QCRI-MES Submission at WMT13: Using Transliteration Mining to Improve Statistical Machine Translation

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 219–224, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics QCRI-MES Submission at WMT13: Using Transliteration Mining to Improve Statistical Machine Translation Hassan Sajjad1, Svetlana Smekalova2, Nadir Durrani3, Alexander Fraser4, Helmut Schmid4 1Qatar Computing Research Institute – hsajjad@qf.org.qa 2University of Stuttgart – smekalsa@ims.uni-stuttgart.de 3University of Edinburgh – dnadir@inf.ed.ac.uk 4Ludwig-Maximilians University Munich – (fraser|schmid)@cis.uni-muenchen.de Abstract This paper describes QCRI-MES’s sub- mission on the English-Russian dataset to the Eighth Workshop on Statistical Ma- chine Translation. We generate improved word alignment of the training data by incorporating an unsupervised translitera- tion mining module to GIZA++ and build a phrase-based machine translation sys- tem. For tuning, we use a variation of PRO which provides better weights by optimiz- ing BLEU+1 at corpus-level. We translit- erate out-of-vocabulary words in a post- processing step by using a transliteration system built on the transliteration pairs extracted using an unsupervised translit- eration mining system. For the Russian to English translation direction, we apply linguistically motivated pre-processing on the Russian side of the data. 1 Introduction We describe the QCRI-Munich-Edinburgh- Stuttgart (QCRI-MES) English to Russian and Russian to English systems submitted to the Eighth Workshop on Statistical Machine Trans- lation. We experimented using the standard Phrase-based Statistical Machine Translation System (PSMT) as implemented in the Moses toolkit (Koehn et al., 2007). The typical pipeline for translation involves word alignment using GIZA++ (Och and Ney, 2003), phrase extraction, tuning and phrase-based decoding. Our system is different from standard PSMT in three ways: • We integrate an unsupervised transliteration mining system (Sajjad et al., 2012) into the GIZA++ word aligner (Sajjad et al., 2011). So, the selection of a word pair as a correct alignment is decided using both translation probabilities and transliteration probabilities. • The MT system fails when translating out-of- vocabulary (OOV) words. We build a statis- tical transliteration system on the translitera- tion pairs mined by the unsupervised translit- eration mining system and transliterate them in a post-processing step. • We use a variation of Pairwise Ranking Op- timization (PRO) for tuning. It optimizes BLEU at corpus-level and provides better feature weights that leads to an improvement in translation quality (Nakov et al., 2012). We participate in English to Russian and Rus- sian to English translation tasks. For the Rus- sian/English system, we present experiments with two variations of the parallel corpus. One set of experiments are conducted using the standard par- allel corpus provided by the workshop. In the sec- ond set of experiments, we morphologically re- duce Russian words based on their fine-grained POS tags and map them to their root form. We do this on the Russian side of the parallel corpus, tuning set, development set and test set. This im- proves word alignment and learns better transla- tion probabilities by reducing the vocabulary size. The paper is organized as follows. Section 2 talks about unsupervised transliteration mining and its incorporation to the GIZA++ word aligner. In Section 3, we describe the transliteration sys- …

Content truncated to 500 words