Illustration by Ruth Gwily
When AI learns to create puns:
Can machines become comedians?
AI that knows how to create puns doesn't seem to be serious for an artificial intelligence researcher. But for He He who did this during her postdoctoral research at Stanford, it was the starting point for a very strange problem in machine learning. He He's goal is to build an AI that can speak naturally and humorously - not merely reading news or informing us about the weather but it can joke, compose a good poem or even tell a fascinating story. But according to He, to achieve this we need to go against the limits of how the AI usually learns.
Neural networks are natural imitators who learn language types by scanning large amounts of text. If the goal is coherency, then the approach works very well. In fact, recent advances have sparked a moral debate about whether humans can take advantage of AI to create compelling fake news. But the resulting passage are also as dry as press texts and Wikipedia articles often used to train them. In other words, neural networks are bound by rules when avoiding errors and this makes them very bad when it comes to jokes. He said an ingenious joke will lie at the edge of coherence but not become meaningless, and neural networks are simply incapable of reaching this balance. Besides, the most important thing of creation is to be novel. According to He, "Even if we have a long list of puns that AI can learn, it has lost its point of creation."
Instead, He and her team including Nanyun Peng and Percy Liang, tried to give their AI a bit of witty creativity when using insights into humor theory. For example, with the sentence "The greyhound stopped to get a hare cut", the phrase "stopped to get a hare cut" has the meaning itself (“get a hare cut” sounds like “get a haircut”), but it still makes people laugh when linking all in the sentence again thanks to the word "greyhound". Using the principle of "local-global surprisal principle", the neural network is given a pair of homophones ( which is hare-hair in this situation) and creates a normal- sounding sentence with the first word but surprises people when the second word is swapped. Then, in order to avoid a grammatical sentence, the sentence will be inserted with other words to bring about more logic in whole.
He He created the pun method by editing a bland sentence to create humorous tension.
Next, He held a pun contest where AI competed with humorous people. According to reviewers of the pun games, the results for machines are not good, at least according to human standards. While He's system produces much funnier puns than the AI control effort as before, it only defeated human 10% of the total game time. In addition, the puns are stuck in a rather crude structure, suggesting that AI sometimes struggles with grammar while joking.
That’s because negotiator got my car back to me in one peace.
Even from the outside, I could tell that he’d already lost some wait.
Well, gourmet did it, he thought, it’d butter be right.
“We’re nowhere near solving this,” He says.
However, Roger Levy, director of MIT's computational psychology lab, said the method is a promising step to build AI with a bit of personality. "Humor is a challenging aspect when studying the mind, but it also underpins what makes us human," says Levy. Four years ago, Levy described a computational method to predict whether a pun was humorous or not, the achievement of which eventually became the foundation for He He's method of creating jokes. Levy said that he planned to test something like "local-global surprisal principle", which was better adjusted than the theories used in his article. This concept is reasonable based on Levy's intuition but he had not had enough data to prove it. So, "it's great to see that really growing," Levy said.
Looking more broadly, humor research underscores the need to put human intelligence into neural networks more. In recent times, Levy has used surprisal as a way to study different aspects of how AI understands language. According to Levy, "Surprisal is one of the most central concepts in both AI and cognitive science." That makes surprisal a useful way to compare how the human brain and machine deduce in their own way through language. By bringing the neural network into a series of psychological tests to study how humans handle words with ambiguous meaning, Levy begins to realize that machines overcome challenges in a way unlike humans. Adjusting those differences, he says, may be the key to designing AI that is closer to human behavior.
Meanwhile, He said that she hopes to apply her pun creating approach to more difficult creative tasks such as telling stories. He said the idea here was for the neural network to do what it did well and then correct the results with human intelligence. For example, a neural network can be trained to create a series of perfect, coherent sentences, and then learn how to edit that output into a creative short story based on narrative theories . He's goal is that AI will be able to create more creative and interesting stories. "I want AI to write stories about things people have not thought to write”
Source: Wired, techinfinite.in, ideas.ted.com, newsroom.cisco.com, hautil.us, wsj.com