Explicit Relation-Enhanced AMR for Document-Level Event Argument Extraction with Global-Local Attention

In document-level event argument extraction, arguments are often more scattered than at the sentence level. Current methods have addressed the issue of argument scattering by incorporating additional information, but they still have limitations: a) the additional information introduced also generates noise. It leads to a weakening of the model’s focus; b) existing methods rely on simple concatenation operations. They are unable to capture the complex interactions between argument roles and triggers. To solve these problems, we present a novel Explicit Relation-Enhanced AMR (EREA) model. The EREA model effectively selects and utilizes critical information, improving focus on relevant event-related elements. In addition, the attention mechanism effectively integrates both local and global information, capturing complex relationships and dependencies to address the complexity of argument roles and trigger interaction. This module also improves the model’s efficiency in resource allocation and enables a more refined focus on relational data, which optimizes performance in event argument extraction. Empirical evidence from experiments conducted on WIKIEVENTS shows that our model, enhanced by the Explicit Relation-Enhanced AMR module and the Attention Fusion module, outperforms the state-of-the-art in Event Argument Extraction.