Enhanced Fusion of AMR Graph Representations for Event Argument Extraction

Event Argument Extraction (EAE) is a key task in natural language processing that involves identifying and classifying events and their arguments in text. With advancements in language models, the quality of semantic representation is becoming more crucial for the success of EAE. Abstract Meaning Representation (AMR), as a graphical semantic representation form, effectively captures the semantic structures of complex sentences, providing valuable contextual information for the EAE task. Current research indicates that AMR plays an important role in EAE by offering detailed event and argument relationships, helping models better understand the semantic aspects of text. However, existing models often face inefficiencies in integrating AMR graph information, leading to subpar performance in extracting from complex texts. Additionally, many models fail to fully utilize multimodal information, limiting their performance improvements. Therefore, we propose EAGF, an Enhanced AMR Graph Fusion method, which aims to improve EAE effectiveness through enhanced AMR graph representation fusion. EAGF first conducts in-depth analysis of the AMR graph using Graph Attention Networks (GAT), extracting graph embeddings with rich context. Then, these embeddings undergo cross-modal attention fusion with the input text embeddings, and the fusion results are injected into the attention mechanism of the encoder in the generative model. Through this approach, EAGF effectively integrates information from different sources, improving the accuracy of event argument extraction. Experimental results show that EAGF achieves significant performance improvements across multiple benchmark datasets.