When multiple knowledge bases are built separately, each either singly or multiply authored, they are bound to differ in details of vocabulary and other aspects of expression, even though they are meant to cover the same expert knowledge.
Faced with such diversity, there will be a need for systems that can help to recognize similarities and to ferret out gratuitous differences so as to combine the multiple knowledge bases into one, more robust and reliable knowledge base.
There has been, alas, little work in this area, perhaps because a total solution would necessarily be ``AI Complete.'' That is, a total solution would require a system with not only great expert knowledge at the human level, but also a great deal of commonsense intelligence.
Nevertheless, it appears that good work can be done with systems that rely principally on structural matching mechanisms. In particular, we are encouraged by the CARTER system of Davis et al., described in our Previous Accomplishments section, and we propose to base our work in this area on that precedent. This work characterizes some 35 different structural disparities that occur between near-miss encodings of knowledge and provides procedures for how a system like HAWK can then manage a collaborative process leading to consensus.
The context for HAWK's application will differ somewhat from this earlier work. In Davis' earlier system, two complete rule-based systems were constructed in isolation and then compared. In our situation, HAWK will continuously acquire bits and pieces of knowledge structure from many contributors. The reconciliation process in HAWK will be dynamic and incremental.
Another key difference is that HAWK will not assume that consensus is possible; indeed, it will encourage multiple differing perspectives on a problem. It is only when the knowledge structures are near misses that HAWK will attempt to create consensus. If there is no consensus achievable, then HAWK will maintain multiple contextualized perspectives.
A final difference from the earlier work has to do with what is compared and with the time at which the comparisons are made. Davis' CARTER system only made static, design time comparisons between the structure of the knowledge representations. We believe that HAWK will also be able to make dynamic comparisons which are made at problem solving time and which compare the stucture of the problem solving behavior evolved by those knowledge structures.