Making grounding decisions: Data-driven estimation of dialogue costs and confidence thresholds

This paper presents a data-driven decisiontheoretic approach to making grounding decisions in spoken dialogue systems, i.e., to decide which recognition hypotheses to consider as correct and which grounding action to take. Based on task analysis of the dialogue domain, cost functions are derived, which take dialogue efficiency, consequence of task failure and information gain into account. Dialogue data is then used to estimate speech recognition confidence thresholds that are dependent on the dialogue context. 1 Making grounding decisions One of the greatest challenges when building spoken dialogue systems is to deal with uncertainty, which stems partly from the ambiguity of language itself and partly from the error-prone speech recognition process. Speakers handle uncertainty in dialogue by exchanging positive and negative evidence of understanding, a process commonly referred to as grounding (Clark, 1996). In order for a speaker to evaluate the addressee’s understanding of the concepts involved, two kinds of evidence are especially useful: display of understanding (DISPLAY) and clarification requests (CLARIFY). The addressee can also choose not to provide any direct evidence, but to simply ACCEPT or REJECT the hypothesis. These options are exemplified in the following alternative system responses in a human-computer dialogue, where the system is uncertain about the red building: (1) U: I can see a red building. S (ACCEPT): Ok, can you see a tree in front of you? S (DISPLAY): Ok, a red building, can you see a tree in