Use of Combined Topic Models in Unsupervised Domain Adaptation for Word Sense Disambiguation

Use of Combined Topic Models in Unsupervised Domain Adaptation for Word Sense Disambiguation Shinya Kunii Ibaraki University 4-12-1 Nakanarusawa, Hitachi, Ibaraki 316–8511, Japan 13nm707sl@hcs.ibaraki.ac.jp Hiroyuki Shinnou Ibaraki University 4-12-1 Nakanarusawa, Hitachi, Ibaraki 316–8511, Japan shinnou@mx.ibaraki.ac.jp Abstract Topic models can be used in an unsuper- vised domain adaptation for Word Sense Disambiguation (WSD). In the domain adaptation task, three types of topic mod- els are available: (1) a topic model con- structed from the source domain corpus: (2) a topic model constructed from the tar- get domain corpus, and (3) a topic model constructed from both domains. Basically, three topic features made from each topic model are added to the normal feature used for WSD. By using the extended features, SVM learns and solves WSD. However, the topic features constructed from source do- main have weights describing the similar- ity between the source corpus and the entire corpus because the topic features made from the source domain can reduce the accuracy of WSD. In six transitions of domain adap- tation using three domains, we conducted experiments by varying the combination of topic features, and show the effectiveness of the proposed method. 1 Introduction In this paper, we propose an unsupervised method of domain adaptation for Word Sense Disam- biguation (WSD) using topic models. An inductive learning method is used in many tasks of natural language processing. In inductive learning, training data is created from corpus A, and a classifier learns from the training data. A original task is solved by using the classifier. Dur- ing this analysis, the data for the task is in cor- pus B that differs from the domain of corpus A. In cases, the classifier learned from corpus A (i.e., the source domain) cannot analyze the data of corpus B (i.e., the target domain). This problem is called the domain adaptation problem, which is also re- garded as a component of transfer learning in the field of machine learning. The domain adaptation problem has been extensively researched in recent years. The methods of domain adaptation can be di- vided into two groups from the viewpoint of whether labeled data is to be used in the target do- main. When using labeled data, it is called super- vised learning, while unsupervised learning does not use labeled data. There is substantial research on supervised learning techniques. Conversely, not much attention has been paid to unsupervised learning because of low precision; however, we adopt the unsupervised learning approach because it is does not require labeling. Shinnou and Sasaki examined the unsupervised domain adaptation for WSD (Shinnou and Sasaki, 2013). In their study, the topic model is built from the target domain corpus, and topic features con- structed from the topic model are added to train- ing data in both source and target domains. As a result, the accuracy of the classifier made by train- ing data in the source domain is improved; how- ever, in their study, the topic model is made by only the target domain. As indicated by …

Content truncated to 500 words