Transformation and Decomposition for Efficiently Implementing and Improving Dependency-to-String Model In Moses

Abstract

Dependency structure provides grammatical relations between words, which have shown to be effective in Statistical Machine Translation (SMT). In this paper, we present an open source module in Moses which implements a dependency-to-string model. We propose a method to transform the input dependency tree into a corresponding constituent tree for reusing the tree-based decoder in Moses. In our experiments, this method achieves comparable results with the standard model. Furthermore, we enrich this model via the decomposition of dependency structure, including extracting rules from the substructures of the dependency tree during training and creating a pseudo-forest instead of the tree per se as the input during decoding. Large-scale experiments on Chinese‐English and German‐English tasks show that the decomposition approach improves the baseline dependencyto-string model significantly. Our system achieves comparable results with the state-of-the-art hierarchical phrase-based model (HPB). Finally, when resorting to phrasal rules, the dependency-to-string model performs significantly better than Moses HPB.