Bayesian learning of n-gram statistical language modeling

The n-gram language model adaptation is typically formulated using deleted interpolation under the maximum likelihood estimation framework. This paper proposes a Bayesian learning framework for n-gram statistical language model training and adaptation. By introducing a Dirichlet conjugate prior to the n-gram parameters, we formulate the deleted interpolation under maximum a posterior criterion with a Bayesian learning procedure. We study the Bayesian learning formulation for n-gram and continuous n-gram language models. The experiments on North American News Text corpus have validated the effectiveness of the proposed algorithms.