| The rapid growth of user demand has led to the rapid development of mobile Internet.A large number of applications have been developed to meet the needs of users.Understanding the existing program code is the basic step of many software development tasks.How to quickly analyze the functions of the code,to reduce the time of develop and maintenance program as much as possible,has become a hot issue in the field of software engineering.It has very important practical and economic significance.Traditional program classification tasks can only rely on a large number of human resources for manual labeling.Some scholars learn from the field of natural language processing,introduce deep learning into the field of programming language processing,improve the speed of understanding of program code.Previous studies have used traditional recurrent neural networks or convolutional neural networks as classification models.Because the program has strong local semantic structure characteristics and long-term dependence,this strategy effect is not good.Some researchers have proposed to characterize the code by using vector expression based on abstract syntax tree,and to preserve the local semantic structure features in the code through abstract syntax tree.Based on previous research,this paper puts forward a program code classification research model based on ConvTransformer.It preserves the local semantic structure and long-term dependencies of program code by adding location coding to word embedding and using Transformer's attention mechanism.In this paper,the program language category classification experiment and the program function category classification experiment are carried out,in order to compare a variety of models.In the program language category classification experiment.the classification accuracy rate reached 97.5%,and in the program function category classification experiment,the accuracy rate reached 85.6%.These two experiments prove the effectiveness and advancement of the model. |
PDF Full Text Request