Parsing Dependency Paths to Identify Event-Argument Relations

International Joint Conference on Natural Language Processing, pages 699–705, Nagoya, Japan, 14-18 October 2013. Parsing Dependency Paths to Identify Event-Argument Relations Seung-Cheol Baek CS Dept., KAIST 291 Daehak-ro, Yuseong-gu, Daejeon, 305-701, Republic of Korea scbaek@nlp.kaist.ac.kr Jong C. Park CS Dept., KAIST 291 Daehak-ro, Yuseong-gu, Daejeon, 305-701, Republic of Korea park@nlp.kaist.ac.kr Abstract Mentions of event-argument relations, in par- ticular dependency paths between event- referring words and argument-referring words, can be decomposed into meaningful compo- nents arranged in a regular way, such as those indicating the type of relations and the others allowing relations with distant arguments (e.g., coordinate conjunction). We argue that the knowledge about arrangements of such com- ponents may provide an opportunity for mak- ing event extraction systems more robust to training sets, since unseen patterns would be derived by combining seen components. How- ever, current state-of-the-art machine learning- based approaches to event extraction tasks take the notion of components at a shallow level by using n-grams of paths. In this paper, we propose two methods called pseudo-count and Bayesian methods to semi-automatically learn PCFGs by analyzing paths into compo- nents from the BioNLP shared task training corpus. Each lexical item in the learned PCFGs appears in 2.6 distinct paths on aver- age between event-referring words and argu- ment-referring words, suggesting that they contain recurring components. We also pro- pose a grounded way of encoding multiple parse trees for a single dependency path into feature vectors in linear classification models. We show that our approach can improve the performance of identifying event-argument re- lations in a statistically significant manner. 1 1 Introduction Event extraction tasks can be viewed as identify- ing event-argument relations between tokens by mapping events onto tokens, to be called hence- forth triggers, even though events may have oth- 1 All the datasets and codes used in this study are available at http://www.biopathway.org/ijcnlp2013 er events as arguments in contrast to average re- lation extraction tasks, leading to interdependen- cies between events. On looking into mentions of event-argument relations, in particular the short- est dependency path between triggers and argu- ments, one may find that they can be decom- posed into intuitively meaningful components arranged in a regular way, such as core compo- nents indicating the type of relations and subor- dinate components making it possible for events to take arguments further away from triggers (e.g., coordinate conjunction). We anticipate that the knowledge about arrangements of compo- nents provides an invaluable opportunity for making event extraction systems more robust to the choice of training sets, for example by as- sembling seen components into unseen patterns. Towards this goal, we propose in this paper a way of automatically learning and exploiting in- ternal structures of dependency paths for a robust extraction of biological events from the biologi- cal literature with the corpora provided by a se- ries of BioNLP shared tasks (Kim et al., 2009 and Kim et al., 2011). For example, the following sentence has anno- tated positive regulation events, including the induction of IP-10 by IFN, …

Content truncated to 500 words