BERT

A new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks