[Colloquium] JOB TALK TODAY: Sam Brody, Rutgers University

[Katie Casey]

DEPARTMENT OF COMPUTER SCIENCE

UNIVERSITY OF CHICAGO

Date: Monday, January 24, 2011
Time: 2:30 p.m.
Place: Ryerson 251, 1100 E. 58th Street

----------------------------------------------

Speaker:		Sam Brody

From:		Rutgers University

Web page:	http://people.dbmi.columbia.edu/~sab7012/index.html

Title: 		Statistical Semantics: from Text to Understanding

Abstract:  	Teaching machines to understand and reason about natural language is a holy grail for AI. The approaches used to reach this goal vary from realm of formal logic and rule-based systems, through manually defined semantics, to machine learning classification. However, these techniques are difficult to apply on a global scale. Supervised systems are highly accurate in the specific setting for which they were designed, but suffer dramatically when transitioning to a different domain or task, and the cost of retraining is often prohibitive. Unsupervised NLP can provide systems for language understanding and reasoning that are truly agnostic with regard to the target domain and even language, and can adapt themselves to the tasks to which they are applied.

In this talk, I will briefly describe the challenges unsupervised NLP faces on the way from raw text to understanding and reasoning. I will focus on recent work addressing a subset of these challenges in the domain of online reviews, where supervised method are often impractical. I will present a flexible and robust system which employs statistical semantics to analyze the textual content of the reviews. The system infers the relevant aspects of the product or service automatically, and detects the reviewer's sentiment regarding each aspect. The aspect inference method achieves state-of-the-art performance, despite using a simpler and more general model than previous partially-supervised approaches. Sentiment detection correlates strongly with human judgment, and automatically tunes itself to the relevant context. Both components are independent of annotation, and represent the power of statistical semantics to learn directly from raw text.

Host: 		John Goldsmith

Refreshments will be served following the talk at 3:30 in Ryerson 255.