Video Summary

The paper proposes a new method for training neural machine translation models. The method is based on the idea of self-supervised learning, which means that the model is trained on data that does not have human-generated labels. In this case, the model is trained on a corpus of parallel text, where each sentence in one language is paired with a translation in another language. The model learns to predict the translation of a given sentence in the source language, without being explicitly told what the correct translation is. This method has been shown to improve the performance of neural machine translation models on a variety of tasks.