It was trained on 238M messages! This just goes on to show the strategic importance of data in modern machine learning. No wonder Google doesn't mind making the algos public: without the vast training data, they will not get you very far :(
I wonder what it would look like if instead of going after breadth with superficial responses, they restricted the system to one narrow domain, like scheduling and went on to generate deeper and more contextual responses.
Paper title: Smart Reply: Automated Response Suggestion for Email