We present an innovative approach to traffic incident modelling through the integration of large language models (LLMs) with established machine learning techniques. This method enhances the classification accuracy of traffic incident severity, leveraging the nuanced understanding of unstructured text provided by language models.