En datadriven beslutsteoretisk metod för val av återkopplingsstrategier i dialogsystem
One of the greatest challenges when building dialogue systems is to deal with uncertainty. Uncertainty comes not only from the ambiguity of language, but in the case of spoken dialogue, from imperfect speech recognisers, from which the system designer must expect a certain amount of errors. When humans communicate, they deal with uncertainty by giving negative and positive evidence of understanding to each other - a process commonly referred to as \"grounding\". Examples of such evidence, or \"grounding actions\", are display of understanding and clarification requests. The problem is how to choose which hypotheses to accept and which grounding actions to take. A common solution is to base this decision on static hand-crafted speech recognition confidence thresholds.
In my talk, I will first discuss how the selection of grounding actions can be modelled in the well known framework of \"decision making under uncertainty\", where the utility and costs of different actions are weighted against the probability of different outcomes. I will then show how a unified cost measure (efficiency) can be estimated from real dialogue data from the Higgins spoken dialogue systems developed at KTH, to derive dynamic confidence thresholds for choosing grounding actions. I will also show how the information gain of different concepts and the consequence of task failure will affect these thresholds.