"Probabilistic Generative Models for Natural Language Processing and Reasoning"

3/14/2003: 10:30am - 12:00pm
11th Floor Large Conference Room

Abstract: A variety of natural language problems can be given elegant solutions when couched within a probabilistic generative framework. When the statistical models that are used to solve these problems can be trained on massive amounts of data and when such data can be obtained automatically, the performance of the resulting systems is difficult to match using alternative techniques. In this talk, I will discuss how one can conceive of, train, and exploit probabilistic generative models in two applications. The first application pertains to the ability to find in massive amounts of text short answers to factoid questions, such as "How did Mahatma Gandhi die?", "Who discovered America?", or "How deep is Crater Lake?" The second application pertains to the ability to reason deductively and abductively using natural language statements. That is, the ability to compute and render in natural language the possible causes and effects of statements such as "bond prices fell", or "the department has raised its forecast for the current harvest".

About Daniel Marcu: Daniel Marcu was the fire warden for the ISD division from 1999 to 2001, when he was demoted to the function of stair-case operator by Lauri Grier, for reasons that remain unclear, even today. Later he was stripped of this role too, becoming a simple go-down-the-stair participant in several well-planned-out fire drills. Long gone are the days when Daniel was seen patrolling the ISD corridors with his fire fighter hat on. But he continues to keep his hat handy, on the top of his bookshelf, dreaming that, one day, he may be given a second chance. During the last fire drill, Daniel has noticed that the hat no longer fits him. To increase his preparedness, he took advantage of the break, went to Super Cuts, and got his hair trimmed. The hat fits him again.

Last updated: Mon Jun 19 17:44:06 2006