Chapter 04 · Section I · 17 min read
Translating English curriculum to Nepali (and mother tongues)
The useful translation is not English-to-Nepali; it is a parallel bilingual version that uses the NEB textbook term and is honest about where the tool stops.
Most Nepali teachers do translation work every single class period, and most of it goes unrecorded. The textbook says photosynthesis; you say प्रकाश-संश्लेषण, then you say it again in slow English, then you write the equation on the board, then you take a question in Nepali from the back row and answer it half in Nepali and half in English while pointing at the diagram. By the end of forty minutes you have done dozens of small translations that no curriculum document mentions. AI does not replace this work — your students need you in the room — but it can take the prep load off your evenings and produce parallel materials that, until recently, only a very well-resourced school could afford. The catch is that doing this well is not the same as asking ChatGPT to “translate this chapter into Nepali.”
The bilingual-first pattern
The instinct of most teachers, the first time they try this, is to ask the model for a Nepali translation of an English passage. The output looks impressive — fluent Devanagari, paragraphs that scan — and then you put it in front of students and it lands flat. The Nepali is a wall of text in the wrong register. The English terms the students will see on the exam paper have disappeared. And because the prose is generic, you cannot point at a sentence and say “this line, in the textbook, is on page 47.”
The pattern that actually works is bilingual-first — ask the model to produce the same content explained in plain Nepali and in simple English, side by side, with the key terms shown in both languages on first mention. The research on bilingual education is consistent on one point: students learn the concept in the language they think in, and acquire the term in the language they will be tested in. If a Class 7 student understands photosynthesis in Nepali, learning that the English word is “photosynthesis” takes thirty seconds. If they meet “photosynthesis” first in an English paragraph they cannot parse, they spend a week confused about both the word and the idea.
So the workflow is: teach the concept in Nepali, introduce the English term once the concept is clear, drill the English term so the exam paper does not become a translation problem on top of a science problem. The model is good at producing the parallel version that supports this. Ask it explicitly: “Explain this concept for a Class 7 student in plain Nepali, then in simple English. On first mention, show the key term in both languages. Keep the Nepali and English explanations aligned paragraph by paragraph.” The output is a teaching document, not a translation.
Terminology must match the textbook
Here is the mistake that quietly poisons the whole workflow. You ask the model to translate a unit on photosynthesis and it confidently produces उज्यालो-निर्माण or प्रकाश-निर्माण or some other reasonable-sounding compound. It is a perfectly defensible Nepali word. It is also not the word in the NEB textbook, which is प्रकाश-संश्लेषण. When your student opens the SEE practice paper in eight months and sees प्रकाश-संश्लेषण, they will not recognise the term you have been teaching all year. You have just spent a unit’s worth of class time training them to fail a question.
The model does not know what is in the NEB textbook. It is guessing from general Nepali patterns, and the guess is often a different word than the one Nepali educationists agreed on twenty years ago. The fix is to feed the textbook term in as part of the prompt — every time. “Use the following Nepali terms exactly, because these are the NEB textbook terms my students will see in the exam: photosynthesis → प्रकाश-संश्लेषण, chlorophyll → हरितलवक, stomata → रन्ध्र, glucose → ग्लुकोज.” The model will then hold the line through the rest of the document.
Build the glossary unit by unit, the first time you teach each chapter. Keep it in a single document on your phone or laptop. The next year, the work is already done, and the model produces parallel materials that match the textbook from the first prompt.
A worked example — Class 8 photosynthesis
Take a concrete unit. Class 8 NEB Science, the photosynthesis section. The English textbook gives you a paragraph definition, a labelled diagram of a leaf cross-section, the chemical equation, and three review questions. You want a parallel Nepali version your students can read alongside the English textbook at home.
The prompt: “I am teaching Class 8 NEB Science, the photosynthesis unit. Produce a parallel bilingual study sheet — plain Nepali on the left, simple English on the right, aligned paragraph by paragraph. Use exactly these Nepali terms because they are in the NEB textbook: photosynthesis → प्रकाश-संश्लेषण, chlorophyll → हरितलवक, stomata → रन्ध्र, glucose → ग्लुकोज, carbon dioxide → कार्बन डाइअक्साइड. Cover: (1) what photosynthesis is in one sentence, (2) where it happens in the leaf, (3) the inputs and outputs, (4) why it matters for life on Earth. End with three review questions in both languages, with the English term in brackets after the Nepali noun on first mention.”
