Gaussian alignments in statistical translation models

Abstract

Machine translation software has been under development almost since the birth of the electronic computer. Current state-of-the-art methods use statistical techniques to learn how to translate from one natural language to another from a corpus of hand-translated text. The success of these techniques comes from two factors: a simple statistical model and vast training data sets. The standard agenda for improving such models is to enable it to model greater complexity; however, it is a byword within the machine learning community that added complexity must be supported with more training data. Given that current models already require huge amounts of data, our agenda is instead to simplify current models before adding extensions. We present one such simplification, which results in fewer than 10% as many alignment model parameters and produces results competitive with the original model. An unexpected benefit of this technique is that it naturally gives a measure for how difficult it is to translate from one language to another given a data set. Next, we present one suggestion for adding complexity to model new behavior.