The last speakers in this Bots Building Bridges workshop session are Felix Gumbert and Rob Ackland. Felix starts by outlining three hypothesis about political talk on social media: first, social media might provide a space for productive deliberation; second, social media might serve as a hostile environment where constructive deliberation is impossible; and third, social media might create isolated communication environments (‘echo chambers’, ‘filter bubbles’) were people with different views no longer even encounter each other.
Empirical research on these possibilities tend to employ a classical sender-receiver model of communication, and utilise fairly simplistic datasets of online communication (e.g. hashtag collections); reciprocity is then often assessed at the level of posts or pairwise interactions. Research on ‘echo chambers’ often also employs network analysis to identify levels of homogeneity or cross-ideological interaction.
This tends to ignore the extent to which multi-turn conversations between participants are present in the data. There is a need to introduce more sophisticated approaches to conversation analysis into the methodological toolkit, for instance by identifying reply trees in lengthy conversations and detecting instances of agreement or disagreement, especially when they occur repeatedly between the same participants.
Felix’s study explored this for a dataset from a 2020 US presidential debate, gathered through the #debate2020 hashtag and extended by gathering the full conversations attached to the tweets in the datasets. This identified some 2.5 million reply trees; demonstrating the limitations of hashtag-only collections, the initial hashtag dataset itself represents only 1% of this full dataset.
The project then took a non-random sample focussing on longer reply trees, working with some 215 reply trees in total. 75% contained at least one uncivil turn (40% of them two consecutive uncivil turns); 65% contained at least two consecutive disagreement turns. The length of reply chains was a predictor for greater disagreement, but not for greater incivility.
Reply chains were then clustered for their similarity in terms of agreement and incivility turns within the discussion; it is more likely that expressions of agreement are followed by discussions than that discussions are followed by agreement. Some discussions continue despite, some even because of the presence of incivility; sometimes incivility terminates discussions early on; sometimes it gradually occurs more prominently further downthread.
Rob now zooms out further from individual reply chains to the overall reply trees that they are embedded in. It is possible to classify these by their depth of argumentation (length of chains of reciprocal engagement) and width of representation (number of diverse branches from the early posts); ideally, deliberation features both: turn-taking argumentation between participants, and broad representation of a range of perspectives.
Reply trees can then be classified by which combination of these two features is most prominent here; it is then also possible to explore the extent to which highly partisan actors from the left and/or right of US politics were present here. Most trees in the sample were more representative than argumentative, and only 11% mixed both features.
This approach to the study and classification of interaction sequences can make a significant contribution to the analysis of deliberative processes in online environments; it is especially well suited to platforms that support complex reply trees (Twitter, Reddit; not necessarily Facebook or YouTube).
