BanStereoSet: A Dataset to Measure Stereotypical Social Biases in LLMs for Bangla

2024-09-18

arXiv.org

AI Summary

This study presents the first systematic evaluation of social bias in multilingual large language models (LLMs) for Bengali, covering nine dimensions: race, gender, occupation, age, physical attractiveness, geography, caste, religion, and “attractiveness-in-occupation.” To address the scarcity of non-English bias evaluation resources, we introduce BanglaStereo—the first stereoscopic social bias benchmark for Bengali—comprising 1,194 expert-validated, culturally adapted, and manually translated sentences. We innovatively localize mainstream bias evaluation frameworks (e.g., StereoSet) to the South Asian context by incorporating region-specific bias categories, including caste, regional identity, and religion. Empirical evaluations across multiple multilingual LLMs reveal statistically significant biases across all dimensions. Our work establishes a reproducible, quantitative benchmark for bias measurement in Bengali NLP, enabling rigorous fairness assessment and facilitating equitable AI deployment in Bengali-speaking communities.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

This study presents BanStereoSet, a dataset designed to evaluate stereotypical social biases in multilingual LLMs for the Bangla language. In an effort to extend the focus of bias research beyond English-centric datasets, we have localized the content from the StereoSet, IndiBias, and Kamruzzaman et. al.'s datasets, producing a resource tailored to capture biases prevalent within the Bangla-speaking community. Our BanStereoSet dataset consists of 1,194 sentences spanning 9 categories of bias: race, profession, gender, ageism, beauty, beauty in profession, region, caste, and religion. This dataset not only serves as a crucial tool for measuring bias in multilingual LLMs but also facilitates the exploration of stereotypical bias across different social categories, potentially guiding the development of more equitable language technologies in Bangladeshi contexts. Our analysis of several language models using this dataset indicates significant biases, reinforcing the necessity for culturally and linguistically adapted datasets to develop more equitable language technologies.

Problem

• Measure social biases in Bangla multilingual LLMs
• Extend bias research beyond English-centric datasets
• Evaluate biases across 9 categories in Bangla

Innovation

• Localized datasets for Bangla bias evaluation
• Covers 9 social bias categories comprehensively
• Analyzes multilingual LLMs for cultural biases