Plenary debates in parliament represent a democratic ideal. They are the forum in which parties discuss their policy positions. Government parties use them to explain and justify their policies to the public, while opposition parties can scrutinise them. In these debates parties present their positions in front of their fellow parliamentarians but also the wider public. But which topics are members of parliament talking about in plenary debates and who talks about what? This paper presents a new annotated text corpus of the plenary debates of the German Bundestag between 1990 and 2013 and discusses the methods used to automatically classify each of the speeches into a policy topic. Each speech has been assigned to one ministerial portfolio using supervised classification methods. Such methods outperform simple topic models, as they give more steering capacity to researchers and the results can be evaluated against hold-out data. But they come with the obstacle that they require large amounts of training data which normally need to be created in time-consuming work by human coders. By using minister speeches as training data the paper proposes a time-efficient solution to this problem. Special attention is being devoted to evaluating the validity of the newly produced data. The paper’s contribution is thus twofold. First, it proposes a new way of annotating a large parliamentary text corpus that can be easily adapted to other countries or time frames. Second, it presents a new data set that offers many possibilities for research on the behaviour of parliamentarians.