On 4/27/07, Matt Mahoney <[EMAIL PROTECTED]> wrote:
I think learning in layers (A) is the correct approach, but also that it
can
be done from a corpus of adult level language, at least if you are
training a
pure, ungrounded language model. When parents use baby talk, they are
actually using fragments of adult level language, like holding a ball and
saying "ball". In an ungrounded model (no vision, robotics, etc), the
model
only needs to know that "ball" is likely to appear in immediate contexts
like
"a" or "the" and more distant contexts like "play" or "catch". It doesn't
need to know what these words mean in the same sense that we do.
Parents might use baby talk like "goo goo ga ga" but I don't think that
helps
children learn language. It might help them learn to associate their own
utterances with what they hear, but you don't need that type of training
for a
statistical model trained on text.
I know you can't completely solve any of the AI problems like language,
vision, hearing, or robotics without solving all of them, but I do think
it is
possible to induce ungrounded semantics (word associations) and grammar in
a
statistical model without having to train on children's books.
"Layered learning" is not just better, it's actually the only
computationally feasible approach.
We may talk to a baby like:
"MILK?"
"You want to play BALL?"
"Oh you POO-POO again" etc.
And these things are said simultaneously as some *physical* events (eg milk,
ball, poo) are happening, which allows the baby to correctly *bind* the
words to concepts, ie achieve grounding.
Contrast this with something from Wall Street Journal:
Headline: "Employees of a new plan to get Dell back on the road to growth,
including streamlining management and looking at new methods of distribution
beyond the computer company's direct-selling model."
Can a baby really learn from THIS ^^^ ?
YKY
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
