When the home language has no training data

The bottom of the gradient

The first lesson in this block laid out the gradient: languages the model has read a great deal of, some of, little of, almost none of. This lesson is about the last group, because a large share of the world’s students live there: in languages with a few million speakers, or a few hundred thousand, that are spoken daily and written rarely, and that appear on the web as a handful of pages.

For those languages the model is not a translator. It has seen too little to have learned the grammar, the vocabulary, the register. What it produces is fluent-looking text assembled from fragments and from the patterns of neighbouring languages, and it is wrong in ways that a speaker sees at once and a non-speaker never can. The check question’s teacher did everything right (detailed prompts, glossaries) and it did not help, because prompting shapes what the model does with what it knows, and here it knows almost nothing.

Nothing in this course’s earlier techniques changes that. It is a limit of the data, not of the prompt.

What still works

The English side, fully. Every technique in this course works for the English (or Hindi, or Kiswahili) material: the passages, the questions, the explanations, the feedback. The model does that part; the crossing into the home language is done by people.

A speaker with a draft. For the languages just above the bottom (some text exists, the model has some of it) a model draft checked by a speaker is faster than a speaker starting from nothing. The back-translation check tells you whether you are in this zone: if the round trip comes back roughly right, a speaker’s check makes the draft usable. If it comes back as nonsense, the draft is not saving anyone time.

Students as the bridge. A class that speaks the language is a room full of translators. A worked example in English, translated by the students into the language they think in, explained to each other in it, and written up in both: that is a bilingual lesson that needs no model at all for the home-language half, and it is better pedagogy than a machine translation would have been.

Regional models. A few projects train models on specific language groups (African languages, Indian languages, Indonesian languages) and some are free to use. They are worth checking, because a model trained deliberately on a language does far better than a general one that saw it by accident. Coverage is patchy and changing.

Where the data would come from

The reason the model has not read the language is that it is not written down on the web, and the reason for that is history: which languages were printed, taught, digitised, and which were not. The absence is not an accident of technology. It is the shape of who has been writing on the internet for thirty years.

Projects exist to change it. Mozilla’s Common Voice collects recorded speech in hundreds of languages from volunteers, and it is what makes speech recognition possible for a language it did not previously exist for. Masakhane is a network of African researchers building translation for African languages. Several Indian and South-East Asian initiatives do the same for their regions. A teacher and a class can contribute (an hour of read sentences in Common Voice is real data) and a student who has done it understands something about where the model’s abilities come from that no lesson conveys.

The political question, said plainly

Whose languages get models is decided by where the data is, which is decided by history and money. A student in a small-language community is offered tools that work beautifully in English and not at all in the language of their home, and the message that carries, that their language is not one that machines are for, is a message worth naming in class and refusing. The tools are unequal because the data is unequal, and the data is unequal because of choices people made, and some of those choices can be unmade by people who care to. That is a citizenship lesson and a true one.

For the teacher tonight

Practically: use the model for the English side and your students and colleagues for the rest. Do not put a machine translation in a small language in front of a class without a speaker having read it. And if a company or a ministry offers a tool that claims to support the language, test it with three sentences and a speaker before believing the claim. The gradient is real, and the bottom of it is where the claims are least reliable and the marketing most confident.

The honest limit

This will improve. Language coverage in models grows each year, and a language with three million speakers may be reasonably served within a few years if someone builds the data. But a teacher cannot wait, and a student in Class 5 now will be in Class 10 before the model learns their language. For now the tool is for the English, and the language of the home is the students’ and yours.