Embedding Context as Code Dependencies for Neural Program Repair
Noor Nashid, Mifta Sintaha, Ali Mesbah · 2023
Deep learning-based program repair has received significant attention from the research community lately. Most existing techniques treat source code as a sequence of tokens or abstract syntax trees. Consequently, they cannot incorporate semantic contextual information pertaining to a buggy line of code and its fix. In this work, we propose a program repair technique called GLANCE that combines static program analysis with graph-to-sequence learning for capturing contextual information. To represent contextual information, we introduce a graph representation that can encode information about the buggy code and its repair ingredients by embedding control and data flow information. We employ a fine-grained graphical code representation, which explicitly describes code change context and embeds semantic relationships between code elements. GLANCE leverages a graph neural network and a sequence-based decoder to learn from this semantic code representation. We evaluated our work against six state-of-the-art techniques, and our results show that GLANCE fixes 52% more bugs than the best performing technique.