MILE: Multi-Instance Learning for Document Event Argument Extraction

Current research on event extraction primarily utilizes intra-document information and overlooks inter-document relationships. Meanwhile, existing instance learning methods in event extraction overly rely on similarity-based retrieval, failing to capture event complexities and document semantics. To address these issues, we propose an event extraction model based on multi-instance learning, exploring the connections between documents through event types and event arguments. We design multiple instance selection strategies and construction methods to facilitate a comprehensive grasp of event semantics. Furthermore, we implemented a two-stage training approach to optimize the model’s ability to learn from instances obtained through various selection strategies. Experiments conducted on the RAMS and WIKIEVENTS datasets demonstrate that our method surpasses the current state-of-the-art models in certain metrics, validating its effectiveness and superiority. ¹