Enhancing Event Detection with Inter-Event Dependencies in Large Ontologies

Event Detection (ED), a crucial component of comprehensive text analysis tools, is a well-established task within the fields of Natural Language Processing (NLP) and Information Extraction (IE). Current state-of-the-art models for ED primarily focus on identifying a limited set of predefined event types. Recently, the challenge of detecting a broad array of predefined event types has garnered increasing interest within the IE community. However, a significant gap in existing research on ED with extensive ontologies is the inadequate exploration of how interactions between event types affect ED model performance. One of the hindrances for this purpose is the lack of resources to encode event-event dependencies for large ontologies. This study introduces a novel approach that leverages existing inter-event dependency resources to provide this information for extensive ontologies. Specifically, a solution based on Optimal Transport is proposed to map event-event dependency from existing resources to a large ontology. We conduct extensive experiments on multiple benchmark datasets to assess the effectiveness of our approach. Our findings, supported by a thorough analysis, demonstrate that this innovative technique significantly enhances the performance of ED models, especially for ontologies with a large number of event types.