A Discriminative Alignment Model for Abbreviation Recognition

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 657–664 Manchester, August 2008 A Discriminative Alignment Model for Abbreviation Recognition Naoaki Okazaki† okazaki@is.s.u-tokyo.ac.jp Sophia Ananiadou‡ sophia.ananiadou@manchester.ac.uk Jun’ichi Tsujii†‡ tsujii@is.s.u-tokyo.ac.jp †Graduate School of Information Science and Technology University of Tokyo 7-3-1 Hongo, Bunkyo-ku Tokyo 113-8656, Japan ‡School of Computer Science, University of Manchester National Centre for Text Mining (NaCTeM) Manchester Interdisciplinary Biocentre 131 Princess Street, Manchester M1 7DN, UK Abstract This paper presents a discriminative align- ment model for extracting abbreviations and their full forms appearing in actual text. The task of abbreviation recognition is formalized as a sequential alignment problem, which finds the optimal align- ment (origins of abbreviation letters) be- tween two strings (abbreviation and full form). We design a large amount of fine- grained features that directly express the events where letters produce or do not pro- duce abbreviations. We obtain the optimal combination of features on an aligned ab- breviation corpus by using the maximum entropy framework. The experimental re- sults show the usefulness of the alignment model and corpus for improving abbrevia- tion recognition. 1 Introduction Abbreviations present two major challenges in nat- ural language processing: term variation and am- biguity. Abbreviations substitute for expanded terms (e.g., dynamic programming) through the use of shortened term-forms (e.g., DP). At the same time, the abbreviation DP appearing alone in text is ambiguous, in that it may refer to different concepts, e.g., data processing, dirichlet process, differential probability. Associating abbreviations and their full forms is useful for various applica- tions including named entity recognition, informa- tion retrieval, and question answering. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. The task of abbreviation recognition, in which abbreviations and their expanded forms appearing in actual text are extracted, addresses the term vari- ation problem caused by the increase in the num- ber of abbreviations (Chang and Sch¨utze, 2006). Furthermore, abbreviation recognition is also cru- cial for disambiguating abbreviations (Pakhomov, 2002; Gaudan et al., 2005; Yu et al., 2006), pro- viding sense inventories (lists of abbreviation def- initions), training corpora (context information of full forms), and local definitions of abbreviations. Hence, abbreviation recognition plays a key role in abbreviation management. Numerous researchers have proposed a variety of heuristics for recognizing abbreviation defini- tions, e.g., the use of initials, capitalizations, syl- lable boundaries, stop words, lengths of abbrevia- tions, and co-occurrence statistics (Park and Byrd, 2001; Wren and Garner, 2002; Liu and Fried- man, 2003; Okazaki and Ananiadou, 2006; Zhou et al., 2006; Jain et al., 2007). Schwartz and Hearst (2003) implemented a simple algorithm that finds the shortest expression containing all alpha- numerical letters of an abbreviation. Adar (2004) presented four scoring rules to choose the most likely expanded form in multiple candidates. Ao and Takagi (2005) designed more detailed condi- tions for accepting or discarding candidates of ab- breviation definitions. However, these studies have limitations in dis- covering an optimal combination of heuristic rules from manual observations of a …

Content truncated to 500 words