Mitigating Extrinsic Gender Bias for Bangla Classification Tasks

Researchers have developed RandSymKL, a debiasing technique for Bangla language models that mitigates gender bias in classification tasks like sentiment analysis and hate speech detection. The study introduces four manually annotated benchmark datasets with gender-perturbation testing and demonstrates that the approach effectively reduces bias while maintaining competitive accuracy compared to existing methods.

This research addresses a critical gap in AI fairness for low-resource languages, specifically Bangla, where gender bias in machine learning models remains largely unstudied despite widespread deployment of NLP systems. The authors' contribution extends beyond theoretical exploration by providing practical, reproducible solutions through publicly available datasets and implementation code, establishing a foundation for bias mitigation in South Asian language processing.

The methodology employs a sophisticated approach using minimal-pair testing with gender perturbations—systematically swapping gendered names while preserving semantic meaning—to isolate gender-driven prediction shifts. This technique reveals how models encode and perpetuate societal gender stereotypes. The proposed RandSymKL strategy combines randomized debiasing with symmetric KL divergence and cross-entropy loss, offering a unified framework that balances bias reduction with model performance.

The implications extend across both research and practical domains. For developers deploying NLP systems in Bangla-speaking regions, this work provides validated benchmarks to audit model fairness before production deployment. The publicly released resources democratize access to debiasing techniques for low-resource language communities that typically lack such tools. This matters particularly given that gender bias in language models can amplify discrimination in downstream applications like hiring systems, content moderation, and legal document processing.

Future work should explore whether RandSymKL generalizes to other low-resource languages and examine intersectional biases beyond gender. The research also highlights the need for similar bias mitigation studies across underrepresented language communities where fairness remains a secondary concern.

• →RandSymKL effectively mitigates gender bias in Bangla classification tasks while maintaining competitive accuracy against baseline methods
• →Four new manually annotated benchmark datasets enable minimal-pair evaluation of gender-driven prediction shifts in Bangla NLP models
• →Publicly available implementation and datasets accelerate bias mitigation research for low-resource language communities
• →The work addresses a critical fairness gap in South Asian NLP systems where gender bias has been largely unstudied
• →Symmetric KL divergence combined with cross-entropy loss provides a unified framework for extrinsic gender bias reduction