Schema-Constrained Document-Level Event Argument Extraction with Lightweight LLM Fine-Tuning

2026-07-18

AI Summary

This work addresses boundary errors and schema inconsistencies in document-level event argument extraction caused by long-range dependencies, cross-sentence context, and strict role constraints. The authors propose a lightweight end-to-end framework that injects role-set prompts to ensure outputs conform to predefined schemas, employs parameter-efficient fine-tuning of a medium-scale open-source large language model via LoRA, and integrates a deterministic decoding pipeline with post-processing steps—including JSON validation, invalid role filtering, argument deduplication, and span alignment. This approach demonstrates for the first time that medium-scale open-source models can achieve reliable extraction under strong schema constraints, surpassing existing GPT-based baselines on the MAVEN-ARG benchmark with an F1 score of 42.39% across mention, entity coreference, and event coreference metrics.

Abstract

Event Argument Extraction (EAE) converts documents into structured event records by identifying argument spans and assigning them schema-defined roles. Document-level EAE is challenging due to long-range dependencies between triggers and arguments, cross-sentence context, and strict role constraints, which often lead to boundary errors, uncertainty in roles, and inconsistencies with restricted schemas. In this paper, we study whether mid-sized open LLMs can perform schema-constrained EAE reliably at the document level on MAVEN-ARG. Our approach combines (i) role-set injection in prompts for schema compliance, (ii) parameter-efficient supervised fine-tuning (LoRA) using the same JSON-only interface used at inference, and (iii) deterministic decoding with post-processing that validates JSON, filters invalid roles, de-duplicates arguments, and aligns spans to the document window. Under the official MAVEN-ARG evaluator, fine-tuned mid-sized open models outperform previously reported GPT baselines across mention, entity-coreference, and event-coreference evaluations; our best model (Phi-4, 14B) reaches 42.39\% F1 at the event-coreference level. Code to reproduce experiments is publicly available at https://github.com/dessertlab/EAE/.

Problem

• Document-level Event Argument Extraction
• Schema Constraints
• Long-range Dependencies
• Role Assignment
• Event Schema

Innovation

• schema-constrained EAE
• lightweight LLM fine-tuning
• LoRA
• deterministic decoding
• document-level event extraction