Directing the violence or admonishing it? A survey of contronymy and androcentrism in Google Translate and some recommendations

Anonymous

Keywords: Google translate, NLP, Machine translation, bias, ethics

TL;DR: Translation biases in Google Translate

Abstract:

The recent raft of high-profile gaffes involving neural machine translation technologies has brought to light the unreliability of this evolving technology. A worrisome
facet of the ubiquity of this technology is that it largely operates in a use-it-at-yourown-peril mode where the user is often unaware of either the idiosyncratic brittleness of the underlying neural translation model or when it is, that the translations
be deemed trustworthy and when they wouldn’t. These revelations have worryingly
coincided with other developments such as the emergence of large language models
that now produce biased and erroneous results, albeit with human-like fluency, the
use of back-translation as a data-augmentation strategy in so termed ’low-resource’
settings and the emergence of ’AI-enhanced legal-tech’ as a panacea that promises
’disruptive democratization’ of access to legal services. In the backdrop of these
quandaries, we present this cautionary tale where we shed light on the specifics
of the risks surrounding cavalier deployment of this technology by exploring two
specific failings: Androcentrism and Enantiosemy. In this regard, we empirically
investigate the fate of the pronouns and a list of contronyms when subjected to
back-translation using Google Translate. Through this, we seek to highlight the
prevalence of ’defaulting-to-the-masculine’ phenomenon in the context of engendered profession-related translations and also empirically demonstrate the scale and
nature of threats pertaining to contronymous phrases covering both current-affairs
and legal issues. Based on these observations, we have collected a series of recommendations that constitute the latter half of this paper. All of the code and datasets
generated in this paper have been open-sourced for the community to build on here:
https://github.com/rteehas/GT_study_recommendations.