User Simulations for Spoken Dialogue Interfaces in Self-Help systems
Speaker: Srini Janarthanam
Institution: University of Edinburgh
Abstract:
Spoken dialogue interfaces are becoming a popular choice for customer contact solutions for businesses. These interfaces allow the customers to interact using speech on specific tasks like finding a restaurant in the city or getting information about flights, etc. User simulations have been used in a variety of ways for developing and deploying these interfaces. They can be used for training and testing these systems during development and also for predicting customer's moves in deployed systems. They have been used to train dialogue management policies using reinforcement learning. These policies decide system's utterance in response to the user's utterance. In the current systems, user simulation models interact with the dialogue manager using dialogue acts and only simulate an average group of users.
Self-help dialogue systems are those that customers could call to resolve issues like broken broadband connections themselves. In contrast to the air-ticket and town info systems mentioned above, self-help systems have to deal with a variety of users ranging from experts to complete novices. Therefore the system must dynamically adapt to the users during the conversation. It must give detailed instructions to novices and avoid jargon. With experts, it must ideally use jargon and avoid unnecessary simplified instructions. In order for the dialogue system to train to choose the appropriate instruction types and referring expressions, user simulations must be sensistive to such choices.
In this talk, I will present my proposed model which will be sensitive to the dialogue system's choice of instruction types and referring expressions and simulate diverse groups of users ranging from novices to experts.
