eTranslation’s Submissions to the WMT 2021 News Translation Task

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 172–179 November 10–11, 2021. ©2021 Association for Computational Linguistics 172 eTranslation’s Submissions to the WMT 2021 News Translation Task Csaba Oravecz† Katina Bontcheva† David Kolovratník† Bhavani Bhaskar† Michael Jellinghaus⋆ Andreas Eisele⋆ DG Translation – DG CNECT, European Commission †firstname.lastname@ext.ec.europa.eu ⋆firstname.lastname@ec.europa.eu Abstract The paper describes the 3 NMT models sub- mitted by the eTranslation team to the WMT 2021 news translation shared task. We de- veloped systems in language pairs that are actively used in the European Commission’s eTranslation service. In the WMT news task, recent years have seen a steady increase in the need for computational resources to train deep and complex architectures to produce compet- itive systems. We took a different approach and explored alternative strategies focusing on data selection and filtering to improve the per- formance of baseline systems. In the domain constrained task for the French–German lan- guage pair our approach resulted in the best system by a significant margin in BLEU. For the other two systems (English–German and English-Czech1) we tried to build competitive models using standard best practices. 1 Introduction The eTranslation team is behind the translation ser- vices of the European Commission’s eTranslation project2. This is a building block of the Connecting Europe Facility (CEF), with the aim of support- ing European and national public administrations’ information exchange across language barriers in the EU. The project is described in more details in (Oravecz et al., 2019). The team’s participation in the WMT shared tasks has provided valuable insights to improve the quality of our production systems and allowed us to explore languages and domains beyond the formal language of EU institutions, leading to a continuous extension of the eTranslation service and helping in the search for the right balance between the use of resources in production environments and the best possible performance of models. 1Due to returning problems of resource availability, the En→Cs experiments did not finish until the submission dead- line so we could finally only submit last year’s system. 2https://ec.europa.eu/cefdigital/wiki/ display/CEFDIGITAL/eTranslation This year the team participated in the news trans- lation shared task with 3 different language pairs: English → German, English → Czech and French → German. The selection was motivated by the fact that these language pairs can all be consid- ered as high or medium resource, which is the main scenario in the eTranslation service, while the constrained domain in Fr→De offered a good opportunity to focus on and experiment with data selection and filtering techniques, which is a more viable alternative in our environment than the re- source demanding (brute-force) increase in model complexity. 2 Data Preparation Here we briefly describe the base data sets, the gen- eral selection and filtering methods we applied to prepare these initial data sets used to train the first models. Further data selection and augmentation methods to improve the quality of baseline models are described in Section 3.2. For all models we only used the provided parallel and monolingual data, so our …

Content truncated to 500 words