Relation Discovery with Out-of-Relation Knowledge Base as Supervision

Proceedings of NAACL-HLT 2019, pages 3280–3290 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3280 Relation Discovery with Out-of-Relation Knowledge Base as Supervision Yan Liang1, Xin Liu1, Jianwen Zhang2, and Yangqiu Song1 1Department of CSE, Hong Kong University of Science and Technology, HK 2Microsoft, USA 1{yliangav, xliucr, yqsong}@cse.ust.hk 2{jiazhan}@microsoft.com Abstract Unsupervised relation discovery aims to dis- cover new relations from a given text cor- pus without annotated data. However, it does not consider existing human annotated knowl- edge bases even when they are relevant to the relations to be discovered. In this pa- per, we study the problem of how to use out- of-relation knowledge bases to supervise the discovery of unseen relations, where out-of- relation means that relations to discover from the text corpus and those in knowledge bases are not overlapped. We construct a set of constraints between entity pairs based on the knowledge base embedding and then incor- porate constraints into the relation discovery by a variational auto-encoder based algorithm. Experiments show that our new approach can improve the state-of-the-art relation discovery performance by a large margin. 1 Introduction Relation extraction has been widely used for many applications, such as knowledge graph construc- tion (Dong et al., 2014), information retrieval (Liu et al., 2014), and question answering (Ravichan- dran and Hovy, 2002). Traditional supervised ap- proaches require direct annotation on sentences with a relatively small number of relations (Roth and Yih, 2002; Kambhatla, 2004).1 With the de- velopment of large-scale knowledge bases (KBs) such as Freebase (Bollacker et al., 2008), rela- tion extraction has been extended to larger scales comparable to KBs using the distant supervi- sion (Mintz et al., 2009). However, when the train- ing corpus does not support the annotated relations showing in the KB, such approach could fail to find sufficient training examples. Distant super- vision assumption can be violated by up to 31% 1We distinguish a relation (e.g., a predicate in a knowl- edge base) from the relation expression (e.g., the text surface between entities in a sentence) throughout the paper. for some relations when aligning to NYT corpus (Riedel et al., 2010). More importantly, either traditional supervised learning or distantly super- vised learning cannot discover new relations un- seen in the training phase. Unsupervised relation discovery tries to over- come the shortcomings of supervised or distantly supervised learning approaches. Existing ap- proaches either extract surface or syntactic pat- terns from sentences and use relation expressions as predicates (which result in many noisy rela- tions) (Etzioni et al., 2004; Banko et al., 2007), or cluster the relation expressions based on the ex- tracted triplets to form relation clusters (Yao et al., 2011, 2012; Marcheggiani and Titov, 2016). How- ever, these approaches do not use existing high- quality and large-scale KBs when they are relevant to the relations to be discovered. In this paper, we consider a new relation dis- covery problem where both the training corpus for relation clustering and a KB are available, but the relations in …

Content truncated to 500 words