Exploiting Background Knowledge for Relation Extraction

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 152–160, Beijing, August 2010 Exploiting Background Knowledge for Relation Extraction Yee Seng Chan and Dan Roth University of Illinois at Urbana-Champaign {chanys,danr}@illinois.edu Abstract Relation extraction is the task of recog- nizing semantic relations among entities. Given a particular sentence supervised ap- proaches to Relation Extraction employed feature or kernel functions which usu- ally have a single sentence in their scope. The overall aim of this paper is to pro- pose methods for using knowledge and re- sources that are external to the target sen- tence, as a way to improve relation ex- traction. We demonstrate this by exploit- ing background knowledge such as rela- tionships among the target relations, as well as by considering how target rela- tions relate to some existing knowledge resources. Our methods are general and we suggest that some of them could be ap- plied to other NLP tasks. 1 Introduction Relation extraction (RE) is the task of detecting and characterizing semantic relations expressed between entities in text. For instance, given the sentence “Cone, a Kansas City native, was origi- nally signed by the Royals and broke into the ma- jors with the team.”, one of the relations we might want to extract is the employment relation between the pair of entity mentions “Cone” and “Royals”. RE is important for many NLP applications such as building an ontology of entities, biomedical in- formation extraction, and question answering. Prior work have employed diverse approaches towards resolving the task. One approach is to build supervised RE systems using sentences an- notated with entity mentions and predefined target relations. When given a new sentence, the RE sys- tem has to detect and disambiguate the presence of any predefined relations that might exist between each of the mention pairs in the sentence. In build- ing these systems, researchers used a wide variety of features (Kambhatla, 2004; Zhou et al., 2005; Jiang and Zhai, 2007). Some of the common fea- tures used to analyze the target sentence include the words appearing in the sentence, their part-of- speech (POS) tags, the syntactic parse of the sen- tence, and the dependency path between the pair of mentions. In a related line of work, researchers have also proposed various kernel functions based on different structured representations (e.g. de- pendency or syntactic tree parses) of the target sentences (Bunescu and Mooney, 2005; Zhou et al., 2007; Zelenko et al., 2003; Zhang et al., 2006). Additionally, researchers have tried to au- tomatically extract examples for supervised learn- ing from resources such as Wikipedia (Weld et al., 2008) and databases (Mintz et al., 2009), or at- tempted open information extraction (IE) (Banko et al., 2007) to extract all possible relations. In this work, we focus on supervised RE. In prior work, the feature and kernel functions em- ployed are usually restricted to being defined on the various representations (e.g. lexical or struc- tural) of the target sentences. However, in recog- nizing relations, humans are not …

Content truncated to 500 words