Aarhus University Seal

Researchers train AI model to recognise emotional responses on social media

Researchers from DATALAB – Center for Digital Social Research have further developed a model to predict six primary emotions and their imprint on social media. The hope is that it can contribute to improving crisis monitoring and help fact-checkers in the future.

Anja Bechmann, Professor på medievidenskab og direktør for DATALAB - Center for Digital Social Research. Foto: Aarhus Universitet

What can nearly 60 million tweets across the four largest Nordic countries tell us about our emotional responses on Twitter during the Covid-19 crisis? And what role do different communication environments on Twitter play in this context?

A group of researchers from DATALAB – Center for Digital Social Research at Aarhus University have studied this using artificial intelligence. They have analysed tweets by users in Sweden, Norway, Denmark and Finland during the second wave of the Covid-19 crisis. More specifically, they have analysed the emotional content of hashtagged, non-hashtagged and threat-hashtagged (in this case hashtags related to misinformation) tweets with a view to assessing the presence of anger, fear, sadness, disgust, joy and optimism.

The results of the study, which was recently published in PLOS ONE, revealed certain patterns in emotional expressions across different types of tweets. Crisis-hashtagged tweets (#Covid-19) expressed more negative emotions, such as anger, fear, disgust and sadness, and fewer positive emotions, such as optimism and happiness, compared to non-hashtagged Covid-19 tweets. Furthermore, threat-hashtagged tweets expressed even more negative emotions and fewer positive emotions – compared to crisis-hashtagged tweets – and anger was found to have a particularly large effect.

The study therefore indicates that the amplification of emotions through the use of hashtags was particularly effective in tweets related to threats – in this connection misinformation – during the pandemic.

“With this new knowledge and by adding to the methods we use in our field to gain a more detailed understanding of different emotions and the combination of emotions, we can better inform crisis communication and intervention strategies,” says Anja Bechmann, professor of media studies and head of the research project.

Emotions as collective behaviour on social media
The point of departure of the study was a wish to investigate how Twitter users in Nordic high-trust societies expressed and amplified their emotions in response to the crisis. In particular, how ambient affiliations and hashtag use shaped emotional expressions during the pandemic, Bechmann points out and continues:

“The aim was to try to understand the dynamics of collective behaviour on social media, and particularly on Twitter, during the pandemic. How emotions on Twitter can be a crisis imprint in Nordic high-trust societies.”

“More specifically, it’s about being able to use this knowledge and model to monitor emotions in the future, as the model predicts emotions in tweets. The model allows fact-checkers, for example, to look at which tweets they should pay extra attention to, because they are more likely to spread and harm citizens and society,” she points outs.

Recognising emotions in national languages
In order to make such research breakthroughs with regard to collective emotional expressions, the researchers had to move away from the most prevalent method: investigating whether communication has a negative or positive valence, explains Anja Bechmann:

“What’s different about this research article is that existing research has primarily focused on sentiment and whether something has a positive, negative or neutral valence. However, we worked with 11 specific emotions and their intensity.”

In addition, the AI model, which normally works in English, was trained by the researchers to be sensitive to the four Nordic languages and recognise emotional expressions in the national languages: “We need stronger models which can handle small languages without translating them into English,” she continues.

With or without hashtags
Another aim of the project was to contribute to a greater understanding of the differences between hashtagged and non-hashtagged online communication and thereby also to basic research in the field, Anja Bechmann points out:

“To understand whether it makes sense for researchers to continue to conduct hashtag analyses, we need to change the way we see Twitter and other platforms. How does the collective behavioural imprint differ in the various ways of communicating?” Anja Bechmann asks and makes a suggestion:

“Most Twitter studies are hashtag analyses because they are most easily accessible, but if we are to investigate how Twitter can be used in crisis management and whether it’s possible to predict crises, we also need to look at the feeds that don’t have hashtags. Social media communication generally takes place without the use of hashtags.”

Better crisis management in the future
As Twitter (now X, ed.) continues to serve as a central medium for the public exchange of information, understanding the emotional dynamics is crucial for promoting resilience and handling misinformation during crises such as the Covid-19 pandemic, according to Anja Bechmann, head of both the EU project Nordic Observatory for Digital Media and Information Disorder (NORDIS) and the Independent Research Fund Denmark project Social Media Influence, which are the projects behind the article.

“The fact that the model is available now will make it easier to respond to some of the things that are being spread if they are considered inappropriate for society during a crisis,” she says and continues:

“There’s a strong need to provide fact-checkers and other stakeholders with some tools that can help predict what will go viral and what won’t, and what could be harmful to society and individuals.” 

The research results – more information

Link to the research article:

Emotions on Twitter as crisis imprint in high-trust societies: Do ambient affiliations affect emotional expression during the pandemic

Type of study: 
Interdisciplinary research team with backgrounds in neuroscience, sociology, cognitive science, biophysics and media science.

External partners: 
One fact-checker from each of the Nordic countries is part of the NORDIS consortium, and the model is now freely available in the European Digital Media Observatory, which also includes certified fact-checkers across Europe.

Independent Research Fund Denmark 
SHAPE (internal funding by Aarhus University) 

For further information: 

Professor Anja Bechmann
Email: anjabechmann@cc.au.dk
Phone: +4551335138