Polarization on social media: Comparing the dynamics of interaction networks and language-based opinion distributions

When people share information and converse on social media, they create “echo chambers” through preferential attachment to like-minded people and opinions they already support. A great deal of research uses interactional ties between people—created by retweeting and following—to identify and study polarization in social networks. Some of this work then uses language analysis to characterize the opinions and concerns of subcommunities in the network. We used machine learning to create “speaker landscapes” that can identify polarization in user language (in tweets about COVID-19 vaccination) independently of the social networks created by user interactions via retweeting. In contrast to the prevailing assumptions, we found that distances between users in interaction networks did not predict their language similarity very well. We then compared the effect of a polarizing event (the declaration of the COVID-19 pandemic) on polarization between communities in the retweet networks and the speaker landscapes. We found that retweeting was done both to support and to criticize claims and that polarization between Democrats and Republicans emerged much more strongly in the speaker landscapes than the interaction networks. The results suggest that different cognitive-motivational dynamics affect who we interact with and what we say on social media, raising questions about how language is used to promote polarization.