Title: Detection of Increased Time Intervals of Anti-Vaccine Tweets for COVID-19 Vaccine with BERT Model

Abstract: The most effective of the solutions against Covid-19 is the various vaccines developed. Distrust of vaccines can hinder the rapid and effective use of this remedy. One of the means of expressing the thoughts of society is social media. Determining the time intervals during which anti-vaccination increases in social media can help institutions determine the strategy to be used in combating anti-vaccination. Recording and tracking every tweet entered with human labor would be inefficient, so various automation solutions are needed. In this study, The Bidirectional Encoder Representations from Transformers (BERT) model, which is a deep learning-based natural language processing (NLP) model, was used. In a dataset of 1506 tweets divided into four different categories as news, irrelevant, anti-vaccine, and vaccine supporters, the model was trained with a learning rate of 5e-6 for 25 epochs. To determine the intervals in which anti-vaccine tweets are concentrated, the categories to which 652840 tweets belong were determined by using the trained model. The change of the determined categories overtime was visualized and the events that could cause the change were determined. As a result of model training, in the test dataset, the f-score of 0.81 and AUC values for different classes were obtained as 0.99,0.91, 0.92, 0.92, respectively. In this model, unlike the studies in the literature, an auxiliary system is designed that provides data that institutions can use when determining their strategy by measuring and visualizing the frequency of anti-vaccine tweets in a time interval, different from detecting and censoring such tweets.