Example-Based Arabic-English Machine Translation Kfir Bar, Nachum Dershowitz, and Yaacov Choueka June 2006 We describe an implementation of an Example-Based Machine Translation system that translates phrases from Arabic to English. The system uses a large parallel corpus, aligned at the paragraph level, and built using many parallel Arabic-English documents from the United-Nations database. This is a non-structural system, so examples are stored in their surface forms, with some additional morphological and part-of-speech information. Each new input phrase is matched to example patterns by using various levels of morphological data. Matched phrases are transferred to English using a rough word-level alignment algorithm, which makes use a bilingual dictionary, along with WordNet. There is no syntax parser involved in the matching and/or transfer processes, although we do use a shallow parser to help with specific situations. Currently, the system first fragments any newly introduced input phrase and then translates each segment separately; recombining those translations into final coherent form is left for future work. We encountered several problems in the matching and the transfer steps, some of which were solved, partially or totally, sometimes by using linguistic tools for both languages. We discuss those problems and our proposed solutions. The system has been implemented and manually evaluated. Preliminary results are encouraging.