The Zeno's Paradox of `Low-Resource' Languages

• The NLP community has focused on a small set of languages, while many Global South languages remain neglected and are labeled “low-resource.”Source: §1 · p. 1
• The high-resource/low-resource framing resembles Zeno’s paradox because less-studied languages continually pursue a research goalpost established for better-resourced languages.Source: §1 · p. 1
• Data alone oversimplifies low-resourcedness because languages also differ in speakers, non-digital archives, and available experts.Source: §1 · pp. 1–2
• “Low-resource” lacks consensus, complicating progress tracking, intervention assessment, classification boundaries, and evaluation of whether technologies meet community needs.Source: §1 · p. 2
• The authors qualitatively analyze 150 papers to examine definitions, studied languages, and classification criteria for “low-resource” and “under-resource.”Source: §1 · p. 2
• The analysis identifies four interacting aspects of resourcedness and uses real-world examples to connect them to interventions, followed by stakeholder recommendations.Source: §1 · p. 2

2 Methodology

• Keyword searches across the ACL Anthology, INTERSPEECH, and ICASSP produced a final corpus of 868 unique papers.Source: §2 · p. 2
• The analysis excluded papers treating low resource solely as missing labeled data or simulated under-sampling, while retaining studies combining simulated and actual low-resource languages.Source: §2 · p. 2
• The researchers manually labeled 541 papers and randomly sampled 150 papers published from 2017–2023 for qualitative analysis of non-simulated low-resource languages.Source: §2 · p. 2
• Researchers annotated how papers define low-resource, which languages they study, and what additional challenges they associate with low-resourcedness, using inductive thematic analysis.Source: §2 · p. 2
• Table 1 separates keyword-search results into task/domain data scarcity, simulated low-resource settings, and studies of low-resource languages.Source: Table 1 · p. 2

3 What is a ‘low-resource’ language?

• Definitions draw on four aspects: socio-political conditions, human and digital resources, linguistic and technological artifacts, and community agency.Source: §3 · p. 3

3.1 Socio-Political

• Socio-political constraints include financial barriers and limited representation in mainstream media, government, and education.Source: §3.1 · p. 3

3.2 Resource

• Resource scarcity includes limited native speakers, linguistic experts, researchers, digital access, and data suitable for crawling or scraping; large speaker populations do not necessarily remove low-resourcedness.Source: §3.2 · pp. 3–4

3.3 Artifacts

• Artifact-related low-resourcedness includes missing computational tools, pretrained-model coverage, compute, linguistic descriptions, standardization, and labeled, unlabeled, or benchmark data.Source: §3.3 · pp. 4–5

3.4 Agency

• Community agency connects resource and artifact conditions to whose values guide technology, with detached tools offering minimal utility and active engagement embedding community values.Source: §3.4 · p. 5
• Figure 2 shows diverse language-family coverage, with Indo-European languages most represented in the sample, followed by Niger-Congo and Austronesian families.Source: Figure 2 · p. 5

• Studies range from detailed monolingual accounts to broad multilingual inventories spanning 101 languages.The Gondi study describes historical, political, and linguistic context, whereas the 101-language study reports standardized metadata and available data.Source: §4 · p. 5
• The sample covers diverse language families, with Indo-European languages most represented, followed by Niger-Congo and Austronesian.Source: §4 · p. 5
• Figure 3 summarizes how criteria are distributed across the top 20 most frequently studied languages.Source: Figure 3 · p. 6, §4 · p. 6
• Lack of labeled data is the most common criterion, appearing in 39.8% of cases, while digitized text and online data appear less often.The reported shares are 1.7% for lack of digitized text and 6.9% for lack of online-available data.Source: §4 · p. 6
• Although data is the most common criterion, papers also invoke computational tools and the number of native speakers when categorizing languages.Source: §4 · p. 6

