A real-time software defect prediction method based on code representation learning

Abstract

The invention relates to a real-time software defect prediction method based on code representation learning. With the help of the natural language model idea, the problem that traditional neural networks cannot solve long-distance dependencies is solved; first, code change information is reviewed and decomposed into two parts: submission information and code change ; Then encode the code change information to obtain word vector sequences based on word embeddings, paragraph embeddings and position embeddings; then use the attention mechanism to build the Transformer encoder model; Finally, pre-train the previous coding sequence to construct code defect prediction based on code representation learning Model. Compared with the prior art, the present invention uses the advantages of the attention mechanism to construct a Transformer encoder as a model to perform two pre-training tasks, so that the generated language model is more robust, can better represent code change information, and further improves software defects detection rate.

Description

Technical field[0001]The present invention relates to a real-time software defect prediction method based on a code representation, belonging to software analysis and defect prediction technology in software engineering.Background technique[0002]The software defect prediction technology was born in the 1970s, and the main role is to reduce the cost of software, and further protect the development quality of software. In recent years, the development of software field has developed rapidly, and the demand for high quality software at home and abroad has also increased significantly. With the development of software, the coupling of coupling in software development has also produced great hidden dangers. The software system acts as an important part of promoting my country's economic and social development, which will have a huge impact on enterprises and people when facing the above hidden dangers. In order to improve the quality of the software system, developers need to invest a lo...

AI Technical Summary

Problems solved by technology

However, in the existing traditional deep learning software defect prediction methods, there is a problem that it cannot deal with long-distance dependencies. This problem often occurs in software systems, especially in a single source code file. For example, method calls may be separated by thousands of lines. code, which leads to long-distance dependency problems

Images