THE PSYCHOLOGY OF LANGUAGE

THE PSYCHOLOGY OF LANGUAGE

9781841693828_1_pre.PM5 9/21/07, 9:55 AM1

Copyright © 2008 Psychology Press http://www.psypress.com/harley/

9781841693828_1_pre.PM5 9/21/07, 9:55 AM2


The Psychology of LanguageFrom Data to Theory

Third Edition

Trevor A. Harley

9781841693828_1_pre.PM5 9/21/07, 9:55 AM3


First published 2008 by Psychology Press, an imprint of Taylor & Francis27 Church Road, Hove, East Sussex BN3 2FA

Simultaneously published in the USA and Canada by Psychology Press270 Madison Avenue, New York, NY 10016

Psychology Press is an imprint of the Taylor & Francis Group,an informa business

© 2008 Psychology Press

All rights reserved. No part of this book may be reprinted orreproduced or utilized in any form or by any electronic,mechanical, or other means, now known or hereafter invented,including photocopying and recording, or in any informationstorage or retrieval system, without permission in writingfrom the publishers.

The publisher makes no representation, express or implied, with regardto the accuracy of the information contained in this book and cannotaccept any legal responsibility or liability for any errors or omissionsthat may be made.

British Library Cataloguing in Publication DataA catalogue record for this book is available from the British Library

Library of Congress Cataloging-in-Publication DataHarley, Trevor A.

The psychology of language : from data to theory / TrevorHarley.—3rd ed.

p. cm.Includes bibliographical references and indexes.ISBN 978-1-84169-381-1—ISBN 978-1-84169-382-8

1. Psycholinguistics. I. Title.BF455.H2713 2008401′.9—dc22

2007021225

ISBN 978-1-84169-381-1 (Hb)ISBN 978-1-84169-382-8 (Pb)

Typeset in 10/12pt Times by Graphicraft Limited, Hong KongPrinted and bound in the UK by Ashford Colour Press Ltd, Gosport, Hampshire

9781841693828_1_pre.PM5 9/21/07, 9:55 AM4


For Siobhan, without whom everythingwould be impossible

9781841693828_1_pre.PM5 9/21/07, 9:55 AM5


9781841693828_1_pre.PM5 9/21/07, 9:55 AM6


Contents

Preface xiHow to use this book xv

SECTION A: INTRODUCTION 1

1. The study of language 3Introduction 3What is language? 5The history and methods of

psycholinguistics 9Language and the brain 14Themes and controversies in

modern psycholinguistics 19Summary 24Some questions to think about 25Further reading 25

2. Describing language 27Introduction 27How to describe speech sounds 27Linguistic approaches to syntax 34Summary 45Some questions to think about 46Further reading 46

SECTION B: THE BIOLOGICALAND DEVELOPMENTAL BASESOF LANGUAGE 49

3. The foundations of language 51Introduction 51

Where did language come from? 51Do animals have language? 54The biological basis of language 67The cognitive basis of language:

The cognition hypothesis 79The social basis of language 82The language development of

visually and hearing-impairedchildren 84

What is the relation betweenlanguage and thought? 87

Summary 99Some questions to think about 100Further reading 100

4. Language development 103Introduction 103The driving forces of language

development 105Do children learn any language

in the womb? 119Phonological development 120Lexical and semantic development 125Early syntactic development 136Summary 150Some questions to think about 151Further reading 151

5. Bilingualism and second languageacquisition 153Introduction 153Bilingualism 153

vii

9781841693828_1_pre.PM5 9/21/07, 9:55 AM7


Second language acquisition 158Summary 162Some questions to think about 162Further reading 163

SECTION C: WORD RECOGNITION 165

6. Recognizing visual words 167Introduction 167Basic methods and basic findings 168Meaning-based facilitation of

visual word recognition 185Morphology: Processing

complex words 191Models of visual word recognition 192Coping with lexical ambiguity 199Summary 207Some questions to think about 208Further reading 208

7. Reading 209Introduction 209A preliminary model of reading 211The processes of normal reading 212The neuropsychology of adult

reading disorders: Acquireddyslexia 220

Models of word naming 227Summary 239Some questions to think about 240Further reading 240

8. Learning to read and spell 241Introduction 241Normal reading development 241Developmental dyslexia 249Summary 255Some questions to think about 255Further reading 256

9. Understanding speech 257Introduction 257Recognizing speech 257Models of speech recognition 267The neuropsychology of spoken word

recognition 281Summary 282

Some questions to think about 283Further reading 283

SECTION D: MEANING ANDUSING LANGUAGE 285

10. Understanding the structure ofsentences 287Introduction 287Dealing with structural ambiguity 289Early work on parsing 291Processing structural ambiguity 298Gaps, traces, and unbounded

dependencies 313The neuroscience of parsing 315Summary 319Some questions to think about 320Further reading 320

11. Word meaning 321Introduction 321Classic approaches to semantics 324Semantic networks 325Semantic features 328Family resemblance and

classification 335Combining concepts 338Processing figurative language 340The neuropsychology of semantics 342Connectionist approaches to

semantics 353Summary 359Some questions to think about 360Further reading 360

12. Comprehension 361Introduction 361Memory for text and inferences 363Reference, co-reference, and

ambiguity 373Models of text representation and

processing 378Individual differences in

comprehension skills 388The neuropsychology of text

and discourse processing 390Summary 392

viii CONTENTS

9781841693828_1_pre.PM5 9/21/07, 9:55 AM8


Some questions to think about 393Further reading 393

SECTION E: PRODUCTION ANDOTHER ASPECTS OF LANGUAGE 395

13. Language production 397Introduction 397Speech errors and what they tell us 399Syntactic planning 404Lexicalization 412Phonological encoding 428The analysis of hesitations 432The neuropsychology of speech

production 435Writing and agraphia 447Summary 448Some questions to think about 450Further reading 450

14. How do we use language? 453Introduction 453Making inferences in conversation 454The structure of conversation 458Collaboration in dialog 459Sound and vision 460Summary 462Some questions to think about 462Further reading 462

15. The structure of the languagesystem 463Introduction 463What are the modules of language? 464How many lexicons are there? 465Language and short-term memory 471Summary 476Some questions to think about 477Further reading 477

16. New directions 479Introduction 479Themes in psycholinguistics

revisited 479Some growth areas? 482Conclusion 483

Appendix: Connectionism 485Interactive activation models 485Back-propagation 487Further reading 489

Glossary 491Example of sentence analysis 500

References 501Author index 575Subject index 591

CONTENTS ix

9781841693828_1_pre.PM5 9/21/07, 9:55 AM9


9781841693828_1_pre.PM5 9/21/07, 9:55 AM10


I would like to thank Psychology Press for pro-viding the welcome opportunity for me to writethe third edition of my text. If back in 1994 whenI finished writing the first edition someone hadtold me that more than 10 years later I’d be writ-ing the third, I would have been extremely happy.

As I remarked in the preface to the first edi-tion, although language might not be all that makesus human, it is hard to imagine being human with-out it. Given the importance of language in ourbehavior, it is perhaps surprising that until not solong ago, relatively scant attention was been paidto it in undergraduate courses. Often at best it wasstudied as part of a general course on cognitivepsychology. That situation has changed. Further-more, the research field of psycholinguistics isblossoming, as evinced by the growth in thenumber of papers on the subject, and indeed, inthe number of journals dedicated to it. With thisgrowth and this level of interest, it is perhapssurprising that there are still relatively few text-books devoted to psycholinguistics. I hope thisbook fills this gap. It is aimed at intermediate andadvanced level undergraduates, although new post-graduates might also find it useful, and I wouldbe delighted if it found other readers.

I have tried to make as many of the citationsas possible refer towards easily obtainable ma-terial. I have therefore avoided citing unpublishedpapers, doctoral theses, and conference papers.New papers are coming out all the time, and if

I were going to make this book completely up todate, I would never stop. Therefore I called a haltat material published in early 2006. Of course,given current publication lags, much of this ma-terial would actually have been written some yearsbefore, and the current state of people’s thinkingand work, as discussed in conferences and semi-nars, might be very different from the positionsattributed in this book. This outcome is mostunfortunate, but unavoidable.

It is now impossible to appreciate psycholin-guistics without some understanding of connec-tionism. Unfortunately, this is a topic that manypeople find difficult. The formal details ofconnectionist models are in an Appendix: I hopethis does not mean that it will not be read. I toyedwith a structure where the technical details weregiven when the class of model was first intro-duced, but a more general treatment seemed moreappropriate.

I have been very gratified with the positivefeedback I have received on the first two editionsof this book, and the number of suggestions andcomments I have received. I have tried to takethese into account in this revision. “Taking intoaccount” does not mean “agreeing to everything”;after consideration, there are some suggestionsthat I decided against implementing. One of thesewas numbering sections, which I find estheticallydispleasing. I take the general point that thereare many cross-references, but I take this as a

Preface to the Third Edition

xi

9781841693828_1_pre.PM5 9/21/07, 9:55 AM11


positive point: we should try to foster as manyconnections between parts of the subject as pos-sible. I have also found that it is impossible toplease everybody. Indeed, people’s suggestions areoften contradictory. In such cases I can only rely onmy own inclinations. One example of this is thematerial on the Sapir-Whorf hypothesis: somepeople have suggested deleting it, whereas othershave found it to be one of the most useful sec-tions, and wanted more information on it. Almosteveryone wanted more material on their particularspecialty, and probably everyone is going to be dis-appointed. The book is quite long enough as it is.

It is difficult to believe that it is 6 years sinceI finished writing the second edition. Much hashappened in psychology since then—particularlythe widespread use of brain imaging. I have triedto incorporate this research where appropriate.

Whereas the main theme of the first editionwas interaction versus independence, and that ofthe second, broadly speaking, was connectionism,the main theme that has emerged in this thirdedition is that of the role of language in a widercontext. Are language processes specific to lan-guage, and do we have innate linguistic knowl-edge and dedicated regions of the brain doinglanguage processing? Or are they a reflectionof more general cognitive processes? In a waythis theme is a synthesis of modularity andconnectionism.

The structure of the book has settled down—whereas the second edition involved major sur-gery on the first, all I have done in the third editionis to split the lengthy chapter on reading into adultand child reading, and added a chapter about howwe use language.

Students often find the study of the psychol-ogy of language rather dry and technical, andmany find it difficult. In addition to making thisedition as comprehensible as possible, I have alsotried to make it fun. I have also tried to emphas-ize applications of research.

The American Psychological Associationnow recommends the use of the word “partici-pants” instead of “subjects.” I have followed thisrecommendation

There is a web site associated with thisbook. It contains links to other pages, details of

important recent work, and a “hot link” to contactme. It is to be found at: http://www.dundee.ac.uk/psychology/language. I still welcome any correc-tions, suggestions for the next edition, or discus-sion on any topic. My email address is now:[email protected]. Suggestions on topicsI have omitted or under-represented would beparticularly welcome. The hardest bit of writingthis book has been deciding what to leave out.I am sure that people running other courses willcover some material in much more detail than hasbeen possible to provide here. However, I wouldbe interested to hear of any major differences ofemphasis. If the new edition is as successful asthe second, I will be looking forward (in a strangesort of way) to producing the fourth edition in5 years time.

I would like to thank all those who havemade suggestions about both editions, particularlyJeanette Altarriba, Gerry Altmann, Elizabeth Bates,Helen Bown, Paul Bloom, Peer Broeder, Gor-don Brown, Hugh Buckingham, Lynne Duncan,the Dundee Psycholinguistics Discussion Group,Andy Ellis, Gerry Griffin, Zenzi Griffin, Annettede Groot, Francois Grosjean, Evan Heit, LaoragHunter, Lesley Jessiman, Barbara Kaup, AlanKennedy, Kathryn Kohnert, Annukka Lindell,Nick Lund, Siobhan MacAndrew, Nadine Mar-tin, Randi Martin, Elizabeth Maylor, Don Mitchell,Wayne Murray, Lyndsey Nickels, Jane Oakhill,Padraig O’Seaghdha, Shirley-Anne Paul, MartinPickering, Julian Pine, Ursula Pool, EleanorSaffran, Lynn Santelmann, Marcus Taft, JeremyTree, Roger van Gompel, Carel van Wijk, BethWilson, Alan Wilkes, Suzanne Zeedyk, and PienieZwitserlood. I would also like to thank severalanonymous reviewers for their comments; hope-fully you know who you are. Numerous peoplepointed out minor errors and asked questions:I thank them all. George Dunbar created the soundspectrogram for Figure 2.1 using MacSpeechLab.Lila Gleitman gave me the very first line; thanks!Katie Edwards, Pam Miller, and Denise Jacksonhelped me to obtain a great deal of material,often at very short notice. This book would bemuch worse without the help of all these people.I am of course responsible for any errors or omis-sions that remain. If there is any one else I have

xii PREFACE

9781841693828_1_pre.PM5 9/21/07, 9:55 AM12


forgotten, please accept my apologies. Manypeople have suggested things that I have thoughtabout and decided not to implement, and manypeople have suggested things (more connection-ism, less connectionism, leave that in, take thatout, move that bit there, leave it there) that are theopposite of what others have suggested.

I would also like to thank Psychology Pressfor all their help and enthusiasm for this project,particularly Lucy Kennedy, Tara Stebnicky, andMandy Collison. Holly Loftus made up the ancil-lary materials: Shirley-Anne Paul helped me tocheck them. Finally, I would like to thank BrianButterworth, who supervised my PhD. He prob-ably doesn’t realize how much I appreciated hishelp; without him, this book might never haveexisted.

Finally, perhaps I should state my bias aboutsome of the more controversial points of psy-cholinguistics: I think language processing is

massively interactive, I think connectionistmodeling has contributed enormously to our under-standing and is the most profitable direction togo in the near future, and I think that the study ofthe neuropsychology of language is fundamentalto our understanding. Writing this edition hasfostered and reinforced these beliefs. I realize thatmany will disagree with me, and have tried tobe as fair as possible. I hope that any bias there isin this book will appear to be the consequenceof the consideration of evidence rather than ofprejudice.

Trevor A. HarleySchool of PsychologyUniversity of Dundee

Dundee DD1 4HNScotland

March 2007

PREFACE xiii

9781841693828_1_pre.PM5 9/21/07, 9:55 AM13


9781841693828_1_pre.PM5 9/21/07, 9:55 AM14


How to Use This Book

This book is intended to be a stand-alone intro-duction to the psychology of language. It is myhope that anyone could pick it up and finish read-ing it with a rich understanding of how humansuse language. Nevertheless, it would probably beadvantageous to have some knowledge of basiccognitive psychology. (Some suggestions for booksto read are given in the “Further reading” sectionat the end of Chapter 1.) For example, you shouldbe aware that psychologists have distinguishedbetween short-term memory (which has limitedcapacity and can store material for only shortdurations) and long-term memory (which is vir-tually unlimited). I have tried to assume that thereader has no knowledge of linguistics, althoughI hope that most readers will be familiar withsuch concepts as nouns and verbs. The psycho-logy of language is quite a technical area full ofrather daunting terminology. I have italicized tech-nical terms when they first appear, and short defi-nitions are given in boxes on the pages concerned.There is also a glossary near the end of the bookwith short definitions of the technical terms.

Connectionist modeling is now central to mod-ern cognitive psychology. Unfortunately, it is alsoa topic that most people find extremely difficultto follow. It is impossible to understand the de-tails of connectionism without some mathemati-cal sophistication. I have provided an Appendixthat covers the basics of connectionism in moremathematical detail than is generally necessary to

understand the main text. However, the generalprinciples of connectionism can probably beappreciated without this extra depth, although itis probably a good idea to look at the Appendix.

In my opinion and experience, the material insome chapters is more difficult than others. I donot think there is anything much that can be doneabout this, except to persevere. Sometimes com-prehension might be assisted by later material,and sometimes a number of readings might benecessary to comprehend the material fully. For-tunately the study of the psychology of languagegives us clues about how to facilitate understand-ing. Chapters 7 and 11 will be particularly usefulin this respect. It should also be remembered thatin some areas researchers do not agree on theconclusions, or on what should be the appropriatemethod to investigate a problem. Therefore it issometimes difficult to say what the “right answer,”or the correct explanation of a phenomenon, mightbe. In this respect the psychology of language isstill a very young subject.

The book is divided into sections, each cover-ing an important aspect of language. Section A isan introduction. It describes what language is, andprovides essential background for describing lan-guage. It should not be skipped. Section B is aboutthe biological basis of language, the relationshipof language to other cognitive processes, and lan-guage development. Section C is about how we rec-ognize words. Section D is about comprehension:

xv

9781841693828_1_pre.PM5 9/21/07, 9:55 AM15


how we understand sentences and discourse. Sec-tion E is about language production, and alsoabout how language interacts with memory. It alsoexamines the grand design or architecture of thelanguage system. The section concludes with abrief look at some possible new directions in thepsychology of language.

Each chapter begins with an introduction out-lining what the chapter is about and the mainproblems faced in each area. Each introductionends with a summary of what you should knowby the end of the chapter. Each chapter concludeswith a list of bullet points that gives a one-sentence summary of each section in that chapter.This is followed by questions that you can thinkabout, either to test your understanding of thematerial, or to go beyond what is covered, usuallywith an emphasis on applying the material. If youwant to follow up a topic in more detail than iscovered in the text (which I think is quite richlyreferenced, and should be the first place to look),then there are suggestions for further reading atthe very end of each chapter.

One way of reading this book is like a novel:start here and go to the end. Section A shouldcertainly be read before the others because itintroduces many important terms, without whichlater going would be very difficult. However,after that, other orders are possible. I have tried tomake each chapter as self-contained as possible,so there is no reason why the chapters cannotbe read in a different order. Similarly, you mightchoose to omit some chapters altogether. In eachcase you might find you have to refer to theglossary more often than if you just begin at thebeginning. Unless you are interested in just a fewtopics, however, I advise reading the whole bookthrough at least once.

Each chapter looks at a major chunk of thestudy of the psychology of language.

OVERVIEW OF THIS BOOK

Chapter 1 tells you about the subject of thepsychology of language. It covers its history andmethods. Chapter 2 provides some important

background on language, telling you how we candescribe sounds and the structure of sentences. Inessence it is a primer on phonology and syntax.

Chapter 3 is about how language is related tobiological and cognitive processes. It looks atthe extent to which language depends on thepresence and operation of certain biological,cognitive, and social precursors in order to beable to develop normally. We will also examinewhether animals use language, or whether theycan be taught to do so. This will also help toclarify what we mean by language. We will lookat how language is founded in the brain, and howdamage to the brain can lead to distinct typesof impairment in language. We will examine indetail the more general role of language, byexamining the relation between language andthought. We will also look at what can be learnedfrom language acquisition in exceptional cir-cumstances, including the effects of linguisticdeprivation.

Chapter 4 examines how children acquirelanguage, and how language develops through-out childhood. Chapter 5 examines how bilingualchildren learn to use two languages.

We will then look at what appear to be thesimplest or lowest level processes and worktowards more complex ones. Hence we will firstexamine how we recognize and understand singlewords. Although these chapters are largely aboutrecognizing words in isolation, in the sense thatin most of the experiments we discuss only oneword is present at a time, the influence of thecontext in which they are found is an importantconsideration, and we will look at this as well.

Chapter 6 addresses how we recognize wordsand how we access their meanings. Although theemphasis is on visually presented word recogni-tion, many of the findings described in this chap-ter are applicable to recognizing spoken words aswell. Chapter 7 examines how we read and pro-nounce words, and looks at disorders of reading(the dyslexias). Chapter 8 looks at how we learnto read and spell, how reading should best betaught, and at the difficulties some children facein learning to read—developmental dyslexia.Chapter 9 looks at the speech system and how weprocess speech and identify spoken words.

xvi HOW TO USE THIS BOOK

9781841693828_1_pre.PM5 9/21/07, 9:55 AM16


HOW TO USE THIS BOOK xvii

We then move on to how words are ordered toform sentences. Chapter 10 looks at how we makeuse of word order information in understandingsentences. These are issues to do with syntax andparsing. Chapter 11 examines how we representthe meaning of words. Chapter 12 examines howwe comprehend and represent beyond the sentencelevel; these are the larger units of discourse or text.In particular, how do we integrate new informa-tion with old to create a coherent representation?How do we store what we have heard and read?

In Chapter 13 we consider the process inreverse, and examine language production andits disorders. By this stage we will have an

understanding of the processes involved in under-standing language, and these processes must belooked at in a wider context. Chapter 14 looksat how we use language. It examines how wecontrol conversation, and how speakers and lis-teners cooperate to make dialogue efficient. It alsoexamines how listeners draw inferences aboutwhat the speaker means from apparent violationsof the normal expectation of how to run a goodconversation.

In Chapter 15 we look at the structure of thelanguage system as a whole, and the relationbetween the parts. Finally, Chapter 16 looks atsome possible new directions in psycholinguistics.

9781841693828_1_pre.PM5 9/21/07, 9:55 AM17


9781841693828_1_pre.PM5 9/21/07, 9:55 AM18


7. READING 209

209

7

Reading

• Appreciate how different types of dyslexiarelate to the dual-route model, and also theproblems they pose for it.

• Know about connectionist models of readingand how they account for dyslexia.

