Event Logic Graph Augmented Large Language Models for Fall Accident Risk Reasoning

Abstract

Falls represent the most prevalent type of construction accident, emphasizing the critical importance of implementing effective risk management strategies. The advent of large language models (LLMs) has introduced novel perspectives to address this challenge. This paper presents an original methodology for enabling risk reasoning in fall accidents through the utilization of an event logic graph (ELG) to enhance LLMs. First, event relations and elements are extracted from accident reports to construct an ELG using semantic matching and deep learning techniques. Second, actual events are mapped to the ELG by applying a text similarity method. Then, accident casualties and corresponding medical procedures are identified to calculate loss values and risk metrics for fall accidents. Ultimately, the ELG augmented Qwen3-Max model is employed to generate risk assessments and operational recommendations based on relevant safety regulation provisions. This approach is validated through a real case. Results demonstrate that it effectively supports risk analysis and provides practical preventive recommendations for fall accidents.