Model Card for German Hate Speech Classifier

Model Details

Introduction

This model was developed to explore the potential of German language models in multi-class classification of hate speech in German online journals. It is a fine-tuned version of the GBERT model from (Chan, Schweter, and Möller, 2020).

Dataset

The dataset used for training is a consolidation of three pre-existing German hate speech datasets:

• RP (Assenmacher et al., 2021)
• DeTox (Demus et al., 2022)
• Twitter dataset (Glasenbach, 2022)

The combined dataset underwent cleaning to minimize biases and remove redundant data.

Performance

Our experiments delivered promising results, with the model reliably classifying comments into:

• No Hate Speech
• Other Hate Speech (Threat, Insult, Profanity)
• Political Hate Speech
• Racist Hate Speech
• Sexist Hate Speech

The model achieved a macro F1-score of 0.775. However, to further reduce misclassifications, improvements are essential. Short comments are overproportionally classified as Sexist Hate Speech.