OD-RTE: A One-Stage Object Detection Framework for Relational Triple Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11120–11135 July 9-14, 2023 ©2023 Association for Computational Linguistics OD-RTE: A One-Stage Object Detection Framework for Relational Triple Extraction Jinzhong Ning, Zhihao Yang∗, Yuanyuan Sun, ZhiZheng Wang, Hongfei Lin School of Computer Science and Technology, Dalian University of Technology, China (Jinzhong_Ning, wzz_dllg)@mail.dlut.edu.cn (yangzh, syuan, hflin)@dlut.edu.cn Abstract The Relational Triple Extraction (RTE) task is a fundamental and essential information ex- traction task. Recently, the table-filling RTE methods have received lots of attention. De- spite their success, they suffer from some in- herent problems such as underutilizing regional information of triple. In this work, we treat the RTE task based on table-filling method as an Object Detection task and propose a one-stage Object Detection framework for Relational Triple Extraction (OD-RTE). In this frame- work, the vertices-based bounding box detec- tion, coupled with auxiliary global relational triple region detection, ensuring that regional information of triple could be fully utilized. Besides, our proposed decoding scheme could extract all types of triples. In addition, the neg- ative sampling strategy of relations in the train- ing stage improves the training efficiency while alleviating the imbalance of positive and neg- ative relations. The experimental results show that 1) OD-RTE achieves the state-of-the-art performance on two widely used datasets (i.e., NYT and WebNLG). 2) Compared with the best performing table-filling method, OD-RTE achieves faster training and inference speed with lower GPU memory usage. To facilitate future research in this area, the codes are pub- licly available at https://github.com/ NingJinzhong/ODRTE. 1 Introduction The Relational Triple Extraction (RTE for short) aims to extract triples of the form (head, relation, tail) consisting of entity pairs and their relations from unstructured text, which is an important task of information extraction. In the early stage, tradi- tional pipeline methods (Zelenko et al., 2003; Zhou et al., 2005; Chan and Roth, 2011) usually decom- pose the RTE task into two independent steps of named entity recognition and relation extraction. ∗* Corresponding author “Born in New York City , Tom Luna lives in Las Vegas.” (Tom Luna, place_of_birth, New York City) (Tom Luna, place_lived, Las Vegas) Object Detection Relational Triple Extraction …… in New York City , …… in Las Vegas Tom Luna lives , PB-1 PB-2 PB-3 PB-4 PL-1 PL-2 PL-3 PL-4 …… …… …… …… …… …… …… …… …… …… …… …… Figure 1: A Comparison of Object Detection and Rela- tional Triple Extraction based on table-filling method. PB denotes relation "place_of_birth" and PL denotes relation "place_lived". Note that the table cells (repre- sentations of token pairs), the table, and the table regions occupied by the triples are aligned to the pixels, the im- age, and the objects on the visual side, respectively. Although the pipelined approach is flexible, it ig- nores the correlation between the two tasks and suffers from error propagation (Ren et al., 2022). To overcome this problem, some researchers try to use end-to-end joint entity and relation extraction models to solve …

Content truncated to 500 words