WWR-CS: Chinese Spelling Error Correction Based on Whole Words Replacement fusing Confusion Set

Chinese Spelling Error Correction, which aims to detect and correct spelling mistakes in Chinese text, is an essential branch of Natural Language Processing (NLP) field. However, the similarity of character sounds and shapes poses considerable challenges in Chinese spelling correction. Although significant progress has been made in the BERT-based Masked Language Models (MLMs) incorporating phonological and morphological information, further research is needed on improving correction accuracy in multiple error scenarios. Therefore, the paper proposes the Chinese Error Correction based on Whole Words Replacement fusing Confusion Set (WWR-CS) model which utilizes the whole words replacement method, incorporating confusion set in the noise generation module. Furthermore, due to the fact that most spelling errors stem from phonetic misuse, the model incorporates the pronunciation of Chinese characters and comprehensively considers character correction and pinyin prediction to obtain the final correction result. The experimental results on the Sighan datasets exhibits improved performance compared to the models proposed in recent years, verifying the effectiveness of the proposed model.