Research On Chinese Text Error Correction

Chinese text error correction is a challenging task in Chinese natural language processing,which aims to detect and correct textual errors made by users when using Chinese.With the development of natural language processing and the emergence of seq2 seq models,Chinese text error correction research has made significant progress.However,the size of the existing annotated corpus for the Chinese grammatical error correction task is still difficult to support the training of seq2 seq models.Therefore,recent research has mainly focused on data augmentation strategies,and limited research tried to adapt the existing corpus size by reducing the training difficulty of the models.In addition,the current evaluation system for Chinese grammatical error correction is inadequate,and the evaluation results are to some extent influenced by the Chinese word segment algorithm,so a more reasonable evaluation system for Chinese grammatical error correction tasks needs to be explored.For the Chinese spelling error correction task,the existing research only focuses on the correction ability of the model,but neglects to pay attention to the model’s ability of identifying confusing words.In order to provide a more reasonable evaluation method for Chinese grammatical error correction tasks and to solve the problem in training Chinese grammatical error correction models,this paper investigates the evaluation framework and model optimization methods for Chinese grammatical error correction tasks and uses spelling error correction models to assist grammatical error correction tasks.The main research of the thesis includes:(1)An evaluation framework for Chinese grammatical error correction.In this paper,the test set provided by NLPCC is manually annotated and expanded to address the shortcomings of the existing evaluation metric and evaluation framework,so that the gold-standard correction annotations of the original test set samples can correspond to more correction results.To address the shortcomings of the Chinese grammar error correction evaluation system,three new evaluation metrics are proposed for the Chinese grammatical error correction task.(2)Transformer-based grammatical error correction model.This paper introduces the application of the Transformer model to grammatical error correction tasks,and designs several strategies to improve the effectiveness of the grammatical error correction model,and verifies the superior performance of the Transformer model and the effectiveness of the designed strategies through comparative experiments.(3)A grammatical error correction model based on model compression and grammatical generalization.This paper proposes a Chinese grammatical error correction model that is easier to learn.Experimental results validate the effectiveness of the proposed model.In addition,the effectiveness of the proposed model is also verified on the English grammatical error correction dataset.Meanwhile,ablation experiments and parameter exploration experiments were conducted in this paper to verify the effectiveness of each strategy and the optimal values of the parameters.(4)A spelling error correction model based on reverse contrastive learning.In this paper,a reverse contrastive learning strategy is proposed and applied to build a spelling error correction model to assist in improving the performance of grammatical error correction.The reverse contrastive learning strategy is described in detail,and the constructed spelling error correction model is used in the pre-processing and postprocessing stages of the Chinese grammatical error correction task to verify the effectiveness of the spelling error task for improving the effectiveness of the Chinese grammatical error correction task.This paper achieves certain performance improvements in both Chinese grammatical error correction and Chinese spelling error correction tasks,and provides new research ideas and research directions for Chinese grammatical error correction and Chinese spelling error correction tasks.