Continuous Task - Speci c Categories for Disambiguation : Putting Lexical Constraints ( back ) in the Lexicon

Syntactic ambiguity resolution often appears to be guided by the use of lexical constraint information (e. A traditional linguistic framework of discrete syntactic categories can complicate the modeling of such lexical constraints. There is evidence that syntactic ambiguity resolution is exposure-based, i.e. present analyses are innuenced by past experiences (e.g. Mitchell and Cuetos (1991)). An important feature of such exposure-based models is that similar instances are recognized in order to be able to make the appropriate generalizations. We have developed a method to construct a continuous representation of lexical similarities { the Lexical Space { based on local (within several words to the right and left) distributional properties of words in a corpus. The distributional organization of LexSpace can be exploited for a task-speciic syntactic categorization or clustering of words with an arbitrary degree of granularity according to the demands of a speciic language processing task. In this paper we describe several models which perform a disambiguation task on the basis of learning from examples. We compare the use of syntactic representations of words from LexSpace with task-speciic generalizations inherent to the Example-based models. We consider two tasks, prepositional phrase attachment and direct object/sentence complement ambiguity resolution, and compare the generalization performance of the methods on unseen test material. The results suggest that Lexical Space provides us with rich lexical representations that can also be used for similarity-based analysis as required in exposure based strategies. Moreover, the LexSpace representation can be used as input to connectionist models. The Lexical Space representations were found to work successfully, and it was shown that the Example-based method generalizes to new test material with an accuracy (84.1%) that is equal to the best method known for this task in the literature (Collins and Brooks, 1995). However, these optimal results were not obtained with the Lexical Space method but with the Overlap IG metric.