Measuring South Asian Biases in Large Language Models

TL;DR

The paper addresses culturally grounded, intersectional biases in multilingual South Asian LLMs by introducing a novel bias lexicon and a large-scale 10-language dataset spanning gender, religion, marital status, and number of children. It employs a multilingual generation pipeline with primary models mT0-xxl and IndicTrans2, and evaluates two self-debiasing prompts (simple and complex) using a Bias TF-IDF framework. Findings show substantial biases, especially in Indo-Aryan languages for task-oriented outputs, and reveal that self-debiasing yields limited reductions, underscoring the need for culturally informed debiasing strategies. The work contributes a first-of-its-kind framework and dataset for South Asian bias analysis, with implications for fair multilingual NLP and bias mitigation across culturally diverse contexts.

Abstract

Evaluations of Large Language Models (LLMs) often overlook intersectional and culturally specific biases, particularly in underrepresented multilingual regions like South Asia. This work addresses these gaps by conducting a multilingual and intersectional analysis of LLM outputs across 10 Indo-Aryan and Dravidian languages, identifying how cultural stigmas influenced by purdah and patriarchy are reinforced in generative tasks. We construct a culturally grounded bias lexicon capturing previously unexplored intersectional dimensions including gender, religion, marital status, and number of children. We use our lexicon to quantify intersectional bias and the effectiveness of self-debiasing in open-ended generations (e.g., storytelling, hobbies, and to-do lists), where bias manifests subtly and remains largely unexamined in multilingual contexts. Finally, we evaluate two self-debiasing strategies (simple and complex prompts) to measure their effectiveness in reducing culturally specific bias in Indo-Aryan and Dravidian languages. Our approach offers a nuanced lens into cultural bias by introducing a novel bias lexicon and evaluation framework that extends beyond Eurocentric or small-scale multilingual settings.

Figures (12)

• Figure 1: Pipeline with templates and debiasing prompts (output arrows: black=original, red=simple, blue=complex).
• Figure 2: Identities and Their Highest Bias TF-IDF Terms in Story Generations.
• Figure 3: Identities and Their Highest Bias TF-IDF Terms in Hobbies and Values Generations.
• Figure 4: Identities and Their Highest Bias TF-IDF Terms in To-do List Generations.
• Figure 5: Average Gender Bias Score by Language Family.