Adapting a Probabilistic Disambiguation Model of an HPSG Parser to a New Domain

Adapting a Probabilistic Disambiguation Model of an HPSG Parser to a New Domain Tadayoshi Hara1, Yusuke Miyao1, and Jun’ichi Tsujii1,2,3 1 Department of Computer Science, University of Tokyo, Hongo 7-3-1, Bunkyo-ku, Tokyo 113-0033, Japan 2 CREST, JST (Japan Science and Technology Agency), Honcho, 4-1-8, Kawaguchi-shi, Saitama 332-0012, Japan 3 School of Informatics, University of Manchester, POBox 88, Sackville St, Manchester, M60 1QD, UK Abstract. This paper describes a method of adapting a domain-inde- pendent HPSG parser to a biomedical domain. Without modifying the grammar and the probabilistic model of the original HPSG parser, we develop a log-linear model with additional features on a treebank of the biomedical domain. Since the treebank of the target domain is limited, we need to exploit an original disambiguation model that was trained on a larger treebank. Our model incorporates the original model as a reference probabilistic distribution. The experimental results for our model trained with a small amount of a treebank demonstrated an improvement in parsing accuracy. 1 Introduction Natural language processing (NLP) is being demanded in various fields, such as biomedical research, patent application, and WWW, because an unmanage- able amount of information is being published in unstructured data, i.e., natural language texts. To exploit latent information in these, the assistance of NLP technologies is highly required. However, an obstacle is the lack of portability of NLP tools. In general, NLP tools specialized to each domain were developed from scratch, or adapted by considerable human effort. This is because linguistic resources for each domain, such as a treebank, have not been sufficiently devel- oped yet. Since dealing with various kinds of domains is an almost intractable job, sufficient resources can not be expected. The method presented in this paper is the development of disambiguation models of an HPSG parser by combining a disambiguation model of an original parser with a new model adapting to a new domain. Although the training of a disambiguation model of a parser requires a sufficient amount of a treebank, its construction requires a considerable human effort. Hence, we exploit the original disambiguation model that was trained with a larger, but domain-independent treebank. Since the original disambiguation model contains rich information of general grammatical constraints, we try to use its information in developing a disambiguation model for a new domain. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 199–210, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 200 T. Hara, Y. Miyao, and J. Tsujii Our disambiguation model is a log-linear model into which the original disam- biguation model is incorporated as a reference distribution. However, we cannot simply estimate this model, because of the problem that has been discussed in studies of the probabilistic modeling of unification-based grammars [1,2]. That is, the exponential explosion of parse candidates assigned by the grammar makes the estimation intractable. The previous studies solved the problem by applying a dynamic programming algorithm to a packed representation of parse trees. In this paper, we borrow their idea, and define reference distribution on a …

Content truncated to 500 words