The Machine That Japed: Microsoft’s Humor-Emulating AI

Ten years ago, highbrow culture magazine The New Yorker started a contest. Each week, a cartoon with no caption is published in the back of the magazine. Readers are encouraged to submit an apt and hilarious caption that captures the magazine’s infamous wit. Editors select the top three entries to vie for reader votes and the prestige of having captioned a New Yorker cartoon.

The magazine receives about 5,000 submissions each week, which are scrutinized by cartoon editor [Bob Mankoff] and a parade of assistants that burn out after a year or two. But soon, [Mankoff]’s assistants may have their own assistant thanks to Microsoft researcher [Dafna Shahaf].

[Dafna Shahaf] heard [Mankoff] give a speech about the New Yorker cartoon archive a year or so ago, and it got her thinking about the possibilities of the vast collection with regard to artificial intelligence. The intricate nuances of humor and wordplay have long presented a special challenge to creators. [Shahaf] wondered, could computers begin to learn what makes a caption funny, given a big enough canon?

[Shahaf] threw ninety years worth of wry, one-panel humor at the system. Given this knowledge base, she trained it to choose funny captions for cartoons based on the jokes of similar cartoons. But in order to help [Mankoff] and his assistants choose among the entries, the AI must be able to rank the comedic value of jokes. And since computer vision software is made to decipher photos and not drawings, [Shahaf] and her team faced another task: assigning keywords to each cartoon. The team described each one in terms of its contextual anchors and subsequently its situational anomalies. For example, in the image above, the context keywords could be car dealership, car, customer, and salesman. Anomalies might include claws, fangs, and zoomorphic automobile.

The result is about the best that could be hoped for, if one was being realistic. All of the cartoon editors’ chosen winners showed up among the AI’s top 55.8%, which means the AI could ultimately help [Mankoff and Co.] weed out just under half of the truly bad entries. While [Mankoff] sees the study’s results as a positive thing, he’ll continue to hire assistants for the foreseeable future.

Humor-enabled AI may still be in its infancy, but the implications of the advancement are already great. To give personal assistants like Siri and Cortana a funny bone is to make them that much more human. But is that necessarily a good thing?

[via /.]

18 thoughts on “The Machine That Japed: Microsoft’s Humor-Emulating AI

  1. What is it with people trying to solve big AI problems by just “getting a big enough dataset” ?

    You’ll never be able to build a dataset large enough to cover many of the issues with AI, specifically because many of the unsolved problems resist reduction to simple rules and examples, and have solutions which are so nuanced and varied in how they are reached that the dataset required to derive/emulate the behavior under even a large number of the possible circumstances would be impractically huge.

    Not to mention, with a spread as wide as 55%, you have to wonder if the computer is really much better than a random number generator. (who’s to say that it wouldn’t have been an even larger spread, given a larger number of tests?)

    1. Just to point out.

      90 years of jokes? Seriously? If you look at any long running comic, the humor does change. If you look at many of the long running comics like Garfield and Peanuts you’ll find some central jokes have disappeared or evolved over time.

    2. >What is it with people trying to solve big AI problems by just “getting a big enough dataset” ?

      Well, it’s easier than writing code.

      Still, many of the processes of thinking, you can be aware of from noticing them in yourself (a small history of LSD will help with that). Then there’s the masses of psychology literature. It might be best to spend the time trying to code these rules in by hand. Perhaps in some sort of versatile framework that’s good at drawing relationships between things, that can generalise somewhat.

      There’s also the great missing fact, that humans evolved in a physical world. Where we verb nouns, where we shoot the brown deer with the sharp arrow. These rules come from real life (think about that, Chomsky). We run and eat and breathe and think. “We” is us, me and the guys over there. The basic structures of language mirror the structure of reality. Adjectives exist because there are different types of tree, deer, face, etc. Adverbs exist cos there’s many ways to do a thing.

      Might help to either simulate a physical reality, or much better, give an AI a body. If we’re going to go the neural net route, and let them figure it out for themselves, give them the same start we had. Put them in a world, give them rich sensory input.

  2. Title seems a bit mis-leading…no humor is created by the AI…therefore stating that the machine has ‘japed’ is outright balderdash. Try to attract readers without promising things you won’t deliver

  3. That rag and its cartoons are not nearly as funny or intellectual as it proclaims. Surprised to find that there is any kind of vetting process based on the past. I end up laughing at the articles and the lack of a genuine picture of problems or concern and of course flip past the non-funny cartoons like most folks. Hell the only time I see the New Yorker is in my dentist’s office.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.