PROVERB: The Probabilistic Cruciverbalist

From: Raphael Hoffmann (raphaelh_at_u.washington.edu)
Date: Sat Dec 06 2003 - 13:59:26 PST

  • Next message: Patrick Haluptzok: "Paper Review #6 - by Patrick Haluptzok - CSE 573 Fall 2003"

    "PROVERB: The Probabilistic Cruciverbalist"
    by Greg A. Keim, Noam M. Shazeer, Michael L. Littman et. al.

    [Summary:]
    PROVERB is a system that combines various data sources and AI techniques to
    solve crossword puzzles. The paper presents its probabilistic design and
    analyzes its performace.

    [Pros:]
    PROVERB has a two-stage architecture that allows the combination of several
    AI techniques (CSP, hillclimbing, ...) and makes the system very flexible.
    New data sources can easily be integrated as "Expert modules" and the value
    of each can be quickly determined by comparing the performance of using all
    modules vs. all-but-one modules.

    Additionally, applying probability theory to solving crossword puzzles is
    definetly a creative approach. Thinking about the problem, one might first
    try a simple and direct way of searching for solutions. However, I find that
    their idea resembles the way humans tackle this problem: experienced
    crossword puzzle solvers first check the words that are most likely to fit
    according to prior experiences. This leads to solutions a lot faster.

    [Cons:]
    Not surprisingly, the performance of PROVERB depends heavily on the CWDB
    database. The reason is that the same questions appear in crossword puzzles
    over and over. This fact makes the system somehow weak and limits its
    capabilities in solving new, unusual and hard cw puzzles.

    Furthermore, the authors used crossword puzzles from various sources, e.g.
    NYT, TV-Guide etc., and built a single system, PROVERB. However, as they
    stated themselves, the difficulty of the crossword puzzles depends on their
    source. I believe that the kind of questions is correlated with the source,
    e.g. TV-Guide crossword puzzles might more often ask about actors, stars
    etc., whereas the NYT crossword puzzles more often ask about Greek
    philosophers or Jewish customs. Therefore, it could be better not to train a
    single set of weights for the merging component of PROVERB.

    [Future ideas:]
    I am wondering what would happen if the authors made their software publicly
    available. Most probably, NYT editors would try to fool the program by
    exploiting some of its weaknesses: PROVERB can only be as good as its data
    sources. Although the architecture is flexible by allowing the integration
    of new data sources, that can be time-consuming and difficult. I believe a
    solution would be to use the web as a data source. However, unlike Google
    that simply searches for a site containing the specified words, we are now
    looking for words with a specified length and whose appearance close to our
    search string (question) is statistically significantly higher than overall.

    As stated above, it might make sense to set PROVERB's parameters according
    to the type of crossword puzzle. A classification of the puzzle bevore
    trying to solve it could be advantageous.


  • Next message: Patrick Haluptzok: "Paper Review #6 - by Patrick Haluptzok - CSE 573 Fall 2003"

    This archive was generated by hypermail 2.1.6 : Sat Dec 06 2003 - 13:59:28 PST