Scott Preece on retrieval

3 Oct 91
(See tracing, Navigation)

My own "vision" of information retrieval models the whole database as

a network of objects. Some of the objects are words, some of them are

index terms (from a controlled vocabulary), some of them are documents some of them are pieces of documents, some of them are authors, etc. There are also typed links between nodes in the graph -- words are connected to documents by occurrence links, words are tied to words by dictionary links, document pieces are tied to documents by "is section of" links, etc. Searching then becomes a process of turning some of the nodes on, then turning on the nodes attached to them by certain kinds of links, and so forth.

So a dictionary expansion of the query works by activating a set of terms and then following all the dictionary links from those terms to other terms; a "search" works by activating a set of terms, then following all the occurrence links to the documents they appear in; relevance feedback works by starting with a set of activated documents and following the links back to the terms that occur in them.

If you use appropriate rules for calculating the level of activation of a node you can implement many of the similarity functions that have been reported in the literature and do a pretty effective job of seaching. For instance, suppose you have a term node which is activated with a

weight of 1. Suppose the spreading rule is that the weight is split among all the occurrence links leading from it to documents and the

combining rule is that all weights coming into a node are summed. Then after one spreading cycle each active document will have a weight equal to the sum of the inverse frequency of the terms in contains, which is a pretty reasonable search strategy. One enhancement is to have each link also weighted -- for term occurrence links it makes sense for that weight to be the number of occurrences of the term in the document.

It is true that doing this effectively requires doubly inverting the database, so that each document points to all its terms as well as vice versa, although you can finesse that by encoding the document as a list of terms rather than as Ascii text, with a slightly higher cost of

rebuilding the text when you need to display the document.

[My dissertation, describing this in excruciating detail, is *A Spreading Activation Model for Information Retrieval*, University of Illinois, 1981. You might be able to get it from University Microfilms if you're really interested. If you're at Thinking Machines, Dave Waltz had a copy once, but may well have shed it in the last decade. The machine readable form, alas, no longer exists (it lived on a long-dead PDP10)]


-- scott preece motorola/mcg urbana design center 1101 e. university, urbana, il 61801 uucp: uunet!uiucuxc!udc!preece, arpa: phone: 217-384-8589 fax: 217-384-8550