2 Dec 2016 - Nate Chambers: Seminar
Learning Commonsense Event Schemas from Unlabeled Text
The early years of research in natural language understanding focused on big picture representations like scripts and frames to drive language understanding. These formalisms motivated a rich line of research, but they suffered from brittle hand-coded structure. Most recently, several lines of research have focused on new computational methods to learn script-like knowledge. Event schema induction, for instance, is the task of learning high-level representations of complex events (e.g., a bombing) and their entity roles (e.g., perpetrator and victim) from unlabeled text. As is usual, many methods have been proposed, and different event representations and learning approaches have been applied. This talk will motivate why such schema learning is important to NLU, and present a particular generative model that learns schemas from unlabeled text without human supervision. It achieved state of the art performance on an information extraction domain. Further, the talk will briefly summarize the latest work in the area, and describe a new dataset to motivate future research in event schema learning.
Nate Chambers is an Associate Professor at the United States Naval Academy (USNA), and co-founder of their Center for High Performance Computing. He is visiting the Informatics Forum while on sabbatical this year. He enjoys teaching undergraduate computer science and researching models of temporal reasoning, unsupervised information extraction, and commonsense knowledge extraction.