Mining the Web for Sense Discrimination Patterns

In this paper we present a method for mining the Web in order to extract lexical patterns that help in discriminating the senses of a given polysemic word. These patters are defined as sets and sequences of words strongly related to each sense of the word. To discover the patterns, the method first determines the different senses of the word from a reference lexical database, and then it uses the set of synonyms from each sense as search patterns on the Web. The purpose is to create a corpus of usage cases per sense, downloading snippets via fast search engines. Finally, it applies a well-known association discovery data mining technique to select the most relevant lexical patterns for each word sense. The preliminary results indicate that making sense out of the Web is possible and the discovered patters should be of great benefit in tasks such as information retrieval and machine translation.