Lost in Translation: Loss and Decay of Linguistic Richness in Machine Translation

Lost in Translation: Loss and Decay of Linguistic Richness in Machine Translation Eva Vanmassenhove Dimitar Shterionov ADAPT, School of Computing, Dublin City University, Dublin, Ireland firstname.lastname@adaptcentre.ie Andy Way Abstract This work presents an empirical approach to quantifying the loss of lexical rich- ness in Machine Translation (MT) systems compared to Human Translation (HT). Our experiments show how current MT sys- tems indeed fail to render the lexical di- versity of human generated or translated text. The inability of MT systems to gen- erate diverse outputs and its tendency to exacerbate already frequent patterns while ignoring less frequent ones, might be the underlying cause for, among others, the currently heavily debated issues related to gender biased output. Can we indeed, aside from biased data, talk about an algo- rithm that exacerbates seen biases? 1 Introduction Berman (2000) observed that the translation pro- cess consists of deformation processes, one of which he refers to as ‘quantitative impoverish- ment’, a loss of lexical richness and diversity. Al- though mitigated by a human translator, this loss is to some extent inevitable as it is hard to re- spect the multitude of signifiers and constructions when translating one language into another. While Berman (2000) studied the decrease of lexical rich- ness of human translations (HTs) from a theoret- ical point of view, Kruger (2012) demonstrated using empirical methods that there is indeed a lex- ical loss when comparing translations to original texts. In the field of Machine Translation (MT), Klebanov and Flor (2013) showed that Statistical c⃝ 2019 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. MT (SMT) suffers considerably more from lexical loss than HTs in a study focused on lexical tight- ness and text cohesion. We are not aware of any other research in this direction. As generating accurate translations has been the main objective of current MT systems, maintaining lexical richness and creating diverse outputs has understandably not been a priority. Nevertheless, the issue of lexical loss in MT might at the same time be a symptom and a cause of a more serious issue underlying the current systems. The differ- ence between a one-to-many relationship such as the one illustrated in Figure 1, is very different from the one illustrated in Figure 2 or Figure 3 from a (human) translator point of view. How- ever, from a statistical point of view, they are not always clearly distinguishable. When presented with an ambiguous sentence, like ‘I am intelli- gent’ or ‘See?’ where there is little context to decide on a particular target variant of the same source word, it essentially boils down to the same thing: picking the translation that maximizes the probability over the entire sentence. As such, the loss of richness and diversity and the exacerba- tion of already frequent patterns might not sim- ply be limited to the loss of (near) synonyms or rare words, but could also be the underlying cause of, for example, the inability of statistical …

Content truncated to 500 words