Skip to content

Seed catalogs for the polysynthetic and syllabary languages of the Americas #1655

Description

@dqnykamp

#1654 seeded eight languages indigenous to the Americas: Haitian Creole, Quechua, Guarani, Aymara, Nahuatl, Kʼicheʼ, Mapudungun and Ojibwe.

Five more were in that batch's original scope and were deliberately left out: Kalaallisut (kl), Inuktitut (iu), Cree (cr), Cherokee (chr) and Navajo (nv).

Why they were held back

Every other catalog in this repository is an unreviewed machine-generated seed, and that is a workable state because a seed is checkable: a speaker can read it against the English beside it, and lint:i18n reports what is missing. These five would not have met that bar, for three distinct reasons.

Polysynthesis means a seed would be inventing words, not recalling them. Kalaallisut, Inuktitut and Cree build a sentence's worth of meaning into single words through long affix chains. There is no dictionary form to look up for "no attempts remaining" or "invalid attribute value" — the word has to be constructed, and constructing it wrong produces something that looks like the language and is not. That is a different failure from a wrong word choice, because a reader cannot tell the two apart.

Tone and diacritics carry meaning and would be wrong at a high rate. Navajo marks tone, nasalization and vowel length, and all three are contrastive. A seed that gets them mostly right is not mostly correct; it is a text a learner would be actively harmed by reading.

Three of the five are written in scripts a reader cannot check the seed against. Cherokee uses its own syllabary, and Inuktitut and Cree are published in syllabics — those are the orthographies their school boards and governments actually use, so writing them in Latin would be the wrong answer too. packages/i18n/README.md already gives exactly this reason for omitting the Khmer, Lao and Sinhala chemistry names rather than guessing them:

An unreviewed guess written in a script the reader cannot check against the English beside it is worse than the English

That argument was made about 130 chemistry keys. It applies with more force to a whole catalog of 562.

What would unblock each

These are not equally hard, and they do not all need the same thing.

  • kl Kalaallisut — has a stable official orthography, a language secretariat (Oqaasileriffik), and a real body of published administrative and educational text. The blocker is word-building rather than the writing system. A corpus of Greenlandic software or school UI strings would probably be enough to seed from.
  • iu Inuktitut — Nunavut publishes government and school material in syllabics, and there is existing localized software UI to draw on. Note the script decision if this is picked up: CLDR's default for iu is Cans, and a iu-Latn reader would then reach the syllabics catalog, which is the bs-Cyrl asymmetry the README already describes.
  • cr Cree — same script question, plus cr is a macrolanguage; whoever seeds it should decide which variety the catalog is written in and say so in its header, the way locales/qu names Southern Quechua. Note that Seed catalogs for eight indigenous languages of the Americas #1654's MACROLANGUAGE_MEMBERS will need a cr entry, and oj's existing entry is the model.
  • chr Cherokee — the Cherokee Nation runs an active language program with published materials and existing software localizations. The syllabary is well supported in Unicode; the blocker is vocabulary, not encoding.
  • nv Navajo — the hardest of the five. Needs a speaker or a substantial verified corpus, because tone and the verb system are where a machine pass goes wrong invisibly.

Suggested approach

Rather than another machine pass, these want either an existing localization corpus to seed from, or a speaker willing to review. #1521's translation platform is the natural home for the second. Splitting them is fine — kl and chr look most tractable and could ship well before nv.

Whoever picks this up: packages/i18n/README.md is the place to start, and the eight catalogs in #1654 are the closest models for headers that state honestly what a seed does and does not know.

🤖 Generated with Claude Code

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    Priority

    None yet

    Effort

    None yet

    Projects

    Status
    No status

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions