NOS news•
If you startle someone and they calmly reply “Oh my God. I’m having a heart attack”, you know: this is not serious, on the other hand, computers have great difficulty recognizing such a comment as sarcasm. But that’s about to change: researchers at the University of Groningen have developed an algorithm that detects sarcasm.
Linguists Xiyuan Gao, Shekhar Nayak and Matt Coler – from China, India and the US – met and trained the new algorithm using data from sarcastic language. The so-called sarcasm detector then heard the sound of a comedy show Friends in Big Bang Theory. The detector knew this very well: 74% of the time it recognized the sarcastic comments of the characters Penny, Raj, Joey and Chandler.
And not insignificant: the algorithm was able to point about as successfully when it was not sarcasm. The researchers presented their results today at a meeting of the Acoustical Society of America.
Some data
The research was not easy because little data was available to train the model. The researchers had to settle for just four hours of audio, enriched with all kinds of information about the speakers’ sarcastic intentions. “That was a big problem for us,” said speech technology’s Matt Coler, one of the researchers.
To reach a successful conclusion, the researchers looked at as many variables as possible. What is the height of the speakers? How fast do they talk? How much variation is there on the field? What is the intensity of the speech? Does the meaning of the words correspond to the emotions of the sound?
It became clear that people in all languages often speak a little slower when they make sarcasm comments. But the pitch varies from language to language. In English and German, the pitch decreases and varies less. Like Chandler in the American series Friends is stunned and says slowly and steadily that he is having a heart attack, then this is a clear indication to the algorithm that it is sarcasm.
The expressions
In Italian, French, and Cantonese, the pitch actually increases, and so does the variation. A good sarcasm detector therefore needs to be trained according to the language. In any case, the researchers want to develop their detector further. “The challenge is to look not only at the sound of the speech, but also at the context in which the statement is made,” says Coler. “And facial expressions are also a very accurate predictor of sarcasm.”
The research is part of efforts to make communication between humans and computers more human. The idea is that if computers can better understand the nuances of human communication, they can also respond better to it. American technology company OpenAI took a step in this direction earlier this week with a new version of ChatGPT.
The sarcasm detector should also have concrete applications in the long term. Because people with autism and certain forms of dementia and schizophrenia often have trouble recognizing sarcasm. The detector could help them with this, the researchers think, for example by vibrating the wristband: watch out for sarcasm!