SMT: A case study of Kazakh-English word alignment

Abstract

In this paper, we present results from a set of experiments to determine the effect on translation quality, it depends on the particular kind of morphological preprocessing that can be represented by finite-state transducers. A high agglutinative nature of the Kazakh language under the condition of poor language resources makes an issue in the processing of derivational morphology. Our methods are focused on useful phrase pairs in word alignment and provide a most language independent approach, which may improve a translation into other morphological complex languages. We processed our algorithms over the Kazakh Wikipedia base of about 1.5 million unique lexeme and 230 million words overall. Our best translation system increases 3 BLEU points over the Kazakh-English baseline on a blind test set.