TPLinker: Single-stage Joint Extraction of Entities and Relations Through Token Pair Linking

Proceedings of the 28th International Conference on Computational Linguistics, pages 1572–1582 Barcelona, Spain (Online), December 8-13, 2020 1572 TPLinker: Single-stage Joint Extraction of Entities and Relations Through Token Pair Linking Yucheng Wang1, Bowen Yu1∗, Yueyang Zhang2 Tingwen Liu1, Hongsong Zhu1, Limin Sun1 1School of Cyber Security, University of Chinese Academy of Sciences 1Institute of Information Engineering, Chinese Academy of Sciences 2Baidu, Inc. {wangyucheng,yubowen,liutingwen,zhuhongsong,sunlimin}@iie.ac.cn zhangyueyang@baidu.com Abstract Extracting entities and relations from unstructured text has attracted increasing attention in recent years but remains challenging, due to the intrinsic difficulty in identifying overlapping relations with shared entities. Prior works show that joint learning can result in a noticeable performance gain. However, they usually involve sequential interrelated steps and suffer from the problem of exposure bias. At training time, they predict with the ground truth conditions while at inference it has to make extraction from scratch. This discrepancy leads to error accumulation. To mitigate the issue, we propose in this paper a one-stage joint extraction model, namely, TPLinker, which is capable of discovering overlapping relations sharing one or both entities while immune from the exposure bias. TPLinker formulates joint extraction as a token pair linking problem and introduces a novel handshaking tagging scheme that aligns the boundary tokens of entity pairs under each relation type. Experiment results show that TPLinker performs significantly better on overlapping and multiple relation extraction, and achieves state-of-the-art performance on two public datasets1. 1 Introduction Extracting entities and relations from unstructured texts is an essential step in automatic knowledge base construction (Takanobu et al., 2019). Traditional pipelined approaches first extract entity mentions and then classify the relation types between candidate entity pairs. However, due to the complete separation of entity detection and relation classification, these models ignore the interaction and correlation between the two subtasks, being susceptible to cascading errors (Li and Ji, 2014). In the last few years, there has been increasing research interest in building joint models to simulta- neously extract entities and relations. Recent works show that joint learning approaches can effectively integrate the information of entity and relation, and therefore achieve better performance in both sub- tasks (Dai et al., 2019; Tan et al., 2019). Zheng et al. (2017) proposed a unified tagging scheme to convert joint extraction to a sequence labeling problem but lacks the elegance to identify overlapping relations: one entity may participate in multiple relations in the same text (Figure 1). Most existing models in handling EntityPairOverlap (EPO) and SingleEntiyOverlap (SEO) cases can be categorized into two classes: decoder-based and decomposition-based. Decoder-based models use encoder-decoder architecture where the decoder extracts one word or one tuple at a time like machine translation models (Zeng et al., 2018; Nayak and Ng, 2020). Decomposition-based models first distin- guish all the candidate subject entities that may be involved with target relations, then label corresponding object entities and relations for each extracted subject (Li et al., 2019; Yu et al., 2020; Wei et al., 2020). Although these methods have achieved reasonable performance, they all suffer from the same …

Content truncated to 500 words