London Review of Education

Vol. 4, No. 1, March 2006, pp. 5–17

Dual route and connectionist models of

reading: an overview
Max Coltheart*
Macquarie University, Australia
0MaxColtheart
max@[Link]
00000March
London
10.1080/13603110600574322
CLRE_A_157415.sgm
1474-8460
Original
Taylor
412006 and
&Review2006
Article
Francis
(print)/1474-8479
Francis
of Education
Ltd (online)

Reading researchers seek to discover exactly what kinds of information-processing activities go on

in our minds when we read; to discover what the structure and organization is of the cognitive
system skilled readers have acquired from learning to read. Little is known about how the most
elaborate aspects of this system work, but much has been learned about its basic building blocks
such as letter identification, visual word recognition and knowledge of letter-sound rules. I contrast
two approaches to theorizing about these basic reading components, the dual route approach and
the connectionist approach, and offer reasons for believing that the dual route approach is to be
preferred.

Introduction
Cognitive psychology views reading as an information processing activity: reading
aloud is transforming print into speech, and reading comprehension is transforming
print into meaning. Cognitive psychologists interested in reading seek to understand
the nature of the mental information-processing systems people use to perform these
transformations; and cognitive psychologists interested in learning to read seek to
understand how children acquire these mental information-processing systems.
It does not seem likely that much progress would be made if we started off by
investigating ‘real reading’, seeking for example to discover how readers, as they read
The Brothers Karamazov, develop an understanding of what life might have been like
in Imperial Russia. No one has any idea about how to carry out such an investiga-
tion; so more tractable reading situations have to be studied first. This is done by
breaking up ‘real reading’ into simpler component parts that are more immediately
amenable to investigation, with the hope that as more and more of these component

*Macquarie Centre for Cognitive Science, Macquarie University, Sydney NSW 2109, Australia.
Email: max@[Link]

ISSN 1474-8460 (print)/ISSN 1474-8479 (online)/06/010005–13

© 2006 Institute of Education, University of London
DOI 10.1080/13603110600574322
6 M. Coltheart

parts come to be understood we will get closer and closer to a full understanding of
‘real reading’.
One of these component parts of reading is visual word recognition. When we open
a novel and read

ALEXEY Fyodorovitch Karamazov was the third son of Fyodor Pavlovitch Karamazov,
a landowner well known in our district in his own day, and still remembered among us
owing to his gloomy and tragic death, which happened thirteen years ago, and which I
shall describe in its proper place.

we encounter many highly familiar words—third, son, landowner, district—which

carry most of the meaning of the sentence and which we recognize immediately
and effortlessly (if we are skilled readers). Part of understanding how people
comprehend a whole book is understanding how they comprehend whole
sentences; and part of understanding how people comprehend whole sentences is
understanding how they recognize whole words. So if we knew how people recog-
nize whole words on the page, we would know part of what we need to know in
order to understand how people comprehend whole printed sentences, and even a
little part of what we need to know in order to understand how people comprehend
The Brothers Karamazov.
Consequently we need a task which measures visual word recognition and which
will allow us to test theories about how this component of the reading process works;
and the most frequently used task here is lexical decision. This is the task of deciding
whether a printed letter string is a real word or a non-word.
Now, if I ask you which of horse and zloty is a real word of English, you don’t need
to consult your knowledge of whole words: the mere fact that one of these two items
begins zl is enough to indicate that it isn’t an English word, and therefore that the
other letter string must be the word. So the non-word items in a lexical decision task
must be word-like non-words, such as eam, biddle or cloot. There’s nothing about the
letter string biddle that rules it out as a word; the only way you can decide that riddle
is a word and biddle isn’t is to consult your entire list of the words you know—your
mental lexicon—to discover that riddle is in that lexicon whereas biddle is not. Finding
a printed word in your mental lexicon is exactly what we mean by visual word recog-
nition. So if we knew how people perform the lexical decision task, that could tell us
how people recognize whole words as they are doing real reading.
An obvious way in which people might perform the visual lexical decision task is
simply to search through their mental lexicons one word at a time, comparing each
word they find in the mental lexicon with the letter string they are looking at. If they
find a match they respond ‘Yes’; if they get to the end of the lexicon and have not
found a matching word, they respond ‘No’. But this can’t be the way the task is
done. Skilled language users have large vocabularies; there may be 20,000 or more
words a skilled reader is familiar with in print. In lexical decision experiments, it
takes people around 600 milliseconds—six tenths of a second—to respond ‘No’. If
responding ‘No’ requires checking through 20,000 words one after the other, then
each check of a word is being completed in at most .03 milliseconds—3 100,000ths
 Dual route and connectionist models of reading 7

