Density Maximization in Context-Sense Metric Space for All-words WSD

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 884–893, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Density Maximization in Context-Sense Metric Space for All-words WSD Koichi Tanigaki†‡ Mitsuteru Shiba† Tatsuji Munaka† Yoshinori Sagisaka‡ † Information Technology R&D Center, Mitsubishi Electric Corporation 5-1-1 Ofuna, Kamakura, Kanagawa 247-8501, Japan ‡ Global Information and Telecommunication Institute, Waseda University 1-3-10 Nishi-Waseda, Shinjuku-ku, Tokyo 169-0051, Japan Abstract This paper proposes a novel smoothing model with a combinatorial optimization scheme for all-words word sense disam- biguation from untagged corpora. By gen- eralizing discrete senses to a continuum, we introduce a smoothing in context-sense space to cope with data-sparsity result- ing from a large variety of linguistic con- text and sense, as well as to exploit sense- interdependency among the words in the same text string. Through the smoothing, all the optimal senses are obtained at one time under maximum marginal likelihood criterion, by competitive probabilistic ker- nels made to reinforce one another among nearby words, and to suppress conflicting sense hypotheses within the same word. Experimental results confirmed the superi- ority of the proposed method over conven- tional ones by showing the better perfor- mances beyond most-frequent-sense base- line performance where none of SemEval- 2 unsupervised systems reached. 1 Introduction Word Sense Disambiguation (WSD) is a task to identify the intended sense of a word based on its context. All-words WSD is its variant, where all the unrestricted running words in text are expected to be disambiguated. In the all-words task, all the senses in a dictionary are potentially the target destination of classification, and purely supervised approaches inherently suffer from data-sparsity problem. The all-words task is also character- ized by sense-interdependency of target words. As the target words are typically taken from the same text string, they are naturally expected to be inter- related. Disambiguation of a word should affect other words as an important clue. From such characteristics of the task, knowledge-based unsupervised approaches have been extensively studied. They compute dictionary-based sense similarity to find the most related senses among the words within a certain range of text. (For reviews, see (Agirre and Edmonds, 2006; Navigli, 2009).) In recent years, graph-based methods have attracted considerable attentions (Mihalcea, 2005; Navigli and Lapata, 2007; Agirre and Soroa, 2009). On the graph structure of lexical knowledge base (LKB), random-walk or other well-known graph-based techniques have been applied to find mutually related senses among target words. Unlike earlier studies disambiguating word-by-word, the graph-based methods obtain sense-interdependent solution for target words. However, those methods mainly focus on modeling sense dis- tribution and have less attention to contextual smoothing/generalization beyond immediate context. There exist several studies that enrich immedi- ate context with large corpus statistics. McCarthy et al. (2004) proposed a method to combine sense similarity with distributional similarity and config- ured predominant sense score. Distributional sim- ilarity was used to weight the influence of context words, based on large-scale statistics. The method achieved successful WSD accuracy. Agirre et al. (2009) used …

Content truncated to 500 words