Content and Network Feature in Attention-based Neural Network for Stance Detection on COVID-19 Vaccination Tweets

Abstract

Stance detection categorizes tweet posts based on user attitudes toward vaccines. The research implies that the government needs help educating people about the COVID-19 vaccine. The dataset is extracted using user metadata to obtain content and network features necessary to represent users' interest in tweets. The content parts representations of text based on the tweet, and the network features define support arguments in the tweet related to the user account following or unfollowing. Attention mechanisms are added in the neural network model to identify essential information in a sentence's stance. Two-word embedding models are used with Word2Vec and FastText to represent the text into vectors. The main contribution is employing a combination of tweet text and network features with a combination of Attention and Long-short term memory (AttLSTM) to determine stance detection. The results reported are AttLSTM and Word2Vec provided better classification results than FastText, which can catch the minority class defined as against stance inside text tweets. Future studies are recommended to employ topic modelling to classify documents by various aspect categories. They are also suggested to address the issue of imbalanced classes in the text data and enhance the current attention model. Additionally, this research can explore users who shift their stance from anti-vaccine to pro-vaccine.