Munich-Edinburgh-Stuttgart Submissions of OSM Systems at WMT13

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 122–127, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Munich-Edinburgh-Stuttgart Submissions of OSM Systems at WMT13 Nadir Durrani1, Helmut Schmid2, Alexander Fraser2, Hassan Sajjad3, Rich´ard Farkas4 1University of Edinburgh – dnadir@inf.ed.ac.uk 2Ludwig Maximilian University Munich – schmid,fraser@cis.uni-muenchen.de 3Qatar Computing Research Institute – hsajjad@qf.org.qa 4University of Szeged – rfarkas@inf.u-szeged.hu Abstract This paper describes Munich-Edinburgh- Stuttgart’s submissions to the Eighth Workshop on Statistical Machine Transla- tion. We report results of the translation tasks from German, Spanish, Czech and Russian into English and from English to German, Spanish, Czech, French and Rus- sian. The systems described in this paper use OSM (Operation Sequence Model). We explain different pre-/post-processing steps that we carried out for different language pairs. For German-English we used constituent parsing for reordering and compound splitting as preprocessing steps. For Russian-English we transliter- ated the unknown words. The translitera- tion system is learned with the help of an unsupervised transliteration mining algo- rithm. 1 Introduction In this paper we describe Munich-Edinburgh- Stuttgart’s1 joint submissions to the Eighth Work- shop on Statistical Machine Translation. We use our in-house OSM decoder which is based on the operation sequence N-gram model (Durrani et al., 2011). The N-gram-based SMT frame- work (Mari˜no et al., 2006) memorizes Markov chains over sequences of minimal translation units (MTUs or tuples) composed of bilingual transla- tion units. The OSM model integrates reordering operations within the tuple sequences to form a heterogeneous mixture of lexical translation and 1Qatar Computing Research Institute and University of Szeged were partnered for RU-EN and DE-EN language pairs respectively. reordering operations and learns a Markov model over a sequence of operations. Our decoder uses the beam search algorithm in a stack-based decoder like most sequence-based SMT frameworks. Although the model is based on minimal translation units, we use phrases dur- ing search because they improve the search accu- racy of our system. The earlier decoder (Durrani et al., 2011) was based on minimal units. But we recently showed that using phrases during search gives better coverage of translation, better future cost estimation and lesser search errors (Durrani et al., 2013a) than MTU-based decoding. We have therefore shifted to phrase-based search on top of the OSM model. This paper is organized as follows. Section 2 gives a short description of the model and search as used in the OSM decoder. In Section 3 we give a description of the POS-based operation se- quence model that we test for our German-English and English-German experiments. Section 4 de- scribes our processing of the German and English data for German-English and English-German ex- periments. In Section 5 we describe the unsuper- vised transliteration mining that has been done for the Russian-English and English-Russian experi- ments. In Section 6 we describe the sub-sampling technique that we have used for several language pairs. In Section 7 we describe the experimental setup followed by the results. Finally we summa- rize the paper in Section 8. 2 System Description 2.1 Model …

Content truncated to 500 words