AustroTox: A Dataset for Target-Based Austrian German Offensive Language Detection
This provides a resource for interpretable toxicity detection in Austrian German, addressing a gap in non-English datasets, though it is incremental as it extends existing methods to a new dialect.
The authors tackled the lack of token-level annotations for offensive language detection in non-English languages by introducing AustroTox, a dataset of 4,562 Austrian German comments annotated for offensiveness, vulgar language, and targets, and found that large language models outperformed fine-tuned models in detecting offensiveness.
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 language 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. We publish the data and code.