Morphological Generation of German for SMT

We participated in the ACL WMT 2009 shared task for translation of German to English, and English to German. We used the Moses open source system, combined with morphological processing. For German to English, we had the only constraint system comparable with the open-data systems. One of the reasons the system performed well was strong reduction of the German vocabulary, through a simplistic corpusdriven algorithm with minimal linguistic knowledge, which performed aggressive inflection removal and compound splitting. For the English to German task, we submitted a two-step system. In the first step we translated English to the reduced German representation (the same representation which we used as the input to our German to English system). In the second step, we built another Moses system to “translate” the reduced German representation back to normal German through the addition of inflection and merging of split compound words. This two-step system was the worst constraint system submitted (and in fact the only constraint system that differed from all other constraint systems which otherwise scored as a group). Our presentation at the Research Workshop of the Israel Science Foundation on Machine Translation and Morphologically-rich Languages will describe work done as a direct reaction to the poor performance of our English to German system, and to the poor performance of SMT systems in general for this translation task. Our current English to German system for news translation improves by 0.84 BLEU over the baseline and uses sophisticated morphological generation based on SMOR, the University of Stuttgart morphological analyzer/generator of German, and BitPar, a state-of-the-art parser of German also developed in-house.