1, Budapest. Fradin, B., N. Hathout and F. Meunier (2003). La suffixation en –ET et la question de la productivité. Langue Française, 140, 56-78. Frauenfelder, U. H. and R. Schreuder (1992). Constraining psycholinguistic models of morphological processing and representation: the role of productivity. Yearbook of Morphology 1991, 165-183. Gaeta, L. (2005). Word formation and typology : which language universals? In: Morphology and Linguistic Typology (G. Booij et al., eds.). On-line Proceedings of the 4th Mediterranean Morphology Meeting (Catania). http://morbo-lingue.unibo.it/mmm/mmm4-proceedings.php Gaeta, L. and D. Ricca (2006). Productivity in Italian word formation: a variable-corpus approach. Linguistics, 44, 57-89. Gatauova, E. (2006). Verbale Wortbildung in der kasachischen und russischen Sprache. PhD thesis, Universität Wien. Gillis, S. (Ed.). (1998). Studies in the Acquisition of Number and Diminutive Marking. Antwerp Papers in Linguistics 95. University of Antwerp, Antwerp. Haspelmath, M., E. König, W. Österreicher and W. Raible (Eds.). (2001). Language Typology and Language Universals. de Gruyter, Berlin. Hay, J. (2001). Lexical frequency in morphology: is everything relative? Linguistics, 39, 1041-1070. Hay, J. (2002). From speech perception to morphology: affix ordering revisited. Language, 78, 527555. Hay, J. and H. Baayen (2003). Phonotactics, parsing and productivity. Italian Journal of Linguistics, 15, 99-130. Helasvuo, M.-L., A. Klippi and M. Laakso (2001). Grammatical structuring in Broca’s and Wernicke’s aphasia in Finnish. Journal of Neurolinguistics, 14, 231-254. Hohenaus, P. (2005). Lexicalization and institutionalization. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 353-373. Springer, Dordrecht. Järvikivi, J. (2003). Allomorphy and Morphological Salience in the Mental Lexicon. University of Joensuu, Joensuu. Kastovsky, D. (2005). Hans Marchand and the Marchandeans. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 99-124. Springer, Dordrecht.

182

The Mental Lexicon: Core Perspectives

Kertesz, A. and J. Osmán-Sagi (2001). Manifestations of aphasia symptoms in Hungarian. Journal of Neurolinguistics, 14, 313-319. Kilani-Schoch, M. and W. U. Dressler (2005). Morphologie naturelle et flexion du verbe français. Narr, Tübingen. Krishnamurti, B. (2003). The Dravidian Languages. Cambridge University Press, Cambridge. Krott, A., R. Schreuder, R. H. Baayen and W. U. Dressler (2007). Analogical effects on linking elements in German compounds. Language and Cognitive Processes, 22, 25-57. Laaha, S., D. Ravid, K. Korecky-Kröll, G. Laaha and W. U. Dressler (2006). Early noun plurals in German: regularity, productivity or default? Journal of Child Language, 33, 271-302. Libben, G., G. Jarema, W. U. Dressler, J. Stark and C. Pons (2002). Triangulating the effects of interfixation in the processing of German compounds. Folia Linguistica, 36, 23-43. Loskant, S. (1998). Das neue Trendwörter-Lexikon. Bertelsmann, Ort. MacWhinney, B. (1978). The Acquisition of Morphophonology. University of Chicago Press, Chicago. Manova, S. and W. U. Dressler (2005). The morphological technique of conversion in the inflectingfusional type. In: Approaches to Conversion/Zero-Derivation (L. Bauer and S. Valera, eds.), pp. 67101. Waxmann, Münster. Marchand, H. (1969). The categories and types of present-day English word-formation. Beck, München. Matussek, M. (1994). Wortneubildung im Text. Buske, Hamburg. Motsch, W. (1981). Der kreative Aspekt in der Wortbildung. In: Wortbildung (L. Lipka, ed.) pp. 94118. Wissenschaftliche Buchgesellschaft, Darmstadt. Nicoladis, E. (2006). Preschool children’s acquisition of compounds. In: The Representation and Processing of Compound Words (G. Libben and G. Jarema, eds.) pp. 96-124. Oxford University Press, Oxford. Osman, N. (1994). Lexikon untergegangener Wörter. Beck, München. Peschel, C. (2002). Zum Zusammenhang von Wortneubildung und Textkonstitution. Niemeyer, Tübingen. Peuser, G. and M. Fittschen (1977). On the universality of language dissolution: The case of a Turkish aphasic. Brain and Language, 4, 196-207. Pinker, S. (1999). Words and Rules. The Ingredients of Language. Weidenfeld and Nicolson, London. Pinker, S. and M. T. Ullman (2002). The past and future of the past tense. Trends in Cognitive Sciences, 6, 456-463. Plag, I. (1999). Morphological Productivity: Structural Constraints on English Derivation. Mouton de Gruyter, Berlin. Plag, I. (2003). Word-formation in English. Cambridge University Press, Cambridge. Plag, I. (2004). Syntactic category information and the semantics of derivational morphological rules. Folia Linguistica, 38, 193-225. Pöchtrager, M., C. Bodó, W.U. Dressler and T. Schweiger (1998). On some inflectional properties of the agglutinating type, illustrated from Finnish, Hungarianand Turkish inflection. Wiener Linguistische Gazette, 62-63, 57-92. Pounder, A. (2000). Processes and Paradigms in Word Formation Morphology. Mouton de Gruyter, Berlin. Rainer, F. (1993). Spanische Wortbildungslehre. Niemeyer, Tübingen.

Productivity in Word Formation

183

Rainer, F. (1997). Vers une contrainte sur les stades dérivationnelles virtuelles. In : Mots possibles et mots existants. Silexicales, no 1 (D. Corbin et al., eds.), pp. 231-240. Lille. Rainer, F. (2005). Constraints on productivity. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 335-352. Springer, Dordrecht. Rubba, J. E. (1993). Discontinuous Morphology in Modern Aramaic. PhD thesis, University of California, San Diego. Sabatini, F. and V. Coletti (1997). DISC-Dizionario Italiano Sabatini Coletti. Giusti, Firenze. Sánchez Miret, F. M. (2006). Productivity of the weak verbs in Romanian. Folia Linguistica, 40, 29-50. Savickiene, I. and W. U. Dressler (Eds.). (2007). The Acquisition of Diminutives: Cross-linguistic Perspective. Benjamins, Amsterdam. Savickiene, I., W. U. Dressler, V. Barcza, P. Bodor, N. Ketrez, K. Korecky-Kröll, M. Palmović, U. Stephany and E. Thomadaki (2006). Diminutives as pioneers of derivational and inflectional development – a cross-linguistic perspective. In: Emergence of nominal and verbal morphology from a typological perspective, (S. Laaha and S. Gillis, eds.), Antwerp Papers in Linguistics, Antwerp. Schultink, H. (1961). Produktiviteit als morfologisch fenomeen. Forum der Letteren, 2, 110-125. Skalička, V. (1979). Typologische Studien. Vieweg, Braunschweig. Steever, S. B. (Ed.). (1998). The Dravidian Languages. Routledge, London. Štekauer, P. and R. Lieber (2005). Handbook of Word-Formation. Springer, Dordrecht. Szymanek, B. (2005). The latest trends in English Word-Formation. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 429-448. Springer, Dordrecht. Tuggy, D. (2005). Cognitive approach to word-formation. In: Handbook of Word-formation (P. Štekauer and R. Lieber, eds.), pp. 233-264. Springer, Dordrech. Ülkü, V. (1980). Affixale Wortbildung im Deutschen und Türkischen. Ankara Üniversitesi BasImevi. Vannest, J., R. Bertram, J. Järvikivi and J. Niemi (2002). Counterintuitive cross-linguistic differences: more morphological computation in English than in Finnish. Journal of Psycholinguistic Research, 31, 83-106. van Marle, J. (1985). On the Paradigmatic Dimension of Morphological Creativity. Foris, Dordrecht. Werner, F. (1983). Die Wortbildung der hebräischen Adjektiva. Wiesbaden. Wierzbicka, A. (1991). Cross-Cultural Pragmatics. de Gruyter, Berlin. Zemskaja, E. (Ed.). (1996). Russkij jazyk konca XX stoletija (1985 - 1995). Jazyki Russkoj Kul'tury, Moskva.

This page intentionally left blank

9 BILINGUAL LEXICA

Loraine K. Obler, The Graduate School and University Center of The City University of New York and Boston University School of Medicine and the Boston VA Healthcare System, USA Mira Goral, The Graduate School and University Center of The City University of New York and Lehman College of the City University of New York, USA

1. THE BILINGUAL LEXICON Studying the bilingual lexicon is interesting in and of itself, however it can also help us answer questions about lexical organization and processing in humans in general. Indeed, what we learn from lexical studies in bilinguals can be extrapolated to both monolinguals and to multilinguals. For example, the bilingual’s mastery of two systems that must, in some situations, be kept distinct may be seen to require an exaggerated form of the inhibition required for monolinguals to select and express the precise lexical item they target rather than selecting semantically close substitutions. In the middle of the 20th century, research on the bilingual lexicon focused on whether or not the bilingual’s two languages were shared or separate systems. In the final decades of that century, that discussion evolved to consider as well lexical connections between words in a given language and their translation-equivalents in another of bilinguals’ languages. In the past decade, a discussion on bilingual processing has complemented the one concerning representation. That is, researchers have asked whether bilinguals process their languages selectively for certain tasks or whether they always have both languages available (non-selective processing).

1.1. Shared or Separate Systems In the 1950s, Weinreich (1953) proposed three configurations of two languages (by which he seems to have intended not exclusively their lexical systems) that he considered typified different types of bilinguals, reflecting their different experiences in learning their two languages. Coordinate bilinguals learned the two languages separately, and kept their meanings

185

186

The Mental Lexicon: Core Perspectives

separate; compound bilinguals learned the two languages in a single environment, and thus had a single meaning unit for a given word and its translation equivalent; subordinate bilinguals were in the earlier stages of acquiring a second language (L2) and processed L2 lexical items via their first language (L1) equivalents (see Figure 1).

(a) Coordinate book ‫׀‬

kniga ‫׀‬

/buk/

/kn’iga/

(b) Compound book = kniga / \ /buk/ /kn’iga/

(c) Subordinate book /buk/ ‫׀‬ /kn’iga/ Figure 1. Types of bilingualism (Weinreich, 1953)

These categories that Weinreich proposed – particularly the first two – were borrowed by a number of researchers for their research (e.g., Ervin and Osgood, 1954). Lambert pursued these distinctions in a series of studies (Lambert, 1969; Lambert and Fillenbaum, 1959), studying recovery patterns in bilingual individuals with aphasia. He found that those who could be characterized as coordinate in their learning styles were more likely to support neither primacy nor recency explanations for differential aphasia while those who would be characterized as compound were relatively more likely to support either or both explanations.

Bilingual Lexica

187

Others ignored the manner-of-acquisition factors implicit in Weinreich’s categories, setting up studies to ask the falsely dichotomous questions assuming bilinguals all either evidenced compound or coordinate processing, finding an answer linked to the tasks they employed. List-recall tasks, for example, demonstrated more coordinate representation, whereas free-association tasks suggested more compound representation (e.g., Kolers and Gonzalez, 1980). With the introduction of hierarchical models in the 1980s, however, the discussion became more sophisticated again. The Revised Hierarchical Model of Kroll and her colleagues (Kroll and Stewart, 1994), most particularly, first addressed questions of possibly asymmetric directionality in lexical connectivity in the bilingual mental lexicon.

1.2. Lexical Connections in the Bilingual Mental Lexicon The Revised Hierarchical Model (see Figure 2) posits a developmental curve over the course of which connections between words in the second language shift from linking to their meaning via translation-equivalents in the first language, to establishing direct (independent of L1) relationship with their meanings in the concept store. Thus, in the early period of acquiring a second language, words will be strongly linked to their translation-equivalents in L1, but at a later stage, they will be less strongly linked to the translation equivalents and more directly linked, presumably, not only to their meanings, but also to other semantically and idiomatically (and, perhaps phonologically) related words in the second language.

lexical links

L2

L1 conceptual links

conceptual links

concepts

Figure 2. The Revised Hierarchical Model (from Kroll and Stewart, 1994) A parallel in the monolingual’s situation would be when new words are first encountered in a vocabulary list for study, they are linked, first, to a single-word synonym definition, if there is one, or to a multi-word definition. When the word is encountered again later, more direct links to refinements of meaning will be constructed directly rather than via

188

The Mental Lexicon: Core Perspectives

definitions. More likely, of course, the monolingual, like the advanced second-language learner acquires new words in the context of reading or listening, and intuits a meaning for them from context, perhaps locating a known single-word synonym for them, but, (like the secondlanguage learner learning outside the classroom, we would argue) by attaching those words directly to their apparent meanings. We note that the hierarchical model posits independent but connected lexica. This assumption is challenged in the Bilingual Interaction Activation model (BIA, BIA+), that assumes one integrated lexicon. This approach extends interactive activation models, of the type McClelland and Rumelhardt (1981) proposed, to account for bilingual processing. It was developed by Dijkstra and his colleagues (Dijkstra and Van Heuven, 1998, 2002) to explain how interaction among levels operates to permit processing of words (written words, in this model) from bottom-up analyses of letter-features, letters, bigrams (and presumably morphemes, though these are not mentioned) to words. As the authors of these papers based much of their thinking on Dutch-English speakers, they confronted the problem that, when two languages share an orthography – the Roman alphabet in this case – the same written word may have two distinct meanings in each of the two languages (e.g., pain in English and pain meaning bread in French, or glad which means slippery in Dutch). At some point, the reader will need to identify in which language the intended meaning is to be derived, at least in experimental situations where conversational context has not made it clear. This problem led to substantial discussion of whether bilinguals operate in a selective or nonselective manner, that is, whether they can, when reading a monolingual text, restrict processing from the earliest levels to only one of their languages, or whether both are always not only accessible but also accessed, up to a point. To address this problem, proponents of this model propose a layer of language nodes that operate in top-down fashion to inhibit words in the language that should not be operating. Such a layer, one supposes, might operate in monolingual circumstances for reading words with higher and lower register homographs (e.g., crack when something is broken vs. the drug) or, indeed, with different meanings in different contexts (traffic, cars vs. drugs), or, in production, for inhibiting colloquial terms in academic discourse, or (in earlier times), inhibiting curse-words in front of elders.

