Abstract
Word Sense Disambiguation (WSD) is the task of selecting the meaning of a word based on the context in which the word occurs. The principal statistical WSD approaches are supervised and unsupervised learning. The Lesk method is an example of unsupervised disambiguation. We present a measure for sense assignment useful for the simple Lesk algorithm. We use word co-occurrences of the gloss and the context, which is statistical information retrieved from the Web. In the SemCor data our method always gives an answer. On the Senseval 2 data, our variant of the Lesk method outperformed some other Lesk-based methods.