How Human is Machine Translationese? Comparing Human and Machine Translations of Text and Speech

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 280–290 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 280 How Human is Machine Translationese? Comparing Human and Machine Translations of Text and Speech Yuri Bizzoni1, Tom S Juzek1, Cristina Espa˜na-Bonet2, Koel Dutta Chowdhury1, Josef van Genabith1,2, and Elke Teich1 1 Department of Language, Science and Technology, Saarland University, Germany 2 DFKI GmbH, Saarbr¨ucken, Germany {yuri.bizzoni,tom.juzek,koel.duttachowdhury}@uni-saarland.de {cristinae,Josef.Van Genabith}@dfki.de,{e.teich}@mx.uni-saarland.de Abstract Translationese is a phenomenon present in human translations, simultaneous interpreting, and even machine translations. Some trans- lationese features tend to appear in simulta- neous interpreting with higher frequency than in human text translation, but the reasons for this are unclear. This study analyzes trans- lationese patterns in translation, interpreting, and machine translation outputs in order to explore possible reasons. In our analysis we (i) detail two non-invasive ways of detecting translationese and (ii) compare translationese across human and machine translations from text and speech. We find that machine trans- lation shows traces of translationese, but does not reproduce the patterns found in human translation, offering support to the hypothesis that such patterns are due to the model (human vs. machine) rather than to the data (written vs. spoken). 1 Introduction In recent years, a growing body of work has pointed to the presence, across different genres and do- mains, of a set of relevant linguistic patterns with respect to syntax, semantics and discourse of hu- man translations. Such patterns make translations more similar to each other than to texts in the same genre and style originally authored in the target language. These patterns are together called “trans- lationese” (Teich, 2003; Volansky et al., 2015). Linguists classify translationese in two main cat- egories: (i) source’s interference, or shining-though as put forward by Teich (2003). For example, a translation replicating a syntactic pattern which is typical of the source language, and rare in the target language, displays a typical form of shining- through; (ii) aherence or over-adherence to the tar- get language’s standards, that is normalisation. For example, translating a sentence displaying marked order in the source with a sentence displaying stan- dard order in the target it a typical example of over- normalization. Nonetheless, translationese’s main causes remain unclear (Koppel and Ordan, 2011; Schmied and Sch¨affler, 1996). Translationese displays different patterns de- pending on the translation’s mode and register: a typical example is simultaneous interpreting, which shows translationese patterns distinct from those observed in written translations (Bernardini et al., 2016). We can interpret differences as either (i) an effect of the limitations of the human language ap- paratus that constrain translations, (ii) an inevitable effect of the structural and semantic differences be- tween languages; or (iii) a combination of the two. To test these hypotheses, it is common to compare human translations (HT) produced under different circumstances, e.g. written translation versus simul- taneous interpreting, following the assumption that, if translationese is a product of human cognitive limitations, translations produced under higher cog- nitive constrains should present more evident trans- …

Content truncated to 500 words