Multi-dynamic bayesian networks for machine translation and nlp

Many important applications such as document summarization, speech recognition, DNA sequence alignment, sentence labeling, syntactic parsing, and statistical machine translation can be cast as structured classification problems, where we want to learn a mapping between structured sets of inputs and outputs, such as sequences, trees, or arbitrary networks. A general framework for easily expressing and reasoning about this type of problems is highly desirable but is not without challenges. We present a general-purpose graphical model-based framework for structured classification and show its expressive power through two main applications: learning new generalized string edit distance dynamic Bayesian models (DBNs), and statistical machine translation (MT) using multi-dynamic Bayesian networks (MDBNs), which address several shortcomings of DBNs in modeling multi-stream stochastic processes. We discuss inference, learning and search algorithms developed for the MDBN framework. We also present a novel multilingual MT word alignment model, the general idea behind which can be applied in other multiple string alignment settings.