Mitigate Gender Bias Using Negative Multi-task Learning

Deep learning models have showcased remarkable performances in natural language processing tasks. While much attention has been paid to improvements utility, privacy leakage and social bias are two major concerns arising trained models. In this paper, we address both protection gender mitigation classification simultaneously. We first introduce a selective privacy-preserving method that obscures individuals’ sensitive information by adding noise word embeddings. Then, propose negative multi-task framework mitigate bias, which involves main task prediction task. The employs positive loss constraint for utility assurance, while the utilizes remove gender-specific features. analyzed four existing embeddings evaluated them sentiment analysis medical text tasks within proposed framework. For instances, RoBERTa achieves best performance with an average accuracy of 95% sentiment, 1.1 disparity score 1.6 respectively, GloVe 96.42% 0.28 Our experimental results indicate our can effectively maintaining model classification.