Automatic Domain Adaptation for Word Sense Disambiguation Based on Comparison of Multiple Classifiers

Copyright 2012 by Kanako Komiya and Manabu Okumura 26th Pacific Asia Conference on Language,Information and Computation pages 80–88 Automatic Domain Adaptation for Word Sense Disambiguation Based on Comparison of Multiple Classifiers Kanako Komiya Tokyo University of Agriculture and Technology 2-24-16 Naka-cho, Koganei Tokyo, 184-8588 Japan kkomiya@cc.tuat.ac.jp Manabu Okumura Tokyo Institute of Technology 4259 Nagatsuta Modori-ku Yokohama 226-8503 Japan oku@pi.titech.ac.jp Abstract Domain adaptation (DA), which involves adapting a classifier developed from source to target data, has been studied intensively in re- cent years. However, when DA for word sense disambiguation (WSD) was carried out, the optimal DA method varied according to the properties of the source and target data. This paper proposes automatic DA based on com- paring the degrees of confidence of multiple classifiers for each instance. We compared three classifiers for three DA methods, where 1) a classifier was trained with a small amount of target data that was randomly selected and manually labeled but without source data, 2) a classifier was trained with source data and a small amount of target data that was ran- domly selected and manually labeled, and 3) a classifier was trained with selected source data that were sufficiently similar to the tar- get data and a small amount of target data that was randomly selected and manually la- beled. We used the method whose degree of confidence was the highest for each instance when Japanese WSD was carried out. The av- erage accuracy of WSD when the DA methods that were determined automatically were used was significantly higher than when the origi- nal methods were used collectively. 1 Introduction Classifiers in standard supervised machine learning have been trained for data in domain A using manu- ally annotated data in domain A, e.g., to train classi- fiers for newswires using newswires. However, clas- sifiers for data in domain B have sometimes been necessary when there have been no or few manu- ally annotated data, and there have only been man- ually annotated data in domain A, which has been related to domain B. Domain adaptation (DA) in- volves adapting the classifier that has been trained from data in domain A (source domain) to data in domain B (target domain). This has been studied in- tensively in recent years. However, the optimal method of DA varied ac- cording to the properties of the data in the source domain (the source data) and the data in the target domain (the target data) when DA for word sense disambiguation (WSD) was carried out. There are many methods of DA for WSD but we assume that the optimal method varies according to each instance. This paper proposes automatic DA based on comparison of the degrees of confidence of multiple classifiers for each instance when Japanese WSD is performed. Our experiments show that the average accuracy of WSD when the DA methods that were determined automatically were used was significantly higher than when the original methods were used collectively. This paper is organized as follows. Section 2 re- views …

Content truncated to 500 words