NLG, Multi-task learning, Parameter efficiency, Retrieval-enhanced Transformer