| Statistical machine translation (SMT) treats the translation of natural language as a machine learning problem. By examining many samples of human-produced translations, SMT algorithms automatically learn how to translate. In this thesis, we discuss the automatic machine translation from English to Arabic using a statistical phrase-based approach employing a parallel Arabic-English corpus that was developed manually by more than one translator. Statistical machine translation (SMT) consists of two phases: The training phase and the decoding phase. In the training phase, the statistical language model and the translation model are built. In the decoding phase, the best possible translation is chosen depending on a comprehensive search process. We built a sizable parallel corpus spanning various categories of topics from the Meedan website, and later compared the results of Meedan with that of the other two corpora: LDC and UN. The performance was compared based on the Bilingual Evaluation Understudy (BLEU). Our experimentation shows that, overall, the Meedan corpus outperformed the other two corpora in most categories. We, also, compared the performance of the Moses decoder and the Pharaoh decoder. We conclude that although the response time for the pharaoh decoder is better than that of the Moses decoder, the quality of the translation of the Moses decoder exceeds that of the Pharaoh decoder. |
