A View from Henry Lieberman
Watson on Jeopardy, Part 3
IBM’s machine improves its performance and turns the game into a rout.
The second night of the IBM/Jeopardy match was a rout for the Watson program. The questions in this episode were harder, and, in my view, the match was more interesting.
Watson performed admirably. As I explained in my previous essays, the real game isn’t knowing facts; it’s the ability to do enough natural-language understanding to figure out what the question is asking for.
Clue: An étude is a composition that … The name is French for this
Here, it has to figure out that what’s being asked for is the English translation of the French word étude. Then do the translation. It got it. Bravo!
Even in its mistakes, I was impressed. And looking at mistakes that people make, too, tells you a lot. In one case, both Watson and the humans had trouble interpreting the question.
Clue: In May 2010, five paintings … left Paris’s art museum of this period.
Watson: Picasso (Wrong)
Ken Jennings: Cubism (Wrong)
Brad Rutter: Impressionism (Wrong)
Correct answer: Modern Art.
The question isn’t asking for the name of an artist or an art period—it’s asking for the name of a museum! (Musée d’Art Moderne—Museum of Modern Art). There is a Picasso Museum in Paris, but there is no Museum of Cubism nor Museum of Impressionism. Modern Art was Watson’s third choice. Kudos to Watson for at least getting that far!
The Final Jeopardy question built up the suspense.
Category: US Cities
Clue: Its largest airport was named after a World War II hero; its second largest for a World War II battle.
Human contestants’ answer: Chicago (Correct) Watson’s answer: Toronto (Wrong)
The question was interesting because it requires doing an intersection—figuring out which items several sets have in common. You think of every airport you know. Which ones are named after World War II heroes? Probably you can’t remember too many war heroes, but you do know which airports are named after people. Which cities have two airports? New York? Kennedy a hero, sure, but (I’m a New Yorker) LaGuardia was a mayor, not a battle. And so on. Cognitive science experiments show that people aren’t very good at doing intersections (What city is named after an animal? Answer: Buffalo).
I don’t know the names of the Toronto airports, but the answer is sure to antagonize Canadians, since the category was US Cities. I don’t know why Watson made that mistake. But it illustrates the problem of
overconstrained situations—you can’t find any really good answer, and you’ve got to break the rules. A better choice would have been to stay in the US and take the risk that you would be wrong about LaGuardia.
But Watson did have some game strategy that saved the day for it. It realized it was ahead and it couldn’t be caught by the other contestants, so it didn’t bet much on the final outcome. If it figured that out, that’s more impressive than knowing the facts of the answer. Strategy sometimes trumps knowledge.
Henry Lieberman is a research scientist who works on artificial intelligence at the Media Laboratory at MIT.