Title:  The Impact of Arabic Morphological Segmentation on Broad-Coverage English-Arabic Statistical MT

Abstract:
The relative morphological richness and complexity of Arabic in
comparison with English imposes challenges on the design of
phrase-based SMT systems between the two languages.  Morphological
segmentation impacts multiple components and resources along the MT
training pipeline and runtime decoder.  We explore the full spectrum
of segmentation schemes ranging from full word forms to
fully-segmented morphemes, and examine their impact, all within the
context of training broad-scale phrase-based SMT systems between
English and Arabic.  Our results and analysis indicate a complex
picture, where some morphological decomposition choices help improve
performance whereas others have no impact or hurt performance.
Furthermore, the resulting variation in system output can be further
leveraged for improving performance via system combination.