Causality-aware LLM-agent framework for document-level news event extraction with heterogeneous arguments

With the explosion of unstructured text data in news reports, there is an urgent need for automatic event extraction to support timely decision-making. This paper tackles the challenging task of document-level multi-type news event extraction with heterogeneous argument structures. Unlike sentence-level extraction, this task must handle long documents containing multiple events of different types, each with a distinct set of roles (arguments). Existing event extraction approaches struggle with incomplete information and widely scattered clues in lengthy news articles. To address these issues, we propose a causality-aware LLM-agent framework that leverages large language models (LLMs) and an agent-based extraction strategy. First, our method uses an LLM-driven event detector to identify event triggers and types in a document, even when event mentions are sparse or spread across paragraphs. Second, an argument extraction agent performs state-aware role-wise extraction for each event. It initializes a schema-specific role queue and uses QA feedback, including returned spans, confidence scores, and NULL decisions, to adjust subsequent query formulation and local context selection. This yields more complete arguments while keeping predictions grounded in source spans. Third, we incorporate a causal inference module that links events with cause-effect relations, using these relations to infer or validate certain arguments (e.g. shared locations or times) when explicit mentions are missing. We conduct extensive experiments on two public datasets for document-level event extraction from news documents. Results show that our approach achieves state-of-the-art performance, outperforming strong baselines by 3-5% F1. Notably, it maintains high recall on long news documents with multiple events, demonstrating robustness to incomplete and scattered information. These improvements validate the effectiveness of integrating LLM reasoning, agent-based extraction, and causal inference. Our work provides a novel solution for accurate multi-event extraction from complex news documents, and we discuss its significance for downstream news understanding tasks and future extensions to broader domains.