Identification of Speaker Actions in Mixed Initiative Dialogue

Today’s state-of-the-art spoken language dialogue systems (SLDSs) are mainly system-directed leaving very little initiative to the user. This paper discusses how to enable controlled steps in the direction of mixedinitiative SLDSs. The discussion is based on experience from the design, implementation and test of system-directed dialogue for a spoken language dialogue system and on first Wizard of Oz experiments towards achieving implementable mixed-initiative dialogue. To this end, we present a categorisation of utterances in simulated human-machine dialogue based on an identification of the specific actions users perform with them. The categorisation builds on Speech Act Theory and holds the promise that limited mixed-initiative human-machine dialogue may be implemented in current SLDSs.