1.3. Language-Selective vs. Non-Selective Processing As the twentieth century ended with more psycholinguistic focus on language processing to complement the linguists’ return to discussion of lexical structure, the models of bilingual lexical processing turned to a discussion particularly motivated by the BIA concerning whether or not – or in more sophisticated versions, to what extent – access to both languages may be narrowed when this would be desirable. Evidence for the activation of one language and the inhibition of the other during processing comes from two types of findings. One is interference from the non-target language during processing in the target language. Such interference, if found, suggests that both languages are activated. The other is what has been referred to as

Bilingual Lexica

189

switching cost. If bilinguals slow down when they have to switch between their languages as compared to when they are confronted with unilingual stimuli, separate processing of the two languages can be deduced. The experiments to determine the answer to this question date back to Kolers’ (1963) work and to later bilingual studies of Stroop (1935)-based tasks (e.g., Albert and Obler, 1978), demonstrating a negative interaction between proficiency levels in the second language and the likelihood that it will interfere in tasks where ignoring the second language is strategic (e.g., Abunuwara, 1992; Miller and Kroll, 2002). Such proficiency findings may be related to word-frequency effects in monolinguals: For the bilingual, arguably, all words are of lower frequency than they are for age- and education-matched monolingual controls in their languages, and, as Lehtonen and Laine (2003) have reported, even quite proficient Swedish-Finnish bilinguals respond to mid-frequency words like monolinguals respond to low-frequency words, whereas for monolinguals those mid-frequency words are more similar to high-frequency words. Recent experimental techniques have also demonstrated that at least for the more dominant language, the need to “switch off” the non-target language slows down performance (Kroll and de Groot, 2005). In language perception tasks, Dijkstra and others (e.g., Dijkstra et al., 2000; de Groot et al., 2000) have demonstrated that bilinguals’ performance in one language is influenced by the status of the stimuli in their other, non-target language. For example, when performing a lexical-decision task in one language, bilinguals’ response latencies to nonwords in the target language that are real words in their other language are slower than when the nonwords do not exist in the non-target language (Dijkstra et al., 2000). Recent investigations have examined the question of language selection in production tasks; to date, the degree to which both languages are active during language production is still under debate (e.g., Costa et al., 2006; Kroll et al., 2006). Furthermore, researchers have engaged in a discussion about the processing level at which cross-language competition during production can be detected. For example, are words from the non-target language active only initially, at the level of speech planning (e.g., Costa et al., 1999) or do they remain active until the later stages of lexeme selection and the execution of the articulation plan (e.g., Kroll et al., 2006). Those who maintain that production processes are language non-selective would need to explain how bilinguals do not mix their languages when talking to monolinguals. Green (1998) has proposed an Inhibitory Control Model that operates to decrease activation of a given language when it is not being employed. Grosjean (1997, 2001) proposes a continuum of activation from monolingualism to bilingualism (see Figure 3) depending on contextual cues in the environment of the speaker/listener. Because the ability to select and maintain the language appropriate to monolingual interlocutors seems to only break down with dementias (Friedland and Miller, 1999; Hyltenstam and Stroud, 1989), we may speculate that frontal-lobe executive systems develop in childhood to handle such situations, consistent with Green’s model (see also Hernandez and Meschyan, 2006). A model that incorporates frontal-system executive control, moreover, is consistent with Grosjean’s model in that the system would operate on the

190

The Mental Lexicon: Core Perspectives

“dimmer” model, rather than an on/off toggle. As de Groot and Christoffels (2006) point out, furthermore, such models must extend to account for the cross-language selective behavior of simultaneous interpreters, who must listen in one language and speak in another. They critique Paradis’ (1994) discussion of how control could work in the process of simultaneous interpretation. He suggested that it was necessary to posit different levels of activation for the source and target languages of the interpreter, such that the threshold for the source language was relatively heightened, but, they point out, precise comprehension is crucial for the task of interpreters. Rather, they suggest, one may consider inhibition of the source-language for production but not for comprehension, whereas both modalities may remain available for the target language. Linguists, of course, prefer to consider a single lexicon for comprehension and production, but, as Albert and Obler pointed out in 1978, distinction between the processes involved in comprehension and production may or may not rely on a single underlying representation for both.

LANGUAGE A (base language)

MONOLINGUAL LANGUAGE MODE

1

2

3

BILINGUAL LANGUAGE MODE

LANGUAGE B

Figure 3. Bilingual’s Language Modes (from Grosjean, 2001)

2. METHODOLOGICAL HURDLES In their paper “Models of Bilingual Representation and Processing: Looking Back and to the Future”, Kroll and Tokowicz (2005) discuss three problems with “early” bilingual models. The first of these is a confounding of the various levels of representation, in which they include the orthographic alongside phonology, syntax and semantics. To these levels, we would add morphology as well, important for the purposes of this volume, and, for a more comprehensive listing, discourse and other pragmatics. The second concern they point to is that in both theorizing that focuses on lexical representation and theorizing that focuses on access, authors tend to assume that separation for representation must align with separation for access. In fact, however, it should be “possible to have shared memory representations with selective access or

Bilingual Lexica

191

separate representations with parallel and nonselective aspects”. As well, they detail how papers regularly ignore distinctions among the different cognitive requirements for comprehension as compared to production of language. Both processes, moreover, link to, but are different from, memory as it interacts in language performance. While endorsing the distinctions they raise, here we address a complementary set of concerns: the assumptions that “most words have translation equivalents” across languages, that structural differences among specific pairs of languages (e.g., between pairs that are structurally quite similar and ones that are structurally more distant) can be ignored, and that “distance” between languages can be treated merely as a matter of shared cognates, language-family, or typology. As in Kroll and Tokowicz we discuss, as well, the issues of heterogeneity on the part of bilingual populations that researchers must confront.

2.1. Assumption of Universal Lexical Translation Equivalence While it is true that many words in a given language have single-word translation-equivalents in simple dictionaries, in fact even word-pairs that are used in experiments as translation equivalents may share some but not all of their full range of meanings and connotations, not to mention their usage possibilities. Consider the word bank, which translates to the noun banque in French, but can also be a semantically related verb in English (e.g., as in “to bank at Chase Manhattan”), and participate in idioms such as “to bank on something” in English but not in another language, even if that language uses a cognate for bank for the basic meaning. One’s abstract sense of what a bank will look like may differ based on one’s experience, so only the most core meaning “a place to store money” will be shared across markedly different cultures. This holds true for non-cognate translation “equivalents” as well: the French maison (house) will have a different range of likely images associated with it that do not altogether overlap with those of English house (and, English house will have a not-fully overlapping set in England and in the U.S.). As we observe below, moreover, morphological subconstituents of words are rarely discussed in the literature on bilingualism. Yet as they enter substantially into discussions on the human lexicon that focus on monolinguals, and as they appear to determine aspects of bilingual representation and processing when, indeed, they are studied in bilingualism, such components must not be ignored in research on the bilingual lexicon. Consider, also, that some words have single meanings in one language whereas their translation equivalents have two or more in a second language (e.g., English bank, French rive and banque). For the bilingual, it has been demonstrated (e.g., de Groot et al., 1994; Goral et al., 2006) that such differences enter into lexical representation and processing, yet rarely are such phenomena taken into account when generating stimuli for studies of the bilingual lexicon. Cognate status, by contrast, has received substantial focus in studies of the bilingual lexicon (Costa et al., 2000; Costa et al., 2005; Gollan and Acenas, 2004; Kirsner et al., 1993).

192

The Mental Lexicon: Core Perspectives

Linguistic specificities of what contributes to cognate effects are virtually ignored in the literature, however. That is, a given pair of words are assumed to either be, or not be, cognates, whereas even within a single language, speakers can benefit more or less by awareness of items coming from the same source (e.g., two, twins, between). Some measure of degree and weighting of phonological, orthographic, syntactic, semantic, and idiomatic dimensions of similarity between cognate words and among cognate sets should yield greater precision in discussions of cognate effects. Hoshino and Kroll (in press) have addressed the issue of phonological cognates versus phonological and orthographic cognates comparing Japanese and Spanish L2 users of English and found that the added value of orthography contributed to processing in some conditions.

2.2. Differences among Language-Pairs 2.2.1. Phonological. One dimension on which languages differ is the phonological. Not only will some languages have phonemic elements, even classes, that a bilingual’s other language may not; at phonology’s borders, phonetics and morphophonology, languages will have differing systems that should contribute to the greater or lesser overlap of bilinguals’ lexical organization and processing. When we were testing voice-onset-time (VOT) in highly balanced young-adult Hebrew-English speakers (Obler, 1982), we employed words that are phonemically homonymous across the two languages (e.g., /bet/: bait in English, house in certain constructions in Hebrew). In non-language-specific environments, participants would ask us which language they should write the word in, suggesting they heard the words as similar. Indeed, consonant VOT cross-over values for these bilinguals were broader than those of their respective monolinguals on comprehension tasks. The same bilinguals, however, produced stop consonants differently in Hebrew or English environments, consistent with, but not exactly the same as, the distinctions monolingual speakers made. As such phonetic differences across languages must be reflected in lexical processing for production and comprehension, we should ask whether they are reflected in the lexicon, or in some earlier level, for comprehension, and some later level, for production. That is, does the Hebrew-English bilingual “hear” /bet/ phonemically first, and then process the phonetic distinctions that would assign it to one or the other language, or never process those phonetic distinctions at all outside of experimental conditions, but, rather, comprehend the appropriate meaning in the language of the linguistic context and broader environment? And, for production, perhaps independently of the answer to the question concerning comprehension, does the speaker call up the word /bet/ as necessary, perhaps “read” a “language tag” that indicates which language to produce it in, then modify the production appropriately? 2.2.2. Morphological. Several interesting questions remain unresolved concerning how two languages’ morphological structures interact in the bilingual’s lexicon and lexical processing. One is the extent to which the morphological structure of one language influences that of the

Bilingual Lexica

193

second. This has been addressed in a series of studies by Lehtonen and Laine (e.g., 2003; Lehtonen, in press) in which they studied early Finnish-Swedish bilinguals’ processing of inflected items. Because Finnish has so many inflections possible for a given stem (something in the order of 2000), and Swedish markedly fewer, the authors expected that early, proficient bilinguals might process each of these two languages as the respective monolinguals did, and that is what they report finding. A related question is how the different morphological structure (e.g., monomorphemic vs. multimorphemic) of translation equivalents enters into their lexical connection. Data from Hebrew-English bilinguals performing a lexical-decision task with compound words as the target items revealed that bilinguals’ processing of compounds in one language influences their recognition of compounds in their other language (Goral et al., in preparation). For example, bilinguals’ judgments that pseudo-compounds were not real words when their components, if replaced with their translation equivalents, formed real compound words in the other language, yielded longer response times and higher error rates than their judgments of nonword compounds that did not exist (mutatis mutandi) as compounds in the other language. Furthermore, cross-language priming effects in which the L2 translation equivalent of a constituent of the L1 target compound was the prime were weaker for compounds that were compound translation equivalents in the two languages and were composed of the exact same constituents (e.g., sand box - ‘argaz chol) than for target compounds that had a single word translation equivalent in the prime language (e.g., the Hebrew compound lu’ach shana equivalent of the English single word calendar). The authors hypothesized that interference from the compound equivalent of the constituent prime could account for these results. Clearly, morphological equivalence or lack thereof in cross-language translation equivalents plays a role in bilingual language processing. 2.2.3. Orthographic. Orthographic differences themselves pose challenges for second-language learners, and account for part of the difficulty in learning to read a second language that does not share the same alphabet as the first one learned to read. In addition, the relation between the orthography and the lexical phonology and morphology plays a role in determining the ways in which orthographic words are represented and processed in the bilingual (and across languages as well). Frost et al. (2005), for example, demonstrated that form-priming effects are greater for the English of Hebrew-English bilinguals than they are for their Hebrew. Hebrew, like other Semitic languages, is pervasively structured around morphological roots. Their Hebrew-speaking participants, like their participants speaking another Semitic language with yet a third orthography, Arabic, showed much less form-priming and much more morphological-structure priming than was seen for the English of the bilinguals. One must conclude, then, that language-specific orthographic systems may interact with language-specific morphological structures in determining writtenword lexical organization. A similar finding was seen in the research mentioned above comparing Finnish monolinguals’ reading of lexical items and that of Finnish-Swedish bilinguals in Lehtonen and

194

The Mental Lexicon: Core Perspectives

Laine (2003). Because Finnish is an extremely affixational language one might argue, it does not make sense for Finnish readers to recognize Finnish words as whole words, but rather to decompose all but the most frequent forms, which is precisely what Lehtonen and Laine found. Their Finnish-Swedish participants, by contrast, treated even the most-frequent affixed Finnish forms as decomposable words. Because they would be expected to treat Swedish words more as full-form words, the authors suggest that the bilinguals’ treatment of frequent inflected Finnish words similarly to less frequent ones results from the likelihood that such words are not as frequent in the bilinguals’ experience of Finnish as they are in monolinguals’.

2.3. Measuring Language Distance If, as we have argued, differences among language pairs appear necessary to enter into the bilingual-lexicon calculus, then a sizable hurdle we face is the difficulty in doing so. While typologists have been studying such differences from a theoretical perspective, what we see as necessary is a measure of their psycholinguistic consequences. If, for example, Hebrew and Arabic employ markedly different orthographic systems that, nevertheless, share the characteristic that short vowels are rarely represented, what consequence does that have on the languages’ orthographic closeness as it affects language processing and production in biliterate Arabic-Hebrew bilinguals? If the pronunciation of cognate words in Modern Standard Arabic and Moroccan Arabic is so different as to be opaque to bidialectal speakers, what effect, if any, does the cognate status of the terms have on the speaker? Furthermore, similarities at the lexical level may or may not be consistent with similarities at morphological or syntactic levels and therefore, language similarities must be considered relative to the language skills examined. For example, Hindi and Arabic share a sizable number of cognate words, mostly those borrowed from Arabic into Hindi, but the syntactic and morphological structure of the two languages are quite different. Undoubtedly the multidimensional nature of characteristics that need be taken into account by the theorist of the bilingual lexicon provides a substantial challenge to research in this arena. Little research that we are aware of has considered the processing similarities and differences between bilingualism and bidialectalism, yet linguists have long been aware of the relatively arbitrary, political, nature of what is called a language as opposed to a dialect of another language (Chomsky, 2000; Jones, 2002). Jones, for example, reviews the subset of data from published case-studies of bidialectal aphasia and concludes that the patterns evidenced in bidialectal aphasia are as varied and no different from those evidenced in bilingual aphasia. In Jones et al. (submitted), moreover, it becomes clear that in bidialectal aphasia, as in bilingual aphasia, the patterns of breakdown in a given dialect can be construed as agrammatic if production conforms to neither of the participant’s two dialects. While there is substantial linguistic work on dialect structures, there is little to date that we know of on bidialectal processing, which we expect would be a fertile area for research among scholars of bilingualism.

Bilingual Lexica

195

2.4. Heterogeneous Participant Variables While the hurdles to advancing study of the bilingual lexicon that we have discussed to this point (the assumption that morphology can be ignored, the assumption that differing pairs of languages provide pretty much similar challenges to the speaker/listener, the difficulty in measuring distance between languages in different pairs) are relatively ignored in the literature, the question of participant heterogeneity is not. Nevertheless, it remains a substantial challenge to researchers. Age of acquisition of the second language has been demonstrated through a large literature to matter substantially for processing abilities (see Hyltenstam and Abrahamsson, 2003, for an excellent review). Lenneberg (1967) opened the discussion of brain bases underlying such differences by proposing a critical period cut-off occurring approximately at puberty. Others have set the timing earlier and postulated different critical periods for different language skills, or suggested a sensitive rather than critical period (e.g., Johnson and Newport, 1989; Krashen, 1973). Some researchers, Snow most prominent among them, have suggested that there is no critical period reflecting brain-substrate differences; rather, she maintains, the adult can never be immersed in language learning to the exclusion of much else, as the very young child is (Birdsong, 1998; Jia and Aaronson, 2003; Snow and Hoefnagel-Höhle, 1978). Snow’s argument is a strong form, moreover, of discussions in the mid-late twentieth century of the different representations (and, we would add, processing) that must result from different manners of second-language acquisition, most crucially for L2s acquired in the classroom as compared to via immersion in a community speaking the L2. Even within the classroom, these studies suggested, a grammar-translation method would result in tighter linking of translation-equivalent items than a total-immersion one. Learning to speak a second language with no written support to interfere would result in more native-like pronunciation of lexical items, proponents of the audio-lingual method argued. If there have been large bodies of literature on the topics of age and manner of secondlanguage acquisition, there have been fewer on years of second-language use, and virtually nothing on the extent and type of second-language use, though these are now entering some research papers (e.g., years in the country where one’s second language is spoken is only a coarse measure of one’s exposure to it or practice using it, which, presumably are the more basic phenomena affecting lexical and other levels of organization and accessing). Indeed, years in the country rarely independently contributes to performance on language tasks (Hulsen et al., 2002; Obler et al., 2005). Moreover, years in the country interacts, often, with age at time of testing, and age-related lexical-access effects are well-known in the literature on monolinguals. While they can be separated out in studies of bilinguals (e.g., Goral et al., in press), they rarely are. Wei (2000) has developed an effective measure of dominance in language use which requires individuals to list the 10 people they are closest to and then which language or languages they use to communicate with those individuals. Finally, precisely because some positive effects of bilingualism on age-related cognitive declines have been

196

The Mental Lexicon: Core Perspectives

reported (e.g., Bialystok et al., 2004), attention must be paid to current age alongside years, extent, and type of second-language use in research on the bilingual lexicon. More likely to be taken into account in the literature on bilingualism is one’s secondlanguage proficiency at the time of testing. Debates have even been conducted on whether results that have been attributed to age of second-language learning might not be better explained by proficiency at the time of testing (e.g., Abutalebi et al., 2001; Perani et al., 1998). We recommend measuring peak proficiency as well, as it hard to imagine that an individual who evidences attrition from a substantially high peak proficiency achieved decades ago has the same system of lexical (or other) representation and processing as an individual of the same age, with the same current proficiency who has just achieved that level of proficiency recently. The research hurdle in comparing these two different measures of proficiency is that it is hard enough to get researchers to agree on a single measure of current proficiency (see Bahrick et al., 1994; Galletta and Goral, 2005; Pray, 2005), and, to our knowledge, no one has yet begun the work on conceptualizing and developing measures of peak proficiency. A final issue that has been understudied with respect to the lexicon is the potential differences in lexical organization between good and poor second-language learners. There is now substantial literature on students of a second language who have problems as adolescents and young adults (e.g., Ijalba, in preparation; Ijalba and Obler, in press; Lamm and Epstein, 1999; Sparks et al., 2006), but little of it focuses on the lexicon per se. Overall, this literature suggests that individuals who have particular difficulties with second-language learning have processing deficits similar to those of dyslexia, even if they were not identified as dyslexic when learning to read in their first language due to its having particularly transparent orthography (Ijalba, in preparation), inattentive teachers, or an environment unfamiliar with dyslexia as a brain-based phenomenon. If children with learning differences are exposed to two languages from early childhood, however, there is evidence that they may acquire them both with relatively little difficulty. One must question, however, whether the structure of their lexicon or lexica is similar to that of non-dyslexics, a topic that has not, to our knowledge, been studied at all. Ignoring such a possibility, however (or for that matter, the possibility that people whose brains permit them to be particularly talented as second-language learners; Novoa et al., 1988) runs the risk of adding inter-individual variability to studies of lexical organization and processing in bilinguals.

