Despite their reputation in IR circles as cumbersome, inefficient and suitable only for small databases, at least one IR researcher has used knowledge-based indices successfully [FHL+91]. Fuhr et al's AIR/X performs automatic indexing of documents using terms (descriptors) from a restricted vocabulary. Probabilistic classification determines indexing weights for each descriptor using rule-based inference.
After building a system similar to AIR/X, Jacobs [Jac93] determined that ``the combination of statistical analysis and natural language based categorization is considerably better than either alone.'' His paper describes an automated set of statistical methods for pattern acquisition that operate inside a knowledge-based approach for news categorization (an area closely related to document classification and other information retrieval tasks).
Both of these systems attempt to automate the process of generating index terms. KEYNET, by contrast, is concerned not with how the index terms are obtained but instead with their structural relationships.
Several distinct families of databases for semantic networks have been developed. Such databases are often called knowledge-base systems. Some of the best known of these are: Conceptual Dependency, ECO, KL-ONE, NETL, Preference Semantics, PSN and SNePs (see [Leh92]). All of these support link types, frame systems and so on, but few if any explicitly concern themselves with performance measures familiar in traditional database work, such as minimizing the number of disk accesses required to retrieve complex structures (i.e., graphs assembled from multiple frames and their typed relations). Contemporary knowledge-base systems traverse semantic networks one frame/relation at a time. Unless the knowledge-base fits entirely in the main memory of a single processor, these traversals result in large amounts of virtual memory paging.
In [Lev91], Levinson describes a technique for pattern-oriented (graph) retrieval. Levinson's approach discovers common structures among graphs in a database, and then uses pattern associativity to reduce the required number of tests for subgraph isomorphism. A separate paper, [Lev92], compares techniques for subgraph isomorphism testing suitable for pattern-oriented retrieval. The baseline retrieval method described in [Lev92] considers a flat set of N networks with no pattern associativity information, in which each query requires N isomorphism tests. The first improvement indexes commonly occurring substructures in the graphs, and tests only a subset of the entire database. A second improvement creates a multilevel index of subgraph relationships between successive levels of substructures and the domain of graphs in the database (establishes a partial order). A final elaboration applies multilevel indexing to connectivity and label information required during subgraph isomorphism testing (via the refinement method), rather than to the graph substructures themselves. This produces a tree of ``node descriptors'' in order of increasing specificity, with actual index pointers at its leaves.
In KEYNET we approximate subgraph isomorphism testing to determine the relevance of information objects that were previously annotated with content labels. Our approach compiles fragments of content labels into a hash table, reducing lookup computation (while increasing database construction overhead). Although some ambiguity remains possible when matching fragments instead of whole graphs, vertex overlap between components that contain vertex-incident edges ensures similarity of graph structure between a query and a content label that is retrieved using its fragments. This approximation further allows ``simultaneous'' (multiprocess, distributed or parallel) matching of different regions of the query and content labels (rather than serialized path-oriented inspection).
The KNOWIT system of Sølvberg, Nordbø and Aamodt [SNA92] embeds a semantic network in a front-end query refinement system. Queries to the ESA-QUEST bibliographic database are expanded according to a semantic model that describes the meaning of a concept entirely in terms of its relations to other concepts in the model. The KNOWIT system is a front-end to a traditional IR system, whereas KEYNET uses semantic networks as part of its search strategy.
The Government Information Locator Service (GILS)[GIL] is an example of a second-level retrieval mechanism in which the result of a search is another search engine rather than an information object. The KEYNET model can also be applied to to instances of itself, producing a quasi-encyclopedic classification of information objects. The following diagram suggests how one could organize search engines in the NII:
Using currently available technology, each search engine could support collections having a few million information objects. The top-level search engine could, in turn, support one million search engines. The entire structure would therefore index information objects having an aggregate storage size of bytes, i.e., 10,000 Terabytes.