DeText is a “Deep Text understanding framework” for natural language tasks such as ranking, classification, and language generation. Designed for use in search and recommendation systems among other use cases, It leverages neural components such as word embeddings, text encoding layers, an interaction layer, and an MLP layer to combine various wide and deep feature processing steps. The end-to-end model is trained to optimize user click probability in the final task. Since it is intended to be a general framework, it also provides flexibility in the choice of deep models, loss functions, and source and target fields.