Training a Better Chinese Spelling Correction Model via Prior-knowledge Guided Teacher

Recent advancements in Chinese Spelling Correction (CSC) predominantly leverage pretrained language models (PLMs). However, a notable challenge with fine-tuned PLM-based CSC models is their tendency to over-correct, leading to poor generalization for error patterns outside the standard distribution. To address this, we developed a teacher network guided by prior knowledge for distillation learning of CSC models. Unlike traditional teacher networks, which depend on task-related pretraining, our method infuses task-related prior information into the teacher network, offering guidance beyond mere labels to the student network. This strategy significantly enhances the CSC model's language modeling capabilities, crucial for minimizing over-correction. Importantly, our approach is model-independent and the teacher network does not require taskrelated pre-training, making it broadly applicable for enhancing various PLM-based CSC models with minimal additional computational resources. Extensive experiments on widely used benchmarks demonstrate that our method achieves new state-of-the-art results. Additionally, we explored the potential of generalizing our method to other non-autoregressive textgeneration tasks.