Friday, November 11, 2011
12:45 PM
CSB 703
James Allen
University of Rochester
Overview of new DARPA grant
One of the largest obstacles to improving natural language understanding systems is the lack of comprehensive deep semantic knowledge that allows words and phrases to be grounded in the world. The knowledge that an everyday human language user requires appears so vast that it is implausible to think it could be hand coded. So we need methods for automatically learning deep semantic knowledge. Prior attempts at language learning have either learned very shallow knowledge (e.g., associating words as an uninterpreted atoms with a scene) or domain-specific (e.g., learning to associate sentences with actions in a GUI). In addition, all these approaches require significant amounts of training data and, even then, offer little insight into how to produce generic linguistic knowledge for use by other NLP systems across different applications.
In contrast, human learners appear to rapidly learn new words, concepts, and phrasing, often from a single experience, and then readily apply this new knowledge to new situations and activities. The key goal in this project is developing a capability for such rapid learning of generic linguistic knowledge. We will show how the underlying semantic structure of language, coupled with a cognitive architecture initialized with a simple common-sense theory of the physical world, enables rapid acquisition of new words and concepts, as well as new linguistic constructions.