Artificial Intelligence (AI) is playing an increasingly significant role. Robots are, for example, being used more and more often as companions for elderly people experiencing loneliness. Kruijt investigated whether it is possible to teach a robot how humans speak and to adapt to that, so that communication with such a robot becomes more natural and personal—perhaps even resulting in a unique way of speaking between human and robot.
Experimental Method
“To study this, it’s important that humans and robots talk to each other over an extended period,” says Kruijt. For this purpose, he and his colleagues developed an experimental method in which a robot and a human play a game together. “The game is a bit like ‘Guess Who?’ or ‘Spot the Difference’,” he explains. “Participants talk with the robot about who they see in their pictures. This encourages them to describe people and use references.”
Interaction
Kruijt examined how the way people referred to others changed over the course of the game. “We wanted to know whether these references became shorter, more personal and more efficient as the interaction progressed – similar to what we see between humans.”
A Model for the Robot
To explore this, he developed a special model for the robot, which allowed him to influence how it interpreted human descriptions. Rather than using large language models such as ChatGPT, Kruijt opted for custom-built methods. “This allowed us to better control and explain the robot’s choices.”
Human-Robot Communication
The research showed that people tended to continue using fairly detailed descriptions, even after longer interactions with the robot. “Between humans, nicknames or brief references tend to emerge quickly. We saw this much less with robots, so my research highlights a clear difference in human-robot communication.”
Context
The study also revealed that numerous factors influence how people choose words to describe someone, and how well the robot is then able to understand those descriptions. “The context of a conversation is the most important factor. When a reference could be interpreted in multiple ways, it tends to become more detailed to avoid miscommunication. Humans are more aware of this than robots. As a result, they are less likely to agree on which reference to use, and miscommunication remains a risk,” Kruijt explains.
Progress
Nevertheless, Kruijt does see progress. “Communication between humans and robots did become more successful. Over time, participants and robots understood each other better, even though a truly shared language, as we see among humans, did not emerge.”
Future
“The results of my research are relevant to the development of more social, personal forms of artificial intelligence,” Kruijt states. “Think, for instance, of social care robots, which provide companionship and keep people socially engaged in care homes.” Kruijt emphasises that his method offers better control over the robot’s decision-making process than the widely used large language models. “This makes the robot’s behaviour easier to explain. At the same time, it shows that there is still much work to be done before social robots can genuinely function as good companions for humans.”
Photo: Peggy van Minkelen