Department of Computer Science
University of Illinois, Urbana-Champaign
With tremendous amounts of texts across the Internet nowadays, various Natural Language Processing (NLP) systems are built to help people seek for valuable knowledge from massive corpora, by performing knowledge-intensive tasks like text retrieval, concept organization, commonsense reasoning, and question answering. Despite the remarkable success, most existing NLP systems still rely on large amounts of task-specific training data, which are costly to obtain.
My research designs principled approaches for label-efficient, knowledge-based NLP applications which rely on minimal human supervision. In this talk, I will introduce a general framework for textual knowledge extraction and utilization: (1) concept ontology construction by transforming generic linguistic knowledge encoded in pre-trained language models into hierarchical structures connecting entities; (2) entity extraction by replacing manual prompt template designs with automatic soft verbalizer learning; (3) commonsense reasoning via entity knowledge prompting and iteratively optimizing reasoning paths generated by language models.