This repository contains the Arabic-English parallel contexts and cultural entities of the CAMeL-2 dataset for measuring cultural biases in language models.
For more details, see the accompanying paper:
"On The Origin of Cultural Biases in Language Models: From Pre-training Data to Linguistic Phenomena", NAACL 2025
The folder entities contains the parallel Arabic-English entities for 7 different entity types, annotated for broad association with Arab or Western cultures.
The folder qa-contexts provides long contexts with implicit reference to the [MASK] which supports evaluation on extractive QA.
All QA contexts are parallel (provided in Arabic and English versions).
The folder camelco-contexts provides two culturally-contextualized contexts from the previous CAMeL benchmark, where only Arab entities are appropriate [MASK] fillings
We provide two versions of the contexts:
- a version for
masked-lmswhere the [MASK] is placed anywhere in the context - a version for
causal-lmswhere we rewrite sentences for the cultural context to appear behind the [MASK]
The masked-lms contexts are annotated for sentiment (positive, negative, neutral) to support fairness evaluation on sentiment analysis.
All culturally-grounded contexts are parallel (provided in Arabic and English versions).
The fine-tuned Arabic NER models based on XLMR-Large have been uploaded to huggingface:
Names and Author names: https://huggingface.co/tareknaous/xlmr-large-ar-ner-per
Locations: https://huggingface.co/tareknaous/xlmr-large-ar-ner-loc
Food and Beverage: https://huggingface.co/tareknaous/xlmr-large-ar-ner-food
Sports Clubs: https://huggingface.co/tareknaous/xlmr-large-ar-ner-sportsclubs
@article{naous2025origin,
title={On The Origin of Cultural Biases in Language Models: From Pre-training Data to Linguistic Phenomena},
author={Naous, Tarek and Xu, Wei},
journal={arXiv preprint arXiv:2501.04662},
year={2025}
}
Tarek Naous: Scholar | Github | Linkedin | Research Gate | Personal Wesbite | tareknaous@gatech.edu