Self supervised learning works much better for natural language than it does for image recognition and video because it's easier to account for uncertainty in the prediction process.