The icing on the cake: Enhancing Code Pre-trained models-based program repair with multi-source inputs and fix templates

Automated Program Repair (APR) aims to automatically repair software bugs with minimal or no human intervention, enhancing software reliability. Template-based approaches have been widely studied among APR techniques and have shown promising results. However, these methods are limited by the scope of predefined templates, making it challenging to handle out-of-template bugs. To this end, researchers have integrated Code Pre-trained Models (CodePTMs) to augment template-based APR. Existing methods typically rely on a single representation of the source code, which limits their ability to capture complex semantic and syntactic features. To address this issue, we propose FTPR, a novel approach that combines multi-source inputs (i.e., comments, code segments, and ASTs) and fix templates to improve CodePTM-based APR. Our evaluation on Defects4J-v1.2 shows that FTPR outperforms previous state-of-the-art approaches, successfully fixing 85 bugs. Furthermore, we validate the generality of FTPR on Defects4J-v2.0 and QuixBugs, the latter of which includes two programming languages. FTPR fixes 53 and 32 bugs on these datasets, respectively, demonstrating the significant potential and value of applying multi-source inputs and fix templates to automated program repair.