Learning to Generate and Extract: A Multi-Agent Collaboration Framework For Zero-shot Document-level Event Arguments Extraction

ABSTRACT

Extracting structured event arguments from massive, unstructured document repositories is highly error-prone using single-pass LLMs. This paper proposes a collaborative multi-agent architecture featuring specialized Generator, Extractor, and Verifier agents. By debating and iteratively refining their extraction patterns, the swarm achieves unprecedented accuracy in zero-shot environments, drastically reducing the need for labeled training data.