How Ghana texts
We’re building a model that reads Ghanaian messages the way they’re meant, and a dataset to teach it with.
Started . Updated .
A message in Ghana often carries more than its words. People switch between English, Pidgin and Twi in one sentence. An emoji can be a thank-you or a jab, and a compliment can be a complaint. A Ghanaian reader gets all of this in a second. A tool, however, has to be taught each one.
We’re teaching one. The model will read four things in a message: the feeling behind it, whether it means the opposite of what it says, how soon it needs an answer, and which languages it mixes. Figure 1 shows what that looks like.
To a landlord
massa, water start dey drip from ma ceiling since yestee. adey beg, send somebody make e come check today
- What it means
- Water is coming through my ceiling. Send a plumber today.
- Feeling
- Worried
- Means the opposite?
- No
- Needs an answer
- Right now
To a school
Ei, so the fees went up again and nobody told us? Nice one 👏
- What it means
- You raised the fees without warning us, and I’m not pleased.
- Feeling
- Fed up
- Means the opposite?
- Yes
- Needs an answer
- Soon
Family chat
Mummy says if you people don’t come for Sunday lunch she won’t forgive anybody 😂
- What it means
- Come for lunch on Sunday. Nobody is really in trouble.
- Feeling
- Happy
- Means the opposite?
- No
- Needs an answer
- It can wait
To a friend
Wo ho te sɛn? Wo da so wɔ Kumasi? Mɛba Friday
- What it means
- How are you? Are you still in Kumasi? I’m coming on Friday.
- Feeling
- Calm
- Means the opposite?
- No
- Needs an answer
- It can wait
Why this is hard
The text that tools learn from is mostly not ours. Ghana NLP keeps a register of datasets for Ghanaian languages; it lists Ghanaian Pidgin, Ewe and Fante as having no data, and no sentiment or social-media sets for any Ghanaian language.
Source: Ghana NLP, Ghanaian NLP datasets and models, github.com/GhanaNLP, read November 2026.
What does exist is hard to label. AfriSenti, a public set of Twi tweets labelled for feeling, had three people label each tweet. When we counted their published labels, all three agreed on 41% of them.
Source: AfriSenti (SemEval-2023), Twi individual labels, github.com/afrisenti-semeval; our count, 18 November 2026.
Sarcasm is the hardest part, because a reader can only guess what the writer meant. So we ask the writer. Each person who sends a message tells us how they felt and whether they meant the opposite, the method the iSarcasmEval benchmark used.
Source: iSarcasmEval (SemEval-2022), github.com/iabufarha/iSarcasmEval.
What we’re collecting
- What’s in it
- Messages people type the way they’d really send them: to companies, schools, landlords, friends, family and group chats. Some were really sent; some are made up for us. We keep the two apart, and test the model only on the real ones.
- Who it’s from
- Anyone aged 18 or older. We don’t ask for a name, number or email, and people write NAME or NUMBER in place of real ones.
- How it’s labelled
- The writer tells us how they felt, whether they meant the opposite and how urgent it was. Then we label each message ourselves, following a written guide.
- Consent
- Nobody can send a message without agreeing that we may keep it and share it.
- How it’s released
- We may publish the messages and our labels as an open dataset. That isn’t decided yet; everyone who sends a message agrees to it in case we do.
- What never goes in
- Messages from the businesses we work for, not even with names removed, and other people’s datasets, which we use under their own licences.
- Removing a message
- Tell us what you wrote at privacy@asledatalabs.com and we’ll find it and delete it.
Where it stands
- The form opened to anyone who texts.
- NextThe first 500 messages labelled, and the guide checked against them.
- After thatA first model, tested only on real messages. We’ll post how it does here, good or bad.
How to help
Type a message the way you’d really send it, then tell us what you meant. It takes about a minute, and you can add as many as you like.
Once it reads Ghanaian messages well, a model like this could sort a business’s messages by urgency and topic. That’s part of our Machine Learning Development work.