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

Modelling Cultural and Socio-Economic Dimensions of Political Bias in German Tweets Aishwarya Anegundi1 and Konstantin Schulz1 and Christian Rauh2 and Georg Rehm1 1 Deutsches Forschungszentrum für Künstliche Intelligenz GmbH (DFKI), Berlin, Germany aishwarya.anegundi@dfki.de, konstantin.schulz@dfki.de, georg.rehm@dfki.de 2 Wissenschaftszentrum Berlin für Sozialforschung (WZB), Berlin, Germany christian.rauh@wzb.eu Abstract We introduce a new bi-dimensional classifica- tion scheme for political bias. In particular, we collaborate with political scientists and iden- tify two important aspects: cultural and socio- economic 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 clas- sifiers (F1 scores: 0.92 and 0.86), compared to one-dimensional approaches. We investi- gate the validity by applying the new classi- fiers to the whole dataset, including previously unseen data from other parties. Additional ex- periments highlight the importance of dataset size and balance, as well as the superior per- formance of transformer language models as opposed to older methods. Finally, an exten- sive error analysis confirms our hypothesis that lexical overlap, in combination with high atten- tion values, is a reliable empirical predictor of misclassification for political bias. 1 Introduction Political radicalization is linked to a society’s sense of insecurity (Bartoszewicz, 2016). Such a feel- ing may arise especially in times of crisis, such as financial crashes, large migration movements, or pandemics. In this setting, citizens’ trust in a country’s government or into the political system more generally can decline quickly (Easton, 1975; Dostal, 2015), leading to further radicalization. The effects of such a development are visible not only in terms of elections (Funke et al., 2016; Recuero et al., 2020) and media coverage (Bender et al., 2021), but also in general public political discourse and corresponding language use: Politi- cally biased texts tend to exhibit a wording that is different from their neutral counterparts (Krestel et al., 2012; Fairbanks et al., 2018). At times, this lexical deviation is hard to detect because the texts are positioned in seemingly neutral environments like technological or scientific sections of a news- paper (Kang and Yang, 2022). Furthermore, there are additional factors beyond wording: The filter- ing and selection of information to be presented in a given spot is a bias in its own right, but can directly affect or reflect political discourse: Pre- senting quotes by famous hyperpartisan politicians often serves as a subtle disguise for an author’s own political motives (Fan et al., 2019). Besides, the media coverage of political parties or crime-related ethnical aspects is indicative of the current govern- ment, the popularity of specific parties (Lazaridou and Krestel, 2016) and the trust in the executive’s impartiality (Pfeiffer et al., 2018). By training language models on such tenden- tious texts, we tend to reproduce and spread their bias (Bender et al., 2021), even if the resulting mod- els are used in rather neutral contexts (Liu et al., 2021). Since political bias (PB) is closely related to credibility (Su et al., 2020; Vargas et al., 2020; Ak- senov et al., …

Content truncated to 500 words