Bias mitigation in text classification through cGAN and LLMs

Investigating bias, fairness, and ethics in AI/ML models is essential in the digital era, as automated systems are increasingly used to enhance people’s daily lives. Evaluating these systems in terms of fairness has become crucial. Our proposed work focuses on identifying and mitigating gender bias in text classification models. Specifically, we are evaluating a hate speech classification (HSC) model using fairness metrics and mitigating bias through pre- and in-processing debiasing techniques. In the pre-debiasing stage, we address bias by applying gender swap techniques and removing gender-related terms from the dataset. For in-processing debiasing, we use conditional generative adversarial networks (cGAN) and large language models (LLMs) to generate text similar to that of minority classes, helping balance the dataset and reduce bias during model training. We assess the impact of both methods on bias reduction in the HSC model, with each targeting bias at different stages of model development. While the classification performance of the HSC model showed a slight decrease, fairness evaluation metrics improved by 3.03% and 3.42%, demonstrating the methods’ effectiveness. Addressing gender bias in HSC models applied to social media posts is vital for fostering a healthier digital society.