Microsoft Creates an AI That Can Spot a Joke In a New Yorker Cartoon

An anonymous reader writes: For over a decade Bob Mankoff, the cartoon editor at the New Yorker, and his assistants have gone through 5,000 cartoon entries for the magazine's caption contest each week. Needless to say, the burnout rate of his assistants is quite high, "The process of looking at 5,000 caption entries a week usually destroys their mind in about two years, and then I get a new one," Mankoff says. But now thanks to a collaboration with Microsoft, Bob may finally have found the perfect helper. Researchers have been working on an artificial intelligence project to teach a computer what's funny. Fortune reports: "Dafna Shahaf, a researcher at Microsoft, used the database of cartoons to train the program to understand commonalities and differences in the millions of cartoons, which lets the AI run through the entries the New Yorker receives each week for its back-of-magazine cartoon caption contest. About 55.8% of the time the humans agree with the captions the AI selects, which is a pretty good percentage."

Comments

What they actually did.

[paskie]

The paper is a bit confusing at first, and the /. summary doesn't help. Basically, they developed a sorting criteria to reduce the amount of work for the editors. In an isolated comparison of two jokes, the funnier joke wins 64% of them on average; this is quite better than a coin!

To get a sorted list, they run a "comparison tournament" between the jokes. The 55.8% number means that the funniest joke is in the top 55.8% of the list on average; if we are willing to occasionally miss a brilliant joke, we can cut the list in a little more than half and still keep most of the great jokes.

The full paper is http://research.microsoft.com/...