Who is Polluting the Debate? Tracing Back Hate Speech to User Groups in Facebook Communication during the German Federal Election Campaign 2017

Social media are broadly seen as drivers of polarisation and radicalisation. At first glance, this view is much supported by the daily observation of offensive and discriminatory speech towards individuals or groups, so-called hate speech, in online communication and social media. According to current debates, hate speech apparently has considerable detrimental effects on the quality of political discourse. However, as empirical studies show for political online discourse and campaigning, it is only a minority of users frequently commenting on party political posts. Hate speech in particular seems to originate in the activity of a relatively small number of hyperactive users. This observation would be in line with the assumption that we nowadays see new kinds of strategic information operations exerted by either particularly active users that more or less intentionally draw online discourses into an offensive and conflictual direction or by supposedly automated activity (so-called social bots). In the German context, it is of particular interest whether the newcomer in the party political system, the right-wing populist party Alternative für Deutschland (AfD) and its affiliates have an overall negative impact on discourse quality in online campaigns. In this paper, we apply Explainable AI approaches for hate speech detection to different sub-sets of a large dataset of Facebook communication, in order to learn more about the sources of hate speech in political online discourse. Can we trace back hate speech to certain user groups? The complete dataset consists of 2.9 mio. comments users made on the public Facebook pages of major German parties and their leading candidates in the period of the German federal election campaign 2017 (Jan-Sep 2017). For hate speech detection, we use a dictionary approach with the dictionary composed of elements from Ruppenhofer’s German Twitter embeddings and further elements added manually (2052 words in sum). We developed a fine-grained annotation scheme, including intensity and type of offensive speech. The presentation will focus on this current study as work in progress.