Proposed a novel method of contextual embedding mapping and revealed the tight relationship of isotropy, isometry and isomorphsim in contextual embedding spaces.
Developed a new algorithm to reduce the peak energy of data transmission based on Unimodular Quadratic Programming.
Proposed novel data augmentation approach for domain adaptation, which surpasses state-of-the-art performance in
Dialogue State Tracking and Event Extraction tasks.
Built a deep learning method based on GANs to naturally predict and expand the boundaries of incomplete images
Designed a human-computer interaction GUI system to recognize signatures of users with high precision, where users sign their names in the air (invisibly).
Investigated a zero-shot approach for dependency parsing by building a multilingual concept-shared semantic space, which
achieves state-of-the-art performance.