---
license: cc-by-4.0
datasets:
- jagoldz/gahd
- Paul/hatecheck-german
language:
- de
metrics:
- f1
library_name: transformers
pipeline_tag: text-classification
tags:
- hate-speech-detection
- hate-speech
---

# Model Card

## Model Description

We fine-tuned this [gelectra-large model](https://huggingface.co/deepset/gelectra-large) for four rounds of dynamic adversarial data collection to create the GAHD dataset. In each round annotators created examples by trying to trick the model into a misclassification. We explored different ways of supporting annotators in finding model-tricking examples during the data collection. This is the final model (R4) in our paper. The model classifies text into "hate speech" (1) or "not-hate speech" (0).

Please check out our [paper](https://arxiv.org/abs/2403.19559) for further details about the training procedure (Appendix C) or evaluation (Section 4).

- paper: https://arxiv.org/abs/2403.19559
- GAHD dataset on Huggingface: https://huggingface.co/datasets/jagoldz/gahd
- GAHD dataset on GitHub: https://github.com/jagol/gahd

## Citation

When using this model or the GAHD dataset, please cite our paper:

```
@inproceedings{goldzycher-etal-2024-improving,
    title = "Improving Adversarial Data Collection by Supporting Annotators: Lessons from {GAHD}, a {G}erman Hate Speech Dataset",
    author = {Goldzycher, Janis  and
      R{\"o}ttger, Paul  and
      Schneider, Gerold},
    editor = "Duh, Kevin  and
      Gomez, Helena  and
      Bethard, Steven",
    booktitle = "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = jun,
    year = "2024",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.naacl-long.248",
    doi = "10.18653/v1/2024.naacl-long.248",
    pages = "4405--4424",
    abstract = "Hate speech detection models are only as good as the data they are trained on. Datasets sourced from social media suffer from systematic gaps and biases, leading to unreliable models with simplistic decision boundaries. Adversarial datasets, collected by exploiting model weaknesses, promise to fix this problem. However, adversarial data collection can be slow and costly, and individual annotators have limited creativity. In this paper, we introduce GAHD, a new German Adversarial Hate speech Dataset comprising ca. 11k examples. During data collection, we explore new strategies for supporting annotators, to create more diverse adversarial examples more efficiently and provide a manual analysis of annotator disagreements for each strategy. Our experiments show that the resulting dataset is challenging even for state-of-the-art hate speech detection models, and that training on GAHD clearly improves model robustness. Further, we find that mixing multiple support strategies is most advantageous. We make GAHD publicly available at https://github.com/jagol/gahd.",
}
```