Transductive learning for text classification using explicit knowledge models

We present a generative model based approach for transductive learning for text classification. Our approach combines three methodological ingredients: learning from background corpora, latent variable models for decomposing the topic-word space into topic-concept and concept-word spaces, and explicit knowledge models (light-weight ontologies, thesauri, e.g. WordNet) with named concepts for populating latent variables. The combination has synergies that can boost the combined performance. This paper presents the theoretical model and extensive experimental results on three data collections. Our experiments show improved classification results over state-of-the-art classification techniques such as the Spectral Graph Transducer and Transductive Support Vector Machines, particularly for the case of sparse training.