3. FUTURE DIRECTIONS 3.1. Expanding Lexicon-Investigations beyond the Monomorphemic Word As we indicated above, morphology has been substantially undervalued in discussions of the bilingual lexicon, yet there are signs that it is starting to be treated in greater depth, especially by the research team of Laine and his colleagues who have the advantage of working in a country with two official languages that differ substantially in their morphological weights

Bilingual Lexica

197

(Finnish and Swedish, e.g., Lehtonen and Laine, 2003) and by the research team of Libben, Jarema, and their colleagues. The Semitic languages, as well, provide interesting morphological contrasts when speakers speak non-Semitic languages as well (e.g., Frost et al., 2005; Wade-Wooley and Geva, 1999). For example Goral and her colleagues have studied compound words in HebrewEnglish bilinguals, noting the role of cross-language morphological equivalence in complex word processing. Compound words have been studied also in bilingual language development (e.g., Nicoladis, 1999, 2006). Nicoladis has demonstrated that children are sensitive to the morphological structure of the words they acquire and to whether it is equivalent in their two languages. In sum, morphologically complex words will, we predict, provide rich soil for work on the bilingual lexicon in the coming decades. In addition to comparing monomorphemic and multimorphemic words in bilinguals’ lexica, it appears that the study of idioms in bilinguals’ lexica will expand. Sets of linguists and psycholinguists whose research focuses on idioms now get together for conferences, and journals devoted to this field of study cannot lag far behind. Precisely because we know that idioms are late acquisitions on the proficiency continuum, for monolinguals as well as for second-language learners, they are of interest vis-à-vis the bilingual lexicon (e.g., MonereauMurray, 2005; Wray, 2002). Also because idioms participate in lexical systems in a variety of ways, some functioning fully as inflexible unitary lexical items, others with more, but constrained, possibility for component substitutability, they provide complex, but, we maintain, useful avenues of further study, not only for understanding how the bilingual brain works, but also for getting further hints of how the lexicon works in the human brain more generally.

3.2. Increasing Sophistication of Studies 3.2.1. Bilingual aphasia. Early studies of bilingual aphasia focused not on specific questions of the lexicon, but, rather, on questions of which language returned first and why. From the middle of the twentieth century when laterality became a substantial focus in neurolinguistics, the question motivating researchers of bilingual aphasia was whether more crossed aphasia was seen among bilingual aphasics than among monolinguals (e.g., Albert and Obler, 1978; Gloning and Gloning, 1965). None of these studies provided anywhere near the level of language-history detail, nor stimulus control, that would suffice to permit linguistic conclusions concerning the lexicon. Even aphasia-therapy studies such as those of Fredman (1975) and of Watamori and Sasanuma (1976) focused on broader questions, such as whether overall, therapy in a bilingual’s second language might change their performance in the first, or whether it might do so for only production or only comprehension. In the past few years, however, several aphasia therapy studies have been designed in ways that permit greater understanding of the bilingual lexicon. Edmonds and Kiran (2006), for example, demonstrated that treating naming difficulties in the second language of bilingual individuals with aphasia

198

The Mental Lexicon: Core Perspectives

led to improvement in naming abilities in their first language as well. Furthermore, the special status of cognate words found further support in studies of bilingual aphasia that demonstrated greater cross-language treatment transfer for cognate than for non-cognate words (Kohnert, 2004). From such studies we glean that cognates have a different status from non-cognates among lexical items in that cognate-status can facilitate lexical access across the participants’ two languages in ways that mere translation equivalence cannot, and that aphasia recovery provides evidence of asymmetric directionality of the lexicon as previous studies had shown for healthy individuals as well (e.g., Kroll and Stewart, 1994). Such promising papers permit one to anticipate that more will be contributed in the future to further fill out our understanding. 3.2.2. Factors in participant selection and language-pair selection. Many of the participantselection issues we raised as hurdles are currently being considered in better-designed studies, with the exception of the distinction between current proficiency and peak proficiency, the question of changes in lexical processes associated with later adulthood, and the distinction between good and poor second-language learners. More work will need to be done to determine how peak proficiency can best be measured, and the extent to which it affects current performance. With respect to language-pair selection, we feel, a level of awareness has been achieved in the past decade or two, so overgeneralization is less of an issue than it was previously. However for pragmatic reasons researchers tend to select bilingual populations that are easily available to them, rather than those that might best answer specific questions they would wish to address. Perhaps, with advances in European Community funding, further progress may be made in this direction. Conferences and workshops such as the one on Intelligibility of Closely Related Languages (Groningen, the Netherlands, 2007) promise to complement the substantial research on receptive multilingualism (e.g., Haugen, 1966; Braunmüller and Zeevaert, 2001) by bringing focus on linguistic aspects of the process. Here, too, the insights that arise from study of bilinguals and multilinguals should contribute to issues of the lexicon more generally, in this instance the issues of cross-language differences as they influence – and interact with – speaker variability. 3.2.3. Neuroimaging. In the neuroimaging literature, as in the bilingual aphasia literature, the original questions asked were relatively broad: are the same areas of the left hemisphere involved in both of a bilingual’s languages? To what extent is right-hemisphere processing activated in bilingualism? (for reviews of this literature, see Vaid and Hull, 2002; Goral et al., 2002; Grosjean et al., 2003; and Obler et al., in press). In recent years, however, focus on specific levels of language has become possible. With respect to the lexicon, one finds such studies as that of Wartenberger et al. (2003) and De Bleser et al. (2003; see also Hernandez and Meschyan, 2003, on this study) which focus on cognate versus non-cognate status of lexical items on a naming task conducted under PET. Only for non-cognate naming in the second language did brain activation extend into the prefrontal area on the left, beyond the overlapping language areas employed for cognate and

Bilingual Lexica

199

non-cognate naming for the first language, and cognate naming in the second. This finding is not completely consistent with the shared neural networks hypothesized for the lexica of bilinguals (e.g., Ullman, 2006; Ullman et al., 2005). Ullman and others have hypothesized that lexical knowledge, attributed to explicit memory mechanisms in both L1 and L2, is processed by overlapping, posterior areas of the left hemisphere. By contrast, syntactic knowledge in L1 is hypothesized to be acquired by implicit memory mechanisms and to engage frontal areas in the left hemisphere, but not so for a later-learned L2. Therefore, syntactic skills are expected to be subserved by differing neural networks in the two languages whereas the lexicon is expected to have a shared representation for both languages. Although imaging studies present a variety of problems of interpretation, with their additional refinement, we trust, they will add precision to our understanding of bilingual lexical organization and processing.

4. CONCLUSION In the second half of the 20th century, the primary focus on the bilingual lexicon was on representation of lexical items: whether the lexicon was compound, coordinate or subordinate (e.g., Weinreich, 1953; Ervin and Osgood, 1954). Towards the end of the century, focus turned to questions of processing in the bilingual lexicon, for example, whether word-association across languages predominates, or whether, by contrast, lexical items in the two languages are linked through their shared meaning (e.g., Potter et al., 1984). These, then, were superceded by a developmental model, the Revised Hierarchical Model (Kroll and Stewart, 1994), whereby, with advancing proficiency, L2 words are processed for meaning more directly, rather than via their L1 translation equivalents. Currently, a major focus is on whether processing is language selective or not (e.g., the BIA model of Dijkstra and colleagues, cf. Costa, Sebastian-Galles, and colleagues), that is, whether both languages are available (and compete) when words from one are being processed, or whether the bilingual can constrain availability to one while processing the other for production or comprehension. Whereas these models consider levels of processing (semantic, phonological), for example by comparing and contrasting the processing of cognate and non-cognate translation equivalents, the investigations conducted to support them have assumed substantial vocabulary equivalences between pairs of languages. Only recently have investigators started to take into account differing morphological systems underlying different languages’ lexica (e.g., Frost et al., 2005, Lehtonen and Laine, 2003). In addition, they typically ignore the likelihood that the lexicon is composed not only of monomorphemic words, but also of multimorphemic words and perhaps fixed idioms as well. The challenges facing future work in this area, then, include addressing questions concerning semantic and morphological non-overlapping aspects of translation equivalents. One must also ask how the greater or lesser sharing of cognates between language pairs influences lexical representation and processing; how more and less similar phonological, morphophonological, and morphological systems do; how differing morphological systems

200

The Mental Lexicon: Core Perspectives

(e.g., root-based or not) do; and how different orthographic systems reinforce or diminish shared aspects of the lexicon. Additional hurdles include the issue of measuring distance between languages, the extent to which bidialectalism is necessarily a form of bilingualism, and how to study these aspects of lexical representation and processing in the face of extremely heterogeneous subject populations. Confronting such issues in study of bilingualism, of course, should inform studies of monolingualism as well, as currently studies of the lexicon may generalize beyond the specific language (or dialect) spoken, and may assume that all adult “normals” had similar language exposure prior to testing. The solutions lie in expanding the investigations to include morphologically complex words and words that share only part of their meaning, phonology, and morphophonology with their nearest translation equivalents, as well as including language pairs of relatively more and less differing underlying morphology and orthography. Continuing to conduct comparable experiments across various language pairs can further our understating of the effects of structural and lexical differences and similarities on bilingual processing. In addition, more can be gleaned about the variables that shape the bilingual lexicon by examining the effects of individuals' language learning history and language use. Finally, systematic evidence from the manifestation of aphasia in bilingual individuals and from patterns of recovery following treatment in bilingual aphasia (e.g., Edmonds and Kiran, 2006; Kohnert, 2004), as well as data from neuroimaging will complement behavioral data to teach us more about bilingualism and, thus, about human lexical capacities.

Bilingual Lexica

201

REFERENCES Abunuwara, E. (1992). The structure of the trilingual lexicon. European Journal of Cognitive Psychology, 4, 311-322. Abutalebi, J., S. Cappa and D. Perani (2001). The bilingual brain as revealed by functional neuroimaging. Bilingualism: Language and Cognition, 4 (2), 179–190. Albert, M. and L. K. Obler (1978). The Bilingual Brain: Neuropsychological and neurolinguistic aspects of bilingualism. Academic Press, New York. Bahrick, H. P., L. K. Hall and J. P. Goggin (1994). Fifty years of language maintenance and language dominance in bilingual Hispanic immigrants. Journal of Experimental Psychology: General, 123 (3), 264-283. Bialystok, E., F. I. M. Craik, R. Klein and M. Viswanathan (2004). Bilingualism, aging, and cognitive control: Evidence from the Simon task. Psychology and Aging, 19 (2), 290-303. Birdsong, D. (Ed.). (1998). Second language acquisition and the Critical Period Hypothesis. Erlbaum, Mahwah, NJ. Braunmüller, K. and L. Zeevaert (2001). Semikommunikation, rezeptive Mehrsprachigkeit und verwandte Phänomene. Eine bibliographische Bestandsaufnahme (Arbeiten zur Mehrsprachigkeit, Folge B,19). Universität Hamburg, Sonderforschungsbereich Mehrsprachigkeit. Hamburg. Chomsky, N. (2000). The Architecture of Language. Oxford University Press, Oxford/New Delhi. Costa, A., A. Caramazza and N. Sebastian-Galles (2000). The cognate facilitation effect: Implication for models of lexical access. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 1283-1296. Costa, A., W. La Heij and E. Navarrete (2006). The dynamics of bilingual lexical access. Bilingualism: Language and Cognition. 9, 137-152. Costa, A., M. Miozzo and A. Caramazza (1999). Lexical selection in bilinguals: Do words in the bilingual’s two languages compete for selection? Journal of Memory and Language, 41, 365-397. Costa, A., M. Santesteban and A. Caño (2005). On the facilitatory effects of cognate words in bilingual speech production. Brain and Language, 94, 94-103. De Bleser, R., P. Dupont, J. Postler, G. Bormans, D. Speelman, L. Mortelmans et al. (2003). The organization of the bilingual lexicon: A PET study. Journal of Neurolinguistics, 16, 439-456. De Groot, A. and I. Christoffels (2006). Language control in bilinguals: Monolingual tasks and simultaneous interpreting. Bilingualism: Language and Cognition, 9, 189-202. De Groot, A. M. B., L. Dannenburg and J. G. van Hell (1994). Forward and backward translation by bilinguals. Journal of Memory and Language, 33, 600-629. De Groot, A. M. B., P. Delmaar and S. J. Lupker (2000). The processing of interlexical homographs in translation recognition and lexical decision: Support for non-selective access to bilingual memory. The Quarterly Journal of Experimental Psychology, 53a, 397-4128. Dijkstra, T., M. Timmermans and H. Schriefers (2000). On being blinded by your other language: Effects of task demands on interlingual homograph recognition. Journal of Memory and Language, 42, 445–464.

