A compact, multi-dimensional notation for quantifying a language's NLP
resource profile. Written as S/M/L1-B1/L2-B2/…,
each integer is an order of magnitude derived from a canonical source:
Ethnologue speakers, OSCAR 23.01 monolingual corpora, OPUS parallel
corpora. Designed first as a communication tool — a
reader of 6/0/en-6/es-3 (Quechua) or 1/1/en-3
(Owens Valley Paiute) immediately grasps a resource profile that a
single “low-resource” label would flatten.
Paper draft: paper.pdf (updated live from the LaTeX source).
“Low-resource” appears in thousands of NLP papers with no shared definition. A paper may use the term whether a language has 500 parallel sentences or 500,000, whether it has 10 fluent speakers or 10 million. This makes it difficult to compare results across papers, to assess whether a technique might transfer to a new language, or to communicate the difficulty of a task. Existing classifications (e.g. Joshi et al. 2020) provide coarse, single-dimensional categories that are hard to reproduce from first principles and too imprecise for quantitative reasoning.
A RAN score is written as
S / M / L1-B1 / L2-B2 / …
where every integer is a floor of a base-10 logarithm:
preprocessing=moses. Pairs are listed in descending order.
Compact (fits in an abstract), multi-dimensional (separates speakers,
monolingual, and bilingual resources), reproducible (derived from
canonical, citable sources), and interpretable (each integer is an
order of magnitude). When a paper reports evaluation on a
3/2/en-4 language (Cherokee) versus a
7/4/en-7/fr-7 language (Swahili), a reader instantly
grasps the difference across multiple dimensions, without looking up
corpus statistics or consulting a classification table.
20 typologically diverse languages spanning the full RAN range, from
English (9/10/…) to Owens Valley Paiute
(1/1/en-3). Includes four Indigenous languages of the
Americas: Quechua, Guarani, Cherokee, Owens Valley Paiute.
| Language | RAN | S | M | Ben | Bmax | sum |
|---|---|---|---|---|---|---|
| English | 9/10/es-9/fr-9 | 9 | 10 | 0 | 9 | 28 |
| Spanish | 8/9/en-9/fr-8 | 8 | 9 | 9 | 9 | 26 |
| Swahili | 7/4/en-7/fr-7 | 7 | 4 | 7 | 7 | 18 |
| Guarani | 6/1/en-6/es-3 | 6 | 1 | 6 | 6 | 13 |
| Quechua | 6/0/en-6/es-3 | 6 | 0 | 6 | 6 | 12 |
| Cherokee | 3/2/en-4 | 3 | 2 | 4 | 4 | 9 |
| Owens Valley Paiute | 1/1/en-3 | 1 | 1 | 3 | 3 | 5 |
See the paper Table 1 for the full 20-language
inventory. All numbers are derived from published sources and are
independently verifiable from the links in gather_data.py.
For Owens Valley Paiute (ISO mnr), which is
effectively absent from OPUS, bilingual data is sourced from the
community-curated Kubishi
Dictionary (4,484 mnr–en sentence pairs).
We fit simple linear regressions predicting published benchmark scores from RAN components. The aim is not to argue for dropping any component but to show that the components carry complementary information.
| Predictor model | MT | NER | POS |
|---|---|---|---|
| S + M + Bmax | 0.48 | 0.69 | 0.77 |
| S + M + Ben | 0.48 | 0.55 | 0.48 |
| Bmax only | 0.26 | 0.38 | 0.60 |
| M only | 0.26 | 0.22 | 0.35 |
| Ben only | 0.26 | 0.04 | 0.09 |
| RANsum | 0.18 | 0.18 | 0.37 |
| S only | 0.00 | 0.00 | 0.05 |
Training R² — MT: NLLB-200 distilled-600M chrF++, n=17; NER: XTREME WikiANN XLM-R Large F1, n=10; POS: XTREME UD XLM-R Large accuracy, n=9. Three observations follow. First, the full three-component model dominates every task: components carry complementary information. Second, Bmax is the strongest single predictor for NER and POS, while Ben alone explains almost nothing on those tasks (R² = 0.04 and 0.09). Cross-lingual transfer benefits from connections to any high-resource partner, not just English. Third, S alone is uninformative of benchmark performance — but the paper argues S should be kept anyway: corpus size tells you what a model can do today, and S tells you why we should care and what resources could plausibly become available.
A notation is only as useful as the shared numbers it references. The
paper also describes a community-maintained web database
(planned at ran.kubishi.com, not to be confused with this
review site) where anyone can submit an update — ISO 639-3 code,
any subset of S, M, or Li-Bi pairs with citations.
Updates are monotone (max-overwrite); submissions queue for admin
review; GitHub-OAuth-authenticated admin roster. Older citations remain
valid lower bounds.
git clone https://github.com/kubishi/RAN cd RAN uv sync # Full pipeline: data → analyze → plots → report make all # Or rebuild the paper: cd paper && latexmk -pdf paper.tex