Chinese Spelling Check Based On Neural Network

The Chinese spelling Check task is to detect and correct misspelled Chinese characters in sentences,i.e.typos.It has a wide range of applications in daily life and work.For example,it can help search engines to disambiguate keywords,correct typos in voice and optical character recognition,and assist in automatic correction of essays.Meanwhile,Chinese spelling error correction is an extremely challenging task because of the diversity and complexity of the combination of Chinese characters,and the wrong characters can greatly interfere with the semantics of sentences.Therefore,an effective Chinese spelling error correction solution often requires human-level semantic understanding.In the early days,researchers tackled this problem with an end-to-end machine translation model,trying to directly translate wrong sentences into correct sentences.However,as a general sequence-to-sequence model,this method does not specifically consider how to utilize the similarity information of erroneous characters.In response to this problem,this paper launched the first study,trying to use contrastive learning to obtain the similarity information of wrong characters,so that the model can learn more wrong character information.Additionally,research in recent years has focused on combining BERT with external knowledge such as confusion sets and phonetic and morphological features of characters to solve this task.However,the lack of typo detection ability of BERT itself is still the performance bottleneck of such methods.Therefore,this paper develops a second study,which proposes the use of ”input saliency” technique to identify character errors and integrate the error information into the correction model to improve the detection and correction ability of the whole model.The main contributions of this paper can be summarized as follows:· This paper proposes an end-to-end Chinese spelling error correction model based on contrastive learning and pointer networks,namely Contr PN.In order to utilize the similarity information of wrong characters,this paper first uses the confusion set to construct sentences similar to the original wrong sentence,and then uses contrastive learning to make the original wrong sentence and the wrong sentence formed by the confusion set have similar representations,and at the same time make the original wrong sentence and the wrong sentence.The representation distance of random sentences in the training batch becomes farther,so that characters belonging to the same confusion set have more similar representations,increasing the probability of correcting wrong characters.In addition,using a pointer network based on the Seq2 Seq model to let the model learn to copy the correct characters in the sentence instead of generating candidate words from the vocabulary can reduce the false correction rate.· This paper proposes a Chinese spelling Check model based on ”salient information”,namely Spel LM.Aiming at the lack of BERT's own typo detection ability,this paper imitates the way humans correct spelling errors to alleviate this problem.Specifically,humans first identify misspelled characters based on contextual information and then replace them with correct characters.Therefore,this paper proposes a detection-correction two-stage model,where the model identifies spelling errors in the first stage,and then corrects the errors with the help of error detection information in the second stage.The model uses the ”input saliency” technique to identify character errors and integrate more precise error information into the correction model BERT to improve the detection and correction capabilities of the entire model.This method is independent of existing techniques such as similarity filters,confusion sets,phonetic and morphological features,or candidate-dependent features,making it independent and flexibly combined with existing methods and more efficient.· This paper applies the interpretable techniques of deep models to deep models,showing how explainable artificial intelligence(XAI)techniques can be used to improve the performance of deep learning models on specific tasks.Through the task of Chinese spelling error correction,this paper uses the input saliency technology in the Spel LM model to extract the saliency information of characters for sentence error prediction.The higher the saliency is,the more likely it is to be a wrong character.Combined with the wrong model,the success rate of correcting wrong characters is improved.Although Chinese spelling error correction is the key task of this research,we believe that this idea can be transferred to solve other related tasks.To sum up,for the Chinese spelling error correction task,this paper focuses on how to effectively utilize the correlation information of erroneous characters in an end-to-end error correction model,and how to further improve the error detection capability of the BERT-based error correction model.In response to these two problems,this paper proposes their own improvement schemes,namely the above-mentioned Contr PN model and Spel LM model.Combined with the benchmark model,two evaluation matrices are used to evaluate on multiple test sets,which proves the effectiveness and feasibility of the improved scheme.