Patch Validation Approach Combining Doc2Vec and BERT Embedding Technologies

Abstract

Automatic program repair is a research hotspot in recent years and has made some progress.Most of the existing automatic program repair methods use the test suite to validate patch correctness.However,using the test suite to validate a large number of candidate patches will not only bring huge costs,but also lead to the overfitting problem of patches.Therefore,how to improve the efficiency of patch validation and effectively validate patch correctness has become an urgent problem.In order to reduce the cost and improve the patch accuracy,this paper proposes an approach combining two embedding techniques to validate patch correctness.Firstly,this approach uses Doc2Vec model to calculate the similarity between the patch and the error code,then it uses the classifier based on BERT model to filter out the error patches from the patches screened by the similarity.To evaluate the effectiveness of this approach,experiments are carried out based on five open source Java benchmarks.Experimental results show that this approach can effectively validate patch correctness and improve the efficiency of patch validation.

Keywords

• automatic program repair|patch validation|code similarity|embedding technology