We’ve bought cutting-edge effects on a collection of various language duties with a scalable, task-agnostic machine, which we’re additionally liberating. Our method is a mix of 2 present concepts: transformers and unsupervised pre-training. Those effects supply a powerful instance that pairing supervised finding out strategies with unsupervised pre-training works rather well; that is an concept that many have explored previously, and we are hoping our outcome motivates additional analysis into making use of this concept on higher and extra various datasets.
