DBCERT: Reconstruct the BERT Model for Chinese Spelling Correction

Chinese spelling correction (CSC) constitutes a research field of high practical significance. However, most studies proposed a large number of auxiliary modules to fulfill the design goals, resulting in increased complexity in the model structure and the number of parameters. In this paper, we propose a model named DBCERT, which constructs detection and correction modules by reconstructing BERT, endowing it with strong capabilities to perceive and correct spelling errors. Additionally, similarity samples are proposed to train the model's ability to perceive the similar relationship between characters. Without adding any auxiliary modules, DBCERT achieves excellent Chinese spelling correction performance, which represents a significant improvement compared to other works.