Smart Reply: Learning a Model of Conversation from Data
summary
Smart Reply generates one-tap responses for Inbox by Gmail by learning directly from data: a sequence-to-sequence recurrent neural network encodes incoming email and decodes replies, trained end-to-end on historical email–reply pairs. Inference uses beam search and production limits responses to a vetted set; semantic clustering and post-processing enforce diversity and safety. Deployed results show real use—assisting in over 10% of mobile replies—but the model has no user-specific persona or external state and can only mimic knowledge from its training data.
AI-written summary · about
chapters (11)
- 0:00Introduction to Google Brain
- 0:44The smart reply concept
- 2:39Challenges of dialog modeling
- 4:12A data-driven approach
- 5:39Technical architecture
- 9:03Inference and training
- 12:20Learning from data
- 14:34Production and limitations
- 17:02Improving user experience
- 21:46Model limitations
- 22:46Conclusion
related talks
32mDeepCare Chatbot - Generating answers to customers using Deep Learning and NLP
29mEngineering Machine Learning Technologies
10mMachine Learning: The next great leap
41mConversational Agents at Scale: Retrieval and Generative approaches
48mDIY Personalization: How, when why to offer self-made AI
39mCustomize Your App with MLKit
30mConversational AI: Building clever chatbots
28mByron Galbraith, Chief Data Scientist, Talla at MLconf Seattle 2017