Trajectory Based Word Sense Disambiguation

Trajectory Based Word Sense Disambiguation Xiaojie Wang † ‡ Yuji Matsumoto † † Graduate School of Information Science, Nara Institute of Science and Technology 8916-5 Takayama, Ikoma, Nara, 630-0192, Japan ‡School of Information Engineering, Beijing University of Posts and Technology Beijing, 100876, China {xiaoji-w, matsu}@is.naist.jp Abstract Classifier combination is a promising way to improve performance of word sense disambiguation. We propose a new combinational method in this paper. We first construct a series of Naïve Bayesian classifiers along a sequence of orderly varying sized windows of context, and perform sense selection for both training samples and test samples using these classifiers. We thus get a sense selection trajectory along the sequence of context windows for each sample. Then we make use of these trajectories to make final k-nearest-neighbors-based sense selection for test samples. This method aims to lower the uncertainty brought by classifiers using different context windows and make more robust utilization of context while perform well. Experiments show that our approach outperforms some other algorithms on both robustness and performance. 1 Introduction Word sense disambiguation (WSD) has long been a central issue in Natural Language Processing (NLP). In many NLP tasks, such as Machine Translation, Information Retrieval etc., WSD plays a very important role in improving the quality of systems. Many different algorithms have been used for this task, including some machine learning (ML) algorithms, such as Naïve Bayesian model, decision trees, and example based learners. Since different algorithms have different strengths and perform well on different feature space, classifier combination is a reasonable candidate to achieve better performance by taking advantages of different approaches. In the field of ML, ensembles of classifiers have been shown to be successful in last decade (Dietterich 1997). For the specific task of WSD, classifier combination has been received more and more attention in recent years. Kilgarriff and Rosenzweig (2000) presented the first empirical study. They combined the output of the participating SENSEVAL1 systems via simple voting. Pedersen (2000) built an ensemble of Naïve Bayesian classifiers, each of which is based on lexical features that represent co-occurring words in varying sized windows of context. The sense that receives majority of the votes was assigned as the final selection. Stevenson and Wilks (2001) presented a classifier combination framework where three different disambiguation modules were combined using a memory-based approach. Hoste et al. (2002) used word experts consisted of four memory-based learners trained on different context. Output of the word experts is based on majority voting or weighted voting. Florian et al.(2002) and Florian and Yarowsky (2002) used six different classifiers as components of their combination. They compared several different strategies of combination, which include combining the posterior distribution, combination based on order statistics and several different voting. Klein et al. (2002) combined a number of different first-order classifiers using majority voting, weighted voting and maximum entropy. In Park (2003), a committee of classifiers was used to learn from the unlabeled examples. The label of an unlabeled example is predicted by weighted majority voting. Frank …

Content truncated to 500 words