The writing system

The basic unit of written language is the letter.The name grapheme is given to the letter or com-bination of letters that represents a phoneme. Forexample, the word “ghost” contains five lettersand four graphemes (“gh,” “o,” “s,” and “t”),representing four phonemes. There is much morevariability in the structure of written languagesthan there is in spoken languages. Whereas allspoken languages utilize a basic distinctionbetween consonants and vowels, there is no suchcommon thread to the world’s written languages.

INTRODUCTION

In Chapter 6 we looked at how we recognizewords; this chapter is about how we read them.How do we gain access to the sounds and mean-ings of words? We also examine the effects ofbrain damage on reading (giving rise to acquireddyslexia), and show how reading disorders can berelated to a model of reading. The next chapterlooks at how children learn to read.

Reading aloud and reading to oneself areclearly different, but related, tasks. When we readaloud (or name words), we must retrieve the soundsof words. When we read to ourselves, we read toobtain the meaning, but most of us, most of thetime, experience the sounds of the words as “innerspeech.” Is it possible to go to the meaning of aword when reading without also accessing itssounds? By the end of this chapter you should:

• Know how different languages translate wordsinto sounds, and understand the alphabeticprinciple.

• Understand the motivation for the dual-routemodel of reading, and know about its strengthsand weaknesses.

KEY TERM

Grapheme: a unit of written language that correspondsto a phoneme (e.g., “steak” contains four graphemes:s t ea k).

9781841693828_4_007.PM5 9/21/07, 10:03 AM209


210 THE PSYCHOLOGY OF LANGUAGE

Types of written languages

Examples

EnglishOther European languages

HebrewArabic

CherokeeJapanese kana

ChineseJapanese kanji

Alphabetic script

Consonantal script

Syllabic script

Logographic /ideographic script

Features

The basic unit represented by a graphemeis essentially a phoneme.

Not all sounds are represented, asvowels are not written down.

Written units represent syllables.

Each symbol represents a whole word.

The sorts of written language most familiar tospeakers of English and other European languagesare alphabetic scripts. English uses an alphabeticscript. In alphabetic scripts, the basic unit rep-resented by a grapheme is essentially a phoneme.However, the nature of this correspondence canvary. In transparent languages such as Serbo-Croat and Italian there is a one-to-one grapheme–phoneme correspondence, so that every graphemeis realized by only one phoneme and everyphoneme is realized by only one grapheme. Inlanguages such as English this relation can beone-to-many in both directions. A phoneme canbe realized by different graphemes (e.g., compare“to,” “too,” “two,” and “threw”), and a graphemecan be realized by many different phonemes(e.g., the letter “a” in the words “fate,” “pat,” and“father”). Some languages lie between theseextremes. In French, correspondences betweengraphemes and phonemes are quite regular, but aphoneme may have different graphemic realiza-tions (e.g., the graphemes “o,” “au,” “eau,” “aux,”and “eaux” all represent the same sounds). Inconsonantal scripts, such as Hebrew and Arabic,not all sounds are represented, as vowels are notwritten down at all. In syllabic scripts (such asCherokee and the Japanese script kana), thewritten units represent syllables. Finally, some lan-guages do not represent any sounds. In ideographiclanguages (sometimes also called logographic lan-guages), such as Chinese and the Japanese scriptkanji, each symbol is equivalent to a morpheme.

One consequence of this variation in writingsystems is that there must be differences in pro-cessing between readers of different languages.

There is much more variability in the structure of writtenlanguages than there is in spoken languages. Inconsonantal scripts, such as Hebrew (above) and Arabic,not all sounds are represented.

Hence this chapter should be read with the cau-tion in mind that some conclusions may be trueof English and many other writing systems, butnot necessarily all of them.

Unlike speech, reading and writing are arelatively recent development. Writing emergedindependently in Sumer and Mesoamerica, andperhaps also in Egypt and China. The first writingsystem was the cuneiform script printed on clayin Sumer, which appeared just before 3000 bc.The emergence of the alphabetic script can betraced to ancient Greece in about 1000 bc. Thedevelopment of the one-to-many correspondencein English orthography primarily arose betweenthe fifteenth and eighteenth centuries as a conse-quence of the development of the printing pressand the activities of spelling “reformers” who triedto make the Latin and Greek origins of wordsmore apparent in their spellings (see Ellis, 1993,for more detail). Therefore it is perhaps not sur-prising that reading is actually quite a complex

9781841693828_4_007.PM5 9/21/07, 10:03 AM210


7. READING 211

cognitive task. There is a wide variation in readingabilities, and many different types of readingdisorder arise as a consequence of brain damage.

A PRELIMINARY MODEL OF READING

Introspection can provide us with a preliminarymodel of reading. Consider how we might nameor pronounce the word “beef.” Words like this aresaid to have a regular spelling-to-sound corres-pondence. That is, the graphemes map onto pho-nemes in a totally regular way; you need no specialknowledge about the word to know how to pro-nounce it. If you had never seen the word “beef ”before, you could still pronounce it correctly. Someother examples of regular word pronunciationsinclude “hint” and “rave.” In these words, thereare alternative pronunciations (as in “pint” and“have”), but “hint” and “rave” are pronounced inaccordance with the most common pronunciations.These are all regular words, because all the graph-emes have the standard pronunciation.

Not all words are regular, however. Some areirregular or exception words. Consider the word“steak.” This has an irregular spelling-to-sound(or grapheme-to-phoneme) correspondence: thegrapheme “ea” is not pronounced in the usualway, as in “streak,” “sneak,” “speak,” “leak,” and“beak.” Other exceptions to a rule include “have”(an exception to the rule that leads to the regularpronunciations “gave,” “rave,” “save” and so forth)and “vase” (in British English, an exception tothe rule that leads to the regular pronunciations“base,” “case,” and so forth). English has manyirregular words. Some words are extremely irregu-lar, containing unusual patterns of letters thathave no close neighbors, such as “island,” “aisle,”“ghost,” and “yacht.” These words are sometimescalled lexical hermits.

Finally, we can pronounce strings of letterssuch as “nate,” “smeak,” “fot,” and “datch,” eventhough we have never seen them before. Theseletter strings are all pronounceable nonwords orpseudowords. Therefore, even though they arenovel, we can still pronounce them, and we alltend to agree on how they should be pronounced.If you hear nonwords like these, you can spell

them correctly; you assemble their pronunciationsfrom their constituent graphemes. (Of course, notall nonwords are pronounceable—e.g., “xzhgh.”)

Our ability to read nonwords on the one handand irregular words on the other suggests the pos-sibility of a dual-route model of naming. We canassemble pronunciations for words or nonwordswe have never seen before, yet also pronouncecorrectly irregular words that must need informa-tion specific to those words (that is, lexical informa-tion). The classic dual-route model (see Figure 7.1)has two routes for turning words into sounds.There is a direct access or lexical route, whichis needed for irregular words. This must at leastin some way involve a direct link between printand sound. That is, the lexical route takes usdirectly to a word’s entry in the lexicon and weare then able to retrieve the sound of a word.There is also a grapheme-to-phoneme conversion(GPC) route (also called the indirect or non-lexical or sublexical route), which is used forreading nonwords. This route carries out what iscalled phonological recoding. It does not involvelexical access at all. The non-lexical route wasfirst proposed in the early 1970s (e.g., Gough,1972; Rubenstein, Lewis, & Rubenstein, 1971).Another important justification for a grapheme-to-phoneme conversion route is that it is usefulfor children learning to read by sounding outwords letter by letter.

KEY TERM

Sublexical: correspondences in spelling and soundbeneath the level of the whole word.

FIGURE 7.1

Print

Lexicon

Pronunciation

Grapheme–phonemeconversion rules

The simplified version of the dual-route model of reading.

9781841693828_4_007.PM5 9/21/07, 10:03 AM211



Given that neither route can in itself adequatelyexplain reading performance, it seems that wemust use both. Modern dual-route theorists seereading as a “race” between these routes. Whenwe see a word, both routes start processing it. Forskilled readers, most of the time the direct routeis much faster, so it will usually win the race andthe word will be pronounced the way that itrecommends. The indirect route will only be appar-ent in exceptional circumstances, such as whenwe see a very unfamiliar word; in that case, if thedirect route is slower than normal, then the directand GPC routes will produce different pronuncia-tions at about the same time, and these wordsmight be harder to pronounce.

Relation of the dual-route model toother models

In the previous chapter we examined a number ofmodels of word recognition. These can all be seenas theories of how the direct, lexical access read-ing route operates. The dual route is the simplestversion of a range of possible multi-route or par-allel coding models, some of which posit morethan two reading routes. Do we really need a non-lexical route at all for routine reading? Althoughwe appear to need it for reading nonwords, itseems a costly procedure. We have a mechanismready to use for something we rarely do—pronouncing new words or nonwords. Perhaps itis left over from the development of reading, orperhaps it is not as costly as it first appears. Wewill see later that the non-lexical route is alsoapparently needed to account for the neuropsy-chological data. Indeed, whether or not two routesare necessary for reading is a central issue of thetopic of reading. Models that propose that we canget away with only one (such as connectionistmodels) must produce a satisfactory account ofhow we can pronounce nonwords.

Of course, except for reading aloud, the pri-mary goal of reading is not getting the sound of aword, but getting the meaning. As we shall see inChapter 8, in the early stages of learning to readchildren get to the meaning through the sound;that is, they spell out the sound of the words, andthen access meaning as they recognize thosesounds. Some researchers believe that even skilled

adults primarily get to meaning by going fromprint to phonology and then to meaning, an ideacalled phonological mediation (discussed in moredetail below). Most researchers however believethat in skilled adults, most of the time, there is adirect route from print to semantics. Indeed, aswe shall see below, most researchers believe thatthere is a direct route from print to sound, and adirect route via semantics; what is debated is therole of the indirect route in normal reading (seeTaft & van Graan, 1998, for further discussion ofthese issues).

THE PROCESSES OF NORMAL READING

According to the dual-route model, there are twoindependent routes when naming a word andaccessing the lexicon: a lexical or direct accessroute and a sublexical or grapheme–phonemeconversion route. This section looks at how wename nonwords and words.

Reading nonwords

According to the dual-route model, the pronun-ciation of all nonwords should be assembled usingthe GPC route. This means that all pronounceablenonwords should be alike and their similarity towords should not matter. However, pronounce-able nonwords are not all alike.

The pseudohomophone effect

Pseudohomophones are pronounceable non-words that sound like words when pronounced(such as “brane,” which sounds like the word“brain” when spoken). The behavior of the pseudo-homophone “brane” can be compared with the verysimilar nonword “brame,” which does not soundlike a word when it is spoken. Rubenstein et al.

KEY TERM

Pseudohomophones: a nonword that sounds like aword when pronounced (e.g., “nite”).

9781841693828_4_007.PM5 9/21/07, 10:03 AM212


7. READING 213

(1971) showed that pseudohomophones are moreconfusable with words than other types of non-words are. Participants are faster to name them,but slower to reject them as nonwords than con-trol nonwords.

Is the effect caused by the phonological orvisual similarity between the nonword and word?Martin (1982) and Taft (1982) argued that it isvisual similarity that is important. Pseudohomo-phones are more confusable with words than othernonwords are because they look more similar towords than non-pseudohomophones do, ratherthan because they sound the same. Pring (1981)alternated the case of letters within versus acrossgraphemes, such as the “AI” in “grait,” to produce“GraIT” or “GRaiT.” These strings look differentbut still sound the same. Alternating letter caseswithin a grapheme or spelling unit (aI) eliminatesthe pseudohomophone effect; alternating letterselsewhere in the word (aiT) does not. Hence weare sensitive to the visual appearance of spellingunits of words.

The pseudohomophone effect suggests thatnot all nonwords are processed in the same way.The importance of the visual appearance of thenonwords further suggests that something elseapart from phonological recoding is involved here.It remains to be seen whether the phonologicalrecoding route is still necessary, but if it is, thenit must be more complex than we first thought.

Glushko’s (1979) experiment: Lexicaleffects on nonword reading

Glushko (1979) performed a very important experi-ment on the effect of the regularity of the word-neighbors of a nonword on its pronunciation.Consider the nonword “taze.” Its word-neighborsinclude “gaze,” “laze,” and “maze”; these are allthemselves regularly pronounced words. Now con-sider the word-neighbors of the nonword “tave.”These also include plenty of regular words (e.g.,“rave,” “save,” and “gave”) but there is an excep-tion word-neighbor (“have”). As another example,compare the nonwords “feal” and “fead”: bothhave regular neighbours (e.g., “real,” “seal,” “deal,”and “bead”) but the pronunciation of “fead” isinfluenced by its irregular neighbour “dead.”Glushko (1979) showed that naming latencies to

nonwords such as “tave” were significantly slowerthan to ones such as “taze.” That is, reaction timesto nonwords that have orthographically irregularspelling-to-sound correspondence word-neighborsare slower than to other nonword controls. Also,people make pronunciation “errors” with suchnonwords: “pove” might be pronounced to rhymewith “love” rather than “cove”; and “heaf ” mightbe pronounced to rhyme with “deaf ” rather than“leaf.” In summary, Glushko found that the pro-nunciation of nonwords is affected by the pro-nunciation of similar words, and that nonwordsare not the same as each other. Subsequent re-search has shown that the proportion of regularpronunciations of nonwords increases as thenumber of orthographic neighbours increases(McCann & Besner, 1987). In summary, there arelexical effects on nonword processing.

More on reading nonwords

The nonword “yead” can be pronounced to rhymewith “bead” or “head.” Kay and Marcel (1981)showed that its pronunciation can be affected bythe pronunciation of a preceding prime word:“bead” biases a participant to pronounce “yead”to rhyme with it, whereas the prime “head”biases participants to the alternative pronuncia-tion. Rosson (1983) primed the nonword by asemantic relative of a phonologically related word.The task was to pronounce “louch” when pre-ceded either by “feel” (which is associated with“touch”) or by “sofa” (which is associated with“couch”). In both cases “louch” tended to be pro-nounced to rhyme with the appropriate relative.

Finally, nonword effects in complex experi-ments are sensitive to many factors, such as thepronunciation of the surrounding words in the list.This also suggests that nonword pronunciationinvolves more than just grapheme-to-phonemeconversion.

Evaluation of research on readingnonwords

These data do not fit the simple version of thedual-route model. The pronunciation of nonwordsis affected by the pronunciation of visually similarwords. That is, there are lexical effects in nonword

9781841693828_4_007.PM5 9/21/07, 10:03 AM213



processing; the lexical route seems to be affect-ing the non-lexical route.

Word processing

According to the dual-route model, words areaccessed directly by the direct route. This meansthat all words should be treated the same in respectof the regularity of their spelling-to-sound corres-pondences. An examination of the data reveals thatthis prediction does not stand up.

One problem for the simple dual-route modelis that pronunciation regularity affects responsetimes, although in a complex way. Baron andStrawson (1976) provided an early demonstrationof this problem, finding that a list of regular wordswas named faster than a list of frequency-matchedexception words (e.g., “have”). This task is a sim-plified version of the naming task, with responsetime averaged across many items rather than takenfrom each one individually. There have been manyother demonstrations of the influence of regularityon naming time (e.g., Forster & Chambers, 1973;Frederiksen & Kroll, 1976; Stanovich & Bauer,1978). A well-replicated finding is that of aninteraction between regularity and frequency:regularity has little effect on the pronunciationof high-frequency words, but low-frequency regu-lar words are named faster than low-frequencyirregular words (e.g., Andrews, 1982; Seidenberg,Waters, Barnes, & Tanenhaus, 1984), even whenwe control for age-of-acquisition (Monaghan& Ellis, 2002). Jared (1997b) found that high-frequency words can be sensitive to regularity,but the effect of regularity is moderated by thenumber and frequencies of their “friends” and“enemies” (words with similar or conflicting pro-nunciations). That is, it is important to controlfor the neighborhood characteristics of the targetwords as well as their regularity in order toobserve the interaction. On the other hand, it isnot clear whether there are regularity effects onlexical decision. They have been obtained by, forexample, Stanovich and Bauer (1978), but not byColtheart et al. (1977), or Seidenberg et al. (1984).In particular, a word such as “yacht” looks un-usual, as well as having an irregular pronunciation.The letter pairs “ya” and “ht” are not frequent in

English; we say they have a low bigram frequency.Obviously the visual appearance of words is goingto affect the time it takes for direct access, so weneed to control for this when searching for regu-larity effects. Once we control for the generallyunusual appearance of irregular words, regularityand consistency only seem to affect naming times,not lexical decision times. Age-of-acquisition hasa similar effect to frequency, and gives rise to asimilar interaction: Consistency has a much big-ger impact on naming time for late-acquired thanearly-acquired words (Monaghan & Ellis, 2002).Why do late-acquired and low-frequency incon-sistent words stand out? One possibility is thatlate-acquired low-frequency consistent words canmake use of the network structure of other con-sistent words; inconsistent items cannot, and neednew associations to be learned between input andoutput (Monaghan & Ellis, 2002).

In general, regularity effects are more likelyto be found when participants have to be moreconservative, such as when accuracy rather thanspeed is emphasized. The finding that regularityaffects naming might appear problematic for thedual-route model, but makes sense if there is arace between the direct and indirect routes. Re-member that there is an interaction between regu-larity and frequency. The pronunciation of commonwords is directly retrieved before the indirect routecan construct any conflicting pronunciation. Con-flict arises when the lexical route is slow, as whenretrieving low-frequency words, and when the pro-nunciation of a low-frequency word generated bythe lexical route conflicts with that generated bythe non-lexical route (Norris & Brown, 1985).

Glushko’s (1979) experiment: Resultsfrom words

Glushko (1979) also found that words behave ina similar way to nonwords, in that the namingtimes of words are affected by the phonologicalconsistency of neighbors. The naming of a regularword is slowed down relative to that of a controlword of similar frequency if the test word hasirregular neighbors. For example, the word “gang”is regular, and all its neighbors (such as “bang,”“sang,” “hang,” and “rang”) are also regular.

9781841693828_4_007.PM5 9/21/07, 10:03 AM214


7. READING 215

KEY TERM

Body: the same as a rime: the final vowel and terminalconsonants.

TABLE 7.1

Classification of word pronunciations depending on regularity and consistency (based on Patterson &Morton, 1985)

Word type Example Characteristics

Consistent gaze All words receive the same regular pronunciation of the bodyConsensus lint All words with one exception receive the same regular pronunciationHeretic pint The irregular exception to the consensusGang look All words with one exception receive the same irregular pronunciationHero spook The regular exception to the gangGang without a hero cold All words receive the same irregular pronunciationAmbiguous: conformist cove Regular pronunciation with many irregular exemplarsAmbiguous: independent love Irregular pronunciation with many regular exemplarsHermit yacht No other word has this body

Consider on the other hand “base”; this itself hasa regular pronunciation (compare it with “case”),but it is inconsistent, in that it has one irregularneighbour, “vase” (in British English pronuncia-tion). We could say that “vase” is an enemy of“base.” This leads to a slowing of naming times.In addition, Glushko found true naming errors ofover-regularization: for example “pint” was some-times given its regular pronunciation—to rhymewith “dint.”

Pronunciation neighborhoods

Continuing this line of research, Brown (1987)argued that the number of consistently pronouncedneighbors ( friends) determines naming times,rather than whether a word has enemies (that is,whether or not it is regular). It is now thought thatthe number of both friends and enemies affectsnaming times (Brown & Watson, 1994; Jared,McRae, & Seidenberg, 1990; Kay & Bishop, 1987).

Andrews (1989) found effects of neighborhoodsize in both the naming and the lexical decisiontasks. Responses to words with large neighbor-hoods were faster than words with small neigh-borhoods (although this may be moderated byfrequency, as suggested by Grainger, 1990). Notall readers produce the same results. Barron (1981)found that good and poor elementary school read-ers both read regular words more quickly thanirregular words. However, once he controlled forneighborhood effects, he found that there was nolonger any regularity effect in the good readers,although it persisted in the poor readers.

Parkin (1982) found more of a continuumof ease-of-pronunciation than a simple divisionbetween regular and irregular words. All this worksuggests that a binary division into words withregular and irregular pronunciations is no longeradequate. Patterson and Morton (1985) provideda more satisfactory but complex categorizationrather than a straightforward dichotomy betweenregular and irregular words (see Table 7.1).This classification reflects two factors: first, theregularity of the pronunciation with reference tospelling-to-sound correspondence rules; second,the agreement with other words that share the samebody. (This is the end of a monosyllabic word,comprising the central vowel plus final consonantor consonant cluster; e.g., “aint” in “saint” or “us”in “plus.”) We need to consider not only whethera word is regular or irregular, but also whether itsneighbors are regular or irregular. The same clas-sification scheme can be applied to nonwords.

In summary, just as not all nonwords behavein the same way, neither do all words. The regu-larity of pronunciation of a word affects the easewith which we can name it. In addition, thepronunciation of a word’s neighbors can affectits naming. The number of friends and enemiesaffects how easy it is to name a word.