202

The Mental Lexicon: Core Perspectives

Dijkstra. A. and W. J. B. Van Heuven (1998). The BIA model of bilingual word recognition. In: Localist Connectionist Approaches to Human Cognition (J. Grainger and A.M. Jacobs, eds.), pp. 198-225. Lawrence Erlbaum Associates, Mahwah, NJ. Dijkstra. T., and W. J. B. Van Heuven (2002). The architecture of the bilingual word recognition system: From identification to decision. Bilingualism: Language and Cognition, 5 (3), 175-197. Edmonds, L. and S. Kiran (2006). Effects of semantic naming treatment on crosslinguistic generalization in bilingual aphasia. Journal of Speech Language and Hearing Research, 49, 729748. Ervin, S. and E. Osgood (1954). Psycholinguistics: A survey of theory and research problems. In: Psycholinguistics (C. Osgood and T. Sebeok, eds.), pp. 136-146. Waverly Press, Baltimore. Fredman, M. (1975). The effect of therapy given in Hebrew on the home language of the bilingual or polyglot adult in Israel. British Journal of Disorders of Communication, 10, 61-69. Friedland D. and N. Miller (1999). Language mixing in bilingual speakers with Alzheimer's dementia: A conversation analysis approach. Aphasiology, 13, 427-444. Frost, R., T. Kugler, A. Deutsch and K. Forster (2005). Orthographic structure versus morphological structure: Principles of lexical organization in a given language, Journal of Experimental Psychology: Learning, Memory and Cognition, 31, 1293-1326. Galletta, E. and M. Goral (2005). Self-assessment and objective measures of proficiency in second language users. Poster presented at the Second Language Research Forum (SLRF), New York. Gloning, I. and K. Gloning (1965). Aphasien bei Polyglotten, Wiener Zeitschrift fuer Nervenheilkunde, 22, 362-397. Gollan, T. and L.-A. R. Acenas (2004). What is a TOT? Cognate and translation effects on tip-of-thetongue states in Spanish-English and Tagalog-English bilinguals. Journal of Experimental Psychology: Learning, Memory, and Cognition, 30, 246-269. Goral, M., E. S. Levy and L. K. Obler (2002). Neurolinguistic aspects of bilingualism. International Journal of Bilingualism, 6, 411-440. Goral, M., E. S. Levy, L. K. Obler and E. Cohen (2006). Cross-language lexical connections in the mental lexicon: Evidence from a case of trilingual aphasia. Brain and Language, 98, 235-247. Goral, M., G. Libben and G. Jarema (in preparation). Constituent access in bilingual compound processing. Goral, M., G. Libben, K. Ohayon and L. Obler (in press). Lexical attrition in older and younger bilinguals. In: Bilingualism in Aging: Emerging alternative perspectives (S. Makoni, N. Müller and R. Schrauf, eds.). Amsterdam: John Benjamins. Green, D. (1998). Mental control of the bilingual lexico-semantic system. Bilingualism: Language and Cognition, 1, 67-81. Grosjean, F. (1997). The bilingual individual. Interpreting, 2, 163-187. Grosjean, F. (2001). The bilingual’s language modes. In: One mind, two languages: Bilingual language processing (J. Nicol, ed.), pp. 1-22. Blackwell, Oxford. Grosjean, F., P. Li, T. Münte and A. Rodriquez-Fornells (2003). Imaging bilinguals: When the neurosciences meet the language sciences. Bilingualism: Language and Cognition, 6,159-165. Haugen, E. (1966). Semicommunication: the language gap in Scandinavia. Sociological Inquiry, 36, 280–97.

Bilingual Lexica

203

Hernandez, A. and G. Meschyan (2006). Executive function is necessary to enhance lexical processing in a less proficient L2: Evidence from fMRI during picture naming. Bilingualism: Language and Cognition, 9, 177-188. Hoshino, N. and J. F. Kroll (in press). Cognate effects in picture naming: Does cross-language activation survive a change of script? Cognition. Hulsen, M., K. de Bot and B. Weltens (2002). Between two worlds. Social networks, language shift and language processing in three generations of Dutch migrants in New Zealand. International Journal of the Sociology of Language, 153, 27-52. Hyltenstam, K. and N. Abrahamsson (2003). Maturational constraints in SLA. In: The handbook of second language acquisition (C. Doughty and M. Long, eds.), pp. 539-588. Blackwell, Rowley, MA. Hyltenstam, K. and C. Stroud (1989). Bilingualism in Alzheimer’s dementia: Two case studies. In: Bilingualism Across the Liefspan (K. Hyltenstam and L. K. Obler, eds.), pp. 202-226. Cambridge University Press, NY. Ijalba, E. (in preparation). Dyslexia in Spanish-speaking learners of English. Doctoral dissertation at City University of New York Graduate Center. Ijalba, E. and L. K. Obler (in press). Reading acquisition in Spanish-speaking learners of English. In: Clinical Communication Studies in Spanish Speakers: From research to clinical practice (J. Centeno, R. Anderson and L. K. Obler, eds.). Multilingual Matters, Clevedon, England. Jia, G. and D. Aaronson (2003). A longitudinal study of Chinese children and adolescents learning English in the United States. Applied Psycholinguistics, 24, 131-161. Johnson, J. and E. Newport (1989). Critical period effects in second language learning: The influence of maturational state on the acquisition of English as a second language. Cognitive Psychology, 21, 60-99. Jones, J. (2002). Agrammatism in a Bidialectal Speaker of AAVE and SAE. Unpublished doctoral dissertation, City University of New York Graduate School and University Center. Jones, J., M. Gitterman and L. K. Obler (submitted). A case study of a bidialectal (African-American vernacular English/Standard American English) speaker with agrammatism. Kirsner, K., E. Lalor and K. Hird (1993). The bilingual lexicon: Exercise, meaning, and morphology. In: The Bilingual Lexicon (R. Schreuder and B. Weltens, eds.), pp. 215-246. John Benjamins, Amsterdam. Kohnert, K. (2004). Cognitive and cognate-based treatments for bilingual aphasia: A case study. Brain and Language, 91, 294-302. Kolers, P. (1963). Interlingual word associations. Journal of Verbal Memory and Verbal Behavior, 2, 291-300. Kolers, P. A. and E. Gonzalez (1980). Memory for words, synonyms, and translations. Journal of Experimental Psychology: Human Learning and Memory, 6 (1), 53-65. Krashen, S. (1973). Lateralization, language learning and the critical period: Some new evidence. Language Learning, 59, 440-441. Kroll, J., S. Bobb and Z. Wodniecka (2006). Language selectivity is the exception, not the rule: Arguments against a fixed locus of language selection in bilingual speech. Bilingualism: Language and Cognition, 9, 119-136. Kroll, J. F. and A. M. B. de Groot (Eds.). (2005). Handbook of Bilingualism: Psycholinguistic Approaches. Oxford University Press, Oxford.

204

The Mental Lexicon: Core Perspectives

Kroll, J. F. and E. Stewart (1994). Category interference in translation and picture naming: Evidence for asymmetric connections between bilingual memory representations. Journal of Memory and Language, 33, 149-174. Kroll, J. F. and N. Tokowicz (2005). Models of bilingual representation and processing: Looking back and to the future. In: Handbook of Bilingualism: Psycholinguistic approaches (J. F. Kroll and A. M. B. de Groot, eds.) pp. 531-553. Oxford University Press, Oxford. Lambert, W. (1969). Psychological studies of the interdependencies of the bilingual’s two languages. In: Substance and Structure of Language (J. Puhvel, ed.), pp. 99-126. University of California Press, Berkeley and LA. Lambert, W. and S. Fillenbaum (1959). A pilot study of aphasia among bilinguals. Canadian Journal of Psychology, 13, 28-34. Lamm, O. and R. Epstein (1999). Left-handedness and achievements in foreign language studies. Brain and Language, 70, 504-517. Lehtonen, M. (2006). Morphological processing in monolinguals and bilinguals: behavioral and neural approaches. Unpublished doctoral dissertation, Abo Akademi University. Lehtonen, M. and M. Laine (2003). How word frequency affects morphological processing in monolinguals and bilinguals, Bilingualism: Language and Cognition, 6, 213-225. Lenneberg, E. (1967). Biological foundations of language. John Wiley, New York. McClelland, J. L. and D. E. Rumelhart (1981). An interactive activation model of context effects in letter perception, Part 1: An account of basic findings. Psychological Review, 88, 375-405. Miller, N. A. and J. F. Kroll (2002). Stroop effects in bilingual translation. Memory and Cognition, 30 (4), 614-628. Monnereau-Murray, M. (2005). Trilingual production: Producing common American-English Expressions. Poster presented at NYSHLA, Huntington, New York. Novoa, L., D. Fein and L. K. Obler (1988). Talent in foreign languages: A case study. In: The Exceptional Brain: Neuropsychology of Talent and Special Abilities (L. K. Obler and D. Fein, eds.), pp. 294-302. Guilford, NY. Nicoladis, E. (1999). "Where is my brush-teeth?" Acquisition of compound nouns in a French-English bilingual child. Bilingualism: Language and Cognition, 2, 245-256. Nicoladis, E. (2006). Preschool children’s acquisition of compounds. In: The Representation and Processing of Compound Words (G. Libben and G. Jarema, eds.), pp. 96-124. Oxford University Press, New York, NY. Obler, L. K. (1982). The parsimonious bilingual. In: Exceptional Language and Linguistics (L. K. Obler and L. Menn, eds.), pp. 339-346. Academic Press, New York. Obler, L. K., I. Anema, P. Conner, J. Hyun and B. O’Connor (in press). Brain organization in bilinguals, in: Speech and language disorders in bilinguals (A. Ardila and E. Ramos, eds.). Nova Science International, New York. Obler, L. K., A. Stavans, B. Kotik, E. Olshtain and G. Goldzweig (2005, October). Second-Language Comprehension and Age: Roles of age-of-acquisition and current age. Aging and Speech Communication Conference, Bloomington, IN. Paradis, M. (1994). Toward a neurolinguistic theory of simultaneous translation: The framework. International Journal of Psycholinguistics, 10, 319-335.

Bilingual Lexica

205

Perani, D., E. Paulesu, N. S. Galles, E. Dupoux, S. Dehaene, V. Bettinardi, S. F. Cappa, F. Fazio and J. Mehler (1998). The bilingual brain: Proficiency and age of acquisition of the second language. Brain, 121, 1841-1852. Potter, M.C., K.-F. So, B. von Eckardt and L. B. Feldman (1984). Lexical and conceptual representation in beginning and proficient bilinguals. Journal of Verbal Learning and Verbal Behavior, 23, 23-38. Pray, L. (2005). How well do commonly used language instruments measure English oral-language proficiency? Bilingual Research Journal, 29, 387-409. Snow, C. and M. Hoefnagel-Höhle (1978). The critical period for language acquisition: Evidence from second language. Child Development, 49, 1114-1128. Sparks, R., J. Patton, L. Ganschow, N. Humbach and J. Javorsky (2006). Native language predictors of foreign language proficiency and foreign language aptitude. Annals of Dyslexia, 56, 129-160. Stroop, J. (1935). Studies of interference in serial verbal reactions. Journal of Experimental Psychology, 18, 642-662. Ullman, M.T. (2006). Contribution of memory brain systems to first and second language. Paper presented at The Neurocognition of Second Language, Georgetown University. Ullman, M.T., R. Pancheva, T. Love, E. Yee, D. Swinney and G. Hickok (2005). Neural correlates of lexicon and grammar: Evidence form the production, reading, and judgment of inflection in aphasia. Brain and Language, 93, 185-238. Vaid, J. and R. Hull (2002). Re-envisioning the bilingual brain using functional neuroimaging: Methodological and interpretive issues. In: Advances in the neurolinguistics of bilingualism: A festschrift for Michel Paradis (F. Fabbro, ed.), pp. 315-355. Udine University Press, Udine, Italy. Wade-Wooley, L. and E. Geva (1999). Processing inflected morphology in second language word recognition: Russian-speakers and English-speakers read Hebrew. Reading and Writing, 11, 321343. Wartenberger, I., H. Heekeren, J. Abutalebi, S. Cappa, A. Villringer and D. Perani (2003). Early setting of grammatical processing in the bilingual brain. Neuron, 37, 159-170. Watamori, T. S. and S. Sasanuma (1976). The recovery process of a bilingual aphasic. Journal of Communication Disorders, 9, 157-166. Wei, L. (2000). Methodological questions in the study of bilingualism. In: The Bilingualism Reader (L. Wei, ed.), pp. 475-486. Routledge, London. Weinreich, U. (1953). Languages in contact. The Linguistics Circle of New York, New York. Wray, A. (2002). Formulaic Language and the Lexicon. Cambridge University Press, Cambridge.

This page intentionally left blank

10 SKILLS AND REPRESENTATIONS IN LEARNING TO SPELL AND IN EXPERIENCED SPELLERS

Dominiek Sandra, University of Antwerp, Antwerp, Belgium

1. CAN THE STUDY OF SPELLING INFORM US ON THE MENTAL LEXICON? Upon reading the title to this chapter one may wonder whether the study of spelling behavior really fits within the study of the mental lexicon. In order to see why it does, let’s reflect for a moment on the goals of research on the mental lexicon. The two main goals are (i) to achieve insight into the nature of our mental representations of words and (ii) to understand how these representations are accessed during real-time processing. The representations under study are of different kinds, reflecting the different properties of words: its meaning (semantic representation), its syntactic class and function (syntactic representation) and its interfaces to the sensory modalities of vision (orthographic representation) and hearing (phonological representation). Studies of the on-line access processes attempt to unravel the parameters that affect lexical access, like, for instance, form similarity (effect of so-called word neighbours) or morphological relatedness. As it turns out, some words can also be processed outside the mental lexicon, by relying on compositional or decompositional procedures. For instance, lowfrequency words may be recognized faster through grapheme-phoneme recoding than through lexical access (see Coltheart’s dual route model; Coltheart, 1978; Coltheart et al., 2001) and novel compounds and derivations are computed by a form decomposition process and a semantic composition process, as they have no lexical representation. So, part of the study of the mental lexicon also involves the identification of what does not occur in the mental lexicon itself but is nevertheless required for successful word processing. The study of spelling can shed light on both what language users store about the orthographic representations of words in their mental lexicon, and what kind of extra-lexical knowledge sources they need for adequate spelling. These knowledge sources comprise an awareness of the constituent units of words that are mapped onto graphemes, such as

207

208

The Mental Lexicon: Core Perspectives

phonemes and morphemes. The literature survey that will be reported below highlights two basic insights that have been gathered from this type of research. First, in the process of learning to spell, children need to become consciously aware of the phonemes and morphemes in words, that is, they have to learn compositional processes based on these linguistic concepts. This awareness is a precursor of being a good speller, like phoneme awareness is a reliable predictor of becoming a good reader (see below). Second, despite this awareness of rulegoverned or sometimes statistical relationships between linguistic units and graphemes, a lot of evidence points in the direction of stored orthographic patterns and spelling by analogy, both in children and experienced spellers, even in the most rule-governed domains like inflectional morphology. Thus a study of the spelling process can shed light on some of the crucial questions in research on the mental lexicon.

2. WHAT WRITTEN WORDS REPRESENT In alphabetic orthographies written words are representations of their spoken counterparts. In such orthographies, an alphabetic inventory of letter-sound correspondences is used to transcode the phonemes in spoken words into the representational units of the alphabet. Because many languages often contain more phonemes than there are letters in their alphabet (e.g., English and French; see Caravolas, 2004), letter combinations are often necessary to represent all available phonemes (e.g., digraphs like ea, ee, ck in English). This set of letters and letter combinations onto which individual phonemes can be mapped are known as graphemes. Although alphabetic languages essentially represent the linguistic level of phonemes, languages differ in the extent to which they do so. In languages with a so-called shallow orthography, like Czech, there is an almost one-to-one correspondence between phonemes and graphemes, such that many words can be spelled correctly by the mechanical application of phoneme-grapheme mappings. However, in languages with deep orthographies, the situation is far more complex. In such languages a number of factors cause different word spellings than the ones that would be produced by the phoneme-grapheme correspondences. Often, morphology is one of these factors. The spelling of a morpheme is often kept constant across the words in which it occurs, even if the pronunciation of this morpheme may vary. For instance, despite the different pronunciations of the stem in the pairs heal-health, nationnationality, relate-relation, the morpheme’s spelling is preserved in the derived member of the pairs. Spelling consistency of morphemes also occurs at the level of suffixes. For instance, even though the pronunciation of the English regular past tense suffix depends on the phonological properties of the stem-final consonant, its spelling is invariably –ed, whether its pronunciation is [t] (shocked), [d] (planned) or [Id] (loaded). The morphographic spelling principle not only ignores pronunciation variation of a morpheme, it also ignores pronunciation identity. When two different morphemes happen to have the same pronunciation, this principle

Skills and Representations in Learning to Spell and in Experienced Writers

209

leads to different spellings. For instance, the French suffixes –ent and –ant are both pronounced as [ãn] but are spelled differently because their written forms encode their different morpho-grammatical functions: the first turns adjectives into adverbs (rapidement; fast), whereas the second turns infinitives into present participles (arrivant; arriving). An even more surprising example of the claim that morphemes belong to the major representational units in written language comes from the spelling of what has been called “silent morphology” (Fayol et al., 1994). Some languages apply the principle of morphographic spelling so consistently that it is even called upon when a morpheme has no phonetic realization at all in the spoken word. For instance, in French, nouns, verbs and adjectives all take plural suffixes: s for nouns (des bouteilles; bottles), -s for adjectives (des murs blancs; white walls), -nt for verbs (ils chantent; they sing). However, despite the omnipresence of these morphemes in writing, they are not heard in the spoken language. Similar phenomena are found in Dutch spelling (Sandra et al., 1999). The observation that silent morphemes are nevertheless spelled emphasizes the fact that orthographies adopting the morphographic principle not only spell phonemes in the word’s sound structure but also morphemic units in the word’s linguistic representation. Apparently, rendering the morphemic components in words is considered so important that it occurs at the expense of faithfully representing the word’s component phonemes. Apart from representing phonemes and morphemes, written words are also characterized by specific orthographic conventions. These conventions, which may sometimes be described in the form of rules and sometimes in the form of mere tendencies, give rise to orthographic patterns. Picking up these patterns is an additional task for children who learn to spell. An example of clear orthographic rules is the doubling of letters. In orthographies like those for English and French, some consonants can be doubled (e.g., ss and nn occur in English) whereas others cannot (e.g., hh and jj do not occur in English). Moreover, doubling is restricted to certain positions in the word: in English consonant doublets can occur in final but not in initial position (mass, *mmass), whereas in French they can only occur in word-medial position. Learning which consonants can be doubled involves idiosyncratic learning, whereas learning about the allowance or disallowance of consonant doublets in a word position can be described with rules. Both must be learnt in order to become a proficient speller. Apart from such orthographic rules, there are also orthographic tendencies, which do not guarantee a flawless spelling but can still be quite helpful. For instance, the spelling of vowel sounds is notoriously difficult in English. The vowel [i:] is spelled ee in deep and reel but as ea in clear and mean. However, the nature of the following consonant considerably reduces the ambiguity of the vowel spelling (Treiman et al., 1995). When the coda consonant is a [p] or [l], the [i:] sound is often spelled as ee but when it is [r] or [n], it is often spelled as ea. However, words like beer and keen or cheap and meal demonstrate that these are only tendencies and that many individual words will require the memorization of their spelling pattern.

