Automatic Generation Method Of Handwritten Chinese Character Evaluation Based On Template Extraction

Handwritten Chinese character evaluation is to realize the evaluation of Chinese characters and put forward suggestions on how to improve the writing level according to the difference between the characteristics of handwritten Chinese characters and standard Chinese characters.It plays an important role in primary and secondary schools,and can effectively guide students to understand the defects of handwritten Chinese characters,so as to improve the writing quality of handwritten Chinese characters.The task of data to text generation can use the difference between different handwritten Chinese character features as input to get the evaluation of handwritten Chinese characters.This thesis summarizes the current situation of data to text generation methods,analyzes the relationship between generated text and template,and further considers how to effectively apply deep learning to the automatic generation task of handwritten Chinese character evaluation based on template extraction.（1）Based on Hidden semi-Markov models（HSMM）,a template is extracted for handwritten Chinese character evaluation.The key to generating smooth and controllable text according to the characteristics of handwritten Chinese characters is how to obtain high-quality templates.The manually constructed template constructs the corresponding template according to the possible situation in a specific field,and then the handwritten Chinese character features can be used to select the appropriate template to generate the evaluation.The text generation based on neural network can not directly manipulate the content selection and generation,and there are some problems,such as repeated data records in the output text.In order to overcome this disadvantage,this thesis uses the Hidden semi-Markov model to extract the handwritten Chinese character evaluation template,which can improve the accuracy of text generation,and solves the problem of low segmentation frequency but its importance in the template by adding the replication mechanism,which effectively improves the quality of the generated evaluation text.（2）Based on the Hidden semi-Markov model and replication mechanism,the bidirectional long short term memory based on attention mechanism can effectively improve the accuracy of template generation.Due to the problem of incomplete information acquisition due to the long short term memory（LSTM）before segmentation,the two-way short term memory network based on attention mechanism can overcome the above disadvantages,and give a large weight value to the important information in the handwritten Chinese character evaluation text before segmentation,which improves the accuracy of the generated template and the quality of the generated text.（3）The evaluation text is generated for different number of handwritten Chinese character feature deviation values,and the experimental results are analyzed from multiple angles.In the process of text generation,the generation of current words depends on the words that have been generated before,that is,auto regressive model（AR）,but this condition limits the speed of text generation,resulting in non-auto regressive model（NAR）.According to the above two different generation methods,verify the impact of different template extraction methods on the evaluation text generation under different number of handwritten Chinese character feature deviation values,and confirm the advantages of this method.Furthermore,the commonly used evaluation indexes BLEU-4,CIDER,METER and ROUGE are used for evaluation.The experimental results show that on 10 kinds of handwritten Chinese character feature deviation data sets,using BLEU-4 evaluation index,the text generated by non-auto-regressive and auto-regressive is 0.58 and 0.59 respectively,and on 4 kinds of handwritten Chinese character feature deviation data sets,it is 0.56 and 0.58 respectively.