Deep Learning Models for Predicting Political Tendency in Facebook Posts Incorporating Texts and Emojis

COVID-19 has brought massive challenges to the world, altering people's lives. We conducted a study on people's minds during Taiwan's National Epidemic 3-Level Alert in 2021. On May 22, 2021, during a regular press conference held by the Taiwan Centers for Disease Control (CDC), the Minister of Health and Welfare introduced the term ‘Retrospective Adjustment’, which left the entire population in shock. This study focuses on analyzing social media posts during the outbreak, specifically 6,022 selected Facebook posts that mention ‘Retrospective Adjustment’ between May 22, 2021, and May 25, 2021. Various models are utilized to classify the sentiment categories of these posts, considering both texts and emojis. We compare the performance of the classification models using posts with only texts and both texts and emojis. The LSTM and BiLSTM are suitable for processing posts containing texts and emojis. Conversely, the BERT model performs better when it includes only text. In the case of the BERT model with only text, the F1-score reaches 0.8 for the positive and objective posts. However, the BERT model does not perform well for negative posts. Our results indicate a lack of sensitivity towards the contextual effects of negation in the BERT model.