An AI needs to retain the meaning of a teacher’s English explanation when the teacher moves into Twi, then respond to the full idea in the language mix the teacher chose. Translation alone can miss the example, the correction, and the teaching goal carried across both languages.
At 7:18 p.m., Efua stood beside a whiteboard in a small after-school class in Kumasi, rubbing a fading marker between her fingers. Her pupils had followed the English definition of evaporation, but their faces changed when she asked why wet uniforms on the line become dry.
“Water turns into vapour,” she began, then paused. “Nanso, sɛ wokae wo ntoma a, ɛnyɛ sɛ nsuo no ayera. Ɛkɔ wim.”
One boy pointed at the drawing and asked, in Twi, whether the water had “gone inside the sun.” Efua needed to correct the idea before it settled. The lesson was almost over, the rain had started outside, and the next day’s class depended on them understanding what the picture did and did not mean. If the explanation broke here, they could carry the wrong answer into their work.
She tried an AI companion to help shape a short follow-up explanation. A word-by-word translation could repeat “water goes into the sky” and make the confusion worse. What Efua needed was an answer that held the thread: evaporation, heat, the child’s question, and her switch into Twi to make the idea feel close to home.
A language switch often carries the teaching move
Teachers switch languages for a reason. English may introduce the formal term. Twi may make the example land. A return to English can help pupils recognise the word when it appears in a textbook or exam question.
That sequence is part of the explanation. When an AI treats each turn as a separate language task, it can lose the point that connects them.
Efua was not asking for a dictionary definition of “vapour.” She was trying to move her pupils from a visible thing, wet cloth on a line, to an invisible process. Her Twi sentence was doing important work: it reassured them that the water had not vanished, while leaving room to explain where it had gone.
A useful conversational AI should follow that movement. It should understand that “nsuo no ayera” refers back to the water in the uniform, not introduce a new topic. It should recognise a correction when it appears in Twi after an English science term. Then it should answer in language that helps the teacher continue, rather than forcing a fresh start.
Context gives the reply its shape
The difference becomes clear in the next prompt. Efua says, “Explain it again, but make it simple. Mma wɔnnwene sɛ owia no nom nsuo no.”
The request contains an instruction, a concern, and a classroom image. An AI that follows the context can reply with something like: “The sun does not drink the water. Heat helps tiny bits of water leave the cloth and mix with the air. That is evaporation.”
The wording can be adjusted, but the job stays the same. Keep the earlier misconception in view. Use the teacher’s chosen mix of English and Twi. Give an explanation short enough to say aloud before the class packs up.
This matters beyond science. A teacher might introduce “main idea” in English, then use Twi to ask a child to explain what the story is really saying. They may change languages while checking a calculation, calming a nervous pupil, or finding an example from home. Each switch can signal emphasis, care, correction, or a request for a clearer path.
The work around context-aware, curriculum-aligned conversational AI for teacher education, including GenAITEd Ghana, points to the same practical need: an assistant has to follow the teaching situation, not only the words placed in front of it.
Voice makes it easier to keep the lesson moving
In the classroom, hands are often busy. Efua has a marker cap in one hand and exercise books stacked on a desk. Speaking a follow-up question can be easier than stopping to type it.
Nkomo supports natural Twi and Ghanaian English conversation by voice or text. In hands-free voice mode, you hold to speak, hear a reply, and can interrupt when the answer has gone on long enough. That gives a teacher room to ask for a shorter version, change the example, or correct the direction of the explanation while the original context is still fresh.
Concise replies matter here. A long answer can leave the teacher searching for the one sentence worth repeating to pupils. The better reply gives them a usable next line.
That is also why silence is a problem. If a request cannot be handled, an assistant should show the error rather than leaving a teacher wondering whether it heard the Twi switch at all. What happens when your voice assistant goes quiet after a Twi switch? explores what that interruption can cost in an ordinary conversation.
Context needs boundaries, too
A teacher may be discussing a child’s difficulty, a draft lesson plan, or a voice note they do not want retained beyond the moment. Context is useful when it serves the conversation, but people should know what happens to it.
Nkomo makes cloud consent explicit. You can choose never, ask each time, or this session. Its history stays on the device under your control; turn history off and it is purged immediately. Data export and account deletion are available in one tap.
Later that evening, Efua asked for one final version: two sentences, Twi first, then the English term in brackets. She copied it onto tomorrow’s lesson note. The rain had eased. On the board, beneath the drawing of the uniform, she wrote one new line for the class: “Nsuo no nkyerɛ sɛ ayera. It has changed into water vapour.”
Comments
No comments yet.