Incorporating Relation Paths in Neural Relation Extraction

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1768–1777 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Incorporating Relation Paths in Neural Relation Extraction Wenyuan Zeng1, Yankai Lin2, Zhiyuan Liu2∗, Maosong Sun2 1Department of Physics, Tsinghua University, Beijing, China 2State Key Laboratory of Intelligent Technology and Systems Tsinghua National Laboratory for Information Science and Technology Department of Computer Science and Technology, Tsinghua University, Beijing, China Abstract Distantly supervised relation extraction has been widely used to find novel rela- tional facts from plain text. To predict the relation between a pair of two target entities, existing methods solely rely on those direct sentences containing both en- tities. In fact, there are also many sen- tences containing only one of the target entities, which also provide rich useful in- formation but not yet employed by rela- tion extraction. To address this issue, we build inference chains between two target entities via intermediate entities, and pro- pose a path-based neural relation extrac- tion model to encode the relational seman- tics from both direct sentences and infer- ence chains. Experimental results on real- world datasets show that, our model can make full use of those sentences contain- ing only one target entity, and achieves significant and consistent improvements on relation extraction as compared with strong baselines. The source code of this paper can be obtained from https:// github.com/thunlp/PathNRE. 1 Introduction Knowledge Bases (KBs) provide effective structured information for real world facts and have been used as crucial resources for several nat- ural language processing (NLP) applications such as Web search and question answering. Typical KBs such as Freebase (Bollacker et al., 2008), DB- pedia (Auer et al., 2007) and YAGO (Suchanek et al., 2007) usually describe knowledge as multi- relational data and represent them as triple facts. As the real-world facts are infinite and increasing ∗Corresponding author: Z. Liu (liuzy@tsinghua.edu.cn). every day, existing KBs are still far from com- plete. Recently, petabytes of natural-language text containing thousands of different structure types are readily available, which is an important re- source for automatically finding unknown rela- tional facts. Hence, relation extraction (RE), de- fined as the task of extracting structured informa- tion from plain text, has attracted much interest. Most existing supervised RE systems usually suffer from the issue that lacks sufficient labelled relation-specific training data. Manual annotation is very time consuming and labor intensive. One promising approach to address this limitation is distant supervision. (Mintz et al., 2009) gener- ates training data automatically by aligning a KB with plain text. They assume that if two target entities have a relation in KB, then all sentences that contain these two entities will express this relation and can be regarded as a positive train- ing instance. Since neural models have been ver- ified to be effective for classifying relations from plain text (Socher et al., 2012; Zeng et al., 2014; dos Santos et al., 2015), (Zeng et al., 2015; Lin et al., 2016) incorporate neural networks method with distant …

Content truncated to 500 words