Morphological Processing and Word Reordering for Statistical MT 
of Highly Inflected Languages

Marcello Federico, FBK Trento, Italy

Abstract

It is well known that phrase-based statistical MT performs best when
source and target languages have similar sentence structure and
morphology.  My talk will overview recent work done at FBK on
statistical MT from German, Arabic and Turkish to English.  These
languages are characterized by different morphological phenomena which
require specific pre-processing methods to reduce data sparseness, to
improve automatic word alignment at training time, and to induce
better word reordering at decoding time.  In particular, I will survey
morphological decomposition and reduction techniques, as well as verb
reordering methods that showed to positively impact on MT accuracy.
Experimental results will be presented for the Arabic-English NIST
2009 task, the German-English WMT 2010 task, and the Turkish-English
IWSLT 2010 task.