Classifier Optimization and Combination in the English All Words Task

Abstract

We report on the use of machine learning techniques for word sense disambiguation in the English all words task of SENSEVAL2. The task was to automatically assign the appropriate sense to a possibly ambiguous word form given its context. A "word expert" approach was adopted, leading to a set of classifiers, each specialized in one single word form-POS combination. Experts consist of multiple classifiers trained on Semcor using two types of learning techniques, viz. memory-based learning and rule-induction. Through optimization by cross-validation of the individual classifiers and the voting scheme for combining them, the best possible word expert was determined. Results show that especially memory-based learning in a word-expert approach is a feasible method for unrestricted word-sense disambiguation, even with limited training data.