The Choice of Features for Classification of Verbs in Biomedical Texts

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 449–456 Manchester, August 2008 The Choice of Features for Classification of Verbs in Biomedical Texts Anna Korhonen University of Cambridge Computer Laboratory 15 JJ Thomson Avenue Cambridge CB3 0FD, UK alk23@cl.cam.ac.uk Yuval Krymolowski Dept. of Computer Science Haifa University Israel yuvalkry@gmail.com Nigel Collier National Institute of Informatics Hitotsubashi 2-1-2 Chiyoda-ku, Tokyo 101-8430 Japan collier@nii.ac.jp Abstract We conduct large-scale experiments to in- vestigate optimal features for classification of verbs in biomedical texts. We intro- duce a range of feature sets and associated extraction techniques, and evaluate them thoroughly using a robust method new to the task: cost-based framework for pair- wise clustering. Our best results compare favourably with earlier ones. Interestingly, they are obtained with sophisticated fea- ture sets which include lexical and seman- tic information about selectional prefer- ences of verbs. The latter are acquired au- tomatically from corpus data using a fully unsupervised method. 1 Introduction Recent years have seen a massive growth in the scientific literature in the domain of biomedicine. Because future research in the biomedical sciences depends on making use of all this existing knowl- edge, there is a strong need for the development of natural language processing (NLP) tools which can be used to automatically locate, organize and man- age facts related to published experimental results. Major progress has been made on information retrieval and on the extraction of specific rela- tions (e.g. between proteins and cell types) from biomedical texts (Ananiadou et al., 2006). Other tasks, such as the extraction of factual information, remain a bigger challenge. Researchers have recently begun to use deeper NLP techniques (e.g. statistical parsing) for im- 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. PROTEINS: p53� p53� Tp53� Dmp53� ...� ACTIVATE� suggests� demonstrates� indicates� implies...� GENES: WAF1� WAF1� CIP1� p21� ...� It� INDICATE� that� activates� up-regulates� induces� stimulates...� ...� Figure 1: Sample lexical classes proved processing of the challenging linguistic structures (e.g. complex nominals, modal subordi- nation, anaphoric links) in biomedical texts. For optimal performance, many of these techniques require richer syntactic and semantic informa- tion than is provided by existing domain lexicons (e.g. UMLS metathesaurus and lexicon1). This par- ticularly applies to verbs, which are central to the structure and meaning of sentences. Where the information is absent, lexical classi- fication can compensate for it, or aid in obtaining it. Lexical classes which capture the close rela- tion between the syntax and semantics of verbs provide generalizations about a range of linguis- tic properties (Levin, 1993). For example, con- sider the INDICATE and ACTIVATE verb classes in Figure 1. Their members have similar subcatego- rization frames SCFs (e.g. activate / up-regulate / induce / stimulate NP) and selectional preferences (e.g. activate / up-regulate / induce / stimulate GENES:WAF1), and they can be used to make sim- ilar statements describing similar events (e.g. PRO- TEINS:P53 ACTIVATE GENES:WAF1). Lexical classes can be used to abstract away from individual words, …

Content truncated to 500 words