The application of hazard analysis techniques is essential for identifying failures within safety-critical systems that impact both human lives and property, particularly in disciplines such as aeronautical engineering. System-Theoretic Process Analysis (STPA) is notable for its effectiveness in detecting hazards in complex systems, however, it is an extensive technique that demands time and manual effort. As STPA is mostly a text-based analysis, it could benefit from the capabilities of language models. This study introduces BEDS (BERT Error Detection for STPA), a machine learning pipeline leveraging the well-established Bidirectional Encoder Representations from Transformers (BERT) language model to support the initial phase of STPA analysis. The integration of BERT delivers significant advantages by facilitating high-accuracy classification of core STPA categories - System Losses, Hazards, and Constraints - through its semantic understanding capabilities. In addition, BERT is employed to verify compliance with technique guidelines, identify common sentence formulation errors, and recommend expert-validated alternatives. This methodology is designed to enhance the quality and consistency of the descriptions used in safety assessments. The work further contributes a specialized dataset of STPA sentences, providing a foundation for continued research and tailored Natural Language Processing model development for STPA applications. Experimental results substantiate the efficacy of the proposed approach, yielding an average accuracy of 95.20% in STPA sentence classification, 88.51% in sentence validation, and 83.44% in fault detection. These findings underscore the value of transformer-based models as effective tools in the development of complex systems, offering automated support for requirements verification and mitigation of human error during safety analyses.