Safety risk management is one of the important guarantees to ensure the safety of railway operation. The rapid development of Intelligent Transporation System (ITS) genereates large amounts of Multi-source heterogeneous data. Using traditional identification methods to identify safety risk of railway can not full fill with the needs of ITS. Knowledge driven safety risk management becomes an important direction in the development of ITS. In this paper, we propose a novel method to identify safety risk from textual data. Inspired by the knowledge graph of natural language processing community, we identify safety risk by relation extraction mode. In order to reduce the severe dependence on manual feature engineering, we adopt convolutional neural networks to automatically capture semantics of textual data. Then a softmax classifier is used to recognize the relation types. The experimental results on real dataset show that, our method achieves considerable performance and provides a novel direction to intelligent safety risk management.