The Science of Reading: A Handbook

Contents

List of Contributors

Preface

Acknowledgments

PART I Word Recognition Processes in Reading

Editorial Part I

1 Modeling Reading: The Dual-Route Approach

In the Beginning…

Lexical and Nonlexical Reading Routes

Phenomena Explained via the Dual-Route Model

Computational Modeling of Reading

The Dual-Route Cascaded (DRC) Model

What the DRC Model Can Explain

Conclusions

2 Connectionist Approaches to Reading

Principles of Connectionist Modeling

Realist Versus Fundamentalist Approaches

Connectionist Modeling of Reading

Conclusion

3 Visual Word Recognition: Theories and Findings

Historical Context

The Models

The Orthographic-Phonological Interaction

Interactions with Semantics

Two Other Emerging Issues

Parting Thoughts

4 The Question of Phonology and Reading

How Evidence Fuels the Controversy

Giving Up Ether

Spelling and Phonology in an Interactive System

Remainders

Summary and Conclusions

5 Eye Movements During Reading

The Basic Characteristics of Eye Movements During Reading

Reading Skill and Eye Movements

Eye Movements and Measures of Processing Time in Reading

Basic Issues Regarding Eye Movements in Reading

Recent Trends and Current Issues

Models of Eye Movement Control in Reading

Summary

PART II Learning to Read and Spell

Editorial Part II

6 Theories of Learning to Read

Two Background Issues

Frameworks for Identifying the Work of Theories of Learning to Read

Learning Revisited

Applications of Theory

Concluding Remarks

7 Writing Systems and Spelling Development

Principles of Writing Systems

Learning to Spell

8 Development of Sight Word Reading: Phases and Findings

Ways to Assess Sight Word Reading

Memory Processes That Enable Sight Word Reading

Developmental Theories

Synopsis of the Theories

Phase Theory of Sight Word Reading

Transition from the Partial Alphabetic to Full Alphabetic Phase

Development of Automaticity, Speed, and Unitization

Concluding Comments

9 Predicting Individual Differences in Learning to Read

Methodological Issues

Key Predictors of Early Reading Ability

Conclusions

10 Social Correlates of Emergent Literacy

Development of Emergent Literacy

Early Childhood Education

Socioeconomic Status

Family Beliefs and Values

Home Language Stimulation

Home Literacy Environment

Summary and Conclusions

11 Literacy and Cognitive Change

Literacy, Schooling, and Education

The Impact of Literacy on Nonlinguistic Capacities

The Impact of Literacy on Linguistic Capacities

Conclusions

PART III Reading Comprehension

Editorial Part III

12 Comprehension

Processes Underlying Text Comprehension

Textbase Formation

The Situation Model

Summary

13 The Acquisition of Reading Comprehension Skill

Introduction: Simple Ideas about Reading Comprehension

A Framework for Comprehension

Higher-Level Factors in Comprehension

The Linguistic-Conceptual Machinery for Comprehension

Word Identification, Decoding, and Phonological Awareness

Comprehension Instruction

Conclusion: A More General View of Comprehension Development

14 Children’s Reading Comprehension Difficulties

Specific Deficits in Reading Comprehension?

What Causes Poor Reading Comprehension?

Summary and Conclusions

PART IV Reading in Different Languages

Editorial Part IV

15 Orthographic Systems and Skilled Word Recognition Processes in Reading

Overview of Writing Systems

Models of Skilled Reading

Orthographic Depth and How Phonology Is Represented in Different Orthographies: The Case of English, Hebrew, and Serbo-Croatian

Orthographic Depth and Visual Word Recognition

Empirical Evidence for the Orthographic Depth Hypothesis

Languages with Two Writing Systems: The Cases of Serbo-Croatian, Korean, and Japanese

Summary and Conclusions

16 Early Reading Development in European Orthographies

Introduction

Causation

Language Effects

Linguistic Units

Cross-Language Differences in the Development of Linguistic Awareness

Models of Literacy Acquisition

Language Effects on Orthographic Development

Conclusions

17 Learning to Read in Chinese

The Chinese Writing System

The Teaching of Reading in China

How Children Read Compound Characters

Phonological Awareness and Learning to Read Chinese

Learning to Read Chinese As a Second Language

Developmental Dyslexia

Conclusions

18 The Nature and Causes of Dyslexia in Different Languages

Characteristics of Different Writing Systems

Dyslexia among English Speakers: A Point of Reference

Dyslexia in Non-English Languages with Alphabetic Orthographies

Dyslexia in the Logographic Writing System of Chinese

Conclusions

PART V Disorders of Reading and Spelling

Editorial Part V

19 Developmental Dyslexia

Manifest Causes of Dyslexia: Deficiencies in Reading Subskills

Underlying Causes of Dyslexia: Cognitive Deficit Theories

Conclusions

20 Learning to Read with a Hearing Impairment

Do People with a Hearing Impairment Use Phonological Coding in Reading and Spelling?

Reading Processes

Individual Differences in the Use of Phonological Codes by People with a Hearing Impairment

Spelling Processes

Phonological Awareness

The Role of Visible Language in Reading Development of the Hearing Impaired

Cued Speech

The Linguistic Advantages for Hearing-impaired Children Born to Hearing-impaired Parents

The Use of Orthographic Coding by Children with Hearing Impairment

Neural Systems Underlying Reading in People with Hearing Impairment

Conclusions

21 Learning to Read with a Language Impairment

Models of Reading Development

Developmental Disorders of Reading

Reading Development in Children with Oral Language Impairments

Language and Reading Impairments in Children with General Learning Difficulties

Conclusions

22 Acquired Disorders of Reading

Introduction

Semantic Memory and Surface Dyslexia

Phonology and Phonological-Deep Dyslexia

Visual Processing and Pure Alexia (Letter-by-Letter Reading)

23 Spelling Disorders

The Dual-Route Model of Spelling

Differences between Reading and Spelling

Acquired Disorders of Spelling

Developmental Spelling Disorders

Differences between Acquired and Developmental Disorders

Conclusions

PART VI The Biological Bases of Reading

Editorial Part VI

24 Genetics of Dyslexia

Brief History

Behavioral Genetic Approaches

Molecular Genetic Approaches

Summary and Conclusions

25 Functional Brain Imaging Studies of Skilled Reading and Developmental Dyslexia

Functional Imaging Techniques: How PET and fMRI Work

How Can Functional Imaging Inform Cognitive Models of Reading?

