I would say that with enough data the RNN (or more complex models for that matter) could learn a good approximation of the correct sentence associations. At the same time want to point out that from a practical point of view, many clients or use-cases are still reluctant to rely on the generative unpredictability of a RNN and prefer to rely on hard coded pieces of text. Make yes use of models to pick the best piece of text, but definitely don’t allow models to generate text on their own, which I believe is a real pity..