At the 2nd International Workshop on Intelligent Bug Fixing (IBF 2020, co-located with SANER 2020) in London, Ontario, Canada, we presented “Utilizing Source Code Embeddings to Identify Correct Patches” (Csuvik, Horváth, Horváth and Vidács).
A patch that passes all tests can still be wrong, so we looked for ways to predict which patches are correct from the textual and structural similarity between the buggy program and each patch, using Doc2Vec and BERT embeddings on source code. Patches that introduce new variables or change more code tended to be the incorrect ones. On the QuixBugs dataset of 40 bugs, the approach filtered out 45% of the incorrect patches.
Read the paper: publisher page