The Neural Systems for Skilled Reading

Relating Anatomy to Current Cognitive Models of Reading

Functional Imaging Studies of Reading in Developmental Dyslexia

Conclusions and Future Directions

PART VII Teaching Reading

Editorial Part VII

26 Teaching Children to Read: What Do We Know about How to Do It?

Controversies about Teaching Reading

The Introduction of Science and Orthodoxy into Pedagogical Decision Making

Problems with Horse Race Studies

Back to the Reading Skirmishes

27 Recent Discoveries on Remedial Interventions for Children with Dyslexia

Defining the Target of Intervention

An Early Case Study and Other Discouraging Examples

A Recent Study with a Different Outcome for Children with Severe Reading Disabilities

Reading Gains in Other Studies of Intensive Interventions

What about the Remaining Problems in Fluency?

Additional Areas of Knowledge from Intervention Research with Older Children

A Final and Significant Remaining Gap in Our Knowledge

Conclusion

Glossary of Terms

References

Author Index

Subject Index

Blackwell Handbooks of Developmental Psychology

This outstanding series of handbooks provides a cutting-edge overview of classic research, current research and future trends in developmental psychology.

Each handbook draws together 25–30 newly commissioned chapters to provide a com­prehensive overview of a subdiscipline of developmental psychology.

The international team of contributors to each handbook has been specially chosen for its expertise and knowledge of each field.

Each handbook is introduced and contextualized by leading figures in the field, lending coherence and authority to each volume.

The Blackwell Handbooks of Developmental Psychology will provide an invaluable overview for advanced students of developmental psychology and for researchers as an authorita­tive definition of their chosen field.

Published

Blackwell Handbook of Infant Development

Edited by Gavin Bremner and Alan Fogel

Blackwell Handbook of Childhood Social Development

Edited by Peter K. Smith and Craig H. Hart

Blackwell Handbook of Childhood Cognitive Development

Edited by Usha Goswami

Blackwell Handbook of Adolescence

Edited by Gerald R. Adams and Michael D. Berzonsky

The Science of Reading: A Handbook

Edited by Margaret J. Snowling and Charles Hulme

Forthcoming

Blackwell Handbook of Early Childhood Development

Edited by Kathleen McCartney and Deborah A. Phillips

© 2005 by Blackwell Publishing Ltd

except for editorial material and organization © 2005 by Margaret J. Snowling and Charles Hulme

BLACKWELL PUBLISHING

350 Main Street, Malden, MA 02148-5020, USA

9600 Garsington Road, Oxford OX4 2DQ, UK

550 Swanston Street, Carlton, Victoria 3053, Australia

The right of Margaret J. Snowling and Charles Hulme to be identified as the Authors of the Editorial Material in this Work has been asserted in accordance with the UK Copyright, Designs, and Patents Act 1988.

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, except as permitted by the UK Copyright, Designs, and Patents Act 1988, without the prior permission of the publisher.

First published 2005 by Blackwell Publishing Ltd

1 2005

Library of Congress Cataloging-in-Publication Data

The science of reading : a handbook / edited by Margaret J. Snowling and Charles Hulme. p. cm. — (Blackwell handbooks of developmental psychology)

Includes bibliographical references and indexes.

ISBN 13: 978-1-4051-1488-2 (alk. paper)

ISBN-10: 1-4051-1488-6 (alk. paper)

1. Reading. 2. Reading—Research. 3. Reading, Psychology of. I. Snowling, Margaret J. II. Hulme, Charles. III. Series.

LB1050.S365 2005

428.4—dc22

2005001421

A catalogue record for this title is available from the British Library.

For further information on

Blackwell Publishing, visit our website:

www.blackwellpublishing.com

Contributors

Judith A. Bowey

School of Psychology

University of Queensland

St Lucia

Queensland 4072

Australia

email: [email protected]

Brian Byrne

School of Psychology

University of New England

Armidale NSW 2351

Australia

email: [email protected]

Markéta Caravolas

Department of Psychology

University of Liverpool

Liverpool L69 7ZA

UK

email: [email protected]

Max Coltheart

Macquarie Centre for Cognitive Science

Macquarie University

Sydney NSW 2109

Australia

email: [email protected]

Anna Maria Di Betta

Neurosciences Research Institute

School of Life and Health Sciences

Aston University

Birmingham B4 7ET

UK

email: [email protected]

Linnea C. Ehri

Graduate Center of the City University of New York

Program in Educational Psychology

CUNY Graduate Center

365 Fifth Ave.

New York, NY 10016

USA

email: [email protected]

Jack M. Fletcher Center for Academic and Reading Skills

Department of Pediatrics

University of Texas Health Science Center at Houston

7000 Fannin UCT 2478

Houston TX 77030

USA

email: [email protected]

Ram Frost

Department of Psychology

The Hebrew University

Jerusalem 91905

Israel

email: [email protected]

J. Richard Hanley

Department of Psychology

University of Essex

Wivenhoe Park

Colchester CO4 3SQ

UK

email: [email protected]

Charles Hulme

Department of Psychology

York University

York Y010 5DD

UK

email: [email protected]

Connie Juel

School of Education

Stanford University

485 Lasuen Mall

Stanford, CA 94305-3096

USA

email: [email protected]

Barbara J. Juhasz

Department of Psychology

University of Massachusetts

Amherst, MA 01003

USA

email: [email protected]

Brett Kessler

Psychology Department

Washington University in St. Louis

Campus Box 1125

One Brookings Drive

St. Louis, MO 63130-4899

USA

email: [email protected]

Walter Kintsch

Department of Psychology

University of Colorado

Boulder, CO 80309-0344

USA

email: [email protected]

Heidi Kloos

Center for Cognitive Sciences

Ohio State University

211 G Ohio Stadium East

1961 Tuttle Park Place

Columbus, OH 43210

email: [email protected]

Régine Kolinsky

UNESCOG (CP 191)

Université libre de Bruxelles

50 Av. F. D. Roosevelt

B-1050 Brussels

Belgium

email: [email protected]

Matthew A. Lambon Ralph

Department of Psychology

University of Manchester

Manchester M13 9PL

UK

email: [email protected]

Nicole Landi

Learning Research and Development Center

University of Pittsburgh

Pittsburgh, PA 15260

USA

email: [email protected]

Jacqueline Leybaert

LAPSE

Université libre de Bruxelles

50 Av. F. D. Roosevelt

B-1050 Brussels

Belgium

email: [email protected]

Christopher J. Lonigan

Department of Psychology

Florida State University

One University Way

Tallahassee, FL 32306-1270

USA

email: [email protected]

Stephen J. Lupker

Department of Psychology

University of Western Ontario

London

Ontario N6A 5C2

Canada

email: [email protected]

Eamon McCrory

Department of Psychology

Institute of Psychiatry

De Crespigny Park

London SE5 8AF

UK

email: [email protected]

