Domain Adaptation with Active Learning for Word Sense Disambiguation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 49–56, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Domain Adaptation with Active Learning for Word Sense Disambiguation Yee Seng Chan and Hwee Tou Ng Department of Computer Science National University of Singapore 3 Science Drive 2, Singapore 117543 {chanys, nght}@comp.nus.edu.sg Abstract When a word sense disambiguation (WSD) system is trained on one domain but ap- plied to a different domain, a drop in ac- curacy is frequently observed. This high- lights the importance of domain adaptation for word sense disambiguation. In this pa- per, we first show that an active learning ap- proach can be successfully used to perform domain adaptation of WSD systems. Then, by using the predominant sense predicted by expectation-maximization (EM) and adopt- ing a count-merging technique, we improve the effectiveness of the original adaptation process achieved by the basic active learn- ing approach. 1 Introduction In natural language, a word often assumes different meanings, and the task of determining the correct meaning, or sense, of a word in different contexts is known as word sense disambiguation (WSD). To date, the best performing systems in WSD use a corpus-based, supervised learning approach. With this approach, one would need to collect a text cor- pus, in which each ambiguous word occurrence is first tagged with its correct sense to serve as training data. The reliance of supervised WSD systems on an- notated corpus raises the important issue of do- main dependence. To investigate this, Escudero et al. (2000) and Martinez and Agirre (2000) con- ducted experiments using the DSO corpus, which contains sentences from two different corpora, namely Brown Corpus (BC) and Wall Street Jour- nal (WSJ). They found that training a WSD system on one part (BC or WSJ) of the DSO corpus, and applying it to the other, can result in an accuracy drop of more than 10%, highlighting the need to per- form domain adaptation of WSD systems to new do- mains. Escudero et al. (2000) pointed out that one of the reasons for the drop in accuracy is the dif- ference in sense priors (i.e., the proportions of the different senses of a word) between BC and WSJ. When the authors assumed they knew the sense pri- ors of each word in BC and WSJ, and adjusted these two datasets such that the proportions of the differ- ent senses of each word were the same between BC and WSJ, accuracy improved by 9%. In this paper, we explore domain adaptation of WSD systems, by adding training examples from the new domain as additional training data to a WSD system. To reduce the effort required to adapt a WSD system to a new domain, we employ an ac- tive learning strategy (Lewis and Gale, 1994) to se- lect examples to annotate from the new domain of interest. To our knowledge, our work is the first to use active learning for domain adaptation for WSD. A similar work is the …

Content truncated to 500 words