From: Masaharu Kobashi (mkbsh_at_cs.washington.edu)
Date: Sun Dec 07 2003 - 23:39:03 PST
Title: PROVERB: The Probabilistic Cruciverbalist
Authors: Greg A. Keim, Noam M. Shazeer, Michael L. Littman et. al.
[Summary]
The paper reports the mechanism and its superb results
of a new cross-word solver, PROVERB, which incorporates
multiple AI related techniques, such as constraint
satisfaction, state-space search, natural language processing,
machine learning and probabilistic decision making.
[Most Important Ideas]
It is the way to integrate the candidates generated by as many as
30 expert modules. For the effective merger of the results from
those modules it uses variety of methods such as the probabilistic
objective function, the scoring system used in human tournament,
their own grid-filling algorithm which they claim very effective,
although the algorithm itself is not explained at all.
Second, the selection of the expert modules and the use of
probabilistic scheme in generating candidates by those modules
are important ideas, although detailed mechanism is not well described
in the paper.
[Largest Flaws]
First, there is no analysis of causes and results. In other words
they simply presents the results of their system and have
not investigated the effect of each method and strategy on
the performance of the whole system.
Second, while they use substantial portion of the paper for
describing the expert modules, they do not explain enough
about the key mechanisms such as the object functions, the
scoring system, weight-assigning rules of each module, and
the details of the implicit distribution modules which are
the keys to the use of effective probabilistic strategies.
[Important Open Research Questions]
More effective search strategies and more effective methods of
merging different knowledge are still open questions, since the
current success rate is far from perfect and far below human
experts' level.
This archive was generated by hypermail 2.1.6 : Sun Dec 07 2003 - 23:39:03 PST