Unmasking Online Polarization : Automated Detection of Topics and Stances in Social Networks

Social media has emerged as a primary source for accessing news due to its capacity to swiftly disseminate information from a myriad of sources, often without stringent filtration. This accessibility is particularly beneficial during exigent circumstances, affording individuals diverse perspectives on unfolding events. Consequently, a growing number of individuals rely on social media alongside traditional news outlets. However, the nature of information transmission within social media platforms fosters an echo chamber effect, wherein users are exposed predominantly to content aligned with their existing beliefs, leading to several deleterious consequences. Primarily, social media exacerbates societal divisions by amplifying pre-existing ideological segregation. Moreover, the susceptibility of social media content to manipulation renders it a potent tool for the dissemination of misinformation and propaganda, a concern underscored by numerous scholars. This dissertation delves into the phenomenon of social network polarization at a multi-level. The fist study examines social network polarization through the lens of activity generated by social bots. The second study investigates the relationship between social network polarization and external influences such as governmental announcements and vaccine availability in Kuwaiti twitter dataset. Building upon these macro-level analyses, the dissertation introduces methodologies for micro-level diagnosis of social network shifts, utilizing tweet-level textual analysis. Lastly, a masked aspect-based stance detection model is developed using weakly labeled datasets. This model facilitates the expedient prediction of individuals' stances on specific topics, offering a pragmatic alternative to labor-intensive human labeling systems. Through these multifaceted analyses and model developments, this research aims to provide insights into the detection of stances within real-world social network datasets, contributing to the understanding of and ability to navigate social media polarization.