Improving Word Sense Induction by Exploiting Semantic Relevance

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1387–1391, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Improving Word Sense Induction by Exploiting Semantic Relevance Zhenzhong Zhang Institute of Software, Graduate University, Chinese Academy of Sciences, Beijing, China Zhenzhong@nfs.iscas.ac.cn Le Sun Institute of Software, Chinese Academy of Sciences, Beijing, China sunle@iscas.ac.cn Abstract Word Sense Induction (WSI) is the task of au- tomatically inducing the different senses of a target word from unannotated text. Traditional approaches based on the vector space model (VSM) represent each context of a target word as a vector of selected features (e.g. the words occurring in the context). These approaches assume that the words occurring in the context are independent and do not exploit semantic relevance between words. In this paper we propose a WSI method which can exploit se- mantic relevance between words by incorpo- rating a word graph into the framework of clustering of context vectors. The method is evaluated on the testing data of the Chinese Word Sense Induction task of the first CIPS- SIGHAN Joint Conference on Chinese Lan- guage Processing (CLP2010). Experimental results show that our method significantly out- performs the baseline methods. 1 Introduction It has been shown that using word senses instead of surface word forms can improve performance on many natural language processing tasks such as machine translation (Vickrey et al., 2005) and information retrieval (Uzuner et al., 1999; Véro- nis, 2004). Historically, using word senses usual- ly involved the use of manually compiled re- sources in which word senses were represented as a fixed list of definitions. However, there seem to be some disadvantages associated with such fixed list of senses paradigm. Firstly, since dictionaries usually contain general definitions, they can not reflect the exact contents of the con- texts where target words appear (Véronis, 2004). Secondly, because the “fixed list of senses” pa- radigm makes the fixed granularity assumption of the senses distinction, it may not be suitable in different applications (Kilgarriff, 1997; Brody and Lapata, 2009). To overcome these limitations, some tech- niques like WSI have been proposed for disco- vering words senses automatically from unanno- teted corpuses. WSI algorithms are usually based on the Distributional Hypothesis which shows that words with similar meanings appear in simi- lar contexts (Harris, 1954). This concept can be leveraged to induce different senses of a target word by clustering the contexts where the target word appears. Much work in WSI is based on the vector space model, in which each context of a target word is represented by a vector of selected fea- tures (e.g. the words occurring in the context). These context vectors are clustered and the re- sulting clusters are taken to represent the induced senses. However, when constructing context vec- tors, the approaches based on VSM assume that the words occurring in the contexts are indepen- dent and do not exploit semantic relevance be- tween words. This will cause the problem that two contexts using semantically related but dis- …

Content truncated to 500 words