Affective polarization, the emotional divide between members of different political groups, is a growing concern in democracies. It reduces cooperation, fuels hostility, and erodes trust in institutions. While social media is seen as an amplifier of this polarization, there is debate about its extent and impact. This study addresses this gap by quantifying affective polarization on social media through an analysis of emotional content and toxicity in reply interactions on Twitter. The research hypothesizes that in-group interactions will be positive, while out-group interactions will be negative and toxic, and that these emotions will vary systematically with network distance between users. The importance of the study lies in providing a better understanding of the complex dynamics of affective polarization in the digital age and its implications for political discourse. By analyzing large datasets of online discussions, the researchers aim to develop an instrument to measure affective polarization on social media, establish the empirical relationship between emotions and network structure, and analyze differences in information spread across ideologically divided groups. This goes beyond previous research which mainly focused on how people *talk* about out-group members, instead concentrating on how they *interact* with them.

Existing research highlights that affective polarization is influenced by factors such as news media, political elites, and demographics. While the role of social media in amplifying partisan animosity remains debated, many scholars recognize its significant influence. Social media discourse often promotes inflammatory language and moral outrage directed at opposing groups, and echo chambers exacerbate polarization by exposing users to extreme and divisive content. However, exposure to opposing views also occurs and might worsen polarization. This study builds upon this existing literature by focusing on the emotional dynamics of online interactions, leveraging advanced language models to analyze the emotional content and toxicity of conversations. It seeks to provide a more nuanced understanding of affective polarization by moving beyond the simple in-group/out-group dichotomy to consider the role of network structure.

The study utilized two large datasets of Twitter discussions: one on the COVID-19 pandemic and another on the overturning of Roe v. Wade. User ideology (liberal or conservative) was classified using a text-based model trained on data incorporating political bias scores from URLs shared by users. State-of-the-art transformer-based language models (Demux and Detoxify) were employed to measure emotions (anger, disgust, fear, sadness, joy, love, trust, pessimism, optimism) and toxicity in replies. Retweet networks were constructed to represent online social networks, with network distance measured using both shortest path length and Euclidean distance in an embedding space generated by LargeVis. Contentious issues within discussions were identified using a keyword-based approach and a transformer-based model. The researchers analyzed the spread of information by examining daily retweet volumes for each issue within each ideological group, calculating autocorrelation functions to characterize the dynamics of information diffusion. Statistical tests (Mann-Whitney U Test with Bonferroni correction and linear regression) were used to assess statistical significance.

The analysis revealed strong evidence of affective polarization. Out-group replies (interactions between users with opposing ideologies) exhibited significantly higher scores for anger, disgust, and toxicity compared to in-group replies. In-group replies, conversely, displayed higher joy scores, consistent with in-group favoritism and out-group animosity. These findings were consistent across both datasets. Interestingly, out-group replies were also significantly shorter than in-group replies, suggesting a focus on expressing animosity rather than conveying information. Furthermore, the study showed a strong correlation between emotions and network distance. Anger, disgust, and toxicity increased with network distance, while joy and fear decreased, demonstrating that affective polarization is a structural property of online networks. This held true regardless of partisan labels, implying that emotional responses are influenced by network proximity. The analysis of information spread revealed that the dynamics of information sharing differed across ideological groups and across time. Some issues exhibited random bursts of retweets, likely driven by news cycles, while others persisted over longer periods, highlighting the role of emotional salience in maintaining attention within polarized groups. This was evident in the COVID-19 dataset, where, for example, discussions about education were more persistent among liberals, while discussions about lockdowns were more persistent among conservatives. In the abortion dataset, the overturning of Roe v. Wade shifted the dynamics of information spread, with certain issues resonating more strongly within specific groups post-ruling.

The findings directly address the research questions by demonstrating the existence and pervasiveness of affective polarization in online networks. The observation that affective polarization extends beyond the in-group/out-group dichotomy, being influenced by network structure, provides a novel perspective on this phenomenon. The variation in information spread across groups depending on the emotional resonance of issues highlights the complex interplay of emotions, partisanship, and network dynamics. The results underscore the importance of considering emotional factors and network structures when analyzing online political discourse. The study's findings contribute significantly to understanding the mechanisms behind political polarization in the digital age, providing insights relevant to mitigating its negative consequences.

This study provides robust empirical evidence for affective polarization in online networks and its interaction with network structure and information spread. The findings demonstrate that emotions in online interactions are not simply a matter of in-group versus out-group dynamics, but also reflect the network distance between interacting users. The different dynamics of information spread across ideological groups highlight the role of emotional salience in shaping online discussions. Future research could explore how these dynamics are influenced by specific platform algorithms, the role of bots and other automated accounts, and the development of interventions to mitigate the negative effects of affective polarization.

The study's datasets, while large, represent a specific sample of online discussions and may not fully capture the diversity of online interactions. The accuracy of ideology classification, emotion detection, and issue identification could affect the results. The retweet network, while a commonly used proxy, might not fully capture the complexity of follower relationships. Finally, the study does not fully disentangle the directionality of the relationship between network distance and emotive expression.