This paper summarizes our participation in the Harmful Content Detection in Social Media shared task. The task is divided into three subtasks: Call2Action (C2A), Attacks on the Democratic Basic Order (DBO), and Violence Detection (VIO). Following the organizers’ rec- ommendation, we used data augmentation to create more instances of each underrepresented category. To this end, we asked three large lan- guage models to rewrite a tweet in six different styles. We used a German BERT model for classification. The macro F1 scores obtained for the three subtasks were 79 % for C2A, 62 % for DBO, and 75 % for VIO.     «

This paper summarizes our participation in the Harmful Content Detection in Social Media shared task. The task is divided into three subtasks: Call2Action (C2A), Attacks on the Democratic Basic Order (DBO), and Violence Detection (VIO). Following the organizers’ rec- ommendation, we used data augmentation to create more instances of each underrepresented category. To this end, we asked three large lan- guage models to rewrite a tweet in six different styles. We used a German BERT model for class...     »