String Editing Based Chinese Grammatical Error Diagnosis

Proceedings of the 29th International Conference on Computational Linguistics, pages 5335–5344 October 12–17, 2022. 5335 String Editing based Chinese Grammatical Error Diagnosis Haihua Xie1, Xiaoqing Lyu2, Xuefei Chen3 1Yanqi Lake Beijing Institute of Mathematical Sciences and Applications / Beijing, China 2Wangxuan Institute of Computer Technology, Peking University / Beijing, China 3AI Research Center, Cloopen Group Holding Ltd / Beijing, China haihua.xie@bimsa.cn, lvxiaoqing@pku.edu.cn, cxfff168@163.com Abstract Chinese Grammatical Error Diagnosis (CGED) suffers the problems of numerous types of grammatical errors and insufficiency of training data. In this paper, we propose a string editing based CGED model that requires less training data by using a unified workflow to handle vari- ous types of grammatical errors. Two measures are proposed in our model to enhance the per- formance of CGED. First, the detection and correction of grammatical errors are divided into different stages. In the stage of error de- tection, the model only outputs the types of grammatical errors so that the tag vocabulary size is significantly reduced compared with other string editing based models. Secondly, the correction of some grammatical errors is converted to the task of masked character in- ference, which has plenty of training data and mature solutions. Experiments on datasets of NLPTEA-CGED demonstrate that our model outperforms other CGED models in many as- pects. The project of our approach is available at https://github.com/xiebimsa/se-cged. 1 Introduction The traditional methods of Grammatical Error Di- agnosis (GED) are pipeline-based, which has three main steps in series, including error detection, se- lection of candidate corrections, and ranking of candidate corrections (Lee et al., 2015). With the wide application of seq2seq models, seq2seq based GED models are designed to combine error detec- tion and error correction (Wan et al., 2020; Kiyono et al., 2019). Considering the low efficiency of decoding in seq2seq and the uncontrollability of the generated results, researchers tend to apply the mode of sequence tagging for GED which outputs the correction operations instead of direct correc- tion results. The sequence tagging based GED methods are designed based on edit distance algorithms, such as Levenshtein distance (Levenshtein, 1966), which produces edit operations that are used to trans- form one sentence to the other. Because the in- tersection of an ungrammatical sentence and the corrected sentence is often large, it is feasible to transform them to each other based on a few edit operations. Several string editing based GED meth- ods for English have been recently proposed, such as LaserTagger (Malmi et al., 2019), PIE (Awasthi et al., 2019) and GECToR (Omelianchuk et al., 2020). The string editing based GED methods are highly accurate, controllable and interpretable (Ro- zovskaya and Roth, 2021; Parnow et al., 2021). Fine-grained data processing and edit tags design- ing are crucial. However, the size of the tag vocab- ulary is generally very large, which makes the cur- rent string editing based GED methods work better on closed sets, but difficult to deal with the com- plexities and long-tail problems on open datasets (Omelianchuk et al., 2020). The current string …

Content truncated to 500 words