Datasets:
metadata
license: cc-by-4.0
task_categories:
- text-classification
language:
- de
tags:
- hate-speech-detection
- hate-speech
pretty_name: GAHD
size_categories:
- 10K<n<100K
Dataset Card for GAHD
Dataset Description
GAHD is a German Adversarial Hate speech Dataset containing 10,996 examples. We collected the dataset via four rounds of Dynamic Adversarial Data Collection and explored various methods of supporting annotators in finding adversarial examples.
- Paper: https://arxiv.org/abs/2403.19559
- Repository: https://github.com/jagol/gahd
Dataset Structure
gahd.csv
contains the following columns:
gahd_id
: unique identifier of the entrytext
: text of the entrylabel
:0
= "not-hate speech",1
= "hate speech"round
: round in which the entry was createdsplit
: "train", "dev", or "test"contrastive_gahd_id
:gahd_id
of its contrastive example
gahd_disaggregated.csv
contains the following additional columns:
source
:- if annotators entered the entry via the Dynabench interface:
dynabench
- if the entry was translated from the Vidgen et al. 2021 dataset:
translation
- if the entry stems from the Leipzit news corpus:
news
- if annotators entered the entry via the Dynabench interface:
model_prediction
: label predicted by the target model,0
or1
annotator_id
: unique identifier of the annotator that created the entryannotator_labels
: a string containing a forward slash-separated list of all labels by annotatorsexpert_labels
:0
or1
if an expert annotator annotated the entry, otherwise empty
Citation
When using GAHD, please cite our preprint on Arxiv:
@misc{goldzycher2024improving,
title={Improving Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a German Hate Speech Dataset},
author={Janis Goldzycher and Paul Röttger and Gerold Schneider},
year={2024},
eprint={2403.19559},
archivePrefix={arXiv},
primaryClass={cs.CL}
}