High WSD Accuracy Using Naive Bayesian Classifier with Rich Features

High WSD accuracy using Naive Bayesian classifier with rich features Cuong Anh Le and Akira Shimazu School of Information Science, Japan Advanced Institute of Science and Technology (JAIST) 1-1, Asahidai, Tatsunokuchi, 923-1292, Ishikawa, Japan {cuonganh,shimazu}@jaist.ac.jp Abstract Word Sense Disambiguation (WSD) is the task of choosing the right sense of an ambiguous word given a context. Using Naive Bayesian (NB) classifiers is known as one of the best methods for supervised approaches for WSD (Mooney, 1996; Pedersen, 2000), and this model usually uses only a topic context represented by unordered words in a large context. In this paper, we show that by adding more rich knowledge, represented by ordered words in a local context and collocations, the NB classifier can achieve higher accuracy in comparison with the best previously published results. The features were chosen using a forward sequential selection algorithm. Our experiments obtained 92.3% accuracy for four common test words (interest, line, hard, serve). We also tested on a large dataset, the DSO corpus, and obtained accuracies of 66.4% for verbs and 72.7% for nouns. 1. Introduction WSD is always a difficult and important task in natural language processing. Its task is to determine the most appropriate sense for an ambiguous word given a context. Approaches for this work include supervised learning, unsupervised learning, and combinations of them. Except for the expense involved in building labeled datasets, supervised based methods generally give results with higher precision. Many supervised learning algorithms have been applied, such as Bayesian learning, Exemplar-Based learning, Decision Trees, Decision Lists, and Neural Networks. Despite their simplicity, NB methods are still effective when applied to WSD (Mooney, 1996; Pedersen, 2000). Before presenting the previous related studies and describing our approach, we need to define some terms that are used throughout in this paper. These are topic context, local context, and collocation. The first kind of information, which is always used for determining the senses of a word, is the topic context represented by a bag of surrounding words in a large context of the ambiguous word. The other informative resource is collocation. There are various definitions of collocation, and for our approach we define collocation as a sequence of words including the ambiguous word. Several studies, such as Leacock and Chodorow (1998), used local context for disambiguating word senses. Like them, we define local context as the words (or tags of words) assigned with their position in relation to the ambiguous word in a local context. For example, suppose that we have a context of the ambiguous word interest as follows: “yields on money-market mutual funds continued to slide, amid signs that portfolio managers expect further declines in interest rates.” Then the topic context includes the words: yields, money-market, mutual, funds, continued, . . .; Collocations include the expressions: interest rates, declines in interest, in interest rates, further declines in interest rate ,. . .; Local context is represented by the pairs: (declines,-2), (in,-1), (rates,1), (further, -3), . . . Note that words in collocations and local contexts …

Content truncated to 500 words