DCT-Centered Temporal Relation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2087–2097 October 12–17, 2022. 2087 DCT-Centered Temporal Relation Extraction Liang Wang, Peifeng Li ∗and Sheng Xu School of Computer Science and Technology Soochow University, Jiangsu, China docy@vip.qq.com, pfli@suda.edu.cn, sxu@stu.suda.edu.cn Abstract Most previous studies on temporal relation ex- traction focus on extracting temporal relations among events and suffer from the issue of dif- ferent forms of events, timexes and Document Creation Time (DCT) in a document. More- over, DCT can act as a hub to semantically connect the other events and timexes. Unfor- tunately, previous work cannot fully use such critical and helpful information. To address the above issues, we propose a unified DCT- centered Temporal Relation Extraction model DTRE to identify temporal relations among events, timexes and DCT. Specifically, we first introduce sentence-style DCT to unify the ex- pressions of event, timex and DCT. Then, we apply a DCT-aware graph to obtain their con- textual structural representations. Furthermore, we propose a DCT-anchoring multi-task frame- work to jointly predict three tasks of temporal relation extraction in a batch. Finally, we pro- vide a DCT-guided global inference to further enhance the global consistency among different relations. Experimental results on three pop- ular datasets TBD, TDD-man and TDD-Auto show that our DTRE outperforms several SOTA baselines on E-E, E-T and E-D significantly. 1 Introduction Temporal relation extraction focuses on the occur- rence order (TLINK) of event mentions, time ex- pressions (timexes) and Document Creation Time (DCT). Most previous studies only focus on the event-centered tasks and consider three TLINKs: event-event (E-E), event-timex (E-T), and event- DCT (E-D). As a crucial component of relation extraction, temporal relation extraction can help many downstream NLP tasks, such as question an- swering (Ning et al., 2020), summarization (Noh et al., 2020) and timeline construction (Li et al., 2021). ∗Corresponding author DCT: 1998-02-27 The Organization of African Unity <e2 said /e2> <t41 value=”1998-02-27” Friday /t41> it would investigate the Hutu-organized genocide… ... to investigate who <e9 shot /e9> down Rwandan President Juvenal Habyarimana's plane on April 6, 1994 . DCT Simultaneous Is Included After E2 T41 E9 Is included Before After DCT Simultaneous Is Included After E2 T41 E9 Is included Before After DCT: 1998-02-27 The Organization of African Unity <e2 said /e2> <t41 value=”1998-02-27” Friday /t41> it would investigate the Hutu-organized genocide… ... to investigate who <e9 shot /e9> down Rwandan President Juvenal Habyarimana's plane on April 6, 1994 . DCT Simultaneous Is Included After E2 T41 E9 Is included Before After Figure 1: Examples of temporal relations among events, timex and DCT. Most previous studies (Mathur et al., 2021; Liu et al., 2021) only focus on the single E-E task, ig- noring the other E-T and E-D tasks. The main barricade is the hardness of combining the E-T or E-D task with the E-E task due to their different expression forms. Since most event mentions 1 are sentences or clauses, their rich information is helpful for a neural network model to identify the specific relation between …

Content truncated to 500 words