UC Berkeley

This work represents a confluence of three lines of research: The OASIS program of studies in adaptive searching, led by Michael Buckland, which from 1990 has developed prototypes for supporting improved use of existing metadata [Buckland et al. 1992, and http://www.sims.berkeley.edu/research/oasis/]; Ray Larson's development of "classification clustering" to create entry vocabulary modules for the Library of Congress Classification in CHESHIRE, a next-generation online catalog and full-text information retrieval system using advanced IR techniques [http://cheshire.lib.berkeley.edu/]; and Fredric Gey's research and development on access to numeric databases and use of probabilistic retrieval techniques in TREC [http://ucdata.berkeley.edu/gey.html]. The rapid increase in network-accessible repositories increases opportunities for searching. Unfamiliar metadata is difficult to search and, increasingly, what is accessible is unfamiliar. Providing indexes to metadata vocabularies offers a remedy. We seek to exploit the potential significance of combining linguistic analysis with statistical methods to help searchers. We are designing an amenity that can easily be used on regular workstations and integrated into actual work environments with little investment of time and effort in order to improve performance when searching and thereby to improve the return on the investment in those systems. Advantages of this work are that it provides an alternative to the expensive human crafting of links within and between vocabularies, it is based on searching of fragments existing within the metadata and databases, it uses advanced probabilistic techniques, it allows the searcher to start a search using familiar search terms from a familiar vocabulary, it is designed for use in a networked environment, and, since only a training set is required, it allows for rapid deployment with unfamiliar metadata schemes.