We propose to extract knowledge not only from cooperating people using HAWK, but also from existing data in relational databases. Given the enormous amount of data available in that form, we feel it is essential to exploit this data to whatever extent possible during knowledge base construction.
Our approach emphasizes the extraction and use of temporally-related information, and is based on use of the Transition Space representation [Borchardt, 1992][Borchardt, 1994]. This representation captures information about time, change, events and causality, and is motivated by research in human cognition and in language. As such, the representation can formalize intuitive, human-generated models of what happens during the temporal unfolding of events, and humans can specify this knowledge easily using stylized English that converts directly into assertions in the represenatation.
In summary, the approach for extracting temporally-related information from relational databases involves a four-step process of (1) reorganizing database relations to express underlying, cognitively-motivated, unary and binary subrelations, (2) translating the subrelations into sets of Transition Space assertions, (3) performing a comparative analysis of changes occuring between successive time points, also identifying event occurences as particular combinations of these changes, and (4) using our IMPACT system to perform a range of event-related reasoning operations on the resultant, extracted knowledge.
IMPACT is an interactive planning and monitoring tool developed at our laboratory. IMPACT's lower-level processing capabilities enable the system to be used as a vehicle for merging information from multiple relational databases or relaying information between relational databases with different underlying data models. Merging of information from databases can be accomplished by successive applications of the extraction process described above. Relaying of information between databases can be accomplished through a three-step process of extracting information from a first database, performing inference within IMPACT, and reversing the extraction process to re-express the encoded knowledge in the format of the second database. Central to IMPACT's merging and relaying capabilities is the common, intuitive and language-based representational grounding provided by the Transition Space representation.
IMPACT's higher-level processing capabilities enable human operators to construct, modify and monitor the execution of plans, as well as monitor ongoing activity related to those plans. IMPACT works like a ``spreadsheet for events'' that can: verify that reported events match other known occurrences, check action preconditions and explore possible instantiations for proposed actions, perform causal prediction and explanation, and perform inference and event recognition. Together the lower-level capabilities of the system, IMPACT's higher-level capapabilities provide a foundation for advanced application of extracted information in solving real-world problems.