Informational characterization of dialogue states

We introduce multidimensional feature structures as a generalization of standard slot/filler representations commonly employed in spoken language dialogue systems. Nodes in multidimensional feature structures contain an n dimensional vector of values instead of one single filler element. The additional elements serve to represent, among other information, confidence measures of speech recognizers or the number of times a filler has been queried. We demonstrate the application of multidimensional feature structures to spoken dialogue systems. We show that unification based dialogue processing can be retained as long as the elements of the fillers are drawn from partially ordered sets. The dialogue manager employs a variant of constraint logic programming for representing dialogue strategies and update rules. The constraint logic program partitions the space of possible dialogue states in sets of states that are equivalent for the dialogue strategy.