From Symptoms to Fixes: Towards Review-Driven Automated Program Repair

## About This ProjectModern automated program repair (APR) techniques mainly rely on structured supervision such as bug reports and test cases. However, in real-world scenarios, many software failures are first reported by end users through app reviews, where defects are described as observable symptoms rather than technical causes.To address this gap, we construct a **review–issue–patch aligned dataset** that connects user reviews, developer-reported issues, and corresponding bug-fixing commits from open-source Android applications. This dataset captures cross-level relationships between symptom-level descriptions, technical issue reports, and code-level fixes, enabling the study of defect localization and patch generation under weak supervision.Since explicit links between these artifacts are not available, we establish connections through semantic alignment, combining representation learning and structured symptom matching. As a result, the dataset provides weakly aligned but practically meaningful instances that bridge user-perceived failures and developer-level repair actions.Based on this dataset, we further propose **ReAPR**, a review-augmented program repair framework that models user feedback as semantic constraints and leverages cross-level alignment to guide repair generation.