A Natural Diet: Towards Improving Naturalness of Machine Translation Output

Findings of the Association for Computational Linguistics: ACL 2022, pages 3340 - 3353 May 22-27, 2022 c⃝2022 Association for Computational Linguistics A Natural Diet: Towards Improving Naturalness of Machine Translation Output Markus Freitag, David Vilar, David Grangier, Colin Cherry, George Foster Google Research {freitag, vilar, grangier, colincherry, fosterg}@google.com Abstract Machine translation (MT) evaluation often fo- cuses on accuracy and fluency, without paying much attention to translation style. This means that, even when considered accurate and flu- ent, MT output can still sound less natural than high quality human translations or text originally written in the target language. Ma- chine translation output notably exhibits lower lexical diversity, and employs constructs that mirror those in the source sentence. In this work we propose a method for training MT systems to achieve a more natural style, i.e. mirroring the style of text originally written in the target language. Our method tags par- allel training data according to the naturalness of the target side by contrasting language mod- els trained on natural and translated data. Tag- ging data allows us to put greater emphasis on target sentences originally written in the tar- get language. Automatic metrics show that the resulting models achieve lexical richness on par with human translations, mimicking a style much closer to sentences originally writ- ten in the target language. Furthermore, we find that their output is preferred by human ex- perts when compared to the baseline transla- tions. 1 Introduction Machine translation has made tremendous progress in recent years with the advent of neural methods (Bahdanau et al., 2015; Vaswani et al., 2017). This is especially true for language pairs with a large amount of available bilingual text for training (Bar- rault et al., 2020a). However MT output still can be improved: it currently trails human translators in expert evaluation (Toral et al., 2018; Freitag et al., 2021) and its language is perceived as poorer and more synthetic (Vanmassenhove et al., 2021). In this work, we aim to produce machine translation output that has a more natural style. Although difficult to define precisely, we con- sider a translation to be natural if it is an adequate Source Es wird befürchtet, dass die Opferzahlen noch deutlich in die Höhe gehen. Translationese It is feared that the number of victims will increase significantly. Natural It is feared that the death toll will rise significantly. Figure 1: Example De→En translations: This work sets the goal to generate more natural translations like death toll/rise in comparison to literal translations like number of victims/increase. and fluent translation, whose style matches that of high quality monolingual text. Such a transla- tion should contain few translationese constructs and use a rich vocabulary. This is exemplified in Figure 1. The translationese sentence uses the construct “number of victims”, which is a literal translation for the German “Opferzahlen”. Al- though correct (i.e. adequate and fluent), “death toll” shows a much more natural word choice for this translation. Our objective in this paper is to study how the naturalness of …

Content truncated to 500 words