Corpus-based semantic lexicon induction with web-based corroboration

Various techniques have been developed to automatically induce semantic dictionaries from text corpora and from the Web. Our research combines corpus-based semantic lexicon induction with statistics acquired from the Web to improve the accuracy of automatically acquired domain-specific dictionaries. We use a weakly supervised bootstrapping algorithm to induce a semantic lexicon from a text corpus, and then issue Web queries to generate co-occurrence statistics between each lexicon entry and semantically related terms. The Web statistics provide a source of independent evidence to confirm, or disconfirm, that a word belongs to the intended semantic category. We evaluate this approach on 7 semantic categories representing two domains. Our results show that the Web statistics dramatically improve the ranking of lexicon entries, and can also be used to filter incorrect entries.