Global Relation Embedding for Relation Extraction

Proceedings of NAACL-HLT 2018, pages 820–830 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Global Relation Embedding for Relation Extraction Yu Su∗, Honglei Liu∗, Semih Yavuz, Izzeddin G¨ur University of California, Santa Barbara {ysu,honglei,syavuz,izzeddingur}@cs.ucsb.edu Huan Sun The Ohio State University sun.397@osu.edu Xifeng Yan University of California, Santa Barbara xyan@cs.ucsb.edu Abstract We study the problem of textual relation em- bedding with distant supervision. To combat the wrong labeling problem of distant super- vision, we propose to embed textual relations with global statistics of relations, i.e., the co- occurrence statistics of textual and knowledge base relations collected from the entire cor- pus. This approach turns out to be more ro- bust to the training noise introduced by distant supervision. On a popular relation extraction dataset, we show that the learned textual re- lation embedding can be used to augment ex- isting relation extraction models and signifi- cantly improve their performance. Most re- markably, for the top 1,000 relational facts dis- covered by the best existing model, the preci- sion can be improved from 83.9% to 89.3%. 1 Introduction Relation extraction requires deep understanding of the relation between entities. Early studies mainly use hand-crafted features (Kambhatla, 2004; Zhou et al., 2005), and later kernel methods are intro- duced to automatically generate features (Zelenko et al., 2003; Culotta and Sorensen, 2004; Bunescu and Mooney, 2005; Zhang et al., 2006). Recently neural network models have been introduced to embed words, relations, and sentences into con- tinuous feature space, and have shown a remark- able success in relation extraction (Socher et al., 2012; Zeng et al., 2014; Xu et al., 2015b; Zeng et al., 2015; Lin et al., 2016). In this work, we study the problem of embedding textual relations, defined as the shortest dependency path1 between two entities in the dependency graph of a sentence, to improve relation extraction. Textual relations are one of the most discrimina- tive textual signals that lay the foundation of many ∗ Equally contributed. 1We use fully lexicalized shortest dependency path with directional and typed dependency relations. relation extraction models (Bunescu and Mooney, 2005). A number of recent studies have explored textual relation embedding under the supervised setting (Xu et al., 2015a,b, 2016; Liu et al., 2016), but the reliance on supervised training data lim- its their scalability. In contrast, we embed textual relations with distant supervision (Mintz et al., 2009), which provides much larger-scale training data without the need of manual annotation. How- ever, the assertion of distant supervision, “any sen- tence containing a pair of entities that participate in a knowledge base (KB) relation is likely to ex- press the relation,” can be violated more often than not, resulting in many wrongly labeled training ex- amples. A representative example is shown in Fig- ure 1. Embedding quality is thus compromised by the noise in training data. Our main contribution is a novel way to com- bat the wrong labeling problem of distant supervi- sion. Traditional embedding methods (Xu et al., 2015a,b, 2016; Liu …

Content truncated to 500 words