Towards Robust High Performance Word Sense Disambiguation of English Verbs Using Rich Linguistic Features

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 933 – 944, 2005. © Springer-Verlag Berlin Heidelberg 2005 Towards Robust High Performance Word Sense Disambiguation of English Verbs Using Rich Linguistic Features Jinying Chen and Martha Palmer Department of Computer and Information Science, University of Pennsylvania, Philadelphia, PA, 19104, USA {jinying, mpalmer}@cis.upenn.edu Abstract. This paper shows that our WSD system using rich linguistic features achieved high accuracy in the classification of English SENSEVAL2 verbs for both fine-grained (64.6%) and coarse-grained (73.7%) senses. We describe three specific enhancements to our treatment of rich linguistic features and pre- sent their separate and combined contributions to our system’s performance. Further experiments showed that our system had robust performance on test data without high quality rich features. 1 Introduction Word sense disambiguation (WSD) has been regarded as essential or necessary in many high-level NLP applications that require a certain degree of semantic interpreta- tion, such as machine translation, information retrieval (IR) and question answering, etc. However, previous investigations into the role of WSD in IR have shown that low accuracy in WSD negated any possible performance increase from ambiguity resolu- tion [1,2]. This suggests that improving the performance of WSD systems is crucial for applications to attain benefits from WSD. Much effort has been aimed at the creation of sense tagged corpora that can be used to develop supervised WSD systems with high accuracy. 1 However, highly polysemous words with subtle sense distinctions still pose major challenges for auto- matic systems, as evidenced in SENSEVAL2 [3]. This problem seems more serious for verbs, as indicated by the relatively poorer performance achieved by the best sys- tem in the SENSEVAL2 English lexical sample task for verbs: 56.6% accuracy, in contrast with the 64.2% accuracy for all parts-of-speech [4,5]. On the other hand, disambiguating verb senses accurately is very important for lexical selection in MT. It is also helpful for information retrieval, especially for fact retrieval systems that take full-sentence queries as their input. Therefore, this paper will focus on improving the accuracy of our supervised WSD system for verbs. We are using a linguistically rich approach for verb sense disambiguation. Linguis- tically rich approaches [5-9] utilize syntactic and/or semantic features, e.g., syntactic relations, selectional preferences, and semantic information of NP arguments of verbs, 1 http://www.senseval.org/ 934 J. Chen and M. Palmer etc. In verb sense disambiguation, Dang and Palmer's work [5] demonstrated that their system, which achieved 59.6% accuracy (62.5% in a recent report [10]) in disambigu- ating the SENSEVAL2 English verbs, benefited substantially from using rich linguis- tic features that capture information about a verb's lexical semantics. On the other hand, the performance of a system using rich linguistic features relies heavily on the quality of preprocessing, such as part-of-speech tagging, parsing, fea- ture extraction and generation, etc. How accurate and how robust can such a system be? In particular, we are interested in the following three questions: How much ad- vantage can we gain from the rich-feature approach by careful extraction and …

Content truncated to 500 words