Syllabification and parameter optimisation in Zulu to English machine translation

Abstract

We present a series of experiments involving the machine translation of Zulu to English using a well-known statistical software system. Due to morphological complexity and relative scarcity of resources, the case of Zulu is challenging. Against a selection of baseline models, we show that a relatively naive approach of dividing Zulu words into syllables leads to a surprising improvement. We further improve on this model through manual configuration changes. Our best model significantly outperforms the baseline models (BLEU measure, at p < 0.001) even when they are optimised to a similar degree, only falling short of the well-known Morfessor morphological analyser that makes use of relatively sophisticated algorithms. These experiments suggest that even a simple optimisation procedure can improve the quality of this approach to a significant degree. This is promising particularly because it improves on a mostly language independent approach — at least within the same language family. Our work also drives the point home that sub-lexical alignment for Zulu is crucial for improved translation quality.