Corpus-Based Statistical Sense Resolution

CORPUS-BASED STATISTICAL SENSE RESOLUTION Claudia Leacock, 1 Geoffrey Towell, 2 Ellen Voorhees 2 1Princeton University, Cognitive Science Laboratory, Princeton, New Jersey 08542 2Siemens Corporate Research, Inc., Princeton, New Jersey 08540 ABSTRACT The three corpus-based statistical sense resolution methods studied here attempt to infer the correct sense of a polyse- mous word by using knowledge about patterns of word co- occurrences. The techniques were based on Bayesian decision theory, neural networks, and content vectors as used in in- formation retrieval. To understand these methods better, we posed s very specific problem: given a set of contexts, each containing the noun line in a known sense, construct a classi- fier that selects the correct sense of line for new contexts. To see how the degree of polysemy affects performance, results from three- and slx-sense tasks are compared. The results demonstrate that each of the techniques is able to distinguish six senses of line with an accuracy greater than 70%. Furthermore, the response patterns of the classifiers are, for the most part, statistically indistinguishable from one another. Comparison of the two tasks suggests that the degree of difficulty involved in resolving individual senses is a greater performance factor than the degree of polysemy. 1. INTRODUCTION The goal of this study is to systematically explore the effects of such variables as the number of senses per word and the number of training examples per sense on corpus-based statistical sense resolution methods. To en- able us to study the effects of the number of word senses, we selected the highly polysemous noun line, which has 25 senses in WordNet. 1 Automatic sense resolution systems need to resolve highly polysemous words. As Zipf [2] pointed out in 1945, frequently occurring words tend to be polysemous. The words encountered in a given text will have far greater polysemy than one would assume by simply tak- ing the overall percentage of po]ysemous words in the language. Even though 86% of the nouns in WordNet have a single sense, the mean number of WordNet senses per word for the one hundred most frequently occurring nouns in the Brown Corpus is 5.15, with only eight words having a single sense. 1WordNet is a lexical database developed by George Miller and his colleagues at Princeton Urdversity.[l] 2. PREVIOUS WORK Yarowsky [3] compared the Bayesian statistical method with the published results of other corpus-based statisti- cal models. Although direct comparison was not possible due to the differences in corpora and evaluation criteria, he minimizes these differences by using the same words, with the same definition of sense. He argues, convinc- ingly, that the Bayesian model is as good as or better than the costlier methods. As a pilot for the present study, a two-sense distinction task for line was run using the content vector and neural network classifiers, achieving greater than 90~ accuracy. A three-sense distinction task was then run, which is re- ported in Voorhees, st. al. [4], and discussed in Section 5. 3. METHODOLOGY The training and testing contexts were …

Content truncated to 500 words