Identification of Political Leaning in German News Articles

Bagrat Ter-Akopyan

Abstract:Media coverage plays an important role in shaping personal and public opinion, as it is often a primary source of information. Especially in today’s world where ideological division with subsequent polarization of the political discourse is constantly increasing across social networks, the responsibility of media professional to ensure an unbiased reporting is growing. However, the coverage often exhibits a specific political leaning that is reflected in the articles and is commonly referred to as media bias . Many works already exist in computational social science that explore this phenomenon using English-language newspaper articles. This thesis aims to investigate the phenomenon in the German media landscape. This work uses a new methodology to attribute newspaper articles to a political leaning based on the party affiliation of the politician who shared the article on social media. We used AbgeordnetenWatch — a platform containing short profiles of politicians — to generate a list of German MPs including their Twitter profiles. We extract their tweets, using the Twitter API, and select all tweets that reference an article in the German media landscape. We extract all articles, including the title, the summary, the actual text and several meta information. In the end, we obtain a dataset with a total of over 40 thousand articles from the most pop-ular German media houses and the respective politician who shared the article. We examine the resulting dataset with respect to three research questions: (1) How effective are standard classification approaches with frequency-based features and the standard neural approach in detecting bias from the right-wing political fringe, when using the articles shared by all other parties as negative samples? (2) How effective are the same standard approaches in detecting bias from right-wing political fringe, when training against the articles shared by other parties separately? (3) How do the same approaches perform when applying to the detection of the left-wing party DIE LINKE , i.e.: a) when using the articles shared by all other parties as negative samples, analogous to (1); b) when training articles of DIE LINKE separately against articles of every other party, analogous to (2). The results of the experiments regarding the second and the third research questions are promising: the state-of-the-art model,

Political Science