Learning Bayesian Belief Networks from Aviation Accidents with a Generative Pre-Trained Transformer Model

Unstructured aviation accident narratives are difficult to convert into causal, decision-oriented models. This research presents and validates a generative pre-trained transformer (GPT)-powered automated analysis framework for aviation safety risk. GPT models systematically analysed 23,620 unstructured accident reports from the Aviation Safety Network (ASN) using a templated prompt to extract cause events, identifying 12 effect events, 33 cause events, and their probabilistic dependencies. A Bayesian belief network (BBN) was developed through a simulated annealing structure search guided by the Bayesian Dirichlet equivalent uniform (BDeu) score, and its parameters were estimated through maximum likelihood estimation. The model was validated through five-fold cross-validation, achieving an overall accuracy of 86.58%. Tornado-based sensitivity analysis and composite risk ranking identified aircraft design defect, crew disregard of regulations, dust haze, hydraulic failure, and thunderstorms as the five most critical cause events, forming the basis for targeted risk-control recommendations. This research contributes a GPT-driven extraction and mapping workflow that turns narratives into cause events, and is deployed as a BBN for aviation risk inference on 23,620 ASN accidents with recommendations. This research shows that the GPT-powered extraction process supports systematic risk analysis; on an expert-adjudicated benchmark, it achieved higher recall, micro-F1, and macro-F1 than keyword and TF–IDF baselines, while the BBN translated the extracted events into risk-control priorities.