Domain Adaptation for Coreference Resolution: An Adaptive Ensemble Approach

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 744–753, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Domain Adaptation for Coreference Resolution: An Adaptive Ensemble Approach Jian Bo Yang†‡∗, Qi Mao†, Qiao Liang Xiang†, Ivor W. Tsang†, Kian Ming A. Chai§, Hai Leong Chieu§ † School of Computer Engineering, Nanyang Technological University, Singapore ‡ Electrical and Computer Engineering Department, Duke University, USA § DSO National Laboratories, Singapore jianbo.yang@duke.edu, {qmao1,qlxiang,ivortsang}@ntu.edu.sg, {ckianmin,chaileon}@dso.org.sg Abstract We propose an adaptive ensemble method to adapt coreference resolution across domains. This method has three features: (1) it can op- timize for any user-specified objective mea- sure; (2) it can make document-specific pre- diction rather than rely on a fixed base model or a fixed set of base models; (3) it can auto- matically adjust the active ensemble members during prediction. With simplification, this method can be used in the traditional within- domain case, while still retaining the above features. To the best of our knowledge, this work is the first to both (i) develop a domain adaptation algorithm for the coreference reso- lution problem and (ii) have the above features as an ensemble method. Empirically, we show the benefits of (i) on the six domains of the ACE 2005 data set in domain adaptation set- ting, and of (ii) on both the MUC-6 and the ACE 2005 data sets in within-domain setting. 1 Introduction Coreference resolution is a fundamental component of natural language processing (NLP) and has been widely applied in other NLP tasks (Stoyanov et al., 2010). It gathers together noun phrases (mentions) that refer to the same real-world entity (Ng and Cardie, 2002). In the past decade, several corefer- ence resolution systems have been proposed, e.g., (Ng and Cardie, 2002), (Denis and Baldridge, 2007) and (Stoyanov et al., 2010). All of these focus on the within-domain case — to use the labeled doc- uments from a domain to predict on the unlabeled ∗The work is done during postdoc in NTU, Singapore. documents in the same domain. However, in prac- tice, there is usually limited labeled data in a specific domain of interest, while there may be plenty of la- beled data in other related domains. Effective use of data from the other domains for predicting in the do- main of interest is therefore an important strategy in NLP. This is called domain adaptation, and, in this context, the former domains is called the source do- mains, while the latter domain is called the target domain (Blitzer et al., 2006; Jiang and Zhai, 2007). Based on the type of the knowledge to be trans- ferred to the target domain, domain adaptation learn- ing can be categorized as instance-based method, feature-based method, parameter-based method or relational-knowledge-based method (Pan and Yang, 2010). Previously, domain adaptation learning has been successfully used in other NLP tasks such as relation extraction (Jiang, 2009) and POS tagging (Jiang and Zhai, 2007), semantic detection (Tan et al., 2008), name …

Content truncated to 500 words