[Top] [All Lists]

Re: [ontolog-forum] IBM Watson on Jeopardy

To: ontolog-forum@xxxxxxxxxxxxxxxx
From: "John F. Sowa" <sowa@xxxxxxxxxxx>
Date: Sat, 12 Feb 2011 15:23:08 -0500
Message-id: <4D56EC2C.4030303@xxxxxxxxxxx>
On 2/12/2011 9:53 AM, Anders Tell wrote:
> By training the system too hard, the system becomes very good
> at explaining the past but poor at predicting the future.    (01)

That is also true of people.  As the saying goes, generals are
very good at fighting the last war, but not the current one.    (02)

> Is the Watson system good at visiting the future?    (03)

I was not involved in the development, but I know some of
the people who were, and I read some of the publications
about it.  My impression is that Watson (a) uses state-of-
the-art NLP software, (b) adds UIMA software to manage
large volumes of resources, (c) uses statistics and other
methods of machine learning, (d) generates large numbers
of hypotheses (abduction) about possible interpretations,
(e) uses deduction to determine the implications of each
hypothesis, (f) does a "reality check" to test each of the
implications, (g) uses highly specialized heuristics for
answering jeopardy questions, and (h) uses a supercomputer
to analyze a huge number of alternatives in parallel.    (04)

Everything except step (g) is general, and steps (c)
through (f) are basically Peirce's Cycle of Pragmatism,
which I have presented in various talks and papers for
over 20 years.  Following is a one-slide summary:    (05)

    http://www.jfsowa.com/figs/soup5a.gif    (06)

Jeopardy questions span a very wide range of topics.
(That's general.)  But they are often stated in tricky
ways that depend on untangling jokes, metaphors, and
wordplay. (That has general and specialized aspects.)    (07)

As I said in my previous note, I recommend listening
to the 2006 talk by David Ferrucci, reading a little bit
about the UIMA software (which IBM donated to open source,
under the care of the Apache Foundation), and then listening
to Ferrucci's recent talk at the IBM Yorktown auditorium.
Both talks get into more detail than the TV documentary.
(See below for a copy of the URLs from my previous notes.)    (08)

Your question about accurately predicting the future is one
that everybody would like to answer.  My guess is that the
Cycle of Pragmatism (soup5a.gif) is the best approach, and
improvements would depend on better methods of guessing
(abduction), belief revision, and testing.  I would add
that good methods of analogy finding could improve all
those steps.  That's the theme of the following talk:    (09)

    http://www.jfsowa.com/talks/pursue.pdf    (010)

As for the specific question about Watson: Since Jeopardy
questions never ask about the future, I doubt that the
developers spent any effort on that aspect.    (011)

____________________________________________________________    (012)

Following are Ferrucci's slides about UIMA from 2006:    (013)

http://ontolog.cim3.net/file/resource/presentation/DavidFerrucci_20060511/UIMA-SemanticWeb--DavidFerrucci_20060511.pdf    (014)

Following is the talk that explains those slides:    (015)

http://ontolog.cim3.net/file/resource/presentation/DavidFerrucci_20060511/UIMA-SemanticWeb--DavidFerrucci_20060511_Recording-2914992-460237.mp3    (016)

And following is his recent talk about the DeepQA project for
building and extending that foundation for the Watson system:    (017)

http://www-943.ibm.com/innovation/us/watson/watson-for-a-smarter-planet/building-a-jeopardy-champion/how-watson-works.html    (018)

Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/  
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/  
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/ 
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J
To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx    (019)

<Prev in Thread] Current Thread [Next in Thread>