eTranslation’s Submissions to the WMT 2020 News Translation Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 254–261 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 254 eTranslation’s Submissions to the WMT 2020 News Translation Task Csaba Oravecz† Katina Bontcheva† L´aszl´o Tihanyi† David Kolovratnik† Bhavani Bhaskar† Adrien Lardilleux† Szymon Klocek⋆ Andreas Eisele⋆ DG Translation – DG CNECT, European Commission †firstname.lastname@ext.ec.europa.eu ⋆firstname.lastname@ec.europa.eu Abstract The paper describes the submissions of the eTranslation team to the WMT 2020 news translation shared task. Leveraging the experi- ence from the team’s participation last year we developed systems for 5 language pairs with various strategies. Compared to last year, for some language pairs we dedicated a lot more resources to training, and tried to follow stan- dard best practices to build competitive sys- tems which can achieve good results in the rankings. By using deep and complex archi- tectures we sacrificed direct re-usability of our systems in production environments but eval- uation showed that this approach could result in better models that significantly outperform baseline architectures. We submitted two sys- tems to the zero shot robustness task. These submissions are described briefly in this paper as well. 1 Introduction The European Commission’s eTranslation project1, a building block of the Connecting Europe Facility (CEF), has been set up to help European and na- tional public administrations exchange information across language barriers in the EU. More details about the project can be found in (Oravecz et al., 2019). Our participation in last year’s WMT shared task marked an important step towards opening the service to the coverage of additional, non-EU lan- guages and to domains beyond the formal language of EU institutions. Due to the encouragement and insights we received from WMT 2019, a complete set of general domain MT engines has meanwhile been implemented and incorporated into the eTrans- lation service. This year the team participated in the news trans- lation shared task with five different language pairs: English → German, Japanese → English, 1https://ec.europa.eu/cefdigital/wiki/ display/CEFDIGITAL/eTranslation English → Polish, Russian → English and English → Czech. The varying performance of these sys- tems reflects the amount of resources dedicated to their developments. 2 Data Preparation This section briefly describes the data sets, the se- lection, and filtering methods applied to the pro- vided parallel and monolingual data in order to increase the quality of trained models. We primar- ily focused on constrained submissions, but due to the low quality of our first En→Pl models trained only on the constrained data set we switched to the unconstrained scenario and chose to submit only the unconstrained En→Pl system (see Section 4.3). 2.1 Data Selection and Filtering In general, we made use of all provided original parallel (OP) data to build baseline models for ref- erence or back-translation. Some brief experiments were made with the exclusion of one or the other data set. However, the best baseline models were trained when we used all OP data (except for the UN Parallel Corpus for Ru→En, which, like last year, did not improve the results). This year, where …

Content truncated to 500 words