José Morais

UNESCOG (CP 191)

Université libre de Bruxelles

50 Av. F. D. Roosevelt

B-1050 Brussels

Belgium

email: [email protected]

Kate Nation

Department of Experimental Psychology

University of Oxford

South Parks Road

Oxford OX1 3UD

UK

email: [email protected]

Jane Oakhill

Department of Psychology

University of Sussex

Falmer House

Brighton BN1 9RH

UK

email: [email protected]

Andrew Olson

Department of Psychology

University of Birmingham

Edgbaston

Birmingham B15 2TT

UK

email: [email protected]

Richard K. Olson

Department of Psychology

University of Colorado, UCB 345

Boulder, CO 80309

USA

email: [email protected]

Karalyn Patterson

MRC Cognition and Brain Sciences Unit

15 Chaucer Road

Cambridge CB2 2EF

UK

email: [email protected]

Bruce F. Pennington

Department of Psychology

University of Denver

2155 S. Race St.

Denver, CO 80210-4638

USA

email: [email protected]

Charles A. Perfetti

Learning Research and Development Center

University of Pittsburgh

Pittsburgh, PA 15260

USA

email: [email protected]

Beth M. Phillips

Florida Center for Reading Research

City Centre Building Suite 7250

227 North Bronough St.

Tallahassee, FL 32301

USA

email: [email protected]

David C. Plaut

Departments of Psychology and Computer Science and Center for the Neural Basis of Cognition

Carnegie Mellon University

5000 Forbes Ave.

Pittsburgh, PA 15213-3890

USA

email: [email protected]

Alexander Pollatsek

Department of Psychology

University of Massachusetts

Amherst, MA 01003

USA

email: [email protected]

Cathy J. Price

Wellcome Department of Imaging Neuroscience

University College London

Institute of Neurology

12 Queen Square

London WC1N 3BG

UK

email: [email protected]

Katherine Rawson

Kent State University

Department of Psychology

P.O. Box 5190

Kent, OH 44242-0001

USA

email: [email protected]

Keith Rayner

Department of Psychology

University of Massachusetts

Amherst, MA 01003

USA

email: [email protected]

Cristina Romani

Department of Psychology

Aston University

Aston Triangle

Birmingham B4 7ET

UK

email: [email protected]

Philip H. K. Seymour

Department of Psychology

University of Dundee

Dundee DD1 4HN

UK

email: [email protected]

Catherine E. Snow

Harvard Graduate School of Education

Larsen 3

Cambridge, MA 02138

USA

email: [email protected]

Margaret J. Snowling

Department of Psychology

York University

York YO10 5DD

UK

email: [email protected]

Joseph K. Torgesen

Florida Center for Reading Research at Florida State University

227 N. Bronough St., Suite 7250

Tallahassee, FL 32301

USA

email: [email protected]

Rebecca Treiman

Psychology Department

Washington University in St. Louis

Campus Box 1125

One Brookings Drive

St. Louis, MO 63130-4899

USA

email: [email protected]

Guy C. Van Orden

Department of Psychology

Arizona State University

Tempe, AZ 85287-1104

USA

email: [email protected]

Frank R. Vellutino

Department of Psychology

State University of New York

1535 Western Avenue

Albany, NY 12203

USA

email: [email protected]

Preface

To completely analyse what we do when we read would almost be the acme of the psychologist’s achievements, for it would be to describe very many of the most intricate workings of the human mind

(Huey, 1968).

The science of reading is mature and healthy as the contributions to this volume make clear. Together they provide an assessment of how far we have come in meeting the chal­lenge laid down by Huey more than a century ago. Different chapters illustrate how some old issues remain alive, how new questions have been raised and how some problems have been solved. Many of the issues discussed here would undoubtedly have been familiar to Huey. Discussions of how skilled readers recognize printed words rapidly, of how eye movements in reading are controlled, the factors limiting reading comprehension, and arguments about how best to teach reading, all featured prominently in early studies of reading. These are important topics and ones that remain current, as several chapters in this book attest. There is little doubt that the technical advances made in many of these areas would be a source of pleasure to Huey and his contemporaries in the field of reading research. On the other hand, a number of issues dealt with in this book would probably have seemed totally foreign to people in the field of reading a century ago. For example, studies imaging the brain while it reads, studies examining the molecular genetics of reading disorders, and computational models of different aspects of the reading process would have seemed like science fiction a hundred years ago.

This Handbook provides a state-of-the-art overview of scientific studies of reading. The book is divided into seven sections. Part I deals with word recognition processes and is concerned largely with theories developed in studies of fluent adult reading. Such the­ories have heavily influenced (and been influenced by) studies of reading development, which are dealt with in Part II. Efficient word recognition processes are necessary, but not sufficient, for reading comprehension (Gough & Tunmer, 1986) and the chapters in Part III go beyond single word processing to consider reading comprehension processes in both adults and children, with an emphasis on the problems that may be encountered in children learning to comprehend what they read. Studies of reading and reading devel­opment have until recently been concerned only with reading English. Gough and Hillinger (1980) suggested that learning to read was an unnatural act; if that is true there is growing evidence that learning to read in English is a particularly unnatural act! Part IV of the book brings together work exploring how reading and reading develop­ment may differ across languages. This section highlights a number of issues and con­fronts the question of whether we can hope for a universal cognitive theory of reading and reading development – such a hope seems closer than some may have believed.

One justification for much research in psychology is that it helps us to understand, and in turn to prevent and to treat, disorders in psychological processes. The chapters in Part V look at our understanding of developmental and acquired disorders of reading and spelling. An important question here is the extent to which common forms of explana­tion may be valid for both acquired and developmental disorders. Part VI of the book examines the biological substrates of reading. It brings together work on brain imaging, which has revealed with new clarity the brain regions involved in different aspects of reading, with work on the genetic basis of dyslexia. The final section of the book, Part VII, examines how scientific studies of reading can contribute to improving the teaching of reading both in normally developing children and children with dyslexia.

We hope that the overviews of research presented here will be of value to psycholo­gists and educationalists studying reading, their students, and to practitioners and others who want to find out about the current status of The Science of Reading.

Acknowledgments

We would like to thank Mark Seidenberg who played an invaluable role in helping to shape the form of this book in the early stages of its development.

We have learned a great deal from editing this book and would like to thank all our contributors for their excellent chapters, which made our task so easy and pleasurable.

