Abstract
Automatic extraction of term semantic classes is an interesting task in the Information Retrieval field. Models such as Latent Dirichlet Allocation or Probabilistic Semantic Indexing are able to provide the probability that a term belongs to a given semantic class. However such models do not provide neither Euclidean coordinates for terms nor a hierarchical structure to organize the latent semantic classes, which makes difficult to visualize the information under consideration. In this work we propose a hierarchical latent topic extraction method that exploits the information contained in asymmetric term similarity matrices. Our method produces Mercer Gram matrices for terms organized by frequency levels and then hierarchically combines classes belonging to different levels. Euclidean coordinates for terms can be recovered from the proposed kernel matrices. Our proposal also provides explicit conditional probabilities, as the Latent Dirichlet Allocation model does, but avoiding the computational burden usually involved in the iterative step of such probabilistic models. Finally, we analyze a real data base showing the advantages of the new approach.
