Graph-based statement-level program repairing method and system

Abstract

The invention discloses a graph-based statement-level program repairing method and system, and belongs to the field of software debugging. The method comprises the following steps of: firstly, extracting defect codes, patches and standard codes to construct training and pre-training data sets; preprocessing the data sets and pre-training a programming language model; performing data embedding by utilizing the programming language model, and constructing and training a translation model based on a Graph-to-Sequence architecture; and generating a patch of a defect statement by using the trained translation model. According to the invention, codes are represented by using a code graph fusing various features of a source code, the code specification is learned by using a pre-trained model, the training convergence speed of the translation model is increased, the context representation of a defect statement can be optimized, the translation model can better learn grammar and semantic association information between the defect statement and a correct statement, so that the semantic meaning of defect repair is better represented, a high-quality repair patch following the programming language specification is generated to automatically repair a defect program, and the cost of defect repair can be greatly reduced.

Description

technical field [0001] The invention belongs to the field of software debugging, in particular to a graph-based statement-level program repair method and system. Background technique [0002] Program defects are inevitable in the software development process, and developers need to spend a lot of energy to fix these defects. As the scale of modern software continues to increase, the number of program defects and the difficulty of repairing increase accordingly, and program defects have caused huge economic losses to enterprises. To improve software reliability and reduce development costs, researchers have proposed many automatic program repair (APR) techniques to automate defective programs. [0003] Traditional defect repair methods rely on expert knowledge and require domain experts to spend a lot of effort to build repair templates or repair strategies, so they do not have the ability to generalize. Due to the recurring characteristics of software defects, the research...

AI Technical Summary

Problems solved by technology

Although NMT-based methods have great advantages over traditional techniques, there are still deficiencies in such methods.

The code representations adopted by current NMT-based methods are still unable to preserve rich syntactic and semantic information

Meanwhile, since they tend to represent the source code as a sequence and apply a sequence-to-sequence model to generate patches, such methods ignore the implicit semantics in the source code

Additionally, these models learn inefficiently when the input sequence is too long

Images