R. Osgood and R. Bareiss
As aids to human problem solving, CBR systems typically rely on feature matching to retrieve cases that are likely to be useful. Human experts in contrast provide much richer problem solving assistance. Experts not only can recall relevant experiences, they can provide a helpful interpretive framework of information. Because these kinds of information are highly interrelated, there are memory advantages to indexing them relative to one another---chiefly, making them more easily retrievable by human users. Therefore, we are studying the relative indexing of stories (elaborate cases) in the context of a type of hypermedia system called an ASK System. ASK Systems are designed to simulate conversations with experts. These systems provide access to manually indexed, multimedia databases of story units. Indexers (knowledge engineers) link these units together to form conversationally coherent threads. This paper discusses the theory of relative indexing employed in ASK Systems and the practical index-construction process that we have devised. However, as these system grow in size finding appropriate relative indices manually becomes increasingly difficult. We call this the indexer saturation problem. Our solution is to provide automated assistance to indexers. We describe an approach that uses a theory of conversation to propose relative links between units, eliminating the need for exhaustive manual unit-to-unit comparison. Initial results suggest that the approach provides a practical solution to the saturation problem balancing the strengths of humans ( e.g., feature extraction and categorization) and machines ( e.g., rapid search and record keeping).