Learning Probabilistic Models of Word Sense Disambiguation