Maggie Snowling and Charles Hulme

PART I

Word Recognition Processes in Reading

Editorial Part I

Word recognition is the foundation of reading; all other processes are dependent on it. If word recognition processes do not operate fluently and efficiently, reading will be at best highly inefficient. The study of word recognition processes is one of the oldest areas of research in the whole of experimental psychology (Cattell, 1886). The chapters in this section of the Handbook present an overview of current theories, methods, and findings in the study of word recognition processes in reading.

What do we mean by recognition here? Recognition involves accessing information stored in memory. In the case of visual word recognition this typically involves retrieving information about a word’s spoken form and meaning from its printed form. The first two chapters, by Coltheart and Plaut, outline the two most influential theoretical frameworks for studies of visual word recognition.

Coltheart outlines the history and evolution of dual-route models of reading aloud (i.e., how the pronunciation of a printed word is generated). These dual-route models posit that there are two routes from print to speech: a lexical and nonlexical route. Broadly the lexical route involves looking up the pronunciation of a word stored in a lexicon or mental dictionary. In contrast, the nonlexical route involves translating the graphemes (letters or letter groups) into phonemes and assembling the pronunciation of a word from this sequence of phonemes. Such a process should work just as well for nonwords as for words, just so long as the word follows the spelling pattern of the language (a nonlexical reading of YACHT, will not yield the pronunciation for a kind of boat with a sail on it). This idea is embodied in an explicit computational model (the DRC model) that Coltheart describes in detail. It may be worth emphasizing that this highly influential model is a model of how adults read aloud; it is not concerned with how the knowledge allowing this to happen is acquired. A major focus of the model is how different disorders of reading aloud, which arise after brain damage in adults, can be accounted for.

Plaut gives an overview of a different class of models of reading aloud that employ connectionist architectures (models that learn to pronounce words by training associations between distributed representations of orthography and phonology). One particularly influential model of this type is the so-called triangle model (Plaut, McClelland, Seidenberg, & Patterson, 1996; Seidenberg & McClelland, 1989). This model abandons the distinction between a lexical and nonlexical procedure for translating visual words into pronunciations; instead the same mechanism is used to convert words and nonwords into pronunciations, based on patterns of connections between orthographic inputs and phonological outputs. One other critical difference between the triangle model and the DRC model is that the triangle model explicitly embodies a learning procedure and thus can be considered a model of both adult reading and reading development. It is clear that these are very different conceptions of how the mind reads single words. Both approaches deal with a wide range of evidence. Arguably, the DRC model is more successful in dealing with the detailed form of reading impairments observed after brain damage in adults, while the ability to think about development and adult performance together in the triangle model is a considerable attraction. There is no doubt that differences between these models will be a source of intense interest in the coming years.

Lupker’s chapter moves on to review a huge body of experimental evidence concerned with how adults recognize printed words. Many of these experiments investigate what is a remarkably rapid and accurate process in most adults, by measuring reaction time, or by impairing performance by using masking (preventing participants from seeing a word clearly by superimposing another stimulus immediately after the word has been presented). Any complete model of word recognition ultimately will have many phenomena from such experiments to explain. These include the fact that people perceive letters more efficiently when they are embedded in words, that high-frequency (i.e., more familiar) words are recognized easier than less familiar words, and that recognition of words is influenced by previously presented words (seeing a prior word that is related in form or meaning helps us to recognize a word that follows it). One conclusion that emerges powerfully from Lupker’s review is the need for interactive models in which activation of orthographic and phonological information reciprocally influence each other. This is an issue that Van Orden and Kloos take up in detail, presenting a wealth of evidence that converges on the idea that there is intimate and perpetual interaction between representations of orthography and phonology (spelling and sound) during the process of recognizing a printed word.

Moving on from the recognition of isolated words, Rayner, Juhasz, and Pollatsek discuss eye movements in reading. Eye movements provide a fascinating window on how word recognition processes operate in the more natural context of reading continuous text. It appears that the pattern of eye movements in reading is heavily influenced by the cognitive processes subserving both word recognition and text comprehension. The majority of words in text are directly fixated (usually somewhere in the first half of the word). For readers of English the area of text processed during a fixation (the perceptual span) is about 3 or 4 letters to the left of fixation and some 14 or 15 letters to the right of fixation. This limit seems to be a basic one determined by acuity limitations, and useful information about letter identity is extracted only from a smaller area, perhaps 7 or 8 letters to the right of the fixation point. It appears that only short, frequent, or highly predictable words are identified prior to being fixated (so that they can be skipped). However, partial information (about a word’s orthography and phonology but typically not its meaning) about the word following the fixation point often is extracted and combined with information subsequently extracted when the word is directly fixated. These studies are consistent with the view that the speed and efficiency of word recognition processes (as well as higher-level text-based processes) place fundamental constraints on how quickly even skilled readers read text.

Arguably the central question in the study of word recognition in reading is the role of phonology. All of the chapters in Part I address this issue explicitly. It appears that a consensus has been reached: phonological coding is central to word recognition, though opinions are divided on many details of how phonology is accessed and its possible importance in providing access to semantic information.

1

Modeling Reading: The Dual-Route Approach

Max Coltheart

Reading is information-processing: transforming print to speech, or print to meaning. Anyone who has successfully learned to read has acquired a mental information-processing system that can accomplish such transformations. If we are to understand reading, we will have to understand the nature of that system. What are its individual information-processing components? What are the pathways of communication between these components?

Most research on reading since 1970 has investigated reading aloud and so sought to learn about the parts of the reading system that are particularly involved in transforming print to speech. A broad theoretical consensus has been reached: whether theories are connectionist (e.g., Seidenberg & McClelland, 1989; Plaut, this volume) or nonconnectionist (e.g., Coltheart, Curtis, Atkins & Haller, 1993), it is agreed that within the reading system there are two different procedures accomplishing this transformation – there are dual routes from print to speech. (The distinction between connectionist and nonconnectionist theories of cognition is discussed later in this chapter.)

In the Beginning…

The dual-route conception of reading seems first to have been enunciated by de Saussure (1922; translated 1983, p. 34):

there is also the question of reading. We read in two ways; the new or unknown word is scanned letter after letter, but a common or familiar word is taken in at a glance, without bothering about the individual letters: its visual shape functions like an ideogram.

