Abstract: This paper proposed a comprehensive survey of deep learning-based grammatical error correction (GEC) from 2015 to 2021. GEC research could be divided into universal and adaptive paradigms under which the basic approaches were analyzed. The types and effects of pre-trained language models, public corpus data and evaluation metrics were introduced. Furthermore, a comprehensive evaluation on major GEC systems' performance was conducted. It is suggested that future research should focus on the following aspects: 1) building a suitable and personalized error correction system; 2) exploring methods to enhance the reasoning ability of deep analysis models from the following two aspects: (1) exploring the application methods of pre trained language models, and (2) constructing multi model hybrid systems.