Revisiting fisher kernels for document similarities

Abstract

This paper presents a new metric to compute similarities between textual documents, based on the Fisher information kernel as proposed by T. Hofmann. By considering a new point-of-view on the embedding vector space and proposing a more appropriate way of handling the Fisher information matrix, we derive a new form of the kernel that yields significant improvements on an information retrieval task. We apply our approach to two different models: Naive Bayes and PLSI. © Springer-Verlag Berlin Heidelberg 2006.