Political parties emphasis different issues in their public communication efforts to address the topics of the day and to strengthen their policy profiles. In this article, we develop a dynamic supervised machine learning model to measure parties' evolving issue agendas from press releases. The model combines a language model for the classification of texts into pre-specified issue categories and a dynamic hierarchical model for the evolution of each parties' emphasis of these issue-categories in published press-releases. We estimate the model jointly on the labelled and unlabeled press-release data to evaluate the overall issue agendas. A simulation study shows that this model outperforms existing approaches to estimate category prevalence in text corpora when issue agendas evolve smoothly over time. An application to press releases of German and Austrian parties from 2012-2018 illustrates the usefulnesses of our approach in studying dynamic party competition, communication and behaviour.