高效推理和训练引擎LightSeq
NAACL 2019: Submodular optimization-based diverse paraphrasing and its effectiveness in data augmentation