Datasets:
metadata
task_categories:
- text-classification
language:
- de
tags:
- toxic
- offensive
- hate
- perspectivism
- subjective
- spans
AustroTox: A Dataset for Target-Based Austrian German Offensive Language Detection
This is the data for the paper AustroTox: A Dataset for Target-Based Austrian German Offensive Language Detection.
Dataset description
Index
The index of the comment
Article Title
The title of the article under which it was posted
Comment
The text from the Jigsaw Toxic Comment Classification Challenge which was classified
Label
Denotes whether the aggregated label is Toxic / Offensive (1) or Not toxic / Not offensive (0)
Annotators not toxic
The annotator ids of the annotators who labelled the text as not toxic / not offensive
Annotators toxic
The annotator ids of the annotators who labelled the text as toxic / offensive
Manually cleaned
Indicates whether the annotations were manually cleaned
Label fine
The fine-grained label of the post
Tags
The spans and the labels of the spans
Citation
If you use this dataset, please cite us:
@inproceedings{pachinger-etal-2024-austrotox,
title = "{A}ustro{T}ox: A Dataset for Target-Based {A}ustrian {G}erman Offensive Language Detection",
author = "Pachinger, Pia and
Goldzycher, Janis and
Planitzer, Anna and
Kusa, Wojciech and
Hanbury, Allan and
Neidhardt, Julia",
editor = "Ku, Lun-Wei and
Martins, Andre and
Srikumar, Vivek",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.findings-acl.713/",
doi = "10.18653/v1/2024.findings-acl.713",
pages = "11990--12001",
abstract = "Model interpretability in toxicity detection greatly profits from token-level annotations. However, currently, such annotations are only available in English. We introduce a dataset annotated for offensive language detection sourced from a news forum, notable for its incorporation of the Austrian German dialect, comprising 4,562 user comments. In addition to binary offensiveness classification, we identify spans within each comment constituting vulgar language or representing targets of offensive statements. We evaluate fine-tuned Transformer models as well as large language models in a zero- and few-shot fashion. The results indicate that while fine-tuned models excel in detecting linguistic peculiarities such as vulgar dialect, large language models demonstrate superior performance in detecting offensiveness in AustroTox."
}