Language - Computational Explorations in Cognitive Neuroscience - page 325

rect pathway connecting orthography and phonology,
so that preserved reading goes via semantics, which
does not have representations for nonwords. In con-
trast, in the intact system, the quasi-regular mappings
between word spelling and pronunciation are learned
by the direct pathway, enabling nonwords to be pro-
nounced. As with all brain lesions, there can be varying
degrees of damage and thus varying degrees of impair-
ment in reading nonwords among people with phono-
logical dyslexia.
People with deep dyslexia are like those with phono-
logical dyslexia in that they cannot read nonwords, but
they also exhibit significant levels of semantic errors,
in which they mistakenly read words as semantically
related words (e.g., reading “dog” as “cat”). People
with deep dyslexia also make visual errors (e.g., read-
ing “dog” as “dot”), and sometimes even make com-
bined visual and then semantic errors (e.g., “sympathy”
read as “orchestra,” presumably via “symphony”). Note
that it is difficult to distinguish visual and phonological
errors, so we count these as the same (and the distinc-
tion is not relevant for our model). They also tend to
make more errors on abstract words (e.g., “truth”) com-
pared to more concrete words (e.g., “chair”), which may
reflect something about the richness and robustness of
the semantic representations for these items.
One explanation of deep dyslexia involves a direct
pathway lesion as in phonological dyslexia to account
for nonword reading impairments, with additional dam-
age in the semantic pathway to account for semantic
errors (Plaut & Shallice, 1993; Plaut, 1999). Alter-
natively, a somewhat more parsimonious account can
be given based on the learned division of labor phe-
nomenon described in the previous section. In this
account, deep dyslexia results just from more severe
damage to the direct pathway than that which leads to
phonological dyslexia (see also, Friedman, 1996). Be-
cause the semantic pathway comes to depend on the di-
rect pathway to produce correct phonological outputs
during learning, a severe direct pathway lesion can re-
veal this underlying dependence by eliminating the di-
rect pathway contribution. We expect that the newly in-
dependent semantic pathway would make errors (which
would have otherwise been prevented by the direct path-
way) based on similarities among semantic representa-

tions (i.e., semantic errors), which is the key charac-
teristic of deep dyslexia. We will see that our model
instantiates this alternative account of deep dyslexia (as
well as the original one studied by Plaut & Shallice,

1993).
Surface dyslexia, in contrast with both phonologi-
cal and deep dyslexia, is characterized by the preserved
ability to read nonwords, but impairments in retrieving
semantic information from written words, and difficulty
in reading exception words, especially low-frequency
ones like “yacht.” People with surface dyslexia also
make visual errors, but do not make semantic errors.
Thus, we can interpret surface dyslexia as resulting
from damage to the semantic pathway in our distributed
lexicon model (figure 10.5), with preserved ability to
use the direct orthography to phonology pathway for
reading nonwords, regular words, and high-frequency
exception words.
The learned division of labor phenomenon is neces-
sary to explain the low-frequency exception word read-
ing impairment in surface dyslexia, as demonstrated by
the PMSP model. Because the pronunciation of excep-
tion words does not follow regularities present in other
words, it is easier to learn this exceptional mapping
via semantics instead of directly in the orthographic to
phonological representations. Thus, the direct pathway
will come to depend relatively heavily on the indirect
pathway for the pronunciation of these low-frequency
exceptions. High frequency exceptions are learned in
the direct pathway simply because the high levels of ex-
posure to these words will drive learning in the direct
pathway. Thus, when the semantic pathway is damaged,
the underlying dependence of low-frequency exceptions
on this pathway is revealed. Note that PMSP did not
actually simulate the full set of pathways. Instead, they
simulated the effect of a semantic pathway by provid-
ing partial correct input to the appropriate phonological
representations during training of their direct pathway
model, and then removed these inputs to simulate se-
mantic damage.
In the simulation exercises below, we will see how
these forms of dyslexia — phonological, deep and sur-
face — emerge from just these kinds of damage.