Enhanced human factor and causation analysis in maritime accidents using large language models

Journal article, 2026

Maritime accidents result from the interaction of environments, ships, and human factors, with communication breakdowns often initiating cascades of human errors. Marine Accident Investigation Reports (MAIRs) from various maritime safety agencies are recognized as the most common and efficient source to study human-related risk influencing factors (RIFs). However, these reports differ in standards, languages, and structures, making manual review, traditional natural language processing and expert judgment inadequate. Focusing on MAIRs involving communication breakdowns, this study applies large language models (LLMs) to analyze MAIRs from eight agencies in 2000-2024. By addressing report heterogeneity, the method highlights the embedded human factors and clarifies their causation roles. Based on the built LLM-assisted MAIR selection and RIF identification, a Bayesian Network (BN) is developed for accident scenario modeling and causation analysis. Results show that communication breakdowns appear in 224 of 682 MAIRs (32.84%) and exert great influence on both collisions (65%) and groundings (36%). Collisions are linked to more complex causal chains where human errors dominate, whereas groundings follow simpler pathways driven by improper use of navigational aids. Unlike prior literature that emphasizes standardized communication protocols, this study finds that enhancing very high frequency (VHF) communication is more relevant for collision prevention. By integrating LLM-assisted analysis with BN modeling, this study provides data-driven insights into the causal roles of human factors in maritime accidents and offers evidence to guide targeted risk control measures.