Properties of extended Boolean models in information retrieval

Reviewer: Danny B. Lange

Ranked document retrieval outputs in decreasing order of query-document similarities are obviously useful to control the size of a retrieved document set. Conventional Boolean retrieval systems, however, do not provide such ranked output because they cannot compute similarity coefficients between queries and documents. The author of this paper analyzes several extended Boolean models in order to determine which one is the most suitable for achieving high retrieval effectiveness. Although extended Boolean models use document term weights to calculate query-document similarities, ranking is often not satisfactory. The author demonstrates with clear examples that some models (fuzzy set models) can generate incorrectly ranked output that does not agree with human behavior. Positively compensatory operators and binary soft Boolean operators in other models (Waller-Kraft, Paice, P-Norm [1], and Infinite-One) are shown to overcome this problem. The author continues to demonstrate with a new set of clear examples that these models (except for P-Norm) still violate the usual assumption that all the terms given in a query are equally important. Lee concludes that, since P-Norm is the only model that solves both the deficiency of fuzzy models and the unequal importance problem, it is more effective than any of the other extended Boolean models. The concluding section is devoted to an analysis of the meaning of query weights. The analysis concludes that P-Norm is superior, since it uses relative query weights, found to be easier for users to write than absolute query weights. The author provides clear examples and presents the analysis in a very readable and convincing form supported by well-written mathematical proofs.