Word Sense Disambiguation by Combining Labeled Data Expansion and Semi-Supervised Learning Method

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 676–685, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Word Sense Disambiguation by Combining Labeled Data Expansion and Semi-Supervised Learning Method Sanae Fujita NTT Communication Science Lab. fujita.sanae@lab.ntt.co.jp Akinori Fujino NTT Communication Science Lab. fujino.akinori@lab.ntt.co.jp Abstract Lack of training data is one of the severest problems facing word sense disambigua- tion (WSD). To overcome the problem, we propose a method that combines automatic labeled data expansion (Step-1) and semi- supervised learning (Step-2). The Step-1 and 2 methods are both effective, but their combination has a synergistic effect. In this paper, in Step-1, we automatically extract reliable labeled data from raw cor- pora using dictionary example sentences, even for infrequent and unseen senses (which do not appear in training data, but appear in a dictionary). Then, in Step-2, we apply a semi-supervised classifier and obtain a improvement using easy-to-get unlabeled data. In this step, we also show that we can guess even unseen senses. We target a SemEval-2010 Japanese lex- ical sample WSD task. Both the Step-1 and Step-2 methods performed better than the best published result (76.4 %). Fur- thermore, the combined method achieved much higher accuracy (84.2 %). 1 Introduction Many words have multiple meanings that change depending on the context. Recently, it has been confirmed that word sense disambiguation (WSD) improves certain NLP applications such as parse selection (Fujita et al., 2007) or Machine Trans- lation (Chan et al., 2007). In international WSD competitions such as SemEval, many tasks have been proposed, which shows that WSD is a prob- lem that attracts a lot of interest. In this paper, we experiment on the Japanese WSD task from the most recent competition, SemEval-2010 (Oku- mura et al., 2010). Various methods have been proposed for WSD (Navigli, 2009). Unsupervised approaches such as clustering based methods (Pedersen, 2006) and ex- tended Lesk (Lesk, 1986) have been shown to do well (Baldwin et al., 2010), although in general, they are beaten by supervised approaches if train- ing data are provided (Tanaka et al., 2007). With the Japanese WSD tasks at SENSEVAL-2 and SemEval-2010, supervised approaches achieved the best results (Murata et al., 2003; Okumura et al., 2010). However, the lack of training data is a severe problem with non-English languages. Also in the Japanese WSD task, there are only 50 given training instances for each target word and this is insufficient. Two main types of methods have been proposed to compensate for a lack of training data. One type is the semi-supervised learning method (Niu et al., 2005; Pham et al., 2005) or bootstrapping (Mihal- cea, 2002, 2004; Yarowsky, 1995). These meth- ods use labeled data and unlabeled data, and this is beneficial because unlabeled training data is easy to obtain. These methods are effective and have high applicability, but unfortunately, there is one problem in that this method cannot obtain training data for senses in the lexicon that do not appear in the training data (we call …

Content truncated to 500 words