Unsupervised Classification of Biomedical Abstracts Using Lexical Association

Unsupervised Classification of Biomedical Abstracts using Lexical Association⋆ Jonathon Read, Jonathan Webster, and Alex Chengyu Fang The Dialogue Systems Group Department of Chinese, Translation and Linguistics City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong {jlread,ctjlw,acfang}@cityu.edu.hk Abstract. The task of text classification is the assignment of labels that describe texts’ char- acteristics, such as topic, genre or sentiment. Supervised machine learning techniques such as Support Vector Machines or the simple but effective Na¨ıve Bayes have been successfully ap- plied to this task. However, it is not always practical to acquire a sufficient corpus of labelled examples to train these methods. For these cases we describe an unsupervised method for text classification based on two hypotheses. Firstly, we propose that the class of a document may be determined by calculating its constituent features’ similarity with prototypical examples of each class. Secondly, we note the importance of class priors in Na¨ıve Bayes classifiers, and hypothesize that class distributions might be estimated using the relative frequency of prototype words. Performing experiments on a corpus of biomedical abstracts with topic information derived from the Medical Subject Headings (MeSH), we investigate the charac- teristics of the method when used in conjunction with basic, linguistic and knowledge-based features, and find that the performance of the unsupervised method is approximately 80% that of Na¨ıve Bayes. Our research is significant in that it highlights a candidate method with good potential for further improvement when training on unlabelled data. Keywords: Pointwise mutual information, Text classification, Unsupervised methods 1 Introduction The RAMCORP project aims to design and construct a telephony-based dialogue system that pro- vides interactive dissemination of knowledge in a variety of domains. In particular, it is intended for use by domain experts who are not native English speakers while engaged in dialogue with other domain experts. While the current focus is with respect to the translation of terminology (Webster et al., 2009), the supporting knowledge base is designed to be easily extensible and in- teroperable (Fleissner et al., 2010); future iterations of the project will explore dialogue-based information retrieval and question answering in multiple domains. We aim to develop techniques that will automatically construct grammars for automatic speech recognition (ASR) that guide users through topics in various domains. Initially this will employ information automatically gen- erated from articles and terminologies, but eventually we intend for dialogues to be driven by automatically constructed ontologies. A particular challenge faced by the project is the discrimination of a large number of terms through ASR (Fang et al., 2008); it is desirable to reduce the vocabulary size in order to improve recognition accuracy and promote user satisfaction. One approach to accomplish this is to organise ⋆ The work reported in this paper was supported in part by research grants from the City University of Hong Kong (Project Nos 9610053, 9610126, 7008002, 7002387 and 7002190). We would like to thank the reviewers for their comments. Copyright 2010 by Jonathon Read, Jonathan Webster, and Alex Chengyu Fang PACLIC 24 Proceedings 261 terms …

Content truncated to 500 words