From incident reports to Bayesian inference: NLP driven causal modeling for port occupational safety

Port operators conduct cargo‑handling activities using heavy machinery, exposing workers to substantial safety risks. National regulations require that safety accidents be reported to authorities, creating large collections of narrative incident records that remain underutilized for systematic risk analysis. This study analyzes 3,131 real‑world accident reports published in Korea related to loading, unloading, stacking, transportation, and tool‑handling tasks. A transformer‑based natural language model (i.e. Sentence‑RoBERTa) is employed to extract causal factors from unstructured narratives and construct a consolidated causality graph. This graph is subsequently converted into a Bayesian network to quantify the marginal effects of unsafe behaviors on accident outcomes. Simulated results reveal distinct risk patterns: low‑probability events such as tools mishandling and unloading show nonlinear effects, with a 10% increase leading to nearly an 8% rise in critical‑injury likelihood, while high‑frequency events such as stacking and shipping are more closely associated with severe outcomes. The proposed framework demonstrates how qualitative safety narratives can be transformed into a probabilistic risk model, enabling transparent, data‑driven assessment and supporting targeted interventions in training, operational controls, and human‑factor management.