I was very impressed by a recent paper from a team at Facebook about a production-ready end-to-end neural NLG system. Especially interesting to me was the “engineering” approach to key issues such as accuracy, data collection, and latency.