4 Votes | Average: 4.25 out of 54 Votes | Average: 4.25 out of 54 Votes | Average: 4.25 out of 54 Votes | Average: 4.25 out of 54 Votes | Average: 4.25 out of 5 (4 votes, average: 4.25 out of 5) Loading ... Loading ...

by Mike Zazaian September 4, 2006 - 3:45pm, 1 Comment

WebCrow Logo

A new software program called WebCrow, capable of solving complex word puzzles, is a important step toward realizing Artificial intelligence.A unique competition at the European Conference on Artificial Intelligence in Riva del Garda, Italy pitted 75 bi-lingual human competitiors against two versions of a revolutionary crossword-solving software program dubbed “WebCrow.” Each competitor, human and software alike, were given 90 minutes to complete that day’s editions of both the New York Times and Washington Post crosswords in English, as well as a puzzle in Italian from the Roman newspaper, La Repubblica, and a puzzle from another local Roman newspaper.

Interestingly, the two versions of Webcrow finished in 1st and 2nd place for the American Puzzles, but managed only 21st and 25th on the Italian Puzzles. WebCrow creator Marco Ernandes noted the poorer performance by WebCrow on the Italian crosswords was likely due to the fact that, “The writer of one of those puzzles is well known for having lots of puns and political clues.” Regardless of Webcrow’s showing in the Italian puzzles, it’s amazing that the software’s intelligence was able to reason out such puns and political messages at all.

In order to solve a puzzle, WebCrow uses four seperate approaches on each given clue:

  1. Searching a database of previously solved crosswords.
  2. Consulting a dictionary.
  3. Using proven rules for a specific kind of two-letter Italian clue.
  4. Searching the internet.

WebCrow’s searches consist of key words derived from the crossword clue. Usually it’s enough for WebCrow to simply scan the immediate search engine or database search results, picking out words of the proper length that appear most often in the search results. WebCrow then compiles a list containing possible answers produced by each method, and checks the answers against each other to see how they link together in the crossword.

Granted, the general purpose of a crossword puzzle is lost if a Human were to consult a database, a dictionary or the internet to solve each clue, but obviously a computer doesn’t have the rote memory skills of a human. Despite the fact that some might label this cheating, these electronic means must suffice for now, as this science doesn’t strive to simply beat Humans at crosswords, but to understand relationships between a set of data in the same way that Humans are naturally able to. In actuality, the web is still considered a “shallow” source of data for such programs, because although there is great breadth to the information contained by the internet, it’s currently very difficult for a program like WebCrow to connect and relate the data that it searches for.

The full results of the competiton are posted on the WebCrow homepage.