RSGT: Relational Structure Guided Temporal Relation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2001–2010 October 12–17, 2022. 2001 RSGT: Relational Structure Guided Temporal Relation Extraction Jie Zhou∗ Shenpo Dong∗ Hongkui Tu� Xiaodong Wang Yong Dou College of Computer National University of Defense Technology {jiezhou,dsp,tuhongkui,xdwang,yongdou}@nudt.edu.cn Abstract Temporal relation extraction aims to extract temporal relations between event pairs, which is crucial for natural language understanding. Few efforts have been devoted to capturing the global features. In this paper, we propose RSGT: Relational Structure Guided Temporal Relation Extraction to extract the relational structure features that can fit for both inter- sentence and intra-sentence relations. Specifi- cally, we construct a syntactic-and-semantic- based graph to extract relational structures. Then we present a graph neural network based model to learn the representation of this graph. After that, an auxiliary temporal neighbor pre- diction task is used to fine-tune the encoder to get more comprehensive node representa- tions. Finally, we apply a conflict detection and correction algorithm to adjust the wrongly pre- dicted labels. Experiments on two well-known datasets, MATRES and TB-Dense, demonstrate the superiority of our method (2.3% F1 im- provement on MATRES, 3.5% F1 improve- ment on TB-Dense). 1 Introduction Temporal relation extraction (TRE) is crucial for natural language understanding and can facilitate various downstream applications such as summa- rization (Zhou et al., 2010), question answering (Yu et al., 2017), and clinical diagnosis (Zhou et al., 2021). As shown in Figure 1, the goal of TRE is to determine the temporal order between an event pair (BEFORE, AFTER, etc.). Most early methods were based on statistical ma- chine learning (Mani et al., 2006; Chambers, 2013). In recent years, neural network based methods and large-scale pre-trained language models such as BERT (Devlin et al., 2018) have contributed to a substantial increase in the performance of TRE task (Ning et al., 2019; Wang et al., 2020). *These authors contributed equally to this work After Before Equal Before S1: Former President Nicolas Sarkozy was (e1,informed) Thursday that he would face a formal investigation into whether he (e2,abused) the frailty of Liliane Bettencourt, to get funds for his 2007 presidential campaign. S2: Mr. Sarkozy has (e3,denied) accepting illegal campaign funds from Ms. Bettencourt, either personally or through his party treasurer at the time, Eric Woerth, as (e4,alleged) by her former butler. informed abused denied alleged Figure 1: An example of temporal relation extraction. There are four events in these sentences. The graph below shows the pair-wise event temporal relationships. However, these methods may ignore the global structure features which carry non-consecutive and long-distance semantics (Peng et al., 2018). This shortcoming is obvious in dealing with an event pair that the two events belong to differ- ent sentences (inter-sentences event pair), such as < e1, e4 > in Figure 1. Few previous works differentiate inter-sentence event pairs from intra- sentence ones (where the two events appear in the same sentence). Thus, the performance may be impacted. For example, we test that there is a 5 accuracy points gap between intra-sentences and inter-sentence event …

Content truncated to 500 words