However, it was not until the 1970s that this conception achieved wide currency. A clear and explicit expression of the dual-route idea was offered by Forster and Chambers (1973):

The pronunciation of a visually presented word involves assigning to a sequence of letters some kind of acoustic or articulatory coding. There are presumably two alternative ways in which this coding can be assigned. First, the pronunciation could be computed by application of a set of grapheme–phoneme rules, or letter-sound correspondence rules. This coding can be carried out independently of any consideration of the meaning or familiarity of the letter sequence, as in the pronunciation of previously unencountered sequences, such as flitch, mantiness and streep. Alternatively, the pronunciation may be determined by searching long-term memory for stored information about how to pronounce familiar letter sequences, obtaining the necessary information by a direct dictionary look-up, instead of rule application. Obviously, this procedure would work only for familiar words. (Forster & Chambers, 1973, p. 627)

Subjects always begin computing pronunciations from scratch at the same time as they begin lexical search. Whichever process is completed first controls the output generated. (Forster & Chambers, 1973, p. 632)

In the same year, Marshall and Newcombe (1973) advanced a similar idea within a box-and arrow diagram. The text of their paper indicates that one of the routes in that model consists of reading via putative grapheme–phoneme correspondence rules (Marshall & Newcombe, 1973, p. 191). Since the other route in the model they proposed involves reading via semantics, and is thus available only for familiar words, their conception would seem to have been exactly the same as that of Forster and Chambers (1973).

This idea spread rapidly:

We can… distinguish between an orthographic mechanism, which makes use of such general and productive relationships between letter patterns and sounds as exist, and a lexical mechanism, which relies instead upon specific knowledge of pronunciations of particular words or morphemes, that is, a lexicon of pronunciations (if not meanings as well). (Baron & Strawson, 1976, p. 386)

It seems that both of the mechanisms we have suggested, the orthographic and lexical mechanisms, are used for pronouncing printed words. (Baron & Strawson, 1976, p. 391)

Naming can be accomplished either by orthographic-phonemic translation, or by reference to the internal lexicon. (Frederiksen & Kroll, 1976, p. 378)

In these first explications of the dual route idea, a contrast was typically drawn between words (which can be read by the lexical route) and nonwords (which cannot, and so require the nonlexical route). Baron and Strawson (1976) were the first to see that, within the context of dual-route models, this is not quite the right contrast to be making (at least for English):

The main idea behind Experiment 1 was to compare the times taken to read three different kinds of stimuli: (a) regular words, which follow the rules of English orthography, (b) exception words, which break these rules, and (c) nonsense words, which can only be pronounced by the rules, since they are not words. (Baron & Strawson, 1976, p. 387)

Figure 1.1 An architecture of the reading system (redrawn from Baron, 1977).

Baron (1977) was the first to express these ideas in a completely explicit box-and-arrow model of reading, which is shown in figure 1.1. This model has some remarkably modern features: for example, it has a lexical-nonsemantic route for reading aloud (a route that is available only for words yet does not proceed via the semantic system) and it envisages the possibility of a route from orthography to semantics that uses word parts (Baron had in mind prefixes and suffixes here) as well as one that uses whole words.

Even more importantly, the diagram in figure 1.1 involves two different uses of the dual-route conception. The work previously cited in this chapter all concerned a dualroute account of reading aloud; but Baron’s model also offered a dual-route account of reading comprehension:

we may get from print to meaning either directly – as when we use pictures or maps, and possibly when we read a sentence like I saw the son – or indirectly, through sound, as when we first read a word we have only heard before. (Baron, 1977, p. 176)

Two different strategies are available to readers of English for identifying a printed word. The phonemic strategy involves first translating the word into a full phonemic (auditory and/or articulatory) representation, and then using this representation to retrieve the meaning of the word. This second step relies on the same knowledge used in identifying words in spoken language. This strategy must be used when we encounter for the first time a word we have heard but not seen. The visual strategy involves using the visual information itself (or possibly some derivative of it which is not formally equivalent to overt pronunciation) to retrieve the meaning. It must be used to distinguish homophones when the context is insufficient, for example, in the sentence, Give me a pair (pear). (Baron & McKillop, 1975, p. 91)

The dual-route theory of reading aloud and the dual-route theory of reading comprehension are logically independent: the correctness of one says nothing about the correctness of the other. Further discussion of these two dual-route theories may be found in Coltheart (2000). The present chapter considers just the dual-route approach to reading aloud.

A final point worth making re Baron’s chapter has to do with the analogy he used to illustrate why two routes might be better than one (even when one is imperfect – the nonlexical route with irregular words, for example):

A third – and to me most satisfying – explanation of the use of the indirect path… is that it is used in parallel with the direct path. If this is the case, we can expect it to be useful even if it is usually slower than the direct path in providing information about meaning. If we imagine the two paths as hoses that can be used to fill up a bucket with information about meaning, we can see that addition of a second hose can speed up filling the bucket even if it provides less water than the first. (Baron, 1977, p. 203)

An analogy commonly used to describe the relationship between the two routes in dual-route models has been the horse race: the lexical and nonlexical routes race, and whichever finishes first is responsible for output. But this analogy is wrong. In the reading aloud of irregular words, on those occasions where the nonlexical route wins, according to the horse race analogy the response will be wrong: it will be a regularization error. But what is typically seen in experiments on the regularity effect in reading aloud is that responses to irregular words are correct but slow. The horse race analogy cannot capture that typical result, whereas Baron’s hose-and-bucket analogy can. The latter analogy is equally apt in the case of the dual-route model of reading comprehension.

Lexical and Nonlexical Reading Routes

This use of the terms lexical and nonlexical for referring to the two reading routes seems to have originated with Coltheart (1980). Reading via the lexical route involves looking up a word in a mental lexicon containing knowledge about the spellings and pronunciations of letter strings that are real words (and so are present in the lexicon); reading via the nonlexical route makes no reference to this lexicon, but instead involves making use of rules relating segments of orthography to segments of phonology. The quotation from de Saussure with which this chapter began suggested that the orthographic segments used by the nonlexical route are single letters, but, as discussed by Coltheart (1978), that cannot be right, since in most alphabetically written languages single phonemes are frequently represented by sequences of letters rather than single letters. Coltheart (1978) used the term grapheme to refer to any letter or letter sequence that represents a single phoneme, so that TH and IGH are the two graphemes of the two-phoneme word THIGH. He suggested that the rules used by the nonlexical reading route are, specifically, grapheme–phoneme correspondence rules such as TH → /θ/ and IGH → /ai/.

Phenomena Explained via the Dual-Route Model

This model was meant to explain data not only from normal reading, but also facts about disorders of reading, both acquired and developmental.

Reaction times in reading-aloud experiments are longer for irregular words than regular words, and the dual-route model attributed this to that fact that the two routes generate conflicting information at the phoneme level when a word is irregular, but not when a word is regular: resolution of that conflict takes time, and that is responsible for the regularity effect in speeded reading aloud. Frequency effects on reading aloud were explained by proposing that access to entries for high-frequency words in the mental lexicon was faster than access for low-frequency words. From that it follows, according to the dualroute model, that low-frequency words will show a larger regularity effect, since lexical processing will be relatively slow for such words and there will be more time for the conflicting information from the nonlexical route to affect reading; and this interaction of frequency with regularity was observed.

