LONG BEACH, CALIFORNIA—Robots might not yet make great standup comedians, but computers are learning to predict what we’ll find funny, according to a study presented here last week at the International Conference on Machine Learning.
Researchers conducted online surveys in which Americans rated the humor value of 120,000 words and nonwords, then used computers to analyze the data. Using ratings of some words, an algorithm could predict the humor of others. Predictably, senses of humor were idiosyncratic. But the software did find clusters of people with similar taste. For example, one mostly female cluster liked funny-sounding words (“gobbledegook,” “kerfuffle”), a younger male cluster liked sexual words (“asshattery,” “dong”), and an older group liked scatological and insulting words (“crapola,” “wanker”). (We’re excluding more explicit words.)
People judged words not just on humor, but also on whether they were colloquial (“wee lad”), insulting (“nincompoops”), juxtaposed (“party poopers”), scatological (“dung”), sexual (“foreskins”), or funny sounding (“lollygag”). If a word rated highly on any of these factors, it was more likely to be funny, but sounding funny was the most important factor.
The researchers also identified certain words rated much funnier by women than by men (“whakapapa,” “doohickey”), and others more preferred by men (“sexual napalm,” “poundage”). Given someone’s sense of humor profile—based on which words they found funny—the artificial intelligence (AI) could predict better than chance whether they were a man or a woman. It could also use such profiles to predict which of two people would find a given word funnier.
The research could lead to chatbots that sound more human and might be scaled up to predict the funniness of phrases or sentences, which could lead to writers’ assistants for evaluating or even generating chucklers. To those who think AI is hopeless at getting humor, the machines have one word for you: poppycock.