Learning Aspect Models with Partially Labeled Data

Anastasia Krithara(1,2), Massih-Reza Amini(1,3), Cyril Goutte(3), Jean-Michel Renders(2)
(1) Laboratoire d'Informatique Paris 6              (2)Xerox Research Center Europe                 (3) CNRC
                            5, Place de Jussieu                                 6, Chemin de Maupertuis                  283, Bd Alexandre-Taché
                               75252 Paris, cedex 05                                       38240 Meylan                            Gatineau, QC J8X 3X7          

In this paper, we address the problem of learning aspect models with partially labeled data for the task of document categorization. The motivation of this work is to take advantage of the amount of available unlabeled data together with the set of labeled examples to learn latent models whose structure and underlying hypotheses take more accurately into accountthe document generation processm compared to other mixture-based generative models. We present one semi-supervised variant of the PLSA model. In our approach, we try to capture the possible data mislabeling errors which occur during the training of our model. This is done by iteratively assigning class labels to document collections, as well as over a real world dataset coming from a Business Group of Xerox and show the effectiveness of our approach compared to a semi-supervised version of Naive Bayes, another semi-supervised version of PLSA and to transductive Support Vector Machines.