CUNI Systems for the WMT 22 Czech-Ukrainian Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 352 - 357 December 7-8, 2022 ©2022 Association for Computational Linguistics CUNI Systems for the WMT 22 Czech-Ukrainian Translation Task Martin Popel Jindˇrich Libovický∗ Jindˇrich Helcl∗ Charles University, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostranské námˇestí 25, 118 00 Prague, Czech Republic {popel,libovicky,helcl}@ufal.mff.cuni.cz Abstract We present Charles University submissions to the WMT 22 General Translation Shared Task on Czech-Ukrainian and Ukrainian-Czech ma- chine translation. We present two constrained submissions based on block back-translation and tagged back-translation and experiment with rule-based romanization of Ukrainian. Our results show that the romanization only has a minor effect on the translation quality. Further, we describe Charles Translator, a sys- tem that was developed in March 2022 as a response to the migration from Ukraine to the Czech Republic. Compared to our constrained systems, it did not use the romanization and used some proprietary data sources. 1 Introduction How fast can the machine translation (MT) com- munity react to a sudden need of a high-quality MT system which was previously under low demand? This question motivated the new task at the WMT this year, which is Czech-Ukrainian translation. Both languages belong to the Slavic language family (Czech is western Slavic, Ukrainian is east- ern Slavic), and share some lexical and structural characteristics. Unlike Czech, which uses the Latin script, Ukrainian uses its variant of the Cyrillic alphabet. We submit three systems to the WMT 22 General Translation Shared Task for this lan- guage pair in each translation direction. The first system, CUNI-JL-JH, implemented in Mar- ian (Junczys-Dowmunt et al., 2018), uses tagged back-translation and is a result of our experiments with romanization of Ukrainian. Our second sys- tem, CUNI-TRANSFORMER, implemented in Ten- sor2Tensor (Vaswani et al., 2018), uses block back- translation. Finally, we submit an unconstrained system, CHARLES TRANSLATOR, implemented in Tensor2Tensor, which has been developed in ∗The author order was determined by a coin toss. spring 2022 as a response to the crisis caused by the Russian invasion of Ukraine and the following migration wave. 2 Constrained WMT Submissions We submitted two systems in each translation di- rection that use the same parallel and monolingual data, but different techniques and different toolk- its. This section first describes the shared data processing steps and then the specifics of each of the submissions in separate subsections. 2.1 Training Data We use all parallel data allowed in the constrained task, along with 50 million Czech and 58 million Ukrainian sentences of monolingual data. In the following paragraphs we describe the data cleaning steps when preparing the training data. We fur- ther experiment with romanization of the Ukrainian Cyrillic alphabet and with artificial noising of the data. Parallel data. The data for the constrained trans- lation task consist of OPUS corpora (Tiedemann, 2012) that have a Czech-Ukrainian part, WikiMa- trix (Schwenk et al., 2021) and the ELRC EU acts in Ukrainian.1 We clean the parallel data using rule-based filter- ing in the following …

Content truncated to 500 words