Suppose brain damage in a previously literate person selectively impaired the operation of the lexical route for reading aloud while leaving the nonlexical route intact. What would such a person’s reading be like? Well, nonwords and regular words would still be read with normal accuracy because the nonlexical route can do this job; but irregular words will suffer, because for correct reading they require the lexical route. If it fails with an irregular word, then the response will just come from the nonlexical route, and so will be wrong: island will be read as iz-land, yacht to rhyme with matched, and have to rhyme with cave. Exactly this pattern is seen in some people whose reading has been impaired by brain damage; it is called surface dyslexia, and two particularly clear cases are those reported by McCarthy and Warrington (1986) and Behrmann and Bub (1992). The occurrence of surface dyslexia is good evidence that the reading system contains lexical and nonlexical routes for reading aloud, since this reading disorder is exactly what would be expected if the lexical route is damaged and the nonlexical route is spared.

Suppose instead that brain damage in a previously literate person selectively impaired the operation of the nonlexical route for reading aloud while leaving the lexical route intact. What would such a person’s reading be like? Well, irregular words and regular words would still be read with normal accuracy because the lexical route can do this job; but nonwords will suffer, because for correct reading they require the nonlexical route. Exactly this pattern – good reading of words with poor reading of nonwords – is seen in some people whose reading has been impaired by brain damage; it is called phonological dyslexia (see Coltheart, 1996, for a review of such studies). This too is good evidence for a dual-route conception of the reading system.

The reading disorders just discussed are called acquired dyslexias because they are acquired as a result of brain damage in people who were previously literate. The term developmental dyslexia, in contrast, refers to people who have had difficulty in learning to read in the first place, and have never attained a normal level of reading skill. Just as brain damage can selectively affect the lexical or the nonlexical reading route, perhaps also learning these two routes is subject to such selective influence. This is so. There are children who are very poor for their age at reading irregular words but normal for their age at reading regular words (e.g., Castles & Coltheart, 1996); this is developmental surface dyslexia. And there are children who are very poor for their age at reading nonwords but normal for their age at reading regular words and irregular words (e.g., Stothard, Snowling, & Hulme, 1996); this is developmental phonological dyslexia. Since it appears that difficulties in learning just the lexical and or just the nonlexical route can be observed, these different patterns of developmental dyslexia are also good evidence for the dual-route model of reading.

Computational Modeling of Reading

We have seen that the dual-route conception, applied both to reading aloud and to reading comprehension, was well established by the mid-1970s. A major next step in the study of reading was computational modeling.

A computational model of some form of cognitive processing is a computer program which not only executes that particular form of processing, but does so in a way that the modeler believes to be also the way in which human beings perform the cognitive task in question. Various virtues of computational modeling are generally acknowledged – for example, it allows the theorist to discover parts of a theory that are not explicit enough; inexplicit parts of a theory cannot be translated into computer instructions. Once that problem is solved and a program that can actually be executed has been written, the modeler can then determine how closely the behavior of the model corresponds to the behavior of humans. Do all the variables that influence the behavior of humans as they perform the relevant cognitive task also affect the behavior of the program, and in the same way? And do all the variables that influence the behavior of the program as it performs the relevant cognitive task also affect the behavior of humans, and in the same way? Provided that the answer to both questions is yes, studying the behavior of the computational model has demonstrated that the theory from which the model was generated is sufficient to explain what is so far known about how humans perform in the relevant cognitive domain. That does not mean that there could not be a different theory from which a different computational model could be generated which performed just as well. If that happens, the time has come for working out experiments about which the theories make different predictions – that is, whose outcomes in simulations by the two computational models are in conflict.

Of all cognitive domains, reading is the one in which computational modeling has been most intensively employed. This began with the interactive activation and competition (IAC) model of McClelland and Rumelhart (1981) and Rumelhart and McClelland (1982). This was a model just of visual word recognition, not concerned with semantics or phonology. The latter domains were introduced in the much more extensive computational model developed in a seminal paper by Seidenberg and McClelland (1989). One influence their paper had was to prompt the development of a computational version of the dual-route model: the DRC (dual-route cascaded) model (Coltheart et al., 1993; Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001).

Figure 1.2 The DRC model.

The Dual-Route Cascaded (DRC) Model

The DRC is a computational model that computes pronunciation from print via two procedures, a lexical procedure and a nonlexical procedure (see figure 1.2).

The lexical procedure involves accessing a representation in the model’s orthographic lexicon of real words and from there activating the word’s node in the model’s phonological lexicon of real words, which in turn activates the word’s phonemes at the phoneme level of the model. Nonwords cannot be correctly read by this procedure since they are not present in these lexicons, but that does not mean that the lexical route will simply not produce any phonological output when the input is a nonword. A nonword such as SARE can produce some activation of entries in the orthographic lexicon for words visually similar to it, such as CARE, SORE, or SANE; this in turn can activate the phonological lexicon and hence the phoneme level. Such lexically generated activation cannot produce the correct pronunciation for a nonword, but there is evidence that it does influence the reading aloud of nonwords. For example, a nonword like SARE which is similar to many entries in the orthographic lexicon will be read aloud with a shorter reaction time (RT) than a nonword like ZUCE which is similar to few (McCann & Besner, 1987).

The nonlexical procedure of the DRC model applies grapheme–phoneme correspondence rules to the input string to convert letters to phonemes. It does so in serial left-to-right fashion, initially considering just the first letter in the string, then the first two letters, then the first three letters, and so on, until it gets past the last letter in the input. It correctly converts nonwords from print to sound, and also regular words (those that obey its grapheme–phoneme correspondence rules). Irregular (exception) words are regularized by the nonlexical procedure – that is, their rule-based pronunciations, which will be incorrect.

Processing along the lexical route occurs as follows:

Cycle 0: set all the units for visual features that are actually present in the input string to 1; set all others to zero.

Cycle 1: every visual feature set to 1 contributes activation to all the letters in the letter units to which it is connected. The connections are inhibitory when the letter does not contain that feature, and so the activation contributed is negative; the connections are excitatory when the letter does contain that feature, and so the activation contributed is positive.

Cycle 2: what happens on Cycle 1 again happens here. In addition, every letter unit contributes activation to all the word units in the orthographic lexicon to which it is connected. The connections are inhibitory when the word does not contain that letter, and so the activation contributed from letter unit to word unit is negative; the connections are excitatory when the word does contain that letter, and so the activation contributed from letter unit to word unit is positive.

