Exploring group fairness in news media recommendations: Algorithms, metrics, and grouping

Author: Blake Huebner

Supervisor: Julia Neidhardt, Co-Supervisor: Thomas E. Kolb


The COVID-19 pandemic and the resulting government measures have triggered a wave of protests and demonstrations in Austria. We saw some protestors resorting to populist rhetoric to express their dissatisfaction, which in some cases, led to anti-democratic tendencies. Populist talking points have not been confined to the public sphere but rather emerged from social media and other online platforms, including news comments. Thus, there is a need to develop automated methods to detect populist statements in those texts. While previous research on populism has focused on politicians, recent studies have emphasized the role of citizens as populist actors. However, most of the current methods to detect populism in text rely on manual coding or dictionary-based approaches. Only a few scholars have attempted to employ machine learning for this task, resulting in a shortage of annotated data, particularly for the German language. To address this gap, this thesis adopts a minimalistic ideational definition of populism and performs various experiments using BERT-based transformer models to enhance the detection of populist user-generated content. Additionally, the thesis presents the first annotated dataset for populist news user comments in the German language by conducting an annotation study. The proposed model outperforms the state-of-the-art in this area and is applied in a case study analyzing the correlation between COVID-19 and populism in Austrian news user comments. A large-scale analysis of comments in the news forum of the Austrian daily newspaper Der Standard is conducted and the study reveals that the topic of COVID-19 in news articles attracted more populist comments than other topics during the pandemic. From these findings, implications can be drawn for future crisis management and communication.