Automatic Construction of Predicate-argument Structure Patterns for Biomedical Information Extraction

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 284–292, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Automatic Construction of Predicate-argument Structure Patterns for Biomedical Information Extraction Akane Yakushiji∗ † Yusuke Miyao∗ Tomoko Ohta∗ ∗Department of Computer Science, University of Tokyo Hongo 7-3-1, Bunkyo-ku, Tokyo 113-0033 JAPAN § School of Informatics, University of Manchester POBox 88, Sackville St, MANCHESTER M60 1QD, UK {akane, yusuke, okap, yucca, tsujii}@is.s.u-tokyo.ac.jp Yuka Tateisi∗ ‡ Jun’ichi Tsujii∗ § Abstract This paper presents a method of automat- ically constructing information extraction patterns on predicate-argument structures (PASs) obtained by full parsing from a smaller training corpus. Because PASs represent generalized structures for syn- tactical variants, patterns on PASs are ex- pected to be more generalized than those on surface words. In addition, patterns are divided into components to improve recall and we introduce a Support Vector Machine to learn a prediction model using pattern matching results. In this paper, we present experimental results and analyze them on how well protein-protein interac- tions were extracted from MEDLINE ab- stracts. The results demonstrated that our method improved accuracy compared to a machine learning approach using surface word/part-of-speech patterns. 1 Introduction One primitive approach to Information Extrac- tion (IE) is to manually craft numerous extrac- tion patterns for particular applications and this is presently one of the main streams of biomedi- cal IE (Blaschke and Valencia, 2002; Koike et al., 2003). Although such IE attempts have demon- strated near-practical performance, the same sets of patterns cannot be applied to different kinds of information. A real-world task requires several kinds of IE, thus manually engineering extraction Current Affiliation: † FUJITSU LABORATORIES LTD. ‡ Faculty of Informatics, Kogakuin University patterns, which is tedious and time-consuming process, is not really practical. Techniques based on machine learning (Zhou et al., 2005; Hao et al., 2005; Bunescu and Mooney, 2006) are expected to alleviate this problem in manually crafted IE. However, in most cases, the cost of manually crafting patterns is simply trans- ferred to that for constructing a large amount of training data, which requires tedious amount of manual labor to annotate text. To systematically reduce the necessary amount of training data, we divided the task of construct- ing extraction patterns into a subtask that general natural language processing techniques can solve and a subtask that has specific properties accord- ing to the information to be extracted. The former subtask is of full parsing (i.e. recognizing syntactic structures of sentences), and the latter subtask is of constructing specific extraction patterns (i.e. find- ing clue words to extract information) based on the obtained syntactic structures. We adopted full parsing from various levels of parsing, because we believe that it offers the best utility to generalize sentences into normal- ized syntactic relations. We also divided patterns into components to improve recall and we intro- duced machine learning with a Support Vector Machine (SVM) to learn a prediction model us- ing the matching results of extraction patterns. As an actual …

Content truncated to 500 words