davanstrien/gahd

---
license: cc-by-4.0
task_categories:
- text-classification
language:
- de
pretty_name: GAHD
configs:
- config_name: default
data_files:
- split: train
path: "data/gahd.csv"
- config_name: gahd_disaggregated
data_files:
- split: train
path: "data/gahd_disaggregated.csv"
---

**NOTE** README copied from https://github.com/jagol/gahd

This repository contains the dataset from our NAACL 2024 paper "Improving Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a German Hate Speech Dataset".

`gahd.csv` contains the following columns:
- `gahd_id`: unique identifier of the entry
- `text`: text of the entry
- `label`: `0` = "not-hate speech", `1` = "hate speech"
- `round`: round in which the entry was created
- `split`: "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`
- `model_prediction`: label predicted by the target model, `0` or `1`
- `annotator_id`: unique identifier of the annotator that created the entry
- `annotator_labels`: a string containing a forward slash-separated list of all labels by annotators
- `expert_labels`: `0` or `1` if an expert annotator annotated the entry, otherwise empty

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}
}
```

原始信息汇总

数据集概述

基本信息

• 许可证: CC-BY-4.0
• 任务类别: 文本分类
• 语言: 德语
• 数据集名称: GAHD

配置详情

• 默认配置

  • 数据文件: data/gahd.csv
  • 分割: 训练

• gahd_disaggregated配置

  • 数据文件: data/gahd_disaggregated.csv
  • 分割: 训练

数据集内容

• gahd.csv

  • 列信息:
    • gahd_id: 唯一标识符
    • text: 文本内容
    • label: 标签 (0: "非仇恨言论", 1: "仇恨言论")
    • round: 创建轮次
    • split: 分割类型 ("train", "dev", "test")
    • contrastive_gahd_id: 对比示例的gahd_id

• gahd_disaggregated.csv

  • 额外列信息:
    • source: 数据来源 (dynabench, translation, news)
    • model_prediction: 目标模型的预测标签 (0 或 1)
    • annotator_id: 标注者唯一标识符
    • annotator_labels: 标注者提供的标签列表，以斜杠分隔
    • expert_labels: 专家标注者提供的标签 (0 或 1)，否则为空

引用信息

• 论文: "Improving Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a German Hate Speech Dataset"
• 作者: Janis Goldzycher, Paul Röttger, Gerold Schneider
• 年份: 2024
• 预印本: arXiv:2403.19559
• 类别: cs.CL