Discovering operational risk patterns in maritime accident reports through latent theme extraction

Maritime accidents pose persistent risks to human life, environmental sustainability, and economic infrastructure. Traditional accident analysis methods rely on structured data and predefined taxonomies, limiting their ability to detect emerging risks in unstructured narratives. This study applies Latent Dirichlet Allocation (LDA), an unsupervised topic modelling technique, to 247 maritime accident reports to uncover latent safety themes and operational failures. The optimised LDA model identified ten key topics, including rescue coordination failures, fire and smoke hazards, towing risks, grounding incidents, communication breakdowns, and alarm system deficiencies. Model performance was validated using coherence (0.55), topic diversity (0.68), and perplexity (183), confirming both interpretability and predictive reliability. High-frequency topics such as emergency response and onboard fires reveal gaps in preparedness and crew training. Low-frequency but high-impact issues—such as alarm failures—highlight critical overlooked vulnerabilities. This study offers a scalable, data-driven approach that complements traditional safety frameworks, enabling more effective risk prioritisation, targeted interventions, and resource allocation. Although limited by the subjectivity of narrative reports and absence of structured technical data, the findings provide a robust foundation for integrating multi-source datasets in future research. The approach supports more informed and proactive safety management across the global maritime industry.