Monday, April 27, 2009

AI in Jeopardy

John Markoff writes in the Times about IBM's plans to create a computer contestant for the Jeopardy TV game show. Jeopardy is a basic knowledge and trivia show with few gimmicks (beyond giving questions for answers instead of vice-versa) which is probably why the show has endured so long. The issue for the computer is not knowledge (one could download the entirety of Wikipedia on a USB drive) but of properly interpreting the question.

Back in my NEC days, some of the AI folks there wrote a similar program for the game show Who Wants to be a Millionaire, although the program was never considered for use on the real show. Unlike Jeopardy, Millionaire had multiple choice answers and their program basically did keyword matching on Google searches. Their program did quite well on the very specific high-dollar questions. But their program often stumbled on the low-dollar easy questions which relied on basic knowledge.

This Jeopardy computer experiment still seems to me comparing apples to oranges. Humans have little trouble interpreting the meaning of the "answers" in Jeopardy, they are being tested on their knowledge of that material. The computer has access to all that knowledge but doesn't know how to match it up to simple English sentences. I suspect the computer will win this fight for it has the ultimate Jeopardy advantage: it can buzz in faster.


  1. Obama vows investment in science

  2. Link to the "NEC days" paper (thanks Lance): Shyong K. Lam, David M. Pennock, Dan Cosley, and Steve Lawrence. 1 Billion Pages = 1 Million Dollars? Mining the Web to Play "Who Wants to be a Millionaire?", Conference on Uncertainty in Artificial Intelligence, pp. 337-345, 2003.