Video Summary

The paper proposes a new method for training neural machine translation models. The method uses a self-supervised objective that encourages the model to learn a latent representation of the input text that is useful for translation. The authors show that their method improves translation performance on several benchmark datasets.