Estimating Document Similarity using Auxiliary Category Information

We have developed a novel approach to determine the similarity of documents using probabilistic latent semantic indexing. For each document a probability vector of latent factors is estimated which on the one hand takes into account the distribution of words in the text and on the other hand the distribution of category values. The emphasis can be freely shifted between both aspects and therefore the method allows to select a similarity measure which is more appropriate to the domain. In a preliminary evaluation we determined articles which are most similar to a target document. In spite of the small training set the groups of similar articles exhibits a remarkable coherence.