I'm not sure what methods they use, but the "single sentence reply suggestions" created by Google's Inbox are the highest quality natural language generation that I have come across.
Looks less like NLG, and more like picking existing responses from a (probably huge) corpus using ML. Hard to replicate unless you have access to the kind of data Google has.
If you think about it, all natural language generation is simply picking from a list of possible words to go next. They just do that at a sentence level.
Heh. I suppose, if you have enough sentences to choose from for the various combinations of words, tenses, conjugations, prepositions, and so forth. Can't imagine there's many entities that would have that much data.
