Many popular latent topic models for text documents generally make two assumptions. The first assumption relates to a finite-dimensional parameter space. The second assumption is the bag-of-words assumption, restricting such models to capture the interdependence between the words. While existing nonparametric admixture models relax the first assumption, they still impose the second assumption mentioned above about bag-of-words representation. We investigate a nonparametric admixture model by relaxing both assumptions in one unified model. One challenge is that the state-of-the-art posterior inference cannot be applied directly. To tackle this problem, we propose a new metaphor in Bayesian nonparametrics known as the "Chinese Restaurant Franchise with Buddy Customers". We conduct experiments on different datasets, and show an improvement over existing comparative models.
展开▼
