Classifying Anti-Mask Tweets into Misclassification vs. Rejection: A Year-Long Study

Abstract

The debate over masks has played out vigorously over social media platforms such as Twitter over the course of the Covid-19 pandemic. Anti-maskers oppose the use of face masks on two philosophical grounds. First, they question their effectiveness and second, they reject them as an infringement of their personal liberties and freedoms. Both these narratives can be damaging in their own respective ways; misinformation can mislead people to abandon this simple public health measure, and rejection can incite unrest, disobedience and violence. Different policies, ranging from completely removing the tweet to simply placing a warning label, may be applied to these two types of anti-mask tweets to mitigate their damage. To facilitate these differentiated policy decisions, driven by the state of the pandemic and the surrounding social and political circumstances, this paper proposes a machine learning approach to separate anti-mask tweets into misinformation and rejection. Linguistic, social, auxiliary, and sentiment features are extracted from this corpus of tweets collected over the first year. A combination of these features is used to train ensemble and neural network classifiers. The results show that our machine learning framework can separate between misinformation and rejection tweets with a F1-score of around 0.90. These results are noteworthy because the framework can classify between two groups of tweets that share a common overall theme of anti-masking yet have only subtle differences. Moreover, the data collected over a period of one year implies that this separation is achieved even when the anti-masking rhetoric is embedded in widely varying social and political contexts.