Using Controlled Knowledge Search to Retrieve Cross-Contextual Information

Indexing cases on features has been shown to be a successful method of retrieving experiential knowledge for many single-task systems. However, discovering a general set of indices for retrieving cross-contextual information in a multi-use knowledge base is impractical. We present a method for retrieving crosscontextuM information based on knowledgedirected spreading activation in a semantic network. This method uses task-specific knowledge to guide a spreading activation search to a case or concept in memory which meets a desired similarity condition. We describe a specific instantiation of this method that retrieves analogies for the task of creative design, and also how the behavior of this method will differ from indexing schemes under varying knowledge base conditions. tem is represented in a global semantic network (a large labeled graph): nodes in the network represent primitive concepts, and links between nodes represent relations between those concepts. More complex concepts including, e.g., summary information of documents known to the system are represented as subgraphs within the larger semantic net. Retrieval is a process in which activation is spread outward from the nodes in some retrieval cue subgraph (representing either user request or some target concept to be matched) to neighboring nodes, and repeating this process until some other subgraph in the semantic network accumulates enough activation to be considered retrieved. Both Rau [Rau, 1987] and Cohen and Kjeldsen [Cohen and Kjeldsen, 1987] have applied spreading activation to general IR problems. This general class of approaches has the following advantages over traditional word-based IR approaches: (1) Information can be stored without explicit consideration of how it might be used in the future. I.e., information need not be stored with any specialpurpose indices (except the links to the rest of the knowledge base). And the way the information is represented makes it amenable to a number of other possible AI reasoning techniques to which simple text is not amenable. (2) The user can be given a much more descriptive language in which to express his query. In Rau’s system SCISOR, for example, the user makes his request for information using a restricted-vocabulary English. (3) Spreading activation call be triggered by any concept activation in the semantic net, so information retrieval can be a by-product of many system activities other than explicit user requests. Spreading activation has been applied to IR problems with success in small narrowly-defined domains. However, since spreading activation is essentially a blind breadth-first search mechanism, it has increasingly greater difficulty retrieving information the larger and more unconstrained the knowledge base becomes.