Japanese all-words WSD system using the Kyoto Text Analysis ToolKit

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 392–399 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Hiroyuki Shinnou, Kanako Komiya, Minoru Sasaki and Shinsuke Mori Japanese all-words WSD system using the Kyoto Text Analysis ToolKit Hiroyuki Shinnou Kanako Komiya Minoru Sasaki Shinsuke Mori Ibaraki University, Department of Computer and Information Sciences 4-12-1 Nakanarusawa, Hitachi, Ibaraki 316–8511, Japan hiroyuki.shinnou.0828@vc.ibaraki.ac.jp {kanako.komiya.nlp, minoru.sasaki.01}@vc.ibaraki.ac.jp Kyoto University, Yoshida Honcho, Sakyouku, Kyoto 606–8501, Japan mori.shinsuke.8u@kyoto-u.ac.jp Abstract In this paper, we discuss Japanese all-words word sense disambiguation (WSD) and pro- pose a new system KyWSD to achieve it. Ky- WSD uses the Kyoto Text Analysis ToolKit, a learning system building a Japanese mor- phological analysis model. It accepts plain Japanese text, segments it into words, and as- signs a sense to each segmented word. Ky- WSD is open source software that can serve as the baseline system for a Japanese all- words WSD system. Therefore, it can be useful for several Japanese semantic analysis systems and is an advancement in all-words WSD technology. Furthermore, we show that Japanese all-words WSD involves a peculiar problem different from those of general WSD and that KyWSD is adaptable and highly pre- cise. 1 Introduction Word-sense disambiguation (WSD) is a basic proce- dure of semantic analysis, but it has not been widely used in practice. This is because current WSD sys- tems adopt a supervised learning approach, limit- ing WSD target words. WSD for all words called “all-words WSD” has been studied for a long time (Navigli, 2009). However, a sense in many all- words WSD systems is defined as a concept, result- ing in coarse granularity. Furthermore, the target language is generally English. Japanese all-words WSD has not been achieved, preventing easy access to it. Given this background, we created a Japanese all-words WSD system called KyWSD1. KyWSD is 1KyTea for WSD. useful for several Japanese semantic analysis sys- tems. Using it, we can add sense features when we use a learning method to solve various NLP tasks, thereby improving precision. The substance of KyWSD is a model built using the Kyoto Text Analysis ToolKit (KyTea)2, a learn- ing system. By executing KyTea using this model, KyWSD accepts plain Japanese text, segments it into words, and assigns a sense to each segmented word (Neubig et al., 2011). Briefly KyTea is a sys- tem learning a morphological analysis model. We build KyWSD using KyTea because all-words WSD can be regarded as a kind of morphological anal- ysis. Therefore, KyTea contains a mechanism for learning a model to adapt to a target domain. The ability to use this mechanism provides KyWSD with high adaptability. For example, adding training data to KyWSD, senses to all words, but a target sense. Thus, KyWSD is an appropriate system for domain adaptation. As seen above, KyWSD provides great value as new use of KyTea. We evaluated KyWSD using a Japanese dictio- nary task in Senseval-2 (Kiyoaki Shirai, 2001). Adding training data of this task to its original train- ing data enabled …

Content truncated to 500 words