5 Why does it matter?

• Numma-guhooni and Warlpiri share a data class but differ in speakers, vitality, digital support, and available NLP tools.Warlpiri has four times as many speakers, is stable with emerging digital support, and has tools such as KirrKirr; Numma-guhooni is endangered with no digital support.Source: §5 · p. 6
• Cherokee and Kalaallisut likewise diverge: Cherokee is endangered but has vital digital support, whereas Kalaallisut is institutional with ascending digital support.Cherokee is spoken by around 2,000 people in the cited community, while Kalaallisut is spoken by about 50,000 and is Greenland’s official federal language.Source: §5 · p. 6
• isiZulu and Konkani have similar vitality, digital access, and NLP-tool availability but differ substantially in speaker population and language trajectory.isiZulu has about six times as many speakers and is South Africa’s most common first language, while Konkani has declining speakers in parts of its primary region.Source: §5 · p. 6
• Languages in the same data-availability class show drastic differences in other available resources and language circumstances.Figure 4 profiles six languages across three data-based classes, while the examples compare languages within those classes.Source: §5 · p. 6, Figure 4 · p. 7
• Socio-political constraints can reduce human resources and digital presence, which then restrict artifact creation and community agency.The paper links assimilation, colonial policies, limited funding, scarce digital data, and a shortage of native-speaker researchers across these aspects.Source: §5 · pp. 7–8
• Targeted interventions should follow each language’s specific constraints, such as preservation for endangered Cherokee, digital-data growth for Kalaallisut, and learning or translation tools for Numma-guhooni and Konkani.Cherokee and Kalaallisut differ in vitality and institutional support, while Numma-guhooni and Konkani face assimilation toward dominant local languages.Source: §5 · p. 8

6 What can we do?

• Explicit terminology should identify which resource is limited because differing definitions affect which languages programs include and serve.The paper notes that programs can select different languages even when those languages have the same amount of data.Source: §6 · p. 9
• Researchers should engage speakers, document how language characteristics limit their work, and state the criteria used to call a language low-resource.Community-led organizations can support collaborative technology building and include native speakers outside NLP when deciding which tools to build.Source: §6 · p. 9
• Conferences and academic institutions can improve representation through alternative tracks, reduced participation costs, diversified venues, interdisciplinary collaboration, and partnerships supporting documentation.The recommendations target both access for researchers from underrepresented communities and preservation of marginalized languages.Source: §6 · p. 9
• Industry, governments, and funding agencies can support language diversity through resources, accessible products, preservation policies, digital inclusion, and grants tied to community relevance.The recommendations include subsidizing community resources, prioritizing accessibility, and funding underrepresented researchers and language-specific technologies.Source: §6 · p. 9

7 Conclusion

• The paper identifies four interacting aspects of resourcedness and recommends terminology that explicitly states which resource makes a language low-resource.The authors acknowledge that a language may lack several aspects, especially in multilingual settings, but still call for detailed documentation.Source: §7 · p. 9

• The analysis does not cover definitions from all papers in all searched venues.Source: §8 · p. 9
• The study makes no quantitative claims and focuses instead on nuanced qualitative analysis with direct evidence from sampled papers.Source: §8 · p. 10
• Its language-study findings are limited to the 150-paper sample and could be supplemented by automated extraction at scale.Source: §8 · p. 10
• The sample size prevented longitudinal analysis, leaving evolution in the use of ‘low-resource’ for future work.Source: §8 · p. 10

• The section defines labels for language vitality and digital access used in Figure 4.Source: A Labels for Classifying Languages · p. 20

A.1 Vitality

• A.1 Vitality uses an Ethnologue scale derived from the Expanded Graded Intergenerational Disruption Scale.Source: §A.1 · p. 20
• The vitality labels range from Institutional and Stable through Endangered to Extinct, based on language use, transmission, and institutional support.Source: §A.1 · p. 20

A.2 Digital Access

• A.2 Digital Access classifies languages from Still and Emerging through Ascending, Vital, and Thriving as digital support expands.Source: §A.2 · p. 20

B Criteria used in Studying Languages

• Figure 5 shows the distributions of criteria used to categorize languages as low-resource, while Figure 6 breaks down perspectives on lacking datasets.Source: B Criteria used in Studying Languages · p. 20

C Most frequently studied languages

• Swahili and Telugu lead the sample with 14 papers each, while Indian languages and African languages each contribute seven represented languages.Source: C Most frequently studied languages · pp. 20–21
• The figures present distributions of low-resource criteria and the number of papers per language for the top 20 studied languages.Source: Figure 5 · p. 21, Figure 6 · p. 21, Figure 7 · p. 21

D Categories used to define low-resource

• The paper groups studies by whether they use resources, artifacts, or diverse criteria to define a language as low-resource.Source: D Categories used to define low-resource · p. 21

Resources

• Resource criteria include native speakers, online presence, language experts, and linguistic knowledge.Source: Resources · p. 21, Artifacts · p. 21

Artifacts

• Artifact criteria include data and technology, covering the materials and computational tools available for a language.Source: Artifacts · pp. 21–22