Myanmar Phrase Translation for Myanmar-English Machine Translation System Thet Thet Zin, Khin Mar Soe, Ni Lar Thein University of Computer Studies, Yangon thetthetzin.ucsy@gmail.com Abstract In statistical machine translation (SMT), the currently best performing systems are based in some way on phrases or word groups. In this paper, phrase-based translation system is described. Translation model and word-based lexicon model are used in this system. This system is implemented as a separated part of the Myanmar to English machine translation system. Myanmar language does not use space between words. Therefore, the system uses Myanmar Word Segmenter, which is implemented in UCSYNLP lab and is available for research purpose, to segment Myanmar sentence. Myanmar-English bilingual corpus is also used as a main knowledge source. The vast amount of information is needed to guide the translation process. The large scale Myanmar corpus is unavailable at present. This system aims to increase correct translation results with limited bilingual corpus. The experimental results show that the proposed system seems promising for Myanmar to English machine translation system. 1. Introduction The goal of creating statistical machine translation (SMT) systems incorporating rich, spare, features over syntax and morphology has consumed much recent research attention. In SMT approach, MT is treated as a decision problem: given the source language sentence, we have to decide for the target language sentence that is the best translation. Bayes rule and statistical decision theory are used to address this decision problem. Statistical models estimated parameters based on bilingual text corpora. It differs from example-based MT (EMT) is the ranking between fragments is done with probabilities rather than matching measures. SMT approach relies only on automatically detected word correspondences and alignment patterns. In the source-channel approach, ) | ( f e P depends on two factors ) (e P and reverse translation probability ) | ( e f P . In a direct translation approach, various feature function M m f e h m ,...., 1 ), | (  is developed. In this case, the free model parameters are m 1  . A standard training criterion for the translation model in the source-channel approach is the maximum likelihood criterion and in maximum entropy based translation models. Today’s statistical translation models ) | ( e f P are only rough approximations to the “true” probability distributions ) | ( f e P . Therefore, certain natural language phenomena cannot be handled well. Source language architecture uses language model to get well form of target language sentence. SMT approach has advantages in machine translation. Firstly, In SMT, we have a mathematically well-founded machinery to perform on optimal combination of these knowledge sources. Second, translation knowledge is learned automatically from example data. As a result, MT system based on statistical methods is very fast compared to the rule-based approach. Third, SMT can deal with lexical word ambiguity involving context or