Here is my (non-automated :-) summary of Stuart Russell's talk that I
attended today at CMU. No need to say that the talk was VERY VERY
interesting and thought-provoking... The overall problem was that of
counting how many unique objects exist in the given (toy or real) world,
and the challenge is that we can have multiple objects that satisfy same
description (name) while same object could have multiple descriptions.
For this, he used one toy example (of counting #balls in the urn) and one
real-world example of citation matching (finding multiple citations that
refer to the same publication).
In the toy example, you are given an urn that contains balls. You may see
balls with different colors or there could be multiple balls with the same
color. And, you are supposed to determine (of course automatically) the
actual number of balls in the urn. e.g. if you see one red ball, and one
blue ball, you know there has to be at least 2 balls in the urn, but there
may be 2 red balls and 2 blue balls or there could be just one red ball
and one blue ball (if you see equal number of red and blue balls in your
large #sample trials). For this, he uses a probabilistic bayesian
approach, and plots probabilities (y-axis) that the urn contains some number
of balls (x-axis).
In the real-world problem, he showed the results of searching citeseer for
a query "Russell and Norvig" (hit Citations and not the Documents button).
Here is what I got when I did the same search --
--------------
Context Doc 619 (10): Russell S, Norvig P. Artifical intelligence a
modern approach. Upper Saddle River, NJ: Prentice Hall; 1995.
Context Doc 166 (1): RUSSELL S., NORVIG P.: Artificial
Intelligence: A Modern Approach. Prentice Hall, 1994.
Context Doc 43 (0): Stuart J. Russell and Peter Norvig, editors.
Artificial Intelligence: A Modern Approach. Prentice Hall, 2003.
Context Doc 23 (0): Russell, S. J. and Norvig, P. (1995). Arti
cial Intelligence: A Modern Approach. Prentice-Hall International, Inc.
Context Doc 19 (0): S. J. Russell and P. Norvig. Artificial
Intelligence. A Modern Approach. AI. Prentice Hall, Englewood Cli#s, 1995.
Context Doc 18 (0): Russell and Norvig, Ai: A modern approach,
Prentice Hall, 1995.
Context Doc 11 (1): Peter Norvig and Stuart Russell. Arti cial
Intelligence: A Modern Approach. Prentice-Hall, 1995.
Context Doc 6 (1): P. Norvig and S. Russell. Artificial
Intelligence. A Modern Approach. Prentice Hall Series in Artificial
Intelligence, 2003.
Context Doc 4 (0): S. J. Russell and P. Norvig. Arti cial
Intelligence. A Modern Approach. Prentice-Hall, Englewood Cli s, NJ, 1995.
Context Doc 4 (0): S. Russell and P. Norvig. Introduction to
Artificial Intelligence. Prentice Hall, 1995.
Context Doc 3 (0): Russell, S., and Norvig, P. Artificial
Intelligence A Modem Approach. Prentice Hall, 74, 1995.
Context Doc 3 (0): Stuart Russell and Peter Norvig. Arti cial
Intelligence. Prentice-Hall, 1995.
-------------
As you can see the same book is listed multiple times, because the author
names, titles, publisher are spelled differently, or the book had different
years/editions etc. Ideally, it is supposed to only return only one
citation, as there is just one book that Russell and Norvig wrote, and its
just one same object. Also, it should automatically detect that names such
as "Russel" etc are spelling errors, or "Stuart Russell" is same as
"S.J. Russell" etc.
He also gave some interesting examples like --
"Wauchope, K. Eucalyptus: Integrating Natural Language Input with
Graphical User Interface..."
where its not clear whether "Eucalyptus" is part of the paper title, or
its part of the author's name :-)
But, the same publication when cited as --
"Kennith Wauchope (1994). Eucalyptus: Integrating Natural Language..."
is no more ambiguous!
Thus, we want to use object representations that are minimally ambiguous,
and maximally correct (system should prefer Russell over Russel) etc.
Their system BLOG (stands for Bayesian Logic) uses probabilistic models
for First Order Logical representations of objects, their relations (authors
related to papers etc). And solves the problem of counting unique objects
as that of maximizing probabilities... The interesting property of this
world is that, object descriptions could be corrupt (such as spelling
errors) and ambiguous (multiple styles of citations).
His results show that their probabilistic first-order logic model achieves
higher accuracy (lower error rate) over the current citeseer search
engine.
Those who attended this talk, please feel free to correct/enhance my
summary, if I said something incorrectly or missed any important point.
Enjoy,
Amruta
___________________________________________________________________
Amruta Purandare [EMAIL PROTECTED]
Intelligent Systems Program http://www.cs.pitt.edu/~amruta
University of Pittsburgh (412)-657-1318
___________________________________________________________________
------------------------ Yahoo! Groups Sponsor --------------------~-->
DonorsChoose. A simple way to provide underprivileged children resources
often lacking in public schools. Fund a student project in NYC/NC today!
http://us.click.yahoo.com/KZzaMD/.WnJAA/HwKMAA/x3XolB/TM
--------------------------------------------------------------------~->
Yahoo! Groups Links
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/nlpatumd/
<*> To unsubscribe from this group, send an email to:
[EMAIL PROTECTED]
<*> Your use of Yahoo! Groups is subject to:
http://docs.yahoo.com/info/terms/
