[Defense] Analysis and Modeling of Turn Dynamics in Multiparty Conversations
Monday, November 11, 2024
4:00 pm - 5:30 pm
In
Partial
Fulfillment
of
the
Requirements
for
the
Degree
of
Doctor
of
Philosophy
Meng-Chen
Lee
will
defend
his
proposal
Analysis
and
Modeling
of
Turn
Dynamics
in
Multiparty
Conversations
Abstract
The smooth and efficient flow of turns is fundamental to achieving natural and meaningful communication in conversations. It allows participants to engage fluidly, manage overlaps or interruptions, and maintain a shared understanding. This is particularly crucial in multiparty conversations, where the presence of more than two participants increases interaction complexity, making effective turn management more challenging. Accurate prediction of when a speaker will continue, stop, or yield the floor is essential, not only for human comprehension but also for the development of conversational agents and systems that enable seamless interactions. To explore these dynamics, a detailed analysis of conversation patterns is conducted, focusing on predicting turn-keeping or turn-taking using inter-pausal units (IPUs). A novel metric, Relative Engagement Level (REL), is introduced to quantify engagement, revealing strong relationships between social cues and conversational flow. Next, the study examines turn transitions at sentence boundaries, where transitions typically occur, aiming to identify the next speaker. This approach not only captures changes in turns but also provides insights into potential next speakers. Extensive experiments involving over 10 machine learning models are conducted to assess predictability. Finally, an advanced model for online end-of-turn prediction is developed, detecting turn endings on a frame-by-frame basis, regardless of IPU or sentence boundaries. The PLM-GRU model effectively predicts regular turn-taking, interruptions, and overlaps, providing a more nuanced understanding of conversation dynamics. These results contribute to the development of conversational systems for robots or virtual agents that can respond more naturally and effectively to human interactions.
Monday,
November
11,
2024
4:00
PM
-
5:30
PM
PGH 550
Dr. Zhigang Deng, proposal advisor
Faculty, students, and the general public are invited.