210

The Mental Lexicon: Core Perspectives

To summarize, an alphabetic orthography may represent three types of information. By definition, it represents phonemes. Deep orthographies tend to represent morphemes as well. At the sub-morphemic level there are orthographic restrictions that are imposed by word position or by neighboring sounds. In the next paragraphs we will turn our attention to the question of which cognitive skills beginning spellers need in order to learn how to spell. First, we will consider two key concepts for the development of spelling ability: phonological awareness (section 3) and morphological awareness (section 4). As we will see these are precursors to the development of the orthographic representations referred to in section 1. Following this, we will consider how orthographic patterns are learnt that are not supported by these metalinguistic skills (section 5). This will shed light on the role of lexical storage. Sections 3 through 5 all concern aspects of spelling development. We will end this chapter by considering research on experienced spellers (section 6), more particularly, on the use of error patterns for studying the nature of spelling processes. These, too, will inform us on lexical storage, even for words with descriptively clear-cut computational procedures.

3. PHONOLOGICAL AWARENESS Even though spelling may not be the same kind of automatic process as word recognition, experienced spellers write many words fast and effortlessly, without reflecting much on their individual letters (things are different for really difficult words like hippopotamus, conscientious). Such behavioral characteristics come close to the use of automatic processes, which are performed without much conscious effort and with the allocation of few attentional resources. Instances of such automatic processes outside the context of literacy are walking, riding a bicycle and driving a car, daily activities that many of us perform in an automatic fashion because we have learnt how to sequence and coordinate the individual components of these motor activities quickly and without devoting conscious attention to them. However, such a learning process takes time and its initial stages are laborious. Anyone who has seen a baby trying to make its first steps and then trying to walk longer distances, with more confidence and greater speed, will know that learning the individual components of a motor activity that will eventually become fluent, is a painstaking and often frustrating activity. There is nothing fundamentally different when children learn how to spell the words in their language. In order to learn how spoken words should be transcoded into a sequence of graphemes they must struggle with the sounds to become aware of the units in the spoken words that are orthographically represented. They must also implicitly or explicitly (through classroom instruction) learn the orthographic patterns and rules that are involved. Finally, they have to memorize the spelling patterns of words that cannot be derived from any mapping system (lexical learning). Eventually, most of them will spell effortlessly.

Skills and Representations in Learning to Spell and in Experienced Writers

211

3.1. Phoneme Awareness Predicts Spelling Success Phonological awareness plays a key role in the acquisition of literacy. The evidence showing the importance of phonological awareness for becoming a good reader is overwhelming (Adams, 1990; Bryant and Bradley, 1985; Goswami and Bryant, 1990). Preliterate children with a good phonological sensitivity – as can, for instance, be measured by having them indicate which word in a list of three starts or ends with a different phoneme (oddity task) or by having them delete a phoneme – tend to become better readers than children who experience problems with such tasks. Similarly, a child’s success in spelling is uniquely predicted by his or her phoneme awareness. In a longitudinal study reported by Caravolas et al. (2001) only phoneme awareness and knowledge of letter-sound correspondences independently predicted English-speaking children’s phonologically acceptable (but not necessarily correct) spellings in the first year of formal schooling. Other cognitive measures, like IQ and memory, did not explain unique variance in the spelling data. Moreover, once phonological awareness has made it possible for the child to learn the basics of the spelling process, i.e., rendering the phonological structure of the word, the learner seems to be on the right track for taking the next steps as well. Caravolas et al. found that first-year children who produced phonologically acceptable spellings were also the ones who produced more conventionally correct spellings in their second year. The latter finding may be linked to the phenomenon of orthographic self-teaching (Share, 1995, 1999). It has been shown that children who can successfully decode words have better orthographic representations of the words they have read. The idea is that when children successfully decode a word they know which word they have read and, hence, can store the word’s orthographic properties together with its pronunciation. In an experiment aimed at selfteaching in Grade 1, Cunningham (2006) demonstrates that children’s skill in decoding words predicts how well they can remember the spelling pattern of a word they have seen three days earlier, when the task makes them distinguish the word’s spelling from a homophonic distractor and two other distractors. Since phonological decoding and phonologically accurate spelling both hinge on phonological awareness, this experiment is in line with Caravolas et al.’s finding that phonologically accurate spellers learn the word-specific orthographic patterns faster than their peers.

3.2. The Role of Phoneme Awareness in Consistent Orthographies It may not be all that surprising that children’s early spelling performance is only explained by phoneme awareness and the knowledge of letter-sound correspondences. In an orthography based on the alphabetic principle, phonemes are the basic units in spoken words that must be encoded and graphemes are the basic units of the code system. Hence, knowing how to identify the phonemes and knowing which letter (grapheme) is conventionally linked to these phonemes in the alphabet predicts a high success in initial spelling attempts. However,

212

The Mental Lexicon: Core Perspectives

orthographies differ in their consistency in applying phoneme-grapheme correspondences. A possible consequence could be that phoneme awareness may produce larger gains in consistent, shallow orthographies than in less consistent, deep orthographies. Experiments have indeed confirmed this hypothesis. Caravolas and Bruck (1993) compared the spelling performance of Czech and English children who had been learning to read and write for eight months. Czech has a highly consistent orthography in which the ratio of phonemes to letters is almost one to one (Caravolas, 2004). When the two groups of children had to spell nonwords that were equated for difficulty the Czech children produced more phonologically accurate spellings than the English ones, confirming the effect of phonemic transparency on the rate of spelling development. Wimmer and Landerl (1997) arrived at a similar conclusion. They compared German-speaking Austrian children and British children in grades 2, 3 and 4 on their ability to spell the vowels in words that were matched across the two languages. Performance was highest in German, the language with a higher phoneme-grapheme consistency. Moreover, British children produced more alternatives for the vowel spelling than the German children did, indicating that children’s progress in learning to spell depends on the consistency in the phoneme-grapheme correspondences in the written vocabulary. Learning to spell in an alphabetic orthography always capitalizes on phoneme awareness, but children take more advantage from this awareness if there is a consistency in the mappings. Even though a phonologically consistent orthography boosts the initial learning process, it has been claimed that the high phoneme-grapheme consistency in the orthography of some languages leads to a short-lived effect of phoneme awareness on literacy measures. Wimmer et al. (1991) have proposed that a high phoneme-grapheme consistency quickly leads to a mastery degree of phonological awareness, irrespective of the preliterate child’s awareness level, such that the measurable effect of this factor on spelling dissipates much more rapidly than in less consistent orthographies. A comparable point has been made by Öney and Durguno lu (1997). Still, several studies indicate that phoneme awareness has a more long-lasting predictive relationship on spelling success. In an unpublished study by Bruck et al. (1996), discussed in Caravolas (2004), a comparison was made between the word and nonword spelling performance of English and French children in Grade 3, taking their scores of phoneme awareness and letter knowledge from kindergarten or Grade 1 as predictors. As French is a more consistent language than English, the effect of phoneme awareness might no longer be measurable by Grade 3. However, phoneme awareness in Grade 1 was a strong predictor of spelling accuracy in both languages. Caravolas (2004) demonstrates on the basis of the data from Bruck et al. (1996) that an effect of phoneme awareness on later spelling performance can be demonstrated if the difficulty of the tests allows for a sufficiently large range of scores. In a comparison between Czech and English children in the age range between 7 and 12 years Caravolas, Volín et al. (2005) found that phoneme awareness had the same impact on literacy measures in the two languages: it affected both the children’s reading speed and

Skills and Representations in Learning to Spell and in Experienced Writers

213

spelling performance. Also, Czech dyslexic children in these age groups had problems with phonemic awareness tasks, suggesting that the underlying cause of their reading problem was of a phonological nature (the phonological deficit account of dyslexia is now widely accepted, see Snowling, 2000; Vellutino et al., 2004). Recall that Czech orthography is characterized by highly consistent grapheme-phoneme correspondences. So, this study, too, shows that phonological awareness is a key component in children’s spelling development beyond the first two years of literacy education, even in a highly consistent orthography. These authors also emphasize the point about task difficulty: the tasks used to measure phonological awareness and spelling performance should be sufficiently difficult to create enough variance in the children’s scores (for instance, the child had to delete the second consonant in a CCVC word or the penultimate consonant in a CVCC word). Finally, the recent longitudinal study by Nikolopoulos et al. (2006) leads to the same conclusion. They found that phoneme awareness accounted for unique variance in the spelling data of Grade 2 and Grade 4 Greek children, who can rely on a consistent orthography. Remarkably, one year later the spelling data of these same children were still predicted by their initial phonological awareness scores, even after the effect of their spelling performance in the first test year had been statistically removed. It appears that differences in phoneme awareness continued to increase the gap between good and poor spellers, which emphasizes the importance of this factor in the development of spelling accuracy. It is noteworthy that a measure of phonological processing speed also predicted children’s spelling scores in the Caravolas, Volín et al. (2005) and Nikolopoulos et al. (2006) studies, independently of phoneme awareness. In the former study, the reading speed of both the Czech and English children predicted their spelling accuracy, whereas in the latter study the children’s speech rate (operationalized as the speed at which they could repeat pairs of pseudowords) was a significant predictor. These findings suggest that spelling accuracy is not only determined by children’s conscious awareness of phonemes but also by their fluency in processing phonemes.

3.3. Syllable Awareness and Spelling Success To conclude this section on phonological awareness it is important to point out that the term phonological awareness is not restricted to the awareness of phonemes. Syllables, for instance, are also phonological units and syllable awareness might have an independent impact on literacy scores as well. Aidinis and Nunes (2001) showed that this is indeed the case, at least in Greek. Even though Greek has consistent phoneme-grapheme mappings its vocabulary is characterized by a large number of words with CV syllables, which create clear syllable boundaries. Hence, Greek children may treat syllables as phonological units and link them to their spelling patterns. Aidinis and Nunes found that performance in a syllable oddity task indeed uniquely predicted the scores on the spelling (and reading) performance of Grade 1 and 2 children, after controlling for phoneme awareness measured in a phoneme oddity task. The

214

The Mental Lexicon: Core Perspectives

authors conclude that Greek children’s syllable awareness predicts success in spelling (and reading) independently of phoneme awareness. However, an effect of syllable awareness on spelling does not imply that children go through a syllabic spelling stage, in which each syllable is represented by one letter, before proceeding to a phonemic stage. The Greek study certainly does not warrant this conclusion, as syllable and phoneme awareness were independent predictors. Yet, such a stage has been proposed for Romance languages (Ferreiro, 1990; Vernon and Ferreiro, 1999) based on the observation that children often write one letter per syllable, generally a vowel (e.g., UUU for Portuguese urubu). However, Pollo et al. (2005) demonstrated that this is an artifact of children’s strategy of spelling letters whose name they can hear in a word and the fact that many more vowel letter names appear in Portuguese words than in English words. Thus, these findings emphasize that even preliterate children attempt to spell words by attending to their constituent sounds and use a very sensible strategy for mapping sounds onto letters: “listen whether you recognize a letter name in the word and spell the corresponding letter if you do”. This seems to be a precursor of phoneme awareness. In conclusion, the above review indicates that phonological awareness, particularly phoneme awareness, is a strong predictor of successful spelling in alphabetic orthographies, both consistent and inconsistent ones.

4. MORPHOLOGICAL AWARENESS 4.1. Morphological Awareness Predicts Morphemic Spelling As mentioned in section 1, many orthographies adopt a principle of morphographic spelling. One may wonder whether learning to deal with morphemic spellings also requires a form of metalinguistic awareness. The idea sounds reasonable enough. If one wants to learn a representational code, one may have to become aware of (i) the units that are represented by this code and (ii) the corresponding units in the representational code itself. Section 2 has amply demonstrated the correctness of this idea as far as phonemes are concerned. There are indications that even Grade 1 children already have some form of morphological awareness that helps them spell morphologically complex words correctly. Treiman et al. (1994) found that children in a word spelling completion task spelled flap sounds more correctly when they corresponded to a stem-final phoneme than when they appeared in monomorphemic controls (e.g., the t in dirty versus duty). Treiman and Cassar (1996) also reported evidence for children’s reliance on morphological knowledge in spelling. Even though first graders and kindergartners often omit the first of two consonant phonemes in word-final position, they made fewer omission errors when the first consonant was a stem-final letter (the n in tuned) than when it appeared in a phonologically matched control (the n in brand). This reflects a very early sensitivity to the constituent morphemes of words.

Skills and Representations in Learning to Spell and in Experienced Writers

215

While showing the importance of morphological structure to spelling performance, Treiman’s experiments did not investigate whether variability in children’s morphological awareness predicts variability in their spelling of morphemic patterns. Research that specifically targeted this question revealed such a relationship. This was most clearly demonstrated in a series of studies by Nunes and colleagues (Bryant et al., 2000; Nunes et al., 1997a,b; Nunes et al., 2006). In a first set of experiments, Nunes et al. (1997a,b) studied children’s use of the -ed spelling for the English regular past tense. Children in grades 2, 3 and 4 had to spell ten regular past tenses (in addition to irregular verbs and nonwords) at the onset of the study, once more seven months later, and again thirteen months later. At the onset of the study their morphological awareness was also measured in three tests. In a word analogy test, they heard a word pair that instantiated a morphological relationship between two words and had to apply that same relationship to a new word (e.g., anger-angry, strength-___ ). In a sentence analogy task they had to change the verb tense, and in a morphological production task, they had to inflect a novel word (a variant on the wug test). The scores on the two analogy tasks predicted the children’s correct use of the -ed spelling for the regular past tense morpheme seven months later, even after differences in age, IQ, and phonetic spelling at the beginning of the study (measured on irregular past tenses and nonwords) were removed. The word analogy task still predicted unique variance in the spelling data collected thirteen months later. Two conclusions can be drawn. First, morphological awareness predicts success in spelling morphemic patterns, beyond sound awareness. Second, the best way to measure this variable is to use a test that involves conscious reflection on the morphological relationship between two words. Arguably, a wug test can be performed without metalinguistic awareness, which seems to be the key factor underlying morphological awareness. Nunes et al. (1997b) replicated these results with nonword verb forms. In this study too, they demonstrated that children’s success in using the -ed spelling is predicted by their scores on a word analogy task, even after the effects of age, IQ, and performance in a phoneme oddity task (phonological awareness) had been removed. Phonological awareness also had an independent effect on succesful -ed spelling, as its effect was still significant after removing the effect of morphological awareness. Bryant et al. (2000) reported a study on the use of a different morphemic spelling pattern: the genitive, whose orthographic marker is the apostrophe. As in their study of the -ed spelling, they found that children’s capacity in keeping the spelling patterns of the genitive and the homophonous plural forms apart was predicted by their skill in solving word analogies based on a morphological relationship. Morphological awareness had a unique effect, as it was significant after age, IQ, and reading age had been removed. Hence, in order to distinguish between the spelling patterns for morphologically distinct but homophonous word forms like bird’s and birds, one needs morphological awareness. Note that the items in the word analogy test did not include any genitives or plurals, indicating that the underlying predictor does not involve a particular morphological relationship but rather a general morphological awareness.

