From: Tyler Robison (trobison_at_cs.washington.edu)
Date: Mon Dec 08 2003 - 09:52:39 PST
PROVERB: The Probabilistic Cruciverbalist
Greg A Keim, Noam M Shazeer, Michael L Littman
Summary:
The authors present a system that solves cross-word puzzles by
having 'expert modules' for various knowledge domains guess words based on
given clues.
Important Ideas:
The most interesting part of this in my mind is trying to match
words to the clues given, a problem which seems like it would be very
difficult for a computer, but with which they have had significant
success. To do this, they use several expert modules which return lists
of possibilities along with weights indicating the likelihood.
One thing that struck me is the thorough way in which they
evaluated the system. The problems they tried to solve weren't conceived
for the purpose of the experiment; they were real cross-word puzzles, and
so it is easy to compare the performance of the system against actual
people. In addition, they performed tests after removing certain modules
to see the dependence of the system on them. Overall, it seemed like they
did a good job of evaluating the system.
Flaws:
I did like this paper, but one flaw stands out nonetheless: it is
an application of various techniques, and not really a general tool
itself; it doesn't appear to have much use outside of cross-word puzzles.
Despite this, it does show-off the potential of AI by being very
successful at a problem that seems very 'human' in nature, and so seems
worthwhile.
Research Questions:
Since much of the problem involves figuring out clues, it could be
interesting seeing similar techniques applied other problems that require
guessing the answer to some clue, such as riddles, for example. Of
course, in the cross-word puzzle problem, they have the benefit of knowing
the length (and possibly some letters) of the answer, so this may not work
at all.
One could also think of expanding & improving the technique to the
point that the program is on par with human masters. Since this isn't a
well researched area, there is certainly room for improvement.
This archive was generated by hypermail 2.1.6 : Mon Dec 08 2003 - 09:52:39 PST