Conventional IR systems focus on keyword based automatic searching (in conjunction with Boolean operations), weighting of words based on their statistical properties, ranking of documents according to probability of relevance, automatic relevance feedback for query modification and query languages [Croft et al., 1990]. However, very few (or none) of these methods retrieve complete or accurate information. Too general a query may yield a lot of items and too specific a query may retrieve no items. Thus, traditional IR is an inherently uncertain process. Combining inference techniques could eliminate or minimize uncertainty. In hypertext systems, a weighted keyword search combined with hypertext links can improve IR by finding only a subset of nodes or "hits" whose links can then be followed to other semantically related nodes [Carlson, 1989].
According to Halasz, query and search mechanisms can be classified into content search and structure search [Halasz, 1988]. Content search is standard IR technique extended to hypertext systems. That is, all nodes and links are treated independently and examined for a match to the given query. On the other hand, structure search will yield the hypertext sub-network that matches a given pattern. Query facilities which combine aspects of both content search and structure search will be capable of acting as filters. Based on the user's query, the interface will display only those nodes and links that match the query, filtering out other parts of the network. Filtered browsers have been implemented both for NoteCards and Tektronix's Neptune. In NoteCards, a user can filter out information based on the node or link type. In Neptune, the query can be content-based; if the query is broad enough, a global view of the entire network is displayed; if the query is well refined, the viewing size will be manageable.