New NORDIS preprint: Expressions of joy during the COVID-19 pandemic

Study by NORDIS partners at Aarhus University’s DATALAB detects joy, anger, but little fear in corpus of Nordic Twitter posts.

The following post provides the Executive summary of the report by the DATALAB team. The full report can be accessed here.

The COVID-19 pandemic has allowed for unprecedented studies of public emotions in times of crisis, in particular by using social media data to take the pulse of the population’s emotions. Many studies have highlighted the presence of fear and anxiety during the most uncertain times of the pandemic, especially in those countries most badly affected at the time (e.g., Italy, UK, China). But what happens when we turn our heads towards more resilient countries, such as some of the Nordic countries?

Do uncertainty and health concerns also translate into fear?

This is the first question we asked in our study, in which we bring the focus to four of the Nordic countries (i.e., Denmark, Norway, Finland and Sweden) that share a high level of trust in their democracies, the media, comprehensive healthcare, high levels of education and similar welfare state values.

We implemented a multilingual transformer model to detect emotions in a large sample of the Nordic Twitter, containing over 57 million tweets in Danish, Norwegian, Finish and Swedish, without limiting those tweets to be Covid19-related. The data was collected during the second wave of the pandemic, including a period before hospitalisation numbers started to rise up and until numbers had normalised again after reaching the peak in all countries (i.e., from August 2020 to March 2021).

Strikingly, our model did not detect any tweets containing fear for this time period in any of the languages present in our data. Of note, the model had been able to correctly detect most of the fear-related tweets in the validation data we used (SemEval 2018 dataset, subtask 5), annotated through a crowdsourcing project in which participants inferred the affectual state of a person from their tweet.

Report task 2.3 Predicting COVID-19 related collective anxiety on social media – report on scientific paper with executive summary

Does high trust equal less fear?

A possible explanation for the lack of fear-related tweets are the high levels of trust in the government, healthcare system and the media in the Nordic countries, which make citizens feel like the crisis is under control. In other words, at least to some degree, the high resilience that the population from the Nordic countries expect from their governments, health care system and other institutions, prevent fear from being spread, and even expressed often, on Twitter.

First, the results are produced by a model, with the biases and limitations that derive from the data it was trained on. For one, the data used for the finetuning of the model on the task of emotion detection came from English tweets translated using automatic tools into the Nordic languages of this study, with the loss of information inherent to this process. Also, the training dataset was unbalanced in the amount of emotions inferred from those tweets (e.g., many tweets expressed joy and anger, but not so many expressed sadness), leading to further difficulties in detecting those emotions that were underrepresented – a well known limitation ubiquitous to machine learning algorithms.

Even with these factors in consideration, the results seem to indicate that the amount of fear is much lower in the Nordic Twitter than we would have expected based on results from other countries, but that is not to say that people in the Nordic countries experienced no fear during the pandemic. As we consider the Twitter culture and the profile of Twitter users in the Nordic countries, it also seems unlikely for our data to be an accurate representation of the Nordic population on its own. Complementing our dataset with Facebook data might show different results in terms of the amount of fear being expressed online. Unfortunately, Facebook data is not as widely accessible for academic research as Twitter data.

Anger versus joy

The second question we ask in our study is, if fear is not a predominating emotion in the Nordic Twittersphere during Covid-19, what is? And is that different between the Nordic countries taking part in this study? From our results, we infer that the two main emotions during that period of time were joy and anger in any of the Nordic countries.

While it might be tempting to draw conclusions regarding the presence of joy in what are believed to be among the happiest countries on Earth, a word of caution comes from previous research showing that high-arousal emotions, such as anger and joy, tend to become viral and therefore be overrepresented on social media in contrast to data sampled through other methods (e.g., surveys). And as mentioned above, our model was more able to detect some emotions than others, with joy and anger being the ‘easiest’ emotions for the model to recognise.

We further assessed differences between the countries of study, finding that joy is always the predominant emotion detected across countries, although the difference between the amount of joy and fear is much larger in Finland than in the other countries. This especially goes for Sweden, where the amounts of joy and anger are closest to a balance. However, this analysis was conducted under the assumption that the model performs equally well across languages, but further research would be needed in order to verify this assumption.

Emotions’ evolution over time

Lastly, we focused on the fluctuations of joy across time, to assess potential relationships with variables indicative of the pandemic evolution, such as number of hospitalisations and time spent at home. We found no relationship between fluctuations in joy and neither of these two variables for any of the countries, in contrast to similar studies addressing these questions. Instead, we observed that spikes in joy could be found around national celebrations, such as Christmas and New years eve. Given the comprehensive data collection implemented for this study, which did not involve restricting data to Covid19-related tweets, a hidden relationship between these variables might become visible if we refine our dataset around Covid-19, removing unrelated tweets.

Future work

All in all, this study provides insights about the public emotions in the Nordic countries during the second wave of the pandemic. It does so through an exploratory analysis of – to our knowledge – the largest Nordic Twitter dataset collected during the Covid-19 crisis. Our findings highlight a lack of fear in the Nordic Twittersphere in contrast to what we would have expected based on the results from studies on data from other countries during the pandemic. Instead, we found that most tweets in our data expressed joy or anger, and that joy remained stable independently of the rising number of hospitalisations and increasing number of hours spent at home. This was the case across all Nordic countries studied in this project. Future research could address the generalizability of our findings across social media platforms, and add complementary information through data from other countries during this same time period, in order to see whether additional research supports our preliminary findings.

Leave a Reply

%d bloggers like this: