Learning Curved Multinomial Subfamilies for Natural Language Processing and Information Retrieval
Keith Hall, Thomas Frank Hofmann · 2000
Many problems in natural language learn-ing and information retrieval involve estimat-ing probabilities in very large discrete state spaces. Dimension reduction as well as clus-tering techniques in various avors have been popular choices to deal with the problem of data sparseness. In this paper, we present a general framework for dimension reduc-tion based on curved multinomial subfami-lies. The investigated class of models include dierent geometries as well as various objec-tive functions and algorithms for model t-ting. The pursued goal is twofold { to achieve a systematic understanding of the dierences and similarities between various models and to empirically investigate their generalization performance on a number of representative data sets. 1.