Why Machine Translation Fails on Scheme Names

The citizen wrote 'ayushman card nahi bana, PMJAY wala'. The translator expanded one acronym, dropped the other, and produced a Hindi sentence about a state health insurance that had been subsumed years ago. The agent then retrieved the old guideline. The citizen was told to visit a kiosk that no longer existed.

Nothing in that chain looked like a model failure on a news benchmark. Every step was a proper-noun failure on a government object.

This teardown is about why machine translation — and the agents that secretly pivot through it — break on scheme names, and how to lock the names so the rest of the sentence can be fluent without opening the wrong treasury code.

Why names break MT

Scheme titles are half acronym, half political sentence. They collide across Centre and states. They are amended, renamed, merged and remembered by the old name. They appear in English inside a Hindi sentence and in Hindi inside an English noting.

A translator is trained to produce a fluent sentence. Fluency wants to expand, to avoid repetition, to pick a synonym. Those are vices when the token is a key. PM-KISAN is not improved by becoming a generic phrase about farmer income. A local scholarship named after a leader is not improved by becoming 'the student help scheme'.

Back-translation will not always catch this. The wrong scheme can back-translate into something that still looks like help for farmers. The file needs a glossary check, not only a BLEU loop.

Failure modes we see on real tickets

Expansion into the wrong long form. Collapse of two schemes into one. Revival of a closed scheme because the training data was old. Translation of a person's name inside a scheme title. Translation of a year or a phase into a different phase. A helpful 'similar scheme' that the citizen did not ask for.

Label these on every eval item that contains a scheme.

Mode	What the output did	Mark
Wrong expansion	Acronym opened into a different official title	Hard fail
Merge	Two schemes treated as one benefit	Hard fail
Ghost	Closed or never-existed scheme offered	Hard fail
Gloss only	Official name kept, explanation added	Pass if the official name is intact
Nickname map	Citizen nickname mapped to the official title	Pass if only one official title results

Lock the glossary before you tune the model

Build a list: official title in each authentic language, short title, banned poetic equivalents, citizen nicknames, status (live, closed, merged-into), and the document id the agent must retrieve. Put the list in front of the model as a constraint, in the retriever as aliases, and in the eval as a hard check.

When a new office memorandum renames a scheme, the glossary version changes on the same day as the website. A model that is 'generally good at Hindi' will not notice the rename for months.

The Official Languages Act's bilingual instruments are a gift here. Where Hindi and English authentic titles exist, store both. Do not let the model invent a third.

Architecture that does not 'improve' a name

Detect scheme mentions with the glossary, not with generic named-entity guesswork alone. Retrieve by the official id. If you must translate the rest of the sentence, mask the title, translate, and unmask. If you pivot through English, mask first. An unmasked pivot is how nicknames become ghosts.

If the agent is not sure which scheme was meant, it asks. It does not pick the most famous one.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

If we lock names, the language will sound unnatural.

Citizens already say the acronym. Officers already say the acronym. Naturalness that changes a treasury code is not a public-service value.

The glossary will be stale.

Then put a named owner and a trigger: every OM that creates, renames or closes a scheme updates the list the same week. A stale glossary is visible. A stale model is not.

Our MT vendor says they handle named entities.

Then they can sit a fifty-item scheme pack from your website. If they refuse, you have the answer.

We should let the model explain schemes in simpler words.

Explain after the official title is locked. Simplification without the title is how a benefit becomes a slogan.

Nicknames are aliases, not poetry

A nickname is a second key into the same scheme. It is not an invitation to write a better name. When a citizen says the old mission name, map it and say the official title back. Do not congratulate them on their phrasing and retrieve a neighbour.

If two nicknames collide across schemes, the agent asks. Collision is a glossary fact. It is not a reason to pick the more famous scheme.

Ten days to stop translating the key

Start with the schemes that move money this quarter.

1. Day 1: list live schemes and their official titles in each authentic language.
2. Day 2: add nicknames from last year's tickets.
3. Day 3: mark closed and merged schemes with a pointer.
4. Day 4: write banned equivalents — the poetic phrases you have already seen.
5. Day 5: implement mask-translate-unmask or a constrained decode on titles.
6. Day 6: point retrieval at scheme ids, not at translated phrases.
7. Day 7: build a fifty-item pack that only tests names.
8. Day 8: sit the current stack. Label wrong expansion, merge, ghost.
9. Day 9: name the glossary owner and the OM trigger.
10. Day 10: put hard-fail language into the go-live checklist.

How this shows up in the file

The note should say: scheme titles are constrained by the attached glossary. Machine translation may not expand, merge or poetically replace them. Nicknames map to one official title. Closed schemes are suppressed unless the ticket is about arrears.

Attach the glossary version and the fifty-item result.

This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation, linguistics or engineering advice. Confirm against the current Gazette, Official Languages Act and Rules, state official-language law, GIGW, RPwD Act, DPDP text and Rules, CERT-In directions, departmental manual and your counsel before you file it.

How to test this with real speech, not staff English

“Why Machine Translation Fails on Scheme Names” fails in the field if you only tested officers. A P1 CIO/CTO should hear a first-generation student, a rural caller, or a Hinglish grievance before claiming “translation error government scheme”.

Scheme names are not phrases to be improved. They are keys into money. A translator that helpfully expands an acronym has opened the wrong door. Twenty-two scheduled languages is a Constitution fact, not a model fact. Script support is not language support. Official language rules may require bilingual output even when the model prefers one script.

• Name the languages and scripts in the eval set.
• Include code-mix and scheme-name tests.
• Measure comprehension, not BLEU alone.
• Design a human fallback when language fails.

Close this loop before the next CAB

Put “Why Machine Translation Fails on Scheme Names” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P1 CIO/CTO, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “translation error government scheme” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

What must be true before you file this

If “Why Machine Translation Fails on Scheme Names” is only a heading, it will not survive a file inspection. A P1 CIO/CTO should be able to attach one artefact that proves “translation error government scheme”: a log export, a clause, a scored row, a dated notice, or a refusal rule.

Write three dated sentences: what was decided, who owns it, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the owner of “translation error government scheme” inside the institution.
• Attach one artefact a stranger can open next year.
• Revisit when the model, the notice, or the SI changes.
• Do not treat a vendor slide as evidence.

What the next file must contain

“Why Machine Translation Fails on Scheme Names” earns a line in the noting only if a P1 CIO/CTO can attach proof of “translation error government scheme.” A heading is not proof. A vendor slide is not proof. A workshop photograph is not proof.

Write three dated sentences: what was decided, who owns it after the next posting order, and when it will be re-checked. If you cannot write the three sentences, you are not ready to buy, to sell, or to go live.

Leave unsourced percentages out of the note. DPDP is not a blanket localisation statute. The November 2025 AI governance text is guidance, not an Act. CERT-In’s 28 April 2022 directions still set specified incident and log clocks. A PAC, when lawful, lives in GFR Rule 166.

• Name the designation that owns “translation error government scheme.”
• Attach one artefact a stranger can open next year.
• Record the instrument you are actually using.
• Revisit when the model, the SI, the notice or the posting changes.