Trainable Approaches to Generation in Dialogue Systems 

Dr. Amanda Stent

Abstract: 

I will start by giving a brief overview of the field of natural language processing and 
dialogue systems. I will then talk about several pieces of research that exemplify the 
current tension between knowledge-driven and data-driven approaches to natural language 
processing. As dialog systems have become more complex, there has been increasing interest 
in the use of natural language generation. However, rule based generators often have to be 
tuned to the application to improve efficiency and output quality. I will describe several 
examples of how to adapt natural language generators, including: trainable content planning, 
trainable sentence planning, and trainable multimodal generation. This research was joint 
with researchers from AT&T and the University of Pennsylvania. 

Biography

Dr. Amanda Stent has been at Stony Brook for almost three years. She previously worked briefly 
at AT&T Research. She got her Ph.D. in 2001 from the University of Rochester. She currently 
supervises or \line co-supervises five PhD students, three MS students and three undergraduates. 
Her research focuses on natural language and multimodal generation and dialog systems.