One Small and One Large for Document-level Event Argument Extraction

ACL ARR 2026 January Submission5706 Authors

Keywords: Event Extraction; Event Argument Extraction; Large Language Model

Abstract: Document-level Event Argument Extraction (EAE) is hampered by two key chal- lenges in long texts: ambiguity among co-occurring events and noise from irrel- evant content. To address these issues, we propose CsEAE, a unified framework that comprises two synergistic modules. The co-occurrence-aware module delin- eates ambiguous event boundaries by modeling dependencies among co-occurring events, while the structure-aware module filters noise by modeling trigger-centric sentence relations. We further extend this framework to Large Language Models (LLMs) with CsLLM, which distills these structural and co-occurrence cues into tailored prompts. Trained on multiple datasets, CsLLM enhances the generaliza- tion and performance of LLMs on the EAE task. On the RAMS, WikiEvents, and MLEE benchmarks, CsEAE improves Arg-C F1 scores over the PAIE baseline by 2.1%, 2.3%, and 3.2%, respectively. Our LLM-based approach, CsLLM, achieves even greater performance, demonstrating the effectiveness of our framework

Paper Type: Long

Research Area: Information Extraction and Retrieval

Research Area Keywords: Information Extraction

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: Engilish

Submission Number: 5706