Teaching computers to understand jokes and humor is a funny business, according to Julia Rayz.
“Joke generators have been around for years, and some have been quite successful,” said Rayz, “but there was no system at all for attempting to detect a joke” at the time she started researching computer recognition of humor.
Rayz, associate professor of computer and information technology, is trying to help computers understand and interact with people as naturally as people interact with each other.
“Humor is a good way into this,” she said. “It’s difficult because it’s not enough to just create or find a joke. Recognizing humor in language is complicated, even for what you might call a standard artificial intelligence.”
Although computers have made great strides in natural language recognition, artificial intelligence hasn’t mastered the nuances of context, delivery, emotion, or relevance, to name just a few of the challenges associated with recognizing humor or telling a joke.
“If you’re having a bad day and want a computer to be funny for you, it needs to know what you went through that day and what you’ll react to,” said Rayz. “It has to cater to your mood and your sense of humor. Do you like long jokes or short ones? Do you like puns? It has to learn your preferences.”
That’s just the beginning, said Rayz. Although modern computers aggregate staggering quantities of information posted online every day, artificial intelligence might not recognize humor when it finds it.
“When you open a book and read a short paragraph, is it a joke or not? A computer needs to be able to understand that,” said Rayz.
Improving artificial intelligence for human-like understanding of humor is a challenge which goes well beyond the field of information technology. Rayz said researchers with backgrounds in linguistics, psychology, sociology, anthropology, and other fields are working on the problem together.
Some of the most widely-recognized research into “what makes a joke a joke” was done at Purdue by Victor Raskin, distinguished professor of English and linguistics in the College of Liberal Arts. In 1985, Raskin wrote “Semantic Mechanisms of Humor,” providing the first book-length theoretical framework and linguistic algorithm for interpreting verbal humor.
Additionally, Salvatore Attardo, dean of humanities, social sciences and arts at Texas A&M University–Commerce, studied under Raskin and extended Raskin’s script-based semantic theory of humor into the general theory of verbal humor.
Rayz’ research involves providing artificial intelligence with understandings of Raskin’s and Attardo’s theories along with ontological semantic theories of humor.
“Their theories were written when computers had no ability to analyze human language,” she said. “Today, using a (software-based) ontological reasoner, you can detect whether or not something is a joke, or you can compare the language in different jokes.”
Again, though, for teaching artificial intelligence to understand humor in language, Rayz said that’s just the beginning. Even with the recent advances in language processing, computers haven’t yet reached a child’s level of cognition.
“You can ask people whether something is funny or not. But if you ask people whether something is a joke, they still answer based on whether it’s funny,” said Rayz. “Once you start with a computer, it might be able to tell whether something in the text is a joke, but it can’t tell whether it’s funny. We discovered that the hard way.”
The current limitations of artificial language processing can be seen by using a search engine such as Google, Rayz said. “If you ask Google a question which isn’t just factual, where it has to put two and two together to come up with a result, odds are it will fail miserably and just try to produce facts. And in the world of a joke, we sort of sketch a situation just enough for a human to figure out what’s going on, with all of the inferences that are there.”
Getting computers to do the same thing remains difficult, said Rayz, but research is ongoing.
“People aren’t forgiving when it comes to mistakes in humor. They may laugh at you but not with you,” she said. “Computers need to learn the difference.”