Exemplar-Based Word Sense Disambiguation” Some Recent Improvements

Exemplar-Based Word Sense Disambiguation: Some Recent Improvements Hwee Tou Ng DSO National Laboratories 20 Science Park Drive Singapore 118230 nhweetou@dso, org. sg Abstract In this paper, we report recent improve- ments to the exemplar-based learning ap- proach for word sense disambiguation that have achieved higher disambiguation accu- racy. By using a larger value of k, the number of nearest neighbors to use for de- termining the class of a test example, and through 10-fold cross validation to auto- matically determine the best k, we have ob- tained improved disambiguation accuracy on a large sense-tagged corpus first used in (Ng and Lee, 1996). The accuracy achieved by our improved exemplar-based classifier is comparable to the accuracy on the same data set obtained by the Naive-Bayes al- gorithm, which was reported in (Mooney, 1996) to have the highest disambiguation accuracy among seven state-of-the-art ma- chine learning algorithms. 1 Introduction Much recent research on word sense disambigua- tion (WSD) has adopted a corpus-based, learning approach. Many different learning approaches have been used, including neural networks (Leacock et al., 1993), probabilistic algorithms (Bruce and Wiebe, 1994; Gale et al., 1992a; Gale et al., 1995; Leacock et al., 1993; Yarowsky, 1992), decision lists (Yarowsky, 1994), exemplar-based learning algorithms (Cardie, 1993; Ng and Lee, 1996), etc. In particular, Mooney (1996) evaluated seven state-of-the-art machine learning algorithms on a common data set for disambiguating six senses of the word "line". The seven algorithms that he eval- uated are: a Naive-Bayes classifier (Duda and Hart, 1973), a perceptron (Rosenblatt, 1958), a decision- tree learner (Quinlan, 1993), a k nearest-neighbor classifier (exemplar-based learner) (Cover and Hart, 1967), logic-based DNF and CNF learners (Mooney, 1995), and a decision-list learner (Rivest, 1987). His results indicate that the simple Naive-Bayes al- gorithm gives the highest accuracy on the "line" corpus tested. Past research in machine learning has also reported that the Naive-Bayes algorithm achieved good performance on other machine learn- ing tasks (Clark and Niblett, 1989; Kohavi, 1996). This is in spite of the conditional independence as- sumption made by the Naive-Bayes algorithm, which may be unjustified in the domains tested. Gale, Church and Yarowsky (Gale et al., 1992a; Gale et al., 1995; Yarowsky, 1992) have also successfully used the Naive-Bayes algorithm (and several extensions and variations) for word sense disambiguation. On the other hand, our past work on WSD (Ng and Lee, 1996) used an exemplar-based (or near- est neighbor) learning approach. Our WSD pro- gram, LEXAS, extracts a set of features, including part of speech and morphological form, surrounding words, local collocations, and verb-object syntactic relation from a sentence containing the word to be disambiguated. These features from a sentence form an example. LEXAS then uses the exemplar-based learning algorithm PEBLS (Cost and Salzberg, 1993) to find the sense (class) of the word to be disam- biguated. In this paper, we report recent improvements to the exemplar-based learning approach for WSD that have achieved higher disambiguation accuracy. The exemplar-based learning algorithm PEBLS contains a number of parameters that must be …

Content truncated to 500 words