Modeling texts in semantic space and ensemble topic-models via boosting strategy

Topic model, especially Latent Dirichlet Allocation, is a probabilistic graphic model for feature selection and dimension reduction in text categorization tasks. With the mapping from word space to the latent topic space, there are more benefits, but at the same time, the inference and estimate of the model parameters become a new trouble. This article first improved the traditional LDA through introducing an ontology called CILIN, and then we constructed a novel classification algorithm that combined different models with different parameters together via boosting strategy. Moreover, Naïve Bayes and Support Vector Machine are employed as weak classifier and a weighted method is proposed for improving the accuracy by integrating weak classifiers into strong classifier in a more ration way. Experiment results show our method well perform both in accuracy and generalization.