A Deep Learning Approach for Ideology Detection and Polarization Analysis Using COVID-19 Tweets

Abstract

Polarization analysis is critical for effective policy and strategy implementation. Various aspects of the COVID-19 pandemic are discussed on social media platforms extensively. While social media are used to share factual information and official directives, there is also an abundance of misinformation and beliefs (both personal and political). Some of that misinformation and beliefs are driven by polarized opinions from different ideologies. Consequently, considerable polarization has been observed on widely discussed topics related to Covid-19 such as face masks and vaccines. The study of emotion is essential for polarization detection as positive or negative sentiment towards a topic might indicate favorability or hesitancy. While positive or negative sentiment indicates a polar view toward a subject matter, it is paramount to understand the fine-grained emotion (e.g. Happiness, Sad, Anger, Pessimism) for effective polarization detection. In this research work, we propose a deep learning model leveraging the pre-trained BERT-base to detect the political ideology from the tweets for political polarization analysis. The experimental results show a considerable improvement in the accuracy of ideology detection when we use emotion as a feature. Additionally, we develop a deep learning model accompanied by an adversarial sample generation module to detect the emotion in the tweets. The adversarial sample general module significantly improves the performance of the deep learning model. Finally, we explore the political polarization for the topics “mask” and “vaccine” in the different states of the USA throughout the pandemic.