A Shocking Amount of the Web is Machine Translated: Insights from Multi-Way Parallelism

We show that content on the web is often translated into many languages, andthe low quality of these multi-way translations indicates they were likelycreated using Machine Translation (MT). Multi-way parallel, machine generatedcontent not only dominates the translations in lower resource languages; italso constitutes a large fraction of the total web content in those languages.We also find evidence of a selection bias in the type of content which istranslated into many languages, consistent with low quality English contentbeing translated en masse into many lower resource languages, via MT. Our workraises serious concerns about training models such as multilingual largelanguage models on both monolingual and bilingual data scraped from the web.