This thesis proposes a method to disambiguate the senses of ambiguous words by using distribution information of feature frequency. Recently, corpus-based methods, especially supervised learning methods, are much studied for Word Sense Disambiguation(WSD). The corpus-based methods which display a good performance learn the senses of words by using features extracted from the context of ambiguous words. They use the frequency of features straightly for weighting features with assumption that all features come out independently. However, if we know that the frequency of features shows a certain probability distribution, we can use distribution information for WSD.
We assume that the frequency of features shows the 2-Poisson distribution, the distribution for the correct sense and the distribution for the incorrect senses. Then we apply distribution information to the feature weighting function. Because of difficulity to estimate parameters for 2-Poisson distribution, however, we use the simplified feature weighting function. To apply 2-Poisson distribution to word sense disambiguation, we take each distribution information of features according to the topical feature and the local feature respectively. We complete the feature weighting function which considers the normalization of feature. Finally, we add feature location information by giving more weight to features extracted from same sentence of an ambiguous word.
As the result of experiments, we know that distribution information of feature frequency is useful in word sense disambiguaion. Compared with other systems which don`t use distribution information of feature frequency, the proposed method shows the highest result.
