[ExI] Watson On Jeopardy

[BillK]

On Tue, Feb 15, 2011 at 7:56 PM, David Lubkin  wrote:
> You don't seem to have read what I wrote. The only question I raised about
> Watson's current capabilities was whether it had a module to analyze its
> failures and hone itself. *That* has been possible in software for several
> decades.
>
> (I've worked in pertinent technologies since the late 70's.)
>
>


I think the answer is Yes and No.

No, because Watson doesn't have time to do any learning or
optimisation while the game is actually in progress. Watson doesn't
take any notice of opponents answers. That's why it gave the same
wrong answer as an opponent had already given.


Yes, because it does do learning and optimisation. The programmers
'trained' Watson by asking many Jeopardy questions during training.
Quote:
The team has developed technology based on the latest results of the
statistical learning theory (e.g. kernel methods) applied to natural
language understanding. This has already increased Watson's ability to
learn from the questions it is asked (e.g. automatic Jeopardy cue
classification). Learning to handle the uncertainty in the selection
of the best answer (e.g. ranking the answer list) from those found by
Watson's search algorithms also has been one of their main research
directions.
-------------------------

BillK