A connectionist approach to word sense disambiguation

The title of Cottrell's book mentions only two concepts: connectionism and lexical disambiguation. That's misleading, because the book has much more to offer than just that. Among the topics addressed are parsing, agrammatism, connectionist inheritance hierarchies, and structural ambiguity, and it is the integration of this wide-ranging set of topics that is one of the strengths of the work. The book appears four years after the 1985 University of Rochester dissertation upon which it is based. Thus the flavor of connectionism that Cottrell uses is the coarsegrained localist representations used at Rochester in the early 1980s, in which each node in the network represents a concept. This is in contrast to the distributed representations ("PDP") that became popular in the latter part of the decade, in which many nodes may contribute to the representation of a concept (Rumelhart and McClelland 1986). Cottrell has taken advantage of the delay in publication to restructure the work substantially and to add discussions of the later research. He seems to suggest (p. 7) that distributed representations are generally preferable because they can learn, whereas localist networks like his own need to be individually hand designed. Nevertheless, this research shows that there is considerable appeal in hand-designed, localist networks. Cottrell takes work in psycholinguistics as the starting point for his model of lexical access and disambiguation. In the early 1980s, it was discovered that in many circumstances, people subconsciously consider all meanings of an ambiguous word, even if the preceding context makes one alternative preferable a priori. For example, the floral sense of the word rose is activated even when one hears The congregation rose. Within a few hundred milliseconds all senses but the one chosen as correct become deactivated again. (While subsequent research has qualified these resuits somewhat--see Gorfein 1989--the basic principle has proven to be robust.) The usual explanation for these results is in terms of priming and spreading activation in a semantic network, so a localist model is very natural. The input to Cottrelrs networks is a string of words forming a syntactically simple sentence, such as Bob threw a ball to the dog. This is done by activating the nodes corresponding to the words. The activation of a node causes the activation of those other nodes in the network to which it is connected by excitory links and the deactivation of those to which it is connected by inhibitory links. A node can receive activation and inhibition at the same time; for example, an ambiguous word will send activation to all its senses, but the senses will be mutally inhibitory. Thus the network may be unstable for some time until it settles down into a pattern of activation that represents its "output"; the nodes representing the relevant concepts are activated and other nodes aren't. In the case of an ambiguous word, the correct meaning in context will presumably receive activation from more sources, or be pre-activated by the preceding context, and thus be able eventually to force its competitors into inhibition. This final pattern of activation may be construed as the interpretation of the sentence. After the word-sense selection network, there are two more networks, running in parallel with one another: one for determining case roles and one for syntactic analysis. The case role network uses an "exploded" notion of cases; that is, rather than having one node representing, say, the agent role, Cottrell has one node for the agent of a propel action, one for the agent of a vomit action, and so on. (The topic area of Cottrell's example sentences ranges from baseball to emesis.) This seems counter-intuitive, or unparsimonious at the very least; but I must admit that, modern linguistic theory notwithstanding, I know of no particular psycholinguistic evidence for the reality of a single concept of, say, agency that is activated for any and every sentence that involves an agent. A feature of the parsing network is that it need not be constructed by hand; rather, it is automatically generated from a grammar and lexicon by a Lisp program. It parses only the very simple one-clause sentences needed to test the other parts of the system. Unlike the other parts of the system, the parser has no special claim to psychological reality. However, the minimal-attachment strategy of structural ambiguity resolution (namely, to attach a new constituent in the way that creates the fewest new nodes) "falls out" as a natural consequence of the design. Cottrell includes an interesting discussion of his system's predictions for aphasia. If the system has some psychological reality, then one would expect that "damage" to the network would result in behavior similar to that of aphasic patients. For example, if the connection between the case