Language - Computational Explorations in Cognitive Neuroscience - page 324

more on the semantic pathway (because the semantics-
to-phonology mapping lacks the overwhelming regular-
ities of the orthography-to-phonology pathway, it can
more easily support low frequency irregulars).
In contrast, the rule-based system would not be ex-
pected to handle the high-frequency irregulars. Fur-
ther, instead of having a discrete switch between the two
routes, the network approach involves simultaneous, in-
teractive processing and a much softer division of labor.
See Plaut et al. (1996) (PMSP) for discussion of the
empirical evidence supporting the neural network view.
Interestingly, even exception words (e.g., “blown,”
which does not rhyme with the regulars “clown,”
“down,” etc.) exhibit systematicities in the orthography
to phonology mapping (e.g., rhyming with “grown”),
posing a challenge to the traditional account of a simple
lookup table for exceptions (see section 10.4 for specific
examples and more discussion). Instead, such system-
aticities suggest that the exception mapping appears to
take into account subregularities. The neural network
system naturally handles these subregularities in much
the same way it captures the systematicities of the reg-
ular mapping. In contrast, proponents of the traditional
account have had to revise their models to include a neu-
ral network for performing the exceptional mappings
(Pinker, 1991). Given that there is really a continuum
between the regular and exception mappings, and that
a neural network can handle both the regular and ex-
ception mappings, it seems strange to continue to main-
tain a sharp mechanistic distinction between these map-
pings.
Because the present model uses a small number of
words and is focused more on capturing the broadest
level relationships among the different types of dyslex-
ias, it is not really capable of addressing the frequency
and regularity effects that play such an important role
in the debate about the nature of processing in the di-
rect pathway. We will revisit these issues with the sub-
sequent model (section 10.4) where we explore a much
more elaborated and realistic instantiation of the direct
pathway based on the PMSP model. The results from
this model show that a neural network implementation
of the direct pathway does indeed learn the high fre-
quency irregulars in addition to the regulars.

10.3.2 The Interactive Model and Division of Labor

The interactive (bidirectionally connected) nature of the
model leads to some important phenomena. Regardless
of where the activation originates (orthography, seman-
tics, or phonology), it will flow simultaneously through
the direct and indirect pathways, allowing both to con-
tribute to the activation of the other representations.
These dual interacting pathways allow a division of la-
bor for processing different types of words to emerge
over learning. For example, if the direct pathway reli-
ably produces the correct phonological activations for
high-frequency words, then the indirect pathway will
experience less pressure (i.e., less error-driven learning)
to acquire the ability to produce these phonological ac-
tivations itself.
Unless one of these pathways becomes damaged, it
may be relatively difficult to see effects of this division
of labor. Thus, the model can exhibit complex effects of
damage arising from the premorbid division of labor of
certain classes of words on different pathways. This di-
vision of labor can extend beyond basic parameters like
frequency and regularity to include factors such as the
relative richness of the semantic representations for dif-
ferent words. Thus, the interactive neural network ap-
proach provides a novel and parsimonious explanatory
mechanism for patterns of behavior under damage that
can otherwise appear rather puzzling, requiring improb-
ably complex patterns of coincident damage to various
specialized pathways to explain within traditional mod-
els.

10.3.3 Dyslexia

Dyslexia is a generic term for a reading problem. Many
different types or categories of reading problems have
been identified, with the type best known in popular cul-
ture being developmental dyslexia. We focus on three
main categories of acquired dyslexia (i.e., acquired as
a result of brain damage, not developmental factors):
phonological , deep ,and surface dyslexia.
People with phonological dyslexia have a selec-
tive deficit in reading pronounceable nonwords (e.g.,
“nust”) compared with reading real words. In terms of
our distributed lexicon model (figure 10.5), phonolog-
ical dyslexia can be understood as a lesion to the di-