Solving computationally hard problems, such as those commonly encountered in natural language processing often requires that approximate search methods be used to produce a structured output (eg. translation, transcript, summary, etc.). Unfortunately, this fact is rarely taken into account when machine learning methods are conceived and employed. This leads to complex algorithms with few theoretical guarantees about performance on unseen test data.
I present a machine learning approach that directly solves "structured prediction" problems by considering formal techniques that reduce structured prediction to simple binary classification, within the context of search. This reduction is error-limiting: it provides theoretical guarantees about the performance of the structured prediction model on unseen test data. It also lends itself to novel training methods for structured prediction models, yielding efficient learning algorithms that perform well in practice. I empirically evaluate this approach in the context of two tasks: entity detection and tracking and automatic document summarization.
Hal Daume III is an assistant professor in Computer Science at the University of Utah. His primary research interests are in machine learning (specifically, structured prediction and Bayesian methods), natural language processing (summarization) and computational linguistics (typology). He earned his Ph.D. at the University of Southern California with a thesis on structured prediction for natural language processing problems, under the supervision of Daniel Marcu. He spent the summer of 2003 working with Eric Brill in the machine learning group at Microsoft Research. Prior to graduate school, Hal studied math (mostly logic) at Carnegie Mellon University. It is well known that Hal does not like shoes, but does like activities that are hard on your feet: skiing, badminton, Aikido and rock climbing.
This talk is part of the CLIP Colloquium Series, organized by Jimmy Lin (jimmylin -at- umd .dot. edu). For the complete schedule, please visit http://www.umiacs.umd.edu/research/CLIP/colloq/.