We fit a Latent Dirichlet Allocation (LDA) topic model to COVID-related tweets from English-language newspapers, one country at a time. Tweets are preprocessed (special characters removed, words stemmed), and both unigrams and bigrams are included. Case counts come up again and again among the top terms. That is one reason the paper sorts countries by their share of COVID-19 cases.
Each circle on the left is a topic, sized by how common it is. Click a topic to see its most relevant terms on the right. The slider sets λ, which trades off a term's probability within the topic (λ = 1) against its lift (λ = 0). The paper uses λ = 0.5. Built with LDAvis (Sievert and Shirley, 2014); five topics per country.