My research aims to address Natural Language Processing tasks with (extremely) weak supervision. I strive to answer the following questions:

What knowledge do pre-trained neural language models encode? Is the current fine-tuning or prompting paradigm the best to distill the right part of these knowledge to domain-specific tasks?
What is the best forms of getting supervision from human? How to inform machines about the NLP task goals with minimum human effort?
Can machines go beyond extracting information to discovering novel information to expand human knowledge?