Risk analyses require proper consideration and quantification of the interaction between humans, organization, and technology in high-hazard industries. Quantitative human reliability analysis approaches require the estimation of human error probabilities (HEPs), often obtained from human performance data on different tasks in specific contexts (also known as performance shaping factors (PSFs)). Data on human errors are often collected from simulated scenarios, near-misses report systems, and experts with operational knowledge. However, these techniques usually miss the realistic context where human errors occur. The present research proposes a realistic and innovative approach for estimating HEPs using data from major accident investigation reports. The approach is based on Bayesian Networks used to model the relationship between performance shaping factors and human errors. The proposed methodology allows minimizing the expert judgment of HEPs, by using a strategy that is able to accommodate the possibility of having no information to represent some conditional dependencies within some variables. Therefore, the approach increases the transparency about the uncertainties of the human error probability estimations. The approach also allows identifying the most influential performance shaping factors, supporting assessors to recommend improvements or extra controls in risk assessments. Formal verification and validation processes are also presented.