Graph Convolutional Networks for Event Causality Identification with Rich Document-level Structures

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3480–3490 June 6–11, 2021. ©2021 Association for Computational Linguistics 3480 Graph Convolutional Networks for Event Causality Identification with Rich Document-level Structures Minh Tran Phu VinAI Research Hanoi, Vietnam v.minhtp@vinai.io Thien Huu Nguyen Department of Computer and Information Science University of Oregon, Eugene, Oregon, USA thien@cs.uoregon.edu Abstract We study the problem of Event Causality Iden- tification (ECI) to detect causal relation be- tween event mention pairs in text. Although deep learning models have recently shown state-of-the-art performance for ECI, they are limited to the intra-sentence setting where event mention pairs are presented in the same sentences. This work addresses this issue by developing a novel deep learning model for document-level ECI (DECI) to accept inter- sentence event mention pairs. As such, we pro- pose a graph-based model that constructs inter- action graphs to capture relevant connections between important objects for DECI in input documents. Such interaction graphs are then consumed by graph convolutional networks to learn document context-augmented representa- tions for causality prediction between events. Various information sources are introduced to enrich the interaction graphs for DECI, featur- ing discourse, syntax, and semantic informa- tion. Our extensive experiments show that the proposed model achieves state-of-the-art per- formance on two benchmark datasets. 1 Introduction Event Causality Identification (ECI) is an important problem in Information Extraction that seeks to pre- dict causal relation between a pair of events men- tioned in text. For instance, in the sentence “The building was nearly destroyed by a fire early Tues- day morning.”, an ECI system should be able to rec- ognize the causal relation between the two events triggered by “destroyed” and “fire” (called event mentions), i.e., “fire” cause −−−→ “destroyed”. ECI finds its applications for a wide range of problems in natural language processing (NLP), including ma- chine reading comprehension (Berant et al., 2014), future event forecasting (Hashimoto, 2019), and why-question answering (Oh et al., 2016). The early approach for ECI has involved feature- based methods (Do et al., 2011; Hashimoto, 2019; … Nugroho told Agence-France Presse : "We have recorded 24 people dead and 249 people injured", adding that more than 300 buildings have been damaged1 due to the quake1. Rescuers and other assistance teams have arrived in Bener Meriah, while the air force have dispatched a helicopter and a CN-235 aircraft to the region, Nugroho said. "We are now concentrating on searching for people who may be trapped under the rubble," said Rusli M . Saleh, the deputy district chief of Bener Meriah. He said at least 25 of the injured in his district were hospitalized in intensive care. As the quake hit, villagers in the area ran out of their homes in panic and screamed for help. "I see many houses were damaged2 and their roofs fell onto some people," Bensu Elianita , a 22-year-old resident of Bukit Sama village in Central Aceh district, said shortly after the quake2 hit. … causal …

Content truncated to 500 words