Mitigating Extrinsic Gender Bias for Bangla Classification Tasks
Proposes RandSymKL, combining symmetric KL and cross-entropy, to mitigate extrinsic gender bias in Bangla classification tasks, achieving 90.66% accuracy and bias reduction.
Key Findings
Methodology
This study constructs four manually annotated datasets for sentiment, toxicity, hate speech, and sarcasm detection in Bangla. Gender perturbations are generated by swapping gendered names and terms, assessing model bias via minimal pairs. The RandSymKL algorithm integrates stochastic fine-tuning with joint optimization of cross-entropy and symmetric KL divergence, aligning model outputs across gendered pairs. Experiments across multiple pretrained models demonstrate that RandSymKL effectively reduces bias while maintaining high classification accuracy, outperforming existing methods.
Key Results
- Across four tasks, RandSymKL achieves an average accuracy of 90.66%, only 0.44% below the unmitigated baseline, and reduces bias disparity to 0.29%, outperforming other baselines in balancing fairness and performance.
- Bias metrics like FairScore, SPD, and EOD show significant improvements, indicating effective bias mitigation. The method enhances prediction consistency across gendered inputs, reducing prediction disparities.
- Gender-swapped data evaluation confirms the model's increased gender invariance, with predictions becoming more stable and less biased, demonstrating the method's robustness in low-resource settings.
Significance
This work addresses a critical gap in low-resource language fairness research, providing a scalable, effective method for external gender bias mitigation. It advances the state-of-the-art in fair NLP, with implications for deploying equitable AI systems in diverse socio-cultural contexts, especially where explicit gender cues are absent or implicit.
Technical Contribution
The core innovation is the RandSymKL algorithm, which jointly optimizes classification accuracy and output distribution alignment through stochastic training and symmetric KL regularization. The creation of a 573-term gender lexicon supports precise counterfactual data generation, facilitating bias evaluation. This approach differs from prior embedding or data augmentation techniques by directly aligning output distributions, offering a theoretically grounded, practically effective solution.
Novelty
This is the first comprehensive study on external gender bias mitigation for multiple Bangla classification tasks, introducing a novel stochastic joint loss combining cross-entropy and symmetric KL divergence. Unlike traditional methods, it emphasizes output distribution alignment without relying solely on data augmentation or explicit gender markers, representing a significant methodological advance.
Limitations
- The approach depends on the quality of gender lexicons and accurate NER, which may not generalize well to complex or ambiguous contexts. Multilingual or dialectal variations could pose challenges.
- Training complexity and hyperparameter sensitivity require careful tuning, limiting scalability for large models or real-time applications.
- Current focus is solely on gender bias; other social biases like caste, religion, or ethnicity remain unaddressed, necessitating broader bias mitigation frameworks.
Future Work
Future research will explore multi-bias mitigation, integrating fairness across multiple social dimensions. Extending the approach to multilingual contexts and unsupervised bias detection methods will enhance scalability. Additionally, combining output distribution alignment with contextual embedding techniques may further improve fairness robustness.
AI Executive Summary
The rapid development of large language models (LLMs) like GPT-4 and LLaMA has transformed NLP, but their deployment often reveals embedded biases, especially in low-resource languages such as Bangla. While much research has focused on intrinsic biases within model representations, the external, task-specific biases—manifested as prediction disparities across gendered inputs—remain underexplored. This gap is critical, as biased predictions can reinforce social stereotypes and cause unfair outcomes.
In this context, the authors constructed four benchmark datasets for Bangla classification tasks—sentiment, toxicity, hate speech, and sarcasm detection—by manually annotating gendered texts and generating counterfactual pairs through gender swapping. These datasets enabled precise bias evaluation. To address the bias, they proposed RandSymKL, a novel training strategy that jointly optimizes classification accuracy and output distribution alignment across gendered pairs. The method combines stochastic cross-entropy with symmetric KL divergence regularization, encouraging the model to produce gender-invariant predictions.
Experimental results across multiple pretrained models demonstrated that RandSymKL effectively reduces bias disparities, lowering the accuracy gap to 0.29% while maintaining an average accuracy of 90.66%. Compared to baseline methods like data augmentation and token masking, it offers a superior balance of fairness and performance. This approach not only advances bias mitigation techniques for low-resource languages but also provides a scalable framework for broader social bias reduction.
The significance of this work lies in its potential to promote fairer AI systems in diverse socio-cultural settings, reducing harmful stereotypes and fostering social equity. Limitations include dependency on gender lexicons and computational complexity, suggesting future directions such as multi-bias mitigation, multilingual extension, and contextual embedding integration. Overall, this research marks a substantial step toward equitable NLP in underrepresented languages, with promising implications for ethical AI deployment worldwide.
Deep Dive
Abstract
In this study, we investigate extrinsic gender bias in Bangla pretrained language models, a largely underexplored area in low-resource languages. To assess this bias, we construct four manually annotated, task-specific benchmark datasets for sentiment analysis, toxicity detection, hate speech detection, and sarcasm detection. Each dataset is augmented using nuanced gender perturbations, where we systematically swap gendered names and terms while preserving semantic content, enabling minimal-pair evaluation of gender-driven prediction shifts. We then propose RandSymKL, a randomized debiasing strategy integrated with symmetric KL divergence and cross-entropy loss to mitigate the bias across task-specific pretrained models. RandSymKL is a refined training approach to integrate these elements in a unified way for extrinsic gender bias mitigation focused on classification tasks. Our approach was evaluated against existing bias mitigation methods, with results showing that our technique not only effectively reduces bias but also maintains competitive accuracy compared to other baseline approaches. To promote further research, we have made both our implementation and datasets publicly available: https://github.com/sajib-kumar/Mitigating-Bangla-Extrinsic-Gender-Bias