German Hate Speech Superset

This dataset is a superset (N=50,545) of posts annotated as hateful or not. It results from the preprocessing and merge of all available German hate speech datasets in April 2024. These datasets were identified through a systematic survey of hate speech datasets conducted in early 2024. We only kept datasets that:

are documented
are publicly available
focus on hate speech, defined broadly as "any kind of communication in speech, writing or behavior, that attacks or uses pejorative or discriminatory language with reference to a person or a group on the basis of who they are, in other words, based on their religion, ethnicity, nationality, race, color, descent, gender or other identity factor" (UN, 2019)

The survey procedure is further detailed in our survey paper.

Data access and intended use

Please send an access request detailing how you plan to use the data. The main purpose of this dataset is to train and evaluate hate speech detection models, as well as study hateful discourse online. This dataset is NOT intended to train generative LLMs to produce hateful content.

Columns

The dataset contains six columns:

text: the annotated post
labels: annotation of whether the post is hateful (== 1) or not (==0). As datasets have different annotation schemes, we systematically binarized the labels.
source: origin of the data (e.g., Twitter)
dataset: dataset the data is from (see "Datasets" part below)
nb_annotators: number of annotators by post

Datasets

The datasets that compose this superset are:

RP-Mod & RP-Crowd: Moderator- and Crowd-Annotated German News Comment Datasets (RP-mod-crowd in the dataset column)
- paper link
- raw data link
Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis (refugees in the dataset column)
- paper link
- raw data link
Detecting Offensive Statements Towards Foreigners in Social Media (Bretschneider)
- paper link
- raw data link
Overview of the HASOC track at FIRE 2019: Hate Speech and Offensive Content Identification in Indo-European Languages (hasoc)
- paper link
- raw data link
Improving Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a German Hate Speech Dataset (gahd)

Additional datasets on demand

In our survey, we identified one additional dataset that is not public but can be requested to the authors, namely:

DeTox: A Comprehensive Dataset for German Offensive Language and Conversation Analysis
- paper link
- request link here

Preprocessing

We drop duplicates. In case of non-binary labels, the labels are binarized (hate speech or not). We replace all usernames and links by fixed tokens to maximize user privacy. Further details on preprocessing can be found in the preprocessing code here.

Citation

Please cite our survey paper if you use this dataset.

@inproceedings{tonneau-etal-2024-languages,
    title = "From Languages to Geographies: Towards Evaluating Cultural Bias in Hate Speech Datasets",
    author = {Tonneau, Manuel  and
      Liu, Diyi  and
      Fraiberger, Samuel  and
      Schroeder, Ralph  and
      Hale, Scott  and
      R{\"o}ttger, Paul},
    editor = {Chung, Yi-Ling  and
      Talat, Zeerak  and
      Nozza, Debora  and
      Plaza-del-Arco, Flor Miriam  and
      R{\"o}ttger, Paul  and
      Mostafazadeh Davani, Aida  and
      Calabrese, Agostina},
    booktitle = "Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024)",
    month = jun,
    year = "2024",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.woah-1.23",
    pages = "283--311",
    abstract = "Perceptions of hate can vary greatly across cultural contexts. Hate speech (HS) datasets, however, have traditionally been developed by language. This hides potential cultural biases, as one language may be spoken in different countries home to different cultures. In this work, we evaluate cultural bias in HS datasets by leveraging two interrelated cultural proxies: language and geography. We conduct a systematic survey of HS datasets in eight languages and confirm past findings on their English-language bias, but also show that this bias has been steadily decreasing in the past few years. For three geographically-widespread languages{---}English, Arabic and Spanish{---}we then leverage geographical metadata from tweets to approximate geo-cultural contexts by pairing language and country information. We find that HS datasets for these languages exhibit a strong geo-cultural bias, largely overrepresenting a handful of countries (e.g., US and UK for English) relative to their prominence in both the broader social media population and the general population speaking these languages. Based on these findings, we formulate recommendations for the creation of future HS datasets.",
}

Datasets:

manueltonneau
/

german-hate-speech-superset

You need to agree to share your contact information to access this dataset