At the IEEE/ACM International Workshop on Automated Program Repair (APR 2021, co-located with ICSE 2021), held online, we presented “Exploring Plausible Patches Using Source Code Embeddings in JavaScript” (Csuvik, Horváth, Lajkó and Vidács).
Test suites are imperfect judges of automatically generated patches, so we studied similarity-based patch filtering. We trained a Doc2Vec model on an open-source JavaScript project, generated 465 patches for 10 of its bugs and ranked them by similarity to the original program. Plain document embeddings can misclassify patches because they miss subtle code semantics, though in some cases they still gave useful signals.
Read the paper: publisher page
