Chinese Research On The Method Of Text Error Detection And Error Correction Under Speech Transcription

The goal of text error is to detect and correct errors in the text sequence.The type of text error of voice transcription is mainly caused by voice recognition errors and spelling errors.Commonly deep-learning-based representative learning methods can effectively enhance the error correction ability of text,but there are still a large number of signs of text for learning,excessive dependence on the context representation,the model can only be one or two fonts that cannot be paired with the fragment error text.The problem of identification and correction.For this reason,this article proposes an error correction model based on the fusion fuzzy entity and text intention based on Monte Carlo(Ecmc-Fite Model).Firstly,use Monte Carlo’s optimization of model uncertainty.Introduce MCDropout(Monte-Car Lo Dropout)to replace the training method of GM’s random daily Dropout,and find that the performance on different data sets is better than other algorithms.Secondly,detecting and error correction research in the error entity in the text sequence.For the problem that the segmental error text entities cannot effectively correct,a fuzzy physical detection correction model is designed.Adaptively replace the physical fonts randomly and conduct training.Based on Chinese and Pinyin for corrective training,the ability to effectively improve the ability to detect and correct fuzzy entity.The public data sets in the extraction task achieved a good effect.Thirdly,research on the distribution of text intentions.In response to the problems that cannot be corrected by multiple font errors at the same time,the text intention method is proposed to distribute the training sample to add random errors for training,which effectively alleviates the error correction method expressed by the context.Use multi-tasking to learn the combined training ECMC-FITE model to decompose tasks into detection and error correction of entities and non-real texts.In the open dataset experiments such as conventional text OCR＿DATASET and Sighan,ECMC-FITE can also effectively classify the intention of the text and correct the corresponding error words.The accuracy can reach more than 80%,leading ahead of other comparison models.In the practical application of voice recognition dialogue and question and answer error,the experimental results show that the accuracy of 69.1%is also better than the accuracy of other models.