Unsupervised and Semi-supervised Myanmar Word Segmentation Approaches for Statistical Machine Translation

In statistical machine translation (SMT), word segmentation is generally a necessary step for languages that do not naturally delimit words. For many low-resource languages there are no word segmentation tools, and research on word segmentation for these languages is often quite scarce. In this paper, we study several plausible methods for Myanmar word segmentation for machine translation in order to shed light on promising avenues for future research. We propose a novel Bayesian learning method that can perform semisupervised word segmentation with reference to a dictionary. We applied our method to the task of translating with Myanmar language, and compare our method to the following approaches to segmentation: human lexical/phrasal segmentation, character segmentation, syllable segmentation, purely unsupervised word segmentation, and the method of maximum matching. We found that unsupervised segmentation was the most effective segmentation. It outperformed maximum matching, which in turn was better than syllable segmentation.