MrMep: Joint Extraction of Multiple Relations and Multiple Entity Pairs Based on Triplet Attention

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 593–602 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 593 MrMep: Joint Extraction of Multiple Relations and Multiple Entity Pairs Based on Triplet Attention Jiayu Chen, Caixia Yuan, Xiaojie Wang and Ziwei Bai Center of Intelligence Science and Technology School of Computer Science Beijing University of Posts and Telecommunications {jiayuchen, yuancx, xjwang, bestbzw}@bupt.edu.cn Abstract This paper focuses on how to extract multiple relational facts from unstructured text. Neu- ral encoder-decoder models have provided a viable new approach for jointly extracting re- lations and entity pairs. However, these mod- els either fail to deal with entity overlapping among relational facts, or neglect to produce the whole entity pairs. In this work, we pro- pose a novel architecture that augments the en- coder and decoder in two elegant ways. First, we apply a binary CNN classifier for each re- lation, which identifies all possible relations maintained in the text, while retaining the tar- get relation representation to aid entity pair recognition. Second, we perform a multi- head attention over the text and a triplet at- tention with the target relation interacting with every token of the text to precisely produce all possible entity pairs in a sequential man- ner. Experiments on three benchmark datasets show that our proposed method successfully addresses the multiple relations and multiple entity pairs even with complex overlapping and significantly outperforms the state-of-the- art methods. All source code and documen- tations are available at https://github. com/chenjiayu1502/MrMep. 1 Introduction Extracting relational facts from unstructured text is a significant step in building large-scale knowl- edge graphs. A relational fact is typically repre- sented as a triplet <h, r, t> where h represents a head entity, t represents a tail entity, and r is a relation that connects h to t. A number of neural models for extracting relational facts have been developed in recent years, and the most successful models all have one thing in common: they extract both rela- tion and its corresponding entity pairs in a man- ner of joint learning (Zheng et al., 2017; Zeng et al., 2018; Takanobu et al., 2019). However, jointly extracting relation and entity pairs is far from a trivial task due that there might exist more than one relation within the text, while each target relation might correspond to more than one entity pair. A more challenging aspect of this problem is that there might exist com- plex overlapping among different triplets. As ex- emplified in Figure 1, the entity “Indonesia” overlaps in all triplets within the text, while <Indonesia, leaderName, Jusuf Kalla> and <Indonesia, leaderName, Joko Widodo> have the same relation type “leaderName”, <Bakso, region, Indonesia> and <Bakso, country, Indonesia> share the same head and tail entities. A well-defined relational fact extraction task aims to detect all possible relation types in the text, and extract all candidate entity pairs for each target relation type, while taking into account the com- plicated overlapping among the triplets. Zheng et …

Content truncated to 500 words