Second-year PhD seminar

Time: Fri 2020-09-18 15.15

Lecturer: Erik Ekstedt

Location: Physically at the division and streamed online

Abstract:
In this talk I will discuss what I have been working on for my first two
years as a PhD-student. My area of research is in conversational systems
for spoken dialog with a focus on turn-taking. I want to improve a
conversational systems ability to determine when to speak and when to
listen. In my research I focus on how to learn appropriate turn-taking
behavior through deep learning.

The organization of turns in human communication is efficient and
general. Different cultures with different languages tend to organize
turns in similar ways. There are certain moments in a conversation where
a turn-shift may occur, namely Transition Relevance Places, TRPs.
Defining these places often include the notion of completion. These
completions may be prosodic, syntactic, vision based, etc.

The main contribution of my work aims to learn models of turn-taking, or
the distribution of TRPs, based on dialog data through self-supervised
learning. This means that I train models on dialog data with a "minimal"
need for turn-taking specific annotations. I discuss the models and
learning objectives which define my research with a specific focus on
TurnGPT (submitted to EMLNP2020), a verbal TRP model.

-----

Physical location: Fantum (Lindstedtsvägen 24, floor 5, room no. 522)
Map to physical location:
https://www.kth.se/places/room/id/c9ec01ab-b536-4be6-b82a-0d52ddadb2e6
Zoom URL: https://kth-se.zoom.us/j/61962541367