Break it - Message it - Fix it : Learning to Repair Python Programs using Error Messages without Labelled Data

Anonymous

Abstract: In recent years there is an increasing demand to reduce the gap in development to deployment. It has been estimated that developers spend almost 20\% of their time in fixed problems with their code. Therefore tools which can automatically repair code can help accelerate the DevOps cycles. In this work we build upon recent success of deploying neuro-symbolic approaches for automatic code repair. In our approach, we use a dataset of python code, viz, CodeNet, which represents data distribution for human generated code. We train two neural modules a breaker and a fixer, which are trained iteratively, along with a symbolic module Pylint. The breaker learns to introduce errors in the code, the symbolic module acts as a Critic and is able to fragment the error by identifying the line, as well provide the error type with a specific exception message. The Fixer utilizes the exception message to repair the erroneous line in the code. We are able to cover 32 different syntax errors, and iterative training based on back translation actually helps improve the performance of the Fixer.