Sentiment Analysis of Finnish Twitter Discussions on COVID-19 During the Pandemic

AbstractWith the outbreak of the COVID-19 pandemic, researchers have studied how people reacted on social media during the pandemic. Sentiment analysis has been leveraged to gain insight. However, much of the research conducted on both sentiment analysis and social media analysis of COVID-19 often focuses on widespread languages, such as English and Chinese. This is partly due to the scarcity of resources for natural language processing and sentiment analysis for morphologically complex and less prevalent languages such as Finnish. This paper aims to analyze sentiments on Twitter in the Finnish language during the COVID-19 pandemic. We manually annotate with sentiments a random sample of 1943 tweets about COVID-19 in Finnish. We use it to build binomial and multinomial logistic regression models with Lasso penalty by exploiting ngrams and two existing sentiment lexicons. We also build two similar models using an existing (pre-COVID-19) Twitter dataset for comparison. The best-performing model for the Finnish language is then used to determine the trends of positive, negative, and neutral opinions on a collection of tweets in Finnish extracted between April 21 and June 18, 2020. The best sentiment polarity prediction model for the Finnish language attain 0.785 AUC, 0.710 balanced accuracy, and 0.723 macro-averaged F1 for predicting positive and negative polarity (binomial classification), and 0.667 AUC, 0.607 balanced accuracy, and 0.475 F1 when adding neutral tweets (multinomial classification). On the other hand, the pre-COVID-19 model trained on the same number of tweets exhibits higher accuracy for the multinomial model (0.687 balanced accuracy, and 0.588 F1). We hypothesize that this loss of performance is due to the COVID-19 context that makes sentiment analysis of neutral tweets more difficult for the machine learning algorithm to predict. Running the model on all the extracted Finnish tweets, we observe a decrease in negativity and an increase in positivity over the observed time as the Finnish government lifts restrictions. Our results show that applying an existing general-purpose sentiment analyzer on tweets that are domain-specific, such as COVID-19, provides lower accuracy. More effort in the future needs to be invested in using and developing sentiment analysis tools tailored to their application domain when conducting large-scale social media analysis of specific medical issues, such as a global pandemic.