of a second. No one thinks the brain could possibly work that fast. So we need a
different theory about how people recognize words.
I’ll say more about this later. But first let us return to The Brothers Karamazov. It
has many sentences like this one, from chapter 37:
And when Grigory Vassilyevitch wakes up he is perfectly well after it, but Marfa
Ignatyevna always has a headache from it. So, if Marfa Ignatyevna carries out her inten-
tion to-morrow, they won’t hear anything and hinder Dmitri Fyodorovitch.

Who are these people? Especially, who is Marfa Ignatyevna? This is hard to figure
out, because her name has only appeared once in the book so far, and that was back
in chapter 14. Russian novels characteristically have many characters (The Brothers
Karamazov has 48) and their names are not familiar to the non-Russian reader—that
is, the names would not be present in the mental lexicons of such readers. Despite
this unfamiliarity, the names need to be remembered if the plot is to be followed.
Non-Russian readers characteristically report that they do this by pronouncing the
unfamiliar names to themselves and remembering the pronunciations. So adept
readers, having seen the name Marfa Ignatyevna in chapter 14, will generate a
pronunciation from that unfamiliar letter sequence and store it in memory. Then
when something visually unfamiliar crops up in chapter 37—the letter string Marfa
Ignatyevna—the pronunciation that is generated for this will match the previously
stored pronunciation and allow a connection to be made back to the events of
chapter 14.
Why is it that the pronunciations rather than the visual forms of these unfamiliar
letter strings are stored? Because our visual memories have far lower capacities than
our phonological memories. This is easy to demonstrate. Imagine you are staying in
Paris at a hotel in the Rue du Faubourg Saint-Denis. You are out walking in the
Tenth Arrondissement and now you want to go back to your hotel. So you look at
the street signs to try to find the street of your hotel. Do you compare each street
sign to your memory of the 23 letters in the name of your hotel’s street? No; no
matter how execrable your French, you silently pronounce the name of each street
and compare it with your memory of the pronunciation of the name of the hotel’s
street. The reason is that you can remember the sounds of the street name but not its
letters. The same is true for Marfa Ignatyevna’s name.
I have discussed this just to make a simple point: the skilled reader can translate a
letter string from print to speech even if that string has never been seen before (you
can read aloud letter strings such as Ignatyevna or biddle). The mental machinery
that allows us to do this is therefore a second component part of the whole system
we use for reading. The first part I discussed is visual word recognition, which
consists of locating a familiar printed word in one’s mental lexicon. This allows the
reader not only to recognize a printed word, but also to then read it aloud, because
part of a word’s representation in the mental lexicon is a specification of how it is
pronounced; and so this way of reading aloud is called the lexical procedure for read-
ing aloud. When you read aloud a completely unfamiliar letter string such as biddle,
you are not using your mental lexicon, because there is no representation of that
8 M. Coltheart

letter string in your mental lexicon: so whatever procedure you use to translate this
letter string from print to speech (of which more later) is appropriately referred to as
the non-lexical procedure for reading aloud.
These are pretty elementary components of the act of real reading; but they are
nevertheless crucial components. You won’t get far with The Brothers Karamazov
unless you are a fluent recognizer of familiar printed words; and you’ll struggle with
the gigantic cast of characters unless you can translate the printed name of each new
character into a pronunciation (even though you’ve never seen this name before) to
store for later reference.
Although there’s a great deal about the act of reading The Brothers Karamazov that
reading researchers don’t understand, they have learned a great deal over the past
thirty years about how these two elementary components of the act of real reading
are actually accomplished. Figure 1 represents in diagrammatic form the idea that
there’s a lexical procedure for reading aloud (one that consults the mental lexicon)
and also a non-lexical procedure for reading aloud (one that doesn’t consult the
mental lexicon). This basic idea is known as the dual route theory of reading aloud
because it involves two routes from print to speech.
Our mental lexicons contain at least three kinds of information about words: we
Figure 1. The dual-route theory of reading aloud

know about their spellings, their pronunciations and their meanings. Figure 1
includes all three types of information in a single system, the mental lexicon. This
turns out, however, to be wrong; the results of research, particularly neuropsycho-
logical research with people whose language has been disturbed by brain damage,
compels us to adopt the view that these three forms of information about words are
stored in three separate systems, as shown in Figure 2. One of these three lexicons is
the orthographic lexicon which represents knowledge about the visual forms—the
spellings—of words. A second lexicon is the phonological lexicon which represents

print

Letter
identification

Mental Non-lexical
Lexicon procedure

speech

Figure 1. The dual-route theory of reading aloud

 Dual route and connectionist models of reading 9






 



