LEARNING OF WORD SENSE DISAMBIGUATION RULES BY BELIEF NETWORKS

摘要

This paper uses Belief Networks (BN) to solve word sense disambiguation (WSD) problems. For classification problems, the Naive Bayes (NB) is widely used because it generates high performance rules regardless of the simplicity of the model. We use a little more complex model than the NB to get better rules, that is the BN. In the experiments, we attacked Japanese Dictionary Task in SENSEVAL2 and evaluated the BN by comparing it with the NB. One of the features of our BN is that unlabeled data is available in learning. Here, we report on an experiment in which unlabeled data was used in learning.