Impact of COVID-19 on Trust in Government

• ANR
  XTCOVIF
• Semantic
  Shift
• Predicting the Spread
  of COVID-19
• Trust in
  Government
• Racism and
  COVID-19
• Population
  Mobility
• Publications and
  Datasets

Large scale analysis of mobility data and language during the COVID19 pandemic

Impact of COVID-19 on Trust in Government

Introduction

The main objective of the project is to study the impact of COVID-19 on the French society using data extracted from social media platforms. The following issue is of great value: Are people satisfied overall with how the French government has handled the COVID-19 crisis so far? And do they trust government to keep people safe?

The outbreak of COVID-19 caused a severe disruption to the economy, as well as to society as a whole. This had also a significant impact on the way of living of people is several countries including France. Governments around the world have responded differently to the COVID-19 pandemic. Some prioritized the economy, while others were concerned with the ability of the health system to meet people’s needs and with the high risk of deaths. For instance, in France, the government’s measures have been particularly stringent, while in Sweden, they were much looser. The pandemic is likely to have caused a decline of public trust in the government’s ability to manage risks. Therefore, in the context of this project, we utilized data from social media and investigated trust in French government. We should mention that similar studies have already been conducted in the past [Arunachalam and Sarkar, 2013; Corallo et al., 2015], but not in the context of COVID-19 and not for the French government.

Methodology

To deal with this task, we applied a two-step approach. First, given a corpus of tweets, we identified those tweets that are related to the government, i.e., talk about the government, comment on decisions and measures taken by the government, etc. Second, we applied a sentiment analysis algorithm to the extracted set of tweets to detect if the users of Twitter are positive or negative towards the government’s decisions.

To generate tweet representations, for both aforementioned steps, we used a similar pipeline. Specifically, our approach capitalizes on recent advances in representation learning in the field of natural language processing (NLP) and specifically on BERT [Devlin et al., 2019], a method for generating contextual word embeddings. Since its inception, BERT has brought a revolution to representation learning in NLP, and has led to impressive results in several NLP tasks such as machine translation, text categorization, and question answering. Due to its recent success, we decided to use BERT as our starting point. It is important to note that, as we worked with a French corpora, we used CamemBERT [Martin et al., 2020], a French variation of BERT, in order to produce appropriate word representations. Along with the contextual word representations, we also used some linguistic features, such as the length of a tweet.

Since the detection of tweets that are related to the government is treated as a supervised learning task, we had to create a dataset, i.e., annotate some tweets. To this end, we followed an automatic procedure. We first applied an unsupervised algorithm to cluster hashtags extracted from the tweets and identified a cluster that contained hashtags related to the government. We manually processed these hashtags to ensure that they are indeed related to the government and we then made the assumption that the tweets that contain those hashtags talk about the government. Some examples of those hashtags include #macronavirus, #élysée, #déconfinement, etc. We also generated some negative samples (i.e., tweets not related to the government) by randomly sampling a number of tweets from the corpus.

The dataset was split into a training and a test set, and all those tweets were mapped into vectors using the approach described above, and were fed to a logistic regression classifier. The classification of the employed approach was satisfactory. Specifically, the classifier achieved an accuracy of 79% and an F1-score of 78%, thus providing us with a reliable approach to find tweets that concern the government.

Given the set of government-related tweets, the second step of the employed approach involved analyzing the sentiment of those tweets. As already mentioned, we again capitalized on BERT and its sentiment classifier to study the overall sentiment of French users of Twitter towards the government. Note that BERT has been pre-trained to identify sentiment, and thus we only needed to feed the tweets to the model. However, the model has been trained on movie reviews from the Allociné dataset. Note also that we experimented with different thresholds (the value at which the model would consider a tweet positive or negative) in order to also get an idea of the sentiment extremity contained in those tweets.

Results & Discussion

As already discussed, the classifier that predict if a tweet talks about the government achieved an accuracy of 79% and an F1-score of 78%. With regards to the second task (i.e., sentiment analysis), the results are shown in the following Table.

We can see that not only is the general sentiment towards the government is negative and this becomes more evident as the value of the threshold increases.