Research On Code Generation Based On Deep Learning

With the rapid development of computer technology,demand for new software is exploding.But the development speed of existing software is far behind the speed of people's demand for new software.In order to alleviate this situation,code generation technology came into being.Code generation can assist programmer in software development,reducing the coding burden of developers and shorten the development cycle of software.At present,researchers have proposed many methods of code generation.Firstly,we analyze the current situation of code generation and neural network architecture for code generation,and then analyze the challenges and problems of them,and pay attention to feasible method to improve the accuracy of code generation.Next,the components of a new type of neural network architecture called Transformer that based on the attention mechanism are introduced in detail,and compared with the traditional neural network structure like CNN and RNN.Draw the conclusion that Transformer architecture has advantages and is more suitable for code generation.Based on this conclusion,we put forward a method of code generation based on Transformer for the first time.This method combines Transformer architecture with the benefits of encoder-decoder structure to further optimize the model structure and improving the training speed of the model.Finally,we conducted experiments on three typical benchmark that used for code generation and compared them with the previous methods,which showing that this method can effectively improve the accuracy and BLEU score of code generation.The code generation can effectively help programmer to carry out software development,so the efficiency of the code generation can further shorten the development cycle of new software.This paper proposed an new code generation method based on deep learning,which can effectively improve the accuracy and BLEU score of code generation.