- All Journals
- Al-Qadisiyah University
- Basrah University
- Kirkuk University
- Mosul University
- The Iraqi Borad for Medical Specialization
- University of Anbar
- University of Technology
- Al Mustafa University
- Al-Zahraa University for Women
- Alnoor University
- Kunooz University
- Misan University
- Ninevah University
- Tikrit University
- University of Al-Hamdaniya
- University of Kerbala
Investigating the Impact of Source Code Granularity on File-Level Bug Localization Using Pre-trained Code Models | ||
| Al-Noor Journal for Information Technology and Cybersecurity | ||
| Articles in Press, Corrected Proof, Available Online from 26 September 2026 | ||
| Document Type: Original Article | ||
| DOI: 10.69513/njitcs.2026.174817.1078 | ||
| Authors | ||
| tuqa emad Husein1; mohammad abdulghani taha taha aldabbagh* 2 | ||
| 1موظف حكومي | ||
| 2University of Mosul College of Computer Science and Mathematics | ||
| Abstract | ||
| A file-level bug localization is a tool for finding and ranking the source code files that are most relevant to a bug report so as to reduce the developers' debugging effort. Pre-trained code models have been shown to be efficient in capturing the semantic relationship between bug reports and source code. However, little attention has been paid to the impact of source code granularity on the performance of these code models. Here we investigate how source code representation granularities, AST Chunk and Method-Level AST, affect pre-trained code models for file-level bug localization. We conducted experiments using the Defects4J benchmark which consists of five open-source Java projects: Chart, Closure, Lang, Math, and Time. Four popular pre-trained models, CodeBERT, GraphCodeBERT, UniXcoder, and CodeT5, were tested under both source code representations for the same experimental settings. The retrieved files were evaluated using Top@1, Top@2, Top@3, Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (nDCG). Our experimental results show the comparison between the two source code granularities across projects and models, and the impact the code representation has on bug localization. The results are useful for choosing the source code granularity for pre-trained code models for file-level bug localization and to better understand the impact of code representation on file-level bug localization. | ||
| Keywords | ||
| Bug Localization; Defects4J; Pre-trained Code Models; Source Code Granularity; Abstract Syntax Tree (AST) | ||
|
Statistics Article View: 2 |
||