Effect of Translationese on Machine Translation Performance

Koel Dutta Chowdhury; Cristina España-Bonet; Josef van Genabith

SFB Poster Session, 7/2019.


Translationese refers to the presence of unusual characteristics in translated texts. These characteristics can occur as a result of translation as a communicative process itself, or from the influence of the source language on the target language. Consequently, translated texts are affected by the fingerprint of the source text itself –in what is called "the law of interference". Previous studies using statistical machine translation systems have shown that such translationese signals can be used to improve Machine Translation (MT) performance. We conducted similar experiments using seq2seq neural MT to compare the translation performance between models traineon translated text and models trained on original-language text. Our results show that translationese-based models out-perform original-language models, regardless of the language.

SFB_translationese.pdf (pdf, 642 KB )

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence