Joint Distant and Direct Supervision for Relation Extraction

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 732–740, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Joint Distant and Direct Supervision for Relation Extraction Truc-Vien T. Nguyen and Alessandro Moschitti Department of Information Engineering and Computer Science University of Trento 38123 Povo (TN), Italy {nguyenthi,moschitti}@disi.unitn.it Abstract Supervised approaches to Relation Extrac- tion (RE) are characterized by higher ac- curacy than unsupervised models. Unfor- tunately, their applicability is limited by the need of training data for each rela- tion type. Automatic creation of such data using Distant Supervision (DS) provides a promising solution to the problem. In this paper, we study DS for designing end- to-end systems of sentence-level RE. In particular, we propose a joint model be- tween Web data derived with DS and man- ually annotated data from ACE. The re- sults show (i) an improvement on the pre- vious state-of-the-art in ACE, which pro- vides important evidence of the benefit of DS; and (ii) a rather good accuracy on ex- tracting 52 types of relations from Web data, which suggests the applicability of DS for general RE. 1 Introduction Automatic Relation Extraction (RE) as defined in ACE (Doddington et al., 2004) achieves the high- est accuracy when supervised approaches are ap- plied, e.g., (Zelenko et al., 2002). Unfortunately, they require labeled data and tend to be domain- dependent as different domains involve different relations. Distant supervision (DS), e.g., using Wikipedia (Banko et al., 2007; Mintz et al., 2009; Hoffmann et al., 2010), can be applied for auto- matically acquiring relation types and their train- ing data. The main idea behind DS is to exploit (i) rela- tion repositories, e.g., the Infobox, x, of Wikipedia to define a set of relation types RT(x) and (ii) the text of the page associated with x to produce the training sentences, which are supposed to express instances of RT(x). Previous work has applied DS to RE at corpus level, e.g., (Banko et al., 2007; Mintz et al., 2009): relation extractors are (i) learned using such not completely accurate data and (ii) applied to extract relation instances from the whole corpus. The multiple pieces of evidence for each relation in- stance are then exploited to recover from errors of the automatic extractors. Additionally, a recent approach, i.e., (Hoffmann et al., 2010), has shown that DS can be also applied at level of Wikipedia article: given a target Infobox template, all its at- tributes1 can be extracted from a given document matching such template. Sentence-level RE (SLRE) has been typically modeled with the traditional supervised approach, e.g., using the data manually annotated in ACE (Culotta and Sorensen, 2004; Kambhatla, 2004; Bunescu and Mooney, 2005; Zhang et al., 2005; Zhang et al., 2006; Bunescu and Mooney, 2007; Nguyen et al., 2009). The resulting extractors are very valuable as they find rare relation instances that might be expressed in only one document. For example, the relation President(Barrack Obama, United States) can be extracted from thousands of documents thus there …

Content truncated to 500 words