ICML-98 Submission #124

An Investigation of Transformation-Based Learning in Discourse

Authors with addresses: Ken Samuel, Sandra Carberry, and K. Vijay-Shanker
                        Department of Computer and Information Sciences
                        University of Delaware
                        Newark, Delaware 19716 USA
                        {samuel,carberry,vijay}@cis.udel.edu
                        http://www.eecis.udel.edu/~{samuel,carberry,vijay}

Abstract

This paper presents results from the first attempt to apply the
Transformation-Based Learning (TBL) method to any discourse-level
problem. To address a limitation of the standard algorithm, we
developed a Monte Carlo version of TBL that makes the method tractable
for a wider range of problems without degradation in accuracy. To deal
with the particular demands of discourse processing, we introduce an
entropy-minimization approach that automatically collects cue phrases.
We also devised a committee method for assigning confidence measures
to tags produced by TBL. The paper describes these advances, presents
experimental evidence that TBL is as effective as alternative
approaches, such as Decision Trees, N-Grams, and Hidden Markov Models,
and argues that TBL has desirable features that make it particularly
appealing for discourse-level problems, such as Dialogue Act Tagging.

Keywords: Transformation-Based Learning, discourse, NLP

Email address of contact author: samuel@cis.udel.edu

Phone number of contact author: (302) 738-8950

Multiple submission statement (if applicable): None