Integrated GNN-LLM framework for automated task-level safety risk assessment using retrieval-augmented scenario generation

Conventional construction safety risk assessments frequently lack task-specific granularity, as they heavily rely on qualitative judgment and rigid, standardized formats. This paper proposes an integrated framework that leverages Large Language Model (LLM)-based entity extraction, heterogeneous accident correlation networks, and Graph Neural Networks (GNN) to generate prioritized accident scenarios featuring quantitative probability and severity estimates. Empirical results demonstrate that the extraction module achieved high accuracy, with an F1-score exceeding 0.90. Furthermore, the scenario generator exhibited robust diversity, yielding intra-list scores between 0.407 and 0.541, and aligned closely with historical data, as evidenced by a Jensen–Shannon divergence as low as 0.004. To ensure practical utility, a Retrieval-Augmented Generation (RAG) module transformed these outputs into concise narratives tailored for daily toolbox meetings. By offering a systematic and interpretable approach to safety management, this framework provides a scalable solution for task-level risk identification, particularly for resource-constrained construction environments.