علي مجيدزاده

Software traceability, is the ability to find the links between the artifacts created during software development. Some artifacts include code, requirements and tests. Software traceability links help to improve software quality and code understandability. Since developers don't prioritize on maintaining traceability links manually, automatic traceability link recovery is an important challenge and researchers are trying to propose methods to recover traceability links with high precision and recall. The research work done on traceability link recovery is focused on requirements-to-code and test-to-code links. Different information retrieva‎l and machine learning tools are used to recover traceability links. Also, information extracted from code can be used in the case of test-to-code traceability links, because of the test cases being code themselves. Finding requirement traceability links is essential for code understandability, impact analysis and software development management, while test-to-code traceability links make the changes made to the system safer. Among the latest related work done on requirement-to-code traceability, CodeBERT that is a pretrained language model has been used. This research focuses on making the most out of BERT for requirement-to-code traceability and to improve it for a dataset extracted from software projects with issue-to-method format that is more practical for requirement traceability uses. To use language model more efficiently, code data augmentation and 3-stage training using the extracted data are proposed. Data augmentation for code data is a new way to improve training because of its difficulty in code data. Code data augmentation in this research includes rename variable, swap operands and swap statements. Among these 3 methods, rename variable produces the most amount of data and swap statements produces the least, even though swap statements is the most sophisticated among the 3, but the transformations done by it are limited because of the lack of semantic information from outside the method that's being transformed and being limited to the method code itself. In 3-stage training, instead of one intermediate training and one finetuning, the pretrained language is trained in 2 intermediate training stages with more relevant data in the second stage and 1 final finetuning with the data extracted from one software project for usage on that project. eva‎luation results show that data augmentation improves the model in some cases and 3-stage training using the extracted dataset significantly improves the model for recovering issue-to-method links. Using the proposed method, precision of 0.706 and recall of 0.725 for CodeSearchNet doc-to-code data, 0.706 average precision and 0.725 average recall for issue-to-commit, and 0.51 average precision and 0.545 average recall for issue-to-method data extracted from 14 projects have been achieved.

ردپذيري نرم‌افزار , ردپذيري موضوع به كد , مدل زبان , افزايش داده

Software Traceability , Issue-to-Code Traceability , Language Model , Data Augmentation

Ali Majidzadeh

Dr. Mehrdad Ashtiani