German Misogyny/Sexsim - Austria (GerMS-AT) Dataset

Summary

This dataset contains user comments from an Austrian online newspaper. The comments have been annotated by 4 or more out of 11 annotators as to how strong sexism/mysogyny is present in the comment.

For each comment, the code of the annotator and the label assigned is given for all annotators which have annotated that comment. Labels represent the severity of any sexism/misogyny present in the comment from 0 (none), 1 (mild), 2 (present), 3 (strong) to 4 (severe).

The dataset contains 7984 comments. We provide the data using the same split as was used for the GermEval2024 GerMS-Detect shared task with a training set of 5998 comments and a test set of 1986 comments. No dev set is provided as the choice of dev set may be best left to the machine learning researcher/engineer.

A unique propery of this corpus is that it contains only a small portion of sexist/misogynyst remarks which use strong language, curse-words or otherwise blatantly offending terms, a large number of comments contain more subtle, indirect or at times ambiguous forms of sexism/misogyny.

Data Structure

All comments are in a single JSONL file, one comment per line with the following properties:

• JSONL File: each line contains the JSON representation of a map
• Each map contains the information for one comment
• The map contains the following fields
  • text: the text of the comment. The text may contain umlauts or other special characters and may contain arbitrary whitespace, newlines or carriage return charachters
  • annotations: a list of maps, each map containing the fields "user" (the code of the annotator which provided the label) and "label" (the label assigned, see below).
  • round: comments were annotated in rounds of 100, this gives the round identifier as a string containing a two-digit round number, e.g. "00" or "13"
  • source: the code which identifies how comments which are likely negative and positive examples where selected for the annotation round

Annotator codes - the following table shows the possible annotator codes and the number of comments annotated by each of them for the Train, Test and

combined data:

Annotator code	Train	Test	All
A001	970	328	1298
A002	5998	1986	7984
A003	1242	456	1698
A004	1394	504	1898
A005	1552	542	2094
A007	1246	451	1697
A008	1849	649	2498
A009	2923	971	3894
A010	5998	1986	7984
A011	927	114	1041
A012	5998	1661	7659

Annotor IDs are anonymized and deliberately do not give demographic information. Among the 11 annotators there were 4 male and 7 female annotators. There were 7 annotators who are content moderators and 4 annotators who are not.

Language

The comments are from a German (mostly Austrian variant) language web site, but may contain English terms which are commonly used by German speakers, quotes of English text or other non-German parts.

Anonymization

No metatdata about the comments is provided, the date of when the comment was written is deliberately not provided. The comment texts have been scanned automatically and manually for any occurrence of information about the username or real name of a person. Any such occurrence has been replaced with the placeholder {USER}. In ambiguous cases where it was not clear if a name refers to a user or e.g. somebody mentioned in the newspaper article, the name was replaced by the placeholder. All mentions of web addresses / URLs were replaced with the placeholder {URL}

Datasheet

See the detailled datasheet

Papers

Brigitte Krenn, Johann Petrak, Marina Kubina, and Christian Burger. 2024. Germs-at: A sex-ism/misogyny dataset of forum comments from an Austrian online newspaper. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 7728–7739.

@inproceedings{krenn2024,
  title={{GERMS-AT}: A Sexism/Misogyny Dataset of Forum Comments from an {A}ustrian Online Newspaper},
  author={Krenn, Brigitte and Petrak, Johann and Kubina, Marina and Burger, Christian},
  booktitle={Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  pages={7728--7739},
  year={2024}
}

NOTE: other than described in this paper, the current version of the corpus contains more annotations from additional annotators, has 11 comments removed which consistet only of an URL and/or USER name, and has additional mentions of user names anonymized.