GAN Driven Semi-distant Supervision for Relation Extraction

Proceedings of NAACL-HLT 2019, pages 3026–3035 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3026 GAN Driven Semi-distant Supervision for Relation Extraction Pengshuai Li1, Xinsong Zhang1, Weijia Jia2,1∗and Hai Zhao1∗ 1Dept. of CSE, Shanghai Jiao Tong University, Shanghai, China 2State Key Lab of IoT for Smart City, CIS, University of Macau, Macao, SAR China {pengshuai.li,xszhang0320}@sjtu.edu.cn {jia-wj,zhaohai}@cs.sjtu.edu.cn Abstract Distant supervision has been widely used in relation extraction tasks without hand-labeled datasets recently. However, the automati- cally constructed datasets comprise numbers of wrongly labeled negative instances due to the incompleteness of knowledge bases, which is neglected by current distant supervised methods resulting in seriously misleading in both training and testing processes. To address this issue, we propose a novel semi-distant su- pervision approach for relation extraction by constructing a small accurate dataset and prop- erly leveraging numerous instances without re- lation labels. In our approach, we construct ac- curate instances by both knowledge base and entity descriptions determined to avoid wrong negative labeling and further utilize unlabeled instances sufficiently using generative adver- sarial network (GAN) framework. Experimen- tal results on real-world datasets show that our approach can achieve significant improve- ments in distant supervised relation extraction over strong baselines. 1 Introduction Relation extraction aims to identify relations for a pair of entities in a sentence to construct relation triples like [Steve Jobs, Founder, Apple]. It has been well studied by supervised approaches with hand-labeled data. However, supervised methods are limited to costly hand-labeled training sets and ∗ Corresponding authors: Weijia Jia, Hai Zhao, {jia- wj, zhaohai}@cs.sjtu.edu.cn. This work is supported by Na- tional China 973 Project No. 2015CB352401; Chinese Na- tional Research Fund (NSFC) Key Project No. 61532013 and No. 61872239. 0007/2018/A1, DCT-MoST Joint- project No. 025/2015/AMJ,FDCT,SAR Macau, China, and University of Macau Grant Nos: MYRG2018-00237-RTO, CPG2019-00004-FST and SRG2018-00111-FST, National Key Research and Development Program of China (No. 2017YFB0304100), Key Projects of National Natural Sci- ence Foundation of China (U1836222 and 61733011), and Key Project of National Society Science Foundation of China (No. 15-ZDA041). hard to be extended to large-scale relations. To break the bottleneck of hand-labeled training set, distant supervision (Mintz et al., 2009) automati- cally construct datasets with knowledge bases. It assumes that if two entities have a known rela- tion in a knowledge base, all sentences that men- tion these two entities will probably express the same relation and can be called positive instances. At the same time, it treats sentences as nega- tive instances whose entity pairs do not have a known relation in knowledge bases. Due to the strong assumption, instances are likely to be mis- labeled. To alleviate the wrong labeling prob- lem, distant supervised methods have been im- plemented with multi-instance learning and neu- ral networks (Riedel et al., 2010; Hoffmann et al., 2011; Zeng et al., 2015; Lin et al., 2016, 2017). However, previous works focus on positive in- stances and few methods have addressed the issue of false negative instances. The false negative instances, which contain true relations, …

Content truncated to 500 words