# Back-Translation **Entity class:** Machine-translation training method ## Definition and history Back-translation converts target-language material into synthetic source-language examples, then uses the resulting pairs to train a translation system. [[articles/Intelligence Inside|Intelligence Inside]] names it as a case where a model's output becomes training material for a subsequent model. The method expands usable bilingual training data from monolingual text. Its success depends on the quality and distribution of the synthetic side of those pairs. ## Mechanism and limits This is an outer training cycle, distinct from a model updating its weights with every user interaction at inference time. Synthetic output can introduce errors or stylistic biases as well as useful coverage. Provenance of generated examples matters when evaluating what a later model learned and from where. ## Relationships **Training route:** [[wiki/Machine Learning|Machine Learning]]. **Historical setting:** neural machine translation. **Comparison:** [[wiki/Least Mean Squares Algorithm|Least Mean Squares Algorithm]] uses a different online update mechanism. **Article route:** [[articles/Intelligence Inside|Intelligence Inside]]. ## Sources [Google Research, *Recent Advances in Google Translate*](https://research.google/blog/recent-advances-in-google-translate/); [[articles/Intelligence Inside|Intelligence Inside]].