February 14
Marios Skounakis
UW. Madison Department of Computer Science
Multiple-Resolution Hidden Markov Models

This talk presents the Multiple Resolution Hidden Markov Model (MR-HMM),
a machine learning framework for constructing information extractors
which can exploit grammatical and syntactic information about the
structure of sentences. Traditional uses of hidden Markov models in
information extraction have mostly ignored the structure of sentences
and focused only at the words therein.  A notable exception to this is
the Phrase HMM by Ray and Craven. The MR-HMM extends this earlier work
by providing a natural way for exploiting hierarchical representations
of sentences such as the parse trees constructed by grammatical parsers.

The talk focuses on the extraction of instances of relations between
multiple objects (multi-slot extraction), as opposed to the extraction
of isolated instances of objects (single-slot extraction).  I will
define the task of information extraction, overview the traditional use
of HMMs for extraction, motivate the use of hierarchical sentence
representations and develop the MR-HMM. Finally I will present some
experimental results and discuss some ideas for future extensions.

This is joint work with Prof. Mark Craven.