“Is this document relevant?…probably”: a survey of probabilistic models in information retrieval

Reviewer: Karen Sparck-Jones

This useful review provides a competent, clear, and acces sible account of retrieval models that take probability as their grounding notion in defining the relevance relation between queries and documents. These models are divided into two broad classes: those treating the probability of relevance of a document to a query in a direct way (statistical models) and those allowing indirection through reasoning (inference models). After presenting the pertinent generic information retrieval notions, the paper describes eight approaches in the first class and two in the second. As it covers the source material well and uses a common formalism that makes comparisons between the different approaches easier, the paper will be a valuable reference for students and beginning researchers. The paper has a major weakness, however. There is no detail on, assessment of, or serious literature reference to experiments where the models have been applied. The paper's primary emphasis on models is legitimate. However, there is no proper indication of the extent to which the different models have been implemented, of the range of retrieval tests conducted, or, especially, of direct performance comparisons between them. Much model development has been promoted by actual retrieval experiments, and more than one model from the first class and one from the second have been heavily applied. The NIST/DARPA Text Retrieval Conferences (TREC) [ 1] have, moreover, allowed direct comparisons between these models, which are important in providing evidence both for their relative merits and for their performance when compared with other nonprobabilistic models or even with wholly informal approaches. Even at the time the paper was completed (1997), there was a substantial TREC literature.