IBM’s Watson beat a trio of Harvard business School contestants in a mock Jeopardy competition Monday. The natural language processing computer had already proved its might during a real Jeopardy competition earlier this year when it beat two Jeopardy champions.
Jeopardy is a difficult game for a computer because questions use idioms and word play, things that haven’t been strong points for artificial intelligence.
I think it’s because Jeopardy has this broad domain. You can’t imagine solving a Jeopardy problem by building a big FAQ. It’s not like you’re going to map to one of a thousand questions to answer Jeopardy. So there’s a huge variety in both the topics it talks about as well as the complexity and richness of the language used in the clues. So these two aspects get you to think more broadly about how to reason over large volumes of content. But then the other interesting problem in Jeopardy is the probability aspect. You don’t want to buzz in and get the question wrong so now you have to compute a probability that your answer is correct.
Watson would come up with 3 responses to each clue and rank its certainty on each of them. If Watson was relatively sure of the top responses accurately it would ring in and answer. If Watson was in the lead the threshold would move up so it wouldn’t answer as often, if it was trailing the threshold would move down so that it could try to regain the lead.
Harvard was able to take the lead away from Watson for just a short time when it wagered more than 11,000 dollars on a daily double in the category ‘Having a ball.’
MIT’s Sloan school of Business couldn’t find its rhythm and eventually the competition came down to final Jeopardy where all the teams got the correct response with Mount Rushmore, but Watson wagered enough to win.
Some of the next steps for Watson are to adapt its intelligence to applications in finance, health care and technical support. At the Harvard Business School, Nick Barber IDG News Service.