Bayesian Nonparametrics for Information Processing

Abstract

This talk surveys a series of Bayesian nonparametric (BNP) approaches to model selection and their inference procedures which are applied to build information systems including speech recognition, document classification, document summarization and document retrieval. Our goal is to design a flexible, scalable, hierarchical and robust topic models to meet the heterogeneous and nonstationary environments in the era of big data. Two recent works on BNP learning are introduced. One is the hierarchical Pitman-Yor-Dirichlet process for language modeling. The other is the hierarchical theme and topic modeling for document summarization.