Cycle 3: everything that happens on Cycle 1 and Cycle 2 happens again here. In addition:

(a) Feedforward: each unit in the orthographic lexicon contributes activation to its corresponding unit in the phonological lexicon.

(b) Feedback: every word unit in the orthographic lexicon unit contributes activation back to all the letter units to which it is connected. The connections are inhibitory when the word does not contain that letter, and so the activation contributed from word unit to letter unit is negative; the connections are excitatory when the word does contain that letter, and so the activation contributed from word unit to letter unit is positive.

Cycle 4: everything that happens on Cycles 1, 2, and 3 happens again here. In addition:

(a) Feedforward: every unit in the phonological lexicon contributes activation to all the phoneme units to which it is connected. The connections are inhibitory when the word’s pronunciation does not contain that phoneme, and so the activation contributed from word unit to phoneme unit is negative; the connections are excitatory when the word’s pronunciation does contain that phoneme, and so the activation contributed from word unit to phoneme unit is positive.

(b) Feedback: every unit in the phonological lexicon contributes feedback activation to its corresponding unit in the orthographic lexicon.

Cycle 5: everything that happens on Cycles 1, 2, 3, and 4 happens again here. In addition: every phoneme unit contributes activation back to all the word units in the phonological lexicon to which it is connected. The connections are inhibitory when the word does not contain that phoneme, and so the activation contributed from phoneme unit to word unit is negative; the connections are excitatory when the word does contain that phoneme, and so the activation contributed from phoneme unit to word unit is positive.

And so it goes. As processing cycles progress, inhibitory and excitatory influences continue to flow upwards and downwards in the way described above until the reading-aloud response is ready. How is this readiness determined? As follows. In the description of processing cycles given above, the first cycle on which the phoneme system receives any activation is Cycle 4. At the end of cycle 4, some phoneme units will be activated, but extremely weakly. As processing continues, activation of some of the phoneme units will slowly rise. Quite often, early in processing, some of the phoneme units activated will be incorrect ones. But over time as phoneme activations continue to rise it is the correct phonemes that are the most activated. A reading response is considered to be ready when phonemes have reached a critical level of activation (set to.43 when the model is being used for simulating human reading aloud). The pronunciation generated by the model is taken to consist of the most highly activated phoneme within each of the eight sets of phoneme units (one set per position) that comprise the phoneme system. The processing cycle on which that state of affairs occurs is the DRC model’s reading-aloud latency for the particular letter string that was input.

Processing along the nonlexical route does not begin to operate until cycle 10. Without this time lapse after the lexical route begins to operate, the model would have serious difficulty in reading aloud irregular words. When cycle 10 is reached, the nonlexical route translates the first letter of the string into its phoneme using the appropriate grapheme–phoneme rule, and contributes activation to the phoneme’s unit in the phoneme system. This continues to occur for the next 16 processing cycles. The grapheme–phoneme conversion (GPC) system operates from left to right, so eventually will move on to consider the second letter in the string as well as the first. Every 17 cycles, the GPC system moves on to consider the next letter, translate it to a phoneme, and activate that phoneme in the phoneme system. So with the letter string DESK, the GPC system has no input until cycle 10, deals with just D until cycle 27, deals with just DE from cycle 28 to cycle 44, then DES until cycle 60, DESK until cycle 76 and so on.

Computations on the lexical and nonlexical route occur simultaneously – that is, information from the visual feature level is thought of as flowing simultaneously through the lexical and the nonlexical routes and converging on the phoneme system from these two sources. Whenever the input is an irregular word or a nonword, the two sources of activation conflict at the phoneme level. If the system is to produce correct pronunciations for irregular words and for nonwords, it will have to have a way of resolving these conflicts in favor of the correct pronunciation. Nevertheless, the model reads aloud irregular words and nonwords with high accuracy, so these conflicts are almost always resolved in a way that results in a correct pronunciation (via the interplay of inhibition and activation at various levels of the model). This depends on a judicious choice of values for the parameters of the model, such as the strengths of the inhibitory and the facilitatory connections between components of the model. If the lexical route is too strong relative to the nonlexical route, all words will be read correctly but there will be nonword reading errors. If the lexical route is too weak relative to the nonlexical route, all regular words and nonwords will be read correctly but there will be errors in reading irregular words. A delicate balance between the strengths of the two routes is needed if the model is to perform well with both nonwords and irregular words.

What the DRC Model Can Explain

One way in which Coltheart et al. (2001) evaluated the DRC model was to compare its reaction times to particular sets of stimuli to the reaction times of human readers when they are reading aloud the same stimuli. Do variables that affect human reading-aloud reaction times also affect DRC’s reading-aloud reaction times? Many examples where this was so were reported by Coltheart et al. (2001). For both human readers and the DRC model:

(a) High-frequency words are read aloud faster than low-frequency words.

(b) Words are read aloud faster than nonwords.

(c) Regular words are read aloud faster than irregular words.

(d) The size of this regularity advantage is larger for low-frequency words than for high-frequency words.

(e) The later in an irregular word its irregular grapheme–phoneme correspondence is, the less the cost incurred by its irregularity. So CHEF (position 1 irregularity) is worse than SHOE (position 2 irregularity), which is worse than CROW (position 3 irregularity).

(f) Pseudohomophones (nonwords that are pronounced exactly like real English words, such as brane) are read aloud faster than non-pseudohomophonic nonwords (such as brene).

(g) Pseudohomophones derived from high-frequency words (e.g., hazz) are read aloud faster than pseudohomophones derived from low-frequency words (e.g., glew).

(h) The number of orthographic neighbors a non-pseudohomophonic nonword has (i.e., the number of words that differ from it by just one letter), the faster it is read aloud.

(i) The number of orthographic neighbors a pseudohomophone has does not influence how fast it is read aloud.

(j) The more letters in a nonword there are the slower it is read aloud; but number of letters has little or no effect on reading aloud for real words.

The DRC model was also used to simulate acquired dyslexias. Surface dyslexia was simulated by slowing down rate of access to the orthographic lexicon: this lesioned DRC made regularization errors with irregular words, more so when they were low in frequency, just as is seen in surface dyslexia, whereas its reading aloud of regular words and nonwords remained normal, as in the pure cases of surface dyslexia (Behrmann & Bub, 1992; McCarthy & Warrington, 1986). Phonological dyslexia was simulated by slowing down the operation of the nonlexical route: this lesioned DRC still read words correctly, but misread nonwords, especially if they were nonpseudohomophones, as in the case of phonological dyslexia.

