A corpus-based quantitative risk assessment of language barriers in maritime safety

In maritime transportation, language and communication barriers pose substantial risks. Despite numerous accident reports identifying communication issues and large datasets available through Natural Language Processing (NLP) and Large Language Models (LLMs) enabling detailed investigation, corpus-based quantitative research on this topic remains limited. This study addresses this gap by employing CustomGPTs tailored to extract structured data from a corpus of 435 maritime accident reports between 2012 and 2022. These reports focus on communication issues, human factors, operational contexts, and organizational variables. To achieve precise probability estimation, a Bayesian Network (BN) model based on expert judgment was applied alongside a data-driven Expectation-Maximization (EM) algorithm. Predictive analysis quantified the association between communication barriers and accident probabilities, suggesting a 72% likelihood of severe accidents occurring due to communication challenges. Diagnostic, sequential, and simultaneous analyses revealed that communication barriers and human factors substantially increased accident likelihood, elevating the risk to 93%. Sensitivity analysis identified cognitive load, stress, and language proficiency as key risk drivers, underlining the need for improved communication protocols. Based on these findings, the study recommends implementing tailored training programs to improve personnel language proficiency and standardize communication practices to reduce risks and enhance maritime safety.