This paper aims to address the challenges of gender bias in the field of Chinese Natural Language Processing by proposing a methodology comprising three progressive tasks: gender bias detection, fine-grained bias classification, and bias mitigation. For the bias detection and classification tasks, this study employs the BERT language model as a foundation and introduces a Fast Gradient Method (FGM) adversarial training strategy to enhance model robustness. Through a five-model ensemble strategy based on predetermined model ranking and hierarchical decision-making principles, precise identification of gender bias in text and in-depth analysis of bias types are achieved. For the bias mitigation task, this study leverages Large Language Models (LLMs), guiding them with carefully designed few-shot prompts to rewrite biased text into neutral and fluent expressions. This research aims to provide a systematic solution for gender bias governance in the Chinese NLP domain, promoting the development of fairer and more inclusive language technologies. This progressive framework, moving from detection to detailed classification and then to active mitigation, allows for a more comprehensive approach to tackling gender bias compared to addressing each stage in isolation, reflecting the comprehensiveness and depth of the methodology.