Information Technology Reference
In-Depth Information
ing the effects of brain damage on reading ability, as are
reviewed in the third section below. Then, we go on to
explore the model.
Semantics
Hidden
Hidden
10.3.1
Comparison with Traditional Dual-Route
Models
Orthography
Hidden
Phonology
Traditional, symbolic, rule-based accounts of reading
aloud posit two routes for reading, a direct route based
on explicit rules for mapping orthography to phonology,
and a “lexical” route containing a word-based lookup
table (or similar such mechanism) of exceptions to these
rules (Pinker, 1991; Coltheart, Curtis, Atkins, & Haller,
1993; Coltheart & Rastle, 1994). Note that other tradi-
tional dual-route theories focused on lexical access (i.e.,
reading for meaning, not reading aloud) use the term
“direct” to refer to the orthographic to lexical/semantic
pathway, and “indirect” to refer to an indirect route via
phonology to the “lexicon” — we have adopted the ter-
minology of Plaut et al. (1996), which is more ap-
propriate for reading aloud. Although traditional dual-
route accounts may seem at great odds with our inter-
active, distributed model (and advocates of the two ap-
proaches have engaged in contentious debate), there is
actually quite general agreement about the existence of
the two basic pathways discussed: the direct orthogra-
phy to phonology pathway, and the more indirect path-
way via semantics. Thus, the central issue in the de-
bate is not so much about dual versus single routes, but
rather about the nature of the processing mechanisms
taking place within each of these routes.
The real contrast between the traditional and neural
network models is in the use of explicit, formal rules
in the direct pathway of the traditional model compared
with a neural network that is sensitive to the system-
aticity of mappings. This contrast has implications for
understanding the way that the two routes divide up the
reading task, and the extent to which there is a sharp
dividing line between the two. The neural network
model of the direct pathway is sensitive to both sys-
tematicity (regularity) and frequency , whereas the rule-
based system is only concerned with regularity. Thus,
the direct pathway in a neural network will learn the
regular pronunciations, and the high-frequency irreg-
ulars , whereas the low-frequency irregulars will rely
Figure 10.5: For the purposes of reading, words are rep-
resented in a distributed fashion across orthographic (visual
word recognition), phonological (speech output), and seman-
tic areas. The direct route for reading goes directly from or-
thography to phonology, while the indirect route goes via se-
mantics.
ing aloud, then this number becomes more manageable,
consisting principally of orthographic (i.e., visual word
recognition) representations, phonological (i.e., speech
output) representations, and semantic representations
(e.g., Seidenberg & McClelland, 1989; Van Orden,
Pennington, & Stone, 1990). In this section, we explore
a model (based on one developed by Plaut & Shallice,
1993) that demonstrates the interaction between these
three different areas, with complete bidirectional con-
nectivity between them (figure 10.5). Thus, the model
provides an instantiation of the distributed lexicon idea.
The model has two different pathways leading from
orthographic input to phonological output. The direct
pathway goes directly from orthography to phonology,
while the indirect or semantic pathway goes via se-
mantics. The presence of these two pathways gives rise
to many interesting phenomena in reading, and touches
on some important theoretical issues that have been
much debated in the literature. We review this debate in
the following section. The subsequent section discusses
a novel mechanism for explaining the behavioral effects
of brain damage that arises in an interactive, multipath-
way neural network model such as ours. Our model can
potentially address a number of relevant behavioral and
clinical reading phenomena — we focus on understand-
Search WWH ::




Custom Search