ABSTRACT Chat reference service has been used in academic libraries to more efficiently serve patrons in the digital age. Identifying question topics on chat can help librarians understand patrons' needs and improve reference services. Researchers have used qualitative methods to understand question types in chat records; however, these methods are inefficient to analyze large chat datasets. Here, we conducted a novel research using Latent Dirichlet Allocation (LDA) topic modeling to automatically extract topics from chat transcripts generated in 5 years from a large university library. With little human intervention, the model identified major topics based on statistical distributions of terms‐document relationships in chat transcripts. We also applied VOSviewer to analyze the same dataset and found consistent results. From these results, we found that the most prominent chat topics were about accessing various library resources. This finding can help libraries allocate resources, design educational materials, and provide trainings for future librarians.