This paper describes the continued development of natural language processing (NLP) techniques to support safety management on the GB railways. The work considers machine reading techniques to obtain information from more than 800,000 free text hazard records in the Close Call System (CCS) Our research has found that the non-standard nature of the text means that standard NLP techniques yield accuracies between 0% and 60% when categorising hazard reports. To improve accuracy, a workflow was developed that uses the results from individual techniques in an iterative cycle with a human analyst. The technique, dubbed interactive learning, has achieved substantially improved accuracy of up to 98% and is currently being integrated by the GB railway industry as part of its SMS.