|
Earlier this month, an IBM supercomputer named Watson won
over Americans by competing on and winning the popular game show, Jeopardy.
Watson took on Jeopardy's two top champions and won the 3-day contest and a
million dollars that was later donated to a children's charity.
Before Watson could make its debut, IBM researchers worked
to understand the nuances of "natural language" within the system. They used the company's open-source
framework, "Unstructured Information Management Architecture (UIMA), which
analyzes natural language text, speech, image, and videos. Watson also employs "DeepQA" technology that
IBM describes as a quest for a system that operates more effectively in "human
terms."
Putting Watson to the
Test
According to the IBM research blog, it took two years before
the computer could analyze a Jeopardy clue and provide a response in less than
three seconds. To test Watson, the research team devised 55 "sparring matches"
against former Jeopardy Tournament of Champions contestants.
By using learned algorithms, Watson uses information given
in the category and clue to map probabilistic estimates in the accuracy of
responses. At the same time, Watson relies on getting the clues correct to
verify that it's interpreting a category correctly. The early sparring matches
showed significant learning and in-game adjustments as Watson gained a clearer
understanding of the categories.
Deep QA technology also helps Watson produce answers with an
associated confidence before ringing in. Watson will typically ring-in an
answer when there is a confidence threshold of approximately 50%; this
threshold may change towards the end of the game if it provides a higher change
of winning.
Did you watch Watson on Jeopardy?
Sources: IBM
Research News Blog
| 
Re: Training Watson to Take on Jeopardy Champions
