DOI
https://doi.org/10.25772/Y93V-HV11
Defense Date
2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Computer Science
First Advisor
Dr. Kostadin Damevski
Abstract
Modern software development is extremely collaborative and agile, with unprecedented speed and scale of activity. Popular trends like continuous delivery and continuous deployment aim at building, fixing, and releasing software with greater speed and frequency. Bug localization, which aims to automatically localize bug reports to relevant software artifacts, has the potential to improve software developer efficiency by reducing the time spent on debugging and examining code. To date, this problem has been primarily addressed by applying information retrieval techniques based on static code elements, which are intrinsically unable to reflect how software evolves over time. Furthermore, as prior approaches frequently rely on exact term matching to measure relatedness between a bug report and a software artifact, they are prone to be affected by the lexical gap that exists between natural and programming language.
This thesis explores using software changes (i.e., changesets), instead of static code elements, as the primary data unit to construct an information retrieval model toward bug localization. Changesets, which represent the differences between two consecutive versions of the source code, provide a natural representation of a software change, and allow to capture both the semantics of the source code, and the semantics of the code modification. To bridge the lexical gap between source code and natural language, this thesis investigates using topic modeling and deep learning architectures that enable creating semantically rich data representation with the goal of identifying latent connection between bug reports and source code. To show the feasibility of the proposed approaches, this thesis also investigates practical aspects related to using a bug localization tool, such retrieval delay and training data availability.
The results indicate that the proposed techniques effectively leverage historical data about bugs and their related source code components to improve retrieval accuracy, especially for bug reports that are expressed in natural language, with little to no explicit code references. Further improvement in accuracy is observed when the size of the training dataset is increased through data augmentation and data balancing strategies proposed in this thesis, although depending on the model architecture the magnitude of the improvement varies. In terms of retrieval delay, the results indicate that the proposed deep learning architecture significantly outperforms prior work, and scales up with respect to search space size.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
5-3-2022