Event Role Extraction using Domain-Relevant Word Representations

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1852–1857, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Event Role Extraction using Domain-Relevant Word Representations Emanuela Boros¸†‡ Romaric Besanc¸on† Olivier Ferret† Brigitte Grau‡∗ †CEA, LIST, Vision and Content Engineering Laboratory, F-91191, Gif-sur-Yvette, France ‡LIMSI, rue John von Neumann, Campus Universitaire d’Orsay, F-91405 Orsay cedex ∗ENSIIE, 1 square de la r´esistance F-91025 ´Evry cedex firstname.lastname@cea.fr firstname.lastname@limsi.fr Abstract The efficiency of Information Extraction systems is known to be heavily influenced by domain-specific knowledge but the cost of developing such systems is consider- ably high. In this article, we consider the problem of event extraction and show that learning word representations from unla- beled domain-specific data and using them for representing event roles enable to out- perform previous state-of-the-art event ex- traction models on the MUC-4 data set. 1 Introduction In the Information Extraction (IE) field, event ex- traction constitutes a challenging task. An event is described by a set of participants (i.e. at- tributes or roles) whose values are text excerpts. The event extraction task is related to several sub- tasks: event mention detection, candidate role- filler extraction, relation extraction and event tem- plate filling. The problem we address here is the detection of role-filler candidates and their associ- ation with specific roles in event templates. For this task, IE systems adopt various ways of ex- tracting patterns or generating rules based on the surrounding context, local context and global con- text (Patwardhan and Riloff, 2009). Current ap- proaches for learning such patterns include boot- strapping techniques (Huang and Riloff, 2012a; Yangarber et al., 2000), weakly supervised learn- ing algorithms (Huang and Riloff, 2011; Sudo et al., 2003; Surdeanu et al., 2006), fully supervised learning approaches (Chieu et al., 2003; Freitag, 1998; Bunescu and Mooney, 2004; Patwardhan and Riloff, 2009) and other variations. All these methods rely on substantial amounts of manually annotated corpora and use a large body of lin- guistic knowledge. The performance of these ap- proaches is related to the amount of knowledge engineering deployed and a good choice of fea- tures and classifiers. Furthermore, the efficiency of the system relies on the a priori knowledge of the applicative domain (the nature of the events) and it is generally difficult to apply a system on a different domain with less annotated data with- out reconsidering the design of the features used. An important step forwards is TIERlight (Huang and Riloff, 2012a) that targeted the minimization of human supervision with a bootstrapping tech- nique for event roles detection. Also, PIPER (Pat- wardhan and Riloff, 2007; Patwardhan, 2010) dis- tinguishes between relevant and irrelevant regions and learns domain-relevant extraction patterns us- ing a semantic affinity measure. Another possi- ble approach for dealing with this problem is to combine the use a restricted set of manually anno- tated data with a much larger set of data extracted in an unsupervised way from a corpus. This ap- proach was experimented for relations in the con- text …

Content truncated to 500 words