9781841693828_4_007.PM5 9/21/07, 10:03 AM215



The role of sound in accessing meaning:Phonological mediation

There is some experimental evidence suggestingthat a word’s sound may have some influence onaccessing the meaning (Frost, 1998; van Orden,1987; van Orden, Johnstone, & Hale, 1988; vanOrden, Pennington, & Stone, 1990). In a categorydecision task, participants have to decide if avisually presented target word is a member of aparticular category. For example, given “A typeof fruit” you would respond “yes” to “pear,” and“no” to “pour.” If the “no” word is a homophoneof a “yes” word (e.g., “pair”), participants make alot of false positive errors—that is, they respond“yes” instead of “no.” Participants seem confusedby the sound of the word, and category decisionclearly involves accessing the meaning. The effectis most noticeable when participants have to re-spond quickly. Lesch and Pollatsek (1998) foundevidence of interference between homophonesin a semantic relatedness task (e.g., SAND–BEECH). We take longer to respond to homo-phones in a lexical decision task (e.g., MAID),presumably because the homophones are generat-ing confusion in lexical access, perhaps throughfeedback from phonology to orthography (Pexman,Lupker, & Jared, 2001; Pexman, Lupker, &Reggin, 2002).

Hence there is considerable evidence that therecognition of a word can be influenced by itsphonology. The dominant view is that this influ-ence arises through the indirect route, althoughword recognition is primarily driven by the directroute (or routes)—a view that has been labeledthe weak phonological perspective (Coltheart,Rastle, Perry, Langdon, & Ziegler, 2001; Rastle& Brysbaert, 2006). Most of the models describedin this chapter subscribe to the weak phonologicalview. The alternative, strong phonological view,that we primarily get to the meaning throughsound, is called phonological mediation. The mostextreme form of this idea is that visual wordrecognition cannot occur in the absence of com-puting the sound of the word.

There is a great deal of controversy about thestatus of phonological mediation. Other experi-ments support the idea. Folk (1999) examined

eye movements as participants read sentencescontaining either “soul” or “sole.” Folk foundthat the homophones were read with longer gazeduration—that is, they were processed as thoughthey were lexically ambiguous, even though theorthography should have prevented this. This resultis only explicable if the phonology is in someway interfering with the semantic access.

On the other hand, Jared and Seidenberg (1991)showed that prior phonological access onlyhappens with low-frequency homophones. In anexamination of proof-reading and eye movements,Jared, Levy, and Rayner (1999) also found thatphonology only plays a role in accessing the mean-ings of low-frequency words. In addition, theyfound that poor readers are more likely to have toaccess phonology in order to access semantics,whereas good readers primarily activate semanticsfirst. Daneman, Reingold, and Davidson (1995)reported eye fixation data on homophones thatsuggested the meaning of a word is accessed firstwhereas the phonological code is accessed later,probably post-access. They found that gaze dura-tion times were longer on an incorrect homophone(e.g., “brake” was in the text when the contextdemanded “break”), and that the fixation timeson the incorrect homophone were about the sameas on a spelling control (e.g., “broke”). This meansthat the appropriate meaning must have beenactivated before the decision to move the eyes, andthat the phonological code is not activated at thistime. (If the phonological code had been accessedbefore meaning then the incorrect homophonewould sound all right in the context, and gazedurations should have been about the same.) Thephonological code is accessed later, however, andinfluences the number of regressions (when theeyes look back to earlier material) to the target word.(However, see Rayner, Pollatsek, & Binder, 1998,for different conclusions. It is clear that these ex-periments are very sensitive to the materials used.)

Taft and van Graan (1998) used a semanticcategorization task to examine phonologicalmediation. Participants had to decide whether ornot words belonged to a category of “words withdefinable meanings” (e.g., “plank,” “pint”) or thecategory of “given names” (e.g., “Pam,” “Phil”).There was no difference in the decision times

9781841693828_4_007.PM5 9/21/07, 10:03 AM216


7. READING 217

between regular definable words (e.g., “plank”)and irregular definable words (e.g., “pint”), al-though a regularity effect was shown in a wordnaming task. This suggests that the sound of aword does not need to be accessed on the route toaccessing its meaning.

A number of studies have tried to decidebetween the strong and weak phonological viewsusing masked phonological priming. In this tech-nique, targets (e.g., “clip”) are preceded byphonologically identical nonword primes (e.g.,“klip”). Responses to the targets are faster andmore accurate than when the target is precededby an unrelated word. Several studies have foundpriming effects occur even when the primes havebeen masked and presented so briefly that theycannot be consciously observed and reported,suggesting that the phonological stimulus mustoccur automatically and extremely quickly (e.g.,Lukatela & Turvey, 1994a, 1994b; Perfetti, Bell,& Delaney, 1988). While some researchers inter-pret masked phonological priming as supportingphonological mediation—Why else should earlyphonological activation happen so early unless itis essential?—other researchers point out that theseeffects are very sensitive to environmental con-ditions, and are not always reliably found (seeRastle & Brysbaert, 2006, for a review). In a meta-analysis of the literature, Rastle and Brysbaert(2006) do find small but significant maskedphonological priming effects.

These data suggest that the sound of a word isusually accessed at an early stage. However, thereis much evidence suggesting that phonologicalrecoding cannot be obligatory in order to accessthe word’s meaning (Ellis, 1993). For example,some dyslexics cannot pronounce nonwords, yetcan still read many words. Hanley and McDonnell(1997) described the case of a patient, PS, whounderstood the meaning of words in readingwithout being able to pronounce them correctly.Critically, PS did not have a preserved innerphonological code that could be used to accessthe meaning. Some patients have preserved innerphonology and preserved reading comprehension,but make errors in speaking aloud (Caplan &Waters, 1995b). Hanley and McDonnell arguedthat PS did not have access to his phonological

code because he was unable to access both mean-ings of a homophone from seeing just one in print.Thus PS could not produce the phonological formsof words aloud correctly, and did not have accessto an internal phonological representation of thosewords, yet he could still understand them whenreading them. For example, he could give perfectdefinitions of printed words. In general, a reviewof the neuropsychological literature suggeststhat people can recognize words in the absenceof phonology (Coltheart, 2004). Hence it is un-likely that phonological recoding is an obligatorycomponent of visual word recognition (Rastle &Brysbaert, 2006).

How then can we explain the data showingphonological mediation? There are a number ofalternative explanations. First, although phono-logical recoding prior to accessing meaning maynot be obligatory, it might occur in some circum-stances. Given there is a race between the lexicaland sublexical routes in the dual-route model, iffor some reason the lexical route is slow in pro-ducing an output, the sublexical route might havetime to assemble a conflicting phonological rep-resentation. Second, there might be feedback fromthe speech production system to the semantic sys-tem, or the direct-access route causes inner speechthat interferes with processing. Third, it is pos-sible that lexical decision is based on phonologicalinformation (Rastle & Brysbaert, 2006).

Silent reading and inner speech

Although it seems unlikely that we have to accesssound before meaning, we do routinely seem toaccess some sort of phonological code after ac-cessing meaning in silent reading. Subjective evi-dence for this is the experience of “inner speech”while reading. Tongue-twisters such as (1) takelonger to read silently than sentences where thereis variation in the initial consonants (Haber &Haber, 1982). This suggests that we are accessingsome sort of phonological code as we read.

(1) Boris burned the brown bread badly.

However, this inner speech cannot involve ex-actly the same processes as overt speech because

9781841693828_4_007.PM5 9/21/07, 10:04 AM217



the easiest material (Rayner & Pollatsek, 1989).McCutchen and Perfetti (1982) argued that which-ever route is used for lexical access in reading, atleast part of the phonological code of each wordis automatically accessed—in particular we accessthe sounds of beginnings of words. Although thereis some debate about the precise nature of thephonological code and how much of it is acti-vated, it does seem that silent reading necessarilygenerates some sort of phonological code (Rayner& Pollatsek, 1989). This information is used toassist comprehension, primarily by maintainingitems in sequence in working memory.

The role of meaning in accessing sound

Phonological mediation means that we might ac-cess meaning via sound. Sometimes need to accessthe meaning before we can access a word’s sound.Words such as “bow,” “row,” and “tear,” have twodifferent pronunciations. This type of word is calleda homograph. How do we select the appropriatepronunciation? Consider sentences (2) and (3):

(2) When his shoelace came loose, Vladhad to tie a bow.

(3) At the end of the play, Dirk went to thefront of the stage to take a bow.

Clearly here we need to access the word’smeaning before we can select the appropriatepronunciation. Further evidence that semantics canaffect reading is provided by a study by Strain,Patterson, and Seidenberg (1995). They showedthat there is an effect of imageability on skilledreading such that there is a three-way interactionbetween frequency, imageability, and spelling con-sistency. People are particularly slow and makemore errors when reading low-frequency excep-tion words with abstract meanings (e.g., “scarce”).Although a subsequent study by Monaghan andEllis (2002), suggests that this semantic effectmight be at least in part the result of a confoundwith age-of-acquisition, as abstract low-frequencyexception words tend to have late AOA, this inter-action is still found when we control for AOA(Strain, Patterson, & Seidenberg, 2002). Hence,at least some of the time, we need to access a

The experience of “inner speech” while readingdemonstrates that we can access some sort ofphonological code after accessing meaning in silentreading.

we can read silently much faster than we can readaloud (Rayner & Pollatsek, 1989), and becauseovert articulation does not prohibit inner speechwhile reading. Furthermore, although most peoplewho are profoundly deaf read very poorly, someread quite well (Conrad, 1972). Although thismight suggest that eventual phonological codingis optional, it is likely that these deaf able readersare converting printed words into some sign lan-guage code (Rayner & Pollatsek, 1989). Evidencefor this is that deaf people are troubled by thesilent reading of word strings that correspondto hand-twisters (Treiman & Hirsh-Pasek, 1983).(Interestingly, deaf people also have some diffi-culty with signing phonological tongue-twisters,suggesting that difficulty can arise from lip-reading sounds.)

Hence, when we read we seem to access aphonological code that we experience as innerspeech. That is, when we gain access to a word’srepresentation in the lexicon, all its attributesbecome available. The activation of a phonologi-cal code is not confined to alphabetic languages.On-line experimental data using priming andsemantic judgment tasks suggest that phono-logical information about ideographs is auto-matically activated in both Chinese (Perfetti &Zhang, 1991, 1995) and Japanese kanji (Wydell,Patterson, & Humphreys, 1993).

Inner speech seems to assist comprehension;if it is reduced, comprehension suffers for all but

9781841693828_4_007.PM5 9/21/07, 10:04 AM218


7. READING 219

word’s semantic representation before we canaccess its phonology.

Does speed reading work?

Occasionally you might notice advertisements inthe press for techniques for improving your read-ing speed. The most famous of these techniquesis known as “speed reading.” Proponents of speedreading claim that you can increase your readingspeed from the average of 200–350 words a minuteto 2000 words a minute or even faster, yet retainthe same level of comprehension. Is this possible?Unfortunately, the preponderance of psychologicalresearch suggests not. As you increase your read-ing speed above the normal rate, comprehensiondeclines. Just and Carpenter (1987) compared theunderstanding of speed readers and normal readerson an easy piece of text (an article from Reader’sDigest) and a difficult piece of text (an articlefrom Scientific American). They found that normalreaders scored 15% higher on comprehensionmeasures than the speed readers across bothpassages. In fact, the speed readers performedonly slightly better than a group of people whoskimmed through the passages. The speed readersdid as well as the normal readers on the generalgist of the text, but were worse at details. Inparticular, speed readers could not answer ques-tions when the answers were located in placeswhere their eyes had not fixated.

Speed reading, then, is not as effective asnormal reading. Eye movements are the key towhy speed reading confers limited advantages(Rayner & Pollatsek, 1989). For a word to beprocessed properly, its image has to land close tothe fovea and stay there for a sufficient length oftime. Speed reading is nothing more than skim-ming through a piece of writing (Carver, 1972).This is not to say that readers obtain nothing fromskimming: if you have sufficient prior informa-tion about the material, your level of comprehen-sion can be quite good. If you speed read andthen read normally, your overall level of compre-hension and retention might be better than if youhad just read the text normally. It is also a usefultechnique for preparing to read a book or articlein a structured way (see Chapter 12). Finally,

associated techniques such as relaxing before youstart to read might well have beneficial effects oncomprehension and retention.

Evaluation of experiments on normal reading

There are two major problems with a simple dual-route model. First, we have seen that there arelexical effects on reading nonwords, which shouldbe read by a non-lexical route that is insensitiveto lexical information. Second, there are effectsof regularity of pronunciation on reading words,which should be read by a direct, lexical routethat is insensitive to phonological recoding.

A race model fares better. Regularity effectsarise when the direct and indirect routes producean output at about the same time, so that conflictarises between the irregular pronunciation pro-posed by the lexical route and the regular pronun-ciation proposed by the sublexical route. However,it is not clear how a race model where the indirectroute uses grapheme–phoneme conversion canexplain lexical effects on reading nonwords.Neither is it clear how semantics can guide theoperation of the direct route.

Skilled readers have a measure of attentionalor strategic control over the lexical and sub-lexical routes such that they can attend selec-tively to lexical or sublexical information (Baluch& Besner, 1991; Monsell, Patterson, Graham,Hughes, & Milroy, 1992; Zevin & Balota, 2000).For example, Monsell et al. found that thecomposition of word lists affected naming per-formance. High-frequency exception words werepronounced faster when they were in pure blocksthan when they were mixed with nonwords.Monsell et al. argued that this was becauseparticipants allocated more attention to lexicalinformation when reading the pure blocks. Par-ticipants also made fewer regularization errorswhen the words were presented in pure blocks(when they can rely solely on lexical processing)than in mixed blocks (when the sublexical routehas to be involved).

At first sight, then, this experiment suggeststhat in difficult circumstances people seem ableto change their emphasis in reading from usinglexical information to sublexical information.

9781841693828_4_007.PM5 9/21/07, 10:04 AM219



However, Jared (1997a) argued that people neednot change the extent to which they rely onsublexical information, but instead might beresponding at different points in the processing ofthe stimuli. She argued that the faster pronuncia-tion latencies found in Monsell et al.’s experimentin the exception-only condition could just be theresult of a general increase in response speed,rather than a reduction in reliance on the non-lexical route.

However, there is further evidence for strat-egic effects in the choice of route when reading.Using a primed naming task, Zevin and Balota(2000) found that nonword primes produce agreater dependence on sublexical processing, butlow-frequency exception word primes produce agreater dependence on lexical processing. Coltheartand Rastle (1994) suggested that lexical access isperformed so quickly for high-frequency wordsthat there is little scope for sublexical involve-ment, but with low-frequency words or in diffi-cult conditions people can devote more attentionto one route or the other.

THE NEUROPSYCHOLOGY OF ADULTREADING DISORDERS: ACQUIREDDYSLEXIA

What can studies of people with brain damagetell us about reading? This section is concernedwith disorders of processing written language.We must distinguish between acquired disorders(which, as a result of head trauma such as stroke,operation, or head injury, lead to disruption ofprocesses that were functioning normally before-hand) and developmental disorders (which do notresult from obvious trauma, and which disrupt thedevelopment of a particular function). Disordersof reading are called the dyslexias; disorders ofwriting are called the dysgraphias. Damage tothe left hemisphere will generally result in dyslexia,but as the same sites are involved in speaking,dyslexia is often accompanied by impairments tospoken language processing.

We can distinguish central dyslexias, whichinvolve central, high-level reading processes, from

peripheral dyslexias, which involve lower levelprocesses. Peripheral dyslexias include visual dys-lexia, attentional dyslexia, letter-by-letter reading,and neglect dyslexia, all of which disrupt theextraction of visual information from the page.As our focus is on understanding the central read-ing process, we will limit discussion here to thecentral dyslexias. In addition, we will only lookat acquired disorders in this section, and deferdiscussion of developmental dyslexia until ourexamination of learning to read.

If the dual-route model of reading is correct,then we should expect to find a double dissocia-tion of the two reading routes. That is, we shouldfind some patients have damage to the lexical routebut can still read by the non-lexical route only,whereas we should be able to find other patientswho have damage to the non-lexical route but canread by the lexical route only. The existence ofa double dissociation is a strong prediction ofthe dual-route model, and a real challenge to anysingle-route model.

Surface dyslexia

People with surface dyslexia have a selectiveimpairment in the ability to read irregular (excep-tion) words. Hence they would have difficulty with“steak” compared with a similar regular relativeword such as “speak.” Marshall and Newcombe(1973) and Shallice and Warrington (1980) de-scribed some early case histories. Surface dyslexicsoften make over-regularization errors when try-ing to read irregular words aloud. For example,they pronounce “broad” as “brode,” “steak” as

KEY TERMS

Acquired disorders: a disorder caused by braindamage is acquired if it affects an ability that waspreviously intact (contrasted with developmentaldisorder).Developmental disorders: a disorder where thenormal development or acquisition of a process (e.g.,reading) is affected.Dyslexias: disorder of reading.Surface dyslexia: a type of dyslexia where the personhas difficulty with exception words.

9781841693828_4_007.PM5 9/21/07, 10:04 AM220


7. READING 221

“steek,” and “island” as “eyesland.” On the otherhand, their ability to read regular words and non-words is intact. In terms of the dual-route model,the most obvious explanation of surface dyslexiais that these patients can only read via the indi-rect, non-lexical route: that is, it is an impairmentof the lexical (direct access) processing route. Thecomprehension of word meaning is intact in thesepatients. They still know what an “island” is, evenif they cannot read the word, and they can stillunderstand it if you say the word to them.

The effects of brain damage are rarely local-ized to highly specific systems, and, in practice,patients do not show such clear-cut behavior asthe ideal of totally preserved regular word andnonword reading, and the total loss of irregularwords. The clearest case yet reported is that of apatient referred to as MP (Bub, Cancelliere, &Kertesz, 1985). She showed completely normalaccuracy in reading nonwords, and hence hernon-lexical route was totally preserved. She wasnot the best possible case of surface dyslexia,however, because she could read some irregularwords (with an accuracy of 85% on high-frequency items, and 40% on low-frequencyexception words). This means that her lexical routemust have been partially intact. The pure casesare rarely found. Other patients show consider-ably less clear-cut reading than this, with evenbetter performance on irregular words, and somedeficit in reading regular words.

If patients were reading through a non-lexicalroute, we would not expect lexical variables toaffect the likelihood of reading success. Kremin(1985) found no effect of word frequency, part ofspeech (noun versus adjective versus verb), orwhether or not it is easy to form a mental imageof what is referred to (called imageability), onthe likelihood of reading success. Althoughpatients such as MP, from Bub et al. (1985), showa clear frequency effect in that they make fewregularizations of high-frequency words, otherpatients, such as HTR, from Shallice, Warrington,and McCarthy (1983) do not. Patients also makehomophone confusions (such as reading “pane”as “to cause distress”).

Surface dyslexia may not be a unitary category.Shallice and McCarthy (1985) distinguished

between Type I and Type II surface dyslexia.Patients of both types are poor at reading excep-tion words. The more pure cases, known as TypeI patients, are highly accurate at naming regularwords and pseudowords. Other patients, knownas Type II, also show some impairment at readingregular words and pseudowords. The readingperformance of Type II patients may be affectedby lexical variables such that they are better atreading high-frequency, high-imageability words,better at reading nouns than adjectives and atreading adjectives than verbs, and better at read-ing short words than long. Type II patients musthave an additional, moderate impairment to thenon-lexical route, but the dual-route model cannevertheless still explain this pattern.

Phonological dyslexia

People with phonological dyslexia have a selec-tive impairment in the ability to read pronounce-able nonwords, called pseudowords (such as“sleeb”), while their ability to read matched words(e.g., “sleep”) is preserved. Phonological dyslexiawas first described by Shallice and Warrington(1975, 1980), Patterson (1980), and Beauvois andDerouesné (1979). Phonological dyslexics findirregular words no harder to read than regularones. These symptoms suggest that these patientscan only read using the lexical route, and there-fore that phonological dyslexia is an impairmentof the non-lexical (GPC) processing route. As withsurface dyslexia, the “perfect patient,” who inthis case would be able to read all words but nononwords, has yet to be discovered. The clearestcase yet reported is that of patient WB (Funnell,1983), who could not read nonwords at all;hence the non-lexical GPC route must have been

KEY TERMS

Imageability: a semantic variable concerning how easyit is to form a mental image of a word: “rose” is moreimageable than “truth.”Phonological dyslexia: a type of dyslexia wherepeople can read words quite well but are poor atreading nonwords.

9781841693828_4_007.PM5 9/21/07, 10:04 AM221



completely abolished. He was not the mostextreme case possible of phonological dyslexia,however, because there was also an impairmentto his lexical route; his performance was about85% correct on words.

For those patients who can pronounce somenonwords, nonword reading is improved if thenonwords are pseudohomophones (such as “nite”for “night,” or “brane” for “brain”). Those patientswho also have difficulty in reading words haveparticular difficulty in reading the function wordsthat do the grammatical work of the language.Low-frequency, low-imageability words are alsopoorly read, although neither frequency norimageability seems to have any overwhelming rolein itself. These patients also have difficulty in read-ing morphologically complex words—those thathave syntactic modifications called inflections.They sometimes make what are called deriva-tional errors on these words, where they read aword as a grammatical relative of the target, suchas reading “performing” as “performance.” Finally,they also make visual errors, in which a word isread as another with a similar visual appearance,such as reading “perform” as “perfume.”

