Selecting Data for English-to-Czech Machine Translation

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 374–381, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Selecting Data for English-to-Czech Machine Translation ∗ Aleˇs Tamchyna, Petra Galuˇsˇc´akov´a, Amir Kamran, Miloˇs Stanojevi´c, Ondˇrej Bojar Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´am. 25, Praha 1, CZ-118 00, Czech Republic {tamchyna,galuscakova,kamran,bojar}@ufal.mff.cuni.cz, milosh.stanojevic@gmail.com Abstract We provide a few insights on data selection for machine translation. We evaluate the quality of the new CzEng 1.0, a parallel data source used in WMT12. We describe a simple tech- nique for reducing out-of-vocabulary rate af- ter phrase extraction. We discuss the bene- fits of tuning towards multiple reference trans- lations for English-Czech language pair. We introduce a novel approach to data selection by full-text indexing and search: we select sentences similar to the test set from a large monolingual corpus and explore several op- tions of incorporating them in a machine trans- lation system. We show that this method can improve translation quality. Finally, we de- scribe our submitted system CU-TAMCH-BOJ. 1 Introduction Selecting suitable data is important in all stages of creating an SMT system. For training, the data size plays an essential role, but the data should also be as clean as possible. The new CzEng 1.0 was prepared with the emphasis on data quality and we evaluate it against the previous version to show whether the effect for MT is positive. Out-of-vocabulary rate is another problem related to data selection. We present a simple technique to reduce it by including words that became spurious OOVs during phrase extraction. ∗ This work was supported by the project EuroMatrixPlus (FP7-ICT-2007-3-231720 of the EU and 7E09003+7E11051 of the Czech Republic) and the Czech Science Foundation grants P406/11/1499 and P406/10/P259. Another topic we explore is to use multiple refer- ences for tuning to make the procedure more robust as suggested by Dyer et al. (2011). We evaluate this approach for translating from English into Czech. The main focus of our paper however lies in pre- senting a method for data selection using full-text search. We index a large monolingual corpus and then extract sentences from it that are similar to the input sentences. We use these sentences in several ways: to create a new language model, a new phrase table and a tuning set. The method can be seen as a kind of domain adaptation. We show that it con- tributes positively to translation quality and we pro- vide a thorough evaluation. 2 Data and Tools 2.1 Comparison of CzEng 1.0 and 0.9 As this year’s WMT is the first to include the new version of CzEng (Bojar et al., 2012b), we carried out a few experiments to compare its suitability for MT with its predecessor, CzEng 0.9. Apart from size (which has almost doubled), there are impor- tant differences between the two versions. In CzEng 0.9, the largest portion by far came from movie sub- titles (a data source of varying quality), …

Content truncated to 500 words