Coding assistance with deep learning is an emerging concern that has recently attracted much attention in the software development community. To integrate coding assistance with deep learning compactly, we focus on neural machine translation (NMT), which allows users to translate natural language descriptions into expressions in a programming language such as Python. A rising problem here is the limited availability of parallel corpora, which is essential to train better NMT models.

To overcome the problem, we propose a transcompiler-based back-translation, a data augmentation method that generates parallel corpora from numerous source code repositories. In this paper, we present our initial experimental results by comparing several NMT models that are built upon the existing corpora and our corpora. The resulting BLEU indicates that our proposed model is accurate enough to allow coding assistance in the future.