Pretraining with Extracted Gap-sentences for Abstractive SUmmarization Sequence-to-sequence modelsの略称。Gap Sentences Generation(GSG)という新たな学習タスク以外はBERTと同様。