Hoifung Poon

University of Washington

A long-standing goal of AI and natural language processing is to harness human knowledge by automatically understanding texts. Known as machine reading, it has become increasingly urgent with the rise of billions of web documents. To represent the acquired knowledge that is complex and heterogeneous, we need first-order logic. To handle the inherent uncertainty and ambiguity in extracting and reasoning with knowledge, we need probability. Combining the two has led to rapid progress in the emerging field of statistical relational learning. In this talk, I will show that statistical relational learning offers promising solutions for machine reading. I will present Markov logic, which is a leading unifying framework for statistical relational learning, and has spawned a number of successful applications for machine reading. In particular, I will present USP, an end-to-end machine reading system that can read text, extract knowledge and answer questions, all without any labeled examples. To resolve linguistic variations for the same meaning, USP recursively clusters expressions that are composed with or by similar expressions. In a machine reading experiment, USP extracted five times as many correct answers compared to state-of-the-art systems such as TextRunner, and raised accuracy from below 60% to 91%.

4:15pm

B17 Upson Hall

Tuesday, February 22, 2011

Refreshments at 3:45pm in the Upson 4th Floor Atrium

 

Computer Science

Colloquium

Spring 2011

www.cs.cornell.edu/events/colloquium

Markov Logic in Machine Reading