Reproduce “Gmail Smart Compose: Real-Time Assisted Writing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mia Xu Chen, Benjamin N Lee, Gagan Bansal, Yuan Cao, Shuyuan Zhang, Justin Lu, Jackie Tsay, Yinan Wang and 4 more
Google developed Gmail Smart Compose, a system that generates real-time, context-dependent writing suggestions to assist users during email composition. The system leverages neural language models, optimized for low-latency inference on specialized hardware, leading to a reported saving of over one billion characters typed weekly.
Key insights
- Achieving real-time, low-latency neural language model inference at scale necessitated strategic model selection, prioritizing RNNs over higher-quality Transformers for production due to latency constraints, coupled with extensive hardware acceleration.
- Personalization through lightweight n-gram models significantly improved suggestion relevance and user engagement, demonstrating a 6% relative gain in Click-Through-Rate.
- A multilingual WordPiece model effectively supports diverse languages, especially benefiting low-resource languages by leveraging shared data, despite minor quality trade-offs for high-resource languages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.