There are different types of phonological dys-lexia. Derouesné and Beauvois (1979) suggestedthat phonological dyslexia can result from dis-ruption of either orthographic or phonologicalprocessing. Some patients are worse at readinggraphemically complex nonwords (e.g., CAU,where a phoneme is represented by two letters;hence this nonword requires more graphemicparsing) than graphemically simple nonwords(e.g., IKO, where there is a one-to-one mappingbetween letters and graphemes), but show noadvantage for pseudohomophones. These patientssuffer from a disruption of graphemic parsing.Another group of patients are better at readingpseudohomophones than non-pseudohomophones,but show no effect of orthographic complexity.These patients suffer from a disruption of phono-logical processing. Friedman (1995) distinguishedbetween phonological dyslexia arising from animpairment of orthographic-to-phonological pro-cessing (characterized by relatively poor functionword reading but good nonword repetition) fromthat arising from an impairment of general

phonological processing (characterized by thereverse pattern).

Following this, a three-stage model of sublexicalprocessing has emerged (Beauvois & Derouesné,1979; Coltheart, 1985; Friedman, 1995). First, agraphemic analysis stage parses the letter stringinto graphemes. Second, a print-to-sound conver-sion stage assigns phonemes to graphemes. Third,in the phonemic blending stage the sounds areassembled into a phonological representation.There are patients whose behavior can best beexplained in terms of disruption of each of thesestages (Lesch & Martin, 1998). MS (Newcombe& Marshall, 1985) suffered from disruption tographemic analysis. Patients with disrupted gra-phemic analysis find nonwords in which eachgrapheme is represented by a single letter easierto read than nonwords with multiple correspond-ences. WB (Funnell, 1983) suffered from dis-ruption in the print-to-sound conversion stage;here nonword repetition is intact. ML (Lesch &Martin, 1998) was a phonological dyslexic whocould carry out tasks of phonological assemblyon syllables, but not on sub-syllabic units (onsets,bodies, and phonemes). MV (Bub, Black, Howell,& Kertesz, 1987) suffered from disruption to thephonemic stage.

Why do some people with phonological dys-lexia have difficulty reading function words? Onepossibility is that function words are difficultbecause they are so abstract (Friedman, 1995).However, patient MC (Druks & Froud, 2002) hadgreat difficulty in reading nonwords, morpho-logically complex words, and function words inisolation. Crucially he could read highly abstractcontent words, so it cannot be the abstractnessof the function words that caused his problems.Nevertheless, he could understand the meaningof function words that he could not read, and hisdeficit was confined to reading single words. Hisreading of function words in continuous textwas much better. It is likely that MC at least hasa problem with syntactic processing such thatwhen producing words in isolation he is unable toaccess syntactic information.

People with phonological dyslexia showcomplex phonological problems that have noth-ing to do with orthography. Indeed, it has been

9781841693828_4_007.PM5 9/21/07, 10:04 AM222


7. READING 223

proposed that phonological dyslexia is a conse-quence of a general problem with phonologicalprocessing (Farah, Stowe, & Levinson, 1996; Harm& Seidenberg, 2001; Patterson, Suzuki, & Wydel,1996). If phonological dyslexia arises solelybecause of problems with ability to translateorthography into phonology, then there must bebrain tissue dedicated to this task. This impliesthat this brain tissue becomes dedicated by school-age learning, which is an unappealing prospect.The alternative view is that phonological dyslexiais just one aspect of a general impairment ofphonological processing. This impairment willnormally be manifested in performance on non-reading tasks such as rhyming, nonword writing,phonological short-term memory, nonword rep-etition, and tasks of phonological synthesis (“whatdoes “c – a – t spell out?”) and phonologicalawareness (“what word is left if you take the “p”sound out of “spoon”?). This proposal alsoexplains why pseudohomophones are read betterthan non-pseudohomophones. An important pieceof evidence in favor of this hypothesis is thatphonological dyslexia is never observed in theabsence of a more general phonological deficit(but see Coltheart, 1996, for a dissenting view).A general phonological deficit makes it difficultto assemble pronunciations for nonwords. Wordsare spared much of this difficulty because ofsupport from other words and top-down supportfrom their semantic representations. Repeatingwords and nonwords is facilitated by support fromauditory representations, so some phonologicaldyslexics can still repeat some nonwords. How-ever, if the repetition task is made more difficultso that patients can no longer gain support fromthe auditory representations, repetition perform-ance declines markedly (Farah et al., 1996). Thisidea that phonological dyslexia is caused by ageneral phonological deficit is central to theconnectionist account of dyslexia, discussed later.

Deep dyslexia

At first sight, surface and phonological dyslexiaappear to exhaust the possibilities of the conse-quences of damage to the dual-route model. Thereis, however, another even more surprising type

of dyslexia called deep dyslexia. Marshall andNewcombe (1966, 1973) first described deepdyslexia in two patients, GR and KU, although itis now recognized that the syndrome had beenobserved in patients before this (Marshall &Newcombe, 1980). In many respects deep dys-lexia resembles phonological dyslexia. Patientshave great difficulty in reading nonwords, andconsiderable difficulty in reading the grammati-cal, function words. Like phonological dyslexics,they make visual and derivational errors. How-ever, the defining characteristic of deep dyslexiais the presence of semantic reading errors orsemantic paralexias, when people produce a wordrelated in meaning to the target instead of thetarget, as in examples (4) to (7):

(4) DAUGHTER “sister”(5) PRAY “chapel”(6) ROSE “flower”(7) KILL “hate”

The imageability of a word is an importantdeterminant of the probability of reading successin deep dyslexia. The easier it is to form a mentalimage of a word, the easier it is to read. Note thatjust an imageability effect in reading does notmean that patients with deep dyslexia are betterat all tasks involving more concrete words. Indeed,Newton and Barry (1997) described a patient (LW)who was much better at reading high-frequencyconcrete words than abstract words, but whoshowed no impairment in comprehending thosesame abstract words.

Coltheart (1980) listed 12 symptoms commonlyshown by deep dyslexics: They make semanticerrors, they make visual errors, they substituteincorrect function words for the target, theymake derivational errors, they can’t pronouncenonwords, they show an imageability effect, they

KEY TERMS

Deep dyslexia: disorder of reading characterized bysemantic reading errors.Semantic paralexias: a reading error based on aword’s meaning.

9781841693828_4_007.PM5 9/21/07, 10:04 AM223



find nouns easier to read than adjectives, theyfind adjectives easier to read than verbs, they findfunction words more difficult to read than contentwords, their writing is impaired, their auditoryshort-term memory is impaired, and their readingability depends on the context of a word (e.g.,FLY is easier to read when it is a noun in a sen-tence than a verb).

There has been some debate about theextent to which deep dyslexia is a syndrome (asyndrome is a group of symptoms that clustertogether). Coltheart (1980) argued that the clus-tering of symptoms is meaningful, in that theysuggest a single underlying cause. However,although these symptoms tend to occur in manypatients, they do not apparently necessarily do so.For example, AR (Warrington & Shallice, 1979)did not show concreteness and content wordeffects and had intact writing and auditory short-term memory. A few patients make semantic errorsbut very few visual errors (Caramazza & Hillis,1990). Such patients suggest that it is unlikelythat there is a single underlying deficit. Like phono-logical dyslexics, deep dyslexics obviously havesome difficulty in obtaining non-lexical access tophonology via grapheme–phoneme recoding, butthey also have some disorder of the semanticsystem. We nevertheless have to explain why thesesymptoms are so often associated. One pos-sibility is that the different symptoms of deepdyslexia arise because of an arbitrary featureof brain anatomy: Different but nearby parts ofthe brain control processes such as writing andauditory short-term memory, so that damage toone is often associated with damage to another.As we will see, a more satisfying account is pro-vided by connectionist modeling.

Shallice (1988) argued that there are threesubtypes of deep dyslexia that vary in the preciseimpairments involved. Input deep dyslexics havedifficulties in reaching the exact semantic rep-resentations of words in reading. In these patients,auditory comprehension is superior to reading.Central deep dyslexics have a severe auditorycomprehension deficit in addition to their readingdifficulties. Output deep dyslexics can processwords up to their semantic representations, butthen have difficulty producing the appropriate

phonological output. In practice it can be difficultto assign particular patients to these subtypes, andit is not clear what precise impairment of the read-ing systems is necessary to produce each subtype(Newton & Barry, 1997).

The right-hemisphere hypothesis

Does deep dyslexia reflect attempts by a greatlydamaged system to read normally, as has beenargued by Morton and Patterson (1980), amongothers? Or does it instead reflect the operation ofan otherwise normally suppressed system comingthrough? Perhaps deep dyslexics do not alwaysuse the left hemisphere for reading. Instead, peoplewith deep dyslexia might use a reading systembased in the right hemisphere that is normallysuppressed (Coltheart, 1980; Saffran, Bogyo,Schwartz, & Marin, 1980; Zaidel & Peters, 1981).This right-hemisphere hypothesis is supportedby the observation that the more of the left hemi-sphere that is damaged, the more severe the deepdyslexia observed (Jones & Martin, 1985; but seeMarshall & Patterson, 1985). Furthermore, thereading performance of deep dyslexics resemblesthat of split-brain patients when words are pre-sented to the left visual field, and therefore to theright hemisphere. Under such conditions they alsomake semantic paralexias, and have an advantagefor concrete words. Finally, Patterson, Vargha-Khadem, and Polkey (1989) described the case ofa patient called NI, a 17-year-old girl who had hadher left hemisphere removed for the treatment ofsevere epilepsy. After recovery she retained somereading ability, but her performance resembledthat of deep dyslexics.

In spite of these points in its favor, the right-hemisphere reading hypothesis has never wonwide acceptance. In part this is because the hypo-thesis is considered a negative one, in that if itwere correct, deep dyslexia would tell us nothingabout normal reading. In addition, people withdeep dyslexia read much better than split-brainpatients who are forced to rely on the right hemi-sphere for reading. The right-hemisphere advan-tage for concrete words is rarely found, and theimageability of the target words used in theseexperiments might have been confounded with

9781841693828_4_007.PM5 9/21/07, 10:04 AM224


7. READING 225

length (Ellis & Young, 1988; Patterson & Besner,1984). Finally, Roeltgen (1987) described a pa-tient who suffered from deep dyslexia as a resultof a stroke in the left hemisphere. He later suf-fered from a second left hemisphere stroke, whichhad the effect of destroying his residual readingability. If the deep dyslexia had been a conse-quence of right hemisphere reading, it should nothave been affected by the second stroke in theleft hemisphere.

Summary of research on deep dyslexia

There has been debate as to whether the term“deep dyslexia” is a meaningful label. The cru-cial issue is whether or not its symptoms mustnecessarily co-occur because they have the sameunderlying cause. Are semantic paralexias alwaysfound associated with impaired nonword reading?So far they seem to be; in all reported casessemantic paralexias have been associated with allthe other symptoms. How then can deep dyslexiabe explained by one underlying disorder? In termsof the dual-route model, there would need to bedamage to both the semantic system (to explainthe semantic paralexias and the imageabilityeffects) and the non-lexical route (to explainthe difficulties with nonwords). We would alsothen have to specify that for some reason damageto the first is always associated with damageto the second (e.g., because of an anatomicalaccident that the neural tissue supporting bothprocesses is in adjoining parts of the brain).This is inelegant. As we shall see, connectionistmodels have cast valuable light on this question.A second issue is whether we can make infer-ences from deep dyslexia about the processes ofnormal reading, as we can for the other typesof acquired dyslexia. We have seen that thedual-route model readily explains surface andphonological dyslexia, and that their occurrenceis as expected if we were to lesion that model byremoving one of the routes. Hence it is reason-able to make inferences about normal reading onthe basis of data from such patients. There is somedoubt, however, as to whether we are entitledto do this in the case of deep dyslexia; if theright-hemisphere hypothesis were correct, deep

dyslexia would tell us little about normal read-ing. The balance of evidence is at present thatdeep dyslexia does not reflect right-hemispherereading, but does reflect reading by a greatlydamaged left hemisphere. Deep dyslexia suggeststhat normally we can in some way read throughmeaning; that is, we use the semantic represen-tation of a word to obtain its phonology. Thissupports our earlier observation that with homo-graphs (e.g., “bow”) we use the meaning to selectthe appropriate pronunciation.

Non-semantic reading

Schwartz, Marin, and Saffran (1979), andSchwartz, Saffran, and Marin (1980a) describedWLP, an elderly patient suffering from progress-ive dementia. WLP had a greatly impaired abil-ity to retrieve the meaning of written words; forexample, she was unable to match written animalnames to pictures. She could read those wordsout aloud almost perfectly, getting 18 out of 20correct and making only minor errors, even onlow-frequency words. She could also read irregu-lar words and nonwords. In summary, WLP couldread words without any comprehension of theirmeaning. Coslett (1991) described a patient, WT,who was virtually unable to read nonwords, sug-gesting an impairment of the indirect route ofthe dual-route model, but who was able to readirregular words quite proficiently, even thoughshe could not understand those words. These casestudies suggest that we must have a direct accessroute from orthography to phonology that doesnot go through semantics.

Summary of the interpretation of theacquired dyslexias

We have looked at four main types of adult cen-tral dyslexia: surface, phonological, deep, and non-semantic reading. We have seen how a dual-routemodel explains surface dyslexia as an impairmentof the lexical, direct access route, and explainsphonological dyslexia as an impairment of thenon-lexical, phonological recoding route. Theexistence of non-semantic reading suggest thatthe simple dual-route model needs refinement. In

9781841693828_4_007.PM5 9/21/07, 10:04 AM225



Acquired dyslexia in other languages

Languages such as Italian, Spanish, or Serbo-Croat,which have totally transparent or shallow alpha-betic orthographies—that is, where every graphemeis in a one-to-one relation with a phoneme—canshow phonological and deep dyslexia, but notsurface dyslexia, defined as an inability to readexception words (Patterson, Marshall, & Coltheart,1985a, 1985b). However, we can find the symp-toms that can co-occur with an impairment ofexception word reading, such as homophoneconfusions, in the languages that permit them(Masterson, Coltheart, & Meara, 1985).

Whereas languages such as English have asingle, alphabetic script, Japanese has two differ-ent scripts, kana and kanji (see Coltheart, 1980;Sasanuma, 1980). Kana is a syllabic script, andkanji is a logographic or ideographic script. There-fore words in kanji convey no information on howa word should be pronounced. While kana allowssublexical processing, kanji must be accessedthrough a direct, lexical route. The right hemi-sphere is better at dealing with kanji, and the lefthemisphere is better at reading kana (Coltheart,1980). Reading of briefly presented kana wordsis more accurate when they are presented to theright visual field (left hemisphere), but readingof kanji words is better when they are presentedto the left visual field (right hemisphere). Theanalog of surface dyslexia is found in patientswhere there is a selective impairment of readingkanji, but the reading of kana is preserved. Theanalog of phonological dyslexia is an ability toread both kana and kanji, but a difficulty in read-ing Japanese nonwords. The analog of deep dys-lexia is a selective impairment of reading kana,while the reading of kanji is preserved. For exam-ple, patient TY could read words in both kanjiand kana almost perfectly, but she had great diffi-culty with nonwords constructed from kana words(Sasanuma, Ito, Patterson, & Ito, 1996).

Chinese is an ideographic language. Butter-worth and Wengang (1991) reported evidence oftwo routes in reading in Chinese. Ideographs canbe read aloud either through a route that associ-ates the symbol with its complete pronunciation,or through one that uses parts of the symbol.

FIGURE 7.2

When words are misread, arethe errors usually confined to

one half of the word?

No

Yes

Are semantic errors madein reading aloud?

Are regular words readaloud much better than

exception words?

Is naming a letter much harder when it isaccompanied by other, irrelevant letters?

Analyzing acquired dyslexia (adapted fromColtheart, 1981)

Are words often readletter by letter?

Is reading aloud ofnonwords very bad

or impossible?

Yes

Yes

Yes

No

No

No

Yes

Attentionaldyslexia

Deep dyslexia

Phonologicaldyslexia

Yes

Surface dyslexia

Letter-by-letterreading

Neglect orpositional dyslexia

No

particular, the direct route must be split into two.There must be a non-semantic direct access routethat retrieves phonology given orthography, butwhich does not pass through semantics first, anda semantic direct access route that passes throughsemantics and allows us to select the appropriatesounds of non-homophonic homographs (e.g.,“wind”). In non-semantic reading, the semanticdirect route has been abolished but the non-semantic direct route is intact. An analysis ofacquired dyslexia by Coltheart (1981) is shownin Figure 7.2.

9781841693828_4_007.PM5 9/21/07, 10:04 AM226


7. READING 227

First, there are lexical effects for nonwords andregularity effects for words, and therefore readingcannot be a simple case of automatic grapheme-to-phoneme conversion for nonwords, and auto-matic direct access for all words. Single-routemodels, on the other hand, appear to provide noaccount of nonword pronunciation, and it remainsto be demonstrated how neighborhood effectsaffect a word’s pronunciation. Second, any modelmust also be able to account for the pattern ofdissociations found in dyslexia. While surface andphonological dyslexia indicate that two readingmechanisms are necessary, other disorders sug-gest that these alone will not suffice. At first sightit is not obvious how a single-route model couldexplain these dissociations at all.

Theorists have taken two different approachesdepending on their starting point. One possibilityis to refine the dual-route model. Another is toshow how word-neighborhoods can affect pronun-ciation, and how pseudowords can be pronouncedin a single-route model. This led to the develop-ment of analogy models. More recently, a con-nectionist model of reading has been developedthat takes the single-route, analogy-based approachto the limit.

The revised dual-route model

We can save the dual-route model by makingit more complex. Morton and Patterson (1980)and Patterson and Morton (1985) described athree-route model (see Figure 7.3). First, there isa non-lexical route for assembling pronunciationsfrom sublexical grapheme–phoneme conversion.The non-lexical route now consists of two sub-systems. A standard grapheme–phoneme conver-sion mechanism is supplemented with a bodysubsystem that makes use of information aboutcorrespondences between orthographic and pho-nological rimes. This is needed to explain lexicaleffects on nonword pronunciation. Second, thedirect route is split into a semantic and a non-semantic direct route.

The three-route model accounts for the data asfollows. The lexical effects on nonwords and regu-larity effects on words are explained by cross-talkbetween the lexical and non-lexical routes. Two

Kanji (shown here) is a logographic or ideographic script,providing no information on word pronunciation.

(Although Chinese is non-alphabetic, most sym-bols contain some sublexical information on pro-nunciation.) Each route can be selectively impairedby brain damage, leading to distinct types of read-ing disorder.

The study of other languages that have dif-ferent means of mapping orthography ontophonology is still at a relatively early stage, but itis likely to greatly enhance our understanding ofreading mechanisms. The findings suggest thatthe neuropsychological mechanisms involved inreading are universal, although there are obvi-ously some differences related to the uniquefeatures of different orthographies.

MODELS OF WORD NAMING

Both the classic dual-route and the single-route,lexical-instance models face a number of problems.

9781841693828_4_007.PM5 9/21/07, 10:04 AM227



FIGURE 7.3

Graphemes

Lexicon(visual input logogens)

Phonology(speech output logogens)

Semantic system

Printed language

Speech

Non-semanticreading

Sublexical recoding(graphemes and bodies)

The revised dual-route modelof reading.

types of interaction are possible: interferenceduring retrieval, and conflict in resolving multiplephonological forms after retrieval. The two sub-systems of the non-lexical route also give themodel greater power. Surface dyslexia is the lossof the ability to make direct contact with theorthographic lexicon, and phonological dyslexia isthe loss of the indirect route. Non-semantic read-ing is a loss of the lexical-semantic route. Deepdyslexia remains rather mysterious. First, we haveto argue that these patients can only read throughthe lexical-semantic route. While accounting forthe symptoms that resemble phonological dyslexia,it still does not explain the semantic paralexias.One possibility is that this route is used normally,but not always successfully, and that it needsadditional information (such as from the non-lexical and non-semantic direct route) to succeed.So when this information is no longer availableit functions imperfectly. It gets us to the rightsemantic area, but not necessarily to the exact item,hence giving paralexias. This additional assump-tion seems somewhat arbitrary. An alternativeidea is that paralexias are the result of additionaldamage to the semantic system itself. Hence acomplex pattern of impairments is still necessary

to explain deep dyslexia, and there is no reasonto suggest that these are not dissociable.

Multi-route models are becoming increasinglycomplicated as we find out more about the read-ing process (for example, see Carr & Pollatsek,1985). Another idea is that multiple levels ofspelling-to-sound correspondences combine indetermining the pronunciation of a word. InNorris’s (1994a) multiple-levels model, differentlevels of spelling-to-sound information, includingphoneme, rime (the final part of the word givingrise to the words with which it rhymes, e.g., “eak”in “speak”), and word-level correspondences, com-bine in an interactive activation network to deter-mine the final pronunciation of a word. Such anapproach develops earlier models that make useof knowledge at multiple levels, such as Brown(1987), Patterson and Morton (1985), and Shallice,Warrington, and McCarthy (1983).

