杏吧原创

My PC predicted you鈥檇 laugh (Image: Dan Kitwood/Getty)

My battery had an alkaline problem, so it went to AA meetings. Hilarious, eh? Don鈥檛 blame me, a computer model predicted that you would find this pun amusing. Its makers hope to use it to give robots a sense of humour.

We use humour every day, whether to ease tensions, cheer people up or forge romantic relationships.

鈥淎 big part of language is making other people laugh,鈥� says a cognitive psychologist at Stanford University in California.

But computers have always been terrible at understanding humour. The ability to comprehend or tell a joke relies on a detailed knowledge of language, culture, stereotypes and personal experiences 鈥� something computers do not have. Another issue is that there aren鈥檛 very clear definitions of what makes a joke funny, says Kao.

So she and her colleagues decided to begin with a particularly clear-cut type of humour: the pun. 鈥淲e started out using our intuition,鈥� she says.

Kao鈥檚 team believes that there are two key ingredients to a good pun. Firstly, to be a pun, a sentence must have a degree of ambiguity 鈥� you should be able to draw multiple meanings from it.

To be funny, a pun should also score highly on what the team calls 鈥渄istinctiveness鈥� 鈥� the degree to which each of the multiple meanings is supported. 鈥淭he magician got so mad he pulled his hare out鈥� would score highly, for example, because both potential interpretations of hair and hare are likely.

Puns with higher scores of distinctiveness are funnier, says Kao. For example, the pun: 鈥淭he rower could not choose either oar鈥� works, but adding the word 鈥渋ndecisive鈥� before 鈥渞ower鈥� to support the 鈥渙r鈥� interpretation more strongly makes it funnier. 鈥淵ou鈥檇 bump up the distinctiveness,鈥� says Kao.

With this in mind, Kao and her colleagues aimed to develop a program that could identify whether some text is a pun, and then predict how funny it is, based on its ambiguity and distinctiveness.

To help learn about ambiguity, the team asked a group of 100 volunteers to judge the similarity of pairs of words used in puns taken from a joke website, phonetically and in meaning.

The team also asked the participants to rate how funny they found each of the puns, so that they could test and hone their model by comparing the predicted funniness to the answers from the volunteers.

鈥淚t is the first model to give a fine-grained prediction of funniness,鈥� says Kao. She intends to develop the model further to enable it to generate its own puns, as well as other types of jokes.

While Kao hopes to use her model to give robots a sense of humour, it could also be incorporated into an app that allows people to test out their puns and even improve them.

Not everyone likes that idea, however. 鈥淚 absolutely would never use a computer to tell me what is funny,鈥� says Tim Vine, famous for his one-liners and puns. 鈥淚 use one thing and one thing only. The audience. And it has to be an audience of humans.鈥� See below for a list of Vine鈥檚 10 favourite puns

Given the importance of humour in forming relationships, robots鈥� inability to grasp wordplay and sarcasm could hamper their ability to integrate into our lives, whether as toys, appliances or carers.

鈥淲e have talking computers all around,鈥� says Julia Taylor at Purdue University in Lafayette, Indiana, who is also investigating ways to enable computers to communicate informally with people. 鈥淔or that communication to be on a human level, humour has to be integrated, no matter what.鈥�

Journal reference: Cognitive Science, doi.org/6n9