Syntax Guided Neural Program Repair

Automated Program Repair(APR) helps to improve the efficiency of the software development and maintenance. Although a lot of APR approaches have been proposed recently, they have their own limitations. Pattern-based approaches need the hand-written patterns which need a lot of hard work of programmers. DL-based approaches can generate the complex statements when adopting NMT models. In this paper, we propose a framework to integrate these two kinds of APR approaches. Then, we introduce a novel neural model named Recoder designed for this framework to generate the patches based on the given buggy method. We conducted several experiments to evaluate Recoder on several datasets, Defects4j v1.2, and Defects4j v2.0. Our results show that Recoder achieved 23.1% improvements over the previous stateof-the-art approaches and repaired 52 bugs on Defects4j v1.2. Furthermore, Recoder achieved 200% performance on Defects4j v2.0 compared with the top two pattern-based APR tools.