One thing is for sure. The most likely path to a succesful "chat bot" is the system I described above, which to make it clear I certainly did not think up myself. It is the path a large number of programmers have undertaken. The flaw as I percieve it is they are trying to get the chat bots to build up entire sentences and their definitions, as opposed to one word at a time (at least those I've seen). From what I've seen of how children learn language, I think the best way to go is to somehow provide an AI with some other context for judging a word and then simply using single words. To that end, other programmers are designing visual recognition systems which I think should be combined with a chat bot that works with single words and compares them to both memories and current input like that visual data.
Well, we do it in two ways... as infants, we have a limited (but growing) choice of words combined with points and gestures. As we learn more about how sentences are constructed, we eliminate the need for additional gestures. Proper grammar (not that I'm a reliable source on info) is, by nature, designed to pre-emptively answer questions.
Such as the sentence, "Bring me the towel on the counter." Consists of three independent thoughts strung into one action, "Bring me what? the towel where? on the counter."
So along with individual words and usage, you'll have to feed Alive the 3rd grade grammar concerning nouns, subject, predicate, different forms of verbs, etc. so that it can string several independent ideas into a single sentnece or action.