Machine Learning and Friends Lunch





home
past talks
resources
conferences

Learning in Search: Practical Learning Theory for Hard Language Problems


Hal Daume III
ISI

Abstract

Solving computationally hard problems, such as those commonly encountered in natural language processing and computational biology, often requires that approximate search methods be used to produce a structured output (eg., machine translation, speech recognition, protein folding). Unfortunately, this fact is rarely taken into account when machine learning methods are employed. This leads to complex algorithms few theoretical guarantees about performance on unseen test data.

I will discuss recent work that directly solves such "structured prediction" problems, by considering formal techniques to reduce the structured prediction problem to simple binary classification, within the context of search. This reduction is error-limiting: it provides guarantees about the performance of the structured prediction model on unseen test data. It also suggests novel training methods for structured prediction models, yielding efficient learning algorithms that perform well in practice. I will present recently published results on the entity detection and tracking problem as well as preliminary results on a summarization task, and conclude with directions for future work.

Much of this is joint work with my adviser Daniel Marcu (USC/ISI), and/or John Langford (TTI-C). (I'll also spend a few minutes mentioning other work that goes on at ISI these days.)

About the speaker:

Hal Daume III is a Ph.D. candidate ("all but dissertation") in Computer Science at the University of Southern California, expecting to graduate in May 2006. He received his Bachelor's degree from the Mathematical Sciences department at Carnegie Mellon University in 2001. His research interests are in developing and applying advanced machine learning techniques to problems that exhibit complex structure, such as those found in natural language processing.

Back to ML Lunch home