SEMANTIC NETWORKS: A STOCHASTIC MODEL OF THEIR PERFORMANCE IN INFORMATION RETRIEVAL

Recent advances in computer technology have made possible the implementation of information retrieval strategies adopting semantic network approaches from artificial intelligence. This research investigates the claim that document based information retrieval strategies which automatically extend the user's query by inferring, through the use of a semantic network, the concepts that the user wishes to retrieve, can outperform traditional Boolean retrieval strategies. The claim reflects a growing concern among researchers and commercial information retrieval services that many relevant documents are not being retrieved in systems using Boolean retrieval strategies. Research by Furnas, and Blair and Maron attribute this to problems in user query formulation, particularly the use of multiple terms to express similar meanings.; This research compares two classes of document based information retrieval strategies--Boolean and "concept" extended. The latter strategy is adapted from research by Kochen on semantic network based document retrieval. The Kochen algorithm generates an equivalence class of terms called "concepts" from a semantic network which are used to disjunctively extend a document database query.; This research analyzes Boolean and "concept" extended retrieval strategy performance via a stochastic model of user querying and information retrieval. Stochastic processes for querying, retrieving and constructing semantic networks are proposed, and resulting probability distributions are constructed in the analysis. The research also investigates various alternatives for measuring the performance of document retrieval systems.; The research results show improved performance measured by most performance statistics for the "concept" extended retrieval strategy over the traditional Boolean strategy. "Concept" extended queries were shown to improve expected recall as much as 20%, and expected precision as much as 10%. This is attributed to that strategy's injection of new information about document database content into the user's query. The research also provides evidence for the positive correlation of precision and recall, and for the superior informativeness of precision and recall over composite and Neyman-Pearson measures of retrieval performance.; The research additionally provides a limited microeconomic analysis of the costs and benefits of document based information retrieval strategies in a competitive marketplace. This provides economic guidelines for commercial implementations of document retrieval systems using "concept" extended retrieval strategies.