Login
Computer Science @ Rochester
Friday, October 29, 2010
12:45 PM
CSB 209
Catherine Havasi
MIT
Digital Intuition: Understanding the World Using Common Sense
When we communicate, we rely on a large body of shared assumptions about the world that we use to understand each other. We use this intuition everywhere: when we go into a cafe, we know we can buy coffee there, we know our computer mouse won't gnaw on our fingers, and we know not to tell people that when you drop something it will fall down. The Open Mind Common Sense (OMCS) project began collecting common sense from volunteers on the internet in 1999 and has since collected over a million pieces of common sense. The collected information is converted, using automatic NLP techniques, to a semantic network called ConceptNet, which is made of relations between concepts.

Using singular value decomposition on the graph structure of ConceptNet yields AnalogySpace, a vector space representation of common sense knowledge. In this technique, the analogical closure of a semantic network is formed through dimensionality reduction. It self-organizes concepts around dimensions that can be seen as making distinctions such as "good vs. bad" or "animate vs. inanimate", and generalizes its knowledge by judging where concepts lie along these dimensions. This representation reveals large-scale patterns in the data, while smoothing over noise, and predicts new knowledge that the database should contain. Additional techniques extend this model to streaming data sets, association-based similarity, metacognition, and information that varies over time.