216

The Mental Lexicon: Core Perspectives

Nagy et al. (2006) also found a relationship between children’s morphological awareness and their spelling skill. In contrast to the studies by Nunes and colleagues, the spelling test was not restricted to a particular grammatical category (e.g., past tense), but consisted of a list of progressively more difficult words. The authors found that children who knew which suffix made a word, or novel poly-morphemic word, fit in a sentence (e.g., dogless) and who were good at solving morphological relatedness questions (e.g., does mother come from moth?) were also good spellers. This relationship obtained after phonological awareness effects had been removed from the data, again indicating a unique role for morphological awareness. The effect was stable across grades 4 through 9. Note that, as in the experiments by Nunes and colleagues, the morphological awareness tests required deliberate reflection on the morphological structure of words. All the above findings were obtained in English but similar results have been reported for French. In French many word-final letters are not pronounced. Interestingly, the identity of the final letter can sometimes be inferred from the pronunciation of a derived word. For instance, although the last letter of the word galop is not pronounced, one can derive that it is a p on the basis of the derived verb galopper. Hence, thinking of morphological relationships can be an effective spelling strategy in French. Sénéchal et al. (2006) compared the spelling performance of Grade 4 children on French words that can be spelled by applying phonemegrapheme correspondences (lac; lake), words that can be spelled by relying on a morphological relationship (galop; gallop), and words that can only be spelled on the basis of memory retrieval (tabac; tobacco). They observed the best performance on phonological words but also found higher scores on words with morphological relatives than on words requiring a lexically based spelling. Children reported the use of a morphological spelling strategy (i.e., using morphological relationships between words) most often for words with morphological relatives. When they reported to have used this strategy, their spellings were highly accurate, demonstrating the effectiveness of the strategy. The authors demonstrated that children’s spelling accuracy on words with morphological relatives, and their use of morphological strategies, were predicted by their morphological awareness. This relationship remained after the effect of general spelling skill was removed, indicating a unique role of the skill in detecting morphological relationships. This skill was measured by asking the children to generate a derived word on the basis of an example word pair (e.g., gris-grise, gray; blond-___, blond). Production of the correct response implied that the silent sound in the probe word was pronounced. Success at this task predicted success in spelling word-final silent letters on the basis of a derived word. The above studies all emphasize the impact of children’s morphological awareness on their spelling performance. However, learning to deal with morphologically based spelling patterns may in turn increase children’s morphological awareness. As a matter of fact, such a reciprocal relationship has been claimed to hold between phonological awareness and alphabetic literacy (Olson, 1996). Nunes et al. (2006) reported results that suggest a similar bidirectional relationship between morphological awareness and the accuracy on morphemic

Skills and Representations in Learning to Spell and in Experienced Writers

217

spelling patterns. In their study, children’s success at spelling the past tense -ed predicted their performance in morphological awareness tasks one year later. Also, their consistency in spelling morphemic patterns predicted their ability in defining novel words on the basis of their constituent morphemes. Apparently, the ability to spell morphemic patterns predicted success in tasks requiring reflection on morphological relationships. The findings summarized in this section all indicate that children’s explicit awareness of morphological relationships between words has a unique effect on their success at spelling suffixes (Bryant et al., 2000; Nunes et al., 1997a,b) and silent letters that can be inferred from morphological relatives (Sénéchal et al., 2006). This factor predicts variance in spelling accuracy even after the effect of phonological awareness is removed. Clearly, children need different types of metalinguistic awareness when learning to spell.

4.2. Morphological Awareness develops in Stages According to Nunes et al. (1997a) and Bryant (2002) children pass through a series of stages when discovering a morphemic spelling pattern. The nature of these stages was derived from Nunes et al.’s (1997a) spelling data with children in grades 2 through 4. These children had to spell three sets of ten words each: regular past tenses like called and dressed, irregular past tenses like found and felt, and nonverbs like bird and belt. Half of the words in each set ended in the [d] sound, the other half in the [t] sound. These sounds were spelled as d and t respectively in the irregular verb and nonverb group whereas they were both spelled as -ed in the regular verb group. On the basis of the error data the authors derived a stage model of learning to spell the past tense. The first stage is marked by unsystematic spellings. In the second stage, children spell all t and d sounds phonetically, including when they appear at the end of regular past tenses (e.g., kissed is spelled as kist). In the third stage, they realize that the orthographic pattern -ed is one way of representing the t or d sounds, but they do not know that it is restricted to the particular grammatical category of past tense verbs and make overgeneralization errors (e.g., bird is spelled as bired). In the fourth stage, they become aware that the -ed spelling is reserved for the spelling of past tenses but do not know that it is restricted to the spelling of regular forms (e.g., slept is spelled as sleped, possibly in some cases because they make a regularization error in speech). Finally, in the fifth stage, they know the exact application conditions for this spelling pattern and have learnt that irregular forms are exceptions to the general rule. In short, there appear to be two major steps in learning a morphemic spelling pattern. The first step is the discovery of a novel orthographic pattern. The second is the progressive fine-tuning of its usage conditions. According to Bryant (2002) this is the typical way in which this type of learning proceeds.

218

The Mental Lexicon: Core Perspectives

5. LEARNING ORTHOGRAPHIC REGULARITIES As mentioned in section 1, there are orthographic patterns that cannot be explained as the result of simple mappings of phonemes or morphemes onto graphemes. Such patterns involve orthographic conventions, like the fact that consonant doublets never appear in word-initial position in English, or statistical regularities, like the fact that a given vowel phoneme is more frequently spelled in one way than another. Finally, orthographic patterns also involve frequently recurring strings of letters in the written input. Even though stage theories of spelling development propose that beginning spellers only use a phonological spelling strategy, Martinet et al. (2004) showed that children already rely on word-specific orthographic patterns after only three months of literacy education. French children obtained better spelling scores on high-frequency words than on lowfrequency ones. Moreover, when spelling nonwords they applied atypical phoneme-grapheme mappings more often when the nonword resembled a familiar word in which this mapping occurred (e.g., spelling the /o/ sound in /niro/ as op, in analogy with the spelling of the word sirop, syrup). In their study, Martinet et al. had strict control over the written language that the children had been exposed to. The frequency and analogy effects demonstrate that children store orthographic representations of words in their mental lexicon from the beginning of their literacy education. Orthographic learning in early literacy not only occurs at the level of the word. First graders read Dutch CVC pseudowords faster when either the CV or the VC (rime) sequence was a high-frequency spelling pattern in the written language of the children (Geudens and Sandra, 2002), suggesting that they had stored these sublexical orthographic patterns. Children also rely on sublexical orthographic knowledge for spelling. Caravolas, Kessler et al. (2005) studied how very young children cope with the problem of inconsistent vowel spellings in English. They showed that British children in the last year of kindergarten (mean age: 5 years and 7 months) and six months later in mid-year of Grade 1 tended to spell the grapheme that most frequently represents a particular vowel. Their spellings also reflected a smaller but significant effect of word frequency and, to some extent, rime frequency. However, these young children were not sensitive to the conditional probability of the vowel spelling, i.e., the spelling given the coda, which considerably reduces the spelling inconsistency in English (Treiman et al., 1995). These findings show that children’s sensitivity to the frequency of orthographic patterns is limited to patterns that are easy to detect in the input: the association strength between individual phonemes and their various graphemic realizations and the frequency of the entire letter string. The most important finding, of course, is that such young children keep track of orthographic regularities and patterns at all. Whereas the spelling of a vowel in an inconsistent orthography such as English is by definition not rule-governed but probabilistic, some orthographic patterns can be described by simple spelling rules. The positional constraints on consonant doublets are an instance in case. As previously mentioned, consonant doublets in English can occur in word-final but not in

Skills and Representations in Learning to Spell and in Experienced Writers

219

word-initial position. In French, such doublets can only occur in word-medial position. It seems reasonable that, if children can derive statistical tendencies in their written input, they can also derive orthographic rules, as all the facts in the written input support an orthographic rule. Cassar and Treiman (1997) studied children’s implicit knowledge of consonant doublets. They showed that first and second graders knew which consonants could occur as doublets in English words and which ones could not. This effect is, of course, an instance of frequency-based orthographic learning rather than rule induction. However, in the same study the authors found that children rejected the presence of consonant doublets at word beginnings, indicating that they had induced the positional constraints on consonant doublets. Pacton et al. (2001) repeated this study in French and obtained the same results, when using materials that controlled for possible alternative accounts of the Cassar and Treiman data. Although the theoretical import of the Pacton et al. study goes much farther than this (see below), their experimental data reiterate the point that young children learn orthographic rules without being instructed to do so. Orthographic learning is an automatic by-product of being confronted with written words. A remarkable and consistent finding is that, even though children induce an orthographic rule, they still store word-specific properties of the rule-governed words. For instance, even though French children are not explicitly taught that the spelling of the diminutive suffix in French is always -ette, they make this induction themselves. Pacton et al. (2002) found that a diminutive suffix that was attached to a nonword was more often spelled as -ette when the item was presented in a sentence context (e.g., a little /vitar/ is a /vitaret/), which indicates that the children knew how to spell the diminutive suffix. However, despite the children’s sensitivity to the rule, their responses were also determined by the occurrence frequency of orthographic patterns. The suffix was spelled correctly more often after an r than after an f, which reflects the frequency difference of the orthographic patterns rette and fette in French. This, again, indicates the effect of lexical storage on spelling. Comparable findings were reported in Nunes et al.’s (1997b) study of the -ed spelling for spelling the past tense suffix. An untaught rule stipulates that a past tense takes the -ed spelling when its stem has the same pronunciation in the present and past tenses and the irregular -d or -t spelling otherwise (e.g., clean-cleaned, kick-kicked but hear-heard, sleepslept). Nunes et al. introduced pseudowords in their present and past tense forms in sentences, thus making it possible to (implicitly) infer whether the verb was regular or not (e.g., My friend always prells at bedtime. We usually prell in the morning, but last week we /prold/ in the afternoon). Children’s spellings of the final sound of the past tense indicated that they could indeed distinguish regular from irregular verbs, which is another demonstration of implicit learning in the domain of spelling. Interestingly, however, they performed better if the irregular pseudoverbs were analogous to existing verbs in English (e.g., lind-/laund/ along the lines of find-/faund/) than when they were not (e.g., soan-/sand/), even though they distinguished regulars from irregulars for the latter item type as well. This indicates that the rule was not

220

The Mental Lexicon: Core Perspectives

systematically used and that the children also relied on their mental lexicon when determining their spelling response. The impact of the mental lexicon is quite strong. Even when the orthographic rule is explicitly taught in the classroom, children cannot avoid the influence of frequently occurring word forms. This is the conclusion of Pacton and Fayol’s (2003) study of the spelling of the French suffix /ãn/. This homophonous suffix can be spelled as -ent or as -ant, depending on whether it marks an adverb or a present participle. Children in Grade 3 and Grade 5 had to spell frequent and rare adverbs and present participles, in isolation and in sentence contexts. Spelling performance was better in sentence contexts, in both grades, indicating that the children were aware of the association between the grammatical function of the suffix and its spelling. Despite this, the children made fewer spelling errors on the frequent words than on the rare ones, in both spelling conditions. A frequency effect in sentence context would not be expected if children simply applied the spelling rule, especially since rule application for the French participle form is reliably triggered by the word en. The frequency effect again demonstrates that children do not apply rules systematically and are influenced by their lexical knowledge. Interestingly, the frequency effect did not interact with the age of the children, even though their spellings significantly improved between Grades 3 and 5. The stability of this effect indicates that explicit knowledge of an orthographic rule does not easily “block” the influence of the mental lexicon. The findings of frequency effects in domains where rules unambiguously predict the correct spelling emphasize the importance of stored orthographic information for words. Although frequency is expected to assist spelling performance when nothing else can help (“spell what is most probable”), it is not when there is a clear rule. Of course, children in the studies described above may not have developed a sufficiently clear rule yet and therefore kept relying on stored orthographic representations of words. On the other hand, the persistent finding of frequency effects and effects of analogy in these spelling experiments may suggest that the idea of (mental) rules behind spelling is misguided. Do we really rely on mental representations that make complete abstraction of the particular data in the learning experience? Pacton et al. (2001) argue that rule-governed behavior is based on analogical processing. They show that French children’s implicit knowledge of the positional constraints on consonant doublets generalizes to doublets that do not exist in French, but consistently find that violation errors with non-existing doublets are less often detected. This decrement in generalization is found until Grade 5, that is, after years of opportunities for learning the abstract “rule”. If abstraction resulted in the induction of such a rule, the decrement in performance with non-existing doublets should become smaller over time, as the rule would ignore the familiarity of the doublets. As the evidence rejected this prediction, the authors argue for a view in which children’s implicit learning of orthographic regularities is based on stored exemplars and a mechanism that can discover statistical regularities in these exemplars, like the one that has been proposed in the connectionist framework. Rather than representing

Skills and Representations in Learning to Spell and in Experienced Writers

221

abstract mental rules, storing the orthographic patterns of words may actually be the rule of our mind.

6. THE HOMOPHONE FREQUENCY EFFECT AND RULE-GOVERNED WORDS The persistent impact of familiar orthographic patterns in the spelling of rule-governed words that has repeatedly been observed in studies with children (see section 5) has also been observed in the writings of experienced spellers. Despite their much more extended spelling experience and their arguably better knowledge of orthographic rules these spellers still cannot ignore familiar orthographic patterns. In the spellings of experienced writers the use of orthographic representations has largely been diagnosed on the basis of intrusion errors. Conditions favouring intrusion errors occur when a rule-governed target word is homophonous with another rule-governed word that does not fit the grammatical context of the sentence. For example, the Dutch verb forms vind (find) and vindt (finds) create a risk for intrusion errors, even though they require different grammatical subjects (1st and 3rd person singular, respectively), because they are homophonous forms in the present tense paradigm of the verb vinden. Several lines of evidence indicate that in such conditions, experienced spellers tend to make homophone intrusion errors when the syntactically inappropriate form has a higher written frequency than the correct form. Hence, sometimes they do not follow the rule but rely on the frequency of stored orthographic representations. Comparable results have been obtained in Dutch and in French. My colleagues and I reported several experiments in which we studied the spelling of homophonous verb forms in Dutch. We focused on two types of homophone relationships. The first and third person singular of the present tense are pronounced identically if the verb stem ends in a d. In the example above, the first person form vind and the third person form vindt (finds) are both pronounced as /vint/, even though only the latter has the overt inflectional marker t. Another type of homophone relationship involves the third person singular present tense and the past participle of verbs with a so-called weak prefix. For instance, the third person form gebeurt and the past participle gebeurd are both pronounced as /gəbørt/, but the former contains the third person suffix t whereas the latter contains the past participle ending d. For each type of homophony, Sandra et al. (1999) selected verb pairs that differed with respect to the grammatical form that had the highest frequency. Eighteen-year old students had to spell these verb forms in sentences. The verb forms were either adjacent to the word that controlled their suffix spelling (i.e., the subject or the auxiliary in the case of past participles) or separated from that word by four intervening words. The sentences were dictated under time-pressure to prevent response checking. For each grammatical form (and homophone type), intrusion errors were more likely when the target was the lower-frequency member of the homophone pair, indicating that spellers were biased to spell the high-frequency form. Moreover, these errors occurred more often when the verb form was separated from the word determining its suffix spelling (see Assink, 1985, for similar data).

