<p>This article explores the application of Latent Dirichlet Allocation (LDA) to structured tabular insurance data. LDA is a probabilistic approach initially developed in Natural Language Processing (NLP) to uncover the underlying structure of (unstructured) textual data. It was designed to represent textual documents as a mixture of latent topics, and topics as mixtures of words. This study introduces LDA’s document-topic distribution as a soft clustering tool for unsupervised learning tasks in actuarial science. By defining each topic as a risk profile, each insurance policy as a document, and each modality of categorical covariates as a word, we extend LDA’s application beyond textual data. Our experimental results and analysis highlight how the modelling of policies based on topic cluster membership, and the identification of dominant modalities within each risk profile, can give insights into the prominent risk factors contributing to higher or lower claim frequencies.</p>