Using syntax for improving phrase-based SMT in low-resource languages

Abstract

AbstractData driven approaches for machine translation, such as statistical and neural machine translation, suffer from sparsity when dealing with low-resource languages. In these cases, using other sources of information including linguistic information could alleviate the problem. In this article, we focus on the problem of word ordering in translation from a high-resource to a low-resource language and try to improve the quality by using syntactic information from the high-resource side. We propose some syntactic features based on Tree Adjoining Grammar (TAG) to be employed in a phrase-based SMT model in order to improve the word ordering. In this work, a set of synchronous TAG rules is extracted and used to estimate the probability of the phrase orders suggested by the phrase-based model. The main idea of the article is to handle the word ordering by using the extended domain of locality property of TAG and abstracting the long distance dependencies into a local view, which is a TAG elementary tree. The experiments on English–Persian and English–German translation showed that, by combining the proposed TAG-based reordering features with lexical and hierarchical reordering models, we gain significant improvements over the baseline and in comparison with a neural reordering model and a pre-reordering model.