Embedding Semantic Similarity in Tree Kernels for Domain Adaptation of Relation Extraction

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1498–1507, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Embedding Semantic Similarity in Tree Kernels for Domain Adaptation of Relation Extraction Barbara Plank∗ Center for Language Technology University of Copenhagen, Denmark bplank@gmail.com Alessandro Moschitti QCRI - Qatar Foundation & DISI - University of Trento, Italy amoschitti@qf.org.qa Abstract Relation Extraction (RE) is the task of extracting semantic relationships between entities in text. Recent studies on rela- tion extraction are mostly supervised. The clear drawback of supervised methods is the need of training data: labeled data is expensive to obtain, and there is often a mismatch between the training data and the data the system will be applied to. This is the problem of domain adapta- tion. In this paper, we propose to combine (i) term generalization approaches such as word clustering and latent semantic anal- ysis (LSA) and (ii) structured kernels to improve the adaptability of relation ex- tractors to new text genres/domains. The empirical evaluation on ACE 2005 do- mains shows that a suitable combination of syntax and lexical generalization is very promising for domain adaptation. 1 Introduction Relation extraction is the task of extracting se- mantic relationships between entities in text, e.g. to detect an employment relationship between the person Larry Page and the company Google in the following text snippet: Google CEO Larry Page holds a press announcement at its headquar- ters in New York on May 21, 2012. Recent stud- ies on relation extraction have shown that super- vised approaches based on either feature or ker- nel methods achieve state-of-the-art accuracy (Ze- lenko et al., 2002; Culotta and Sorensen, 2004; ∗ The first author was affiliated with the Department of Computer Science and Information Engineering of the Uni- versity of Trento (Povo, Italy) during the design of the mod- els, experiments and writing of the paper. Zhang et al., 2005; Zhou et al., 2005; Zhang et al., 2006; Bunescu, 2007; Nguyen et al., 2009; Chan and Roth, 2010; Sun et al., 2011). How- ever, the clear drawback of supervised methods is the need of training data, which can slow down the delivery of commercial applications in new domains: labeled data is expensive to obtain, and there is often a mismatch between the training data and the data the system will be applied to. Ap- proaches that can cope with domain changes are essential. This is the problem of domain adapta- tion (DA) or transfer learning (TL). Technically, domain adaptation addresses the problem of learn- ing when the assumption of independent and iden- tically distributed (i.i.d.) samples is violated. Do- main adaptation has been studied extensively dur- ing the last couple of years for various NLP tasks, e.g. two shared tasks have been organized on do- main adaptation for dependency parsing (Nivre et al., 2007; Petrov and McDonald, 2012). Results were mixed, thus it is still a very active research area. However, to the best of our knowledge, there is almost no work …

Content truncated to 500 words