文献类型：期刊文献

英文题名：On the Effectiveness of Information Retrieval-Based Bug Localization on Deep Learning Frameworks: An Empirical Study

作者：Tang, Jianhao[1]; Fan, Guisheng[1]; Yu, Huiqun[1]; Ge, Jian[1]; Huang, Zijie[1]

机构：[1] Department of Information Science and Engineering, East China University of Science and Technology, 130 Meilong Road, Shanghai, 200237, China

年份：2023

外文期刊名：SSRN

收录：EI(收录号：20230277708)

语种：英文

外文关键词：Computer software selection and evaluation - Data mining - Deep learning - Information retrieval - Learning systems - Open source software - Quality assurance

摘要：Context: The demand for deep learning frameworks is rising, but defects in these frameworks can hinder model training and performance, making software quality assurance vital. Efficient error recording, analysis, and localization are essential for early detection and improvement of software quality assurance. Objective: Information Retrieval (IR) is crucial for bug localization, but low-quality bug reports can compromise accuracy. Extracting high-quality information from bug reports is essential. This paper proposes a method that combines traditional topic modeling with Information Retrieval-based Bug Localization (IRBL) to enhance bug localization accuracy in deep learning frameworks. Method: Bug reports are treated as queries, and keyword extraction is used to filter non-keywords. A topic model is applied to compute similarity between bug reports and code files, identifying actual defect code files and bug localization. Performance is compared with baselines. Results: Experiments on five open-source deep learning frameworks show that the proposed method outperforms baseline topic modeling in defect prediction across frameworks. Variations in critical keywords influencing bug localization are observed. Conclusion: Extracted keywords from framework documentation and bug reports improve bug localization accuracy by enhancing the similarity between bug reports and actual defect code. ? 2023, The Authors. All rights reserved.