In the first phase of SPP 1999, the project MARDY (Modeling Argumentation Dynamics in Political Discourse) brought together machine learning methods from Natural Language Processing and theoretically grounded analysis frameworks from Political Science in order to better understand how actors in political debates (politicians, parties, demonstrators) articulate their positions by making certain claims (e.g., "migrants should be integrated into the labor market"), how they form ’discourse coalitions’ to achieve shared goals, and how debates evolve over time. To make this vision feasible, MARDY1 limited its focus to the study of a single topically confined debate at a time, namely the German national debates on migration and on pensions, using data from one newspaper source, and restricting network modeling to actors and claims. The interdisciplinary collaboration turned out to be highly successful, showing that the analysis of long-running political debates can indeed be sped up considerably while maintaining quality, and yielding new types of insights. Yet, the necessary limitations restrict the general applicability of the project’s results, and form the starting point for the second phase of MARDY.

The goal of MARDY2 is to generalize in two crucial dimensions. First, to substantially generalize both the methodology and statistical models of MARDY to enable not only the analysis of new debates on arbitrary new topics, but also cross-national (and therefore, cross-lingual) comparisons. This capability will be leveraged to analyse the ongoing debate on COVID-19 and its policy implications in Germany, France, the UK, and the US. Second, to systematically include frames into the models. Frames are argumentative patterns used to substantiate a claim and add an extra level of structure. For example, the claim that migrants should be integrated into the labor market can justified by left-wing actors by referring to concerns for general societal integration, or by center/right wing actors by referring to economic gains. Modeling the dynamics of actors, claims, and frames concurrently creates additional leverage for machine learning based detection and classification of actors’ statements. To take full advantage of the benefits of frames, we extend our research approach to include longer argumentative texts.

Not only can the broadened analytical scope lead to enriched corpus analyses in (computational) social science, it also helps to push the methodological development to a more systematic level, working towards generic workflows and tools supporting the analysis of argumentation in any given complex discourse setting.