Researchers at the University of Strathclyde have published a new study in Ocean Engineering presenting a domain-adapted natural language processing framework for automatically identifying and classifying human and organisational factors in maritime accident reports. The research addresses a longstanding challenge in maritime safety analysis: accident investigations contain large volumes of valuable narrative information, but manual coding is time-consuming and may be affected by subjective interpretation and inconsistent application of safety taxonomies.

At the centre of the proposed framework is SeaSafeBERT, a maritime-specific language model trained on an international collection of accident investigation reports. The model combines domain-adapted BERT representations with TF-IDF lexical features to analyse causal and contributory text segments and assign them to structured human- and organisational-factor categories. The researchers also developed SHIELD-NLP, a machine-learning-ready adaptation of the SHIELD taxonomy, enabling the classification of accident information into 53 detailed factor codes.

The framework was evaluated using 330 maritime accident investigations and achieved strong classification performance for a number of operationally important categories, with several recording F1 scores above 0.80. Association-rule analysis further revealed recurring relationships among perceptual, cognitive, technical and organisational factors across different accident types. These findings demonstrate how artificial intelligence can support faster and more consistent analysis of maritime incidents while helping safety professionals identify recurring risk patterns that may not be readily visible through conventional manual review.

The study provides a new data-driven tool for improving organisational learning, accident prevention and proactive safety management across the maritime sector. The paper, entitled A Domain-Adapted NLP Pipeline for Human Factors Classification of Maritime Accidents, was published on 1 June 2026.