Language models trained to predict deliberately hidden words in a passage from the surrounding context.