Lost in Back-Translation: Emotion Preservation in Neural Machine Translation

Proceedings of the 28th International Conference on Computational Linguistics, pages 4340–4354 Barcelona, Spain (Online), December 8-13, 2020 4340 Lost in Back-Translation: Emotion Preservation in Neural Machine Translation Enrica Troiano, Roman Klinger, and Sebastian Pad´o Institut f¨ur Maschinelle Sprachverarbeitung, University of Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {firstname.lastname}@ims.uni-stuttgart.de Abstract Machine translation provides powerful methods to convert text between languages, and is therefore a technology enabling a multilingual world. An important part of communication, however, takes place at the non-propositional level (e.g., politeness, formality, emotions), and it is far from clear whether current MT methods properly translate this information. This paper investigates the specific hypothesis that the non-propositional level of emotions is at least partially lost in MT. We carry out a number of experiments in a back-translation setup and establish that (1) emotions are indeed partially lost during translation; (2) this tendency can be reversed almost completely with a simple re-ranking approach informed by an emotion classifier, taking advantage of diversity in the n-best list; (3) the re-ranking approach can also be applied to change emotions, obtaining a model for emotion style transfer. An in-depth qualitative analysis reveals that there are recurring linguistic changes through which emotions are toned down or amplified, such as change of modality. 1 Introduction The quality of machine translation (MT) models in some areas follows close behind that of humans (Barrault et al., 2019). MT is deployed widely to support human-to-human communication across languages, e.g., in chat systems, customer support, or (social) media. It is also employed in downstream NLP tasks such as sentence simplification (Xu et al., 2016), error correction (Yuan and Briscoe, 2016), paraphrasing (Mallinson et al., 2017; Wieting and Gimpel, 2018), or cross-lingual resource creation (Barnes and Klinger, 2019). With the increasing use of MT, however, expectations about output quality also grow, and now that the goal of adequacy with regard to propositional content is met more often than not, more subtle aspects start receiving attention. One such aspect is affective content. Establishing common ground is essential for successful MT-assisted communication (Yamashita et al., 2009), but it is still unclear how well MT promotes this, especially when handling the affective qualities of texts. On the one hand, it is able to mostly preserve author sentiment (Balahur and Turchi, 2012). On the other, it is known that translation obfuscates some socio-demographic characteristics of authors, like gender and personality traits (Mirkin et al., 2015; Rabinovich et al., 2017). In this paper, we investigate the question of how well emotions are preserved in MT. Answering this question and, if necessary, increasing the degree of emotion preservation, is important both theoretically (to inform cross-lingual studies that use translation as part of their experimental setup) and practically (to improve the usefulness of MT). The starting point of our research is a study by Rabinovich et al. (2017), who show that some semantic nuances tend to vanish in translation. In fact, just like human translation, MT is not guaranteed to preserve any of the linguistic properties of input texts …

Content truncated to 500 words