Year: 2025
Type: article
Abstract: Large Language Models (LLMs) are increasingly integrated into interactive systems across education, governance, healthcare, and everyday digital communication. However, because these models are trained on large-scale text corpora that reflect societal hierarchies, stereotypes, and prejudices, they often internalize and reproduce biased linguistic patterns. While a considerable body of research has investigated algorithmic bias in English-language LLMs, relatively little attention has been directed toward regional and low-resource languages, where linguistic representation is uneven and culturally specific forms of discrimination—related to caste, ethnicity, religion, gender, and socio-economic stratification—are deeply embedded in textual data. This paper proposes an advanced, multi-layered methodology for identifying, quantifying, and mitigating bias in LLMs operating in regional languages. The approach spans culturally-grounded benchmark construction, cross-lingual transfer learning, counterfactual data augmentation, adversarial representation training, and interpretability-guided conceptual editing. An evaluation framework incorporating both computational metrics and human-in-the-loop cultural assessment is presented. The study posits that bias reduction must be conceptualized as an ongoing socio-technical negotiation rather than a one-time algorithmic adjustment.
Author: Prem Harijan
Institution: Chandigarh University
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
SDG: Reduced inequalities
Open Access status: bronze
OpenAlex ID: W4416142733
FWCI:
Cites:
Cited by:
Related to: