New AI study finds teaching a robot to be funny may be harder than we thought

Two guys walk into a bar, and then a robot totally doesn’t get the punchline. 

A new study put together by researchers from the University of Michigan, Columbia University, Yahoo! Labs and the New Yorker uses a boatload of data to figure out what actually makes something “funny.” Turns out, the answer is a whole lot more complicated than you’d think.

The study compiled data from the New Yorker’s long-running caption contest, and pooled entries from 50 cartoons and 300,000 captions. They took that intel and broke the entries down linguistically, figuring out which ones were about different things (about people, positive emotions, negative emotions, various topics, etc.), as well as connecting them all by topics. Once the algorithm tried to figure out what was funny, they had seven users on Amazon’s Mechanical Turk rank which of two caption options is funnier.

It’s a fascinating concept to try and use math to define humor, and the findings did start to scratch the surface of what makes funny stuff funny. Per the report, they determined that “the methods that consistently select funnier captions are negative sentiment, human-centeredness, and lexical centrality.” Which, yeah, as Popular Science notes, that’s not a huge shocker considering the magazine's upper-class readership.

But could it determine which specific caption was funniest? Not really, apparently. Presumably, it sounds like the algorithm was relatively hit and miss when it came to actually picking the funniest caption, despite spotting some consistent trends in entries. But hey, progress is progress. We’re one step closer to cracking jokes with our own Bender, so we’ll call it a win.

(Via Popular Science, ARXIV)

More from around the web