A Corpus Collection and Annotation Framework for Learning Multimodal Clarification Strategies
Verena Rieser, Ivana Kruijff‐Korbayová, Oliver Lemon · 2005
Current dialogue systems are fairly poor in generating the wide range of clarification strategies as found in human-human dialogue.The overall aim of this work is to learn when and how to best employ different types of clarification strategies in multimodal dialogue systems.This paper describes a framework for learning multimodal clarification strategies for an in-car MP3 music player dialogue system.The framework consists of three major parts.First we collect data on multimodal clarification strategies in a wizard-of-oz study.Second we extract feature in the stateaction space to learn an initial policy from this data.Third we specify a reward function to refine that policy using extensions of existing evaluation schemes.