Ontology-Aware Overlapping Event Extraction

Event extraction is an essential task in natural language processing, where the goal is to identify and categorize events, participants, times and locations from text. This task has become challenging when dealing with complex sentences that encompass overlapping sub-events. To address this issue, we propose a novel ontology-aware neural approach for extracting overlapping events. Our approach consists of an Ontology-Aware Semantic Encoder (OASE) that enriches text representations with event ontology knowledge and semantic features. The decoding process involves a synergistic trio of modules: type detection, trigger extraction and argument extraction, operating in a pipelined fashion. This process is further refined by imposing constraints on event types and argument roles. The model is evaluated on a Chinese financial event extraction benchmark, where it demonstrated state-of-the-art performance, outperforming existing models in terms of accuracy and efficiency. Ablation studies underscore the importance of our model's key components. The integration of ontology-aware modeling and constrained decoding significantly enhances the extraction of complex and overlapping events, showcasing the model's effectiveness in handling intricate event structures.