This paper tries to quantify the ethical dilemma of using culturally toxic training data to improve the performance of AI tools for ultra low-resource languages such as Indigenous languages. Our case study explores the use of Bible data which is both a commonly available source of training pairs for translators of Indigenous languages and a text which has a trail of physical and cultural violence for many Indigenous communities. In the context of fine-tuning a WMT19 German-to-English model into a Guarani Mbya-to-English translator, we first show, with two commonly-used Machine Translation metrics, that using only Bible data is not enough to create successful translators for everyday sentences gathered from a dictionary. Indeed, even fine-tuning with only 3,000 pairs of data from the dictionary produces significant increases in accuracy compared to Bible-only models. We then show that simultaneously fine-tuning with dictionary and Bible data achieves a substantial increase over the accuracy of a dictionary-only trained translator, and similarly happens when using two-step methods of fine-tuning. However, we also observed some, measurable, contaminated text from the Bible into the outputs of the best translator, creating concerns about its release to an Indigenous community. We end by discussing mechanisms to mitigate the negative impacts of this contamination.

In this paper we study the fine-tuning of pre-trained large high-resource language models (LLMs) into many-to-one multilingual machine translators for extremely-low-resource languages such as endangered Indigenous languages. We explore those issues using datasets created from pseudo-parallel translations to English of The Bible written in 39 Brazilian Indigenous languages using mBART50 and WMT19 as pre-trained models and multiple translation metrics. We examine bilingual and multilingual models and show that, according to machine translation metrics, same-linguistic family models tend to perform best. However, we also found that many-to-one multilingual systems have a tendency to learn a “rogue” strategy of storing output strings from the training data in the LLM structure and retrieving them instead of performing actual translations. We show that rephrasing the output of the training samples seems to solve the problem.