The most recent version of the dual-routemodel is the dual-route cascaded, or DRC, model(Coltheart, Curtis, Atkins, & Haller, 1993;Coltheart & Rastle, 1994; Coltheart, Rastle, Perry,Langdon, & Ziegler, 2001). This is a computa-tional model based on the architecture of the dual-route model—although it is in fact misleadingly

9781841693828_4_007.PM5 9/21/07, 10:04 AM228


7. READING 229

so called, as it is really based on the three-routemodel, with a non-lexical grapheme–phoneme rulesystem and a lexical system, which in turn isdivided into one route that passes through thesemantic system and a non-semantic route thatdoes not. The model makes use of cascaded pro-cessing, in that as soon as there is any activationat the letter level, activation is passed on to theword level. The computational model can simulateperformance on both lexical decision and namingtasks, showing appropriate effects of frequency,regularity, pseudohomophones, neighborhood, andpriming. Regularity is now a central motivationof the model; words are either regular, or they arenot. Irregular words take longer to pronounce thanregular ones because the lexical and non-lexicalroutes produce conflicting pronunciations. Themodel accounts for surface dyslexia by makingentries in the orthographic lexicon less available,and for phonological dyslexia by damaging thegrapheme–phoneme conversion route.

There is not uniform agreement that it isnecessary to divide the direct route into two. Inthe summation model (Hillis & Caramazza, 1991b;Howard & Franklin, 1988), the only direct routeis reading through semantics. How does this modelaccount for non-semantic reading? The idea isthat access to the semantic system is not com-pletely obliterated. Activation from the sublexicalroute combines (or is “summated”) with activationtrickling down from the damaged direct semanticroute to ensure the correct pronunciation.

It is difficult to distinguish between these vari-ants of the original dual-route model, althoughthe three-route version provides the more explicitaccount of the dissociations observed in dyslexia.There is also some evidence against the summa-tion hypothesis. EP (Funnell, 1996) could readirregular words that she could not name, and pri-ming the name with the initial letter did not helpher naming, contrary to the prediction of the sum-mation hypothesis. Many aspects of the dual-routemodel have been subsumed by the triangle modelthat serves as the basis of connectionist modelsof reading. The situation is complicated even moreby the apparent co-occurrence of the loss of par-ticular word meanings in dementia and surfacedyslexia (see later).

The analogy model

The analogy model arose in the late 1970s whenthe extent of lexical effects on nonword readingand differences between words became apparent(Glushko, 1979; Henderson, 1982; Kay & Marcel,1981; Marcel, 1980). It is a form of single-routemodel that provides an explicit mechanism forhow we pronounce nonwords. It proposes that wepronounce nonwords and new words by analogywith other words. When a word (or nonword) ispresented, it activates its neighbors, and these allinfluence its pronunciation. For example, “gang”activates “hang,” “rang,” “sang,” and “bang”; theseare all consistent with the regular pronuncia-tion of “gang,” and hence assembling a pro-nunciation is straightforward. When presentedwith “base,” however, “case” and “vase” areactivated; these conflict and hence the assemblyof a pronunciation is slowed down until theconflict is resolved. A nonword such as “taze” ispronounced by analogy with the consistent setof similar words (“maze,” “gaze,” “daze”). Anonword such as “mave” activates “gave,”“rave,” and “save,” but it also activates the con-flicting enemy “have,” which hence slows downpronunciation of “mave.” In order to name byanalogy, you have to find candidate words con-taining appropriate orthographic segments (like“-ave”); obtain the phonological representation ofthe segments; and assemble the complete phono-logy (“m + ave”).

Although attractive in the way they deal withregularity and neighborhood effects, early versionsof analogy models suffered from a number ofproblems. First, the models did not make clearhow the input is segmented in an appropriate way.Second, the models make incorrect predictionsabout how some nonwords should be pronounced.Particularly troublesome are nonwords based ongangs; “pook” should be pronounced by analogywith the great preponderance of the gang com-prising “book,” “hook,” “look,” and “rook,” yet itis given the “hero” pronunciation (see Table 7.1)—which is in accordance with grapheme–phonemecorrespondence rules—nearly 75% of the time(Kay, 1985). Analogy theory also appears to makeincorrect predictions about how long it takes us

9781841693828_4_007.PM5 9/21/07, 10:04 AM229



to make regularization errors (Patterson & Morton,1985). Finally, it is not clear how analogy modelsaccount for the dissociations found in acquireddyslexia. Nevertheless, in some ways the analogymodel was a precursor of connectionist models ofreading.

Connectionist models

The original Seidenberg and McClelland (1989)model evolved in response to criticisms that I willexamine after describing the original model.

Seidenberg and McClelland (1989)’s modelof reading

The Seidenberg and McClelland (1989) model(often abbreviated to SM) shares many featureswith the interactive activation model of letter rec-ognition discussed in Chapter 6. The SM modelprovides an account of how readers recognizeletter strings as words and pronounce them. Thisfirst model simulated one route of a more generalmodel of lexical processing (see Figure 7.4).

Reading and speech involve three types of code:orthographic, meaning, and phonological. Theseare connected with feedback connections. Theshape of the model has given it the name ofthe Triangle Model. As in the revised dual-routemodel, there is a route from orthography tophonology by way of semantics. The key featureof the model is that there is only one other routefrom orthography to phonology; there is no routeinvolving grapheme–phoneme correspondencerules.

Seidenberg and McClelland (1989) just simu-lated the orthographic-to-phonology part of theoverall triangle model. The model has threelevels, each containing many simple units. Theseare the input, hidden, and output layers (seeFigure 7.5). Each of the units in these layers hasan activation level, and each unit is connectedto all the units in the next level by a weightedconnection, which can be either excitatory orinhibitory. An important characteristic of thistype of model is that the weights on these con-nections are not set by the modelers, but arelearned. This network learns to associate aphonological output with an orthographic inputby being given repeated exposure to word-pronunciation pairs. It learns using an algorithmcalled back-propagation. This involves slowlyreducing the discrepancy between the desired andactual outputs of the network by changing theweights on the connections. (See the Appendixfor more information.)

Seidenberg and McClelland used 400 unitsto code orthographic information for input and460 units to code phonological information foroutput, mediated by 200 hidden units. Phonemesand graphemes were encoded as a set of triples,so that each grapheme or phoneme was specifiedwith its flanking grapheme or phoneme. This is

Seidenberg and McClelland’s (1989) “triangle model” ofword recognition. Implemented pathways are shown inbold. Reproduced with permission from Harm andSeidenberg (2001).

FIGURE 7.4

Context

Meaning

Orthography Phonology

MAKE /mAk/ KEY TERM

Back-propagation: an algorithm for learning input–output pairs in connectionist networks. It works byalternately reducing the error between the actualoutput and the desired output of the network.

9781841693828_4_007.PM5 9/21/07, 10:04 AM230


7. READING 231

needed to minimize the differences between thedesired and actual outputs.

After training, the network was tested by pre-senting letter strings and computing the ortho-graphic and phonological error scores. The errorscore is a measure of the average differencebetween the actual and desired output of each ofthe output units, across all patterns. Phonologicalerror scores were generated by applying inputto the orthographic units, and measured by theoutput of the phonological units; they were inter-preted as reflecting performance on a naming task.Orthographic error scores were generated bycomparing the pattern of activation input to theorthographic units with the pattern producedthrough feedback from the hidden units, and wereinterpreted as a measure reflecting the perform-ance of the model in a lexical decision task.Orthographic error scores are therefore a meas-ure of orthographic familiarity. Seidenberg andMcClelland showed that the model fitted humandata on a wide range of inputs. For example,regular words (such as “gave”) were pronouncedfaster than exception words (such as “have”).

Note that the Seidenberg and McClellandmodel uses a single mechanism to read nonwordsand exception words. There is only one set ofhidden units, and only one process is used to nameregular, exception, and novel items. As the modeluses a distributed representation, there is no one-to-one correspondence between hidden units andlexical items; each word is represented by a patternof activation over the hidden units. Accordingto this model, lexical memory does not consistof entries for individual words. Orthographic

a common trick to represent position-specificity(Wickelgren, 1969). For example, the word “have”was represented by the triples “#ha,” “hav,” “ave,”“ve#,” with “#” representing a blank space. Anon-local representation was used: The graphemicrepresentations were encoded as a pattern ofactivation across the orthographic units ratherthan corresponding directly to particular graph-emes. Each phoneme triple was encoded as apattern of activation distributed over a set of unitsrepresenting phonetic features—a representationknown as a Wickelfeature. The underlying archi-tecture was not a simple feed-forward one, in thatthe hidden units fed back to the orthographic units,mimicking top-down word-to-letter connectionsin the IAC model of word recognition. However,there was no feedback from the phonological tothe hidden units, so phonological representationscould not directly influence the processing oforthographic-level representations.

The training corpus comprised all 2897uninflected monosyllabic words of at least threeor more letters in the English language present inthe Kucera and Francis (1967) word corpus. Eachtrial consisted of the presentation of a letter stringthat was converted into the appropriate pattern ofactivation over the orthographic units. This in turnfed forward to the phonological units by way ofthe hidden units. In the training phase, words werepresented a number of times with a probabilityproportional to the logarithm of their frequency.This means that the ease with which a word islearned by the network, and the effect it has onsimilar words, depends to some extent on itsfrequency. About 150,000 learning trials were

FIGURE 7.5

Output layer(phonological units)

Hidden layer

Input layer(visual units)

H S A V E

/h/ /s/ /a/ /v/ /e/

The layers of Seidenberg andMcClelland’s (1989) model ofword recognition (simplified—see text for details). Based onSeidenberg and McClelland(1989).

9781841693828_4_007.PM5 9/21/07, 10:04 AM231



neighbors do not influence the pronunciation of aword directly at the time of processing; instead,regularity effects in pronunciation derive fromstatistical regularities in the words of the train-ing corpus—all the words we have learned—asimplemented in the weights of connections in thesimulation. Lexical processing therefore involvesthe activation of information, and is not an all-or-none event.

Evaluation of the original SM model

Coltheart et al. (1993) criticized importantaspects of the Seidenberg and McClelland (SM)model. They formulated six questions about read-ing that any account of reading must answer:

• How do skilled readers read exception wordsaloud?

• How do skilled readers read nonwords aloud?• How do participants make visual lexical deci-

sion judgments?• How does surface dyslexia arise?• How does phonological dyslexia arise?• How does developmental dyslexia arise?

Coltheart et al. then argued that Seidenberg andMcClelland’s model only answered the first ofthese questions.

Besner, Twilley, McCann, and Seergobin (1990)provided a detailed critique of the Seidenberg andMcClelland model, although a reply by Seidenbergand McClelland (1990) answered some of thesepoints. First, Besner et al. argued that in a sensethe model still possesses a lexicon, where insteadof a word corresponding to a unit, it correspondsto a pattern of activation. Second, they pointedout that the model “reads” nonwords rather poorly—certainly much less well than a skilled reader.In particular, it only produced the “correct,”regular pronunciation of a nonword under 70% ofthe time. This contrasts with the model’s excel-lent performance on its original training set. Hencethe model’s performance on nonwords is impairedfrom the beginning. In reply, Seidenberg andMcClelland (1990) pointed out that their modelwas trained on only 2987 words, as opposed tothe 30,000 words that people know, and that this

may be responsible for the difference. Hence themodel simulates the direct lexical route ratherbetter than it simulates the indirect grapheme–phoneme route. Therefore any disruption of themodel will give a better account of disruption tothe direct route—that is, of surface dyslexia. Themodel’s account of lexical decision is inadequatein that it makes far too many errors—in par-ticular it accepts too many nonwords as words(Besner et al., 1990; Fera & Besner, 1992). Themodel did not perform as well as people do onnonwords, in particular on nonwords that containunusual spelling patterns (e.g., JINJE, FAIJE). Inaddition, the model’s account of surface dyslexiawas problematic and its account of phonologicaldyslexia non-existent.

Forster (1994) evaluated the assumptionsbehind connectionist modeling of visual wordrecognition. He made the point that showing thata network model can successfully learn to per-form a complex task such as reading does notmean that that is the way humans actually do it.Finally, Norris (1994b) argued that a major stum-bling block for the Seidenberg and McClellandmodel was that it could not account for the abilityof readers to shift strategically between relianceon lexical and sublexical information.

The revised connectionist model: PMSP

A revised connectionist model performs muchbetter at pronouncing nonwords and at lexicaldecision than the original (Plaut, 1997; Plaut &McClelland, 1993; Plaut, McClelland, Seidenberg,& Patterson, 1996; Seidenberg, Petersen, Mac-Donald, & Plaut, 1996; Seidenberg, Plaut, Petersen,McClelland, & McRae, 1994). The model, calledPMSP for short, used more realistic input andoutput representations. Phonological representa-tions were based on phonemes with phonotacticconstraints (that constrain which sounds occurtogether in the language), and orthographic rep-resentations were based on graphemes withgraphotactic constraints (that constrain whichletters occur together in the language). The ori-ginal SM model performed badly on nonwordsbecause Wickelfeatures disperse spelling–soundregularities. For example, in GAVE, the A is

9781841693828_4_007.PM5 9/21/07, 10:04 AM232


7. READING 233

represented in the context of G and V, and hasnothing in common with the A in SAVE (rep-resented in the context of S and V). In the revisedPMSP model, letters and phonemes activate thesame units irrespective of context. A mathemat-ical analysis showed that a response to a letterstring input is a function that depends positivelyon the frequency of exposure to the pattern, posi-tively to the sum of the frequencies of its friends,and negatively to the sum of the frequencies ofits enemies. The response to a letter string isnon-linear, in that there are diminishing returns:For example, regular words are so good they gainlittle extra benefit from frequency. This explainsthe interaction we observe between word con-sistency and frequency. As we shall see, therevised model also gives a much better accountof dyslexia.

Accessing semantics

Of course the goal of reading is to access themeaning of words. The PSMP model simulatesthe orthography–phonology side of the triangle.Clearly, according to the model, we can accesssemantics either directly (OS: orthography–semantics) or indirectly (OPS: orthography–phonology–semantics—what we have also calledphonological mediation). Hence there is a divi-sion of labor between the two routes. Harm andSeidenberg (2004) model the access of seman-tics. In the full model, all parts of the systemoperate simultaneously and contribute to the acti-vation of meaning. The Harm and Seidenbergmodel is a complete implementation of the tri-angle model. It is trained to produce the correctpattern of activation across a set of semanticfeatures given an orthographic input. In the firstphase, the model is trained for a while on thephonology–semantics side of the triangle, to simu-late the knowledge of young children who cannotyet read, but who know what words mean. Theseweights were then frozen. In the second phase,the orthography–phonology and orthography–semantics sides of the triangle were then trained.

How does the trained model perform? Perhapsnot surprisingly, in simulations resembling theskilled reader in normal conditions, the OS route

is normally faster, with the OPS route laggingsomewhat behind. Nevertheless, analysis of howactivation of the input determines activation ofthe output shows that activation of the semanticsystem is driven by both pathways. Even ifthe OPS path is slower, it still always contributesto the final output. In addition, because ofinteractivity in the system, activation of thesemantic system activates corresponding pho-nological representations, which in turn affectthe semantic system. Simulations show that therelative contributions of the two pathways (OSand OPS) are modulated by a number of factors,including skill (phonological information is moreimportant early on in training, correspondingto less skilled readers) and word frequency (forhigh-frequency words the OS pathway is moreefficient). The model also simulates the responsetimes of Van Orden (1987), where people are slowto say “no” to “Is it a flower? ROWS.”

Connectionist models of dyslexia

Over the last few years connectionist modelinghas contributed to our understanding of deep andsurface dyslexia.

Modeling surface dyslexia

Patterson, Seidenberg, and McClelland (1989)artificially damaged or “lesioned” the Seidenbergand McClelland (1989) network after the learn-ing phase by destroying hidden units or connec-tion weights, and then observed the behavior ofthe model. Its performance resembled the read-ing of a surface dyslexic. Patterson et al. (1989)explored three main types of lesion: damage tothe connections between the orthographic inputand hidden units (called early weights); damageto the connections between the hidden and output(phonological) units (called late weights), anddamage to the hidden units themselves. Damagewas inflicted by probabilistically resetting aproportion of the weights or units to zero. Thegreater the amount of damage being simulated,the higher the proportion of weights that waschanged. The consequences were measured in twoways. First, the damage was measured by the

9781841693828_4_007.PM5 9/21/07, 10:04 AM233



phonological error score, which as we have seenreflects the difference between the actual andtarget activation values of the phonologicaloutput units. Obviously, high error scores reflectimpaired performance. Second, the damage wasmeasured by the reversal rate. This correspondsto a switch in pronunciation by the model, so thata regular pronunciation is given to an exceptionitem (for example, “have” is pronounced to rhymewith “gave”).

Increasing damage at each location producesnear-linear increases in the phonological errorscores of all types of word. On the whole, though,the lesioned model performed better with regularthan with exception words. The reversal rateincreased as the degree of damage increased,but nevertheless there were still more reversalsoccurring on exception words than on regularwords. Damage to the hidden units in particularproduced a large number of instances whereexception words were produced with a regularpronunciation; this is similar to the result wherebysurface dyslexics over-regularize their pronuncia-tions. However, the number of regularized pro-nunciations that were produced by the lesionedmodel was significantly lower than that producedby surface dyslexic patients. No lesion made themodel perform selectively worse on nonwords.Hence the behavior of the lesioned model re-sembles that of a surface dyslexic.

Patterson et al. also found that word frequencywas not a major determinant of whether a pro-nunciation reversed or not. (It did have someeffect, so that high-frequency words were gen-erally more robust to damage.) As we have seen,some surface dyslexics show frequency effectson reading, while others do not. Patterson et al.found that the main determinant of reversalswas the number of vowel features by which theregular pronunciation differs from the correctpronunciation, a finding verified from the neuro-psychological data.

An additional point of interest is that thelesioned model produced errors that havetraditionally been interpreted as “visual” errors.These are mispronunciations that are not over-regularizations and that were traditionally thoughtto result from an impairment of early graphemic

analysis. If this analysis is correct, then Pattersonet al. should only have found such errors whenthere was damage to the orthographic units in-volved. In contrast, they found them even whenthe orthographic units were not damaged. Thisis an example of a particular strength ofconnectionist modeling; the same mechanismexplains what were previously considered tobe disparate findings. Here visual errors resultfrom the same lesion that causes other char-acteristics of surface dyslexia, and it is unneces-sary to resort to more complex explanationsinvolving additional damage to the graphemicanalysis system.

There are three main problems with this par-ticular account. First, we have already seen thatthe original Seidenberg and McClelland modelwas relatively bad at producing nonwords beforeit was lesioned. We might say that the originalmodel is already operating as a phonological dys-lexic. Yet surface dyslexics are good at readingnonwords. Second, the model does not really over-regularize, it just changes the vowel sound ofwords. Third, Behrmann and Bub (1992) reporteddata that are inconsistent with this model. Inparticular, they showed that the performance ofthe surface dyslexic MP on irregular words doesvary as a function of word frequency. They inter-preted this frequency effect as problematic forconnectionist models. Patterson et al. (1989)were quite explicit in simulating only surface dys-lexia; their model does not address phonologicaldyslexia.

Exploring semantic involvement in reading

The revised model, abbreviated to PSMP, pro-vides a better account of dyslexia. The improve-ments come about because the simulationsimplement both pathways of the triangle modelin order to explain semantic effects on reading.

Surface dyslexia arises in the progressive neu-rological disease dementia (see Chapter 11 onsemantics for details of dementia). Importantly,people with dementia find exception words diffi-cult to pronounce and repeat if they have lost themeaning of those words (Hodges, Patterson,Oxbury, & Funnell, 1992; Patterson & Hodges,

9781841693828_4_007.PM5 9/21/07, 10:04 AM234


7. READING 235

1992; but see Funnell, 1996). Patterson andHodges proposed that the integrity of lexical rep-resentations depends on their interaction with thesemantic system: Semantic representations bindphonological representations together with a se-mantic glue; hence this is called the semantic gluehypothesis. As the semantic system gradually dis-solves in dementia, so the semantic glue gradu-ally comes unstuck, and the lexical representationslose their integrity. Patients are therefore forcedto rely on a sublexical or grapheme–phonemecorrespondence reading route, leading to surfacedyslexic errors. Furthermore, they have difficultyin repeating irregular words for which they havelost the meaning, if the system is sufficientlystressed (by repeating lists of words), but theycan repeat lists of words for which the meaning isintact (Patterson, Graham, & Hodges, 1994; butsee Funnell, 1996, for a patient who does notshow this difference).

PMSP showed that a realistic model of surfacedyslexia depends on involving semantics in read-ing. Support from semantics normally relieves thephonological pathway from having to master low-frequency exception words by itself. In surfacedyslexia the semantic pathway is damaged, and theisolated phonological pathway reveals itself assurface dyslexia.

Plaut (1997) further examined the involvementof semantics in reading. He noted that some pa-tients have substantial semantic impairments butcan read exception words accurately (e.g., DC ofLambon Ralph, Ellis, & Franklin, 1995; DRN ofCipolotti & Warrington, 1995; WLP of Schwartz,Marin, & Saffran, 1979). To explain why somepatients with semantic impairments cannot readexception words but some can, Plaut suggestedthat there are individual differences in the divisionof labor between semantic and phonological path-ways. Although the majority of patients with se-mantic damage show surface dyslexia (Graham,Hodges, & Patterson, 1994), some exceptions arepredicted. He also argued that people use a numberof strategies in performing lexical decision, oneof which is to use semantic familiarity as a basisfor making judgments. The revised model there-fore takes into account individual differences be-tween speakers, and shows how small differences

in reading strategies can lead to different conse-quences after brain damage.

