Elementary Watson

January 6, 2012

Early last year, I wrote several times about IBM’s Watson system, and its victory against two human champions in the popular TV game show, Jeopardy!.  IBM has also announced some initiatives to use Watson’s technology in medical diagnosis.

In December 2011, at the USENIX Large Installation System Administration Conference [LISA] in Boston, MA, Michael Perrone, Manager of Multi-Core Computing at IBM’s Thomas J. Watson Research Center, gave an entertaining talk on Watson’s methods, and how they were developed in the context of the Jeopardy! game.  He presents some statistics and examples to show how the game is difficult, compared to something like chess, and how what Watson does is different from a search engine, like Google.  Inference from natural language is tricky.  For example, what do these three things have in common: shirts, TV remotes, and elevators?  Well, they all have buttons.

Mr. Perrone also gives examples of some of the questions Watson got wrong in its early incarnations (it also made a couple of bloopers in the match against human champions).  For example, in the category “New York Times Headlines”, the clue was:

An exclamation point was warranted for the end of this! in 1918.

Most people would probably realize the correct answer is “What is World War I?” — Watson came up with a silly but eminently logical answer: “What is a sentence?” The talk also includes some information about the hardware used to implement Watson for Jeopardy!.

The talk has now been made available on YouTube; it runs about one hour and twenty minutes.  The slides from the talk [PDF] are also available.

%d bloggers like this: