Munich-Edinburgh-Stuttgart Submissions at WMT13: Morphological and Syntactic Processing for SMT

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 232–239, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Munich-Edinburgh-Stuttgart Submissions at WMT13: Morphological and Syntactic Processing for SMT Marion Weller1, Max Kisselew1, Svetlana Smekalova1, Alexander Fraser2, Helmut Schmid2, Nadir Durrani3, Hassan Sajjad4, Rich´ard Farkas5 1University of Stuttgart – (wellermn|kisselmx|smekalsa)@ims.uni-stuttgart.de 2Ludwig-Maximilian University of Munich – (schmid|fraser)@cis.uni-muenchen.de 3University of Edinburgh – dnadir@inf.ed.ac.uk 4Qatar Computing Research Institute – hsajjad@qf.org.qa 5University of Szeged – rfarkas@inf.u-szeged.hu Abstract We present 5 systems of the Munich- Edinburgh-Stuttgart1 joint submissions to the 2013 SMT Shared Task: FR-EN, EN- FR, RU-EN, DE-EN and EN-DE. The first three systems employ inflectional gen- eralization, while the latter two employ parser-based reordering, and DE-EN per- forms compound splitting. For our ex- periments, we use standard phrase-based Moses systems and operation sequence models (OSM). 1 Introduction Morphologically complex languages often lead to data sparsity problems in statistical machine trans- lation. For translation pairs with morphologically rich source languages and English as target lan- guage, we focus on simplifying the input language in order to reduce the complexity of the translation model. The pre-processing of the source-language is language-specific, requiring morphological anal- ysis (FR, RU) as well as sentence reordering (DE) and dealing with compounds (DE). Due to time constraints we did not deal with inflection for DE- EN and EN-DE. The morphological simplification process con- sists in lemmatizing inflected word forms and deal- ing with word formation (splitting portmanteau prepositions or compounds). This needs to take into account translation-relevant features (e.g. num- ber) which vary across the different language pairs: while French only has the features number and gender, a wider array of features needs to be con- sidered when modelling Russian (cf. table 6). In addition to morphological reduction, we also apply transliteration models learned from automatically 1The language pairs DE-EN and RU-EN were developed in collaboration with the Qatar Computing Research Institute and the University of Szeged. mined transliterations to handle out-of-vocabulary words (OOVs) when translating from Russian. Replacing inflected word forms with simpler variants (lemmas or the components of split com- pounds) aims not only at reducing the general com- plexity of the translation model, but also at decreas- ing the amount of out-of-vocabulary words in the input data. This is particularly the case with Ger- man compounds, which are very productive and thus often lack coverage in the parallel training data, whereas the individual components can be translated. Similarly, inflected word forms (e.g. ad- jectives) benefit from the reduction to lemmas if the full inflection paradigm does not occur in the parallel training data. For EN-FR, a translation pair with a morpho- logically complex target language, we describe a two-step translation system built on non-inflected word stems with a post-processing component for predicting morphological features and the genera- tion of inflected forms. In addition to the advantage of a more general translation model, this method also allows the generation of inflected word forms which do not occur in the training data. 2 Experimental setup The translation …

Content truncated to 500 words