Content uploaded by Sylviane Valdois
Author content
All content in this area was uploaded by Sylviane Valdois
Content may be subject to copyright.
lo
alReviaw
Ttr"
1
8,
k
105,
NO.
4,678-723
Multiple-Trace Memory Model
A
Connectionist
for Polysyllabic
Word
Reading
Bernard
Ans
Serge Carbonnel
Universitd
Pierre
Mend& France and Centre National de la Universit6 de Savoie
and
Centre National de la Recherche
Recherche Scientifique Scientifique
Sylviane
Vddois
Universit6
Pierre
Mend& France and Centre National
de
la Reckhe Scientifique
A
connectionist
feedforward
network implementing
a
mapping
from
orthography
to
phonology is
described.
The
model
develops
a
view
of
the
reading
system that
accounts
for
both
irreguh
word
and
pseudoword reading without relying on any system
of
explicit
or
implicit
conversion
rules.
The
model
assumes,
however, that reading is supported by
2
procedures
that
work
successively:
a global
procedure using knowledge about entire words and
an
analytic procedure
based
on
the
activation
of
word syllabic
segments.
The model
provides
an
account
of
the
basic effects that
characterize
human
skilled
reading
performance including a frequency by consistency
interaction
and
a
position-of-
irregularity effect.
Furthamore,
early
in
training, the network shows a performance
similar
to
that
of
less skilled readers. It also
offers
a plausible account
of
the
patterns
of
acquired
phonological
and
surface dyslexia when lesioned
in
different ways.
Much research in cognitive psychology
and
the neuropsychol-
ogy of language is aimed toward understanding how orthography
is translated to phonology when people read aloud and how
this
transcoding process is acquired. Among the various theoretical
conceptions of
the
architecture
of the reading system that have
been
described,
two
main
classes of models have emerged, dual-
route models (Coltheart, 1978, 1985; Coltheart, Curtis, At-
kins,
&
Hallq
1993; Coltheart
&
Rastle, 1994; Coslett, 1991;
hell,
1983; Patterson
&
Morton, 1985; Patterson
&
Shewell,
1987)
and
parallel distributed processing (PDP) models (Hin-
ton
&
Shallice, 1991; Plaut
&
McClelland, 1993; Plaut, McClel-
land, Seidenberg,
&
Patterson,
19%;
[hereinafter PMSp961;
plaut
&
Shallice, 1993; Seidenberg
&
McClelland, 1989 [here-
in- SM891).
The
Dual-Route
Approach
The
main
feature
of
the dual-route model of reading is that
it postulates
the
existence of
two
different processing routes for
Banard
Ans
and
Sylviane Valdois, Laboramire de Psychologie ExpCr-
imentale,
Univemit6
Pierre France,
and
Centre National de
la
Recherche
Scientifique, Grenoble, France;
Serge
Carbonnel.
Laboramire
de
Psychobgie Exp&imentak,
Th.iversi~
de
Savoie,
and
Centre
National
de
la
Recherche
Scitmtifique,
Chamb&y,
France.
This
research was
funded
by
the
Centm
National de
la
Recherche
Scimtifiquc
and
by
Grant GDR
CNRS
978
in
neuropsychology.
We
thank
Gordon
D.
Logan
and
David C. Plaut
for
their
helpful
comments
on
an
earlier
waxion
of
this
article. We
also
thank
Elwp
C.
Leek
for
the
English
revision
of
the
article.
The
authors
are
listed
in
alphabetical
or&
Correspondence concaning
this
article
should
be
addressed
to
Ba-
nard
Ans,
Laboramire
de Psychologie
Exp&imatale,
Universid
Pierre
Mend&
France,
BP
47,
38040
Grenoble
Cedex
09,
France. Electronic
mail
may be smt
to
Bernard.Ans@upmf-grtmoblerenoble.fr.
converting print to sound. These are a lexical route allowing
access to stored knowledge about familiar (previously learned)
words
and
a separate nonlexical route involving print-to-sound
mapping rules. It is hypothesized within
this
framework
that
the
phonological and orthographic representations
of
all previously
learned
words
a~
stored
in
a mental lexicon in long-term
mem-
ory.
By
this
view, reading aloud by the lexical pathway is
achieved by first accessing the stored orthographic representa-
tion corresponding to
the
input word
and
then retrieving the
whole phonological representation of the word. According
to
some
hypotheses,
the
lexical pathway is further subdivided into
a semantic route and a nonsemantic route.
