MSS: Investigating the Effectiveness of Domain Combinations and Topic Features for Word Sense Disambiguation

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 383–386, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics MSS: Investigating the Effectiveness of Domain Combinations and Topic Features for Word Sense Disambiguation Sanae Fujita Kevin Duh Akinori Fujino Hirotoshi Taira Hiroyuki Shindo NTT Communication Science Laboratories {sanae, kevinduh, taira, a.fujino, shindo}@cslab.kecl.ntt.co.jp Abstract We participated in the SemEval-2010 Japanese Word Sense Disambiguation (WSD) task (Task 16) and focused on the following: (1) investigating domain differences, (2) incorporating topic fea- tures, and (3) predicting new unknown senses. We experimented with Support Vector Machines (SVM) and Maximum Entropy (MEM) classifiers. We achieved 80.1% accuracy in our experiments. 1 Introduction We participated in the SemEval-2010 Japanese Word Sense Disambiguation (WSD) task (Task 16 (Okumura et al., 2010)), which has two new char- acteristics: (1) Both training and test data across 3 or 4 domains. The training data include books or magazines (called PB), newspaper articles (PN), and white papers (OW). The test data also include documents from a Q&A site on the WWW (OC); (2) Test data include new senses (called X) that are not defined in dictionary. There is much previous research on WSD. In the case of Japanese, unsupervised approaches such as extended Lesk have performed well (Bald- win et al., 2010), although they are outperformed by supervised approaches (Tanaka et al., 2007; Murata et al., 2003). Therefore, we selected a su- pervised approach and constructed Support Vector Machines (SVM) and Maximum Entropy (MEM) classifiers using common features and topic fea- tures. We performed extensive experiments to in- vestigate the best combinations of domains for training. We describe the data in Section 2, and our sys- tem in Section 3. Then in Section 4, we show the results and provide some discussion. 2 Data Description 2.1 Given Data We show an example of Iwanami Kokugo Jiten (Nishio et al., 1994), which is a dictionary used as a sense inventory. As shown in Figure 1, each en- try has POS information and definition sentences including example sentences. We show an example of the given training data in (1). The given data are morphologically ana- lyzed and partly tagged with Iwanami’s sense IDs, such as '37713-0-0-1-1( in (1). (1) <mor pos=' � �-� } ( rd=' � � ( bfm=' � �( sense= '37713-0-0-1-1( > 1<</mor> This task includes 50 target words that were split into 219 senses in Iwanami; among them, 143 senses including two Xs that were not defined in Iwanami, appear in the training data. In the test data, 150 senses including eight Xs appear. The training and test data share 135 senses including two Xs; that is, 15 senses including six Xs in the test data are unseen in the training data. 2.2 Data Pre-processing We performed two preliminary pre-processing steps. First, we restored the base forms because the given training and test data have no informa- tion about the base forms. (1) shows an example of the original morphological data, and then we added …

Content truncated to 500 words