Watson and Jeopardy

Lennart Sorensen lsorense-1wCw9BSqJbv44Nm34jS7GywD8/FfD2ys at public.gmane.org
Thu Mar 3 14:55:23 UTC 2011


On Wed, Mar 02, 2011 at 11:31:51PM -0500, marthter wrote:
> Yeah, I only caught about 10 minutes of the middle episode (including  
> the Final Jeopardy question), and I really found the pummelling of the  
> humans pretty surreal.  Like this was the coming out party of our new,  
> friendly-swirly-lines-around-the-big-cyclops-eye computer overlord.  I  
> kept expecting a Saturday Night Live sketch to break out and Sean  
> Connery to pop up in the middle spot.
>
> Although the computer's performance was very impressive, I also  
> sometimes doubted how advanced it really is.  A lot of earlier-round  
> Jeopardy questions, you barely have to know, "Degas' painting blah blah  
> was stolen in 19xx from this French capital".  You don't need to know  
> anything about the artist, the painting, the date of the theft, ... do  
> you "understand" the question, or do you basically keyword search your  
> archive of articles/knowledge, for cities mentioned in the same  
> sentence/paragraph/article as "French" and "capital".
>
> Now certainly there were other more challenging questions on which it  
> did quite well but it seemed to me the more indirect the question, the  
> worse it did, such as the final Jeopardy city-with-two-airports question.

I think the problem with that question was that there probably isn't
really any documents out there that lists every city and the names of
all its airports.

Watson only "knows" whatever is in the data it was given to work with.

Also the developers did say after the fact that Watson had unfortunately
learned over the course of many practice games, that often the category
wasn't that critical to the answer, so for it to ignore the US part
of the category was actually quite reasonable, although either way,
Watson was perfectly aware that its answer was almost certainly wrong,
which is very useful.  At least it means if you ask it a question, it
won't always give you an answer, sometimes it will say "it might be this,
but I am very unsure".

The fact it got the final jeopardy right in the second game was very
impressive and shows just how well it does with indirect references.

-- 
Len Sorensen
--
The Toronto Linux Users Group.      Meetings: http://gtalug.org/
TLUG requests: Linux topics, No HTML, wrap text below 80 columns
How to UNSUBSCRIBE: http://gtalug.org/wiki/Mailing_lists





More information about the Legacy mailing list