Two are Better than One: Joint Entity and Relation Extraction with Table-Sequence Encoders

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1706–1721, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1706 Two are Better than One: Joint Entity and Relation Extraction with Table-Sequence Encoders Jue Wang1 and Wei Lu2 1College of Computer Science and Technology, Zhejiang University 2StatNLP Research Group, Singapore University of Technology and Design zjuwangjue@zju.edu.cn, luwei@sutd.edu.sg Abstract Named entity recognition and relation extrac- tion are two important fundamental problems. Joint learning algorithms have been proposed to solve both tasks simultaneously, and many of them cast the joint task as a table-filling problem. However, they typically focused on learning a single encoder (usually learning rep- resentation in the form of a table) to capture information required for both tasks within the same space. We argue that it can be beneficial to design two distinct encoders to capture such two different types of information in the learn- ing process. In this work, we propose the novel table-sequence encoders where two different encoders – a table encoder and a sequence en- coder are designed to help each other in the representation learning process. Our experi- ments confirm the advantages of having two encoders over one encoder. On several stan- dard datasets, our model shows significant im- provements over existing approaches.1 1 Introduction Named Entity Recognition (NER, Florian et al. 2006, 2010) and Relation Extraction (RE, Zhao and Grishman 2005; Jiang and Zhai 2007; Sun et al. 2011; Plank and Moschitti 2013) are two fundamental tasks in Information Extraction (IE). Both tasks aim to extract structured information from unstructured texts. One typical approach is to first identify entity mentions, and next perform classification between every two mentions to ex- tract relations, forming a pipeline (Zelenko et al., 2002; Chan and Roth, 2011). An alternative and more recent approach is to perform these two tasks jointly (Li and Ji, 2014; Miwa and Sasaki, 2014; Miwa and Bansal, 2016), which mitigates the error propagation issue associated with the pipeline ap- 1Our code is available at https://github.com/ LorrinWWW/two-are-better-than-one. ⊥ ⊥ Edward Thomas is from Minnesota United live_in live_in loc_in B-PER I-PER B-LOC B-LOC O O O ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ ⊥ live_in live_in loc_in , States Edward Thomas is from Minnesota , United States ⊥ ⊥ live_in live_in live_in live_in ⊥ ⊥ I-LOC live_in live_in live_in live_in loc_in loc_in ⊥ ⊥   Figure 1: An example of table filling for NER and RE. proach and leverages the interaction between tasks, resulting in improved performance. Among several joint approaches, one popular idea is to cast NER and RE as a table filling prob- lem (Miwa and Sasaki, 2014; Gupta et al., 2016; Zhang et al., 2017). Typically, a two-dimensional (2D) table is formed where each entry captures the interaction between two individual words within a sentence. NER is then regarded as a sequence labeling problem where …

Content truncated to 500 words