DFKI-LT - Can Out-of-the-box NMT Beat a Domain-trained Moses on Technical Data?

Anne Beyer, Vivien Macketanz, Aljoscha Burchardt, Philip Williams
Can Out-of-the-box NMT Beat a Domain-trained Moses on Technical Data?
2 The 20th Annual Conference of the European Association for Machine Translation, Pages 41-46, Prague, Czech Republic, Charles University, Faculty of Mathematics and Physics, Charles University, Malostranské náměstí 25, 11800 Prague 1, Czech Republic, 2017
In the last year, we have seen a lot of evidence about the superiority of neu- ral machine translation approaches (NMT) over phrase-based statistical approaches (PBMT). This trend has shown for the gen- eral domain at public competitions such as the WMT challenges as well as in the ob- vious quality increase in online translation services that have changed their technol- ogy. In this paper, we take the perspective of an LSP. The questions we want to an- swer with this study is if now is already the time to invest in the new technology. To answer this question, we have collected evidence as to whether an existing state- of-the-art NMT system for the general do- main can already compete with a domain- trained and optimised Moses (PBMT) sys- tem or if it is maybe already better. As it is well known that automatic quality mea- sures are not reliable for comparing the performance of different system types, we have performed a detailed manual evalua- tion based on a test suite of domain seg- ments.
Files: BibTeX, Conference_Booklet_EAMT2017.pdf, Conference_Booklet_EAMT2017.pdf