Using morphology to improve Example-Based Machine Translation

We describe how morphological information was used in an Example-Based Arabic-to-English Machine Translation system to produce significant improvement in translation quality on both small and large corpora. We experimented with different methods of generalizing morphology to obtain more candidate source-side matches, while retaining information about the specific input to be translated. This information was then used with adaptation rules and a language model to generate context-appropriate target-side fragments, select and combine them. We outline essential differences between Statistical MT (SMT) and Example-based MT (EBMT), compare ourselves to other EBMT systems used with morphologically complex languages, and justify our choice of EBMT over SMT.