Assessing the Reliability of LLMs Annotations in the Context of Demographic Bias and Model Explanation

Project page for our ACL 2025 paper at the 6th Workshop on Gender Bias in Natural Language Processing (GeBNLP).

🔗 Live page: tina-shahedi.github.io/GeBNLP

TL;DR

Understanding the sources of variability in annotations is crucial for building fair NLP systems, especially for tasks like sexism detection where demographic bias is a concern. Using a Generalized Linear Mixed Model, we find that annotator demographics — while statistically present — explain only a minor fraction (~8%) of the labeling variance, with tweet content being the dominant factor. We also evaluate generative LLMs as annotators and find that simplistic demographic persona prompting often fails to improve, and sometimes degrades, alignment with human judgments.

Authors

Hadi Mohammadi · Tina Shahedi · Pablo Mosteiro · Massimo Poesio · Ayoub Bagheri · Anastasia Giachanou

Utrecht University

Resources

Resource	Link
📄 Paper (ACL Anthology)	aclanthology.org/2025.gebnlp-1.9
📝 Preprint (arXiv)	arxiv.org/abs/2507.13138
💻 Code & data	mohammadi-hadi/Explainable_Annotations_Reliability
📚 Key resources & author info	tina-shahedi.github.io/ACL2025

Repository structure

.
├── index.html      # project page
└── static/
    ├── css/        # Bulma and custom styles
    ├── js/         # carousel and slider scripts
    ├── images/     # figures and result plots
    └── *.pdf       # paper and CV

Citation

@inproceedings{mohammadi-etal-2025-assessing,
    title = "Assessing the Reliability of {LLM}s Annotations in the Context of Demographic Bias and Model Explanation",
    author = "Mohammadi, Hadi and Shahedi, Tina and Mosteiro, Pablo and Poesio, Massimo and Bagheri, Ayoub and Giachanou, Anastasia",
    booktitle = "Proceedings of the 6th Workshop on Gender Bias in Natural Language Processing (GeBNLP)",
    month = aug,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    pages = "92--104",
    doi = "10.18653/v1/2025.gebnlp-1.9"
}

Acknowledgments

This page is built on the Academic Project Page Template, which is based on the Nerfies project page.