Document-level event argument extraction with dual-paradigm fusion strategy

Abstract

Abstract Document-level Event Argument Extraction (DEAE) is a critical task in the field of information extraction, aiming to identify arguments and their semantic roles for specific event types from documents. While methods based on the third paradigm (classification-based discrimination) and the fourth paradigm (generative) have made progress in DEAE tasks, existing approaches still face challenges such as feature redundancy, inadequate modeling of long-distance dependencies, and difficulties in capturing semantic correlations among arguments. To address these issues, this paper proposes a Document-level Event Argument Extraction method integrating the third and fourth paradigms (DPFS), which breaks through the limitations of single methods through collaborative modeling of dual paradigms. DPFS comprises three core modules: a prompt summary generation module to eliminate redundant information, an argument dependency module to capture cross-text dependencies, and an embedding fusion module to integrate generative and discriminative features. Experiments on the RAMS, WikiEvents and, LEE datasets demonstrate that DPFS achieves significant improvements of 3.0%, 1.2% and 2.4% in argument classification F1 scores, respectively, outperforming existing methods. Its modular design enables flexible enhancement of existing model performance.