Theiling Online    Sitemap    Conlang Mailing List HQ   

Re: Jabberwocky/AI Language Learning

From:H. S. Teoh <hsteoh@...>
Date:Sunday, March 18, 2007, 1:18
On Wed, Mar 14, 2007 at 12:03:37PM -0700, David J. Peterson wrote:
> I ran across this website the other day: > > > > It hosts a "learning" AI chatbot that modifies its responses based > on a growing list of example sentences from which it draws > generalizations. (I don't know exactly how it works, but it has > to do with identifying keywords in entered sentences and gauging > what a likely response would be.)
[...] Actually, what it does is to keep a (huge) database of previous responses users have fed it, and uses some kind of selection algorithm that infers based on context which response is the "best" one. It never actually constructs a sentence on its own; all its responses are previous user input. (This may not be immediately obvious if you chat with it in English, as it has accumulated an incredible amount of responses that would make it seem realistic. But if you try chatting with it in, say, Russian, for which it presumably hasn't collected that many responses yet, you'll quickly see that many of the responses are just parroting what your typical average internet user would type in a chat box: "go to hell", "shut up", "are you male or female", "I am Brezhnev!", "you are the doorknob of a public toilet", and the like. Complete with typos and all.) If you want to teach it your conlang, it's very easy. Just start talking in your conlang, and whenever it gives a "wrong" response (like reply in English), hit the "correct me" button and type in an appropriate response in your lang. After a while (and I do mean a *while*... like hours, or days, or more), it will have accumulated enough responses that it could more-or-less maintain a short conversation in your conlang. Initially, once you've corrected it enough times, you will start seeing it parrot your previous responses back to you. I guess when it has accumulated enough responses, it will have the appearance of carrying out a realistic conversation, although in reality it's really just repeating what you've said to it before. But if you're hoping it will somehow suddenly clue in to your language and start making sentences on its own, well... sorry, it ain't gonna happen. T -- INTEL = Only half of "intelligence".


David J. Peterson <dedalvs@...>
Leon Lin <leon_math@...>