Misalignment Detector — Mining and Modelling Misaligned Code Change Descriptions Using Large Language Models

Code review plays a key role in software development by helping teams understand why a change was made and how it fits into the broader project. A major part of this understanding comes from the commit message, which is supposed to accurately describe the intent and scope of the corresponding code changes. However, in many real-world projects, commit messages are incomplete, outdated, or simply do not match what the code actually does. This misalignment creates confusion for reviewers and increases the time needed to understand and validate a change.

To address this challenge, many open-source communities discuss and correct misleading commit messages during code review. These discussions provide valuable examples of how commit messages should be improved and why certain messages fail to reflect the code.

This project builds on that real-world practice by creating a dataset of aligned and misaligned commit messages drawn from open-source code review discussions. We then train language models to automatically detect when a commit message does not accurately reflect the underlying changes and to suggest improvements.

The goal is to support developers and reviewers by making commit messages clearer, reducing review time, and improving overall code quality.

Faculty Supervisor:

Moataz Chouchen

Student:

Partner:

National School of Computer Science (ENSI), Tunisia

Discipline:

Computer science

Sector:

Artificial Intelligence

University:

Concordia University