Learning topic knowledge to improve Chinese word sense disambiguation

This paper addresses an issue of incorporating topic knowledge to improve Chinese word sense disambiguation. The key is how to learn topic knowledge as features in the design of classifiers for disambiguating word senses. This paper presents two solutions to learn topic knowledge. In the first solution, a Chinese domain knowledge dictionary named NEUKD is used to generate domain feature set. However, due to the limited coverage of the NEUKD, a constrained clustering algorithm is adopted for dictionary expansion. The second method is to build topic feature set by utilizing the Latent Dirichlet Allocation (LDA) algorithm on a large scale unlabeled corpus. Experiments on the SENSEVAL-3 Chinese dataset demonstrated that integrating topic knowledge improve the performance of Chinese word sense disambiguation.

I. Introduction

The goal of word sense disambiguation (WSD) is to assign an appropriate sense to an ambiguous word within a given context. A variety of techniques for supervised WSD have demonstrated reasonable performance, such as exemplar-based learning [1], decision list [2], maximum entropy model [3], Naive Bayes model [4] [5]. Among these supervised approaches, the sense ambiguity of words is resolved with the help of the contexts of their occurrences. Two types of features local collocation features (LCF) and topical contextual features (TCF) are commonly used in WSD studies to represent the contexts [4] [6], such as local words or part-of-speech (POS) tags with position information, bi-gram templates, collocations, and syntactic features. LCF and TCF generally take morphological or syntactic information into account.