Exploiting Diverse Forms of Advice to Guide the Discovery and Extraction of Time-Critical Information

We propose to address the primary challenges to the timely distillation of time-critical data into actionable information by exploiting the synergies amongst information discovery, extraction, and fusion processes. The proposed advisable assistant concept, Sentinel, will be comprised of three primary elements: (1) a user centered approach to context modeling and agent guidance; (2) a unified probabilistic model of information discovery/retrieval, extraction, and fusion; and (3) a predictive model of "interestingness", including representations of time criticality that will effect if, when, and how the user should be alerted. The resulting system will reduce existing barriers to the tasking of an agent through user interfaces that integrate into existing problem solving workflows and through the exploitation of active learning techniques that can make optimal use of any (potentially imperfect) guidance provided by the user. Further, the use of tightly intertwined probabilistic models in which discovery, extraction, and fusion decisions are made with a common pool of evidence and inference procedures will allow much richer forms of inference than possible with the current state of the art. Phase I research and development of a limited prototype will provide a solid foundation for the complete implementation of Sentinel in Phase II and its commercialization.