Improving Word Sense Discrimination with Gloss Augmented Feature Vectors

This paper presents a method of unsupervised word sense discrimination that augments co–occurrence feature vectors derived from raw untagged corpora with information from the glosses found in a machine readable dictionary. Each content word that occurs in the context of a target word to be discriminated is represented by a co-occurrence feature vector. Each of these vectors is augmented with the content words that occur in the glosses of the different possible meanings of the word it represents. Then these vectors are averaged to create a vector that represents that context of the target word. Discrimination is carried out by clustering all of the vectors associated with the contexts in which the target word occurs. We show via an evaluation with the Senseval-2, line, hard and serve corpora that feature vectors augmented with gloss information from WordNet significantly improve discrimination performance when limited data is available.