PMDRSpell: Dynamic Residual Chinese Spelling Check Method Guided by Phonological and Morphological Knowledge

Since the errors in Chinese Spell Correction (CSC) involve phonetically or morphologically confusing Chinese characters, mainstream models have made numerous attempts to fuse phonological and morphological knowledge. We observe that in erroneous sentences where the vast majority of Chinese characters are correctly written, mainstream models may unintentionally increase the difficulty of predicting these correct characters when integrating multi-modal knowledge across all characters. Additionally, these models often overlook the potential relationship between the phonological and morphological modalities of a Chinese character when utilizing multi-modal information. In this paper, we propose an end-to-end model called PMDRSpell, which models erroneous Chinese characters in sentences using their multi-modal knowledge and reduces the use of multi-modal information for correct Chinese characters. Additionally, it uncovers the relationship between phonological and morphological features based on the characteristics of phonograms, enhancing the similarity between similar Chinese characters. Specifically, coarse-grained and hierarchical detection is first employed to localize and mask error locations within sentences, using the original embedding information as residual features. Next, correlation information in the phonological and morphological modalities of the erroneous characters is extracted to construct new representational features, which are then used to update the erroneous Chinese character information within the residual features. Finally, the masked sentences are predicted using the MLM model and classified to generate correct sentences by combining the residual features with the updated multi-modal information. Our model effectively reduces the interference from correct Chinese characters during the inspection process and leverages multi-modal information to accurately correct incorrect Chinese characters. In our comparison experiments with recent state-of-the-art models, PMDRSpell outperforms the optimal baseline in terms of error-corrected F1 scores for Sighan14 and Sighan15 by 1.2 and 1.0 percentage points, respectively.