February 15, 2011

Watson Supercomputer Draws to a Tie in First Round of Jeopardy Match

Michael Feldman

The first day of the Jeopardy match between IBM’s Watson supercomputer and the game’s two top stars ended in a tie between the machine and Brad Rutter, with Ken Jennings a distant third. And if Watson hadn’t been deaf, he would have taken the round easily.

The TV setup was the same as Jeopardy practice matches, with Watson’s spooky little avatar above the middle podium, sandwiched by the two human challengers on each side. After each clue, Watson’s top three guesses (whether it buzzed in or not) were displayed at the bottom of the screen, along with the corresponding confidence level for each choice. In most cases, Watson won’t hit the buzzer unless its confidence level for an answer reaches above 50 percent.

In this initial matchup, the computer jumped out to an early lead, correctly answering 10 of the first 14 clues, plus scoring on a Daily Double that the computer bet to the max — $1,000. Watson managed to decipher even obscurely worded clues like this:”WANTED FOR A 12-YEAR CRIME SPREE OF EATING KING HROTHGAR’S WARRIORS; OFFICER BEOWULF HAS BEEN ASSIGNED THE CASE. To this, Watson immediately buzzed in with “Who is Grendel?” Apparently, the computer was 97 percent sure of that one. In fact, for most of the computer’s responses, its confidence level was north of 90 percent.

By the first commercial break, Watson was out in front with $5,200, with Rutter at $1,000 and Jennings at $200. The humans had managed to answer just two clues apiece.

After the break, Watson made its first mistake, guessing “What is finis?” to the clue “FROM THE LATIN FOR ‘END’, THIS IS WHERE TRAINS CAN ALSO ORIGINATE.” The correct answer was “What is terminus (terminal)?” During this series, the computer stumbled through a few more clues, as Rutter and Jennings made up some lost ground.

Where Watson seemed to have the most trouble was interpreting the category of thing being asked for. For example, in this clue, the computer couldn’t seem to figure out that the correct answer had to be the name of a country (Greece): “IN THE 2004 OPENING CEREMONIES A SOLE MEMBER OF THIS TEAM OPENED THE PARADE OF NATIONS; THE REST OF HIS TEAM CLOSED IT.” Watson’s top three guesses (and corresponding confidence levels): Olympic Games (20 percent), Athens (15 percent) and 2004 Summer Olympics (13 percent).

At one point, Watson offered the same incorrect response Jennings had just said out loud. Since Watson can’t hear, it isn’t able to take advantage of its competition’s mistakes — a big advantage for the mortals. As it turned out, Watson’s second choice on this clue, at a 30 percent confidence level, was the correct response. That cost Watson $2,000.

Even when Jennings and Rutter were racking up the points, Watson seemed to be just a hairbreadth away. For seven clues answered correctly by the humans, the computer’s first displayed choice was actually correct, and in most cases they were breaking the 50 percent confidence level. In three other cases, the correct answer was a second- or third-place choice for Watson.

Here’s the first half of the broadcast from Monday night:

To watch the second half of the broadcast, go here.

The round finished with Watson at $5,000, Rutter at $5,000 and Jennings at $2,000. Tonight the fun continues with double and final jeopardy rounds from Monday night’s game, with a third exhibition game to follow on Wednesday.