Data Augmentation for Code Translation with Comparable Corpora and Multiple References

TL;DR

This paper introduces two data augmentation methods for code translation that leverage comparable corpora and multiple references, significantly improving translation accuracy across Java, Python, and C++.

Contribution

It proposes novel techniques for creating comparable corpora and multiple references, enhancing training data for code translation models.

Findings

Improved CodeT5 translation accuracy by 7.5% CA@1

Generated diverse target translations filtered by unit tests

Enhanced training data with comparable corpora from natural language documentation

Abstract

One major challenge of translating code between programming languages is that parallel training data is often limited. To overcome this challenge, we present two data augmentation techniques, one that builds comparable corpora (i.e., code pairs with similar functionality), and another that augments existing parallel data with multiple reference translations. Specifically, we build and analyze multiple types of comparable corpora, including programs generated from natural language documentation using a code generation model. Furthermore, to reduce overfitting to a single reference translation, we automatically generate additional translation references for available parallel data and filter the translations by unit tests, which increases variation in target translations. Experiments show that our data augmentation techniques significantly improve CodeT5 for translation between Java, Python, and C++ by an average of 7.5% Computational Accuracy (CA@1), which verifies the correctness of translations by execution. The code is available at https://github.com/Veronicium/CMTrans.