Re: Jabberwocky/AI Language Learning

Teoh wrote:
<<
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.
 >>

Oh...  This is disappointing.  See, this is what I thought it could do:

AI: What do you do?
User A: I'm a teacher.

AI: What do you do?
User B: I'm a librarian.

AI: What did you eat today?
User C: I ate a sandwich.

So, if it got that input, I was thinking it'd catalog all the sentences,
and also all the words, and all sequences of two words, and all
sequences of three words, etc.  It could then see that these were
licit two word sequences:

(1) a teacher
(2) a librarian
(3) a sandwich

It could also see that this is a licit three word sequence:

(4) ate a sandwich

Therefore it could, by analogy, predict that the following would
also be licit, and produce them:

(5) ate a teacher
(6) ate a librarian

Then it might prompt the user like this:

AI: Did you eat a librarian today?

Of course, this would be nonsense, so you could correct it and
provide it with some sort of correct noun, and then it would try
to form a generalization (given keywords x, y and z, "librarian"
is not a word that will produce an ordinary response).

Anyway, if it could build up generalizations like this, then it
could figure out what the various classes of words in a created
language were.  It wouldn't label them, or know that nouns
were nouns, or verbs were verbs, but it would know which
words could go after "a" and which couldn't in which circumstances.

So that's what I thought it was doing.  If it's just building up a
database and trying to match what's in its database with what
a user says, then, yes, that isn't very interesting.  As you said,
after a long while, it could hold a conversation with you, but it
wouldn't show you anything you didn't already know.  Maybe
contentwise, but nothing having to do with structure.  Oh well.

-David
*******************************************************************
"sunly eleSkarez ygralleryf ydZZixelje je ox2mejze."
"No eternal reward will forgive us now for wasting the dawn."

-Jim Morrison

http://dedalvs.free.fr/