What Is … a “Jeopardy!”-worthy Computer

In the battle of man vs. machine, even one of humanity’s most charming and knowledgeable representatives has come up short yet again.

A computerized question-answering system (QANTA) built by a team of students and researchers from the University of Maryland and the University of Colorado Boulder recently went head-to-head against “Jeopardy!” champion Ken Jennings in a quiz bowl-style competition and won.

“I’m used to losing to computers in front of a crowd,” said a modest Jennings, who was one of the contestants bested by an IBM computer named Watson on “Jeopardy!” in 2011. (He still holds the record for the longest winning streak on the game show, earning $2.52 million in 75 games, then another $500,000 in the Ultimate Tournament of Champions.)

The match at the University of Washington on Oct. 2 marked the second time that QANTA competed against humans. The first exhibition match, held in Chicago in June, resulted in a tie.

Four UMD computer science students—Mohit Iyyer, Anupam Guha, He He and Stephanie Hwa—and Hal Daumé, an associate professor of computer science and director of the Computational Linguistics and Information Processing laboratory in the university’s Institute for Advanced Computer Studies (UMIACS), helped build QANTA. Much of the early work was led by Jordan Boyd-Graber, an assistant professor at the University of Colorado Boulder and a former member of UMIACS and the College of Information Studies, who also emceed the Oct. 2 competition.

Using cached information from Wikipedia, QANTA analyzes a question, matches it against possible answers and figures out when it has enough information to provide a guess. In this way, it is uniquely built for a quiz bowl, which rewards quick answers and decreases in difficulty over the course of a question. (For example, a question about Venus in the competition started with a Japanese space probe that studied the planet and ended with how it’s second closest to the sun.)

During the contest in a college classroom, QANTA got the first seven questions by correctly identifying subjects like a Hapsburg general and a George Bernard Shaw play before Jennings got on the board with Austrian filmmaker Fritz Lang.

“If it doesn’t start watching movies, we have a chance,” Jennings wryly noted.

QANTA isn’t invincible, however. On the last question, which asked to identify a Dutch physicist, Jennings had the wrong name but QANTA didn’t even guess a person (its response was “natural rubber”).

“At least we ended with both of us looking dumb,” Jennings said.

Boyd-Graber said QANTA tends to do best on difficult questions that are actually about easy things. As the field of potential answers expands, especially in a subject like pop culture, the system has a hard time pinning down a good answer.

But Jennings said he felt like QANTA was coming up with answers faster than he was, whereas Watson’s main advantage was how quickly it could buzz in. “It’s legit faster than me—as well as knowing more stuff.”

The match can be watched below:

Melissa Brachfeld contributed to this story.