Causes and Strategies for Requesting Clarification in Dialogue

We do two things in this paper. First, we present a model of possible causes for requesting clarifications in dialogue, i.e., we classify types of non-understandings that lead to clarifications. For this we make more precise the models of communication of (Clark, 1996) and (Allwood, 1995), relating them to an independently motivated theory of discourse semantics, SDRT (Asher and Lascarides, 2003). As we show, the lack of such a model is a problem for extant analyses of clarification moves. Second, we combine this model with an extended notion of “confidence score” that combines speech recognition confidence with different kinds of semantic and pragmatic confidence, and argue that the resulting processing model can produce a more natural clarification and confirmation behaviour than that of current dialogue systems. We close with a description of an experimental implementation of the model.