Explaining and Generalizing Back-Translation through Wake-Sleep
This work offers a theoretical foundation for a widely used heuristic in machine translation, potentially benefiting researchers and practitioners by enabling more systematic improvements.
The authors tackled the problem of providing a principled interpretation for back-translation in neural machine translation, showing it as approximate inference in a generative model and proposing an iterative generalization that improved results by up to 1.6 BLEU.
Back-translation has become a commonly employed heuristic for semi-supervised neural machine translation. The technique is both straightforward to apply and has led to state-of-the-art results. In this work, we offer a principled interpretation of back-translation as approximate inference in a generative model of bitext and show how the standard implementation of back-translation corresponds to a single iteration of the wake-sleep algorithm in our proposed model. Moreover, this interpretation suggests a natural iterative generalization, which we demonstrate leads to further improvement of up to 1.6 BLEU.