Modeling phonological dyslexia

The triangle model provides the best connectionistaccount of phonological dyslexia. It envisagesreading as taking place through the three routesconceptualized in the original SM model. Theroutes are orthography to phonology, orthographyto semantics, and semantics to phonology (Figure7.3). This approach sees phonological dyslexia asnothing other than a general problem with phono-logical processing (Farah et al., 1996, Sasanumaet al., 1996). Phonological dyslexia arises throughimpairments to representations at the phono-logical level, rather than to grapheme–phonemeconversion. This is called the phonological im-pairment hypothesis. People with phonological dys-lexia can still read words because their weakenedphonological representations can be accessedthrough the semantic level. (Hence this approachis also a development of the semantic gluehypothesis.) We have already noted that theoriginal Seidenberg and McClelland (1989) modelperformed rather like a phonological dyslexicpatient, in that it performed relatively poorlyon nonwords. Consistent with the phonologicaldeficit hypothesis, the explanation for this poorperformance was that the source of these errorswas the impoverished phonological representationsused by the model.

An apparent problem with the phonologicaldeficit hypothesis is that it is not clear that itwould correctly handle the way in which peoplewith phonological dyslexia read pseudohomo-phones better than other types of nonwords(Coltheart, 1996). Furthermore, patient LB ofDerouesné and Beauvois (1985) showed anadvantage for pseudohomophones, but no obvi-ous general phonological impairment. There havealso been effects of orthographic complexityand visual similarity, suggesting that there isalso an orthographic impairment present in pho-nological dyslexia (Derouesné & Beauvois, 1985;Howard & Best, 1996). For example, Howardand Best showed that their patient Melanie-Janeread pseudohomophones that were visually

9781841693828_4_007.PM5 9/21/07, 10:04 AM235



similar to the related word (e.g., GERL) better thanpseudohomophones that were visually more dis-tant (e.g., PHOCKS). There was no effect of visualsimilarity for control nonwords. However, Harmand Seidenberg (2001) show how phonologicalimpairment in a connectionist model can give riseto such effects. A phonological impairment mag-nifies the ease with which different types of stimuliare read.

Modeling deep dyslexia

Hinton and Shallice (1991) lesioned anotherconnectionist model to simulate deep dyslexia.Their model was trained by back-propagation toassociate word pronunciations with a representa-tion of the meaning of words. This model is par-ticularly important, because it shows that one typeof lesion can give rise to all the symptoms ofdeep dyslexia, particularly both paralexias andvisual errors.

The underlying semantic representation of aword is specified as a pattern of activation acrosssemantic feature units (which Hinton & Shallicecalled sememes). These correspond to semanticfeatures or primitives such as “main-shape-2D,”“has-legs,” “brown,” and “mammal.” These canbe thought of as atomic units of meaning (seeChapter 11). The architecture of the Hinton andShallice (1991) model comprised 28 graphemicinput units and 68 semantic output units with anintervening hidden layer containing 40 intermedi-ate units. The model was trained to produce anappropriate output representation given a particu-lar orthographic input using back-propagation.The model was trained on 40 uninflected mono-syllabic words.

The structure of the output layer is quite com-plex. First, there were interconnections betweensome of the semantic units. The 68 semantic fea-ture units were divided into 19 groups dependingon their interpretation, with inhibitory connec-tions between appropriate members of the group.For example, in the group of semantic featuresthat define the size of the object denoted by theword, there are three semantic features: “max-size-less-foot,” “max-size-foot-to-two-yards,” and“max-size-greater-two-yards.” Each of these

features inhibits the others in the group, becauseobviously an object can only have one size. Sec-ond, an additional set of hidden units calledcleanup units was connected to the semantic units.These permit more complex interdependenciesbetween the semantic units to be learned, and havethe effect of producing structure in the outputlayer. This results in a richer semantic space wherethere are strong semantic attractors. An attractorcan be seen as a point in semantic space to whichneighboring states of the network are attracted;it resembles the bottom of a valley or basin, sothat objects positioned on the sides of the basintend to migrate towards the lowest point. Thiscorresponds to the semantic representation ulti-mately assigned to a word.

As in Patterson et al.’s (1989) simulation ofsurface dyslexia, different types of lesion werepossible. There are two dimensions to remember:one is what is lesioned, the other is how it islesioned. The connections involved were thegrapheme–intermediate, intermediate–sememe,and sememe–cleanup. Three methods of lesioningthe network were used. First, each set of connec-tions was taken in turn, and a proportion of theirweights was set to zero (effectively disconnectingunits). Second, random noise was added to eachconnection. Third, the hidden units (the interme-diate and cleanup units) were ablated by destroy-ing a proportion of them.

The results showed that the closer the lesionwas to the semantic system, the more effect ithad. The lesion type and site interacted in theireffects; for example, the cleanup circuit was moresensitive to added noise than to disconnections.Lesions resulted in four types of error: semantic(where an input gave an output word that wassemantically but not visually close to the target;these resemble the classic semantic paralexias ofdeep dyslexics); visual (words visually but notsemantically similar), mixed (where the outputis both semantically and visually close to the

KEY TERM

Attractors: a point in the connectionist attractornetwork to which related states are attracted.

9781841693828_4_007.PM5 9/21/07, 10:04 AM236


7. READING 237

target), and others. All lesion sites and types(except for that of disconnecting the semanticand cleanup units) produced the same broadpattern of errors. Finally, on some occasions thelesions were so severe that the network could notgenerate an explicit response. In these cases,Hinton and Shallice tested the below-thresholdinformation left in the system by simulatinga forced-choice procedure. They achieved thisby comparing the residual semantic output to aset of possible outputs corresponding to a set ofwords, one of which was the target semanticoutput. The model behaved above chance on thisforced-choice test, in that its output semanticrepresentation tended to be closer to that of thetarget than to the alternatives.

Hence the lesioned network behaves like adeep dyslexic patient, in particular in makingsemantic paralexias. The paralexias occur becausesemantic attractors cause the accessing of featureclusters close to the meanings of words that arerelated to the target. A “landscape” metaphor maybe useful. Lesions can be thought of as resultingin the destruction of the ridges that separate thedifferent basins of attraction. The occurrenceof such errors does not seem to be crucially de-pendent on the particular lesion type or site underconsideration. Furthermore, this account providesan explanation of why different error types, par-ticularly semantic and visual errors, nearly alwaysco-occur in such patients. Two visually similarwords can point in the first instance to nearbyparts of semantic space, even though their ulti-mate meanings in the basins may be far apart; ifyou start off on top of a hill, going downhill indifferent directions will take you to very differentultimate locations. Lesions modify semantic spaceso that visually similar words are then attracted todifferent semantic attractors.

Hinton and Shallice’s account is importantfor cognitive neuropsychologists for a number ofreasons. First, it provides an explicit mechanismwhereby the characteristics of deep dyslexia canbe derived from a model of normal reading.Second, it shows that the actual site of the lesionis not of primary importance. This is mainlybecause of the “cascade” characteristics of thesenetworks. Each stage of processing is continu-

ally activating the next, and is not dependent onthe completion of processing by its prior stage(McClelland, 1979). Therefore, effects of lesionsat one network site are very quickly passed on tosurrounding sites. Third, it shows why symptomsthat were previously considered to be conceptuallydistinct necessarily co-occur. Semantic and visualerrors can result from the same lesion. Fourth,it thus revives the importance of syndromes asa neuropsychological concept. If symptoms co-occur as a result of any lesion to a particularsystem, then it makes sense to look for and studysuch co-occurrences.

Plaut and Shallice (1993a) extended this workto examine the effect of word abstractness onlesioned reading performance. As we have seen,the reading performance of deep dyslexic patientsis significantly better on more imageable than onless imageable words. Plaut and Shallice showedthat the richness of the underlying semanticrepresentation of a word is an analog of im-ageability. They hypothesized that the semanticrepresentations of abstract words contain fewersemantic features than those of concrete words;that is, the more concrete a word is, the richerits semantic representation. Jones (1985) showedthat it was possible to account for imageabilityeffects in deep dyslexia by recasting them asease-of-predication effects. Ease-of-predicationis a measure of how easy it is to generate thingsto say about a word, or predicates, and is obvi-ously closely related to the richness of the under-lying semantic representation. It is easier to findmore things to say about more imageable wordsthan about less imageable words. Plaut andShallice (1993a) showed that when an attractornetwork similar to that of Hinton and Shallice(1991) is lesioned, concrete words are read betterthan abstract words. One exception was that severelesions of the cleanup system resulted in betterperformance on abstract words. Plaut and Shalliceargue that this is consistent with patient CAV(Warrington, 1981), who showed such an advan-tage. Hence this network can account for both theusual better performance of deep dyslexic patientson concrete words, and also the rare exceptionwhere the reverse is the case. They also showedthat lesions closer to the grapheme units tended

9781841693828_4_007.PM5 9/21/07, 10:04 AM237



to produce more visual errors, whereas lesionscloser to the semantic units tended to producemore semantic errors. The model also providesan account of the behavior of normal participantsreading degraded words (McLeod, Shallice, &Plaut, 2000). If words are presented very rapidlyto people, they make both visual and semanticerrors. The data fit the connectionist model well.

Connectionist modeling has advanced ourunderstanding of deep dyslexia in particular,and neuropsychological deficits in general. Thefinding that apparently unrelated symptoms cannecessarily co-occur as a result of a single lesionis of particular importance. It suggests that deepdyslexia may after all be a unitary condition. How-ever, there is one fly in the ointment. The findingthat at least some patients show imageability effectsin reading but not in comprehension is trouble-some for all models that posit a disturbance ofsemantic representations as the cause of deep dys-lexia (Newton & Barry, 1997). Instead, in at leastsome patients, the primary disturbance may be tothe speech production component of reading.

Comparison of models

A simple dual-route model provides an inadequateaccount of reading, and needs at least an addi-tional lexical route through imageable semantics.The more complex a model becomes, the greaterthe worry that routes are being introduced on anarbitrary basis to account for particular findings.Analogy models have some attractive features, buttheir detailed workings are vague and they do notseem able to account for all the data. Connectionistmodeling has provided an explicit, single-routemodel that covers most of the main findings, buthas its problems. At the very least it has clarifiedthe issues involved in reading. Its contributiongoes beyond this, however. It has set the challengethat only one route is necessary in reading wordsand nonwords, and that regularity effects in pro-nunciation arise out of statistical regularities inthe words of the language. It may not be a com-plete or correct account, however it is certainly achallenging one.

Currently we are faced with two seriousalternatives: a connectionist model such as thetriangle model, and a variant of the dual-route

model such as the dual-route cascaded model. Theliterature is full of claim and counter-claim, andit would be presumptuous for a text like this tosay that one is clearly right and the other wrong.There are many studies providing support for andagainst one or the other of the models. Many ofthem focus on how we read nonwords (Besneret al., 1990; Seidenberg et al., 1994), because thedivision of labor in the DRC model between alexical route with knowledge of individual wordsand a non-lexical route with spelling rules isabsent in connectionist models, and this differ-ence is the key one between the two sorts ofmodels. The DRC emphasizes regularity (doesthe word obey the rule?), which is a categoricalconcept—either the word obeys the spelling-soundrules or it does not, with nonwords having to bepronounced by the rule. The triangle modelemphasizes consistency of rimes and other units(how often is -AVE pronounced in a certain way?),which is a statistical concept. According to Zevinand Seidenberg (2006), consistency effects suchas those shown in Glushko’s (1979) and Jared’s(1997b) study are the critical test between models.Words like PAVE are regular but inconsistent;according to the DRC model they should be aseasy to pronounce as regular and consistent wordssuch as PANE; according to the triangle modelthey should not. Now of course we know fromGlushko’s study that regular inconsistent wordsare slower to pronounce than regular consistentones, but Coltheart et al. (2001) argue that thesedifferences are an artifact arising from severalconfounding factors (e.g., the presence of excep-tion words in the materials, and an increase inthe number of times it is necessary to reanalyzeinconsistent words as we read them from leftto right). Zevin and Seidenberg (2006) arguedthat graded sensitivity to consistency effects innonwords provides the critical test between themodels, with only connectionist models correctlypredicting the presence of such effects, and beingable to account for individual differences in non-word pronunciation. However, doubtless this de-bate will run and run.

Perhaps the choice between the triangle andthe dual-route cascaded model comes down towhich one values most: explaining a wide rangeof data, or parsimony in design.

9781841693828_4_007.PM5 9/21/07, 10:04 AM238


7. READING 239

Balota (1990) asked it there is a magicmoment when we recognize a word but do notyet have access to its meaning. He argued thatthe tasks most commonly used to study wordprocessing (lexical decision and word naming) areboth sensitive to post-access processes. This makesinterpretation of data obtained using these tasksdifficult (although not, as we have seen, imposs-ible). Furthermore, deep dyslexia (discussed later)suggests that it is possible to access meaningwithout correctly identifying the word, while non-semantic reading suggests that we can recognizewords without necessarily accessing their mean-ing. Whereas unique lexical access is a prerequi-site of activating meaning in models such as thelogogen and the serial search model, cascading

connectionist models permit the gradual activa-tion of semantic information while evidence isstill accumulating from perceptual processing.A model such as the triangle model (Patterson,Suzuki, & Wydell, 1996; Plaut, McClelland,Seidenberg, & Patterson, 1996) seems best ableto accommodate all these constraints.

Finally, all of these models—particularly theconnectionist ones—are limited in that they havefocused on the recognition of morphologicallysimple, often monosyllabic words. Rastle andColtheart (2000) have developed a rule-basedmodel of reading bisyllabic words, emphasizinghow we produce the correct stress, and Ans,Carbonnel, and Valdois (1998) have developed aconnectionist model of reading polysyllabic words.

SUMMARY

• Different languages use different principles to translate words into sounds; languages such asEnglish use the alphabetic principle.

• Regular words have a regular grapheme-to-phoneme correspondence, but exception words do not.• According to the dual-route model, words can be read through a direct lexical route or a sublexical

route; in adult skilled readers the lexical route is usually faster.• The sublexical route was originally thought to use grapheme–phoneme conversion, but now it is

considered to use correspondences across a range of sublexical levels.• There are effects of lexical similarity in reading certain nonwords (pseudohomophones), while

not all words are read with equal facility (the consistency of the regularity of a word’s neighborsaffects its ease of pronunciation).

• It might be necessary to access the phonological code of a word before we can access itsmeaning; this process is called phonological mediation.

• Phonological mediation is most likely to be observed with low-frequency words and with poorreaders.

• Readers have some attentional control over which route they emphasize in reading.• Access to some phonological code is mandatory, even in silent reading, but normally does not

precede semantic access.• Increasing reading speed above about 350 words a minute (by speed reading, for example) leads

to reduced comprehension.• Surface dyslexia is difficulty in reading exception words; it corresponds to an impairment of the

lexical route in the dual-route model.• Phonological dyslexia is difficulty in reading nonwords; it corresponds to an impairment of

the sublexical route in the dual-route model.• Deep dyslexic readers display a number of symptoms including making visual errors, but the

most important characteristic is the presence of semantic reading errors or paralexias.• There has been some debate as to whether deep dyslexia is a coherent syndrome.• Non-semantic readers can pronounce irregular words even though they do not know their meaning.• The revised dual-route model uses multiple sublexical correspondences and permits direct access

through a semantic lexical route and a non-semantic lexical route.

9781841693828_4_007.PM5 9/21/07, 10:04 AM239



• The dual-route cascaded model allows activation to trickle through levels before processing isnecessarily completed at any level.

• Seidenberg and McClelland (SM) produced an important connectionist model of reading; how-ever, it performed poorly on nonwords and pseudohomophones.

• Lesioning the SM network gives rise to behavior resembling surface dyslexia, but its over-regularizations differ from those made by humans.

• The revised version of this model, PSMP, gives a much better account of normal reading andsurface dyslexia; it uses a much more realistic representation for input and output than theoriginal model.

• There are clear semantic influences on normal and impaired reading, and recent connectionistmodels are trying to take these into account.

• The triangle model accounts for phonological dyslexia as an impairment to the phonologicalrepresentations: this is the phonological impairment hypothesis.

• Deep dyslexia has been modeled by lesioning semantic attractors; the lesioned model shows howthe apparently disparate symptoms of deep dyslexia can arise from one type of lesion.

• More imageable words are relatively spared because they have richer semantic representations.• There has been considerable debate as to whether developmental dyslexia is qualitatively differ-

ent from very poor normal reading, and whether there are subtypes that correspond to acquireddyslexias; the preponderance of evidence suggests that developmental dyslexia is on a continuumwith normal reading.

• Connectionist modeling shows how two distinct types of damage can lead to a continuum ofimpairment between development surface and phonological dyslexia extremes.

SOME QUESTIONS TO THINK ABOUT

1. Is there a “magic moment” when we recognize a word but cannot access its meaning?2. Why might reading errors occur? Keep a record of any errors you make and try to relate them to

what you have learned in this and the previous chapter.3. What practical tips could help adult dyslexic readers to read more effectively?

FURTHER READING

Many of the references at the end of Chapter 6 will also be relevant here. There are a number ofworks that describe the orthography of English, and discuss the rules whereby certain spelling-to-sound correspondences are described as regular and others as irregular. One of the best known ofthese is Venezky (1970). For an example of work on reading in a different orthographic system, seeKess and Miyamoto (1999).

For a general introduction to reading, writing, spelling, and their disorders, see Ellis (1993).For more discussion of dyslexia, including peripheral dyslexias, see Ellis and Young (1988). Twovolumes (entitled Deep dyslexia, 2nd edition, by Coltheart, Patterson, & Marshall, 1987, andSurface dyslexia by Patterson, Marshall, & Coltheart, 1985b) cover much of the relevant material.A special issue of the journal Cognitive Neuropsychology (1996, volume 13, part 6) was devotedto phonological dyslexia.

For recent overviews of reading, see Andrews (2006) and Snowling and Hulme (2007).

9781841693828_4_007.PM5 9/21/07, 10:04 AM240


SUBJECT INDEX 591

591

Subject Index

Akeakamai (dolphin), 58Alex (parrot), 57Alexia, 491Algorithms, 485Alignment, 459Allan, 250Allophones, 28, 491Alphabetic principle, 247–248Alphabetic reading, 241Alphabetic scripts, 210Alveolars, 32Alzheimer’s disease (AD), 146,

350–353, 351, 448, 491Ambiguity, 38, 199–208, 289–

291, 298–313, 321, 329,374–377, 460

American English, 29American Sign Language (ASL/

AMESLAN), 59–60, 61, 62,78, 492

Analogy models, 229–230, 238Analysis-by-synthesis, 267–

268Analytic phonics, 248Anaphor resolution, 374Anaphora, 374–277, 491Aneurysms, 443, 491Angular gyrus, 69Animal language, 54–67Animate-inanimate dissociation,

346–350Anomia, 71, 344, 417, 441–442,

468–469, 491Antecedents, 374, 491Antonyms, 323Ants, 54Apes, 53–54, 58–67

Entries given in bold indicateglossary definitions

AB, 351Abstraction-based theories, 338AC, 343Access files, 193Accessibility, 377–378Accommodation, 80, 401, 403Acoustic invariance, 258–259Acoustic phonetics, 29Acoustics, 27, 491Acquired disorders, 220, 491Acquired dyslexia, 220–227Acquisition and learning

distinction hypothesis, 159ACT/ACT* models, 379Activation, 13, 190, 265, 327–

328, 425, 485, 486, 491Active-filler strategy, 314Active filter hypothesis, 160Adjacency pairs, 458Adjectives, 36, 491Adverbs, 36, 491Affirmative declarative sentences,

40Affixes, 403, 491

addition/loss, 411stripping, 191

Affordances, 345Affricatives, 33Age-of-acquisition (AOA), 174–

175, 218Agent, 83, 136, 287, 288, 491Agnosia, 185, 344, 491Agrammatism, 316–318, 437–

439, 491

Aphasia, 68, 70–71, 281, 411,435–437, 439–440,445–446, 491

bilinguals, 158connectionist models, 442–445

Apraxia, 435, 491AR, 224, 343Arabic, 111, 149, 210Arcuate fasciculus, 14, 15, 69Argument structure, 42, 141–144,

287, 492Articulatory apparatus, 52, 492Articulatory phonetics, 29Artificial intelligence (AI), 11–13,

328–329Artificial languages, 58, 118AS, 281ASL, see American Sign

LanguageAspect, 405, 492Aspirated/Aspiration, 28, 492Assimilation, 80, 259, 492Associative priming, 185,

186–187Attachment, 309, 492Attachment preference, 307–308,

492Attentional priming, 180, 181Attentional processing, 178–181,

190, 318, 492Attractors, 236, 354, 492

network, 354Audience design, 291, 459–460Audiolingual method, 159Auditory phonetics, 29Auditory short-term memory

(ASTM), 472–474, 492

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM591


592 SUBJECT INDEX

Augmented transition network(ATN), 492

Austin (chimp), 61, 66Autism, 82Automata theory, 43–45Automatic priming, 180, 181,

