Tabular data, from public opinion surveys to records of interactions with social services, is foundational to the social sciences. One application of such data is to fit supervised learning models in order to predict consequential outcomes, for example: whether a family is likely to be evicted, whether a student will graduate from high school or is at risk of dropping out, and whether a voter will turn out in an upcoming election. While supervised learning has seen drastic improvements in performance with advancements in deep learning technology, these gains are largely lost on tabular data which poses unique difficulties for deep learning frameworks. We propose a technique for transforming tabular data to text data and demonstrate the extent to which this technique can improve the performance of deep learning models for tabular data. Overall, we find modest gains (1.5% on average). Interestingly, we find that these gains do not depend on using large language models to generate text.