| |
Abstract:
Children with Specific Language Impairment (SLI) sometimes
exhibit difficulties in resolving pronominal anaphora (van der Lely
& Stollwerck 1997), which has been taken as evidence that SLI
involves a syntax-specific impairment. Our study explores the
alternative view that grammatical deficits in SLI are sequelae of
impaired phonological processing caused by anomalies in speech
processing (Joanisse & Seidenberg 1998, Leonard 1998, Tallal et
al. 1996). On this account, impaired phonological processing
affects the retention of information in working memory, which
interferes with some aspects of syntactic processing.
We developed a connectionist model of sentence recognition that
learned to resolve pronominal anaphora in sentences like "Mary says
Sally(i) likes herself(i)'' and "Mary(i) says Sally likes her(i).''
The model recognizes sentences by mapping the phonological forms of
words to their meanings, in sequence. The input representation
encoded one- and two-syllable words using distributed phonological
features; the output consisted of semantic feature representations
derived from WordNet (Miller 1990). Thus, the sentence "Mary likes
Sally'' was input to the model as the sequence /mEri/... /lajks/...
/saeli/, which was mapped to the output sequence [+person +female
+Mary]... [+judge +consider+positive]... [+person +female
+Sally].
Anaphoric resolution was represented by adding the semantics of
antecedents to the semantics of bound pronouns. For instance the
"herself'' in "Mary likes herself'' was represented as [+PRO
+person +female +Mary] whereas in ``Mary says Jane likes herself''
it was [+PRO +person +female +Jane].
Two recurrent connectionist networks were trained using a corpus
of 1324 simple and complex, transitive, intransitive and
ditransitive English sentences, including a subset of sentences
containing bound and unbound pronouns. The two models were the same
in all respects except that in one of them, a phonological
impairment was simulated by introducing gaussian noise on the input
during training. This created slightly distorted phonological
inputs such as would result from a perceptual deficit.
When tested on novel sentences, both networks were able to
accurately compute the correct semantics for words, and to use
syntactic context to disambiguate homophonous nouns and verbs
(e.g., "Mary'' vs. ``marry''). However, the impaired network showed
greater difficulty resolving bound anaphors, especially in complex
sentences. Thus, it produced more errors on complex sentences such
as "Mary says Sally gave the ball to herself'' (75% correct)
compared to simple sentences such as ``Mary likes herself'' (85%
correct). The unimpaired network produced only a few errors on
either sentence type (92% and 100% correct, respectively).
These results are consistent with several aspects of the data
concerning syntactic deficits in SLI. As in SLI, the phonologically
impaired network was not globally impaired in sentence processing;
rather, difficulties were limited to the binding of pronouns and
anaphors. Moreover, this deficit was graded rather than
categorical: the model produced correct responses on many anaphoric
sentences, especially in simpler constructions, as observed in
children with SLI, who perform worse than controls on such
sentences, but above chance. The results are discussed in terms of
the importance of phonology and working memory in sentence
processing and theories of SLI.
Joanisse, M.F. & Seidenberg, M.S. (1998), Specific Language
Impairment: A Deficit in Grammar or Processing?, Trends in
Cognitive Sciences 2, 240-247.
Leonard, L. (1998) Specific Language Impairments in Children.
Cambridge MA: MIT Press.
Miller, G.A. (1990) WordNet: An on-line lexical database,
International Journal of Lexicography 3, 235-312.
|