  




 






  












 








Figure 2. Elaboration of the dual-route theory of reading aloud

knowledge about the pronunciations of words. The third lexical system is the
semantic system, where information about the meanings of words is stored.
A crucial distinction inherent in Figure 2 is between regular words and irregular
Figure 2. Elaboration of the dual-route theory of reading aloud

words. Regular words are those that obey the grapheme–phoneme correspondence
rules of English: words like maid or cave. Irregular words are those words which
violate such rules: words like said or have. Regular words can be correctly read by the
lexical and the non-lexical reading routes, but irregular words can be read correctly
only by the lexical reading route: the non-lexical route will get them wrong (it will
read said to rhyme with ‘maid’, have to rhyme with ‘cave’—and yacht to rhyme with
‘matched’).
It is possible to show that in people with different forms of brain damage, any two
of these lexical systems can be intact while the third is damaged.
For example, in some people with dementia (Blazely et al., 2005), knowledge of
word meanings is severely impaired, but the person with dementia can still perform
the visual lexical decision task with normal accuracy (so the orthographic lexicon is
still intact) and can still say words with normal accuracy (so the phonological lexicon
is still intact); here the only one of the three lexical systems that is impaired is the
semantic system. In people with the form of aphasia known as anomia (for review see
Nickels, 1997), it is very difficult to access the pronunciations of words in the
10 M. Coltheart

phonological lexicon, but there can be normal visual word recognition and normal
knowledge of word meanings; here the only one of the three lexical systems that is
impaired is the phonological lexicon. And, finally, in the form of acquired reading
disorder known as surface dyslexia (Patterson et al., 1985), the affected person can
still see perfectly well but can no longer recognize many formerly familiar printed
words, even though still able speak those words and still able to appreciate their
meanings when the words are heard; here the only lexical system that is impaired is
the orthographic lexicon. Such a person may well read said to rhyme with ‘maid’,
have to rhyme with ‘cave’, and yacht to rhyme with ‘matched’, because the lexical
route can no longer be used to read these words. Regular words and non-words will
still be read perfectly because the intact non-lexical route can get them right.
It is hard to see how these various forms of neuropsychological impairment could
be explained unless the mental lexicon is broken up into these three separate systems
as proposed in Figure 2.
A second way in which the elaborated dual route model in Figure 2 differs from
the basic model in Figure 1 is that it makes a specific proposal about how the non-
lexical procedure for reading aloud works, namely, that when people read aloud
without making reference to the mental lexicon, they do this by applying their
knowledge of grapheme-to-phoneme correspondence rules.1
If Figure 2 is a correct account of basic parts of the reading system of skilled read-
ers, then when children are learning to read they are going to have to progress
towards acquiring the mental architecture depicted in Figure 2 if they are to become
skilled readers. That is, possession of this mental processing architecture is an end-
product of successful reading acquisition. There are two different ways of thinking
about reading acquisition in this context. Stage theories of reading acquisition (see
Frith, 1985) propose that as children learn to read they pass through a series of
stages that involve qualitatively different ways of reading. On this view, the Figure 2
diagram does not describe how children read until they have reached the final stage,
i.e., fully skilled reading. An alternative view (see Marshall, 1984; Jackson & Colth-
eart, 2001) is that children at various different points in the course of learning to
read an alphabetic language differ only quantitatively: after they have begun to learn
to read, they all have orthographic lexicons, for example, and differ only in how
many words are represented in these lexicons; they all have a non-lexical procedure
for reading aloud, for example, and differ only in the scope and sophistication of the
knowledge of the relationships between letters and sounds that can be used by that
procedure. On this view, the Figure 2 diagram does describe how children read even
when they are only beginning readers; learning to read does not involve a progres-
sion through qualitatively different stages but instead a progressive quantitative
expansion of the system shown in Figure 2. I consider that the evidence favours this
latter view over a stages view; for further discussion of this see Marshall (1984),
Jackson & Coltheart (2001), Coltheart (2006) and Castles et al. (2006).
One can also apply this particular way of thinking about normal reading
acquisition from the perspective of Figure 2 to the question of abnormal reading
acquisition—that is, developmental dyslexia.2 According to Figure 2, if a child is to
 Dual route and connectionist models of reading 11

