MDCSpell: A Multi-task Detector-Corrector Framework for Chinese Spelling Correction

Findings of the Association for Computational Linguistics: ACL 2022, pages 1244 - 1253 May 22-27, 2022 c⃝2022 Association for Computational Linguistics MDCSpell: A Multi-task Detector-Corrector Framework for Chinese Spelling Correction Chenxi Zhu, Ziqiang Ying, Boyu Zhang, Feng Mao Alibaba Group {mike.zcx, zhangboyu.zby, maofeng.mf}@alibaba-inc.com yingzq0116@163.com Abstract Chinese Spelling Correction (CSC) is a task to detect and correct misspelled characters in Chi- nese texts. CSC is challenging since many Chi- nese characters are visually or phonologically similar but with quite different semantic mean- ings. Many recent works use BERT-based lan- guage models to directly correct each char- acter of the input sentence. However, these methods can be sub-optimal since they cor- rect every character of the sentence only by the context which is easily misled by the mis- spelled characters. Some other works propose to use an error detector to guide the correction by masking the detected errors. Nevertheless, these methods dampen the visual or phono- logical features from the misspelled charac- ters which could be critical for correction. In this work, we propose a novel general detector- corrector multi-task framework where the cor- rector uses BERT to capture the visual and phonological features from each character in the raw sentence and uses a late fusion strat- egy to fuse the hidden states of the correc- tor with that of the detector to minimize the misleading impact from the misspelled char- acters. Comprehensive experiments on bench- marks demonstrate that our proposed method can significantly outperform the state-of-the- art methods in the CSC task. 1 Introduction Chinese Spelling Correction (CSC) is a fundamen- tal task that aims to automatically detect and cor- rect spelling errors in Chinese texts. These spelling errors are typically caused by human writing, au- tomatic speech recognition (ASR) or optical char- acter recognition (OCR) systems (Afli et al., 2016; Wang et al., 2018). CSC is essential since it is cru- cial for many downstream tasks like search engine (Martins and Silva, 2004; Gao et al., 2010) and essay scoring (Burstein and Chodorow, 1999). Despite its recent development, CSC remains a challenging task since many Chinese characters are Input 我这一次写信给你是想跟你安排一 下关(gu¯an)以(yˇı)我们要见面的。 baseline 我这一次写信给你是想跟你安排一 下所(suˇo)以(yˇı)我们要见面的事。 Ground Truth 我这一次写信给你是想跟你安排一 下关(gu¯an)于(yú)我们要见面的事。 Translation I am writing to you this time to make ar- rangements with you about our meeting. Input 为了减少急遍(biàn)的生孩子率，需 要呼吁适当的生育政策。 baseline 为了减少急速(sù)的生孩子率，需要 呼吁适当的生育政策。 Ground Truth 为了减少急变(biàn)的生孩子率，需 要呼吁适当的生育政策。 Translation In order to reduce the rapidly changing rate of childbirth, it is necessary to call for an appropriate childbirth policy. Table 1: Examples of CSC results, the incorrect and correct characters marked in red and blue respectively. visually or phonologically similar, but with great different semantic meanings. According to (Liu et al., 2010), around 83% and 48% of errors belong to phonological and visual similarity respectively. Moreover, the Chinese language usually consists of many characters without word deliminators, which makes the CSC system must recognize spelling errors based on contextual information, rather than just relying on individual words or characters. Many efforts have been put in the CSC task. Early methods are mainly based …

Content truncated to 500 words