Extracting Relations Within and Across Sentences

Proceedings of Recent Advances in Natural Language Processing, pages 25–32, Hissar, Bulgaria, 12-14 September 2011. Extracting Relations Within and Across Sentences Kumutha Swampillai and Mark Stevenson Department of Computer Science Sheffield University S1 4DP, UK (k.swampillai|m.stevenson)@dcs.shef.ac.uk Abstract Previous work on relation extraction has focussed on identifying relationships be- tween entities that occur in the same sentence (intra-sentential relations) rather than between entities in different sen- tences (inter-sentential relations) despite previous research having shown that inter- sentential relations commonly occur in in- formation extraction corpora. This paper describes a SVM-based approach to re- lation extraction that is applied to both types. Adapted features and techniques for counter-acting bias in SVM models are used to deal with specific issues that arise in the inter-sentential case. It was found that the structured features used for intra- sentential relation extraction can be eas- ily adapted for the inter-sentential case and provides comparable performance. 1 Introduction Relation extraction is an established subfield of information extraction concerned with extracting related pairs of entities from text. The majority of research has been applied to extracting rela- tions within single sentences (intra-sentential re- lations), examples include (Chieu and Ng, 2002; Culotta and Sorensen, 2004; Sekine, 2006; Banko and Etzioni, 2008). However, an analysis of the MUC6 corpus (Swampillai and Stevenson, 2010) showed that 28.5% of the relations occur between entities in different sentences (inter-sentential re- lations). This paper describes a SVM-based ap- proach which is applied to the extraction of both inter- and intra-sentential relations. A number of challenges are faced when extract- ing inter-sentential relations. The structured fea- tures, that are based on parse trees and have been successfully used for intra-sentential relation ex- traction, do not naturally apply over multiple sen- tences. The limited research published on inter- sentential relation extraction (Roberts et al., 2008; Hirano et al., 2010) does not employ parse tree features. We address this problem by introducing new structured features (see section 3.2) for the inter-sentential case. There is also a greater data sparsity issue when learning extraction models for inter-sentential re- lations due in part to the smaller number of rela- tions expressed inter-sententially. We investigate a learning approach called threshold adjustment (Shanahan and Roma, 2003) to counter-act the im- balance in the data. The remainder of the paper is organised as fol- lows: Section 2 discusses previous work on re- lation extraction. Section 3 describes a relation extraction system suitable for both inter- and intra- sentential relation extraction that uses both flat and structures features. The MUC6 relation extrac- tion task is described in Section 4. Section 5 in- vestigates whether the bias in the relation extrac- tion SVM models can be mitigated using threshold adaption. Section 6 reports the results of the inter- sentential and intra-sentential relation extraction system described. Finally, Section 7 concludes the paper with a discussion of the effectiveness of a composite kernel approach to inter-sentential rela- tion extraction. 2 Related Work The majority of the work on relation extraction has focused on intra-sentential relations and …

Content truncated to 500 words