A Deep Learning Approach to Fine-Grained Political Ideology Classification on Social Media Texts

Starting with the mid-2010s, the growing impact of social media on politics has been felt worldwide. This impact became even more decisive in instances such as the 2020 US Presidential Election, where a bitter division between the two main US political parties has resulted in an increasingly hostile and combative online discourse on social media platforms such as Twitter (now known as X). Nevertheless, this environment obscures under partisan discourse the actual ideological contents of the debates. We wish to present in this paper a novel dataset we have compiled from publicly available news websites exhibiting a certain political affiliation, as well as a deep learning classification pipeline trained on this dataset. Our best model, based on BERT, can reveal the rough political affiliation of a certain text, on a seven-point left-right political ideology scale, obtaining an F1 score of 90.33% on our validation dataset. We apply this classifier on a sample of 1.5M tweets retrieved from the #Election2020 dataset and perform an n-gram analysis of the classification results, using s-BERT embeddings and DBSCAN clustering to reveal topics of interest to communities of different political persuasions. We highlight the value of such an approach in defining which subjects are of most interest to each political camp and warn about the presence of polarized political discourse propagated through social media. We make the novel dataset available to other researchers as well.