International Journal on Information and Communication Technologies, Vol. 3, No. 3, June 2010 3 Abstract— The field of machine translation has been dominated in the last two decades by statistical and machine learning approaches. Recently, prominent computational linguists [1]-[2]-[3] have expressed misgivings about the exclusive reliance on machine learning approaches to the neglect of the contribution of the symbolic approaches. Furthermore, several machine learning researchers [4]-[5] have recently acknowledged that incorporating linguistic knowledge in their machine learning applications resulted in marked improvement in performance. In this paper, we begin with a brief review of the early attempts at developing machine translation systems and in particular the first English to Arabic machine translation system released in the early eighties. The first generation of machine translation systems followed the direct approach. Part II, is devoted to the rise of the transfer approach in machine translation with an example from the SYSTRAN Arabic to English machine translation system. Part III documents the successes of statistical machine translation systems using the examples of the Language Weaver Arabic-to- English machine translation system and the crowd sourcing Google system. We also talk about the AppTek hybrid approach to machine translation. Part IV concludes the paper. Index Terms— Arabic, rule-based, machine translation, crowdsourcing - Arabic syntax, morphology, machine learning, hybrid systems. I. EARLY MACHINE TRANSLATION SYSTEMS HE invention of the digital computer in the 1940s inspired scientists to think of using the unprecedented speed of the computer to translate texts from one language to another. So inspired, scientists started to take practical steps to realize the dream and vision of Descartes who wrote in 1629 about a mechanical process to convert one human language to another. In 1949, Warren Weaver, the pioneer of machine translation, wrote a memorandum to his colleagues making four proposals for machine translation systems that go beyond word for word translation. Warren realized that many words in language were ambiguous and he proposed in his memorandum to solve this problem by examining the immediate context of the ambiguous word [7]. He also drew attention to the analogy between the structure of the human brain and the “logical machine”. He concluded that the machine translation problem is solvable. He also suggested using the cryptic methods that linguists used in Manuscript received 26 May 2010. Ali Farghaly is a Professor in Computational Linguistics, Senior Member of Technical Staff, Text Group, Oracle USA, CA; Adjunct Professor of Arabic Linguistics, Monterey Institute of International Studies, Monterey, CA, USA. E-mail: afarghal@miis.edu the Second World War for deciphering the German secret code. These cryptographic methods relied heavily on frequencies of letters, combination of letters and letter patterns. He also believed that underlying the statistical regularities of languages, there is a logical and universal foundation which could represent an alternative to translate from one language to another. At the same time with the beginning with the cold war in the 1940s, there was an urgent need for crude machine translation because the United States decided it was essential to scan and interpret every Russian communication coming out of the Soviet Union. However, there weren’t enough translators to keep up with the huge volume of Russian books and papers published in the Soviet Bloc at that time. The urgent need to translate Russian into English coincided with the invention of computers. It was not surprising then, that developing Russian to English machine translation systems would be one of the first tasks these “miracle” machines were set to perform. The first demonstration of the feasibility of fully automated machine translation took place in New York on January 7 th , 1954. On that day, Georgetown University and IBM demonstrated the first non-numerical applications and capabilities of the “new” electronic brain by demonstrating a fully automated Russian English machine translation system. The system embraced the commonly held view that a language consisted of a lexicon and a finite set of rules that could generate an infinite set of sentences. Surprisingly, the first Russian to English machine translation system had only 250 words and 6 syntactic rules. This experiment raised high expectations that probably within five years machine translation systems would be readily available. The promise was to develop a system that does not require pre-editing of the input while produces a reliable translation of the input text in the target language that is clear, intelligible requiring only stylistic modifications. No details were given about the actual linguistic processing in the system. For example, no information about dictionary content and lookup procedures were given. No account of how the syntactic analysis of the Russian sentences was performed and how the target English structure was selected. However, there were some references to reversing the order of pairs of sentences by assigning rules to the lexical items involved. Later on a more detailed description of the system is presented in [8]. Garvin [8] gives more detailed description of the dictionary. For example, the dictionary entries were sometimes stems, endings or full words. Each entry is associated with three codes; the first code indicates which of the six syntactic rules would apply, the Arabic Machine Translation: A Developmental Perspective Ali Farghaly T