A mathematical model to help AIs anticipate human emotions

Humans form a society because we are effectively able to communicate and understand each other. “When speaking to someone who is frustrated or showing another emotion, we adapt the way that we interact with them,” explains Jussi Jokinen, an associate professor of cognitive science at the University of Jyväskylä in Finland. The researcher believes that artificial intelligence systems should be provided with a similar capability. In a context where people are using more and more AI tools, he warns of a “risk that AIs, which behave in a way that is increasingly different to humans, will cause friction and frustration among users.” Providing computer models with solutions that enable them to predict and adapt to human emotions could help to fluidify our interaction with them. “This is a sensitive subject, because it also increases the risk that AIs could manipulate humans.”

A mathematical model based on mental processes

In a bid to help computers understand human emotions, such as irritation, boredom, happiness and so on, the Finnish research team sought inspiration in the field of mathematical psychology. Their goal was to develop a mathematical model based on the functioning of the human mind, which makes calculations and deploys predictions and strategies to achieve particular goals. We also evaluate our performance to ensure that we can do better in the future, which is why these mental calculations come to be associated with emotional states: when they succeed, we experience positive emotions or, conversely, negative emotions when they fail. “The goal was to develop a computer model with the ability to predict users’ emotions. For example, in the event that a computer encounters an error during a task, the software tool could interact in a manner designed to mitigate negative emotion in the user by offering specific instructions.” Integrated into a chatbot, the model should be able to enable an AI system to better respond to users’ feelings. “Our model is not very complicated and could indeed be integrated into AI tools, but we need to decide on what this integration might look like, and the environment created to simulate the model. Both of these questions require further study.”

More appropriate synthetic data 

For Jiayi Eurus Zhang, a member of the research team and a doctoral student in cognitive science, the model could also be used to help users in real-life situations. “Whether it’s in a particular interface or another setting, we should investigate its potential to help users cope with their emotions on a day-to-day basis.” In other words, along with the ambition to predict emotional responses, the goal of this research is also to find out if systems of this kind can actually help users to influence and better manage their emotions when they make use of online services or open particular emails, etc.

Jussi Jokinen is convinced that mathematical models can also improve the quality of synthetic data generated by AIs. “Today, deep learning models are running out of training resources, so publishers are looking to train them with synthetic data. This is not necessarily a good idea, but if models are more in tune with the psychological workings of the human brain, the synthetic data they produce will be more appropriate”