Identifying Degree-of-Concern on COVID-19 topics with text classification of Twitters

The COVID-19 pandemic has various impacts on changing people’s behavior socially and individually. This study identifies the Degree-of-Concern topic of COVID-19 through citizen conversations on Twitter. It aims to help related parties make policies for developing appropriate emergency response strategies in dealing with changes in people’s behavior due to the pandemic. The object of research is 12,000 data from verified Twitter accounts in Surabaya. The varied nature of Twitter needs to be classified to address specific COVID-19 topics. The first stage of classification is to separate Twitter data into COVID-19 and non-COVID-19. The second stage is to classify the COVID-19 data into seven classes: warnings and suggestions, notification of information, donations, emotional support, seeking help, criticism, and hoaxes. Classification is carried out using a combination of word embedding (Word2Vec and fastText) and deep learning methods (CNN, RNN, and LSTM). The trial was carried out with three scenarios with different numbers of train data for each scenario. The classification results show the highest accuracy is 97.3% and 99.4% for the first and second stage classification obtained from the combination of fastText and LSTM. The results show that the classification of the COVID-19 topic can be used to identify Degree-of-Concern properly. The results of the Degree-of-Concern identification based on the classification can be used as a basis for related parties in making policies to formulate appropriate emergency response strategies in dealing with changes in public behavior due to a pandemic.