be normal-for-age at reading, then each of the four components of the system in

Figure 2 will have to have been acquired to an age-appropriate level. If any one of
the components has not been acquired to that level, then reading will not be normal-
for-age; and the way in which reading will be abnormal for age will depend on which
component or components of the system has not been acquired normally. Suppose
just the semantic system is developmentally impaired: such a child will be able to
read aloud normally but will have impaired reading comprehension (and impaired
comprehension of spoken language). This does occur in some children, and is
known as hyperlexia (see Aram, 1997). If just the orthographic lexicon is poor for
age, the child will be normal at reading regular words and non-words, but poor at
reading irregular words; this is developmental surface dyslexia—see Castles and Colth-
eart (1993, 1996) and for review Jackson and Coltheart (2001) and McDougall et al.
(2005). If just the non-lexical procedure for reading is poor for age, the child will be
normal at reading regular words and irregular words, but poor at reading non-words;
this is developmental phonological dyslexia—see Castles and Coltheart (1993) and
Howard and Best (1996), and for review Jackson and Coltheart (2001) and McDou-
gall et al. (2005).
Understanding how children learn to read involves seeking answers to two ques-
tions. The first is: what is it that the children are learning? The second is: how are they
learning it? Figure 2 proposes an answer to the first question: what children are learn-
ing as they learn to read is the dual route mental architecture depicted in Figure 2.
But dual route theorists have said little about how children learn this architecture:
little about, for example, how new representations are introduced into the
orthographic lexicon, or how the child learns letter-sound rules. A different
approach to theorizing about the reading system, the connectionist approach, does
say something about the learning of reading; perhaps for this reason the connection-
ist approach is the superior approach?

Connectionist modeling of reading

This approach (see Plaut et al., 1996; Harm & Seidenberg, 1999, 2004) differs from
the dual route approach in three fundamental ways, as follows.
● The nature of representation: local versus distributed. According to the dual route
approach, words are represented locally in the reading system. What this means is
that a word corresponds to a single unit in the lexicon, a single lexical entry. The
contrasting idea is distributed representation, as used in connectionist models: any
word is represented by the activation of numerous units in the system, and any
unit in the system plays a role in representing many different words.
● The nature of processing: parallel versus serial. In all the connectionist models of read-
ing, all processing goes on in parallel: for example, when a non-word is presented
to the model, all the letters are processed by the model simultaneously. In contrast,
one component of the dual route model in Figure 2 operates serially: the non-lexical
procedure translates letters to sounds one letter after another, from left to right.
12 M. Coltheart

● Learning: the knowledge connectionist models use to carry out the reading-
aloud task is not provided by the modeller, but developed over time by the
model as it is repeatedly exposed to the spellings of words and their correct
pronunciations—the knowledge is developed under the control of a learning
algorithm which gradually and progressively adjusts the strengths of the
connections in the model so that the model’s response to each word in its train-
ing set becomes progressively more accurate. In contrast, while the knowledge
that a dual route model uses to carry out the reading-aloud task is assumed to
be gradually acquired by children, the model is intended just as a description of
the information-processing system that children acquire as a result of this
learning.
We can thus directly compare the two approaches by considering three questions, as
follows.

Are the representations in the reading system local or distributed?

The key issue here is lexical decision. Skilled readers find it very easy to decide
rapidly that riddle is a real word and biddle is not. How do they do this? Models that
propose that representations in the reading system are local offer a simple answer:
any kind of interrogation of the orthographic lexicon will show that there’s a unit in
that system corresponding to the word riddle but no unit corresponding to the non-
word biddle, and so responding ‘Yes’ to riddle and ‘No’ to biddle is simply done.
Serial checking of one item after another in the mental lexicon – which can’t be the
way lexical entries are found, as noted earlier—is not needed: in dual route models
of reading the right lexical entry can be directly activated. If the letter c is connected
to all words beginning with c, the letter a connected to all words with an a in the
second position, and the letter t connected to all words with a t in the third position,
then the letter string cat will activate the right entry in the orthographic lexicon with-
out any need for a search through the lexicon. That is how the dual route model of
reading works (Coltheart et al., 2001).
But according to connectionist models, since there are no local representations,
there’s no unit in the system uniquely responsive to riddle: the word riddle will acti-
vate a lot of units (since its representation is distributed) but so will the non-word
biddle. Therefore:
… given that the current distributed approach to lexical processing does not contain
word-specific representations, it becomes important to establish that distributed models
can, in fact, perform lexical decision accurately and that, in doing so, they are influ-
enced by properties of the words and non-words in the same way as human readers.
(Plaut, 1997, p. 785)

