In November 2020, the Fifth Conference on Machine Translation (WMT20) took place. Collocated with EMNLP 2020, the event was held online and hosted a number of common translation tasks on various aspects of Machine Translation. Every year, the biggest players in the industry get together at WMT to complete a number of shared  translation tasks and to discuss their proceedings. In the end, the results are compared and evaluated, allowing the participants to share their know-how and best practices, but also to assess where their system stands in the field of MT and computational linguistics.

This year, the eTranslation team took part in the news translation shared task1with five different languages pairs: English to German, Czech and Polish on the one hand and Japanese and Russian to English on the other hand. The inclusion of two non-EU languages clearly outlined eTranslation’s mission to extend its language coverage – and the results showed that the European Commission’s Machine Translation system is on a highly promising way:

  • eTranslation reached top positions when ranked by automatic BLEU score.
  • When ranked by humans, eTranslation was typically in the top-ranked cluster (group of engines only separated by expected margins of error).
  • Ranked 2nd for translations into Polish.
  • For Russian to English, eTranslation gave better results than the human translation that was used for comparison.
  • Ranked 3rd for English to German in the test for robustness of MT systems.

These results show that eTranslation can keep up with the best in the field in addition to providing the team with useful insights to further build on this success. The corresponding article on the CEF Digital Portal2 summarises nicely the key takeaway messages: “While parity with human translations has not been reached, the steady improvement in the field means that Machine Translation is having an ever greater impact on the field of translation, both as a tool for human translators and as an alternative when certified accuracy and perfect style are not required”.

The next WMT will come along with a great opportunity and an exciting challenge for eTranslation and the whole research world: ELRC and LT-Bridge will jointly organise a shared task on Multilingual Cultural Heritage Translation covering two families of European languages: Romance and Scandinavian languages. Stay tuned and follow us via Twitter, LinkedIn and Facebook!

Further information about eTranslation is available here: https://ec.europa.eu/cefdigital/wiki/display/CEFDIGITAL/eTranslation.