A Cross-lingual Annotation Projection Approach for Relation Detection

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 564–571, Beijing, August 2010 A Cross-lingual Annotation Projection Approach for Relation Detection Seokhwan Kim†, Minwoo Jeong‡, Jonghoon Lee†, Gary Geunbae Lee† †Department of Computer Science and Engineering, Pohang University of Science and Technology {megaup|jh21983|gblee}@postech.ac.kr ‡Saarland University m.jeong@mmci.uni-saarland.de Abstract While extensive studies on relation ex- traction have been conducted in the last decade, statistical systems based on su- pervised learning are still limited because they require large amounts of training data to achieve high performance. In this pa- per, we develop a cross-lingual annota- tion projection method that leverages par- allel corpora to bootstrap a relation detec- tor without significant annotation efforts for a resource-poor language. In order to make our method more reliable, we intro- duce three simple projection noise reduc- tion methods. The merit of our method is demonstrated through a novel Korean re- lation detection task. 1 Introduction Relation extraction aims to identify semantic re- lations of entities in a document. Many rela- tion extraction studies have followed the Rela- tion Detection and Characterization (RDC) task organized by the Automatic Content Extraction project (Doddington et al., 2004) to make multi- lingual corpora of English, Chinese and Ara- bic. Although these datasets encourage the de- velopment and evaluation of statistical relation extractors for such languages, there would be a scarcity of labeled training samples when learn- ing a new system for another language such as Korean. Since manual annotation of entities and their relations for such resource-poor languages is very expensive, we would like to consider in- stead a weakly-supervised learning technique in order to learn the relation extractor without sig- nificant annotation efforts. To do this, we propose to leverage parallel corpora to project the relation annotation on the source language (e.g. English) to the target (e.g. Korean). While many supervised machine learning ap- proaches have been successfully applied to the RDC task (Kambhatla, 2004; Zhou et al., 2005; Zelenko et al., 2003; Culotta and Sorensen, 2004; Bunescu and Mooney, 2005; Zhang et al., 2006), few have focused on weakly-supervised relation extraction. For example, (Zhang, 2004) and (Chen et al., 2006) utilized weakly-supervised learning techniques for relation extraction, but they did not consider weak supervision in the context of cross-lingual relation extraction. Our key hypoth- esis on the use of parallel corpora for learning the relation extraction system is referred to as cross-lingual annotation projection. Early stud- ies of cross-lingual annotation projection were ac- complished for lexically-based tasks; for exam- ple part-of-speech tagging (Yarowsky and Ngai, 2001), named-entity tagging (Yarowsky et al., 2001), and verb classification (Merlo et al., 2002). Recently, there has been increasing interest in ap- plications of annotation projection such as depen- dency parsing (Hwa et al., 2005), mention de- tection (Zitouni and Florian, 2008), and semantic role labeling (Pado and Lapata, 2009). However, to the best of our knowledge, no work has reported on the RDC task. In this paper, we apply a cross-lingual anno- tation projection approach to binary relation de- …

Content truncated to 500 words