Why Machine Translation Into Low-Resource Languages Still Fails Often

The characteristic failure of low-resource machine translation is not garbled output. It is smooth, grammatical, confident output that says something the source did not say. That specific shape has a specific cause, and the cause is usually the intermediate language.

Why there is a pivot at all

Translation quality is driven by parallel text: pairs of sentences meaning the same thing in two languages. For most language pairs on earth, no such text exists in useful quantity. There is no meaningful Yoruba–Nepali parallel corpus, because nobody has systematically translated between them.

What does exist is English paired with almost everything, because English is what international organisations, software localisation and the translation industry pivot through already. So a system with no direct Yoruba–Nepali data does the available thing: Yoruba to English, English to Nepali. Meta AI’s No Language Left Behind work is explicitly organised around escaping this — training direct pairs across 200 languages rather than routing through a bridge — which is a fair indication of how much damage the bridge does (NLLB, 2022).

The same structure appears when you prompt a general model. Ask for Yoruba to Nepali and you may get no visible intermediate, but the model’s competence in both languages is largely mediated by its English-centric representations, and the failure modes look the same.

What English cannot carry

The pivot is lossy in a precise sense: it is not that meaning evaporates, it is that grammatically obligatory distinctions in the target are optional or absent in English, so after the first leg the information needed for the second leg is simply gone. The second model must supply it, and it supplies the statistically common option.

• Clusivity. Quechua, Tagalog, Cherokee and many others distinguish a “we” including the listener from one excluding them. English has one word. After the pivot, the information is unrecoverable and the target must pick one; it will pick wrong roughly as often as the prior is wrong.
• Evidentiality. Quechua and Turkish, among others, grammatically mark whether the speaker witnessed something, inferred it, or heard it reported. English marks this with optional adverbs that translation routinely drops.
• Honorifics and register. Japanese, Korean, Javanese and many others require a politeness level on every verb. English carries register in word choice, which does not survive.
• Noun class and gender agreement. A Bantu target needs the noun class of every referent to generate concord at all — the machinery described in the Zulu page. English does not mark it, so the class must be re-inferred from an English noun that no longer carries it.
• Tense and aspect granularity. Languages with remoteness distinctions in the past, or obligatory aspect marking, need a distinction English collapses.

Notice the asymmetry. Going into English is comparatively safe — the distinctions are dropped, which loses nuance but rarely asserts something false. Going out of English is where the damage happens, because the target grammar forces a choice that the input no longer determines. Translation into low-resource languages fails harder than translation out of them, and this is why.

How the error compounds

Put a number on it, with the assumption stated. Suppose each leg handles a given phenomenon correctly 85 per cent of the time — a generous figure for a low-resource pair:

Assumption: the two legs' errors are independent.
(This is the load-bearing assumption and it is optimistic;
 in practice a bad first leg makes the second leg worse.)

  P(correct through both) = 0.85 x 0.85 = 0.7225

So a phenomenon that survives 85% of single-leg translations
survives about 72% of pivoted ones. Across a paragraph with
six such phenomena, at independence:

  0.7225 ^ 6 = 0.142

...roughly a one-in-seven chance that a whole paragraph comes
through with every phenomenon intact.

The arithmetic is elementary and it is not the point; the point is the shape. Per-sentence quality metrics look tolerable while document-level correctness collapses, because the phenomena multiply. That is exactly the discrepancy people report when a translation scores acceptably on a benchmark and is unusable in production.

The independence assumption also flatters the pipeline. Errors are correlated: an ambiguous first leg produces an English sentence that is harder to translate, and the second model is most likely to err on precisely the sentences the first model handled worst.

The failure shapes you will actually see

• Fluent and wrong. The output reads well and inverts a relationship, changes a quantity, or picks the wrong participant. This is the dominant shape and the reason review has to be done by someone who reads the source.
• Register collapse. A formal source arrives casual, or the reverse, because the pivot did not carry register and the target defaulted.
• Hallucinated output on degenerate input. Empty strings, very short inputs, numbers alone and repeated tokens regularly produce a fully-formed unrelated sentence in the target, because such inputs are rare in the training data and the model falls back to its prior. Filter degenerate inputs before they reach the model rather than after.
• Silent copy-through. Untranslated source, or English, appearing in the target output — often for named entities and technical terms, where copying is sometimes correct and sometimes not, and nothing distinguishes the two cases downstream.
• Dialect flattening. Output in the standard or most-represented variety regardless of the source variety, which reads as foreign to speakers of any other variety.

What reduces it

None of this is fully fixable without data, but several things move the number, in rough order of effect:

1. Avoid the pivot where a direct model exists. A model trained on direct pairs for your language pair beats a better-known general model routed through English. This is the single largest lever available.
2. If you must pivot, choose the intermediate deliberately. A pivot that marks what the target marks preserves the information the target needs. English is the default and rarely the right one.
3. Supply the lost information explicitly. If your domain always addresses the user formally, or always speaks about a group excluding the listener, put that in the instruction so the target-side choice is not a guess.
4. Prompt natively rather than translating, where you can. Generating directly in the target from structured source data skips one leg entirely — see translate first or prompt natively.
5. Round-trip as a smoke test, not as a metric. A bad round trip reliably indicates a problem; a good one does not indicate correctness, because both legs can share the same misunderstanding.
6. Score at document level on a fixed set, with the phenomena you care about listed as separate checks, so the compounding above is visible in your numbers instead of hidden by a sentence average.