Chunk-based Chinese Spelling Check with Global Optimization

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2031–2040 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2031 Chunk-based Chinese Spelling Check with Global Optimization Zuyi Bao, Chen Li and Rui Wang Alibaba Group {zuyi.bzy,puji.lc,masi.wr}@alibaba-inc.com Abstract Chinese spelling check is a challenging task due to the characteristics of the Chinese lan- guage, such as the large character set, no word boundary, and short word length. On the one hand, most of the previous works only con- sider corrections with similar character pro- nunciation or shape, failing to correct visu- ally and phonologically irrelevant typos. On the other hand, pipeline-style architectures are widely adopted to deal with different types of spelling errors in individual modules, which is difficult to optimize. In order to handle these issues, in this work, 1) we extend the tra- ditional confusion sets with semantical candi- dates to cover different types of errors; 2) we propose a chunk-based framework to correct single-character and multi-character word er- rors uniformly; and 3) we adopt a global op- timization strategy to enable a sentence-level correction selection. The experimental results show that the proposed approach achieves a new state-of-the-art performance on three benchmark datasets, as well as an optical char- acter recognition dataset. 1 Introduction Spelling check is a task to automatically detect and correct spelling errors in human writings. Spelling check is well-studied for languages such as English, and many resources and tools have been developed. However, the characteristics of the Chinese lan- guage make the Chinese spelling check (CSC)1 quite different from the English one in three as- pects: • In contrast to English words that are com- posed of a small set of Latin letters, Chinese has more than three thousand frequently used 1As Chinese spelling check involves both error detection and correction, we do not distinguish between spelling check and spelling correction in this paper. characters. The large character set leads to a huge search space for the CSC models. • For English spelling check, the basic unit is the word. However, Chinese characters are continuously written without word delimiter, and the word definition varies across differ- ent linguistic theories (Xue, 2003; Emerson, 2005). It makes the sentence with spelling errors more ambiguous, and more challenging for the spell checkers to detect and correct the errors. • Chinese words usually consist of one to four characters and are much shorter than the En- glish word. Spelling errors can drastically change the meaning of the word. Thus, the CSC task relies on the contextual semantic information to find the best correction. For the first challenge, previous research demon- strates that most of the Chinese spelling errors come from similar pronunciations, shapes, or mean- ings (Liu et al., 2011; Chen et al., 2011). Previous CSC models usually employ the characters with similar pronunciation or shape as the confusion set to reduce the search space, but the visually and phonologically irrelevant typos cannot be handled. Recent work aims at replacing the pronunciation and shape confusion sets …

Content truncated to 500 words