Pre-training a unified natural language model with corrupted span and replaced token detection | Patent Publication Number 20240062018 | Patexia