---
license: cc-by-sa-4.0
language:
- eu
tags:
- bias
size_categories:
- 100K<n<1M
configs:
- config_name: sentences
  data_files:
  - split: test
    path: holisticbias_eu.csv
task_categories:
- text-generation
---
# HolisticBiasEus
[![GitHub Repository](https://img.shields.io/badge/GitHub-Repository-black?logo=github)](https://github.com/amaiamurillo/holisticbiaseus)

* **Curated by**: HiTZ Center, University of the Basque Country (EHU)
* **Funded by**: Project _Desarrollo de Modelos ALIA_
* **License**: CC-BY-SA-4.0

HolisticBiasEus is a Basque adaptation of the [HolisticBias](https://github.com/facebookresearch/ResponsibleNLP/tree/main/holistic_bias) dataset, a benchmark designed to evaluate social bias in LLMs. The original dataset uses template-based prompts to evaluate model responses to demographic identity language within two-person conversational contexts, and it covers 13 demographic axes: ability, age, body type, characteristics, cultural, gender/sex, nationality, nonce, political ideology, race/ethnicity, religion, sexual orientation, socioeconomic status. The dataset is constructed using 26 sentence templates, more than 600 descriptors in American English representing 13 protected attributes and 30 nouns. 

To create the Basque version, we did not directly translate the instantiated sentence prompts. Instead, we translated and adapted the components separately (sentence templates, descriptors and nouns) taking into account the grammatical features of Basque. During the adaptation process, we filtered out descriptors that were not socially relevant in the Basque context (e.g. _redneck_) and introduced new descriptors to better reflect local sociocultural realities (e.g. _abertzale_). A key difference with respect to the English version is that Basque descriptors may appear either before or after the noun. For this reason, each descriptor in HolisticBiasEus is annotated with its position relative to the noun (1: before the noun, or 2: after the noun). This positional information is used during sentence generation to correctly instantiate prompts and preserve grammaticality in the target language.

As in the original HolisticBias dataset, some automatically generated combinations may result in prompts that are semantically unusual or implausible (e.g. _40 urteko haurra_, '40-year-old child'). Addressing such cases is part of ongoing work on improving the dataset.


## Citation
If you use HolisticBiasEus in your work, please make sure to also cite the original HolisticBias paper:

```bibtex
@inproceedings{smith-etal-2022-im,
    title = "``{I}{'}m sorry to hear that'': Finding New Biases in Language Models with a Holistic Descriptor Dataset",
    author = "Smith, Eric Michael  and
      Hall, Melissa  and
      Kambadur, Melanie  and
      Presani, Eleonora  and
      Williams, Adina",
    editor = "Goldberg, Yoav  and
      Kozareva, Zornitsa  and
      Zhang, Yue",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.625/",
    doi = "10.18653/v1/2022.emnlp-main.625",
    pages = "9180--9211",
    abstract = "As language models grow in popularity, it becomes increasingly important to clearly measure all possible markers of demographic identity in order to avoid perpetuating existing societal harms. Many datasets for measuring bias currently exist, but they are restricted in their coverage of demographic axes and are commonly used with preset bias tests that presuppose which types of biases models can exhibit. In this work, we present a new, more inclusive bias measurement dataset, HolisticBias, which includes nearly 600 descriptor terms across 13 different demographic axes. HolisticBias was assembled in a participatory process including experts and community members with lived experience of these terms. These descriptors combine with a set of bias measurement templates to produce over 450,000 unique sentence prompts, which we use to explore, identify, and reduce novel forms of bias in several generative models. We demonstrate that HolisticBias is effective at measuring previously undetectable biases in token likelihoods from language models, as well as in an offensiveness classifier. We will invite additions and amendments to the dataset, which we hope will serve as a basis for more easy-to-use and standardized methods for evaluating bias in NLP models."
}
```

## Funding

This work has been supported and funded by the Ministerio para la Transformación Digital y de la Función Pública and the Plan de Recuperación, Transformación y Resiliencia – funded by the EU through NextGenerationEU within the project Desarrollo de Modelos ALIA.

## Contact
* Amaia Murillo (amaia.murillo@ehu.eus)
* Olatz Perez de Viñaspre (olatz.perezdevinaspre@ehu.eus)
* Naiara Perez (naiara.perez@ehu.eus)