Title: High-Quality Automated Program Repair

Abstract: Automatic program repair (APR) has recently gained attention because it proposes to fix software defects with no human intervention. To automatically fix defects, most APR tools use the developer-written tests to (a) localize the defect, and (b) generate and validate the automatically produced candidate patches based on the constraints imposed by the tests. While APR tools can produce patches that appear to fix the defect for 11-19% of the defects in real-world software, most of the patches produced are not correct or acceptable to developers because they overfit to the tests used during the repair process. This problem is known as the patch overfitting problem. To address this problem, I propose to equip APR tools with additional constraints derived from natural-language software artifacts such as bug reports and requirements specifications that describe the bug and intended software behavior but are not typically used by the APR tools. I hypothesize that patches produced by APR tools while using such additional constraints would be of higher quality. To test this hypothesis, I propose an automated and objective approach to evaluate the quality of patches and propose two novel methods to improve the fault localization and developer-written test suites using natural-language software artifacts. Finally, I propose to use my patch evaluation methodology to analyze the effect of the improved fault localization and test suites on the quality of patches produced by APR tools for real-world defects.