[Colloquium] Boven/MS Presentation/March 8, 2007

[Margaret Jaffey]

This is an announcement of Bradley Boven's MS Presentation.

-----------------------
Date:  Thursday, March 8, 2007

Time:  3:00 p.m.

Place:  Ryerson 251

M.S. Candidate:  Bradley Boven

M.S. Paper Title:  Dialog Act Tagging in Multi-Party Dialogs

Abstract:
Dialog acts are a key component of language that provide a basis for  
spoken language understanding between humans.  A dialog act (DA)  
represents the meaning of an utterance at the level of an  
illocutionary force [Stolcke et al.,2000].  They convey not the  
semantic meaning of an utterance, but rather the function an  
utterance is intended to serve.  Examples of these labels include  
questions, statements, backchannels, floor controls, and  
disruptions.  This paper presents an analysis of word and feature  
space models which can be used to construct a dialog act tagger using  
both supervised and unsupervised learning techniques.  The study was  
performed using data from the International Computer Science  
Institute (ICSI) Meeting Recorder Dialog Act (MRDA) Corpus.

A natural representation for this tagging problem is a vector space  
model of features. Each vector represents a specific dialog act whose  
values are the co-occurrence frequencies of each word, n-gram, or  
other feature in the model. The resulting ambient space is very high  
dimensional, although most of the matrix is sparse.  However, we  
believe the true latent structure of the space to actually be of a  
much lower dimension. Dimensionality reduction techniques attempt to  
reveal the dominant underlying structure of these models, while at  
the same time also conveniently compact the feature space. This  
results in a more computationally manageable space for the  
classification algorithms.  We augment the feature representation  
with n-gram language models, acoustic features, speaker information,  
and also dialog act tag information. This paper contrasts different  
reduction techniques, including latent semantic analysis and locality  
preserving projections.  The classifiers utilized employ both  
supervised and unsupervised approaches.  Specifically, we use  
unsupervised clustering techniques, such as k-means and spectral  
methods and also a supervised approach using a k-nearest neighbor  
classifier.

This paper attempts to understand what features play an important  
role in the tagging of dialog acts, what reductions are most  
effective at mapping the feature space to improve DA classification  
accuracy, and what classification algorithms accurately separate the  
data given these inputs.  Our best results were achieved using high  
order n-gram features in combination with algorithms using local  
neighborhood spaces.

Advisors:  Professors Gina-Anne Levow and John Goldsmith

A draft copy of Brad Boven's MS Paper will be available soon in Ry 161A.


=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey                             margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 161A)        (773) 702-6011
The University of Chicago                  http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=