Specifically, it was trained with two simultaneous self-supervised objectives, language modeling and link prediction, that encourage deep bidirectional reasoning over text and knowledge graphs. DRAGON ...