What the evidence says
A 2026 benchmark built specifically to test discourse in Indian-language translation — eight languages, with human-annotated alignments for pronouns and lexical cohesion — reached a conclusion worth quoting: models achieve high overall translation quality and still fail to preserve discourse-level phenomena.
That sentence describes the exact failure that sentence-level scoring cannot see. Every segment passes. The document does not hold together.
Why it happens
Three things at once. The source languages are morphologically rich, so a single word carries information English spreads across several. Syntactic structures diverge sharply from English word order. And pronouns frequently do not mark gender, so resolving who a sentence is about requires inference across sentences — which a system translating segment by segment cannot do, because it never sees them together.
Add the data problem. These are low-resource languages in the technical sense — the measure is how much digitised text exists to train on, not how many people speak them. Hindi has around 344 million native speakers and is low-resource. The international evaluation that added Bhojpuri in 2025 had to fall back to a simpler metric, because the organisers judged the reliability of the standard metrics unestablished for it. In 2026, a dedicated low-resource task pairs Arabic with Hindi, Bengali, Indonesian and Urdu specifically because these pairs continue to face limited availability of parallel training data.
What it means for your workflow
Segment-level quality estimation will pass Indic content that is internally incoherent — a referent that changes gender halfway through, a key term rendered three different ways in three sections. The higher your automation, the more of this you ship.
What we do
We review at document level, not segment level: pronoun and referent tracking across the file, and terminology cohesion end to end. For long documents, that is a different check from the one your pipeline is running, and it is the one that catches what your pipeline is built to miss.