Relation Extraction with Relation Topics

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1426–1436, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Relation Extraction with Relation Topics Chang Wang James Fan Aditya Kalyanpur David Gondek IBM T. J. Watson Research Lab 19 Skyline Drive, Hawthorne, New York 10532 {wangchan, fanj, adityakal, dgondek}@us.ibm.com Abstract This paper describes a novel approach to the semantic relation detection problem. Instead of relying only on the training instances for a new relation, we leverage the knowledge learned from previously trained relation detec- tors. Specifically, we detect a new semantic relation by projecting the new relation’s train- ing instances onto a lower dimension topic space constructed from existing relation de- tectors through a three step process. First, we construct a large relation repository of more than 7,000 relations from Wikipedia. Second, we construct a set of non-redundant relation topics defined at multiple scales from the re- lation repository to characterize the existing relations. Similar to the topics defined over words, each relation topic is an interpretable multinomial distribution over the existing re- lations. Third, we integrate the relation topics in a kernel function, and use it together with SVM to construct detectors for new relations. The experimental results on Wikipedia and ACE data have confirmed that background- knowledge-based topics generated from the Wikipedia relation repository can significantly improve the performance over the state-of-the- art relation detection approaches. 1 Introduction Detecting semantic relations in text is very useful in both information retrieval and question answer- ing because it enables knowledge bases to be lever- aged to score passages and retrieve candidate an- swers. To extract semantic relations from text, three types of approaches have been applied. Rule-based methods (Miller et al., 2000) employ a number of linguistic rules to capture relation patterns. Feature- based methods (Kambhatla, 2004; Zhao and Grish- man, 2005) transform relation instances into a large amount of linguistic features like lexical, syntactic and semantic features, and capture the similarity be- tween these feature vectors. Recent results mainly rely on kernel-based approaches. Many of them fo- cus on using tree kernels to learn parse tree struc- ture related features (Collins and Duffy, 2001; Cu- lotta and Sorensen, 2004; Bunescu and Mooney, 2005). Other researchers study how different ap- proaches can be combined to improve the extraction performance. For example, by combining tree ker- nels and convolution string kernels, (Zhang et al., 2006) achieved the state of the art performance on ACE (ACE, 2004), which is a benchmark dataset for relation extraction. Although a large set of relations have been iden- tified, adapting the knowledge extracted from these relations for new semantic relations is still a chal- lenging task. Most of the work on domain adapta- tion of relation detection has focused on how to cre- ate detectors from ground up with as little training data as possible through techniques such as boot- strapping (Etzioni et al., 2005). We take a differ- ent approach, focusing on how the knowledge ex- tracted from the …

Content truncated to 500 words