Thus, the DRC model can explain an impressively large number of findings from studies of normal and disordered reading, far more than any other computational model of reading. Nevertheless, Coltheart et al. (2001) drew attention to a number of limitations of the current implementation of the DRC model: its procedure for performing the lexical decision task was crude, it was not applicable to the pronunciation of polysyllabic words or nonwords, it did not offer any account of one popular paradigm for studying reading (masked priming), the difference between word and nonword reading RTs by the model was probably implausibly large, the amount of variance of word reading RTs that the model could account for, though always significant, was disappointingly low, and the implemented model has nothing to say about semantics. A new version of the DRC model that will correct these and other shortcomings of the existing model is under development.

Connectionist and nonconnectionist modeling

This chapter distinguishes between connectionist models of reading (such as the models of Seidenberg & McClelland, 1989, and Plaut, McClelland, Seidenberg, & Patterson, 1996) and nonconnectionist models of reading (such as the DRC model). The description of the DRC model in Coltheart et al. (2001) uses the term connection and the model in fact contains about 4.5 million connections, in the sense of the term connection used by Coltheart et al. (2001). However, in the DRC model, connections are just expository devices used for talking about how the modules of the model communicate with each other. One could expound this in other ways without using the term connection. In contrast, in connectionist models, the connections are often thought of as neuron-like, the models are referred to as neural networks, and terms like biologically inspired or neurally plausible are often applied. Here a connection is something that is physically realizable as an individual object, in contrast to the DRC model in which there is no such sense to the term.

A second major difference between connectionist and nonconnectionist modeling, at least as those trades have been practiced up until now, is that connectionist models have typically been developed by applying a neural-net learning algorithm to a training set of stimuli, whereas the architectures of nonconnectionist models have typically been specified by the modeler on the basis of the empirical effects that the model is meant to explain.

The Seidenberg and McClelland (1989) connectionist computational model of reading is often presented as an alternative to the dual-route model. Indeed, claims such as The dual-route model has been more recently questioned by a plethora of single-route computational models based on connectionist principles (Damper & Marchand, 2000, p. 13) are common in the literature. But that was not the view of the authors themselves. They were clear about this: Ours is a dual route model, they stated (Seidenberg & McClelland, 1989, p. 559).

Figure 1.3 The Seidenberg and McClelland (1989) model. (The implemented model is in boldface type.)

This is perfectly evident from their diagram of their model (Seidenberg & McClelland, 1989, figure 1, reproduced as figure 1.3 here): it explicitly represents two distinct routes from orthography to phonology, one direct and the other via meaning, and explicitly represents two distinct routes from orthography to semantics, one direct and the other via phonology. One of the two routes for reading aloud (the one via semantics) can only be used for reading words aloud; it would fail for nonwords. The other (nonsemantic) route for reading aloud is required if the stimulus is a nonword. This model has come to be called the triangle model, perhaps because of the reference in Seidenberg and McClelland (1989, p. 559) to the third side of the triangle in Figure 1. More than one subsequent model has been referred to as the triangle model despite being different from Seidenberg and McClelland’s model. So far there have been seven different triangle models, an issue discussed later in this chapter.

What is it that has led to this widespread misunderstanding? The answer is clear: a failure to distinguish between the following two claims:

(a) It is possible for a single processing system to correctly read aloud all irregular words and all nonwords.

(b) The human reading system possesses only one procedure for computing pronunciation from print.

Seidenberg and McClelland (1989) did make claim (a). But they did not make claim (b); indeed, as the quotation in the previous paragraph indicates, they repudiated claim (b). That is why theirs is a dual-route model of reading aloud.

This seminal model turned out not to be able to offer a good account of how people read nonwords aloud because its accuracy on this task was far less than the accuracy that human readers show (Besner, Twilley, McCann, & Seergobin, 1990). The suggestion (Seidenberg & McClelland, 1990, p. 448) that this was because the database of words on which the model was trained was too limited and did not contain enough information for nonword reading to be learned from it was shown to be incorrect by Coltheart et al. (1993). They developed a GPC rule-learning algorithm and applied it to the Seidenberg–McClelland training set. The rule set that this algorithm learned from that training set was then used with 133 nonwords from Glushko (1979). Whereas the Seidenberg and McClelland model scored only 68% correct on a subset of 52 of these nonwords, the DRC read 97.9% of these correctly. This shows that the information needed to learn to be an excellent nonword reader is actually present in the model’s database, and so the poor performance of the PDP model in reading nonwords is a defect not of the database but of the model itself (Coltheart et al., 1993, p. 594). Hence, as noted by Plaut (1997, p. 769) and (Plaut et al., 1996, p. 63), the Seidenberg and McClelland model did not succeed in providing evidence that it is possible for a single processing system to correctly read aloud all irregular words and all nonwords.

Nevertheless, it might well be possible to devise a single processing procedure that can correctly read aloud all irregular words and all nonwords. Plaut et al. (1996) sought to devise such a procedure via training a connectionist network similar in overall architecture to that of the network of Seidenberg and McClelland shown in figure 1.3 (it was, for example, a dual-route model in just the same sense that Seidenberg and McClelland viewed their model as a dual-route model, though training was carried out on only one of the two routes), but differing from the Seidenberg and McClelland model in a number of ways, including in the forms of orthographic and phonological representations used in the network. Input units, which were distributed representations in the Seidenberg and McClelland model, became local representations (each representing a grapheme). Output units, which were distributed representations in the Seidenberg and McClelland model, became local representations (each representing a phoneme).

Plaut et al. (1996) actually presented three different though related models – that is, a second, third and fourth triangle model, the first triangle model being that of Seidenberg and McClelland (1989):

Model 1: purely feedforward, 105 grapheme units, 100 hidden units, 61 phoneme units.

Model 2: as for Model 1 but with feedback from phoneme units back to hidden units: an attractor network.

Model 3: as for Model 1 but adding (unimplemented) external input to the output units, so as to mimic what could happen if there were an implemented semantic system activated by orthography and in turn activating phonology. This approach, discussed further below, was pursued in an attempt to simulate acquired surface dyslexia.

How well do these models read nonwords? Model 1 (which after training scored 100% on reading the 2,972 nonhomographic words in the training set) did quite well on nonword reading (see table 3 of Plaut et al., 1996), almost as well as human readers. However it still fails with items like JINJE, the reason being that there is no word in the training corpus that ends with the final grapheme of this nonword. It follows that careful selection of nonwords which exploits such gaps in the training corpus would produce a set of nonwords on which the model would score at or close to zero. Human readers would be vastly superior to the model on such nonwords. Results with nonword reading by Model 2 were similar, though its nonword reading was slightly worse than that of Model 1. The JINJE problem remained.