186–187Automatic processing, 178–181,

318, 492Autonomous access model, 203Autonomous models, 21, 193–

195, 198–199, 266, 288,289, 299–301

Auxiliary verbs, 40, 492AW, 411

Babbling, 123, 492Baby talk, see Child-directed

speech (CDS)Back-channel communication,

458Back-propagation, 146–147, 230,

485, 487–489, 492Backtrack, 492Backward inferences, 369Backward-looking centers, 378Backward translation, 157Backwards masking, 172Balanced ambiguity, 203Basic color terms, 94, 95Basic level, 130, 336–337,

492Basque, 7Bassa, 94Batonic gestures, 434BC, 346Bees, 54–55Behaviorism, 10Berinmo, 96Berlitz method, 159Bias, 423, 488Bilabial sounds, 32Bilingual, 153, 492Bilingualism, 153–158, 161–162,

260, 417, 423, 492Binding, 314Bins, 193Blindness, 84–86, 87Blocking hypothesis, 417Blocksworld, 12BO, 391, 475Bobby, 249Body, 215, 492

subsytem, 228Boltzman machines, 489Bonding, 372Bonobos, 64, 66

Bootstrapping, 121, 122–123,130, 136, 137, 492

Bottom-up priority, 269Bottom-up processing, 21, 479–

480, 492Bound morphemes, 403, 492Boundary effect, 386Bow-wow theory, 52Box-and-arrow diagrams, 11, 15,

481Box-and-candle problem, 92, 93Boxology, 11, 481Brain

evolution, 52hemisphere dominance, 53–54,

68, 72–75, 158imaging, 16–19, 68, 482lesion studies, 14–16, 68, 73localization of functions, 67–71structure, 15

Bridging inferences, 369Brightness masking, 172British English, 29Broca’s aphasia, 68, 435–437,

445–446, 492Broca’s area, 14, 15, 52, 53, 68,

70, 72, 438Buffers, 492

Canonical sentence strategy, 296Cantonese, 33Capacity theory, 474Cascade models, 21, 420, 421,

423, 425, 427, 492Case grammar, 379CAT (computerized axial

tomography), 17Categorical perception, 260–261,

262–263, 493Categorization, 127, 322–323Category-specific disorders,

346–350, 355–356Causal coherence, 362Causative verbs, 144CAV, 237CB, 440, 444CELEX database, 173Center-embedding, 39Centering theory, 378Central dyslexias, 220Central executive, 439, 471Cerebral vascular accident (CVA),

493Characteristic features, 330, 332Child-directed speech (CDS), 83,

108–110, 121, 148, 493Chimpanzees, 58–67

Chinchillas, 122Chinese, 33, 92, 93, 97, 210, 218,

226–227, 244“Chinese readers”, 251Classification, 335–338Clauses, 37, 294, 493Cleanup units, 236Cleft sentences, 493“Clever Hans” effect, 57Click displacement, 294–295Clitics, 493Closed-class items, 37, 493Closed-class words, see Function

wordsClosure, 297Clue words, 246Co-articulation, 259, 493Coda, 33Code switching, 154–155Cognates, 156, 493Cognition hypothesis, 79–82, 87Cognitive constraint, 323Cognitive cycles, 434Cognitive economy, 322Cognitive intrusions, 404Cognitive linguistics, 43Cognitive neuropsychology,

14–16, 480, 481, 482Cognitive neuroscience, 14Cognitive psychology, 10Cognitive science, 11–13Coherence, 279, 362Coherence graph, 386Cohesion, 362Cohort model, 268–273, 280Co-indexation, 493Collective monologues, 80Color, 94–96Common ground, 377Common-store models, 155–156Communication, 5, 9, 54, 161Communication accuracy, 94Comparative linguistics, 10Compensation for co-articulation,

276Competence, 34–35, 493Competition, 305–306, 309Competition-integration model,

306, 309Competitive queuing, 429Complement, 493Complementizers, 290–291, 493Complex word processing, 191–

192Compound-cue theory, 190Comprehensible input hypothesis,

160

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM592


SUBJECT INDEX 593

Comprehension, 12–13, 218, 361–393

Computational metaphor, 10–11Concepts, 322, 336, 338–339Conceptual dependency theory,

379–380Conceptual mediation, 157Conceptual pacts, 457Conceptual Selection Model

(CSM), 414Conceptualization, 397–398Concrete operational stage, 79–80Conditioned head turn technique,

104–105Conditioning, 10, 44, 106–108Conduction aphasia, 446Conduit d’approche, 446Confidence, 161Conjoint frequency, 327Conjunctions, 36, 493Connectionism, 13, 23–24, 481,

485–489, 493age-of-acquisition, 175aphasia, 442–445dementia, 352–353, 355–356distributional analysis, 139distributional information, 117–

118dyslexia, 233–238lexical ambiguity, 206–207past tense acquisition, 146–148phonological encoding, 429reading, 230–233, 238–239semantic memory, 352–353semantics, 353–358speech production, 425, 426,

429speech recognition, 273–280

Connectionist neuropsychology,481, 482

Connotations, 324Conservation, 79–80, 82Consolidated alphabetic phase,

242Consonantal scripts, 210Consonants, 31–33, 493

processing, 281Constituents, 37, 493Constraints, 43, 298–299, 309Construction-integration model,

386–388Constructivist-semantic view, 137Content words, 36–37, 318, 401–

402, 493Context, 161, 187–190, 203–206,

261, 262–264, 266–267Context-free grammars, 40, 45

Context-guided single-readinglexical access model, 200

Context-sensitive grammars, 40,45

Context-sensitive model, 204–205Contingent negative variation

(CNV), 17Continuity hypothesis, 111, 112,

123Contrastive hypothesis, 134, 158–

159Controlled processing, 178, 493Conversation, 362, 454–460

analysis, 458back channels, 458collaboration, 459–460cooperation, 456–457implicatures, 457, 495inferences, 454–458maxims, 456–458, 493structure, 458–459turn-taking, 83–84, 458–459

Conversational hypothesis, 109Cooperative principle, 456–457Coordinates, 323Copulative verbs, 493Core description, 330, 331Co-reference, 374, 493Counterfactual, 493Creoles, 114, 493Critical period, 72–79Cross-fostering, 59Cross-linguistic differences, 120,

148–149, 307–308, 480, 483,493

Cross-modal priming, 186, 271Cross-nurturing, 59Cross-sectional studies, 105Crossed aphasia, 73, 158, 446Cue validity, 180Cultural transmission, 60Culture, 161Culture-specific information, 368CW, 348

D/P (declarative/procedural)model, 71

Dani, 94–95, 96Dante, 417Data, 4Data-driven, 21Dative verbs, 493David, 112Deafness, 76, 78, 86–87, 109,

112, 114, 123, 218, 281, 468Dean, John, 363Decay, 487

Decompositional theories, 328–335

Deep dysgraphia, 448Deep dyslexia, 223–225, 228,

236–238, 239, 354, 493subtypes, 224

Deep dysphasia, 443, 468, 493Deep structure (d-structure), 41Defining features, 330, 332Degenerate input, 108Deictic function, 312–313Dementia, 234–235, 300, 344,

350–353, 355–356, 391–392,470, 475

Denotation, 324Dentals, 32Dependencies, 314–315, 407–409Derivation, 493Derivational errors, 222Derivational morphology, 5, 6,

494Derivational theory of complexity

(DTC), 292, 293Design features, 55–57Determiners, 36, 494Detransformation, 292, 293Developmental disorders, 220,

494Developmental dysgraphia, 249Developmental dyslexia, 249–255DF, 255Dialects, 29Dialog, 459–460Diary studies, 154Dichotic-listening task, 201Digging-in, 306Ding-dong theory, 52Diphthongs, 31, 494Direct methods, 159Direct object, 37Direct speech acts, 456Disambiguation region, 290Disconnection syndrome, 69Discontinuity hypothesis, 124Discourse, 361–362, 494

analysis, 458neuropsychology of processing,

390–392Discrete stage models, 21, 494Dissociation, 22, 403, 494Distinctive features, 31Distinguishing features, 355, 356Distributional analysis, 24, 138–

139Distributional information, 117–

118, 121, 494Ditransitive verbs, 37

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM593


594 SUBJECT INDEX

DM, 300, 351, 445Dogs, 54, 57Dolphins, 55, 58Domain-specific knowledge

hypothesis (DSKH), 350Double dissociation, 16, 220,

347–348, 411, 445–446, 494Down’s syndrome, 81DP, 445DSMSG model, 443–445DT, 470Dual-code hypothesis, 343Dual-pathway hypothesis, 191Dual-route cascaded (DRC)

model, 228–229, 238Dual-route models

of reading, 211–212, 214, 219,225–226, 227–229, 238

of spelling, 248Dutch, 7, 96, 120, 307Dyirbal, 337Dysarthria, 494Dysgraphias, 220, 448, 494

developmental, 249Dyslexias, 185, 217, 220, 494

acquired, 220–227connectionist models, 233–238developmental, 249–255

Dysphasia, 443, 468Dysprosody, 435, 494

E-language (externalisedlanguage), 35

E-Z Reader model, 170Early-syntax theories, 143Ease-of-predication effects, 237Echolalia, 80EDE, 282EE, 468EEGs (Electroencephalography),

17, 494Egocentric speech, 80Elaborative inferences, 369ELIZA, 11–12, 13Elizabeth (chimp), 61EM, 343, 351Emergentist account, 73Empiricism, 105, 106Endogenous oscillators, 458–459Enemies, 215Energy masking, 172English, 7, 34, 42Entrainment, 457, 459Entrenchment hypothesis, 143Environmental contamination, 404EP, 229Epilinguistic knowledge, 243

Episodic memory, 322, 494Equipotentiality hypothesis, 73ERPs (Event-related potentials),

17, 494lateralization, 74, 158sentence processing, 295, 301speech recognition, 272word recognition, 156, 189

Error-reduction learning, 487Error scores, 231Eskimo, 90, 91EST, 442Evolution, 52–54Excitation, 486Execution, 397, 398Exemplar theory, 337–338Exercise hypothesis, 75Expressive, 494Extension, 324Eye fixation, 169, 216, 297Eye movements

comprehension, 372, 375reading, 168–170, 188–189,

216, 219sentence processing, 295, 299,

300, 303–304, 306, 309–310, 312, 461–462

word recognition, 203, 267Eye-tracking, 149, 157Eye-witness testimony, 363–364,

372

Face management, 457Face perception, 185Facilitation, 14, 177, 486, 494Familiarity, 173–174Familiarity bias, 423Family resemblance, 335–338Fan effect, 379Fast-mapping, 127Featural and Unitary Semantic

Space hypothesis, 335Feature comparison theory, 330–

331, 332Feature lists, 328, 330–331Feature masking, 172Feedback, 21, 83, 106–108, 423–

424, 427–428Felicia, 349Felicity conditions, 454Feral children, 76–77Figurative speech, 340–342, 494Filipino, 90Filled hesitations, 432Filled pauses, 432, 458Fillers, 314, 494Finite-state devices, 44–45

Finnish, 7, 149Finno-Ugric languages, 7First mention, 377–378Fixations, 169, 216Fixed-choice two-stage models,

309Fixed structure, 297Flow diagrams, 11Fluent aphasia, 411, 436Fluent restorations, 270fMRI (functional magnetic

resonance imaging), 18, 146,156, 307, 350

Focal colors, 94–96Focus, 376Foregrounding, 376Form-based priming, 177Formal operations stage, 80Formal paraphasias, 443, 494Formal universals, 112–113Formants, 27–28, 494Formulation, 397, 398Forward-looking centers, 378Forward translation, 157Four Cs, 161Fovea, 169FOXP2 gene, 53, 67, 115Frame-based models, 428–429Frame problem, 369French, 7, 34, 111, 210, 307Frequency attenuation, 176Frequency effect, 173–175, 182–

184Freudian slips, 399Fricatives, 33Friends, 215Full alphabetic phase, 242Full-listing hypothesis, 191Function words, 36 –37, 318,

401– 402, 494Functional architecture, 15Functional core hypothesis, 133Functional fixedness, 92Fuzzy prototypes, 276

Gaps, 313–315, 494Garden path model, 298, 299–

300, 309Garden path sentences, 290–291,

494Gating tasks, 271, 494Gaze, 83, 84, 405, 458Gender, 92, 138, 375, 417–418,

494Generalized delta rule, 488Generative grammar, 494Generative semantics, 379

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM594


SUBJECT INDEX 595

Genetic linguistics, 114–115Genie, 77–78German, 7, 92, 146, 245, 246Gerunds, 494Gestures, 53–54, 58, 86, 434,

458Given–new contract, 378GL, 442Glides, 33Global aphasia, 446Global modularity, 421Globality assumption, 445Glottal stops, 32–33, 494Glottis, 32Goal, 288, 494Gold’s theorem, 115–116Gorillas, 61Government and Binding (GB)

theory, 41GR, 223Grammar, 34–45, 494

learning, 83, 108Grammatical constraints, 323Grammatical differences, 91–92Grammatical elements, 494Grammaticality judgments, 66,

316Grapheme-to-phoneme conversion

(GPC) route, 211, 212Grapheme–phoneme conversion

mechanism, 227Grapheme–phoneme

correspondence, 29, 210,245–246

Graphemes, 209–210, 494analysis, 222

Greek, 97, 246Grounding, 357–358Group monologues, 80Gua (chimp), 59Guessing model, 183

HAL, 11HAM (Human Associative

Memory), 379Hard-wired modules, 22Harm and Seidenberg model, 233Harmonics, 29, 31HDM (high-dimensional memory)

models, 357Head-first/head-last languages, 42Head of phrase, 42Hearing impairment, see DeafnessHeave-ho theory, 52Hebbian learning, 489Hebrew, 138, 169, 210Hemidecortication, 73–74, 495

Hemisphere dominance, 53–54,68, 72–75, 158

Here-and-now model, 385Hesitations, 432–435Heterographic homophones, 199,

495HIC (High Interactional Content),

365Hidden units, 488, 495Holophrastic speech, 136Homographs, 199, 218, 495Homonyms, 495Homophones, 191, 199, 216, 418,

495Honey bees, 54 –55Horizontal information flow, 424HTR, 221Hungarian, 7Hybridization, 339Hypothesis, 4

I-language (internalisedlanguage), 35

Iconic gestures, 434Ideas, 434Identification procedures, 330–

331Identification semantics

hypothesis, 346Ideographic languages, 210, 218Idioms, 341–342, 495IFA, 281IG, 442IL, 442Illocutionary force, 454, 455Imageability, 177, 218, 221, 223,

495Imitation, 106Immersion method, 159, 161Implicatures, 457, 495Importance, 365–366Indirect object, 37Indirect speech acts, 456Indo-European languages, 7Induction, 115Infarct, 495Inference, 361, 368–373, 454–

458, 495Infinitive, 495Inflection, 6, 495Inflectional morphology, 5–6, 495Information flow, 420, 423, 424Information theory, 10Informational load, 376–377Informationally encapsulated, 463Informative signals, 54Inhibition, 14, 177, 486, 495

Initial contact, 264–265Innateness, 22–23, 110–118,

136–137, 350, 480Inner speech, 22, 217–218, 465,

495Input account, 344Instance theories, 337–338, 339Instantiation principle, 338Integration, 203–204, 265Intension, 324Interaction, 262, 266, 288, 290,

301–308, 420Interactive activation and

competition (IAC) model,197–198, 485–487

Interactive activation models,197–198, 424–425, 426,480, 485–487

Interactive direct access models,192–193, 195–199

Interactive processing, 21, 262,480

Intercorrelated features, 332,355–356

Interference hypothesis, 416Interlopers, 416International Phonetic Alphabet

(IPA), 29, 30Interpretative context, 267Interrogative sentences, 40Intersubjectivity, 129Intertranslatability, 90Intonation, 120Intransitive verbs, 37, 38, 495Inuit, 90, 91Invariance hypothesis, 73Irregular forms, 6Irregular words, 211, 220–221Irreversible determinism, 73Irreversible passives, 293ISA link, 325Isabelle, 78Isolated children, 76–78Isolation point, 265Italian, 7, 92, 98, 111, 210, 226Iteration, 39

James IV, 76Japanese, 28, 33, 42, 120, 210,

226Jargon aphasia, 439–440JB, 473JBN, 445JBR, 346, 347, 348JCU, 441–442Jim, 76JJ, 348

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM595


596 SUBJECT INDEX

JM, 252Joint attention, 83, 129JV, 344

Kana, 210, 226Kanji, 210, 218, 226, 227Kannada, 144Kanzi (chimp), 64, 65, 66KC, 440KE (family), 114–115KE (patient), 427Kernel sentences, 40, 292Keywords, 161Kikuyu, 120Kinship terms, 328Koko (gorilla), 61Korean, 113KR, 349KU, 223

L1 and L2, 154, 495Labile stage, 170Labiodentals, 32Lana (chimp), 61, 63Language

biological basis, 67–79before birth, 119–120change, 7–8cognitive basis, 79–82defined, 5, 55described, 27–47design features, 55–57development, 72–79, 84–87,

103–151, 480, 482differences, 8, 19–20functions, 3– 4, 9mixing, 154origin and evolution, 7, 51–54production, 397–450social aspects, 76, 82–84, 129switching, 155system structure, 463– 476,

482–483and thought, 87–98of thought, 288units, 6

Language acquisition device(LAD), 67, 110 –118, 483,495

Language acquisition socializationsystem (LASS), 83

Language bioprogram hypothesis,114

Language-specific processing, 24,480

Larynx, 29Last resort strategy, 339

Late bilingualism, 154Late closure, 298, 307–308Late-syntax theories, 143Latent semantic analysis (LSA),

356–357, 495Laura, 81–82Layering, 453, 456LB, 235Learnability theory, 116Learning difficulties, 81–82Learning rule, 485Learning theory, 106–108, 116Lemmas, 412, 495

selection, 412, 418Lesion studies, 14–16, 68, 73Less-is-more-theory, 79, 118Levels of processing, 11,

19–20Lexemes, 412, 495Lexical, 495Lexical access, 167–168, 265,

495Lexical ambiguity, 199–208Lexical bias, 423Lexical boost, 406Lexical-category ambiguity, 311–

313Lexical causatives, 333–334Lexical decision, 170, 173, 176,

178, 181–182, 183–184, 186,188, 480

Lexical development, 125–136,144 –145

Lexical entrainment, 457Lexical guidance, 300Lexical hermits, 211Lexical identification shift, 262–

263Lexical instance models, 192Lexical neighborhoods, 272Lexical principles, 127–129Lexical route, 211Lexical selection, 264 –265, 412Lexical-semantic anomia, 441–

442Lexicalization, 398, 412–428,

495feedback, 423–424, 427–428interaction, 420–424interactive activation model,

424–425stages, 412– 420time course, 420–423

Lexicons, 7, 71, 322, 495bilingual, 155–157number of, 465– 471

Lexigrams, 61, 62, 64, 65

LIC (Low Interactional Content),365

Limbus tracking, 168Linear-bounded automatons,

44Linguistic determinism,

89–90Linguistic feedback hypothesis,

109Linguistic relativism, 90Linguistic universals, 112–114Linguistics, 4, 10Linking rules, 116, 136Lip-reading, 281, 462Liquids, 33Listening for mispronunciations

task, 270–271Literacy, 168, 243–244Literal language, 340Living-nonliving dissociation,

346–350Local interaction, 421Localization of function, 14, 15Location, 288, 495Locational coherence, 362Locutionary force, 454, 455Logical inferences, 369Logistic function, 488Logogen model, 195–196, 198Logographic languages, 210Longitudinal studies, 105Look-and-say method, 247, 248Loulis (chimp), 60Low bigram frequency, 214LT, 473LW, 223

Macroplanning, 398, 432, 435Macrorules, 386Macrostructure, 386“Magic moment”, 167Main clause, 495Malapropisms, 413, 495Mandarin, 97Manner of articulation, 33, 495Mapping hypothesis, 316–317Mapping problem, 127, 129Marking, 409Masking, 172–173Mass nouns, 130Master file, 193Maturation, 72, 111, 112, 495Maturational state hypothesis,

75–76, 79MC, 222McGurk effect, 462MD, 346

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM596


SUBJECT INDEX 597

Mean length of utterance (MLU),145

Meaningfrequency, 203–205postulates, 333and sound, 216–217, 218–219

Meaning-first view, 137Meaning through syntax (MTS),

311Mediated priming, 182, 190,

420–421, 495MEG (magnetoencephalography),

17Melanie-Jane, 235–236, 252Memory

for color, 94–96inferences, 368–373text, 363–368see also Episodic memory;

Semantic memory; Short-term memory; Workingmemory

Memory organization packets(MOPs), 383–384

Mental dictionary, see LexiconMental encyclopedia, 322Mental grammar, 71Mental models, 384–386MERGE, 280Message-level processes, 397Meta-analysis, 495Metalinguistic knowledge,

243Metaphor, 340–342, 496Metrical segmentation strategy,

259MH, 185, 469Michelangelo, 347Microplanning, 398, 432, 435Microstructure, 386Milliseconds, 496Mini-theories, 338, 339Minimal attachment, 298Minimal pairs, 29, 496Minimalist hypothesis, 370Minimalist program, 35, 42–43Mirror neurons, 54Misreading repetition blindness,

