Document-Level Event Temporal Relation Extraction on Global and Local Cues
Jing Li, Sheng Xu, Peifeng Li · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
Most previous work focused on extracting event temporal relations that the events appear in the same sentence or in two adjacent sentences, failing to address those nonadjacent-sentence event relations, which limits the development of tem-poral relation extraction at document-level and its real-world application. In this paper, we propose a novel Document-level event Temporal Relation Extraction (DTRE) model which can incorporate effective global cues with local cues. In particular, we select both the contextual sentences strongly related to the events and the temporal words in the context as global cues, which can provide additional semantic cues to extract those nonadjacent-sentence event temporal relations. Moreover, we further encode the events and their neighbor words as local cues to extract those intra-sentence relations and enhance the event representation. Experimental results on the English dataset show that our proposed DTRE outperforms several state-of-the-art baselines, especially for handling those nonadjacent-sentence temporal relations.