GerMS-AT Dataset Datasheet

This file contains information about the GerMS-AT Dataset partially structured according to "T. Gebru et.al. (2021): Datasheets for Datasets"

If there is a need for additional information or clarification, please feel free to contact any of the the maintainers of this repository.

Purpose of dataset creation: the corpus was created in the context of project FemDwell in order to build a machine-learning based assistant to help human content moderators identify individual forum posts which may contain sexist or misogynist comments or whole fora, where an unusual large number of such posts were created more easily.
Dataset creators:
- Austrian Research Institue for Artificial Intelligence
- Standard Verlagsgesellschaft
Funding of dataset creation:
- The project was sponsored by Vienna Business Agency

Instance representation: JSON dictionary
Number of instances: 7984
Data per instance:
- id: a unique ID
- text: the text of the forum comment posted, with user names and real person names replaced by the string "{USER}" and web addresses, email address and the like replaced by the string "{URL}"
Label/target per instance:
- for each instance there is a dictionary which contains the label assigned by each of the annotators who were presented the comment. Annotators are represented by anonymized annotator ids.
- the label assigned by each annotator is one of "0" (no sexism/misogyny present), "1" (mild), "2" (present), "3" (severe), "4" (extreme).
Recommended data splits:
- the data is made available as a train and a test file, with the same split that was used in the GermEval2024 GerMS-Detect shared task
Presence of confidential information: there is no confidential information in the dataset
Presence of offensive or otherwise problematic data: comments will contain sexist and misogynist remarks and may also contain other forms of offensive or toxic remarks.
Identifyability of subpopulations: comments have been made by readers of the online news web site
Identifyability of individuals: all information that could identify a user or person has been removed or replaced with placeholders
Presence of sensitive information: none