<p>Urban rail transit systems are expanding rapidly, but the increasing frequency of operational accidents has posed significant challenges to public safety and emergency management. A key issue lies in the structural disconnect between predefined emergency response plans and the measures adopted in real accident scenarios, and the emergency plan text cannot support fast and accurate emergency decision-making. To address this challenge, this study proposes a novel knowledge graph construction framework that integrates multi-source information from emergency plans and accident cases, using the Shanghai urban rail transit system as a case study. 219 accident cases and multiple emergency plan documents were collected, and a Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory-Conditional Random Field model (BERT-BILSTM-CRF model) combined with large language model (LLM) was employed to automatically extract key entities and response measures. Furthermore, a semantic clustering and multi-dimensional mapping strategy was developed to establish fine-grained correspondences between plan-defined measures and accident-specific responses, enabling unified cross-source knowledge modeling. Experimental evaluations demonstrate that the proposed approach substantially improves the standardization and semantic consistency of response measures and achieves high-precision matching between planned strategies and real-world cases. The resulting knowledge graph significantly enhances information retrieval and reasoning capabilities in emergency response scenarios. This work provides a systematic, interpretable, and traceable framework for knowledge representation in urban rail transit safety management and establishes a solid methodological foundation for building intelligent emergency decision-support systems.</p>