Modelling Cultural and Socio-Economic Dimensions of Political Bias in German Tweets

We introduce a new bi-dimensional classification scheme for political bias. In particular, we collaborate with political scientists and identify two important aspects: cultural and socioeconomic positions. Using a dataset of tweets by German politicians, we show that the new scheme draws more distinctive boundaries that are easier to model for machine learning classifiers (F1 scores: 0.92 and 0.86), compared to one-dimensional approaches. We investigate the validity by applying the new classifiers to the whole dataset, including previously unseen data from other parties. Additional experiments highlight the importance of dataset size and balance, as well as the superior performance of transformer language models as opposed to older methods. Finally, an extensive error analysis confirms our hypothesis that lexical overlap, in combination with high attention values, is a reliable empirical predictor of misclassification for political bias.