next up previous
Next: Experimental Rankings Up: Measuring Search Engine Quality Previous: Analytic Models of Performance

Document Rankings from Commercial Retrieval Engines

Any experiments involving Freestyle and Target must of necessity be ``black box" experiments [RHB92], since the algorithms used in these retrieval systems are trade secrets. Based on system documentation, however, we can conclude that both systems employ algorithms based on the vector space and probabilistic models, although the exact values used to calculate relevance remain a mystery. In their evaluation of Target, Tenopir and Cahn [TC94] state that document weights are adjusted for document length, but Keen [Kee94] asserts that he did not detect any clear evidence of such adjustment. Ingwersen [Ing96] suggests that ``Target is applying quorum logic (in the traditional way), document term frequencies and collection term frequencies as elements of its ranking algorithm," (p. 45) but provides no evidence for this claim. We do know, however, that Target's ranking algorithm includes at least four variables [Kee94]:
1.
number of search terms in each record,
2.
proximity of search terms to each other in a record,
3.
frequency of a term in the database, and
4.
length of the document.
Freestyle, on the other hand, provides a little more information about the information retrieval process used. For example, the .WHERE and .WHY screens in Freestyle show that a term's weight is inversely proportional to its frequency in the database. In fact, the Freestyle HELP explanation about query term weights states that ``term importance is based on how frequently the term appears in the file(s) you are searching. The more often a term occurs, the lower its term importance." These facts, then, suggest that the system employs some version of inverse document frequency to calculate term weights [SJ72]. When calculating the inverse document frequency weight (IDF), ``terms with medium to low collection frequencies are assigned high weights as good discriminators, while frequent terms have low weights" [RSJ76, pp. 129-30]. The matching algorithm for Freestyle appears to be derived from the vector space and probabilistic models, where the weight of each document is the sum of the products of term weights and frequencies of the terms in the document. The ranking algorithm for Freestyle, then, appears to involve a minimum of three variables:
1.
frequency of a search term within the database,
2.
frequency of a search term in a record, and
3.
number of search terms in a record.


 
next up previous
Next: Experimental Rankings Up: Measuring Search Engine Quality Previous: Analytic Models of Performance
Bob Losee
1999-07-29