Università di Milano-Bicocca

Title:  The flexible Latent Dirichlet Allocation

Abstract:   Over recent years, text modeling techniques have been employed in several applications, including the detection of latent topics in text documents. A widespread statistical tool for topic modeling is the Latent Dirichlet Allocation (LDA), which allows for a document representation in terms of topic composition. A well-known limitation of the LDA is related to the stiffness of the Dirichlet prior imposed on the topic distributions. To consider a richer dependence structure, we propose a generalization of the Dirichlet distribution as an alternative distribution, namely the flexible Dirichlet (FD). The FD is a distribution defined on the simplex space allowing for a finite mixture structure. This choice introduces additional parameters in the LDA, and ensures more flexibility, still maintaining the model interpretability, as well as conjugacy to the multinomial model. The latter property allows for a Collapsed Gibbs Sampling-based estimation procedure. The generalization of the LDA based on the FD distribution is illustrated via an application to a real dataset.

-----------------------------------