Prompting Neural Machine Translation with Translation Memories

Improving machine translation (MT) systems with memories (TMs) is of great interest to practitioners in the MT community. However, previous approaches require either a significant update model architecture and/or additional training efforts make models well-behaved when TMs are taken as input. In this paper, we present simple but effective method introduce into neural (NMT) systems. Specifically, treat prompts NMT at test time, leave process unchanged. The result slight an existing system, which can be implemented few hours by anyone who familiar NMT. Experimental results on several datasets demonstrate that our system significantly outperforms strong baselines.