Advocates of connectionist modeling of reading have not succeeded in estab-

lishing this. There has been no report of a connectionist model of reading that
can perform the lexical decision task accurately and which in so doing is influ-
enced by the various properties of words and non-words that are known to affect
 Dual route and connectionist models of reading 13

human performance of the lexical decision task. The only work here has sought
to show that, if a connectionist model is equipped with a simulated semantic
system, it can discriminate words from non-words because words activate the
simulated semantic system more strongly (Plaut, 1997; Harm & Seidenberg,
2004). If the correct explanation of how human readers perform the lexical
decision task is that they rely on consulting their semantic systems, then people
whose brain damage has left them with severe semantic impairments could never
achieve normal accuracy in the visual lexical decision task; but some can (for
examples, see Lambon Ralph et al., 1995, 1996, 1998; Ward et al., 2000; Blazely
et al., 2005).

Is all the processing performed by the reading system parallel processing, or

are some reading operations performed in a serial manner?
Rastle and Coltheart (2006) have reviewed this literature. They identify eight differ-
ent effects reported in the literature on reading aloud which cannot be explained
unless there is some form of serial processing occurring in the reading system. I’ll
give just two examples.
● Position of irregularity: skilled readers are slower at reading aloud irregular words
than regular words. The size of this regularity effect depends upon where in the
irregular word the irregular grapheme–phoneme correspondence is. Rastle and
Coltheart (1999) showed that the effect was large when the irregularity involves
the first grapheme of an irregular word (as in the word chef), smaller but still
significant when it involves the second grapheme (as in the word sown), and
absent when it involves the third (as in the word steak). This was confirmed by
Roberts et al. (2003), who also showed that this positional effect could not be
simulated by a connectionist model of reading which had no serial-processing
component but could be correctly simulated by the Dual Route Cascaded (DRC)
model (Coltheart et al., 2001) of reading, a computational realization of the dual
route theory shown in Figure 2, because, in that model, the non-lexical procedure
operates from left to right.
● Position-sensitive Stroop effect: when your task is to name the colour in which a
word is printed, your colour-naming response is faster when the colour name you
have to produce starts with the same phoneme as the word that is carrying the
colour: if the colour name you have to utter is ‘red’ you are faster when the word
is rat than when it is cot. There’s also facilitation from shared final phonemes: the
response ‘red’ is faster to the word bad than to the word cot. But facilitation is
larger when it is the first phoneme that is shared than when it is the last (Coltheart
et al., 1999). This also indicates that there is some serial processing going on as
we read, and this effect too can be simulated by the DRC model of reading in
which the non-lexical procedure operates from left to right. It does not seem that
connectionist models (in which all processing is parallel) could ever explain this
position-sensitive effect.
14 M. Coltheart

Do connectionist models offer a satisfactory account of how children learn

to read?

