Dynamic Assessment-Based Curriculum Learning Method for Chinese Grammatical Error Correction
Ruixue Duan,Zhiyuan Ma,Yangsen Zhang,Zhigang Ding,Xiulei Liu
DOI: https://doi.org/10.3390/electronics13204079
IF: 2.9
2024-10-18
Electronics
Abstract:Current mainstream for Chinese grammatical error correction methods rely on deep neural network models, which require a large amount of high-quality data for training. However, existing Chinese grammatical error correction corpora have a low annotation quality and high noise levels, leading to a low generalization ability of the models and difficulty in handling complex sentences. To address this issue, this paper proposes a dynamic assessment-based curriculum learning method for Chinese grammatical error correction. The proposed approach focuses on two key components: defining the difficulty of training samples and devising an effective training strategy. In the difficulty assessment phase, we enhance the accuracy of the curriculum sequence by dynamically updating the evaluation model. During the training strategy phase, a multi-stage dynamic progressive approach is employed to select training samples of varying difficulty levels, which helps prevent the model from prematurely converging to local optima and enhances the overall training effectiveness. Experimental results on the MuCGEC and NLPCC 2018 Chinese grammatical error correction datasets show that the proposed curriculum learning method significantly improves the model's error correction performance, with F0.5 scores increasing by 0.9 and 1.05, respectively, validating the method's effectiveness.
engineering, electrical & electronic,computer science, information systems,physics, applied