Regional Language Grievance Classification

A state grievance cell showed us a dashboard that had 'improved' after an AI triage went live. Average time-to-bucket was down. A district collector then forwarded a stack of water-supply tickets that had been sitting in a social-welfare queue. The classifier had seen the word for a community and the word for a tap in the same Odia–English SMS and had chosen the community.

The original SMS was not in the officer's screen. Only an English gist was. Nobody could see the tap.

This is a composite case study from work like that — details merged so no inbox is identifiable. It is not a claim that one state endorsed a product. It is the failure mode and the rebuild.

What the old stack did

Inbound regional tickets were machine-translated to English. A classifier trained on English categories assigned a department code. A clerk saw the English gist and the code. The original was in an attachment nobody opened under time pressure.

That design fails in three regular ways. Mix is flattened. Landmarks and scheme nicknames are mistranslated into generic English. And the officer cannot audit the bucket without the original.

The rebuild

Keep the original. Tag language and mix. Classify against the state taxonomy using the original plus a glossary of department and scheme aliases in that language. Show the officer the original, the bucket, and a short reason in the officer's working language. Human review on low confidence and on a list of dangerous pairs — water versus welfare, land versus forest, police versus revenue.

Do not start with twenty-two languages. Start with the two or three that already arrive on paper. The Eighth Schedule is not your year-one taxonomy.

What changed in the file after the rebuild.

Before	After
English gist only on the clerk screen	Original plus optional gist
One blended accuracy number	Per-language confusion matrix
All errors treated equal	Dangerous pairs always human
Vendor model name in the minutes	Taxonomy version + model hash + override log
Photos ignored or dumped in 'other'	Photo in scope only if a human looks, else written out

The taxonomy is the model

Most 'AI failures' were taxonomy failures. Categories overlapped. A leak in a scheduled-caste hostel is both welfare and public works. If the taxonomy cannot express two codes, the classifier will look stupid. Allow multi-label or a primary-plus-secondary, and write which department owns the clock.

Citizen language will not match the secretariat's tree. Glossary the informal names of departments. 'Bijli wale' is not a joke row. It is how the ticket will arrive.

What we measured — without inventing a podium

We do not publish a state's numbers here. The method was: a held-out month, per-language matrices, a count of dangerous-pair overrides, and a count of tickets whose original language never reached the officer. The last count is the one that moved the secretary. It was not an F1 score.

When the English-pivot classifier was compared to the in-language classifier on the same month, the interesting result was not a headline percentage. It was which departments stopped receiving other people's water.

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.

Clerks prefer English. Showing the original will slow them.

Clerks prefer not to be shouted at for a wrong bucket. The original is faster than a recall. Give them both.

We need one model for the whole country.

You need one method. The taxonomy and the glossary are state objects. A national model that has not sat your codes is a draft.

Human review on dangerous pairs will not scale.

Those pairs are why the cell exists. If volume is impossible, narrow the public form so the pair is chosen by the citizen, then still allow a correction.

This is just rules plus a glossary, not an agent.

If rules plus a glossary clear the queue, ship that. Add an agent when the next step is a draft reply or a retrieval, not as a costume for classification.

Do not publish their F1

Someone will want a success story with a percentage. A composite case study does not need one. Publish the method: originals on the screen, dangerous pairs human, per-language matrices on the file. A number without a month and a language column will be reused as a national fact.

If a secretary insists on a figure, give the count of water tickets that left the welfare queue after the rebuild. That figure is a service fact. It is not an Indic leaderboard.

A month to rebuild triage without a press note

Do this quietly. The press note is for when the wrong-department pile shrinks.

1. Week 1: dump last quarter. Keep originals. Tag language and mix.
2. Week 1: list dangerous category pairs with the cell.
3. Week 2: clean the taxonomy. Allow secondary codes.
4. Week 2: build the alias glossary in the incoming languages.
5. Week 3: train or prompt an in-language classifier. Compare to the English pivot on a held-out month.
6. Week 3: change the clerk screen. Original first.
7. Week 4: force human on dangerous pairs and low confidence.
8. Week 4: log overrides. That log is the next glossary update.

How this shows up in the file

The note should say: we classify in the incoming language. We do not use an unofficial English hop to choose a department. Officers see the original. Dangerous pairs are human. The taxonomy version and the model hash are on every routed ticket.

Attach the confusion-matrix method and the dangerous-pair list. Do not attach a fake national accuracy.

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

“Regional Language Grievance Classification” 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 “grievance classification regional language”.

A grievance agent that translates to English and then buckets will send water to caste and pensions to public works. Classify in the incoming language, then route. 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 “Regional Language Grievance Classification” 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. “grievance classification regional language” 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 “Regional Language Grievance Classification” is only a heading, it will not survive a file inspection. A P1 CIO/CTO should be able to attach one artefact that proves “grievance classification regional language”: 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 “grievance classification regional language” 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

“Regional Language Grievance Classification” earns a line in the noting only if a P1 CIO/CTO can attach proof of “grievance classification regional language.” 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 “grievance classification regional language.”
• 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.