222

The Mental Lexicon: Core Perspectives

These results reveal two processes in the spelling of experienced writers. First, the homophone frequency effect indicates that these writers have orthographic representations of verb forms that can be spelled on the basis of rules, and are inclined to spell the high-frequency form of a homophonous target. Second, the probability that they do not apply the rule and fall back on lexical retrieval is determined by the available time for rule application. In order to apply the rule, they have to identify the word that determines the spelling of the verb suffix in their working memory. When the time to retrieve this information, (i.e., the time set by four intervening words in our speeded dictation task), is longer than the available time (i.e., the dictation speed), they will not be able to apply the rule and will spell the most probable form, which is the highest-frequency one. Frisson and Sandra (2002) found that children from Grade 5 to the first year of secondary school progressively showed this effect of homophone frequency. Moreover, Frisson and Sandra (submitted) replicated the effect in an experiment where nonwords were primed with homophonous verb forms. The spelling of the nonword’s final sound mimicked the frequency relationship between the homophonous forms of the prime. This effect was not found with non-homophonous noun primes with the same rimes, indicating that the effect was lexical rather than sublexical. Interestingly, quite similar results were obtained in French (for a systematic comparison between the Dutch and French experiments, see Sandra and Fayol, 2003). Largy et al. (1996) studied verb-noun homophones and found that the spelling of the suffix in inflected verb forms sometimes corresponded to the spelling of a suffix for the homophonous noun. For instance, in sentences like, Le chimiste prend des liquides. Il les filtre. (The chemist takes liquids. He filters them.) participants made errors like filtres, which is the homophonous noun plural. The preceding direct object pronoun les is homographic to the plural article in French and hence could miscue a noun reading of the homophone. Importantly, however, the intrusion of noun spellings occurred more often when the noun had a higher frequency than the verb, which echoes the frequency effect in the Dutch data.

7. WHAT HAVE WE LEARNT AND WHAT DO WE STILL HAVE TO LEARN? Spelling is a complex activity, which requires a number of component skills. In alphabetic orthographies, where written words represent both phonemes and morphemes, children’s phonemic and morphological awareness are both independent predictors of their spelling success. In addition, children implicitly learn general and word-specific orthographic patterns from the texts they are confronted with. They keep relying on orthographic patterns even when they have learnt a spelling rule. Experienced spellers still rely on the frequency properties of orthographic representations when there is limited time for reflection on rule application. In such conditions, they spell the high-frequency form of a homophonous rule-governed word. So, what have we learnt that is relevant for the study of the mental lexicon? Three observations are directly relevant. First, orthographic learning starts as soon as children learn to spell and appears in the form of effects of frequency and analogy (Martinet et al., 2004).

Skills and Representations in Learning to Spell and in Experienced Writers

223

Lexical knowledge keeps influencing word spelling even when the spelling pattern can be unambiguously determined from the word’s syntactic category, as demonstrated by the orthographic frequency effects on the spelling of the French suffixes -ent and -and (Pacton and Fayol, 2003). The impact of familiar orthographic patterns even surfaces in rule-governed domains with which the speller has had years of experience. This is exemplified by the finding that children know that doublets cannot occur in some word positions but are more confident about this restriction when the doublet exists in the language. Even expert spellers rely on lexical knowledge when they have to spell rule-governed words, as demonstrated by the frequency effect on homophone intrusions for inflected forms (Sandra et al., 1999). This dependency on lexical knowledge in experienced spellers particularly occurs in conditions of limited working-memory capacity. A second observation is that spelling does not only depend on lexical knowledge. As a matter of fact, it largely depends on a conscious awareness of the linguistic units that are encoded: phonemes and, in deep orthographies, morphemes as well. This is the point where the study of spelling seems to differ from the study of the other three word-related skills: visual word recognition, spoken word recognition, and spoken word production. The latter skills are prototypical instances of automatized skills, which are performed without any conscious attention (with the exception of noisy input conditions, strange words, tip of the tongue states, etc.). The strong impact of metalinguistic skills on spelling performance suggests that spelling is a much less automatized skill. However, this is only partially true. In this review we have focused a lot on the learning process, in large part because the data on expert performance are rather scarce. It should not come as a surprise that, as for any learning process, learning the skill requires a lot of attentional effort. As a matter of fact, learning to read words, which eventually becomes a reflex-like act (as elegantly demonstrated by the Stroop interference effect), also starts out as a painstaking activity. It seems quite plausible that many components of the spelling process also eventually become automatic once the process has been overtrained, that is, once one has had sufficient spelling experience. Experienced spellers certainly do not have the impression of having to reflect on each and every letter while writing words. The observation that even such writers make frequency-determined homophone intrusions on inflected word forms emphasizes the importance of rapid lexical access during the spelling process and the subsequent left to right read-out of the orthographic pattern. In other words, the effect of metalinguistic awareness on spelling, while essential, only reveals the necessary conditions for getting the skill started and says nothing about the processes underlying written word production in the expert writer. A last conclusion concerns the use of rules. While rules seem to be all over the place in spelling, several experimental findings reported in this review suggest that lexical knowledge can never be in an “off” state. Even when very simple rules concerning positional constraints on consonant doublets are concerned, the fact that even Grade 1 children are already “aware” of these (Cassar and Treiman, 1997) does not imply that they will eventually become abstract mental rules. Indeed, Pacton et al. (2001) demonstrated that the familiarity of orthographic

224

The Mental Lexicon: Core Perspectives

patterns always has an additional effect, suggesting that abstract orthographic rules may not exist in our minds and that the concept of a mental lexicon and a mechanism for the induction of regularities may offer a better account. Even though we know many facts about spelling performance, it is clear that we know far too few such facts to build a processing model for the experienced speller. I am not aware of an experimentally supported spelling model that represents the way in which orthographic representations are retrieved from the mental lexicon, what variables affect this retrieval, and how the retrieved representation is serially output as a sequence of letters in the correct order. In order to find out more about this, the spelling process must be investigated by means of realtime measures. For instance, (more) studies are needed in which typing duration and/or intervals between key strokes are measured to assess the effect of variables that might affect retrieval. At the same time, analyses of spelling errors will continue to provide a window on the spelling process, playing the same role that speech errors play in the study of speech production.

Skills and Representations in Learning to Spell and in Experienced Writers

225

REFERENCES Adams, M.J. (1990). Beginning to read. Thinking and learning about print. MIT Press, Cambridge. Aidinis, A. and T. Nunes (2001). The role of different levels of phonological awareness in the development of reading and spelling in Greek. Reading and Writing: an Interdisciplinary Journal, 14, 145-177. Assink, E.M.H. (1985). Assessing spelling strategies for the orthography of Dutch verbs. British Journal of Psychology, 76, 353-363. Bruck, M., M. Caravolas and F. Genesee (1996). A longitudinal study of early literacy acquisition in English- and French-speaking Canadian children. (discussed in Caravolas, 2004) Bryant, P. and L. Bradley (1985). Children’s reading problems. Blackwell, Oxford. Bryant, P. (2002). Children’s thoughts about reading and spelling. Scientific Studies of Reading, 6, 199216. Bryant, P., T. Nunes and M. Bindman (2000). The relations between children’s linguistic awareness and spelling: the case of the apostrophe. Reading and Writing: an Interdisciplinary Journal, 12, 253276. Caravolas, M. (2004). Spelling development in alphabetic writing systems: A cross-linguistic perspective. European Psychologist, 9, 3-14. Caravolas, M. and M. Bruck (1993). The effect of oral and written language input on children’s phonological awareness: A cross-linguistic study. Journal of Experimental Child Psychology, 55, 130. Caravolas, M., C. Hulme and M. J. Snowling (2001). The foundations of spelling ability: Evidence from a 3-year longitudinal study. Journal of Memory and Language, 45, 751-774. Caravolas, M., B. Kessler, C. Hulme and M. J. Snowling (2005). Effects of orthographic consistency, frequency and letter knowledge on children’s vowel spelling development. Journal of Experimental Child Psychology, 92, 307-321. Caravolas, M., J. Volín and C. Hulme (2005). Phoneme awareness is a key component of alphabetic literacy skills in consistent and inconsistent orthographies : Evidence from Czech and English children. Journal of Experimental Child Psychology, 92, 107-139. Cassar, M. and R. Treiman (1997). The beginnings of orthographic knowledge: Children's knowledge of double letters in words. Journal of Educational Psychology, 89, 631-644. Coltheart, M. (1978). Lexical access in simple reading tasks. In: Strategies of information processing (G. Underwood, ed.), pp. 151–216. Academic Press, London. Coltheart, M., K. Rastle, C. Perry, R. Langdon, and J. Ziegler (2001). DRC: A Dual Route Cascaded model of visual word recognition and reading aloud. Psychological Review, 108, 204–256. Cormier, P. and S. Kelson (2000). The roles of phonological and syntactic awareness in the use of plural morphemes among children in French immersion. Scientific Studies of Reading, 4, 267-293. Cunningham, A. E. (2006). Accounting for children’s orthographic learning while reading text: Do children self-teach? Journal of Experimental Child Psychology, 95, 56-77. Fayol, M., P. Largy and P. Lemaire (1994). Cognitive overload and orthographic errors : When cognitive overload enhances subject-verb agreement errors. A study in French written language. The Quarterly Journal of Experimental Psychology, 47A, 437-464.

226

The Mental Lexicon: Core Perspectives

Ferreiro, E. (1990). Literacy development: psychogenesis. In: How children construct literacy: Piagetian perspectives (Y. Goodman, ed.), pp. 12-25. International Reading Association, Newark, DE. Frisson, S. and D. Sandra (2002). Homophonic forms of regularly inflected verbs have their own orthographic representations: a developmental perspective on spelling errors. Brain and Language, 81, 545-554. Frisson, S. and D. Sandra (submitted). The homophone frequency effect in inflected verb forms: Evidence from spelling primed nonwords. Geudens, A. and D. Sandra (2002). The role of orthographic onset-rime units in Dutch beginning readers. In: Precursors of functional literacy (L. Verhoeven, C. Elbro and P. Reitsma, eds.), pp. 4967. Kluwer, Dordrecht, The Netherlands. Goswami, U. and P. Bryant (1990). Phonological skills and learning to read. Lawrence Erlbaum, Hove. Largy, P., M. Fayol and P. Lemaire (1996). The homophone effect in written French: The case of verbnoun inflection errors. Language and Cognitive Processes, 11, 217-255. Martinet, C., S. Valdois and M. Fayol (2004). Lexical orthographic knowledge develops from the beginning of literacy acquisition. Cognition, 91, B11-B22. Nagy, W., V. W. Berninger and R. D. Abbott (2006). Contributions of morphology beyond phonology to literacy outcomes of upper elementary and middle-school students. Journal of Educational Psychology, 98, 134-147. Nikolopoulos, D., N. Goulandris, C. Hulme and M. Snowling (2006). The cognitive bases of learning to read and spell in Greek: Evidence from a longitudinal study. Journal of Experimental Child Psychology, 94, 1-17. Nunes, T., P. Bryant and M. Bindman (1997a). Morphological spelling strategies: Developmental stages and processes. Developmental Psychology, 33, 637-649. Nunes, T., P. Bryant and M. Bindman (1997b). Learning to spell regular and irregular verbs. Reading and Writing: an Interdisciplinary Journal, 9, 427-449. Nunes, T., P. Bryant and M. Bindman (2006). The effects of learning to spell on children’s awareness of morphology. Reading and Writing: an Interdisciplinary Journal, 19, 767-787. Olson, D. (1996). Towards a psychology of literacy: On the relations between speech and writing. Cognition, 60, 83-104. Öney, B. and A. Y. Durgunoğlu (1997). Beginning to read in Turkish: A phonologically transparent orhography. Applied Psycholinguistics, 18, 1-15. Pacton, S. and M. Fayol (2003). How do French children use morphosyntactic information when they spell adverbs and present participles? Scientific Studies of Reading, 7, 273-287. Pacton, S., M. Fayol and P. Perruchet (2002). The acquisition of untaught orthographic regularities in French. In: Precursors of functional literacy (L. Verhoeven, C. Elbro and P. Reitsma, eds.), pp. 121137. Kluwer, Dordrecht, The Netherlands. Pacton, S., P. Perruchet, M. Fayol and A. Cleeremans (2001). Implicit learning out of the lab: The case of orthographic regularities. Journal of Experimental Psychology: General, 130, 401-426. Pollo, T. C., B. Kessler and R. Treiman (2005). Vowels, syllables, and letter names : Differences between young children’s spelling in English and Portuguese. Journal of Experimental Child Psychology, 92, 161-181.

Skills and Representations in Learning to Spell and in Experienced Writers

227

Sandra, D. and M. Fayol (2003). Spelling errors with a view on the mental lexicon: Frequency and proximity errors in misspelling homophonous regular verb forms in Dutch and French. In: Morphological structure in language processing (R. H. Baayen and R. Schreuder, eds.), pp. 485-514. Mouton de Gruyter, Berlin. Sandra, D., S. Frisson and F. Daems (1999). Why simple verb forms can be so difficult to spell : The influence of homophone frequency and distance in Dutch. Brain and Language, 68, 277-283. Sénéchal, M., M. T. Basque and T. Leclaire (2006). Morphological knowledge as revealed in children’s spelling accuracy and reports of spelling strategies. Journal of Experimental Child Psychology, 95, 231-254. Share, D.L. (1995). Phonological decoding and self-teaching: Sine qua non of reading acquisition. Cognition, 55, 151-218. Share, D.L. (1999). Phonological recoding and orthographic learning: A direct test of the self-teaching hypothesis. Journal of Experimental Child Psychology, 72, 95-129. Snowling, M.J. (2000). Dyslexia: A Cognitive Developmental Perspective, second edition. Blackwell, Oxford. Treiman, R. and M. Cassar (1996). Effects of morphology on children’s spelling of final consonant clusters. Journal of Experimental Child Psychology, 63, 141-170. Treiman, R., M. Cassar and A. Zukowski (1994). What types of linguistic information do children use in spelling? The case of flaps. Child Development, 65, 1310-1329. Treiman, R., J. Mullenix, R. Bijeljac-Babic and E. D. Richmond-Welty (1995). The special role of rimes in the description, use, and acquisition of English orthography. Journal of Experimental Psychology: General, 124, 107-136. Vellutino, F. R., J. M. Fletcher, M. J. Snowling and D. M. Scanlon (2004). Specific Reading Disability (Dyslexia) : What have we learned in the past four decades? Journal of Child Psychology & Psychiatry, 45, 2-40. Vernon, S.A. and E. Ferreiro (1999). Writing development: A neglected variable in the consideration of phonological awareness. Harvard Educational Review, 69, 395-414. Wimmer, H. and K. Landerl (1997). How learning to spell German differs from learning to spell English. In: Learning to spell: Research, theory and practice across languages (C.A. Perfetti, L. Rieben and M. Fayol, eds.), pp. 81-96. Erlbaum, Mahwah, NJ. Wimmer, H., K. Landerl, R. Linortner and P. Hummer (1991). The relationship of phonemic awareness to reading acquisition: More consequence than precondition but still important. Cognition, 40, 219249.

