The importance of score normalization

Generative unigram language models have proven to be a simple though effective model for information retrieval tasks. In contrast to ad-hoc retrieval, topic tracking requires that matching scores are comparable across topics. Several ranking functions based on generative language models: straight likelihood, likelihood ratio, normalized likelihood ratio, and the related Kullback-Leibler divergence are evaluated in two orientations. Best performance is achieved by the models based on a normalized log-likelihood ratio. Key component of these models is the a-priori probability of a story with respect to a common reference dis-