Title: Learning to Read and Reason by Answering Questions
Advisors: Ali Farhadi and Hannaneh Hajishrizi
Supervisory Committee: Ali Farhadi (Co-Chair), Hannaneh Hajishrizi (Co-Chair, EE), Gina-Anne Levow (GSR, Linguistics), and Oren Etzioni
Abstract: Reasoning can be defined as an ability to produce new knowledge from existing knowledge. While reasoning in a formally defined (logical) environment is well-studied, constructing the ontology and annotating training data for grounding surface forms (natural language) require tremendous amount of effort from human experts, still without guarantee of success. In this talk, I will discuss two recent projects on learning to (1) read and comprehend (ground) natural language and (2) reason over multiple facts in a synthetic setup, both with end-to-end training and minimal supervision of question answers. Then I will demonstrate that while language understanding and reasoning could be tackled one by one, there is a clear challenge in learning both at the same time. I will finally end my talk with my thesis proposal for immediate next steps to address the challenge and contribute towards building a general-purpose reasoning model.