基于缺陷报告降噪和抽象语法树表示的软件缺陷定位方法  

Software defect localization method based on defect report denoising and abstract syntax tree representation

作　　者：石翔宇 鞠小林 陈翔[1] SHI Xiangyu;JU Xiaolin;CHEN Xiang(School of Artifical Intelligence and Computer Science,Nantong University,Nantong 226019,China)

机构地区：[1]南通大学人工智能与计算机学院,江苏南通226019

出　　处：《南通大学学报（自然科学版）》2024年第3期23-33,共11页Journal of Nantong University(Natural Science Edition)　

基　　金：国家自然科学基金面上项目(61673384);江苏省现代教育技术研究项目(2022-R-98984)。

摘　　要：自动化的缺陷定位方法能够加快程序员利用缺陷报告定位到复杂软件系统缺陷代码的过程。现有相关缺陷定位方法存在两方面问题:忽略了缺陷报告中噪音信息的影响;程序代码表示的过程中丢失了大量的上下文结构信息。为了解决上述问题,提出一种自动化缺陷定位方法BRFN(bug report fault localization)。该方法首先利用双向的信息传播机制对程序的抽象语法树进行编码;接着针对缺陷报告使用TextCNN和注意力机制学习与缺陷相关的特征;最后通过计算缺陷报告和源代码文件之间的相关性,开展缺陷定位,并基于4个广泛用于缺陷定位研究的软件项目评估BRFN方法的有效性。实验结果表明,相较于BugLocator、LS-CNN和CAST现有的缺陷定位方法,BRFN在多个评价指标上均取得了更好的效果。具体而言,BRFN在4个开源项目上的Acc@1、MRR和MAP性能平均提升了56.3%、43.4%和46%。此外,进一步设计消融实验来验证BRFN中各模块的贡献。结果表明,缺陷报告降噪策略和双向信息传播策略可以提升缺陷定位的准确性。syntax tree of the program using a bidirectional information propagation mechanism.It then employs TextCNN and attention mechanisms to learn defect-relevant features from defect reports.Finally,it calculates the correlation between defect reports and source code files to perform defect localization.The effectiveness of the BRFN method is evaluated based on four widely used software projects for defect localization research.Experimental results show that BRFN outperforms existing methods such as BugLocator,LS-CNN,and CAST across multiple evaluation metrics.Specifically,BRFN improves Acc@1,MRR,and MAP by 56.3%, 43.4%, and 46%, respectively, on four open-source projects. Additionally, ablation experiments are conducted to validate the contribution of each module in BRFN. The results indicate that both the defect report denoising strategy and bidirectional information propagation strategy enhance the accuracy of defect localization.

关 键 词：缺陷定位 深度学习 信息检索 注意力机制 程序表示学习 

分 类 号：TP311.52[自动化与计算机技术—计算机软件与理论]