Abstract: To improve the intelligent management and control capability of railway safety risks, this paper proposed a scenario inference-based similarity calculation method for railway emergency disposal. It used the multi-dimensional scenario space model architecture to conduct knowledge representation of railway emergency scenarios, combined deep learning and natural language processing technologies to construct a Sentence-BERT (SBERT) textual semantic similarity matching model for railway accident report texts, adopted the Token Attention mechanism to dynamically capture and identify internal keywords in text sentences, completed feature extraction, semantic analysis, classification induction, and information retrieval based on railway accident reports, and further achieved high-precision railway accident scenario matching. This method lays a solid foundation for the construction of an emergency case scenario library for railway emergencies.