A Rich Feature Vector for Protein-Protein Interaction Extraction from Multiple Corpora

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 121–130, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP A Rich Feature Vector for Protein-Protein Interaction Extraction from Multiple Corpora Makoto Miwa1 Rune Sætre1 Yusuke Miyao1 Jun’ichi Tsujii1,2,3 1Department of Computer Science, the University of Tokyo, Japan Hongo 7-3-1, Bunkyo-ku, Tokyo, Japan. 2School of Computer Science, University of Manchester, UK 3National Center for Text Mining, UK {mmiwa,rune.saetre,yusuke,tsujii}@is.s.u-tokyo.ac.jp Abstract Because of the importance of protein- protein interaction (PPI) extraction from text, many corpora have been proposed with slightly differing definitions of pro- teins and PPI. Since no single corpus is large enough to saturate a machine learn- ing system, it is necessary to learn from multiple different corpora. In this paper, we propose a solution to this challenge. We designed a rich feature vector, and we applied a support vector machine modi- fied for corpus weighting (SVM-CW) to complete the task of multiple corpora PPI extraction. The rich feature vector, made from multiple useful kernels, is used to express the important information for PPI extraction, and the system with our fea- ture vector was shown to be both faster and more accurate than the original kernel- based system, even when using just a sin- gle corpus. SVM-CW learns from one cor- pus, while using other corpora for support. SVM-CW is simple, but it is more effec- tive than other methods that have been suc- cessfully applied to other NLP tasks ear- lier. With the feature vector and SVM- CW, our system achieved the best perfor- mance among all state-of-the-art PPI ex- traction systems reported so far. 1 Introduction The performance of an information extraction pro- gram is highly dependent on various factors, in- cluding text types (abstracts, complete articles, re- ports, etc.), exact definitions of the information to be extracted, shared sub-topics of the text collec- tions from which information is to be extracted. Even if two corpora are annotated in terms of the same type of information by two groups, the per- formance of a program trained by one corpus is unlikely to be reproduced in the other corpus. On the other hand, from a practical point of view, it is worth while to effectively use multiple existing an- notated corpora together, because it is very costly to make new annotations. One problem with several different corpora is protein-protein interaction (PPI) extraction from text. While PPIs play a critical role in un- derstanding the working of cells in diverse bio- logical contexts, the manual construction of PPI databases such as BIND, DIP, HPRD, IntAct, and MINT (Mathivanan et al., 2006) is known to be very time-consuming and labor-intensive. The au- tomatic extraction of PPI from published papers has therefore been a major research topic in Natu- ral Language Processing for Biology (BioNLP). Among several PPI extraction task settings, the most common is sentence-based, pair-wise PPI ex- traction. At least four annotated corpora have been provided for this setting: AIMed (Bunescu et al., 2005), HPRD50 (Fundel et al., …

Content truncated to 500 words