Research On Technology And Implementation Of Fine-grained Software Bug Localization

Software bug localization is an important research field in software engineering.Due to the increasing number of bugs and the size of the code,how to automatically localize the bugs has become one of the hotspots that researchers have focused on in recent years.The bug report-oriented bug localization approach uses the bug report as a query and the source code of the project as the corpus,and identifies the source code unit corresponding to the bug by analyzing the correlation between the bug report and the source code unit.The current method-level fine-grained software bug localization techniques consider methods as independent code units,ignoring the affiliation between methods and files,resulting in low localization accuracy.In addition,although there has been a lot of research work on bug localization,there is a lack of practical bug localization tools.In order to improve the accuracy of fine-grained bug localization,this thesis studies a fine-grained software bug localization approach and builds a bug localization tool.This work mainly includes the following aspects:(1)We proposed FMBL,a method-level software bug localization technology enhanced with file information.It considers the dependencies between methods and documents to enhance the accuracy of bug localization.In addition,FMBL measures the correlation between code and bug reports by comprehensively considering their lexical similarity,semantic similarity,and code length.We conduct experimental studies on six open-source software projects to evaluate the effectiveness of the approach,and the average Accuracy@10,MAP and MRR on the six projects reach 0.436,0.223 and 0.296,respectively.Compared with BugLocator_m,BLIA,and BugPecker,FMBL has an increase of 153.1%,209.1%,and 22.8%in terms of MAP indicators,respectively.(2)We proposed a query reformulation approach for bug localization based on the pre-training model.We trained a CodeBERT model through a large amount of historical bug data,and use it to supplement relevant bug information for bug reports.The reconstructed query is then been input to FMBL for bug localization,called FMBL-QR.Compared with FMBL,FMBL-QR can identify relevant code units more accurately,and the average MAP on the six experimental projects is increased by 7.2%.(3)A bug localization tool is built,which integrates and encapsulates the functions of automatic collection of historical bug information,automatic bug information processing,bug dataset construction and automatic bug localization.It has been released as a VSCode plug-in for easy usage.