Unbabel announced as the world leader in translation quality at WMT19

Time to read
1 minute
Read so far

Unbabel announced as the world leader in translation quality at WMT19

Tue, 05/21/2019 - 09:23
Posted in:
0 comments

·         Unbabel regains title of best global Machine Translation Quality Estimation system at annual event

·         Won competition in 2016

·         Beats competition including Carnegie Mellon University

London, UK; Lisbon, Portugal. 21st May 2019. Unbabel, the leading enterprise multilingual support provider, has won the ‘shared task’ event at a renowned machine translation conference to re-establish itself as the premier global translation-as-a-service company.

WMT19 is the annual competition where Quality Estimation (QE) systems from several participants, both academic and industry-side, are evaluated and compared on the same data, to establish which is the most accurate at assessing automatically machine translation quality. After winning the competition in 2016, Unbabel achieved the best results across all ‘language pairs’ (English-German, English-Russian, and English-French) and all tracks (word-level, sentence-level, and document-level) this year. Other competitors included Carnegie Mellon University, Nanjing University, the Electronics and Telecommunications Research Institute, and the Moscow Institute for Physics and Technology, among others. Unbabel’s scores this year on English-German outperformed the best system from last year (from Alibaba) on sentence-level by 7 points in Pearson correlation and on word-level by 4 points in the official F1 metric.

This year's WMT shared task on quality estimation consisted of estimating post-editing effort on word and sentence level and also performing diagnostics at a wider document level. A common issue with Machine Translation is the gap in quality between sentences translated word-by-word and how that equates to an entire document.

Unbabel, which recently open-sourced its Quality Estimation software, OpenKiwi, entered a system combining the best of Transformer-based neural machine translation models with transfer learning and ensemble techniques.

João Graça, CTO of Unbabel, commented: “This is great news for Unbabel, and I’m extremely proud of the team who helped us regain our title at WMT19. It demonstrates, and is a good reward for, our continuing work in the fields of quality estimation and machine translation. Quality Estimation is of the highest importance in a world where both individuals and businesses are coming to rely on machine translation to help them in their day to day lives, and where consumers expect and deserve to be accurately served in their native language by organisations.”

Official results of the competition can be found here: http://www.statmt.org/wmt19/qe-results.html.