Multilingual Code Snippets Training for Program Translation

Program translation aims to translate source code from one programming language another. It is particularly useful in applications such as multiple-platform adaptation and legacy migration. Traditional rule-based program methods usually rely on meticulous manual rule-crafting, which costly both terms of time effort. Recently, neural network based have been developed address this problem. However, the absence high-quality parallel data main bottlenecks impedes development models. In paper, we introduce CoST, a new multilingual Code Snippet Translation dataset that contains 7 commonly used languages. The at level snippets, provides much more fine-grained alignments between different languages than existing datasets. We also propose model leverages snippet denoising auto-encoding Multilingual (MuST) pre-training. Extensive experiments show training effective improving performance, especially for low-resource Moreover, our method shows good generalizability consistently improves performance number baseline proposed outperforms baselines snippet-level program-level translation, achieves state-of-the-art CodeXGLUE task. code, data, appendix paper can be found https://github.com/reddy-lab-code-research/MuST-CoST.