You are hereOverview / Research
Learning and Inferring Common Sense Knowledge
Automatic methods of acquiring general knowledge from corpora can miss many kinds of common sense knowledge, because the very knowledge we rely upon for communication is rarely communicated. Some of the most essential common sense knowledge is the knowledge people learn before they can write at all, much less write the text traditionally found in corpora.
We are exploring the problem of inferring new common sense knowledge from an existing knowledge base. Similarities and analogies in existing knowledge can give helpful clues toward filling in the gaps in a knowledge base, or finding connections that help us to unify knowledge from different domains.
Our goal is to create systems that understand the connections between everyday events and objects, people's beliefs, and the way they express them in language. The techniques of principal component analysis and Bayesian mixture models have helped us to discover those connections.
Creating Grounded Applications
Common sense reasoning enable an entirely new breed of applications, ones capable of understanding the user's situation and goals like a real person would. We have built a host of applications demonstrating early versions of our common sense reasoning technologies.
Understanding how People Talk about their Beliefs and Opinions
We are interested in detecting subcultures in people's beliefs. We also are taking recommender systems to another level by predicting not only whether a person will like a product but also what they will think about the product. We are building tools to identify multiple points of view around an issue and then let users explore them in a controlled manner.