StereoHoax-GL

Dataset Summary

StereoHoax-GL is a Galician translated and linguistically revised version of the test partition of the StereoHoax-ES subset included in DETESTS-Dis.

The dataset is intended as an evaluation resource for stereotype and discriminatory-content detection in Galician. It preserves the original identifiers, metadata fields, and annotation labels from the StereoHoax-ES test partition, while replacing the original Spanish text field with its final Galician revised version.

The released dataset is monolingual Galician. It is not a bilingual dataset and does not include the original Spanish texts in the final data files.

Dataset Description

StereoHoax-GL contains Galician social media comments derived from the StereoHoax-ES test partition. The original source data are related to racial stereotypes, discriminatory discourse, immigration-related hoaxes, and related harmful narratives.

The dataset follows the original DETESTS-Dis / StereoHoax-ES annotation format. It includes individual annotator labels and aggregated labels for two annotation dimensions: stereotype and implicit.

No new stereotype or implicitness annotations were added during the Galician adaptation process. The labels are inherited from the original Spanish source data and projected onto the Galician translated version.

This dataset only contains the translated and revised test split of StereoHoax-ES. It should not be interpreted as a complete Galician version of DETESTS-Dis.

Dataset Structure

The dataset is distributed in CSV format.

Each row contains the following columns:

Column	Type	Description
source	string	Source subset or collection name.
id	string	Instance identifier.
comment_id	string	Original comment identifier.
text	string	Final Galician revised text.
level1	string or integer	Original hierarchical or contextual metadata field.
level2	string or integer	Original hierarchical or contextual metadata field.
level3	string or integer	Original hierarchical or contextual metadata field.
level4	string or integer	Original hierarchical or contextual metadata field.
stereotype_a1	integer	Annotation from annotator 1 for stereotype.
stereotype_a2	integer	Annotation from annotator 2 for stereotype.
stereotype_a3	integer	Annotation from annotator 3 for stereotype.
stereotype	integer	Aggregated stereotype label.
stereotype_soft	string or integer	Soft or disagreement-aware stereotype label inherited from the original dataset.
implicit_a1	integer	Annotation from annotator 1 for implicitness.
implicit_a2	integer	Annotation from annotator 2 for implicitness.
implicit_a3	integer	Annotation from annotator 3 for implicitness.
implicit	integer	Aggregated implicitness label.
implicit_soft	string or integer	Soft or disagreement-aware implicitness label inherited from the original dataset.

Supported Tasks and Leaderboards

This dataset is suitable for:

• stereotype detection
• discriminatory-content detection
• racism and xenophobia detection
• evaluation of Galician language models
• multilingual and cross-lingual evaluation
• low-resource NLP evaluation
• analysis of bias, harmful language, and model robustness

Languages

The dataset is in Galician.

The source language used for translation and adaptation was Spanish, but the released dataset is monolingual Galician. The original Spanish texts are not included in the final data files.

Intended Uses

StereoHoax-GL can be used for:

• evaluation of stereotype detection systems in Galician
• evaluation of discriminatory-content detection systems in Galician
• benchmarking multilingual and Galician-specific language models
• analysis of cross-lingual transfer in harmful-content detection
• evaluation of model robustness in low-resource language settings

Out-of-Scope Uses

StereoHoax-GL should not be used to target, profile, or make decisions about individuals or communities.

It should not be used as a standalone content moderation system without additional validation, human oversight, and careful consideration of social, linguistic, and cultural context.

It should not be treated as a complete Galician version of DETESTS-Dis, since it only contains the translated and revised test partition of the StereoHoax-ES subset.

The dataset must not be publicly redistributed or shared outside authorized research contexts.

Dataset Creation

StereoHoax-GL was created from the Spanish StereoHoax-ES test partition.

The creation process consisted of:

1. automatic translation of the original Spanish text field into Galician;
2. linguistic revision of the Galician translations;
3. internal correction of selected revised translations when needed;
4. reconstruction of the original DETESTS-Dis / StereoHoax-ES format.

No new stereotype or implicitness annotations were added during the Galician adaptation process.

Annotation

The stereotype and implicitness labels are inherited from the original DETESTS-Dis / StereoHoax-ES test partition.

The Galician texts were not newly annotated from scratch. Therefore, the labels should be interpreted as projected labels from the Spanish source data onto the Galician translated version.

