Improving Chinese Grammatical Error Detection via Data augmentation by Conditional Error Generation

Findings of the Association for Computational Linguistics: ACL 2022, pages 2966 - 2975 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Improving Chinese Grammatical Error Detection via Data augmentation by Conditional Error Generation Tianchi Yue, Shulin Liu, Huihui Cai, Tao Yang, Shengkang Song, Tinghao Yu Tencent AI Platform Department, China {tianchiyue, forestliu, helenhhcai}@tencent.com {rigorosyang, shengksong, maxwellyu}@tencent.com Abstract Chinese Grammatical Error Detection(CGED) aims at detecting grammatical errors in Chi- nese texts. One of the main challenges for CGED is the lack of annotated data. To alleviate this problem, previous studies pro- posed various methods to automatically gen- erate more training samples, which can be roughly categorized into rule-based methods and model-based methods. The rule-based methods construct erroneous sentences by di- rectly introducing noises into original sen- tences. However, the introduced noises are usually context-independent, which are quite different from those made by humans. The model-based methods utilize generative mod- els to imitate human errors. The generative model may bring too many changes to the original sentences and generate semantically ambiguous sentences, so it is difficult to de- tect grammatical errors in these generated sen- tences. In addition, generated sentences may be error-free and thus become noisy data. To handle these problems, we propose CNEG, a novel Conditional Non-Autoregressive Error Generation model for generating Chinese grammatical errors. Specifically, in order to generate a context-dependent error, we first mask a span in a correct text, then predict an erroneous span conditioned on both the masked text and the correct span. Further- more, we filter out error-free spans by measur- ing their perplexities in the original sentences. Experimental results show that our proposed method achieves better performance than all compared data augmentation methods on the CGED-2018 and CGED-2020 benchmarks. 1 Introduction The goal of Grammatical Error Detection is to de- tect grammatical errors in texts (Rao et al., 2018). It is useful for many NLP applications such as writ- ing assistant (Napoles et al., 2017), search engine (Gao et al., 2010), and speech recognition systems (Wang et al., 2020a), etc. Grammatical errors may Figure 1: An error-correction pair from CGED datasets. The first line is an erroneous sentence, tokens marked in blue color are selection errors, tokens marked in green color are redundant words. The second line is the corrected sentence. appear in all languages (Dale et al., 2012; Bryant et al., 2019). In this paper, we only investigate the problem of Chinese Grammatical Error Detection (CGED). Grammatical Error Detection is usually formu- lated as a sequence tagging task, where each er- roneous token is assigned with an error type, e.g., selection errors and redundant words, as shown in Figure 1. Since annotating grammatical errors re- quires rich linguistic knowledge, it is expensive and time-consuming to annotate a large-scale corpus. Therefore, the scarcity of labeled data is one of the main challenges for this task. To handle this prob- lem, previous works proposed various data aug- mentation methods to automatically generate more training samples (Kiyono et al., 2019; Wang et al., 2019; Lichtarge et …

Content truncated to 500 words