From: Ankur Jain (ankur@cs.washington.edu)
Date: Mon Dec 06 2004 - 12:16:27 PST
PROVERB: The Probabilistic Cruciverbalist by Kein et al.
* One-line summary
The paper presents PROVERB, a system the authors built, that can solve
crossword puzzles with reasonable accuracy and peformance.
* The (two) most important ideas in the paper, and why
One thing that I liked about this work was how the authors identified a
previously (largely) unexplored domain and formulated it as an AI problem.
Moreover the way they broke down a seemingly very challenging problem into
a set of several more-manageable problems was also nice. Another thing
that I liked about the paper was how the authors came up with many
different modules each trying to solve the same problem and yet using very
different approaches. While there are possibly better undiscovered
solutions still out there, the bottom-line is that here is a system that
solves a seemingly very difficult problem with reasonable accuracy and
efficiency!
* The one or two largest flaws in the paper
My only concern with paper was that at places, it didn't look well fleshed
out. Specifically, the authors do not discuss why they took some of the
decisions that ended up taking. For instance, the particular weighting
texhniques that they use to merge candidate lists, or, using expected
overlap with creater's solution ato choose a candidate, etc.
I would consider this as only a minor concern, because being probably the
first ones to take a reasonable cut at the problem, the authors probably
had a lot of other things to discuss in the paper.
* Identify two important, open research questions on the topic, and why they matt
After reading the paper, I was left wondering how many of the techniques such as
Dijkstra modules, LSI, etc could be (most probably, *are being*) used in the Internet
search domain which also has to deal a lot of with "relatedness" of words.
One noteworthy take-away from the paper, which the authors also alluded to, is how faster
processors and larger memory/storage is making it possible to tackle problems that were
previously not tractable. We saw a similar example earlier in the course when somewhere
in the mid-90's SAT-solving became possible. Technological improvisations it seems will
not only help us conquer more domains, but also help drive us to develop newer (AI)
techniques that best leverage the hardware.
This archive was generated by hypermail 2.1.6 : Mon Dec 06 2004 - 12:16:27 PST