Text mining methods for measuring the coherence of party manifestos for the German federal elections from 1990 to 2021

Text mining is an active field of statistical research. In this paper we use two methods from text mining: the Poisson Reduced Rank Model (PRR, see Jentsch et al. 2020; Jentsch et al. 2021) and the Latent Dirichlet Allocation model (LDA, see Blei et al. 2003) for the statistical analysis of party manifesto texts from Germany. For the nine federal elections in Germany from 1990 to 2021, we analyze party manifestos that have been written by the parties to present their political positions and goals for the next legislative period of the German federal parliament (Bundestag). We use the models to quantify distances in the language of the manifestos and in the weight of importance the parties attribute to several political topics. The statistical analysis is purely data driven. No outside information, e.g., on the position of the parties, on the meaning of words, or on currently hot political topics, is used in fitting the statistical models. Outside information is only used when we interpret the statistical results.