Exploring Domain Differences for the Design of a Pronoun Resolution System for Biomedical Text

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 625–632 Manchester, August 2008 Exploring Domain Differences for the Design of Pronoun Resolution Systems for Biomedical Text Ngan L.T. Nguyen Jin-Dong Kim Department of Computer Science, University of Tokyo, Hongo 7-3-1, Tokyo, Japan {nltngan, jdkim}@is.s.u-tokyo.ac.jp Abstract Much effort in the research community has been spent on solving the anaphora resolu- tion or pronoun resolution problem, and in particular for news texts. In order to selec- tively inherit the previous works and solve the same problem for a new domain, we carried out a comparative study with three different corpora: MUC, ACE for the news texts, and GENIA for bio-medical papers. Our corpus analysis and experimental re- sults show the significant differences in the use of pronouns in the two domains, thus by properly considering the characteristics of a domain, we can improve the perfor- mance of pronoun resolution for that do- main. 1 Introduction Pronoun resolution is the task of determining the antecedent of an anaphoric pronoun, or a pro- noun pointing back to some previously mentioned item in a text. For example, in the sentence, “The IL-2 gene displays both T cell specific and in- ducible expression: it is only expressed in CD4+ T cells after antigenic or mitogenic stimulation,” the pronoun “it” should be resolved to refer to “the IL-2 gene,” and thus, we have an anaphora link. Pronoun resolution is an important task in the family of reference resolution tasks, including anaphora resolution and co-reference resolution, which are known as significant parts of text un- derstanding systems. Recently the need to have more powerful information extraction systems for 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. biomedical technical papers has motivated re- searchers to solve the same task for the biomed- ical domain. Castano (Castano et al., 2002) re- solved the sortal and pronominal anaphora, by us- ing a salience measure, which is the sum of all fea- ture scores. Kim and Park (Kim and C.Park, 2004) introduced BioAR, a biomedical anaphora resolu- tion system that relates entity mentions in text with their corresponding Swiss-Prot entries. This sys- tem resolves anaphoric pronouns by using heuris- tic rules and seven patterns for parallelism. How- ever, the sizes of the data sets used in their exper- iments were small. In the former system, 46 and 54 MEDLINE abstracts were used for the devel- opment set, and the test set respectively, and the test set in the latter work contained only sixteen anaphoric pronouns. Contrary to their work, in this work we made use of GENIA, a large co-reference annotated corpus for the bio domain, containing 1999 MEDLINE abstracts. While there are quite a few works on this task for the bio-medical domain, for other domains, and especially for the news domain, a myriad of works on pronoun resolution has been carried out by the NLP researchers (Mitkov, 2002). Since Soon (Soon et al., 2001) started the trend of …

Content truncated to 500 words