Context Management for Embodied Conversational Agents

Dr. Donna Byron

Embodied conversational agents present a new challenge for software models of dialog processing. To participate in a dialog, an agent must manage a complex set of knowledge sources that influence the relationship between linguistic forms and their meanings, including the current state of the discourse itself, tracking which beliefs are mutually known between the agent and its partner, observing the state of the world in which the dialog takes place, and updating the state of the task underway. In contrast to traditional unimodal dialog agents, an agent that is embodied and even mobile within a conversational setting must gather some of this information through perceptual sensors, fuse that information as it arrives in independent data streams with independent timing, and reason appropriately to synchronize this information with the linguistic exchange happening in the dialog.  Although this is a very complex task, we have made some progress on the relevant issues in our first year of development of the OCEANS conversational agent at OSU. The OCEANS project aims to develop an autonomous agent that can assist with search-and-rescue tasks within a simulated graphical world.  This talk will describe the current status of the reasoning component and sensor fusion needed to support context-sensitive language processing in this situated dialog agent.

Corpus of Quake Dialogues