Consistency by Agreement in Zero-shot Neural Machine Translation
This addresses the challenge of reliable translation for low-resource language pairs in multilingual systems, though it is an incremental improvement over existing methods.
The paper tackles the problem of zero-shot generalization in multilingual neural machine translation, where models must translate between language pairs not seen during training, and introduces an agreement-based training method that improves zero-shot translation by 2-3 BLEU points on benchmarks without degrading supervised performance.
Generalization and reliability of multilingual translation often highly depend on the amount of available parallel data for each language pair of interest. In this paper, we focus on zero-shot generalization---a challenging setup that tests models on translation directions they have not been optimized for at training time. To solve the problem, we (i) reformulate multilingual translation as probabilistic inference, (ii) define the notion of zero-shot consistency and show why standard training often results in models unsuitable for zero-shot tasks, and (iii) introduce a consistent agreement-based training method that encourages the model to produce equivalent translations of parallel sentences in auxiliary languages. We test our multilingual NMT models on multiple public zero-shot translation benchmarks (IWSLT17, UN corpus, Europarl) and show that agreement-based learning often results in 2-3 BLEU zero-shot improvement over strong baselines without any loss in performance on supervised translation directions.