Few-shot code translation via task-adapted prompt learning

Abstract

Pre-trained models such as CodeT5 and TransCoder have achieved impressive progress in software engineering. However, fine-tuning PLMs for code translation is confronted with significant challenges owing to the scarce availability of parallel code. Large language models such as ChatGPT have exhibited considerable promise in few-shot learning where only a small number of demonstration examples are given to the LLM. Yet they have not been specifically optimized for domain-specific tasks, and their use often entails significant manual effort in manually curating prompts. In this paper, we propose FSCTrans, a novel parameter-efficient tuning approach for code translation when furnished with only a few demonstration examples. (1) to efficiently reuse prior knowledge during pre-training, FSCTrans employs task-adapted prompt tuning, which freezes the pre-trained CodeT5 while merely updating parameters in a small prompt module; (2) to enable parameter efficient tuning on only a small number of examples, FSCTrans bridges pre-training to the translation task through a new pre-training objective of code-to-code generation. We evaluate FSCTrans on Java↔Python and Java↔C# datasets from both real-world projects and online judge problems. The evaluation results show that FSCTrans is remarkably effective in few-shot code translation: on average, it improves CodeT5 by 54.61% and 31.59% in terms of BLEU-4 and CodeBLEU; notably, FSCTrans demonstrates 14.42% and 18.36% superior performance in Java → C# translations in terms of BLEU-4 and CodeBLEU compared to ChatGPT.