Chinese Text Error Correction Based on PE-T5 Model

Chinese text error correction is to automatically detect errors or inappropriate expressions in a given Chinese text and output correct correction results according to the context and application scenarios. Existing research ideas mainly focus on two categories: rule-based and depth model-based, where depth model-based methods mainly focus on solving alignment-based errors and perform poorly on non-alignment-based errors. And with the scarcity of manually labeled data, few works have considered using self-supervised training methods to pre-train the models. Therefore, for the common multiple and omitted word errors in electric power article writing, this paper proposes a generative model-based error correction method for Chinese text. A large-scale error correction dataset is first constructed using a rule-based self-supervised approach and pretrained using a generalized domain dataset. Then, in the incremental training phase, lexically supervised signals are added inside the model to enhance the model's error detection effect by means of lexical combination. Finally, the model is fine-tuned using a domain-specific self-supervised dataset. The effectiveness of the method in this paper is demonstrated by analyzing the experimental results in comparison with other models.