Mild-Mannered Watson Skewers Human Opponents on Jeopardy
IBM's Watson supercomputer seems like an amiable-enough character, judging from his performance in a practice run against two top Jeopardy players. (Then again, H.A.L. seemed friendly and helpful too, until crossed by a human.) Should the emergence of technology that enables computers to understand and emulate human communications be viewed with any degree of trepidation?
01/14/11 10:41 AM PT
A supercomputer designed by IBM and dubbed "Watson" went up against game show Jeopardy's two all-time champions -- and won.
Pitting champs Ken Jennings and Brad Rutter against Watson, the contest was a practice round for the main event: a televised face-off between man and machine next month on the popular game show.
The games, one of which is being taped on Friday, will air on Feb. 14, 15 and 16. The first place prize is US$1 million dollars, second is $300,000 and third is $200,000.
Language Arts Supercomputer
Given Watson's creds, it is likely to win that those rounds as well. Watson is supercomputer that is anthropomorphized as a small black box sporting a globe avatar.
It is no ordinary supercomputer, though: It has been designed to "understand" language -- that is, Watson uses natural language processing to analyze how words relate to each other, and to grasp all the subtlety and nuance of which human language is capable.
What better venue to test how successful IBM was in manifesting this concept than a match on Jeopardy.
IBM could not anticipate what Jeopardy clues would be given, so it couldn't build and curate databases of all possible answers ahead of time, explained David Gondek, PhD, a researcher on Watson Algorithms and Strategy.
"Watson doesn't look up answers in a specialized structured database," he told TechNewsWorld, "but instead analyzes vast amounts of unstructured content -- text from encyclopedias, newspaper articles, textbooks, literature and more, using natural language processing techniques.
"These techniques have to contend with the complexities of language in order to both interpret what the Jeopardy clue is describing and to analyze text passages that support a given answer," Gondek continued. "This is something humans do all the time without ever explicitly thinking about it, but it turns out to be exceedingly difficult for a computer."
Surrounding vs. Cushioning
For instance, does "surrounding" mean the same thing as "cushioning"? In general, one might think not, Gondek said, but consider the following Jeopardy clue:
"In cell division, mitosis splits the nucleus and cytokinesis splits this liquid cushioning the nucleus."
Watson can find a passage explaining that "cytoplasm is a fluid surrounding the nucleus," but it then has to understand that in this context "surrounding" means the same thing as "cushioning," as well as the fact that a "fluid" can be a "liquid," Gondek explained.
Beyond Game Shows
Ultimately, what all this means is that the technologies used in Watson point to a future when computers can be used to interact with users in the way they actually speak and communicate, Gondek said.
Certainly, that is the point of test-driving Watson in the fun and harmless venue of the Jeopardy show.
"It is exciting to think that a quiz show could be a testing ground for the next generation of artificial intelligence," Phil Zimmerman, Jeopardy's senior publicist, told TechNewsWorld.
The technologies behind Watson can be applied to healthcare, technical support/contact centers, enterprise knowledge management and business intelligence, Gondek theorized -- any area in which the language can be both colloquial and technical, just like the Jeopardy clues.
"Our hope is that computing can move beyond databases requiring precisely constructed queries to a future where computers operate fluently and flexibly to understand questions, gather answers, evaluate evidence and explain it all to users," he said.
"From an AI standpoint, the fact that IBM was able to build a computer that can interact on an intellectual basis with human speech clearly has ramifications that are far beyond Jeopardy," said Scott Testa, a marketing professor for Cabrini College.
"But I think for people to accept that, it has to debut first in places like Jeopardy," he told TechNewsWorld, "taking it to a basic level that people can understand and smile at before they let it make more complicated computations for them."