Dec. 6, 2020

In this talk, we discuss the deep learning–based natural language understanding (NLU) used by Eno, Capital One's intelligent assistant that helps customers manage their money by texting in a conversational way. We discuss the shortcomings of existing options for NLU and why we built our own solution. We present our neural network architecture for NLU. First, we trained word- and character-level word embeddings on web chat conversations. Then we trained an NLU model using labeled data. The result is a model that's robust to common misspellings, abbreviations, and new words. Our model is trained, deployed, and continuously improved entirely in the cloud.