The dataset includes individual annotator labels and aggregated labels for two annotation dimensions:

• stereotype
• implicit

Limitations

• The dataset contains only the test partition of the StereoHoax-ES subset.
• It should not be interpreted as a complete Galician version of DETESTS-Dis.
• The dataset is a translated evaluation resource, not a fully native Galician annotation dataset.
• Because the dataset was translated from Spanish into Galician, some lexical, pragmatic, cultural, or sociolinguistic cues may differ from naturally occurring Galician social media discourse.
• The labels are inherited from the Spanish source data and were not newly annotated on the Galician text.
• Evaluation results may reflect both model performance on harmful-content detection and sensitivity to translation choices.
• The dataset contains sensitive and potentially harmful content and should be used with caution.

Sensitive Content Warning

The dataset may contain discriminatory, offensive, or harmful language because it is derived from social media content related to stereotypes, discrimination, racism, immigration, and hoaxes.

The presence of such content is exclusively for research and evaluation purposes.

Usage

Example with datasets using a local CSV file:

from datasets import load_dataset

ds = load_dataset("csv", data_files={"test": "test.csv"})
print(ds["test"][0])

If the dataset is loaded from a Hugging Face repository with a configured test split:

from datasets import load_dataset

ds = load_dataset("proxectonos/StereoHoax-GL")
print(ds["test"][0])

Access, License and Reuse Conditions

StereoHoax-GL is a Galician translated and linguistically revised version of the test partition of the StereoHoax-ES subset included in DETESTS-Dis.

The original DETESTS-Dis dataset is distributed under restricted access for academic purposes only. Access to the original data requires accepting the corresponding terms and conditions, including the commitment not to redistribute the dataset. StereoHoax-GL follows the same access and reuse conditions.

Permission to publish this Galician derivative version was requested from the people responsible for the original resource. Permission was granted on the condition that the Galician version preserves the same conditions as the original dataset: academic use, proper attribution, preservation of the legal notice/disclaimer, restricted access, and any other applicable terms of use.

Accordingly, StereoHoax-GL is made available under restricted access for academic research and evaluation purposes only. Users must not redistribute, republish, mirror, or share the dataset outside the authorized access mechanism.

The dataset follows the CC BY-NC-SA 4.0 license information associated with the original DETESTS-Dis resource. Access and use are also subject to the restrictions and terms of the original DETESTS-Dis / StereoHoax-ES dataset.

GDPR, IPR and Anonymization Disclaimer

Given the restrictions posed by EU GDPR regulations and to avoid any conflict with the sources of the comments regarding their intellectual property rights (IPR), the original DETESTS-Dis data are made available for academic purposes only and under restricted access, with a commitment not to redistribute the dataset.

The same conditions apply to this Galician derivative version.

Citation

@article{schmeisser_nieto_etal_2024_stereohoax,
  title = {Stereohoax: a multilingual corpus of racial hoaxes and social media reactions annotated for stereotypes},
  author = {Schmeisser-Nieto, W. S. and Cignarella, A. T. and Bourgeade, T. and Frenda, S. and Ariza-Casabona, A. and Laurent, M. and Cicirelli, P. G. and Marra, A. and Corbelli, G. and Benamara, F. and Bosco, C. and Moriceau, V. and Paciello, M. and Taul{\'e}, M. and D'Errico, F.},
  journal = {Language Resources and Evaluation},
  year = {2024},
  publisher = {Springer Nature},
  doi = {10.1007/s10579-024-09791-3}
}

@article{schmeisser_nieto_etal_2024_detests_dis,
  title = {Overview of {DETESTS-Dis} at {IberLEF} 2024: {DETE}ction and classification of racial {ST}ereotypes in {Spanish} - Learning with Disagreement},
  author = {Schmeisser-Nieto, Wolfgang S. and Pastells, P. and Frenda, S. and Ariza-Casabona, A. and Farr{\'u}s, M. and Rosso, P. and Taul{\'e}, M.},
  journal = {Procesamiento del Lenguaje Natural},
  volume = {73},
  pages = {323--333},
  year = {2024},
  issn = {1989-7553}
}

Acknowledgements

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA. Esta publicación del proyecto Desarrollo de Modelos ALIA está financiada por el Ministerio para la Transformación Digital y de la Función Pública y por el Plan de Recuperación, Transformación y Resiliencia – Financiado por la Unión Europea – NextGenerationEU.