The
lexical semantic
route allows access to semantic information from orthographic
representations, prior to phonological retrieval, whereas the non-
semantic lexical pathway allows
the
phonological representation
to
be
directly
addressed from orthography without semantic
mediation.
Only
previously learned words (that have entries
in
the
mental
lexicon) can
be
successfully read aloud by
the
lexical
procedure.
Reading
aloud
of
items
that
have no
entries
in the
mental
lexicon (i.e., new words
or
pseudowords) can only
be
performed
via
the
nodexid pathway. This
route
depends on
the use
of
a set
of
general rules specifying correspondences
between orthography
and
phonology.
The
nonlexical procedure
typically derives pronunciations
that
obey
standard
Orthography-
to-phonology correspondences. It
will
thus produce incorrect
pronunciations (i.e.,
regularization
errors)
when
the
input is
an
imguh
word. Thus,
the
correct pronunciation
of
imgular
words can
be
obtained only through the lexical pathway by
accessing
the
appropriate
stored
representation
in the
ortho-
graphic and phonological lexicons, whereas
the
correct
pnun-
ciation of regular words can
be
achieved by either
the
lexical
or nonlexical routes. The dual-route model
furhex
postulates
that processing orthographic input
starts
simultaneously in
the
678
CONNECTIONIST
READING
MODEL
679
two
routes and works in parallel whatever the nature of the input
letter string (regular word, irregular word, or pseudoword) be-
cause there is no way of determining, prior to lexical access,
whether an orthographic stimulus
is
a familiar word or a new
word. As a consequence, for all words that an individual has
already learned, an “addressed” phonological representation
would be delivered by the lexical route and an “assembled”
pronunciation would
be
computed by the nonlexical route.
In
their dual-route cascaded (DRC) model (a computational ver-
sion of the dual-route model), Coltheart et al.
(1993)
claimed
that phonological information from both the lexical and nonlexi-
cal pathways converges toward a common component of the
model, the phoneme system, which comprises a set of units
representing
all
possible phonemes and their position. These
units are therefore activated by inputs from both pathways. The
DRC model also incorporates assumptions about the time course
of
the
two processing routes. Lexical processing is viewed
as
being typically faster than nonlexical processing. However, non-
lexical processing is not
so
slow that its input never reaches the
phoneme system before activation from the lexical route has
reached
its
full
value. Conflicts can therefore occur at the pho-
neme level when the two routes simultaneously produce differ-
ent outputs (alternative pronunciations in a given position)
as
is the case for low-frequency irregular words.
A wide range of experimental findings from skilled readers
(Beauvois
&
DerouesnC,
1979;
Coltheart
&
Rastle,
1994;
Monsell, Patterson,
Graham,
Hughes,
&
Milroy,
1992;
Paap
&
Noel,
1991)
as well
as
neuropsychological evidence from studies
on acquired reading disorders (Coltheart, Sartori,
&
Job,
1987;
Marshall
&
Newcombe,
1973;
Patterson, Marshall,
&
Coltheart,
1985)
have been interpreted as providing support for the exis-
tence of
two
distinct procedures for translating print to sound
(for discussion,
see
Carbonnel
&
Ans,
1996;
Humphreys
&
Evett,
1985).
The dual-route theory
in
particular accounts for
the regularity by frequency interaction that has been observed
in
a
number of naming latency studies (Seidenberg, Waters,
Barnes,
&
Tanenhaus,
1984;
Taraban
&
McClelland,
1987).
Experimental studies indeed showed longer naming latencies for
irregular
words
than
for regular words
if
they are of low fre-
quency but no such effect for words of high frequency. Within
the DRC model (Coltheart
et
al.,
1993),
this
regularity effect
is interpreted
as
resulting from competition between the lexical
and nonlexical phonological outputs in the following way:
The
rate at which activation rises throughout the lexical route is
a
function
of
word frequency
so
that the more frequent a word
is, the sooner its corresponding phonemes will be activated in
the
phoneme system. As a consequence, activation of
this
system
by the lexical route will be completed for high-frequency words
(be they
regular
or irregular) before substantial activation (if
any) gets to that system from the nonlexical route.
In
this case,
irregularity does not
affect
performance. In contrast, for low-
frequency words, lexical and nonlexical outputs simultaneously
activate
units in the phoneme system,
This
will result in
no
latency cost for low-frequency
regular
words because phonemes
activated by the two sources will be the same but will generate
conflicts whenever the stimulus is a low-frequency hgular
word.
In
this
case, the time taken to resolve the conflict will
affect naming latency
so
that only low-frequency irregular
words will suffer a latency cost (see Coltheart
&
Rastle,
1994).
