+DS IPLE: Case study: Natural Language Processing for Mobile Devices

Mobile devices have become a vital platform for deploying state-of-the-art natural language processing (NLP) techniques, due to their ubiquity and portable nature. This presentation will discuss a case study for how deep NLP models are employed to greatly improve mobile users' productivity while replying to emails on the go. This feature, called Smart Reply, has been actively used on both Gmail and Outlook since being launched. Specifically, the architectural design of this NLP system will be discussed, including the concrete engineering challenges to address in practice. Further, motivated by the rising concern about user/data privacy, a federated learning paradigm is leveraged to serve billions of users without data leaving their devices. Considering the low-resource nature of mobile devices, a special type of word embeddings is introduced to meet the computation and storage requirements. Potential extensions to more application scenarios will also be discussed for this technology.

This session is part of the Duke+Data Science (+DS) program in-person learning experiences (IPLEs). To learn more, please visit https://plus.datascience.duke.edu/