How Well Conditional Random Fields Can be Used in Novel Term Recognition

How Well Conditional Random Fields Can be Used in Novel Term Recognition  Xing ZHANG a, Yan Song b, and Alex Chengyu Fang b a Department of Chinese, Translation and Linguistics, City University of Hong Kong Kowloon, Hong Kong SAR zxing2@student.cityu.edu.hk b Department of Chinese, Translation and Linguistics, City University of Hong Kong Kowloon, Hong Kong SAR {yansong, acfang} @cityu.edu.hk Abstract. In this paper, we describe the construction of a machine learning framework that exploit syntactic information in the recognition of biomedical terms and present the limits of machine learning in generating a novel term candidate list. Conditional random fields (CRF), is used as the basis of this framework. We make an effort to find the appropriate use of syntactic information, including parent nodes, syntactic paths and term ratios under this machine learning framework. The experiment results show that CRF model can achieve good precision in term recognition if trained with known term list. However, with regard to discovering potential novel terms for terminology lexicon editors, CRF model fails to show good performance, if trained with known term list only to predict novel terms in testing corpus. Therefore, this result suggests that more semantic information may be needed to determine a word to be a novel term during a specific period. Keywords: term recognition, novel term recognition, conditional random fields 1 Introduction This paper explores the use of Conditional Random Fields (CRF) in novel term recognition. It investigates the recognition of medical terms using CRF model. A variety of methods have been used in term recognition, some are linguistics focused, some are statistically motivated, and a large part of approaches combine these two together. Recently, with the development of machine learning models, a lot of work has attempted to extract terms using machine learning methods. The usual practice of machine learning depends on training data and a set of discriminating features. Then, machine learning systems use training data to “learn” features useful for term recognition. The widely used standard features for machine learning include orthographic features, POS tags, prefix, and suffix information (Krauthammer and Nenadic, 2004). However, few studies have tried to make use of dynamic linguistic features in respect of term usage in real text. As Zhang and Fang (2009) found out, syntactic functions can be used effectively in selecting and ranking term candidates, which means termhood can be captured by computing term ratios in syntactic paths.  The work reported in this paper was supported in part by research grants from the City University of Hong Kong (Project No. 9610126, 7008002, 7002387 and 7002190). Copyright 2010 by Jianguo Chen, John Smith, and Kinko Yamada PACLIC 24 Proceedings 583 Furthermore, as studies concerning novel terms are not so common, this paper considers how syntactic information integrated under a machine learning framework can be helpful in discovering novel terms. As early as in 1995, Justeson and Katz defined novel terminology as terms that are newly introduced and not yet widely established, or terms that are current only in more advanced …

Content truncated to 500 words