23.1.5 Homonymy, Polysemy, and Synonymy

Since the vector space model is based solely on the use of simple terms, it is useful to consider the effect that various lexical semantic phenomena may have on the model. Consider a query containing the word canine, a word that has senses meaning something like tooth and dog. A query containing canine will be judged similar to documents making use of either of these senses. However, given that users are probably only interested in one of these senses, the documents containing the other sense will be judged non-relevant. Homonymy and polysemy, therefore, can have the effect of reducing precision by leading a system to return documents irrelevant to the user's information need.

Now consider a query consisting of the lexeme dog. This query will be judged close to documents that make frequent use of the term dog, but may fail to match documents that use close synonyms like canine, as well as documents that use hyponyms such as Malamute. Synonymy and hyponymy, therefore, can have the effect of reducing recall by causing the retrieval system to miss relevant documents.

Note that it is inaccurate to state flatly that polysemy reduces precision, and synonymy reduces recall since, as we discuss on page 11, both measures are relative to a fixed cutoff. As a result, every non-relevant document that rises above the cutoff due to polysemy takes up a slot in the fixed size return set, and may thus push a relevant document below threshold, thus reducing recall. Similarly, when a document is missed due to synonymy, a slot is opened in the return set for a non-relevant document, potentially reducing precision as well.

These issues lead naturally to the question of whether or not word sense disambiguation can help in information retrieval. The current evidence on this point is mixed, with some experiments reporting a gain using disambiguation-like techniques (Schütze and Pedersen, 1995), and others reporting either no gain, or a degradation in performance (Krovetz and Croft, 1992; Sanderson, 1994; Voorhees, 1998).