A Noisy Context Optimization Approach for Chinese Spelling Correction

The task of Chinese Spelling Correction (CSC) aims to detect and correct Chinese spelling errors. Recently, BERT-based models have dominated the research on CSC. These methods suffer from unsatisfactory performance when dealing with noisy contexts since incorrect characters are often influenced by other typos. In this paper, we propose A Noisy Context Optimization approach for Chinese Spelling Correction (NCO-Spell). Specifically, first, in the pre-training stage, the multi-character masking strategy is proposed to construct multi-typo texts, enabling the model more robust in noisy environments. Furthermore, we increase the size of the confusion set by dynamically updating. Second, in the inference stage, an iterative algorithm is incorporated to correct the wrong characters one by one, and each iteration will gradually reduce the noise. Extensive experiments and detailed analyses on a widely used benchmark demonstrate that NCO-Spell is effective.