Decopy: detect and correct with pinyin for Chinese spelling correction

Chinese spelling correction (CSC) is a critical and complex task focused on detecting and correcting spelling errors in Chinese text. Previous research has been hampered by issues such as misleading error information, over-reliance on high-frequency characters and scarcity of training data. This article proposes Decopy, a novel CSC model that employs an advanced detection-correction framework and an innovative error masking strategy incorporating pinyin features. Decopy not only captures semantic information (word embeddings) and positional information (position embeddings) but also recognizes phonetic features (pinyin embeddings). By leveraging phonetic information directly from the pinyin level, Decopy minimizes the reliance on confusing elements and reduces misleading information. To address the scarcity of training data, we constructed a new CSC dataset based on THUCNews for pre-training Decopy. This enhances Decopy's comprehensive understanding of the input information, particularly additional pinyin information. The experimental results on SIGHAN15 and three domain-specific datasets-LAW, medical and official document writing-demonstrate that Decopy achieves significant improvements and outperforms previous state-of-the-art methods. Furthermore, we evaluate the performance of several high-performance large language models in the CSC task to assess their capabilities in this field.