The output is a two-column sheet you can photocopy. The Nepali column lets a student who is more comfortable in Nepali grasp the idea fully. The English column gives them the exam vocabulary. The terms match the textbook, so nothing gets unlearned later. Total prep time after the first chapter: ten minutes. The first chapter — building the glossary — took an hour. You will reuse that glossary for years.
The same pattern works for math word problems, social studies units, English grammar lessons explained in Nepali, and economics chapters where the technical vocabulary is mostly English. The shape never changes: bilingual columns, textbook terms passed in, alignment paragraph by paragraph.
Mother tongues — the honest limit
A large share of Nepali children do not speak Nepali at home. They speak Maithili in Janakpur, Tharu in the western Tarai, Tamang in Rasuwa and Sindhupalchok, Newari in the Kathmandu valley, Magar in the hills, Limbu and Rai in the east, Bhojpuri across the Tarai. The official medium of instruction may be Nepali, but the strongest language — the one the child thinks in — is something else. For these children, even the Nepali version is a second language. The translation problem is two layers deep, not one.
The honest news is that AI handles these languages poorly. There is very little Maithili or Tamang or Newari or Tharu or Magar text on the internet, and what exists is mostly outside the model’s training corpus. Ask a current model to translate a science paragraph into Tharu and it will produce something — confidently — and a Tharu speaker will tell you within thirty seconds that half the words are wrong, half the grammar is invented, and the dialect is unrecognisable. This is not going to be fixed by next year’s model release. The data problem is upstream of the model problem.
The realistic workflow is a chain: English → Nepali (with AI) → mother tongue (with a human). You produce the bilingual Nepali-English version with the model. Then you find a community speaker — a parent who comes to pick up their child, a colleague who teaches another section, an older student in Class 10 — and you ask them to read the Nepali version and tell you how they would explain the same idea to a younger child at home. You write down their phrasing. Over a year, you build a small mother-tongue glossary the same way you built the Nepali one. The AI did the first leg. A human did the second. Both are necessary; neither is sufficient alone.
Be honest with the children about this. “This is what the science textbook says in English. This is what it means in Nepali. And this is how Anita’s mother says it in Tharu. We are checking the Tharu with her because the computer does not know Tharu well yet.” You have just taught them three useful things at once: the content, two languages, and an accurate picture of what the tool can and cannot do.
The verify-against-textbook gate
One last discipline, and it protects every other part of the workflow. Anything that goes onto a worksheet, a test, a notice, or a printed handout gets checked against the textbook before it leaves your desk. Read the Nepali column. Find the corresponding term in the textbook. Confirm it matches. This takes a minute per page and it is the single highest-leverage minute you will spend.
The failure mode you are guarding against is subtle. The model produces a unit that is 95% correct, beautifully formatted, and contains one term that drifted — कोष instead of कोशिका, or a date that the model invented, or a unit conversion that was wrong by a factor of ten. The students take it home, study from it, and arrive at the test with one quiet wrong fact in their heads. The cost of catching this is sixty seconds; the cost of not catching it is the rest of the year spent unteaching it. The gate is non-negotiable. The model is your first-draft assistant; the textbook is your authority.
Check your understanding
Quick check
—A Class 7 English-medium science teacher has students whose home language is Nepali. She wants to use AI to support a unit on the digestive system. Which workflow best matches the bilingual-first pattern this section recommends?
Quick check
—You ask an AI to translate a science unit on photosynthesis into Nepali. The output uses the term उज्यालो-निर्माण throughout — a fluent, defensible Nepali compound. The NEB textbook your students will be tested on uses प्रकाश-संश्लेषण. What should you do?
What comes next
The bilingual workflow assumes the student is reading the English material alongside the Nepali version. But many of your students cannot yet read the English material at all — they are in English-medium classes with limited English. The next section is about scaffolding for those students specifically: pre-teaching vocabulary, building bilingual reading guides, and the routines that grow English proficiency over time rather than creating permanent dependency on translation.