Enhanced AMR Graphs for Document-Level Event Argument Extraction

Event extraction, a core task in natural language processing(NLP), aims to identify event arguments in text. While recent studies have advanced document-level extraction with end-to-end models capable of cross-sentence reasoning, these models are still limited by long-distance dependencies and often miss the essential connections between global events. To tackle these issues, we introduce an end-to-end framework leveraging enhanced Abstract Meaning Representation (AMR) graphs. By incorporating AMR graphs as a semantic representation for end-to-end models, our model deciphers complex semantic structures and captures intra-event relationships, effectively addressing long-distance dependencies. We further propose a graph-to-graph external attention mechanism to encode AMR graphs. This mechanism dynamically refines node features through linear transformations, captures the relationships between different graphs, and enables implicit learning of cross-graph interactions. These designs strengthen global event interactions and information flow, improving the model's grasp of long-range event correlations. Experimental results on WikiEvents demonstrate that our framework outperforms existing methods and baselines. Ablation studies validate the contributions of individual components to overall performance.