When a connectionist model can learn to read via repeated exposures to printed
words and their correct pronunciations, that’s clearly a fascinating feature of the
model. But such a model could only be preferred to dual route models, which do not
offer any account of how reading is learned, if the connectionist model learns in the
same way that children learn. If connectionist models do not learn in the same way
as children do, then the fact that these models learn offers no reason to favour them
over models that do not learn. So a key question is: could the way connectionist
models learn to read be the same as the way that children learn to read? I will argue
that the answer to this question is no, for two reasons.
Firstly, the number of exposures to spelling-sound pairs that are needed for the
connectionist models to learn them seems vastly greater than what is needed by chil-
dren. There are numerous different connectionist models of learning to read, all
somewhat different—Coltheart (2005) lists seven such models—but the following
points apply to all of them. The first of the connectionist models of reading
described in Plaut et al. (1996) was trained on a set of 2998 words. Each word and
its pronunciation was presented 300 times to the model; after that number of
presentations, the model could correctly generate the pronunciation of all the words
in the training set. Another of the connectionist models described in Plaut et al.
(1996) needed to have each word presented 1900 times and even after that had not
learned the pronunciations of 25 of the words in the training set. Harm and Seiden-
berg (1999) used a training set of 3123 words for their connectionist model, and
training consisted of presenting a total of ten million words. Harm and Seidenberg
(2004) used a training set of 6103 words and their connectionist model was trained
for 1.5 million word presentations. It is clear from these figures that the PDP models
require a word to be presented along with its correct pronunciation hundreds or
thousands of times for its pronunciation to be learned correctly. Children learning to
read aloud single words do not require to see each word and hear its pronunciation
hundreds or thousands of times before they can correctly read aloud the words.
A possible defence to this criticism is that children are not expected to learn to
read some thousands of words at the same time: perhaps if they were, they might
require hundreds or thousands of exposures to each word (though this seems very
unlikely). Instead, the more usual scenario is that children learn a small set of words
until they have got them right, then another small set and so on. But this introduces
the second and even more serious problem for learning in connectionist models.
Suppose you train a connectionist model on some small set of words—call it Set A—
until all are read perfectly. Then you train the model to perfection on a new set, Set
B, without further training on Set A. If you now retest it on Set A, it will perform
very poorly. This is the problem of catastrophic forgetting in connectionist networks
(McCloskey & Cohen, 1989). It occurs because the connection strengths that are
learned so as to be able to get all the Set A words correct are all liable to be changed
when the only current task is to get all the Set B words right: the Set A words no
 Dual route and connectionist models of reading 15

longer matter to the model now. The only way to avoid the catastrophic forgetting
problem here is to keep presenting the A words and their correct pronunciations
even after training on the B words has begun. But nothing like this happens as
humans learn to read. Once we have learned to read a word aloud, we do not need
to keep receiving training on what its correct pronunciation is; we do not unlearn it
as soon as that training is discontinued. Of course, we keep seeing these words over
and over again as our lives progress, and we also keep hearing them over and over
again. But we do not need constant feedback regarding a printed word’s pronuncia-
tion; once the word is learned, it stays learned even if we never again are taught it by
being shown it as we are told how to say it. In contrast, when PDP networks are
given a new set of words to learn and training on a previously learned set is discon-
tinued, the old set will be forgotten. That doesn’t happen with children.
My conclusion then regarding the connectionist approach to explaining reading is
that at present it has not been successful. The connectionist eschewal of local repre-
sentations makes it impossible for current connectionist models of reading to explain
how people perform the lexical decision task. The connectionist insistence on paral-
lel processing makes it impossible for current connectionist models of reading to
explain the many demonstrations of serial processing as people read aloud. And
though connectionist models do learn to read, this is not an advantage of such
models, because they don’t learn to read the way children do. So at present the dual
route approach described in Figure 2, and converted by Coltheart et al. (2001) into a
computational model of visual word recognition and reading aloud (the DRC
model), seems to offer the most promising initial steps down the long path towards
an understanding of what readers are doing as they are understanding The Brothers
Karamazov.

Acknowledgements
The author thanks Genevieve McArthur, Morag Stuart and Taeko Wydell for help-
ful comments on an initial draft.

Notes
1. The technical term ‘grapheme’ refers to the written representation of a phoneme. So, for
example, the word sheep has five letters but only three graphemes, these graphemes being sh,
ee and p.
2. By ‘developmental dyslexia’ I mean simply difficulty in learning to read. This contrasts with
‘acquired dyslexia’ which refers to any impairment of reading that is caused by brain damage
in a person who had learned to read normally prior to suffering that brain damage.

Notes on contributor
Max Coltheart is Scientific Director of the Macquarie Centre for Cognitive Science
and James Packer Professor of Educational Research. He’s held posts at univer-
sities in Australia, Canada, Italy, the US and England, is author of 229 journal
16 M. Coltheart

articles and book chapters, and has co-authored the books Language processing
in children and adults and Routes to reading success and failure.