190Mixed substitutions, 423MK, 442, 468ML, 222Modality-specific anomia, 344Modality-specific content

hypothesis, 345Modality-specific context

hypothesis, 345

Modality-specific formathypothesis, 345

Model-theoretic semantics, 325Models, 4, 23Modifiers, 42, 496Modularity, 19, 20–22, 297,

301, 463–465, 480,496

Modules, 21Mohawk, 45Monitor hypothesis, 160Monitor model, 159–160Monkeys, 55, 66–67, 122Monologues, 80Monosyllabic, 33, 496Morphemes, 5, 6, 496

stranding, 403Morphing, 409Morphology, 5, 191–192, 410–

412, 496MOSAIC, 139Motherese, see Child-directed

speech (CDS)Motor theory, 267–268Movement, 97Moving masks, 169MP, 221, 234, 344MRI (magnetic resonance

imaging), 17MS, 473Multiple-access model, 200Multiple components, 476Multiple-levels model, 228Multiple meanings, 199Multiple senses, 199Multiple-stores hypothesis,

344Mutual-exclusivity assumption,

128MV, 222

N-statistic, 175–176N3C principle, 129N400, 17, 189, 272, 301Naming errors, 401Naming tasks, 170, 173, 176,

178, 181–182, 183–184,186, 188

Nasal sounds, 33Nativism, 79, 117, 496Natural kind, 325, 496Natural order in acquisition

hypothesis, 159–160Natural selection, 52Nature–nurture debate, 105Navaho, 91NC, 443, 444

Negative evidence, 83, 107–108,116

Negative sentences, 40Neighborhood density, 272Neighborhood effects, 175–176,

184, 213–215, 272Neighborhood size, 176Neologisms, 8, 439–440, 496Network architecture, 485Neural networks, 13, 485Neuroimaging, 16–19, 68, 146Neuromodulatory systems, 343Neuropsychology, 4, 14, 16

cognitive neuropsychology, 14–16, 480, 481, 482

connectionist neuropsychology,481, 482

discourse processing, 390–392dissociation, 22lexicons, 467–471reading, 220–227semantics, 342–353speech production, 435–446speech recognition, 281–282text processing, 390–392writing, 448

Neuroscience, 482cognitive neuroscience, 14parsing, 315–319

New nodes, 296NI, 224Nim Chimpsky, 61, 66Nodes, 38Non-associative semantic priming,

185, 186Noncognate translation

equivalents, 162Non-fluent aphasia, 436Non-linguistic ambiguity, 460Non-nutritive sucking, 119Non-plan-internal errors, 404Non-semantic reading, 225, 228,

239Non-structural context, 266Non-terminal elements, 36Nonwords, 176, 211, 496

processing, 212–214Noun-bar units, 42Noun phrases, 37, 496Nouns, 36, 130, 496

acquisition, 126, 127–128,135

combinatorial history,339–340

Novel name–nameless categoryprinciple, 129

Nucleus, 33

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM597


598 SUBJECT INDEX

Number, 405, 496Number-agreement errors, 407–

409Number systems, 93–94

O, Dr., 469Object, 37, 38, 83, 496Object permanence, 79, 80–81,

85Object recognition, 184 –185,

352–353Obligatory decomposition

hypothesis, 191Obligatory transformations, 292Off-line, 496On-line, 496Onomatopoeia, 51–52Onset, 33, 496Ontogenetic roots, 88Open-class words, see Content

wordsOpen discourse roles, 371–372Open words, 140Optic aphasia, 344, 345Optimality Theory, 43Optional transformations, 292Ordered-access model, 200Orthographic error scores, 231Orthographic priming, 177Orthographic relatedness, 177Orthography/orthographic, 7, 496Orthotactic, 496Orton-Gillingham-Stillman

multisensory method, 254Ostensive model, 127, 496OUCH, 345Output simplification, 125Over-extensions, 131–134, 496Over-regularization errors, 220–

221Overlap of processing, 21, 420,

480

P300, 17P600, 301Palatal sounds, 32Parafovea, 169, 297Paralexia, 496Parallel autonomous models, 289Parallel coding, 212Parallel distributed processing

(PDP), 13, 485Parallel function, 374Parallel transmission, 259Parameters/parameter setting, 41,

42, 111–112, 496Paraphasias, 439–440, 443, 496

Parapraxes, 399Parkinson’s disease, 69–70,

146Parrots, 56–57PARRY, 12Parsing, 287–320, 496

models, 298–299, 308–311neuroscience, 315–319psycholinguistics, 294–298units, 294 –295working memory, 474–476

Partial activation hypothesis, 416–417

Partial alphabetic reading phase,242

Participles, 290, 496Passives, 38, 149, 293Passivization, 40Past tense acquisition, 145–148Patients, 136, 496Pattern masking, 172, 176Pauses, 432–435PC, 346Perception without awareness,

172–173Perceptual code, 343Performance, 34–35, 496Performative verbs, 454Peripheral dyslexias, 220Periphery, 169Perlocutionary force, 454, 455PET (positron emission

tomography), 18, 146, 350,415

Pheromones, 54“Phoenicians”, 251Phoenix (dolphin), 58Phonemes, 28–29, 496

blending, 222constancy, 243monitoring, 200–201, 261–262restoration, 263–264

Phones, 28–29Phonetic encoding, 430Phonetics, 5, 28, 29, 496Phonic mediation theory, 448Phonic method, 247–248Phonological anomia, 442Phonological attractors, 427Phonological awareness, 243–245,

497Phonological blocking, 417Phonological buffers, 472–474Phonological deficit model, 254Phonological development, 75,

76, 120–125, 244Phonological dysgraphia, 448

Phonological dyslexia, 221–223,228, 235–236, 497

Phonological encoding, 398, 412,415, 428–432

Phonological error scores, 231Phonological facilitation, 404Phonological form selection, 412Phonological impairment

hypothesis, 235–236Phonological input store, 468Phonological loop, 471, 472, 474,

475Phonological mediation, 212,

216–217Phonological output store, 468Phonological priming, 217Phonological recoding, 211, 217Phonological transparency, 192Phonological word, 431Phonology, 5, 28, 29, 497Phonotactic cues, 121–122, 497Phrase marker/structure, 39Phrase-structure grammar, 35–40Phrase-structure rules, 35–36, 39Phrases, 37, 497Physical modularity, 22Picture naming, 184

in dementia, 351–353Picture–word interference, 157,

414–415Pidgins, 114, 497Pigeons, 65Piraha, 67, 91Pitch, 33, 34Pivot grammar, 140Pivot words, 140Place of articulation, 31, 32, 497Planning units, 433Planum temporale, 250, 251PMSP model, 232–233Polysemous words, 199, 497Possible-word constraint, 259Post-access effects, 181Postalveolar sounds, 32Postlexical code, 261–262Poverty of the stimulus, 108,

115PQ4R method, 390PR, 448Pragmatic inferences, 369Pragmatics, 5, 453–462, 497Pre-alphabetic phase, 242Pre-intellectual stage, 88Pre-speech, 84Predicate, 37, 497Predicate calculus, 323Predictive validity, 180

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM598


SUBJECT INDEX 599

Preferential looking technique,104

Prefixes, 403, 497Prelexical code, 261–262Preliminary phrase packager

(PPP), 297–298Premackese, 60–61Preoperational stage, 79Prepositional phrases, 309, 497Prepositions, 36, 497Presupposition, 372Primes, 171, 177Priming, 13–14, 171, 177, 459,

497PRIMIR model, 122–123Principle of Economy, 42Principles and parameters theory,

41–42Print-to-sound conversion, 222Prior knowledge, 366–368Privileged information, 457Pro-drop parameter, 111Probabilistic feature model, 330–

331Probabilistic models, 24Processing in cascade, 420Processing modularity, 22, 297,

301Production systems, 379Progressive aphasia, 444Pronounceable nonwords, 211Pronouns, 36, 376, 497Pronunciation, 29Pronunciation neighborhoods, 215Property, 339

inheritance, 326Proportion effect, 180Propositional network models,

379–380Propositions, 379, 497Prosody, 120, 122, 497Prosopagnosia, 185Proto-Indo-European, 7Protolanguage, 52, 67Prototype hypothesis, 133–134Prototypes, 134, 335–337, 497Prototypicality effects, 134, 327PS, 217, 348, 349Pseudohomophone effect, 212–

213, 497Pseudowords, 176, 211, 497PSMP model, 234–235Psycholinguistic grain size theory,

246Psycholinguistics, 4, 9–14, 19–

24, 294–298, 480–481, 483–484, 497

Pure definitional negatives(PDNs), 333

Pure word deafness, 281, 468Purkinje system, 168Push-down automatons, 44PW, 427, 445, 469Pyramids and palm trees test, 441

Queen’s English, 29

Ramu, 77Random noise masking, 172Rank hypothesis, 194Rationalism, 105, 106RB, 444RC, 350RE, 252, 475Reaction times, 11, 13, 170–171,

180Readability, 387–388Reading, 209–240

eye movements, 168–170, 188–189, 216, 219

improving, 254–255, 389–390learning, 241–248models, 211–212, 227–239neuropsychology, 220–227normal processes, 212–220silently, 217–218span, 389, 471speed, 219teaching methods, 247–248

Reanalysis, 309Received Pronunciation (RP), 29,

33Recency, 378Recent-filler strategy, 314Receptive, 497Receptive aphasia, 391Recipient, 287, 288Recognition point, 265, 269Recurrent networks, 278–279,

429, 489, 497Recursion, 39, 67Reduced relative clauses, 290–

291, 497Reduplicated babble, 123Reference, 373–378, 497Referent, 497Referential coherence, 362Referential processing, 362Referential theory, 303–304, 324–

325Refractoriness, 343Refractory period, 343, 497Regressions, 170, 216Regular words, 211, 214, 215

Reinforcement, 10Relatedness effect, 327Relative clauses, 39, 290, 497Relative pronouns, 39Relativism, 22Reordered access model, 203–

204, 205Repair, 401Repetition, 80Repetition blindness, 162, 190Repetition disorder, 472Repetition priming, 155–156,

176, 414, 497Representational modularity, 301Reproduction conduction aphasia,

446Requests, 341–342Resolution, 372, 374Response, 10Response bias, 183Restricted search hypothesis, 377Restructured knowledge, 159Resyllabification, 431–432Reverse flow of information, 420,

423Reversible passives, 293Rewrite rules, 36, 40RG, 448Rhyme, 242Rhythm, 34RIA model, 428Rich interpretation, 140–141Rico (dog), 57Right association, 296Right-hemisphere hypothesis,

224–225Rimes, 33, 242–243, 497RM, 346Romance languages, 7Rote repetition, 161RSVP (rapid serial visual

presentation) technique,189–190

Rules, 23, 24, 34, 140, 480Russian, 149

Saccades, 169, 170, 498Sapir-Whorf hypothesis, 9, 89–98Sarah (chimp), 58–59, 60–61, 63Sausage machine model, 297–

298Scan-copier mechanism, 428–

429, 440Scenes, 383Schema-based theories, 382–384Schemas, 336, 382, 498Scripts, 382–384, 498

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM599


600 SUBJECT INDEX

Second language acquisition, 75–76, 158–162

Segmentation, 259–260, 498Segmentation markers, 386Selection phase, 265Selection restrictions, 329Selective access model, 203Selective adaptation, 264Semantic, 498Semantic assimilation theory, 137,

138Semantic bias, 301–302, 313Semantic bootstrapping, 136, 137Semantic categorization tasks,

171, 216–217Semantic complexity, 135Semantic context, 266–267Semantic dementia, 344, 350–353Semantic development, 126–136Semantic feature hypothesis, 133Semantic features, 328–335, 498Semantic glue hypothesis, 235Semantic memory, 322, 343–346,

350–353, 498Semantic microfeatures, 353–355Semantic networks, 325–328Semantic opaqueness, 192Semantic paralexias, 223, 237Semantic-pragmatic disorder, 84,

391Semantic priming, 14, 177, 178–

181, 185–187, 190, 498Semantic primitives, 328–329Semantic processing, 362Semantic reactivation, 314Semantic roles, 38Semantic sets, 198Semantic transparency, 192Semantic verb class hypothesis,

142Semantics, 5, 321–360, 498

connectionism, 353–358neuropsychology, 342–353

Semantics comes first, 310Sememes, 236Semi-vowels, 33Sense, 324Sensitive period hypothesis, 78–

79Sensorimotor period, 79Sensory sets, 198Sensory–functional theory, 348–

350Sentence, 27, 40, 498

analysis, 500hierarchy, 36parsing, see Parsing

structure supervisor, 297–298surface structure, 295–296verification, 325–326, 330–331,

332Separate-store models, 155–156Sequential bilingualism, 154,

498Serbo-Croat, 98, 210, 226Serial autonomous models, 289Serial search model, 192, 193–

195, 198Sex differences, 71–72Sexist language, 98Shadowing, 270Sherman (chimp), 61, 66Shona, 94Short-term memory (STM), 471–

476, 498aphasia, 446impairments, 391–392

SHORTLIST, 279–280SHRDLU, 12, 13Sign language, 59–60, 61, 68, 76,

86–87, 109, 112, 114, 124Signal detection, 263Signals, 54Silent reading, 217–218Similarity effects, 423–424Simile, 340Simultaneous bilingualism, 154,

498Situation models, 384, 386–387SJD, 418SL, 36, 347SLIP, 401SM model, 230–232Social aspects, 76, 82–84, 129Social interactionism, 83, 84Songbirds, 72Sound

contrasts, 124and meaning, 216–217, 218–

219spectrograms, 27speech sounds, 27–34

Source, 498SP, 255Span, 498Spanish, 7, 226, 246, 307Spatial information, 384–385Spatial terms, 96Specific language impairment

(SLI), 114 –115, 146, 391,498

Speech acts, 454–456, 498Speech apraxia, 435Speech dysfluencies, 432–435

Speech errors, 399–404, 411,413–414, 427

Speech perception, 71, 257, 267in infancy, 120–123

Speech production, 29, 69, 397–450

hesitations, 432–435lexicalization, 412–428models, 401–404, 425, 426,

429neuropsychology, 435–446phonological encoding, 428–

432syntactic planning, 404–412

Speech recognition, 257–283context, 258–259, 261, 262–

264, 266–267, 271–272models, 267–281neuropsychology, 281–282speed, 258, 261time course, 264–265

Speech sounds, 27–34Speed reading, 219Speed–accuracy trade-offs, 170Spelling, 243, 248–249Spelling-to-sound correspondence,

211Spoonerisms, 399, 498Spreading activation, 13, 190,

327–328, 379, 425, 485ST, 473Stage 1 speech, 149Standard theory, 41Stem, 498Stimulus, 10

degradation, 172Stimulus–onset asynchrony

(SOA), 171, 498STM conduction aphasia, 446Stochastic output, 275, 498Stops, 33Stories, 361, 362–363Story grammars, 380–381Stranding, 400, 403Stress, 33–34, 120Stress-based segmentation, 259–

260Stressed-timed languages, 34Stroke, 498Stroop task, 178Structural ambiguity, 289–291,

298–313Structural context, 266Structural priming, 143, 405–406Structuralism, 10Subcategorization frame, 287Subject, 37, 38, 287, 498

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM600


SUBJECT INDEX 601

Sublexical route, 211, 498Subliminal perception, 171, 172Submersion method, 159Subordinate clauses, 498Subordinates, 323Substantive universals, 112Subtraction method, 19Sucking habituation, 74, 104, 498Suffixes, 403, 498Summation model, 229Super-additive impairment, 345–

346Superordinates, 323Suppression, 389, 498Suprasegmental features, 33Surface dysgraphia, 448Surface dyslexia, 220–221, 228,

233–234, 498Surface structure (s-structure), 41Syllabary, 430Syllabic scripts, 210Syllabification, 429–430Syllable-based segmentation, 260Syllable-timed languages, 34Syllables, 6, 33, 498

monitoring tasks, 260onset and rime, 33role in phonological encoding,

431Symbols, 5Syndrome, 68, 499Synecdoche, 340Synonymy, 321Syntactic bootstrapping, 130Syntactic category

ambiguity, 311–313learning, 136–139

Syntactic comprehension, 149Syntactic development, 75–76,

78, 136–149Syntactic persistence, 405–406Syntactic planning, 398, 401–402,

404 – 412chunking/incremental, 409– 410

Syntactic priming, 364, 405– 407Syntactic processing, 288, 299–

308Syntactic roles, 136, 499, 500Syntactically ambiguous

sentences, 38Syntax, 5, 27, 34–43, 56, 288,

499autonomy of, 292, 293

Synthetic phonics, 248

Tabula rasa, 105Tachistoscopes, 171, 499

Targets, 171, 177Taxonomic constraint, 128TB, 475Telegraphic speech, 103, 136,

140, 499Templates, 12, 258, 267

matching, 267Temporal coherence, 362Temporal discreteness, 420Temporal gyrus, 69, 71Tense, 405, 499Terminal elements, 36Text, 361

memory for, 363–368Text processing

models, 378–388neuropsychology, 390–392

Thai, 28Thematic organization points

(TOPs), 383Thematic relations, 339Thematic roles, 136, 287, 288,

499, 500Theme, 38, 287, 288, 499Theory, 4Theory of mind, 82Theory theories, 338, 339Theta role, see Thematic rolesThought, 87–98, 288Time, 97Tip-of-the-tongue (TOT), 416 –

418, 433, 499TMS (transcranical magnetic

stimulation), 18TOB, 343Tone languages, 34Tongue-twisters, 217, 218Top-down processing, 21, 296,

479–480, 499TRACE, 273 –278, 280, 485Trace-deletion hypothesis, 316Traces, 42, 313–315Transcortical aphasia, 446, 473,

499Transformation, 40, 45, 292,

499Transformational grammar,

10, 40, 292 – 294, 499Transition relevance place, 458Transitive verbs, 37, 499Translating between languages,

156 –157Tree diagrams, 38–39Trevor, 249Triangle model, 230–232, 235,

238Truth-theoretic semantics, 325

Truth value, 379TU, 349Tuning hypothesis, 308Turing machines, 44, 45Turn-taking, 83–84, 458–459Twin studies, 117, 252Two-word grammars, 139–141TY, 226Type 1/Type 2 grammars, 44Type 3 language, 44, 116Type-B spelling disorder, 251Tzeltal, 96

U-shaped development, 108, 141,146, 147, 159

Ultra-cognitive neuropsychology,15

Unaspirated, 28, 499Unbalanced ambiguity, 203Unbounded dependencies, 314–

315Under-extensions, 132–134Unfilled pauses, 432–433Unilingual, 499Unimodal store hypothesis, 343Uniqueness point, 265, 269Unit, 499Universal grammar, 111, 112–

114, 499Unrestricted-race model, 309, 310Unrestricted search hypothesis,

377Unvoiced consonants, 32, 499Uralic languages, 7

Variable-choice one-stage models,309

Variegated babble, 123Vegetative sounds, 103Velars, 32Velum, 32Verb-bar units, 42Verbal code, 343Verbal short-term memory, 474Verbs, 36, 499

acquisition, 142argument structure, 141–144,

287, 499, 500auxiliary, 40, 492bias, 305causative, 144copulative, 493dative, 493direct object, 305ditransitive, 37inflections, 145–148intransitive, 37, 38, 495

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM601


602 SUBJECT INDEX

islands, 142NP1/NP2, 375performative, 454sentence complement, 305transitive, 37, 499

Verification models, 198Vertical information, 424Viki (chimp), 58, 59Visual deficits, 250, 255Visual errors, 222, 234Visual impairment, 84–86,

87Visual word form area, 185Visual word recognition,

167–208Visual world, 295, 306–307, 405,

414, 460–462, 483Visuo-spatial sketchpad, 471Vocabulary

differentiation, 90–91learning, 474

Vocal cords, 32Vocal play, 103Vocal tract, 29, 59Voice onset time (VOT), 32, 260–

261, 262, 499Voiceless consonants, 32

Voiceless glottal fricative, 32Voicing, 32, 499Vowels, 29, 31, 499

processing, 281

Washoe (chimp), 59–60, 62–63

WB, 221–222WEAVER++ model, 428, 429–

430, 431Weights, 486Welsh, 93–94Wernicke-Geschwind model, 14,

69–71Wernicke’s aphasia, 69, 391,

436 –437, 445–446, 499Wernicke’s area, 14, 15, 68–69WH-words, 60Whale song, 55Whole-object assumption, 127–

128Whole word method, 247, 248Wickelfeatures, 231, 232–233Wild Boy of Aveyron, 77Williams syndrome, 82, 146WLP, 225WMA, 469

Wolf children, 77Word, 6, 499

association, 157association norms, 185–186class, 36, 499defined, 6exchange errors, 403first words, 125–131frequency, 173–175, 265length, 175order, 113, 404processing, 214–219recognition, 167–208, 264 –265substitutions, 404, 413–414superiority effect, 197

Word meaning deafness, 281, 468Working memory, 22, 82, 387–

388, 389, 391, 471–472,474–476, 499

Writing, 209–211, 447–448WT, 225

X-bar syntax, 42X-rays, 16, 17

Yerkish, 61, 63, 64YOT, 347–348

9781841693828_6_ind.02PM5 9/21/07, 10:19 AM602


Date post:	04-Feb-2022
Category:	Documents
Upload:	others
View:	3 times
Download:	0 times

THE PSYCHOLOGY OF LANGUAGE

Documents