CopySpell: Improving Chinese Spelling Correction Model with Copy Mechanism

The Chinese Spelling Correction (CSC) task aims to detect and correct spelling errors in texts. In recent years, BERT-based models have achieved significant breakthroughs in this field. However, CSC models still suffer from over-correction of the valid character to a similar one, which is caused by the training with downstream tasks. To solve this dilemma, we introduce CopySpell, a novel model that combines a copy mechanism with a new training method. By sharing original character information through a copy block, the model encourages robust and accurate outputs while avoiding over-correction. This approach allows the model to explore associations beyond the similar characters. Experiments are conducted on widely used benchmarks, and our model has achieved state-of-the-art results. We will also release the source code for future research and communication.