首页|基于自注意力机制神经机器翻译的软件缺陷自动修复方法

基于自注意力机制神经机器翻译的软件缺陷自动修复方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
循环神经网络对于代码序列数据有着良好的处理能力,软件缺陷修复的补丁生成模型大多采用循环神经网络实现.然而,基于循环神经网络的补丁生成模型在处理代码序列中长距离依赖问题时仍然具有局限性,其修复成功率和修复效率较低.针对此问题,提出一种基于自注意力神经机器翻译的软件缺陷自动修复方法(Self-attention Neural machine translation based automatic software Repair,SNRepair).首先,为有效缓解源码中的未登录词问题,对数据集引入子词切分技术进行预处理;其次,为解决源代码中棘手的长距离依赖问题并更充分地利用局部信息,构建融合局部建模的Transformer程序补丁生成模型;然后,采用缺陷自动定位技术定位缺陷语句位置,利用参数优化后的Transformer补丁生成模型生成候选补丁;最后,运行测试用例验证候选补丁.在具有395个真实Java软件缺陷的De-fects4J缺陷库上实验评估,结果表明SNRepair方法与对比方法比较,修复成功率和修复效率更高.
Self-Attention Neural Machine Translation for Automatic Software Repair
Recurrent neural network has good ability of the processing for code sequences,and the patch generation model is mostly implemented by it.However,recurrent neural network-based patch generation models still have some limi-tations when dealing with long-distance dependencies in code sequences,and their repair success rate and repair efficiency is low.To address the issue,we present SNRepair,an automatic software fault repair based on self-attention neural machine translation.First,the subword tokenization technology is introduced to preprocess the dataset to alleviate the problem of out of vocabulary.Second,a Transformer program patch generation model that integrates local modeling is constructed to alle-viate the long-distance dependencies in the source code and make better use of local information.Third,the automatic fault localization technology is used to locate the possible fault position and the Transformer patch generation model through pa-rameter optimization is adopted to generate candidate patches.Finally,the candidate patches are verified by test cases.On the 395 real Java software faults in the Defects4J,the results show that the SNRepair has higher repair success rate and re-pair efficiency than the compared ones.

automatic software repairneural machine translationself-attention mechanismsubword tokenizationlocalness modeling

曹鹤玲、刘昱、韩栋

展开 >

粮食信息处理与控制教育部重点实验室(河南工业大学),河南郑州 450001

河南工业大学信息科学与工程学院,河南郑州 450001

河南工业大学河南省粮食信息处理国际联合实验室,河南郑州 450001

软件缺陷自动修复 神经机器翻译 自注意力机制 子词切分 局部建模

国家自然科学基金河南省高等学校重点科研项目河南工业大学青年骨干教师培育项目河南省重大公益专项

6160215422A52002421420158201300311200

2024

10.12263/DZXB.20220734

电子学报
中国电子学会

电子学报

CSTPCD北大核心
影响因子：1.237
ISSN：0372-2112
年,卷(期)：2024.52(3)
  • 36