Polysemous words turn out to be problematic, because word-sense disambiguation still hasn't been solved satisfactorily. Although recent advances have been made using supervised learning techniques [20,86], the amount of training data is prohibitive for application on a large scale. There are other problems related to polysemy, in terms of the organization of the lexicon.
The collection edited by Klavans [67] provides good coverage of these issues. The article by Nirenburg et al. [89] defends the use of ontologies as a basis for lexical entries and shows how this facilitates the encoding of polysemous entries.
Machine readable dictionaries are obviously a good knowledge source for polysemy and can also be used for resolving ambiguity. Vanderwende [115] illustrates an iterative approach for extracting information from MRD's in the face of ambiguity. In the first pass, relations are extracted using unambiguous patterns. Then in subsequent passes, these relations can be used in patterns for ambiguous relations. Relations from earlier passes can also be used for syntactic disambiguation. For instance, PP-attachment might go with the verb if there is a known instrument relation from the with-complement to the verb. Note that Hindle and Rooth [56] would accomplish the same task via lexical associations, as discussed later. Another problem with polysemous words is how to determine the relation among the different senses. Dolan [34] shows how MRD's can be used to detect metaphorical sense extensions. The basic idea is to first check for two senses that are similar in structure but that refer to objects from distinct semantic classes. Then, if significant paths between the objects can be detected in a semantic network (derived from the MRD), a metaphoric extension is likely.