Programma della conferenza

The 2022 Brazilian presidential elections highlighted extreme polarization and widespread online misinformation, culminating in an attempted coup on January 8, 2023 (Bastos & Recuero, 2023). This study addresses the rise of destructive political polarization, characterized by dysfunctional communication driven by negative emotions and antagonism toward opposing political groups (Esau et al., 2024).

Emotions play a critical role in destructive polarization. Anger and mistrust often reinforce group boundaries, escalating polarization (Barnes, 2022; Recuero et al., 2021; Sandvoss, 2020). Social media platforms, particularly Facebook, contribute significantly to this phenomenon. Facebook’s affordances—reactions, comments, and shares—offer valuable metrics to assess these negative emotions (Anwar & Giglietto, 2024). Comments can indicate a willingness to engage, sometimes leading to incivility (McCosker, 2014), while 'angry' reactions often reflect frustration with opposing content (Eberl et al., 2020; Muraoka et al., 2021).

Our research expands on this by analyzing 'angry' and 'love' reactions alongside comments and shares, as well as circulated narratives as antecedents of amplification. While angry reactions and comments may reflect negative sentiments toward the 'other,' shares and 'love' reactions usually denote support and alignment (Eberl et al., 2020). We explore the following research question: What is the role of emotions in amplifying a hyperpartisan political post?

Method

This study is part of a European project that developed a news alert system to detect coordinated link-sharing by actors known for disseminating problematic information on Facebook (Giglietto et al., 2023). Our analysis of flagged links revealed consistent dissemination of content linked to Brazilian political pages and groups.

We mapped Facebook accounts sharing these links and applied a modularity algorithm to identify 58 coordinated pages and groups supporting Bolsonaro. We collected over 12 million posts shared by this network between January 1, 2021, and December 31, 2023. The quantitative analysis examined interactions (comments, shares, love, and angry reactions) to identify engagement trends. Time series analysis of love/angry and share/comment ratios helped assess emotional polarization over time, with notable fluctuations during political events. Seven periods of high volatility, mainly in 2021 and 2023, were identified, narrowing our focus to 1,161,126 posts.

We employed a qualitative grounded approach, analyzing 1,400 posts with the highest engagement during these volatile periods. We categorized 760 posts to fine-tune a Large Language Model (LLM), creating a classification scheme with three groups: target (Bolsonaro, his family, allies and supporters; The Supreme Federal Court and other public institutions; Armed forces / Military Police; Media Mainstream; Lula, his family, allies and supporters; Other), sentiment (positive, negative, neutral), and post typology (user-generated content, or news). This sample will be a gold standard for fine-tuning an OpenAI gpt-4o-mini model and building a regression model to assess how narratives and users’ emotions influence amplification.

Preliminary Findings

Our analysis focuses on three years of content shared by pro-Bolsonaro accounts (N=58). The angry/love and comment/share ratios range between 1 and -1, with values near 1 indicating dominant love reactions and shares and values near -1 showing a predominance of angry reactions and comments.

In 2023, following the attempted coup, these ratios shifted, with comments and angry reactions becoming more prominent. While 2022 remained stable, volatility analysis revealed significant instability in 2021 and 2023.

We will estimate a regression model using total interactions as a proxy for amplification, with narratives and emotional reactions (comments, shares, angry, and love) as independent variables. Control variables include timeframe, posting account, and content type.

By shedding light on the interplay between emotions, narratives, and content amplification, this study enhances our understanding of how destructive polarization evolves within digital ecosystems.