In a mixed-initiative dialogue between multiple interlocutors, the ability to construct, to maintain, and to exploit an explicit model of the dialogue partners' beliefs, goals, and plans is indispensable. An agent model is required for identifying the objects which the dialogue partner is talking about, for planning the appropriate dialogue contributions towards achieving the own goals, and for determining the effects of planned dialogue contributions on the dialogue partner.

If we assume that we have no access to existing models of the dialogue partners, then we have to solve the problem of constructing the initial agent model at the beginning of the dialogue. The approach I propose here is in line with the modal logic approach to agent and stereotype modeling of Allgayer, Ohlbach, and Reddig (1992). The basic idea is to enhance a decidable fragment of first-order logic with modal operators modeling the notions of belief, knowledge, and desire. To provide reasoning capabilities we follow the translation approach of Nonnengart (1992). This amounts to manipulating modal logic formulas by a certain set of transformation rules so that classical, i.e. first-order, proof methods can be applied.