Resource Abundance Notation

RAN (Resource Abundance Notation) is a compact, multi-dimensional, reproducible way to communicate a language's NLP resource profile at a glance.

The term "low-resource" is used constantly in NLP but means almost nothing precise. A paper might call a language low-resource whether it has 500 parallel sentences or 500,000, whether there are 10 fluent speakers or 10 million. Coarse single-number classifications (e.g. Joshi et al. 2020's classes 0–5) don't capture that spread. RAN gives a short score a reader can decode without a lookup table.

The score

A RAN score is written as a fixed sequence of slash-separated components:

S / M / L1-B1 / L2-B2 / …
  • S = ⌊log₁₀(fluent speakers)⌋ — kept for what it says about a language's vitality and community, even though it doesn't predict system performance.
  • M = ⌊log₁₀(monolingual corpus sentences)⌋.
  • each Li-Bi = a bilingual partner language Li and ⌊log₁₀(parallel sentences with it)⌋, listed in descending order of Bi so the strongest resource comes first. A trailing slash with nothing after it (e.g. 0/3/) means no parallel corpus at all.

Every digit is an order of magnitude, so the numbers read directly:

DigitMagnitude
0ones
1tens
2hundreds
3thousands
4tens of thousands
5hundreds of thousands
6millions
7tens of millions
8hundreds of millions
9billions

How to read it

LanguageRANReading
Spanish8/9/en-8/fr-8/pt-8Hundreds of millions of speakers, billions of monolingual sentences, hundreds of millions of parallel sentences each with English, French, and Portuguese
Swahili7/4/en-7/fr-6Tens of millions of speakers, tens of thousands of monolingual sentences, tens of millions of en–sw parallel sentences
Quechua6/0/en-6/es-2Millions of speakers, almost no monolingual corpus, millions of en–qu parallel sentences
Cherokee4/2/en-4Tens of thousands of speakers, hundreds of monolingual sentences, tens of thousands of en–chr parallel sentences
Owens Valley Paiute0/3/en-3Less than ten fluent speakers, thousands of monolingual sentences, thousands of parallel sentences (the Kubishi Dictionary)

Why it's built this way

RAN is designed first as a communication tool. It is:

  • Compact: fits in an abstract, a slide title, or a table cell.
  • Multi-dimensional: separates speaker count, monolingual data, and bilingual data, which behave very differently.
  • Reproducible: apply the same rule and sources and you get the same score.
  • Interpretable: the digits are orders of magnitude.

The bilingual components also turn a set of languages into a graph — partners are edges, weighted by parallel-corpus size — which makes pivot paths explicit. Quechua, for instance, has far more parallel text with Spanish than with English, so Quechua→Spanish→English can beat going direct. Recording the largest partner rather than fixing English as the pivot is deliberate: the RAN Paper finds the best bilingual partner is the strongest predictor of cross-lingual transfer.

How scores are computed

While RAN does not prescribe which source you draw counts from, a score is only as reliable as its source. In the sample provided here, we keep scores comparable by using one consistent set of sources across every language, and each language page states its raw counts and where they came from:

ComponentSourceWhat is counted
S (speakers)EthnologueFluent (L1) speakers
M (monolingual)OSCAR 23.01Sentences in the largest monolingual corpus
Li-Bi (bilingual)OPUSSentences in the largest parallel corpus with partner Li

These are sensible defaults, not a mandate. Corpora grow, so a score is implicitly as of its sources.

Exceptions

Owens Valley Paiute isn't covered by OSCAR or OPUS, so its monolingual and bilingual counts come from community documentation (the Kubishi Dictionary) and its speaker count is a community estimate rather than an Ethnologue figure.

Explore

  • Languages: a sample of scored languages, sortable by component — illustrative, not a definitive registry. Start here to browse.
  • The RAN paper: the full write-up, including how the components correlate with benchmark performance across MT, NER, and POS.
  • Contributing: propose a language for the sample, or correct a score.
About this project

RAN comes from the Kubishi Research Group at Loyola Marymount University (PI: Jared Coleman) and is described in a short paper to appear at AmericasNLP 2026. The sample deliberately includes Indigenous languages of the Americas (Quechua, Guarani, Cherokee, and Owens Valley Paiute) that are routinely bundled under one "low-resource" label despite differing by orders of magnitude.

Created · Updated