Few-shot code translation via task-adapted prompt learning

Xuan Li, Shuai Yuan, Xiaodong Gu, Yuting Chen, Beijun Shen

Abstract: Highlights•Tackles the low-resource problem of code translation in the supervised setting.•Employs task-adapted prompt tuning in CodeT5, updating only prompt module parameters.•Connects pre-training to translation via a new objective in code-to-code generation.•Demonstrates superior performance in Java to C# translation compared to ChatGPT.