Monitoring public concern in the surrounding environment to certain events is done to address changes in public behavior individually and socially. The results of monitoring public attention can be used as a benchmark for related parties in making the right policies and strategies to deal with changes in public behavior as a result of the COVID-19 pandemic. Monitoring public attention can be done using Twitter social media data because the users of the media are quite high, so that they can represent the aspirations of the general public. However, Twitter data contains varied topics, so a classification process is required to obtain data related to COVID-19. Classification is done by using word embedding variations (Word2Vec and fastText) and deep learning variations (CNN, RNN, and LSTM) to get the classification results with the best accuracy. The percentage of COVID-19 data based on the best accuracy is calculated to determine how high the public's attention is to the COVID-19 pandemic. Experiments were carried out with three scenarios, which were differentiated by the number of data trains. The classification results with the best accuracy are obtained by the combination of fasText and LSTM which shows the highest accuracy of 97.86% and the lowest of 93.63%. The results of monitoring public attention to the time vulnerability between June and October show that the highest public attention to COVID-19 is in June.