The dual-route maldel
further
provides a
straightforward
ex-
planation of
two main
forms of acquired dyslexias (reading
disorders caused by
lmain
damage in a previously literate in&-
vidual). The pattern Df
surface
dyslexia
(Coltheart, Masterson,
Byng,
Rioz
&
Riddoch,
1983;
Saffran
&
Marin,
1977;
Shallice,
Warrington,
&
McCarthy,
1983)
that is characterized by a selec-
tive impairment of inregular word reading, resulting
in
the pro-
duction
of
regularizauion
errors,
is viewed as reflecting damage
to the lexical route and preservation
of
the nonlexical pathway
(Coltheart et al.,
1993).
In contrast,
the
pattern of
phonological
dyslexia
(Beauvois
&
Derouesnt,
1979;
Derouesnt
&
Beauvois,
1985;
Shallice
&
McCarthy,
1985;
Shallice
&
Warrington,
1975)
that is characterized by a selective pseudoword reading impair-
ment is interpreted ari resulting from damage to the nonlexical
route and relative preservation
of
the lexical pathway (Coltheart
et al.,
1993).
The theoretical coiaception of the reading system developed
within the framework of dual-route models
has
been challenged
by models that assume that pronouncing irregular words and
pseudowords does not
require
separate lexical and sublexical
procedures. Analogy models of
the
reading system (Glushko,
1979;
Kay
&
Marcel.
1981;
Marcel,
1980)
represented the first
radical departure
from
the
dual-route approach, postulating that
the pronunciation of
regular
words, irregular words, and pseu-
dowords is assigned iby analogy with and by specific reference
to
known
lexical items, novel words being pronounced by gener-
alization from existing words.
The
analogy theory assumes that
only whole word phonology is stored
in
long-term memory
and that orthography .to-phonology conversion rules only exist
implicitly in the integrated activation of words (Glushko,
1979).
However, the ability of an implemented network based on
the
analogy theory to actually reproduce the basic effects that char-
acterize skilled word
and
nonword reading has not been demon-
strated.’ Another conception
of
the reading system has been
developed within the framework of multiple-level models
(Shal-
lice
&
McCarthy,
19185;
Shallice et al.,
1983).
These models
also postulate, apart from
the
semantic pathway, a single nonse-
mantic route, the phonological pathway, fully competent for
reading
all
kinds of letter strings. The phonological route is
described in
this
class of model
as
incorporating spelling-sound
correspondences ranging from single graphemes and phonemes
to
word bodies and entire morphemes
so
that
irregular word
pronunciation could be achieved on the basis of morphemic
correspondences, whcreas pseudoword reading would typically
involve correspondences between smaller graphemic units (see
Norris,
1994,
for an implementation
of
a multiple-level model).
Another alternative to the dual-route model, assuming (apart
from the semantic pathway) a single-route mechanism for read-
’
Coltheart et
al.
(19193)
referred
to
the novel-word pronunciation
system implemented by
Sullivan
and
Damper
(1992)
as
an
explicit
model
of
reading
by
anrdogy.
Sullivan
and
Damper themselves claimed
that
their
model
gives
an
example
of
how
the
analogy
processes
could
be
implemented. However, their model
differs
somewhat from
the
analogy
theory.
For
example,
a
system
of
grapheme-phoneme correspondence
(GPC)
rules
is implemented that provides
“a
first plausible pronuncia-
tion
[of
the
letter
string]
to
be
furtha:
relined at
the
analogy
stage”
(p.
183).
The
precompilatitm
of
a
set
of
GPC
rules seems
in
disagreement
with the very basis
of
analogy
models.
680
ANS,
CARBONNEL,
AND
VALDOIS
ing both words and pseudowords was proposed by SM89, who
developed a PDP connectionist network that learns to map the
orthographic representations of words onto corresponding pho-
nological representations,
The
PDP
Connectionist Approach
The SM89 general lexical processing model is composed of
three layers of units, an orthographic, a phonological, and a
semantic layer, and
three
sets of hidden units mediating between
them. The units
are
fully connected from one layer
to
the
ok.
Information is represented in terms of distributed patterns of
activity over these separate groups of units, and processing is
mediated by connections among the units. Knowledge of the
relations among the orthographic, phonologic, and semantic pat-
terns
is encoded by the weights on connections between units.
The model postulates a phonological pathway by which the
appropriate phonological pattern is computed from the ortho-
graphic pattem activated by the input and a semantic pathway
allowing a word
to
be pronounced by means of a computation
from orthography
to
meaning and then from meaning to phonol-
ogy.
In
fact, only a part of