eTranslation’s Submissions to the WMT22 General Machine Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 346 - 351 December 7-8, 2022 ©2022 Association for Computational Linguistics eTranslation’s Submissions to the WMT22 General Machine Translation Task Csaba Oravecz⋆ Katina Bontcheva† David Kolovratník† Bogomil Kovachev⋆ Christopher Scott⋆ DG Translation – DG CNECT, European Commission ⋆firstname.lastname@ec.europa.eu †firstname.lastname@ext.ec.europa.eu Abstract The paper describes the 3 NMT models sub- mitted by the eTranslation team to the WMT22 general machine translation shared task. In the WMT news task last year, multilingual sys- tems with deep and complex architectures uti- lizing immense amounts of data and resources were dominant. This year with the task ex- tended to cover less domain specific text we expected even more dominance of such sys- tems. In the hope to produce competitive (con- strained) systems despite our limited resources, this time we selected only medium resource language pairs, which are serviced in the Euro- pean Commission’s eTranslation system. We took the approach of exploring less resource intensive strategies focusing on data selection and data filtering to improve the performance of baseline systems. With our submitted systems our approach scored competitively according to the automatic rankings in the constrained category, except for the En→Ru model where our submission was only a baseline reference model developed as a by-product of the multi- lingual setup we built focusing primarily on the En→Uk language pair. 1 Introduction The eTranslation team is responsible for the de- velopment of machine translation systems pro- viding the translation services of the European Commission’s eTranslation project1. This is a building block of the Connecting Europe Facil- ity (CEF), with the aim of supporting 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). During the previous years the team’s participa- tion in the WMT shared tasks allowed us to explore state-of-the-art methods to develop high quality ma- chine translation systems. However, due to strict 1https://language-tools.ec.europa.eu resource constraints, these systems do not normally carry over to production environments and there has been a continuous search for the right balance between the use of resources in production environ- ments and the best performing but more complex architectures. With the news translation shared task extended to being a general MT task the need for more robust- ness, coverage and consequently more complexity and resources has further increased. We expect a strong competition in these areas, where teams with modest resources might have some inherent disadvantages. Therefore, in this year’s experi- ments we did not consider high resource language pairs (specifically English → German, our constant submission in previous years) and opted for the medium resource French → German and English → Ukrainian language directions. The latter sys- tem originated from a multilingual setup including Russian data, so we built and submitted a baseline English → Russian model as well. 2 Data Preparation In this section we briefly describe the base data sets, the general selection and filtering methods we ap- plied to prepare these initial …

Content truncated to 500 words