Burushaski is a language isolate spoken in Hunza, Nagar and Yasin. It is related to nothing — no established family, no sibling language whose resources can be borrowed or adapted.
For a speaker, that is a point of pride. For anyone building software, it is a precise description of the difficulty: every technique that works by transfer from a related language is unavailable from the start.
What "low-resource" actually means
The phrase suggests a smaller version of the same problem. It is a different problem.
- No large digitised corpus, so anything statistical has nothing to count.
- No standard orthography, so the same word appears several ways and every approach that treats strings as identities fragments.
- No tokeniser, tagger or spellchecker, because each of those is trained on data that does not exist.
- No keyboard layout in the places people actually type.
Modern language tooling assumes the corpus is a given and the interesting work starts afterwards. Remove that assumption and the ordering inverts: collection is the work.
The unglamorous first steps
For a language in this position, the highest-value engineering is not a model. It is infrastructure:
- A keyboard layout people can install, on the devices they own.
- A stable orthography with a documented mapping to Unicode.
- A recording and transcription pipeline simple enough for non-technical speakers, because the fluent speakers are mostly not engineers.
- A dictionary that is a database, not a PDF — queryable, correctable, exportable.
None of that is research. All of it is the precondition for research.
Why it is worth doing
There are hundreds of languages in this position, and the number of speakers is falling in most of them. The tooling gap is self-reinforcing: no data means no tools, no tools means the language is harder to use on a phone, and every year that is true pushes more conversation into the dominant language instead.
Breaking that loop does not require a breakthrough. It requires someone building the boring layer — a keyboard, a corpus, a dictionary with an API — and doing it in a form other people can build on.
That is an engineering problem, and engineering problems yield to sustained attention.