Research On Chinese Text Error Correction Method Based On Multimodal Information

Chinese text error correction plays an important role in the field of natural language processing.Common types of errors in Chinese text include spelling errors,redundancy,missing,and disorder.Focusing on spelling errors in Chinese text,including semantically confused misspellings and phonetically and graphically similar words,this thesis proposes an error correction method for Chinese text based on multimodal information,which can effectively deal with spelling errors in Chinese text by encoding semantic,phonetic and image information of Chinese characters and aggregating the obtained multimodal information.In addition,the text error correction model needs to be efficient and real-time in practical application scenarios.To this end,this thesis improves the structure of the original BERT model,proposes a generic model acceleration inference method,and uses it as the backbone model of the text error correction method with multimodal information.The main research contents of this thesis are as follows:(1)To address the problem that pre-trained language models have a large delay in the inference process,this thesis proposes a general model acceleration method based on an early exit mechanism.The method can be used in combination with different pre-trained models or in conjunction with existing popular model compression methods to improve the inference speed of text models.The approach used in this thesis uses BERT as the backbone model and measures the confidence level by the entropy of the prediction distribution to achieve the best trade-off between performance and speed.Compared with existing model speedup methods,this method is more flexible and can achieve different speedup ratios by adjusting the patience value parameters and confidence thresholds to improve the inference speed of the model.In addition,the present method can be combined with other pre-trained models to adapt to more downstream tasks.In this thesis,extensive experiments were conducted on six datasets.The experimental results show that this method performs well on several acceleration ratios,with F1 s of 52.8%,83.4%,86.8%,90.5%,91.2%,and 69.7% for the six datasets of Co LA,MNLI,MRPC,QNLI,QQP,and RTE,respectively,at an acceleration ratio of27%,which is better than existing model acceleration methods such as PABEE,BERxi T.The ablation experiments show that the method in this thesis also outperforms the baseline model on other pre-trained models such as ALBERT and Tiny BERT at different acceleration ratios,which proves the generality of the acceleration method proposed in this thesis.(2)Aiming at the phonetic and graphical similarity characteristics of Chinese characters with spelling errors,this thesis proposes an error correction method for Chinese text based on multimodal information fusion.The method encodes semantic,phonetic and graphic information of Chinese characters and aggregates multimodal information to deal with spelling errors more effectively,especially in the case of high similarity of Chinese characters.In this thesis,training and testing are conducted on several textual error correction datasets,such as SIGHAN and CCTC.The experimental results show that this method has an F1 value of 78.9% for detection level and 77.3%for error correction level on the text error correction dataset,which is better than the existing text error correction methods such as Spell GCN and PLOME.The ablation experiments show that.The introduction of different modal information can improve the accuracy of the model error correction,and there is about 3% improvement of F1 value relative to the unimodal model.