KEYNET is a graph-oriented method for information object indexing and retrieval. Information Objects must be annotated with small semantic networks that represent their key concepts. A larger semantic network (part of a subject-specific ontology) determines which node and link types (basic concepts and relations) are considered pertinent to a subject during information object retrieval.
The query graph actually used for retrieval may substitute more general or specific concepts for those specified by the user. Retrieval does not match large components of the query graph against whole content labels. Instead, the query graph is fragmented into small probes of bounded size. These fragments are matched against content labels, and resulting retrieval sets of potentially relevant information objects are combined using fragment-oriented weights.
The graph representations, their fragmentation, and post-retrieval merging of information object sets associated with distinct fragments naturally introduce a ``fuzziness'' appropriate to information retrieval notions of relevance, and facilitate use of distributed or parallel processing resources (at appropriate stages).
The KEYNET system employs a number of optimizations to ensure that it scales up to large corpora and so that it has high performance. Fragmentation combined with pattern associativity of graph structures and linear hashing techniques produces tractable complexity of communications and computation, despite necessary isomorphism testing and index manipulation. The system is therefore compatible with the requirements for search engines needed in proposals for an NII.