CCTC: A Cross-Sentence Chinese Text Correction Dataset for Native Speakers

Proceedings of the 29th International Conference on Computational Linguistics, pages 3331–3341 October 12–17, 2022. 3331 CCTC: A Cross-Sentence Chinese Text Correction Dataset for Native Speakers Baoxin Wang1,2, Xingyi Duan2, Dayong Wu2,3, Wanxiang Che1, Zhigang Chen2,4, Guoping Hu2 1Research Center for SCIR, Harbin Institute of Technology, Harbin, China 2State Key Laboratory of Cognitive Intelligence, iFLYTEK Research, China 3iFLYTEK AI Research (Hebei), Langfang, China 4Jilin Kexun Information Technology Co., Ltd., Changchun, China {bxwang2,xyduan,dywu2,zgchen,gphu}@iflytek.com car@ir.hit.edu.cn Abstract The Chinese text correction (CTC) focuses on detecting and correcting Chinese spelling er- rors and grammatical errors. Most existing datasets of Chinese spelling check (CSC) and Chinese grammatical error correction (GEC) are focused on a single sentence written by Chinese-as-a-second-language (CSL) learners. We find that errors caused by native speak- ers differ significantly from those produced by non-native speakers. These differences make it inappropriate to use the existing test sets di- rectly to evaluate text correction systems for native speakers. Some errors also require the cross-sentence information to be identified and corrected. In this paper, we propose a cross- sentence Chinese text correction dataset for na- tive speakers. Concretely, we manually anno- tated 1,500 texts written by native speakers. The dataset consists of 30,811 sentences and more than 1,000,000 Chinese characters. It contains four types of errors: spelling errors, redundant words, missing words, and word or- dering errors. We also test some state-of-the-art models on the dataset. The experimental results show that even the model with the best perfor- mance is 20 points lower than humans, which indicates that there is still much room for im- provement. We hope that the new dataset can fill the gap in cross-sentence text correction for native Chinese speakers. 1 Introduction Chinese text correction (CTC) aims at detecting and correcting errors in Chinese text. Text correc- tion has important applications in the domain of education, journalism, and publishing. For many native Chinese speakers, such as journalists, writ- ers, and bloggers, a text correction system for na- tive Chinese speakers will greatly improve the ef- ficiency of their proofreading. In the field of NLP, Chinese text corrections usually includes two tasks: Chinese spelling check (CSC) (Hong et al., 2019; Cheng et al., 2020; Wang et al., 2021) and Chi- non-native examples WRONG: 父母对孩子的爱情在世界上是最重要的。 CORRECT: 父母对孩子的关爱在世界上是最重要的。 TRANSLATION: The love of parents for their children is the most important in the world. WRONG: 这一点可以说是吸烟对个人健康的利益。 CORRECT: 这一点可以说是吸烟对个人健康的好处。 TRANSLATION: This can be said to be a benefit of smoking to the health of the individual. WRONG: 弹奏者只有做到手臂、肘部与腕部都能够完全随着指 尖动作运行，才能有信息弹奏出均匀、流畅的音符， 音色也会更加出挑。 CORRECT: 弹奏者只有做到手臂、肘部与腕部都能够完全随着指 尖动作运行，才能有信心弹奏出均匀、流畅的音符， 音色也会更加出挑。 TRANSLATION: Only when the player is able to run his arms, elbows and wrists completely with the fingertip movements can he have the information confidence to play even, smooth notes and have a more distinguished tone. native examples Figure 1: Comparison between the errors caused by na- tive and non-native speakers. The non-native examples are from CGED 2018, and the native examples are from CCTC. nese grammatical error correction (GEC) (Yuan and Briscoe, 2016; Omelianchuk et al., 2020; Wang et al., 2020). The …

Content truncated to 500 words