I edited my response... I hope it helps... my understanding is that the output gives probabilities for all the words, then one is chosen with some random thrown in (via the #temperature) then fed back in... which to me seems to equate to bag of words. Perhaps I mis-understood the term.