Semantic Representations for Domain Adaptation: A Case Study on the Tree Kernel-based Method for Relation Extraction

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 635–644, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Semantic Representations for Domain Adaptation: A Case Study on the Tree Kernel-based Method for Relation Extraction Thien Huu Nguyen†, Barbara Plank§ and Ralph Grishman† † Computer Science Department, New York University, New York, NY 10003, USA § Center for Language Technology, University of Copenhagen, Denmark thien@cs.nyu.edu,bplank@cst.dk,grishman@cs.nyu.edu Abstract We study the application of word embed- dings to generate semantic representations for the domain adaptation problem of re- lation extraction (RE) in the tree kernel- based method. We systematically evaluate various techniques to generate the seman- tic representations and demonstrate that they are effective to improve the general- ization performance of a tree kernel-based relation extractor across domains (up to 7% relative improvement). In addition, we compare the tree kernel-based and the feature-based method for RE in a compat- ible way, on the same resources and set- tings, to gain insights into which kind of system is more robust to domain changes. Our results and error analysis shows that the tree kernel-based method outperforms the feature-based approach. 1 Introduction Relation Extraction (RE) is an important aspect of information extraction that aims to discover the semantic relationships between two entity men- tions appearing in the same sentence. Previous research on RE has followed either the kernel- based approach (Zelenko et al., 2003; Bunescu and Mooney, 2005; Zhao and Grishman, 2005; Zhang et al., 2006; Bunescu, 2007; Qian et al., 2008; Nguyen et al., 2009) or the feature-based ap- proach (Kambhatla, 2004; Grishman et al., 2005; Zhou et al., 2005; Jiang and Zhai, 2007a; Chan and Roth, 2010; Sun et al., 2011). Usually, in such supervised machine learning systems, it is as- sumed that the training data and the data to which the RE system is applied to are sampled inde- pendently and identically from the same distribu- tion. This assumption is often violated in reality and exemplified in the fact that the performance of the traditional RE techniques degrades signif- icantly in such a domain mismatch case (Plank and Moschitti, 2013). To alleviate this perfor- mance loss, we need to resort to domain adaptation (DA) techniques to adapt a system trained on some source domain to perform well on new target do- mains. We here focus on the unsupervised domain adaptation (i.e., no labeled target data) and single- system DA (Petrov and McDonald, 2012; Plank and Moschitti, 2013), i.e., building a single sys- tem that is able to cope with different, yet related target domains. While DA has been investigated extensively in the last decade for various natural language pro- cessing (NLP) tasks, the examination of DA for RE is only very recent. To the best of our knowl- edge, there have been only three studies on DA for RE (Plank and Moschitti, 2013; Nguyen and Grishman, 2014; Nguyen et al., 2014). Of these, Nguyen et al. (2014) follow …

Content truncated to 500 words