DeepTriage:一种基于深度学习的软件缺陷自动分配方法  

DeepTriage: An Automatic Triage Method for Software Bugs Using Deep Learning

作　　者：宋化志 马于涛[1] SONG Hua-zhi;MA Yu-tao(School of Computer Science,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学计算机学院,武汉430072

出　　处：《小型微型计算机系统》2019年第1期126-132,共7页Journal of Chinese Computer Systems

基　　金：国家重点基础研究发展计划项目(2014CB340404)资助;国家自然科学基金项目(61672387;61702378)资助;武汉市黄鹤英才(现代服务)计划项目资助

摘　　要：在软件开发和维护过程中,缺陷修复工作有一项必不可少的任务,那就是缺陷分配.在大规模的软件项目中,基于文本分类的自动分配技术已被用于提高缺陷分配的效率,从而减少人工分配的等待时间和成本.考虑到缺陷报告文本内容的复杂性,本文提出了一种基于深度学习的缺陷自动分配方法,在词向量化后通过卷积神经网络对缺陷报告文本进行特征提取,然后完成分类任务.在Eclipse和Mozilla两个数据集上的结果表明,与传统的支持向量机和基于递归神经网络的方法相比,文本所提方法在准确率指标上均优于上述基准方法,而且多层平行的卷积神经网络结构比单层的卷积神经网络结构在预测效果上更好.Triaging software bugs is an essential task of bug repair during the process of software development and maintenance. In a fewlarge-scale software projects,automated bug triage techniques based on text categorization have been used to reduce the time and cost of manual allocation. Considering the complexity of the contents of bug reports,we propose an automatic triage method for bugs using deep learning,called DeepTriage. By using a convolutional neural network(CNN),Deep Triage extracts text features of each bug report represented by word embedding and then predicts possible software developers for a given bug. The experimental results on the Eclipse and Mozilla datasets showthat Deeptriage performs better than the traditional support vector machine and recurrent neural networks in term of accuracy. In particular,the multi-level parallel CNN outperforms the single-layer one.

关 键 词：缺陷分配 深度学习 卷积神经网络 递归神经网络 支持向量机 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]