Experiments in Morphosyntactic Processing for Translating to and from German

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 115–119, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Experiments in morphosyntactic processing for translating to and from German Alexander Fraser Institute for Natural Language Processing University of Stuttgart fraser@ims.uni-stuttgart.de Abstract We describe two shared task systems and associated experiments. The German to English system used reordering rules ap- plied to parses and morphological split- ting and stemming. The English to Ger- man system used an additional translation step which recreated compound words and generated morphological inflection. 1 Introduction The Institute for Natural Language Processing (IfNLP), Stuttgart, participated in the WMT-2009 shared tasks for German to English and English to German translation with constrained systems which employed morphological and syntactic pro- cessing techniques. The systems were based on the open source Moses docoder (Koehn et al., 2007). We combined IfNLP tools for syntactic and morphological analysis (which are publicly avail- able and widely used) with preprocessing tech- niques that were successfully used by other groups in WMT-2008, and extended these. For English to German translation, we additionally performed a step which recreated compound words and gener- ated morphological inflection. 1.1 Baseline The baseline is the standard system supplied for the shared task. We used the default parameters of the Moses toolkit, except for a small difference in the generation of the word alignments, see sec- tion 3. 2 Improvements 2.1 Character Normalization We normalize both the English and German by converting all characters to their nearest equivalent in Latin-1 (ISO 8859-1) encoding1, except for the euro sign, which is handled specially. We did not modify the SGML files used for calculating BLEU and METEOR scores in any way. 2.2 German Writing Reform German underwent a writing reform from the alte Rechtschreibung (old spelling rules/orthography) to the neue Rechtschreibung (gloss: new spelling rules/orthography) recently. Early Europarl data are written using the alte Rechtschreibung and hence need to be converted to the neue Rechtschreibung in order to match the news data, which is in the new form. We began the process by mapping all cased vari- ants of a particular word to a single class (such as by mapping two words which are written with ue and ¨u, but are otherwise identical, to a single class). We then tried to automatically identify the correct variant under the writing reform for each class. Initially we tried the linux tool aspell but found that its coverage (the recall of its lexicon) was poor. We used a simple technique for finding the best variant. We separated the Europarl corpus into portions written using the old and new forms. We used the incidence of the word dass (the comple- mentizer meaning that) and its old rules variant daß. We used a chunk size of 70 sentences to segment Europarl into old and new by counting whether there were more instances of daß or dass, respectively, in each chunk. We added the news corpora to the new portion. For each variant …

Content truncated to 500 words