At the IEEE/ACM International Workshop on Automated Program Repair (APR 2023, co-located with ICSE 2023) in Melbourne, Australia, we presented “An Extensive Study on Model Architecture and Program Representation in the Domain of Learning-based Automated Program Repair” (Horváth, Csuvik, Gyimóthy and Vidács).
The study compares deep learning models (T5, CodeT5, RoBERTa and GPT-Neo) for repairing Java and JavaScript bugs, using three code representations: raw text, command sequences and abstract syntax trees. Training on command sequences outperformed most other configurations, reaching 30.64% accuracy on java-small, while combining AST and text information performed far worse, below one percent.
Link: ICSE 2023 speaker profile
Read the paper: publisher page
