Simple and Effective Input Reformulations for Translation

Foundation language models learn from their finetuning input context in different ways. In this paper, we reformulate inputs during finetuning for challenging translation tasks, leveraging model strengths from pretraining in novel ways to improve downstream performance. These reformulations are simple data level modifications, require no additional collection of training data or modification of data at inference time. They can be applied either on single language pair translation tasks or massively multilingual translation tasks. Experiments with these techniques demonstrate significant performance improvements up to 3.5 chrF++ on the Flores200 translation benchmark. We hope our research accessibly improves finetuning data efficiency, enabling more effective training to scalably improve state-of-the-art performance. Our code is released here. Baseline German: Das ist gut. English: Partial Output English Scaffold (POSE) German: Das ist gut. English: That is Parallel Scaffold in English (ParSE) German: Das ist gut. English: That is good. Spanish: Mixed-language Parallel Scaffold (MiPS) German to Spanish: Das ist gut. Chinese to English: 那很好。 Está bien. That is good. Está bien. mT5 That is good. That is good.