IMPROVING THE PERFORMANCE OF BAYESIAN AND SUPPORT VECTOR CLASSIFIERS IN WORD SENSE DISAMBIGUATION USING POSITIONAL INFORMATION

We explore word position-sensitive models and their realizations in word sense disambiguation tasks when using Naive Bayes and Support Vector Machine classifiers. It is shown that a straightforward incorporation of word positional information fails to improve the performance of either method on average. However, we demonstrate that our special kernel that takes into account word positions statistically significantly improves the classification performance. For Support Vector Machines, we apply this kernel instead of the ordinary Bag-of-Words kernel, and for the Bayes classifier the kernel is used for smoothed density estimation. We discuss the benefits and drawbacks of position-sensitive and kernel-smoothed models as well as analyze and evaluate the effects of these models on a subset of the Senseval-3 data.