What we do
Text and language
Language technology has been built mostly in a handful of languages. Many of our labellers grew up speaking two or three of the others.
Tasks
- Classification of documents, messages and posts by topic, intent, sentiment or policy.
- Entity tagging: names, places, organizations, dates, amounts, products.
- Transcription of speech, and review of machine transcripts, with timestamps and speaker turns.
- Translation review: scoring and correcting machine translations against agreed guidelines.
- Document extraction: fields from receipts, invoices, forms and scanned records.
- Writing and rewriting examples for training and evaluating language models.
Languages
| Language | Centre | Labellers able to work in it |
|---|---|---|
| English | Both | 2,964 |
| Kiswahili | Nairobi | 1,690 |
| Sheng (Nairobi urban speech) | Nairobi | 1,105 |
| Gĩkũyũ | Nairobi | 412 |
| Dholuo | Nairobi | 296 |
| Kalenjin | Nairobi | 188 |
| Ekegusii | Nairobi | 131 |
| Somali | Nairobi | 74 |
| Amharic | Nairobi | 22 |
| Filipino | Manila | 1,214 |
| Cebuano | Manila | 233 |
| Ilocano | Manila | 157 |
| Hiligaynon | Manila | 96 |
| Waray | Manila | 48 |
- Labellers tested and approved for text work in each language at 31 December 2025, including Rungwell at Home. Most labellers work in more than one.
Why it matters
A Kiswahili speaker in Nairobi does not speak the Kiswahili of a textbook, and a model that has only read textbooks will not understand her. Our labellers write, classify and correct text in the languages as they are spoken and typed, including the mixtures (Sheng, Taglish) that most real messages are written in, and they are paid and credited as the skilled linguists they are.
Page last reviewed 28 September 2026