Deep Learning-Based Refactoring with Formally Verified Training Data

Refactoring source code has always been an active area of research. Since the uprising various deep learning methods, there have several attempts to perform transformation with use neural networks. More specifically, Encoder-Decoder architectures used transform similarly a Neural Machine Translation task. In this paper, we present learning-based method refactor code, which prototyped for Erlang. Our two major components: localizer and refactoring component. That is, first localize snippet be refactored using recurrent network, then generate alternative Sequence-to- Sequence architecture. could as extension already existing AST-based approaches since it is capable transforming syntactically incomplete code. We train our models on automatically generated data sets, based formally verified definitions by attribute grammar-based sampling.