This paper introduces a semi-automated technique for classifying text-based close call reports from the GB railway industry. The classification schema uses natural language processing techniques to classify close call reports in accordance with the threat pathways shown on bow-tie diagrams. The method enables categorisation of a very large number of unstructured text documents where safety-related information has not previously been extracted due to the infeasibility of analysis by human readers. The results demonstrate mixed accuracy in the categorisation of close calls, with the highest accuracy being for the threat pathways that are more frequently reported. This work paves the way to machine-assisted analysis of text-based safety and risk databases, and provides a step forward in the introduction of data analytics in the safety and risk domain. Others working in this area have speculated that approaches such as this could be mandatory for safety management in the future.