Surrounding Word Sense Model for Japanese All-words Word Sense Disambiguation

Surrounding Word Sense Model for Japanese All-words Word Sense Disambiguation Kanako KOMIYA 1 Yuto SASAKI 2 Hajime MORITA 3 Minoru SASAKI 1 Hiroyuki SHINNOU 1 Yoshiyuki KOTANI 2 1Ibaraki University 2Kyoto University 3Tokyo University of Agriculture and Technology kanako.komiya.nlp@vc.ibaraki.ac.jp, 50010268030@st.tuat.ac.jp morita@nlp.ist.i.kyoto-u.ac.jp, minoru.sasaki.01@vc.ibaraki.ac.jp hiroyuki.shinnou.0828@vc.ibaraki.ac.jp, kotani@cc.tuat.ac.jp Abstract This paper proposes a surrounding word sense model (SWSM) that uses the distri- bution of word senses that appear near am- biguous words for unsupervised all-words word sense disambiguation in Japanese. Although it was inspired by the topic model, ambiguous Japanese words tend to have similar topics since coarse semantic polysemy is less likely to occur than that in Western languages as Japanese uses Chi- nese characters, which are ideograms. We thus propose a model that uses the dis- tribution of word senses that appear near ambiguous words: SWSM. We embed- ded the concept dictionary of an Elec- tronic Dictionary Research (EDR) elec- tronic dictionary in the system and used the Japanese Corpus of EDR for the exper- iments, which demonstrated that SWSM outperformed a system with a random baseline and a system that used a topic model called Dirichlet Allocation with WORDNET (LDAWN), especially when there were high levels of entropy for the word sense distribution of ambiguous words. 1 Introduction This paper proposes a surrounding word sense model (SWSM) for unsupervised Japanese all- words Word Sense Disambiguation (WSD). SWSM assumes that the sense distribution of sur- rounding words varies according to the sense of a polysemous word. For instance, a word “可能性” (possibility) has three senses according to the Electronic Dictionary Research (EDR) electronic dictionary (Miyoshi et al., 1996): (1) The ability to do something well (2) Its feasibility (3) The certainty of something happenings Although sense (3) is the most frequent in the prior distributions, sense (1) will be more likely when the local context includes some concepts like “人間” (man) or “誰々の” (someone’s). is challenging in practice to accurately learn theIt difference in the senses of surrounding words in an unsupervised manner, but we developed an ap- proximate model that took conditions into consid- eration. SWSM is a method for all-words WSD in- spired by the topic model (Section 2). It treats the similarities of word senses using WORDNET- WALK and it generates word senses of ambigu- ous words and their surrounding words (Section 3). First, SWSM abstracted the concepts of the concept dictionary (Section 4) and calculated the transition probabilities for priors (Section 5). Then it estimated the word senses using Gibbs Sam- pling (Section 6) . Our experiments with an EDR Japanese corpus and a Concept Dictionary (Sec- tion 7) indicated that SWSM was effective for Japanese all-words WSD (Section 8) . We dis- cuss the results (Section 9) and concludes this pa- per (Section 10) . 2 Related Work There are many methods of all-words WSD. Ped- ersen et al. (2005) proposed calculation of the se- mantic relatedness of the word senses of ambigu- ous words and their surrounding words. Some papers have reported that methods using topic …

Content truncated to 500 words