Making grounding decisions

Given a speech recognition hypothesis, a dialogue system has the choice of accepting or rejecting this hypothesis, but can also choose to provide evidence of understanding, such as a clarification request, or display its understanding. In 3.3.2.7, this choice was referred to as the grounding decision problem. In the previous chapters, a static model with hand-crafted thresholds was used. In this chapter, we will use a data-driven decision-theoretic model for the grounding decision problem. Based on task analysis of the HIGGINS navigation domain, dialogue cost functions will be derived, which take dialogue efficiency, consequence of task failure and information gain into account. The dialogue data presented in the previous chapter will then be used to estimate parameters for these cost functions, so that the grounding decision may be based on both confidence and dialogue context.