Selecting Optimal Context Sentences for Event-Event Relation Extraction

Understanding events entails recognizing the structural and temporal orders between event mentions to build structures/graphs for input documents. To achieve this goal, our work addresses problems of subevent relation extraction (SRE) (TRE) that aim predict relations two given mentions/triggers in texts. Recent state-of-the-art methods such have employed transformer-based language models (e.g., BERT) induce effective contextual representations mention pairs. However, a major limitation existing SRE TRE is they can only encode texts limited length (i.e., up 512 sub-tokens BERT), thus unable effectively capture important context sentences are farther away In work, we introduce novel method better model document-level with event-event extraction. Our seeks identify most entity pair document pack them into shorter documents be consume entirely by representation learning. The REINFORCE algorithm train where reward functions presented performance, context-based knowledge-based similarity problem. Extensive experiments demonstrate effectiveness proposed performance on benchmark datasets.