This page intentionally left blank

INDEX

ABSL, see Al-Sayyid Bedouin Sign Language (ABSL) Abstractionist models, 98 Acceptability judgments, 107, 108, 112 Acoustic duration, 89, 90, 92, 94–5 Activation, 40–2, 147, 188 Additive model, 107, 112 Adequacy, explanatory vs. descriptive, 33–5 Affix, affixation, 68, 93, 94, 113–6 Affix numerosity, 141 Age of acquisition, 195 Agglutinating languages, 121 Algorithm (Marr), 17, 27, 107, 108 Al-Sayyid Bedouin group, 57 Al-Sayyid Bedouin Sign Language (ABSL), 57–9 Analogical generalization, 86, 97 Analogy, 112, 168–70, 176, 208, 215, 218, 220 Aphasia, 171, 172, 176–8 grammatical productivity in, 171, 172, 176–8 Arabic, 58, 146, 170, 194 Articulation, 59, 88, 90, 94, 95, 136, 189 coarticulation, 91 double, 59 hyper-articulation, 90 hypo-articulation, 90 precision, 91 Assimilation, 94 Attractors, 12 Autonomous lexical processor, 35–40 Autonomy, 39 Bayesian approach, 34 Bidialectalism, 194, 200 Bilingual aphasia, 197–8 Bilingualism, 189, 191, 194, 195, 196, 200

Bilingual lexicon future directions, 196–9 language-selective vs. non-selective processing, 188–90 lexical connections in bilingual mental lexicon, 187–8 methodological hurdles, 190–1 heterogeneous participant variables, 194–6 language distance, measurement, 194 language-pairs, differences among, 192–4 universal lexical translation equivalence, assumption of, 191–2 shared or separate systems, 185–7 “Bin” model of lexical access, 44 Binyanim, see Conjugations Bloomfieldian model, 82–3 Body neighbors, 133 Bound, boundedness (of morphemes), 114, 115, 119, 121 Brain, 195, 196, 197 Cartesian theatre, 16 Causality, 12 CELEX database, 19, 24 Chance performance, 41 Chaotic behaviour, 12 Chinese, 106, 112, 113, 114, 115, 117, 123, 129 Coarticulation, 91 Cognate, 191–2, 194, 198, 199 Cognitive psychology, 9, 107 Cognitive system, 2 Cohort, 89 in comprehension, 89, 91 in production, 92 Competence, 105, 106

229

230

Index

Complexity, structural, 12, 85 Composition, 131, 137, 146, 147–50, 160 Compositionality, 55–7, 61, 107, 109, 110, 113 Compound(s), compounding, 64, 106, 113–23, 138, 168, 173, 175, 193 Comprehension, 2, 84, 89, 90, 91, 92, 93, 95, 111, 190, 191, 192, 199 Computation, 5, 8, 12, 26 maximization of, 82 minimization of, 82 Computational modeling, 31–3 Computation (Marr), 107, 108 Computer program, 32 Conjugations, 72 Construct profligacy, 10 Constructs, 4, 8, 9, 14 Context and morpheme, 121–3 Context effects, 36, 37 Contextual diversity, 43 Corpus, corpora, 86, 90, 95, 106, 111, 112, 113, 117, 119, 120, 121 Corpus analysis, 119–21 Critical period, 195 Cross-language differences in word recognition, origins of, 129–30 frequency and semantic relatedness, effects of, 136–8 morphological relatedness, effects of, 138–9 form and meaning vs. morphological relatedness, under additive effects of, 144–5 list composition on semantic processing, influence of, 147–9 modulation of facilitation, 141–2 regular vs. irregular inflectional relatedness, facilitation due to, 142–4 single words, modulations of morphological processing in, 139–40 word-internal patterning of morphemes, 145–7 orthographic similarity, effects of, 130–1 form effects, absence of, 134–5 form similarity, alternative measures of, 132–3

orthographic neighborhood size, 131–2 prime frequency and lexicality on form priming , effects of, 133 unmasked comparisons across languages, 135–6 Cross-language variation in word recognition, 130 Cross-tense naming, 85 Crossword mode, 38 Dataset-general variance, 19 Declarative memory, 97–8 Decomposition, 3, 18, 87, 88, 94, 95, 97, 108, 121 Deep orthography, 192, 193, 196, 200, 208, 210, 211, 212–3 Degree of complexity, 4 Degree of functional integrity, 2 Derivation, 138, 141, 146, 177 Derivational entropy, 139 Designing experiments, visual word recognition, 46–8 Diminutives, 172, 173, 175, 178 Distributed morphology (DM), 105, 109 “Divide and conquer” approach, 6 DM, see Distributed morphology (DM) Dominance, 195 Domino causality, 12 DRC model, 31, 32, 33 Dual mechanism, 82–3 Dynamic Bayesian networks, 98 Effect sizes, 21, 27 Embodiment, 87, 97 Encapsulation, 87 Entropy, 6 inflectional, 14, 85–6 Exemplar-based models, 98 Experienced spellers, 207–24 Experimental design, 49 Fine phonetic detail, 82, 88–93, 94, 95, 96 syllabification and, 93–6 Finnish, 139, 140, 160, 177, 189, 193 Formal language, 98 as calculus, 98

Index Form and meaning vs. morphological relatedness, under additive effects of, 144–5 Form similarity, 131, 132–3, 149 Fractal pattern, 13–4 Frequency and acoustic duration, 94 and acoustic reduction, 96 effect of, 42–4, 83, 85 Functor, 109, 113, 115, 116 Generalized linear mixed effect modeling (GLMM), 118 Generative grammar, 105 Generative morphology, 105–24 Generative semantics, 109 Genetic programming (GP), 5, 17 German, 169, 177, 178 Germanic languages, 74 GP, see Genetic programming Graded facilitation, 143 Gradedness, 162 Grammar, 35 Grammatical productivity in aphasia, 176–8 language acquisition, in early first, 175–6 and language typology, 170–2 norms, on level of, 172–4 paradigmatic dimension, 166–7 analogy, role of, 168–70 fully productive WFRs, 167–8 WFR, unproductive, 167 syntagmatic dimension, 161–2 loan words with fitting properties, morphological integration of, 164 loan words with unfitting properties, morphological integration of, 163 WFR to abbreviations, application of, 164 WFR to another, shift from one, 165 WFR to indigenous words, application of, 165–6 Head, 13, 19, 47, 119, 120 Hebrew, 146, 148, 151, 170, 192 Hierarchical temporal memory, 5, 97–8, 98 Hindi, 194 Homographic stem, 145

231

Homophone frequency effect, 207–24 Homophone frequency effect and rulegoverned words, 221–2 HTM, see Hierarchical temporal memory (HTM) Hungarian, 171–2, 175, 176, 177–8 Hypothetical constructs, 8, 9 “Ideal speaker-hearer,” 106 Idioms, 5, 191, 197, 199 Implementation (Marr), 113 Implicit learning, 49, 219, 220 Incremental activation, 40 Inflecting languages, 160, 161, 171 Inflectional entropy, 14, 85, 86, 139, 140 Inflectional relatedness, regular vs. irregular, facilitation due to, 142–4 Interactive activation models, 188 Interfix, 94–5 Intervening variables, 8–9 Inverse transformation, 26 Irregular inflection, 138, 142–4 Judgment-based diagnostics, 117–9 Latin, 68–9 deponent verb roots, 70 LDRT, see Lexical decision reaction times (LDRT) Learning orthographic regularities, 218–21 Lemma, 56, 83, 86 Lexeme, 55, 56, 68 formed on Modern Hebrew root KB, 73 formed on Modern Hebrew root ZLP, 72 Lexical competition, 88–93, 94, 95 and computation of fine phonetic detail, 88–93 Lexical decision, 14, 36–7, 38, 41, 42, 43, 44, 45, 46–7, 84, 86, 95, 106, 121 Lexical decision reaction times (LDRT), 14, 18, 19, 20–1, 22, 23–4 predictors, 20–2, 24, 25 Lexical decision task, 14, 36, 37 Lexical frequency effects, 84 Lexical hypothesis, 56–7 Lexical information, 16, 90, 92 Lexicalist hypothesis, 56, 109

232

Index

Lexical processor, 33 Lexical roots, 68–9 Hebrew roots, 69–76 Linguistic level, 76 Linguistic theory, 106 Linking models, 107 List composition, 130, 137, 146, 147–9 on semantic processing, influence of, 147–9 Listeme, 109, 111, 115, 122 Logarithmic transformation, 26, 42 Machine learning, 98 Magnetoencephalography (MEG), 112 Manner (of acquisition), 138 Masked priming, see Priming Mathematical fractal patterns, 13 Meaning interaction and form, 87–8 Memory, 142 for phonological sequences (morphs), 84 for sequences, 84 Mental lexicon, 1–3, 160, 176 core perspectives, 3–4 approach to modeling, 5–6 common architecture, 4 insights, 4–5 definition, 2, 3 goal of, 3 spelling and, 207–8 storage and computation, 81–2 experimental evidence, 82–8 phonetic evidence, 88–96 Mental rules, 220, 221, 224 MiniJudge, 118, 119 Modern Hebrew, 69, 71, 75–6 Modifier, 58, 115, 122, 123 Modularity, 33, 35, 87, 107 Modulation of facilitation, 141–2 Morphemes frequency effects, 121, 150 word-internal patterning of, 145–7 Morphographic spelling, 208–9 Morphological awareness, 214–7 development in stages, 217 morphemic spelling, 214–7 Morphological effects on fine phonetic detail, 93

Morphological family size, 8, 14, 138, 139, 149 Morphologically complex words, meanings of, 59-68 Morphological priming, 139, 146, 147 Morphological richness, 138, 139 Morphological rules, 83, 163, 187 Morphological spelling strategy, 216 Morphonology, 171 Morphosemantic and morphotactic transparency, 178 Morphosemantics, 161 Morphotactics, 161, 168 MROM model, 31, 33 Naming task, 37 Naming time, 13, 14, 37 Native speakers, 3 Naturalistic University of Alberta Nonlinear Correlation Explorer (NUANCE), 18, 19, 21, 22, 23, 24, 27 Natural morphology, 161, 162 ND, see Neighborhood distribution (ND) Neighborhood density positional, 91 Neighborhood distribution (ND), 132 Neural network, 11–2 Neuroimaging, 198–9 Noncompositional meaning, 61 Nonlinear activation functions, 12 Nonlinearity, 11–16 Nonlinear regression, 8 Nonlinear variable effects, synthetic approaches to nonlinearity, 11–6 psycholinguistics , mathematical synthesis in, 16–9 method, 19 results, 20–6 psycho-ontology, 8–9 construct extravagance, 10–1 Nonwords, 42 NUANCE, see Naturalistic University of Alberta Nonlinear Correlation Explorer (NUANCE) Ontology, 8–11 OR function, 12

Index Orthographic activation, 135 Orthographic depth, 129, 130, 136, 137, 138, 150 Orthographic neighborhood: density, 131 size, 131–2, 134, 149 Orthographic neighbors, 131–3 Orthographic patterns, 218, 219, 221, 223 Orthographic priming, 130, 139 Orthography, 196, 200 Overtensing, 84 Paradigmatic structure, 85 bias, 93 likelihood of voicing, 95 Paradigmatic axis, 178 Paradigmatic dimension, 167–70 Parallel activation, 40 Participant selection and language-pair selection, factors in, 198 PDP model, 32 Perceptron, 11 Person/number/gender combinations, 74 Phoneme awareness, 212–4 in consistent orthographies, role of, 211–3 and spelling success, 211 syllable awareness and spelling success, 213–4 Phoneme-grapheme consistency, 212 Phoneme-grapheme correspondences, 208, 212 Phonetic evidence, storage and computation lexical competition and computation of, 88–93 morphological effects on fine phonetic detail, 93 syllabification and fine phonetic detail, 93–6 Phonological awareness, 210–4 phoneme awareness in consistent orthographies, role of, 211–3 and spelling success, 211 syllable awareness and spelling success, 213–4 Phonological neighbors, 91, 95 Phonological neighbourhood, 8

233

Pink noise, 13 Pocket calculator, 82 Positional encoding, 98 Pragmatics, 190 Prelexical, 107, 112 Prime frequency, 133, 135 Priming masked, 45, 48–50 forward, 132, 147, 149, 150 strategies and, 48–50 morphological, 139, 146, 147 relation, 122, 123 semantic, 45, 144, 147 Probabilistic, 218 Processing, 195, 196 Productivity grammatical, see Grammatical productivity in (lexical) phonology, 111 morphological, 111 of seaman schema to grammar, 108–109, 108–9 syntactic, 11 of word formation, 159–61 Proficiency, 189, 196, 197 Psycholinguistics, 8 formalization, 83 mathematical synthesis in method, 19 results, 20–6 variables of, 15 Psycho-ontology, 8–9 construct extravagance, 10–1 R, 118 RAM, 16 Realism, 9 Reciprocally causal hypothesis, 12, 13 Refutation, 39 Regularity, word, 61, 130 Regular verbs, English, 82 Relation priming, 122, 123 Root, 55, 57, 63, 109–10 Root class, 71, 73, 74 Rule-governed words, homophone frequency effect and, 221–2 Rules, symbolic, 159, 160, 162

234

Index

Second-language use, 195 Selectional restrictions, 116, 118 Selective, 188–90 Semantic bleaching, 115–6, 121 Semantic categorization, 44, 45, 48 Semantic decisions, 44–5 Semantic density, 87, 97 Semantic priming, 45, 144, 147 Semantic relatedness, 136–8, 144, 148, 150 Semantic similarity, 87, 139, 144, 145, 146, 147, 150, 151 Semantic transparency, 138, 146, 151 degree of, 148, 138, 146, 151 Semitic, 74, 193, 197 languages, 69, 71 Serbian, 86, 129, 135, 136, 137, 138, 140, 145, 146 Shallow orthography, 208 Sign language, 56, 57–8, 59 Single words, modulations of morphological processing in, 139–40 Speech errors, 84, 89, 112, 224 Spelling and mental lexicon, 207–8 Stage model, 217 Storage maximization of, 82 minimization of, 82 Storage and computation, phonetic evidence lexical competition and computation of fine phonetic detail, 88–93 morphological effects on fine phonetic detail, 93 syllabification and fine phonetic detail, 93–6 Subject verb object (SVO), 58 SVO, see Subject verb object (SVO) Switching, 189 Syllabary, 88 Syllabification, 93 and fine phonetic detail, 93–6 Syllable awareness and spelling success, 213–4 Syntagmatic axis, 178 Syntagmatic dimension, 166, 170, 171, 174 Synthesis, 16

Tautology, 10 Tense, 74, 82–5, 84 Translation, 191–2 Transmission, 108 Triangle model, 35, 87 Typology, language, 170, 175, 191 Variable-predictor relations, 20–1 Verbs class, 73 of position and motion in English and Dutch, 88 Visual word recognition activation, 40–2 adequacy, explanatory vs. descriptive, 33–5 autonomous lexical processor, 35–40 computational modeling, 31–3 frequency effect, form of, 42–4 semantic decisions, 44–5 strategies and masked priming, 48–50 WEAVER model, 83, 89, 92–3 WFR, see Word formation rule (WFR) White noise, 13 Word formation and productivity, 159–61 Word-formation rule (WFR), 160, 161, 162, 163, 165, 166 Word-internal patterning of morphemes, 145–7 Word-level semantics, 109, 123 idiosyncrasies, 109 Word recognition, cross-language differences in, origins, 129–30 frequency and semantic relatedness, effects of, 136–8 morphological relatedness, effects of, 138–9 form and meaning vs. morphological relatedness, under additive effects of, 144–5 list composition on semantic processing, influence of, 147–9 modulation of facilitation, 141–2 regular vs. irregular inflectional relatedness, facilitation due to, 142–4

Index single words, modulations of morphological processing in, 139–40 word-internal patterning of morphemes, 145–7 orthographic similarity, effects of, 130–1 form effects, absence of, 134–5 form similarity, alternative measures of, 132–3 orthographic neighborhood size, 131–2

235

prime frequency and lexicality on form priming , effects of, 133 unmasked comparisons across languages, 135–6 Wortbedeutung, 161 Wortbildung, 160 Wortbildungsbedeutung, 161 Wortgebildetheit, 160 Written word representation, 208–10

This page intentionally left blank