A Semantic Mention Graph Augmented Model for Document-Level Event Argument Extraction

TL;DR

This paper tackles document-level event argument extraction (DEAE) by addressing independent modeling of entity mentions and document-prompt isolation. It introduces GAM, a semantic mention graph that encodes co-existence, co-reference, and co-type relations, an ensembled graph transformer to fuse these signals, and a graph-augmented encoder-decoder that injects topology into a PLM-based generation pipeline. The approach yields state-of-the-art results on RAMS and WikiEvents, with ablations confirming the importance of each relation type and the graph components. The work demonstrates the practical impact of integrating structured graph information with prompt-based DEAE, improving argument identification and classification, and offering interpretability through explicit relations among mentions. It also opens avenues for future expansion with large language models and knowledge-grounded reasoning to further enhance event extraction precision and robustness.

Abstract

Document-level Event Argument Extraction (DEAE) aims to identify arguments and their specific roles from an unstructured document. The advanced approaches on DEAE utilize prompt-based methods to guide pre-trained language models (PLMs) in extracting arguments from input documents. They mainly concentrate on establishing relations between triggers and entity mentions within documents, leaving two unresolved problems: a) independent modeling of entity mentions; b) document-prompt isolation. To this end, we propose a semantic mention Graph Augmented Model (GAM) to address these two problems in this paper. Firstly, GAM constructs a semantic mention graph that captures relations within and between documents and prompts, encompassing co-existence, co-reference and co-type relations. Furthermore, we introduce an ensembled graph transformer module to address mentions and their three semantic relations effectively. Later, the graph-augmented encoder-decoder module incorporates the relation-specific graph into the input embedding of PLMs and optimizes the encoder section with topology information, enhancing the relations comprehensively. Extensive experiments on the RAMS and WikiEvents datasets demonstrate the effectiveness of our approach, surpassing baseline methods and achieving a new state-of-the-art performance.

Figures (6)

• Figure 1: An illustration of DEAE including the relevance among entity mentions with the same color labeled in the document and the co-type relation between the prompt and the document with the same color labeled.
• Figure 2: The architecture of GAM. The left part is an input example of the document and a corresponding prompt. The graph construction module (a) constructs a semantic mention graph including co-existence, co-reference and co-type relations from entity mentions and mask mentions. The ensembled graph transformer module (b) handles the text features combined with three semantic relations. Finally, the graph-augmented encoder-decoder module (c) is utilized to conduct the feature fusion and predict the arguments.
• Figure 3: The illustration of graph transformer. The inputs are the node sequence as well as the node position and the outputs are omitted. The Co-ex, Co-ref and Co-typ semantic mention relations are fused as a attention bias.
• Figure 4: The illustration of ablation study on WikiEvents dataset. The model performances under different $\lambda$.
• Figure 5: The different performance on RAMS dataset under different $\lambda$.