References
Aram, D. M. (1997) Hyperlexia: reading without meaning in young children, Topics in Language
Disorders, 17, 1–13.
Blazely, A., Coltheart, M. & Casey, B. (2005) Semantic dementia with and without surface
dyslexia, Cognitive Neuropsychology, 22, 695–717.
Castles, A., Bates, T. & Coltheart, M. (2006) John Marshall and the developmental dyslexias,
Aphasiology.
Castles, A. & Coltheart, M. (1993) Varieties of developmental dyslexia, Cognition, 47, 149–180.
Castles, A. & Coltheart, M. (1996) Cognitive correlates of developmental surface dyslexia: a single
case study, Cognitive Neuropsychology, 13, 25–50.
Coltheart, M. (2005) Modelling reading: the dual route approach, in: M. J. Snowling & C. Hulme
(Eds) The science of reading (Oxford, Blackwells Publishing).
Coltheart, M. (2006) John Marshall and the cognitive neuropsychology of reading, Cortex.
Coltheart, M., Rastle, K., Perry, C., Langdon, R. & Ziegler, J. (2001) DRC: a dual route cascaded
model of visual word recognition and reading aloud, Psychological Review, 108, 204–256.
Coltheart, M., Woollams, A., Kinoshita, S. & Perry, C. (1999) A position-sensitive Stroop effect:
further evidence for a left-to-right component in print-to-speech conversion, Psychonomic
Bulletin and Review, 6, 456–463.
Frith, U. (1985) Beneath the surface of developmental dyslexia, in: K. E. Patterson, J. C. Marshall
& M. Coltheart (Eds) Surface dyslexia: cognitive and neuropsychological studies of phonological
reading (Hove, Lawrence Erlbaum Associates).
Harm, M. & Seidenberg, M. S. (1999) Reading acquisition, phonology, and dyslexia: insights
from a connectionist model, Psychological Review, 106, 491–528.
Harm, M. & Seidenberg, M. S. (2004) Computing the meanings of words in reading: cooperative
division of labor between visual and phonological processes, Psychological Review, 111, 662–
720.
Howard, D. & Best, W. (1996) Developmental phonological dyslexia: real word reading can be
completely normal, Cognitive Neuropsychology, 13, 887–934.
Jackson, N. & Coltheart, M. (2001) Routes to reading success and failure (Hove, Psychology Press).
Lambon Ralph, M. A., Ellis, A. W. & Sage K. (1998) Word meaning blindness revisited, Cognitive
Neuropsychology, 15, 389–400.
Lambon Ralph, M. A., Sage, K. & Ellis, A. W. (1996) Word meaning blindness: a new form of
acquired dyslexia, Cognitive Neuropsychology, 13, 617–639.
McCloskey M. & Cohen, N. J. (1989) Catastrophic interference in connectionist networks: the
sequential learning problem, in: G. H. Bower (Ed.) The psychology of learning and motivation
(New York, Academic Press).
McDougall, P., Borowsky, R., MacKinnon, G. E. & Hymel, S. (2005) Process dissociation of sight
vocabulary and phonetic decoding in reading: a new perspective on surface and phonological
dyslexias, Brain & Language, 92, 185–203.
Marshall, J. C. (1984) Toward a rational taxonomy of the developmental dyslexias, in: R. N.
Malatesha & H. A. Whitaker (Eds) Dyslexia: a global issue (The Hague, Martinus Nijhoff).
Nickels, L. A. (1997) Words fail me: spoken word production and its breakdown in aphasia (Hove,
Psychology Press).
Patterson, K. E., Marshall, J. C. & Coltheart, M. (Eds) (1985) Surface dyslexia: cognitive and
neuropsychological studies of phonological reading (Hove, Lawrence Erlbaum Associates).
Plaut, D. C. (1997) Structure and function in the lexical system: insights from distributed models
of word reading and lexical decision, Language & Cognitive Processes, 12, 767–808.
 Dual route and connectionist models of reading 17

Plaut, D. C., McClelland, J. L., Seidenberg, M. S. & Patterson, K. (1996) Understanding normal
and impaired word reading: computational principles in quasi-regular domains, Psychological
Review, 103, 56–115.
Rastle, K. & Coltheart, M. (1999) Serial and strategic effects in reading aloud, Journal of
Experimental Psychology: Human Perception & Performance, 25, 482–503.
Rastle, K. & Coltheart, M. (2006) Is there serial processing in the reading system; and are there
local representations?, in: S. Andrews (Ed.) All about words: current issues in lexical processing
(Hove, Psychology Press).
Roberts, M., Rastle, K., Coltheart, M. & Besner, D. (2003) When parallel processing in visual
word recognition is not enough: new evidence from naming, Psychonomic Bulletin & Review,
12, 405–414.
Ward, J., Stott, R. & Parkin, A. J. (2000) The role of semantics in reading and spelling: evidence
for the ‘summation hypothesis’, Neuropsychologia, 38, 1643–1653.