Chinese text correction plays a crucial role in the field of natural language processing. However, its development is hampered by several challenges, including the complexity of error types, the high cost of data collection, and the difficulty of accurately correcting key factual errors in domain-specific texts. This paper focuses on the research of Chinese government text correction based on knowledge bases (KB). Given that general large language models (LLMs) have achieved good semantic understanding capabilities after pre-training, we propose the CGT-Corrector: Chinese Government Text Correction with Knowledge Bases. CGT-Corrector encompasses three key stages. Firstly, a data processing framework that integrates M&T labeling and diversity sampling is constructed. This framework effectively addresses the issue of imbalanced data distribution and significantly reduces the training cost. Secondly, a diverse noise generation and instruction optimization strategy based on large language models is designed, which greatly enhances the model’s generalization ability to handle complex errors. Finally, we have successfully achieved efficient training for Chinese government text correction based on knowledge bases. Our approach has achieved 1st place at the NLPCC-2025 Shared Task 5.