Neural Translation for the European Union (NTEU) Project

Neural Translation for the European Union (NTEU) Project L. Bi´e*, A. Cerd`a-i-Cuc´o*, H. Degroote*, A. Estela*, M. Garc´ıa-Mart´ınez*, M. Herranz*, A. Kohan*, M. Melero‡, T. O’Dowd§, S. O’Gorman§, M. Pinnis†, R. Rozis†, R. Superbo§, A. Vasil¸evskis†∗ ‡Barcelona Supercomputing Center maite.melero@bsc.es §KantanMT{sineadog,tonyod,riccardos}@kantanmt.com *Pangeanic-PangeaMT{l.bie,a.cerda,h.degroote,a.estela,m.garcia,m.herranz,a.kohan}@pangeanic.com †Tilde{roberts.rozis,marcis.pinnis,arturs.vasilevskis}@tilde.lv Abstract The Neural Translation for the European Union (NTEU) project aims to build a neu- ral engine farm with all European offi- cial language combinations for eTransla- tion,1 without the necessity to use a high- resourced language as a pivot. NTEU started in September 2019 and will run un- til August 2021. 1 Introduction Normally, data for translation are available in En- glish from or to another language. With a few ex- ceptions, all eTranslation MT engines include En- glish as either source or target. Thus, to translate between two non-English languages, English must be used as a pivot. The NTEU partners, Pangeanic,2 Tilde,3 Kan- tanMT4 and SEAD5, have been awarded EU funds to build direct machine translation (MT) engines between any of the 24 EU official languages (e.g. Spanish to German, Croatian to Italian, Greek to Polish, etc.) without pivoting through English (around 550 translation engines in total). 2 Approach NTEU will provide a capacity service to eTrans- lation by building a near-human-professional- quality neural engine farm which includes all EU ∗All authors have contributed equally to this work. ∗c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1https://ec.europa.eu/info/resources-partners/machine- translation-public-administrations-etranslation en 2https://pangeamt.com/ 3https://www.tilde.com/ 4https://kantanmt.com/ 5Spanish Agency for Digital Advancement, dependant on the Ministry of Economy language combinations. State-of-the-art technolo- gies such as the transformer (Vaswani et al., 2017) architecture will be implemented. Moreover, lower-resourced languages (for example, Irish or Maltese) will be a challenge, and more effort will be required to obtain well-performing engines for them. In order to obtain the best results, we will experiment with techniques to supplement the original data, such as generating synthetic data by doing back-translation (Sennrich et al., 2016), checking of sentence alignments, transfer learn- ing (Zoph et al., 2016) and unsupervised learning on a monolingual corpus (Artetxe et al., 2019). In addition to providing the trained engines, the NTEU consortium will gather and clean data from all language combinations so that the engines can be retrained with other technologies in the future. As part of the national digital data gathering ef- forts, NTEU will also act as a bridge between pre- vious efforts, putting to work the results of the ELRC6 repository and other European data gather- ing efforts such as the NEC TM7 and ParaCrawl8 projects. Therefore, the project will promote the free flow of data between public administrations themselves and EU bodies. 3 Evaluation methodology The results of the MT will be manually evalu- ated in an open-source platform created by the consortium. Following industry and WMT9 prac- tices for human evaluation, the evaluation dataset has been carefully chosen so as to represent real- world, whole, human-translated documents, pur- posely excluded from the training data. Human …

Content truncated to 500 words