A domain-adapted NLP pipeline for human factors classification of maritime accidents

Tools

Black, Hollie and Kurt, Rafet and Turan, Osman (2026) A domain-adapted NLP pipeline for human factors classification of maritime accidents. Ocean Engineering, 357 (Part 3). 125625. ISSN 0029-8018 (https://doi.org/10.1016/j.oceaneng.2026.125625)

[thumbnail of Black-etal-OE-2026-A-domain-adapted-NLP-pipeline-for-human-factors-classification-of-maritime-accidents]
Preview
 Text. Filename: Black-etal-OE-2026-A-domain-adapted-NLP-pipeline-for-human-factors-classification-of-maritime-accidents.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo
Download (7MB)| Preview

Abstract

Human error continues to account for most maritime accidents, highlighting persistent challenges in safety management and organisational learning. Although post-incident investigations generate rich narrative data, extracting meaningful insights is hindered by manual coding, subjective interpretation, and inconsistent application of human-factors taxonomies. Despite increasing interest in applying NLP to safety analysis, no existing approach provides a taxonomy-ready, domain-adapted framework capable of reliably classifying fine-grained human and organisational factors in maritime accident narratives. This study presents a natural language processing (NLP) and machine learning (ML) framework to automate the identification of human and organisational contributory factors in maritime accident reporting using an adapted SHIELD taxonomy. The approach analyses individual causal and contributory text segments extracted from accident investigation narratives, which are classified into human and organisational factor codes. Central to the approach is SeaSafeBERT, a domain-adapted BERT model fine-tuned on an international corpus of maritime investigations and combined with TF-IDF lexical features to support structured, multi-level classification. To address sparse taxonomy classes, we developed SHIELD-NLP, a consolidated coding scheme preserving SHIELD's intent but ensuring sufficient class representation across a dataset of 330 maritime accident investigations and enabling classification into 53 SHIELD-NLP codes. The system achieved strong performances for many operationally important SHIELD-NLP codes, with several reaching F1 scores above 0.80. Association rule mining further revealed consistent interaction patterns between perceptual, cognitive, and organisational factors across incident types. Together, these contributions represent, to our knowledge, one of the first domain-specific maritime safety language models and integrated pipelines for automated SHIELD-based analysis, demonstrating the potential of AI-enabled methods to accelerate incident learning, improve code consistency, and strengthen proactive safety management in maritime operations.

ORCID iDs

Black, Hollie ORCID logoORCID: https://orcid.org/0009-0002-5099-9302, Kurt, Rafet ORCID logoORCID: https://orcid.org/0000-0002-5923-0703 and Turan, Osman ORCID logoORCID: https://orcid.org/0000-0003-1877-8462;
CORE (COnnecting REpositories)