AI already knocked, now open the door!
Artificial intelligence may be finding its way into the work force, but one place we probably wouldn’t expect AI is on the stand-up stage. Humor, after all, is a distinctively human concept. An AI might learn how to quickly identify prime numbers, navigate Los Angeles without a map, or even act as a therapist but it seems unreasonable to think an algorithm can tell a good joke, right?
Virginia Tech’s Arjun Chandrasekaran et al have addressed this idea head on by attempting to train a machine-learning algorithm to identify – and even create – humorous situations. And MIT Technology Review reports that the research claims to be a success. According to their paper published online, Chandrasekaran and his team created an algorithm that can distinguish funny from unfunny despite being essentially blind to the social context.
In order to train their machine, the VT researchers employed Amazon’s Mechanical Turks to create scenes with clip art, then label them as funny or unfunny, order them by degree of funniness, and describe why certain scenes are humorous. To be sure, these aren’t laugh out loud situations. And it isn’t exactly high-bro humor. But the images do take into account numerous elements. 20 human models of various demographics, 31 animals in different positions, and around 100 everyday objects like windows and trees were used to create 6,400 funny and unfunny scenes.
The researchers then altered the 3,000 or so funny images into 15,000 unfunny images. These changes were sometimes slight, but managed to remove the funny element from the scene. In this way, the researchers could identify – and help their algorithm recognize – what components and associations generated humor and which ones didn’t.
Ultimately the machine was given the task to predict the funniness of a particular scene and then to alter the funniness by replacing one of the objects present within the scene.
Chandrasekaran’s algorithm did surprisingly well, accurately identifying humor over, well over 50 percent of the time. Meanwhile, when asked to alter funniness, the algorithm seems to recognize that certain objects and their associations are more inherently humorous than others. “We observe that the model learns that, in general, animate objects like humans and animals are more likely sources of humor,” the researchers note, “compared to inanimate objects and therefore tends to replace these objects.” More, the researchers say their algorithm learned how to promptly minimize humor by replacing odd objects with standard objects that suited their environment, making the scenes seem more common place. “In human evaluations, scenes made unfunny by our [algorithm] were found to be less funny than the original funny scene 95 percent of the time,” Chandrasekaran et al write.
It’s important to recognize that funniness is subjective and can usually be attributed to things like oddness and unexpectedness. So as Chandrasekaran and his team’s machine learns to identify humor, it’s really recognizing less subjective associations and the way these relations may make someone laugh.
Perhaps unsurprisingly, the algorithm performed successfully when researchers asked it to make an unfunny scene funny or make a funny scene funnier. The machine’s true talent was in making funny scenes unfunny. So, for now, don’t expect to see an AI’s Comedy Central special.
Image credit: qmcogsci/flickr