Better humor through computer science
For those with a refined punchline palate, Jester, the Online Joke Recommender, is a dream come true. The website’s premise is simple: Read eight jokes and rate them according to how funny you find them. After that, Jester will begin suggesting new material tailored to your tastes. Whether you’re a fan of shaggy-dog yarns or snappy one-liners, you’ll find plenty of material for your next social gathering or paper presentation.
Ken Goldberg, robotics and engineering professor and director of Berkeley’s Center for New Media, is the mastermind behind Jester. His patented algorithm, Eigentaste, now in version 5.0, specializes in what is called “collaborative filtering.” If you’ve ever been recommended books while browsing on Amazon or movies on Netflix, you’re already familiar with collaborative filtering. As Tavi Nathanson, Professor Goldberg’s graduate researcher, explains it, “You take a user, you find users who are similar, you recommend items based on what those similar users said they liked.”
That certainly sounds straightforward enough, but under the hood, it’s all hardcore mathematics: The recommendations are based on numbers, not the content of the jokes. “That’s a key point,” stresses Nathanson. “The system works purely based on ratings. It’s this clean statistical approach.”
Goldberg’s algorithm gets its name from something called “eigenvectors.” Don’t ask. All you need to know is that “eigen” in German indicates membership in a group—fitting, since Eigentaste finds people with the same taste in things and groups them accordingly. Once you’ve evaluated the initial jokes, Eigentaste places you into one of 64 different categories of joke lovers—your specific humor profile. Why 64? Nathanson explains that the cluster count could have been any power of 2, but their “gut feeling” was that 64 was broad enough to present a wide array of humor profiles but not so many as to split whiskers.
Unlike the collaborative filters used by Netflix and Amazon, Jester is designed for cases where users can rate a set of items on-the-fly. That doesn’t work so well with books or movies or toaster ovens, which take time to evaluate. But it’s perfect for jokes.
The immediacy of jokes makes Jester addictive, or “sticky” in Web lingo—a much sought-after quality in websites. On average, about 20 users daily rate about 46 jokes on Jester in a sitting. “Jokes have a naturally magnetic property,” says Professor Goldberg. “People are happy to sit down and read them, and you can evaluate [our set of 8] in about a minute.” Happily, that has provided Goldberg’s team with beaucoup data to work with; there are over 4 million ratings so far.
Jester is not without drawbacks. Perhaps the biggest problem is that jokes quickly grow stale; they’re never as funny the second time around. The system, which currently has 128 jokes, has a built-in field where users can submit their own gut-busters, but precious little new humor makes the cut.
So is Jester accurate in matching jokes to visitors’ tastes? It’s a tough question but maybe not the best one to ask, argues Nathanson. He points out that if a user is drawn to, say, Chuck Norris jokes, the safest recommendation would probably be to offer more of the same. Riskier recommendations, on the other hand, while bound to hurt accuracy, “would most likely make the system more useful or enjoyable.” That’s undoubtedly true. Just don’t tell Chuck Norris.