Researchers Have Built an AI-Powered Sarcasm Detector
Researchers at the University of Groningen in the Netherlands have reportedly developed an artificial intelligence model capable of detecting sarcasm in audio.
The research team in the university’s speech technology lab used video clips and other text and audio content from sitcoms like Friends and The Big Bang Theory to train a neural network. The work leveraged a database called the Multimodal Sarcasm Detection Dataset (MUStARD), which another research team from the United States and Singapore had annotated with labels about whether sarcasm was present in given pieces of audio content as part of a push to build their own sarcasm detector.
After training their AI model on the data, the Dutch university’s researchers were able to detect sarcasm in unlabeled exchanges almost 75 percent of the time. Additional work at the lab using synthetic data reportedly improved that accuracy level even further.
“We are able to recognize sarcasm in a reliable way, and we’re eager to grow that,” Matt Coler, a researcher at the University of Groningen’s speech technology lab, told The Guardian. “We want to see how far we can push it.”
Shekhar Nayak, another member of the research project, said the team’s methods and findings could help AI assistants interact more easily with human speakers by detecting negativity or hostility in the speaker’s voice.
Xiyuan Gao, a doctoral student at the university, noted that incorporating visual cues into the AI tool’s training data could further improve the system’s accuracy by enabling it to detect sarcasm conveyed through facial expressions like raised eyebrows or smirks.
But, Gao added, 100 percent accuracy is unlikely. “That’s not something even humans can achieve.”
Nonetheless, the researchers have said that their efforts are valuable. Sarcasm, Coler said, permeates our discourse more than we might appreciate, so understanding it is crucial if humans and machines are to communicate seamlessly.
It’s so crucial, in fact, that the U.S. Department of Defense, through the Defense Advanced Research Projects Agency (DARPA), and researchers at the University of Central Florida have been working for years on an AI model capable of classifying whether text documents like social media posts contain sarcasm.
“Knowing when sarcasm is being used is valuable for teaching models what human communication looks like and subsequently simulating the future course of online content,” the Pentagon agency said in a statement at the time.
Deepfake Detection Gains Momentum
CivAI, a nonprofit that builds an understanding of artificial intelligence and its dangers through interactive software, has launched the Deepfake Sandbox, a free, public tool that allows everyone to experiment with high-quality deepfakes and learn more about the risks generative AI poses.
“Our message is simple: AI is going to hit society like a truck, and people have a right to know,” said CivAI cofounder Lucas Hansen in a statement. “If you’ve never seen a good deepfake, watched AI generate a personalized phishing email in real time, or heard a voice clone of a loved one asking you for money, it’s hard to imagine just how far the technology has come. Our goal is to expose more people to AI and deepfakes in a way that makes clear the fundamental issues we face as it spreads and improves.”
“Surprise AI developments are popping up every month,” said CivAI cofounder Siddharth Hiregowdara in a statement. “Deepfakes like these would have been considered magic just a few years ago. But they’re here now, thrusting us into a world where we can no longer rely on photo and video evidence to determine what’s real. We believe that firsthand experience with AI is key to comprehending this transition and the many changes AI will bring.”
And CivAI is not the only entity making deepfake detection news in the past few weeks. In late July, Pindrop, a voice authentication and security solutions provider, secured $100 million to further develop its audio, voice, and artificial intelligence technologies.
“As cyber threats continue to evolve, our mission to stay ahead of fraudsters and protect our customers is more critical than ever,” said Pindrop Founder and CEO Vijay Balasubramaniyan in a statement.
At the same time, ElevenLabs and Reality Defender announced a partnership to advance the development of audio deepfake detection models.
“This collaboration…will allow us to provide highly advanced audio detection that stays several steps ahead of bad actors and fraudsters,” said Ben Colman, cofounder and CEO of Reality Defender, in a statement.