Document-level Role Filler Extraction exhibits a wide range of application value in natural language processing, including information retrieval, article summarization and trends analysis of world events. Existing document-level event role filler extraction methods face challenges in contextual modeling of long texts and ignore the explicit dependency relationships between event arguments displayed in long texts.
To solve the problems, a research team led by Zhengtao YU published their new research on 15 Feb 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team proposed a novel Element Relational Graph-Augmented Multi-Granularity Contextualized Encoder (ERGM) method for document-level event role filler extraction. Extensive experiments were conducted on the MUC-4 benchmark. Empirical results indicate that ERGM substantially outperforms strong baseline models. Additionally, the team demonstrated that the explicit graph-structured representation, generated by the graph neural network, can more effectively capture the dependency relationships between different event roles.
In the research, they extend the conventional document-level sequence tagging model with an additional graph encoder, enabling the production of an explicit structural representation of the source document while integrating multi-granularity information.
Specifically, the researchers initially construct a structural graph by extracting various elements from the source document, such as keywords, entities, and event triplets. They then utilize separate sentence-level encoders, document-level encoders, and a graph encoder to obtain sentence representations, document representations, and structural representations of the source document, respectively. Furthermore, to enhance the capture of comprehensive semantic information in lengthy texts and the interdependency among event roles, they employ a cross-attention mechanism to seamlessly integrate both document and structural representations. Finally, they dynamically integrate sentence and document representations and leverage a CRF (Conditional Random Field) inference layer for document-level event role extraction. Extensive experiments on the MUC-4 benchmark demonstrate that ERGM substantially outperforms strong baseline models. The study also illustrates that the explicit graph-structured representation generated by the graph neural network effectively captures dependency relationships between different event roles.
Future work can focus on exploring better ways to construct the knowledge graphs based on the source document, demonstrating the modeling of dependencies between event roles.
DOI: 10.1007/s11704-024-3701-4
