WordRelationEE: A Biaffine Approach to Event Extraction

Event extraction is an essential task in the field of natural language processing, aiming to pinpoint predefined event types’ triggers and arguments. Current methods, including sequence labeling and sequence generation techniques, face challenges such as overlapping or exposure bias, hindering effective event extraction. In response, we introduce a novel method that conceptualizes event extraction as classifying relationships between words. This approach employs a Biaffine mechanism to categorize “Start-to-End-” (S2E-) and “Next-to-Next” (N2N) connections among words (for instance, trigger, argument, entity), addressing the issue of overlap. Furthermore, it enhances the decoding process by integrating interactions of features among labels, embedding these interactions for improved performance. Our method demonstrates superior results in all subtasks on the “ACE05-E + ” dataset, outperforming existing event extraction techniques.