Policy actor alignments and social media discourse: a structural topic modeling approach to Twitter/X discussions on German legislation

Internet technologies such as social networking sites (SNSs) have become an integral part of socio-political communication (e.g., Jungherr, 2014; 2015). While the relevant literature mainly focuses on the role of SNSs in elections or protests, a political process-oriented view that considers online political debates is less common (Stier et al., 2017, pp. 19–11; Grossman, 2022). This paper examines the contestation of draft legislation on Twitter/X as a case of policy and professional discourse that enables policy-related interaction between actors such as politicians, interest groups and bureaucrats on a daily basis. It asks to what extent and how alignments between political actors known from the interest group literature (Beyers et al., 2015; Marshall & Bernhagen, 2020; Allern et al., 2021), e.g. between business interests and more right-wing politicians or public interests and more left-wing politicians (Eising et al., 2020, pp. 5–6; see INTERARENA classification of interest group actors by Binderkrantz et al., 2020), translate into systematic differences at the semantic level of the discussions. This study contributes to the literature on policy discourse (e.g., Leifeld, 2016) and political alignments not only by applying its hypotheses to the context of SNSs, but also by including individual professionals (such as science and engineering or health professionals, see European Skills, Competences, Qualifications and Occupations (ESCO) classification) as additional actors and their involvement in the policy negotiation enabled by online spaces. It is based on a retrieval of Twitter discussions (Academic Research Track of the Twitter API v.2) related to selected bills from different policy areas (Breunig, 2014), which were submitted by the German Federal Government to parliament in the most recent, 20th legislative period. Methodologically, the paper contributes in two ways: First, it uses a dataset of thousands of manually coded profile descriptions from a project on legislative Twitter discussions in the 19th legislative period and explores supervised machine learning to classify actor types based on the profiles of users who participated in legislative discussions on bills in the 20th legislative period. Second, it draws on Structural Topic Modeling (STM) (Roberts et al., 2014) and applies it to the analysis of tweets to topically describe the patterns and dynamics of bill-related discussions and to reveal the topical relationships between actor types by treating the latter as metadata (who mentions what topics?). In doing so, it addresses two problems that political scientists face in translating the analysis of discourse structures to the realm of online political discussion: On the one hand, computational methods for detecting semantic structures, such as semantic network analysis, focus on semantic co-occurrences, but are limited in integrating the types of actors on which political scientists rely heavily. On the other hand, methods that combine the analysis of actors and content, such as discourse network analysis (Leifeld, 2016), rely on intensive coding work and can be complemented by approaches that combine automated annotation of actors with topic modeling techniques.