UCSC-CGEC: A Unified Approach For Chinese Spelling Check And Grammatical Error Correction

• Published in: 2024 International Joint Conference on Neural Networks (IJCNN) pp. 1 - 8
• Main Authors: Su, Jindian, Xie, Yunhao, Mou, Yueqi
• Format: Conference Proceeding
• Language: English
• Published: IEEE 30-06-2024
• Subjects: Chinese Grammatical Error Correction; Chinese Spelling Check; Copy Mechanism; Costs; Data integrity; Error correction; Natural language processing; Neural networks; Parameter-Efficient Fine-Tuning; Sequence-to-Sequence Model; Training; Transformers
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN60899.2024.10651014

Chinese Spelling Check (CSC) and Chinese Grammatical Error Correction (CGEC) are two important and challenging tasks in the Natural Language Processing (NLP) field. The former aims to detect and correct Chinese misspellings while the latter focuses on grammatical errors in sentences. Existing methods treat them as two separate tasks, sequence labeling, and conditional text generation respectively. As a consequence, a single encoder is typically selected as the backbone network to handle the CSC task whereas an encoder-decoder structure becomes a requisite for the CGEC task. However, in real-world applications, it is inefficient for a system to determine whether an input sentence contains spelling or grammatical errors and subsequently select different models according to the decision from the previous step. In this paper, to address these two tasks effectively, we propose a unified approach, denoted as UCSC-CGEC, based on a standard Transformer encoder-decoder structure. Notably, we choose to use a recent dataset named CSCD-IME instead of SIGHAN to ensure higher data quality in the CSC task. Additionally, to reduce the training difficulty and enhance generation quality, we introduce Copy Mechanism. Furthermore, to improve training efficiency and reduce cost, we adopt AdaLoRA, a Parameter-Efficient Fine-Tuning (PEFT) method, rather than fine-tuning the model with entire parameter set during the training phase. Experiments are conducted on CSCD-IME and NLPCC2018 datasets, and the results indicate the superiority of our approach when compared to all baseline models.