Selective retrieval improves machine translation by asking only when needed
Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation
Computation and LanguageArtificial IntelligenceInformation Retrieval
Summary
Machine translation systems often use extra information from outside sources to help translate tricky phrases, but they usually retrieve this information all the time, even when it’s not needed. The authors argue that sometimes the translation model already knows enough and only needs to ask for extra help in certain cases. They developed a system that learns when to use outside information and when to translate on its own by trying to produce the best translation quality. This approach outperforms methods that always ask for help or never do, improving results especially for hard phrases or idioms.
What this means in practice
- •For natural language processing engineers: Build machine translation tools that dynamically decide when to retrieve extra information, improving translation accuracy without unnecessary queries.
- •For software localization teams: Improve translation of idioms and terminology by integrating selective retrieval methods, adapting retrieval efforts to specific translation needs.
Authors
Shijun Li, Yang Liu, Najmeh Sadoughi, Dongmei Wang, Xinjie Li, Zhu Liu, Rohith Mysore Vijaya Kumar, Yuzong Liu
Abstract
Retrieving external knowledge can improve large language model (LLM) translation by providing supplementary information such as idiom interpretations and terminology. However, existing retrieval-augmented machine translation systems typically apply predefined retrieval strategies uniformly across all inputs. We argue that translation operates in a partial-evidence regime: LLM's parametric knowledge suffices for many inputs, while external evidence benefits a subset and usually supports rather than directly provides the desired translation. This creates two challenges: deciding when to retrieve and what evidence to acquire. We introduce Selective Retrieval-Augmented Translation (SRAT), an agentic reinforcement learning framework that learns when to translate directly or retrieve, what to retrieve, and how to integrate these decisions by optimizing final translation quality. To align evidence retrieval with downstream translation quality rather than simple semantic similarity, we further develop joint Plackett--Luce optimization and translation-quality contrastive training for retriever-training. Experiments on idiom, knowledge-intensive, and document-level translation with offline or online retrieval show that SRAT consistently outperforms never- and always-retrieve baselines, and quality-aligned retriever training provides further gains when similarity-based retrieval is insufficient. The results demonstrate the benefit of treating retrieval as a learnable translation action rather than fixed preprocessing.