Integrating morpho-syntactic features in English-Arabic statistical machine translation

This paper presents a hybrid approach to the enhancement of English to Arabic statistical machine translation quality. Machine Translation has been defined as the process that utilizes computer software to translate text from one natural language to another. Arabic, as a morphologically rich language, is a highly flexional language, in that the same root can lead to various forms according to its context. Statistical machine translation (SMT) engines often show poor syntax processing especially when the language used is morphologically rich such as Arabic. In this paper, to overcome these shortcomings, we describe our hybrid approach which integrates knowledge of the Arabic language into statistical machine translation. In this framework, we propose the use of a featured language model SFLM (Smaïli et al., 2004) to be able to integrate syntactic and grammatical knowledge about each word. In this paper, we first discuss some challenges in translating from English to Arabic and we explore various techniques to improve performance on this task. We apply a morphological segmentation step for Arabic words and we present our hybrid approach by identifying morpho-syntactic class of each segmented word to build up our statistical feature language model. We propose the scheme for recombining the segmented Arabic word, and describe their effect on translation.