Machine Translationese: Effects of Algorithmic Bias on Linguistic Complexity in Machine Translation

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2203–2213 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2203 Machine Translationese: Effects of Algorithmic Bias on Linguistic Complexity in Machine Translation Eva Vanmassenhoveα Dimitar Shterionovα α Cognitive Science and AI, Tilburg University, The Netherlands β University of Maryland, College Park {e.o.j.vanmassenhove, d.shterionov}@tilburguniversity.edu mgwilliam@umd.edu Matthew Gwilliamβ Abstract Recent studies in the field of Machine Trans- lation (MT) and Natural Language Processing (NLP) have shown that existing models am- plify biases observed in the training data. The amplification of biases in language technol- ogy has mainly been examined with respect to specific phenomena, such as gender bias. In this work, we go beyond the study of gen- der in MT and investigate how bias amplifica- tion might affect language in a broader sense. We hypothesize that the ‘algorithmic bias’, i.e. an exacerbation of frequently observed pat- terns in combination with a loss of less fre- quent ones, not only exacerbates societal bi- ases present in current datasets but could also lead to an artificially impoverished language: ‘machine translationese’. We assess the lin- guistic richness (on a lexical and morpholog- ical level) of translations created by different data-driven MT paradigms – phrase-based sta- tistical (PB-SMT) and neural MT (NMT). Our experiments show that there is a loss of lexical and morphological richness in the translations produced by all investigated MT paradigms for two language pairs (EN↔FR and EN↔ES). 1 Introduction The idea of translation entailing a transformation is widely recognised in the field of Translation Stud- ies (TS) (Ippolito, 2014). Translations are spe- cific communicative acts occurring in a particular context governed by their own laws. Some of the features that characterize translated texts are de- fined as simplification, explicitation, normalization and leveling out (Baker, 1999). The fingerprints left by the translation process and the language this results into, have been referred to as ‘transla- tionese’ (Gellerstam, 1986). Empirical evidence of the existence of translationese can be found in studies showing that machine learning techniques can be employed to automatically distinguish be- tween human translated and original text by look- ing at lexical and grammatical information (Baroni and Bernardini, 2006; Koppel and Ordan, 2011). Translationese differs from original texts due to a combination of factors including intentional (e.g. explicitation and normalization) and unintentional ones (e.g. unconscious effects of the source lan- guage input on the target language produced). Un- like other work on (human) translationese (or even related work on ‘Post-editese’), we delve into the effects of machine translation (MT) algorithms on language, i.e. ‘machine translationese’. So far, generating accurate and fluent transla- tions has been the main objective of MT systems. As such, maintaining the richness and diversity in the outputs has understandably not been a pri- ority (Vanmassenhove, 2020).1 However, as MT systems have reached a quality that is (arguably) close to that of human translations (L¨aubli et al., 2018; Toral et al., 2018) and as such are being used widely on …

Content truncated to 500 words