Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AfriStereo: A Culturally Grounded Dataset for Evaluating Stereotypical Bias in Large Language Models

Yann Le Beux, Oluchi Audu, Oche D. Ankeli, Dhananjay Balakrishnan, Melissah Weya, Marie D. Ralaiarinosy, Ignatius Ezeani

TL;DR

AfriStereo fills a critical gap by delivering the first Africa-grounded stereotype dataset and evaluation framework, built through community-engaged data collection in Senegal, Kenya, and Nigeria and augmented with LLM-generated data to exceed 5,000 stereotype–antistereotype pairs. The work applies a Stereotype–Antistereotype paradigm to eleven open-source LLMs, revealing statistically significant African stereotypes across generations, especially for age, profession, and gender, with domain-specific models showing partial mitigation. Despite recent architectural advances, modern models exhibit biases comparable to or stronger than earlier baselines, underscoring the need for culturally informed bias mitigation beyond generic improvements. The AfriStereo pipeline—combining open-ended surveys, semantic clustering, human-in-the-loop validation, and synthetic augmentation—provides a scalable, reproducible resource for evaluating and debiasing NLP systems in Global South contexts, advancing globally inclusive AI research.

Abstract

Existing AI bias evaluation benchmarks largely reflect Western perspectives, leaving African contexts underrepresented and enabling harmful stereotypes in applications across various domains. To address this gap, we introduce AfriStereo, the first open-source African stereotype dataset and evaluation framework grounded in local socio-cultural contexts. Through community engaged efforts across Senegal, Kenya, and Nigeria, we collected 1,163 stereotypes spanning gender, ethnicity, religion, age, and profession. Using few-shot prompting with human-in-the-loop validation, we augmented the dataset to over 5,000 stereotype-antistereotype pairs. Entries were validated through semantic clustering and manual annotation by culturally informed reviewers. Preliminary evaluation of language models reveals that nine of eleven models exhibit statistically significant bias, with Bias Preference Ratios (BPR) ranging from 0.63 to 0.78 (p <= 0.05), indicating systematic preferences for stereotypes over antistereotypes, particularly across age, profession, and gender dimensions. Domain-specific models appeared to show weaker bias in our setup, suggesting task-specific training may mitigate some associations. Looking ahead, AfriStereo opens pathways for future research on culturally grounded bias evaluation and mitigation, offering key methodologies for the AI community on building more equitable, context-aware, and globally inclusive NLP technologies.

AfriStereo: A Culturally Grounded Dataset for Evaluating Stereotypical Bias in Large Language Models

TL;DR

AfriStereo fills a critical gap by delivering the first Africa-grounded stereotype dataset and evaluation framework, built through community-engaged data collection in Senegal, Kenya, and Nigeria and augmented with LLM-generated data to exceed 5,000 stereotype–antistereotype pairs. The work applies a Stereotype–Antistereotype paradigm to eleven open-source LLMs, revealing statistically significant African stereotypes across generations, especially for age, profession, and gender, with domain-specific models showing partial mitigation. Despite recent architectural advances, modern models exhibit biases comparable to or stronger than earlier baselines, underscoring the need for culturally informed bias mitigation beyond generic improvements. The AfriStereo pipeline—combining open-ended surveys, semantic clustering, human-in-the-loop validation, and synthetic augmentation—provides a scalable, reproducible resource for evaluating and debiasing NLP systems in Global South contexts, advancing globally inclusive AI research.

Abstract

Existing AI bias evaluation benchmarks largely reflect Western perspectives, leaving African contexts underrepresented and enabling harmful stereotypes in applications across various domains. To address this gap, we introduce AfriStereo, the first open-source African stereotype dataset and evaluation framework grounded in local socio-cultural contexts. Through community engaged efforts across Senegal, Kenya, and Nigeria, we collected 1,163 stereotypes spanning gender, ethnicity, religion, age, and profession. Using few-shot prompting with human-in-the-loop validation, we augmented the dataset to over 5,000 stereotype-antistereotype pairs. Entries were validated through semantic clustering and manual annotation by culturally informed reviewers. Preliminary evaluation of language models reveals that nine of eleven models exhibit statistically significant bias, with Bias Preference Ratios (BPR) ranging from 0.63 to 0.78 (p <= 0.05), indicating systematic preferences for stereotypes over antistereotypes, particularly across age, profession, and gender dimensions. Domain-specific models appeared to show weaker bias in our setup, suggesting task-specific training may mitigate some associations. Looking ahead, AfriStereo opens pathways for future research on culturally grounded bias evaluation and mitigation, offering key methodologies for the AI community on building more equitable, context-aware, and globally inclusive NLP technologies.

Paper Structure

This paper contains 50 sections, 3 equations, 2 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Challenges of Fairness in AI for African Contexts
  4. Evolution of Bias Evaluation Benchmarks
  5. Stereotype Benchmarks and Considerations for African Contexts
  6. Methodology
  7. Data Collection through Community Engagement
  8. Participant Demographics
  9. Data Processing Pipeline
  10. Raw Data Extraction
  11. Semantic Clustering
  12. Human-in-the-Loop Verification
  13. Stereotype--Antistereotype Pair Generation
  14. Synthetic Data Augmentation
  15. The AfriStereo Dataset
  16. ...and 35 more sections

Figures (2)

  • Figure 1: Data processing pipeline showing the four stages: raw data extraction, semantic clustering with polarity detection, human-in-the-loop verification, and stereotype--antistereotype pair generation.
  • Figure 2: Most frequent attribute categories after grouping semantically related terms. The top categories reveal patterns in stereotypical associations: cognitive abilities, physical characteristics, behavioral tendencies, and emotional traits dominate the dataset.