In this presentation, we will discuss an application of deep learning to categorize open-ended comments from field interviewers. These comments contain instructions for back-end data technicians to execute manual edits to the data and the goal of the model is to simplify the identification of the correct action. We discuss the development, training, and performance of a model that combines BERT with hand-curated, non-text features. We will also compare the performance of this model with more traditional approaches using lasso.