Abstract
- We present a unified approach to detecting harmful content in German social media, developed for the GermEval 2025 Shared Task: Harmful Content Detection in Social Media (Felser et al., 2025). Our system addresses all three subtasks—(1) calls to action inciting harmful acts (C2A), (2) violence glorification (VIO), and (3) attacks on the free democratic basic order (DBO)—using a single ensemble-based framework. We fine-tuned a three-model ensemble (GBERT-large, XLMRoBERTa-base, and DeBERTa) for each subtask and aggregated their predictions through soft-voting. To mitigate severe class imbalance in the training data, we augmented the dataset with synthetic examples and manually relabeled instances for minority classes, and applied oversampling during training. This unified modeling approach achieved strong performance in the official evaluation: our system obtained macro-F1 scores up to 0.83, ranking 1 st in the DBO subtask, 3rd in C2A, and 4th in VIO.