Towards a better understanding of the relationship between probabilistic models in IR

Abstract

Probability of relevance (PR) models are generally assumed to implement the Probability Ranking Principle (PRP) of IR, and recent publications claim that PR models and language models are similar. However, a careful analysis reveals two gaps in the chain of reasoning behind this statement. First, the PRP considers the relevance of particular documents, whereas PR models consider the relevance of any query-document pair. Second, unlike PR models, language models consider draws of terms and documents. We bridge the first gap by showing how the probability measure of PR models can be used to define the probabilistic model of the PRP. Furthermore, we argue that given the differences between PR models and language models, the second gap cannot be bridged at the probabilistic model level. We instead define a new PR model based on logistic regression, which has a similar score function to the one of the query likelihood model. The performance of both models is strongly correlated, hence providing a bridge for the second gap at the functional and ranking level. Understanding language models in relation with logistic regression models opens ample new research directions which we propose as future work. © 2011 Springer-Verlag.